Tree species classification is a challenging task, especially considering the classification based on two-dimensional digital images. We believe bark images have more consistency for the classification because of the usual consistency of the tree trunk pattern over different seasons.

In our study, we utilized data from a Sony (ILCE-7M2) camera with lens: E PZ 16-50 mm to analyse four species of trees in Slovakia. The dataset consisted of 1369 cropped images (cropping the entire bark) and 527 precisely cropped images (small part of the bark) of Fagus sylvatica L., Quercus petraea (Matt.) Liebl., Picea abies (L.) H. Karst., and Abies alba Mill.). However, the ordinary photographs without cropping from the Slovak dataset were not used in the research due to additional markings and numbers on the barks.

Initially, GLCM was used as a feature extractor, and the features are used with different machine learning algorithms, including Support Vector Machine (SVM). However, it failed to provide better accuracy on SVM. Therefore we implemented Convolutional Neural Networks as a feature extractor to investigate the possibilities of boosting SVM with feature selection and Bagging techniques. The research followed by pre-processing the dataset and applying feature extraction. Then applying feature selection (Lasso, Genetic Algorithms, etc.) and bagging. The final results show that SVM algorithms are provided higher accuracy of 95% and a minimum of 84% after these. The normal method's accuracy was 67% on average. In future work, we plan to develop own kernels for SVM to boost accuracy.

When we successfully identify the best approach to classifly the tree species, we will work on the implementation to the workflows of terrestrial laser scanning, mobile laser scanning and terrestrial photogrammetry.

This abstract is based upon work from COST Action 3DForEcoTech, CA20118, supported by COST (European Cooperation in Science and Technology) and APVV-20-0391, IGA-3168.

DendroNetwork is a research and monitoring network that provides datasets of tree growth and tree water deficit with a high temporal and spatial resolution. It covers around 100 sites and main tree species across a large climatic gradient in the Czech Republic. The study locations include evergreen coniferous and broadleaf deciduous trees commonly found throughout Europe: Picea abies, Pinus sylvestris, Fagus sylvatica, and Quercus spp. In parallel with the measurement of microclimatic and soil characteristics, the monitoring is based on direct observations of the growth and stress response of trees in real time based on the stem dendrometer reading (http://dendronet.cz/).

A complete forest census within a square 30×30 m was performed with the Field-Map technology at each site, when tree positions and DBH values were recorded. In this study, we paired field data with airborne laser scanning (ALS) data, specifically with an estimate of above-ground biomass (AGB) derived from ALS data using an area-based approach. We tested the robustness of AGB models. We applied a model trained on ALS data from other sites in the Czech Republic to DendroNetwork sites, achieving a very good agreement with a median relative error equal to 10 %. Conversely, when we applied a model trained on the DendroNetwork sites to ALS data from five larger forest sites across the Czech Republic, we achieved relative errors ranging from 13 % to 27 %. With these exercises, we can conclude that our models for forest AGB are robust and applicable to various forest species compositions in the Czech Republic / Central Europe.

More laser scanning activities, including terrestrial systems, are planned at DendroNetwork sites under the umbrella of 3DForEcoTech COST action. The sites with their unique field reference can serve as a great network for calibration/validation/benchmarking studies.

Forests play a pivotal role in mitigating climate change but are under increasing pressure from environmental stresses. In Central Europe, drought episodes have increased over the last decades and may lead to hampered forest productivity and increased tree mortality in the future. Forest managers can reduce the risk of drought related tree mortality by controlling tree density in stands, leaving more available soil moisture for remaining trees. In this study, we investigate spatial relationships between available soil moisture and forest structural parameters representing tree density.

Soil moisture content was measured along transects in a Douglas fir stand that was thinned in 2019. Measurements were repeated on three summer days with varying meteorological conditions. Additionally, we derived forest structural parameters (canopy cover, distance-weighted density, local basal area, and plant area index) from terrestrial laser scanning point clouds. We then analysed how these forest structural parameters correlated with soil moisture content along the transects.

Preliminary results showed significant inverse correlations between all forest structural parameters and volumetric soil moisture content. On the driest day, volumetric soil moisture content was on average 5 percent point lower than on the wettest day, but spatial patterns of tree density and soil moisture were consistent between days with different meteorological conditions.

Our findings show that forest structure can drive spatial patterns in soil moisture. This is useful information for forest managers who aim to reduce the risk of drought-induced mortality, or to promote the growth of specific trees. Furthermore, we propose that TLS has the potential to better quantify forest structure and thereby improve our understanding of forest structure as a driver of drought stress in forests. However, this requires further research on 3D forest structural parameters.

The GEDI Forest Structure and Biomass Database (FSBD) will soon be available for download from the Oak Ridge National Laboratory Distributed Active Archive Center (ORNL DAAC). This database represents the publicly available portion of the data used to calibrate the biomass algorithms for GEDI L4A Footprint Level Aboveground Biomass Density data product (additional privately managed datasets were also used in GEDI L4A calibration but cannot be released on ORNL DAAC). The GEDI FSBD is composed of forest inventory datasets from 82 globally distributed sites which were assembled by an international community of researchers. While these field data were collected using different protocols, we have performed extensive standardization and quality control so that the GEDI FSBD can function as a single coherent dataset. In addition to tree- and plot-level biomass estimates, each site had airborne lidar data acquired coincidentally with the forest inventory data. We used this airborne lidar data to produce simulated GEDI waveforms within each field plot, and the relative height metrics from these simulations are included in the GEDI FSBD.

While the field plots contained in the GEDI FSBD have been selected to match the GEDI instrument’s 25-m measurement footprint (the FSBD has ~8500 plots of this size), the database was also designed to accommodate calibration and validation of spaceborne biomass estimates from instruments measuring at a variety of resolutions, such as NISAR (100-m resolution, ~700 plots of this size in the FSBD) and BIOMASS (200-m resolution, ~40 plots of this size). Finally, the GEDI FSBD is planned as a living database; it currently contains data collected as recently as 2021 and we will continue to update it as additional suitable data becomes available.

A tree’s ontogeny manifests the situational, environmental and management history through which it has lived; from stunted individuals in temperate rainforest to sprawling colossi in royal parks. Therefore an ancient tree’s structure can strongly diverge from archetypes that younger specimens may conform to. The SCATTER project has captured, in detailed 3D, a digital inventory of a sample of the UK’s unique ancient oak trees. Using terrestrial laser scanning, 25 trees were captured that capture a range of structural diversity, situational constraints and management histories. Structural metrics were modelled using QSM methods where novel techniques were developed to cope with non-cylindrical branching, particularly in the stem. This database has allowed us to measure the exact shape, size and form of a tree, from the ground to branch tip as well as elicit insight on behaviours such as retrenchment. Project outputs are an unprecedented set of detailed measurements and analysis of UK’s ancient oaks and an open-access online digital archive. We hope this project will allow other groups to study ancient trees, raise awareness of these precarious life forms in a changing climate as well as provide a permanent record once trees have died.

Diameter at breast height (DBH) is crucial in forest inventory surveys, as it highly correlates with a tree's age, volume, and height. Traditional measuring methods are calliper or tape. Recently, modern technologies such as photogrammetry, terrestrial laser scanning (TLS) and handheld laser scanning (HMLS) have come to the fore. This paper examines the use of HMLS for DBH estimation.

HMLS can provide highly detailed and accurate information about the forest structure and composition, including tree heights, diameters, and volumes. The use of HMLS in forest inventory is particularly useful for large-scale surveys, where traditional ground-based inventory methods may be too time-consuming and costly. Compared to TLS, HMLS provides information in real-time and is much less time-consuming.

This study compares the accuracy of Geoslam Zeb Horizon and Stonex X120Go MLS devices for measuring DBH. Geoslam has been around since 2012 and has been widely researched, while Stonex is newer and less studied but has a significantly lower cost. In our experiment, we collected data on five research plots (30x30m), with different trees age and species compositions. Altogether, 436 trees were located on research plots. Pre-processing was done in the scanner manufacturer's software, followed by processing in the Forest Structural Complexity Tool (FSCT) to extract DBH.

Geoslam was easier to use and more reliable in the field. Stonex requires an Android mobile device with an app but allows real-time progress monitoring. DBH estimation from both HMLSs was comparable in accuracy. The RMSE varied from 2.9 cm to 5.1 cm for Stonex and from 1.9 cm to 5.1 cm for Geoslam. The DBH did not differ significantly between these HMLS scanners.

This poster is based on work from COST Action 3DForEcoTech, CA20118, supported by COST and by Slovak Research and Development Agency projects APVV-22-0001 and APVV-20-0391.

With changing weather patterns increasingly dryer summers have resulted in more frequent wildland fire occurrences, even in formerly not fire prone forest systems in central and norther Europe. A prerequisite for fire occurrence is the presence of burnable materials, the fuels. Fuels can be classified according to their vertical position. We differentiate ground fuels (humus, tuff), surface fuels (shrubs, woody debris, logs) and canopy fuels (tree canopies). Much research using remote sensing technology has focused on canopy fuels. In contrast, remote sensing research on surface fuels has been scarce as established optical remote sensing approaches from satellite or UAV platforms are hardly able to collect information from below the canopy.

In a new research project, we want to increase the understanding of the small-area spatial patterns of surface fuels. Fuels are mapped using 3D and RGB data and results are compared to traditional field sampling methods.

For this purpose, we use a) a terrestrial laser scanner for the acquisition of 3D point clouds and b) RGB image mosaics for a continuous representation of the forest floor and the understory. For reference, traditional field sampling is carried out in subplots. Sampling will take place in pine forests in Brandenburg, Germany, during summer 2023.

The research objectives are a) creating a detailed continuous map of different surface fuels using deep learning to b) explore the capacities of remote sensing methods compared to traditional, time-consuming field sampling and c) relate patterns detected by remote sensing to fire behavior models and ecological processes.

At the SilviLaser conference we will present the developed methodical work-flow and first results from the field campaign.

Losses in tree vitality are often assessed by visually estimating crown dieback. These estimates are subjective and biased. Single tree point clouds can be processed into Quantative Structure Models (QSMs) that provide accurate insight into tree branching patterns. In contrast to using attributes calculated directly from laser scans, to our knowledge there are no studies adressing tree vitality by relating it to tree branching patterns using QSMs. We used point cloud data of trees from drought-affected forest stands classified according to the degree of damage. Because fine branches die under drought stress, starting at the crown tip, we expect trees in less damaged stands to have a higher percentage of their fine branches in the upper crown part when compared to trees in more damaged stands, which in turn have a higher percentage of fine branches in the lower crown parts. The ratio of fine-order branches to base-order branches should also be greater in trees in less damaged stands, especially within the top crown. To test this relationship, we compare the average vertical branching distribution pattern of two drought-damage classes (healthy to slightly damaged vs. moderately to severely damaged; 60 trees per class). Our algorithm computes the average of the branch order-dependent branch lengths in one-percent height increments for all tree crowns in a given damage class. Total length values are examined, but also the relative distribution across different tree crown heights. We found a distinct relationship between the damage class and the distribution of branch lengths across the crowns’ vertical profil, as well as between the damage class and the ratio between the cumulative length of higher order branches (fine branches) and base-order branches (main branches). We consider our work as a first step towards an objective assessment of the vitality of individual trees using laser scanning data.

Forest ecosystems are crucial for terrestrial biodiversity, water and nutrient cycling and mitigating the effects of climate change through carbon sequestration. As such, it is vital that forest structure and dynamics can be monitored efficiently. One useful metric for understanding forest ecosystems is tree height; however, it is often challenging to gather information on tree heights because manual measurements are expensive and time-consuming. Google Earth Engine (GEE) [1] is a computing platform providing free and easily accessible remote sensing datasets, including 3-dimensional measurements from the Global Ecosystem Dynamics Investigation (GEDI) mission. These data can be used to estimate tree heights, but the spatial resolution of GEDI data is coarse and there are significant data gaps [2]. Here, we present Tree Height Mapper, a GEE application that estimates global tree heights at 10-m spatial resolution. We downscale GEDI data combining active and passive remotely sensed data (Sentinel 1 and 2) with topographic variables (elevation, slope, and aspect) to produce a high-resolution tree heights map for a user-defined area of interest. In contrast to static maps, the application can dynamically incorporate the most recent GEDI data measurements. Users can also set the time acquisition of remotely sensed data and use dynamic forest masks. Tree height estimates can be derived using one of three algorithms: Random Forest, Gradient Boosting and Classification and Regression Trees. Due to its simplicity and versatility, we hope Tree Height Mapper will appeal to a wide range of users, from forest managers to ecosystem modellers.

Coastal winds are known to have significant effects on tree shape and growth in coastal forests and cause them to bend. These effects vary relative to the distance from the coastline. Observing such bending tree patterns can provide valuable information for understanding coastal forest dynamics. Terrestrial laser scanning (TLS) can obtain accurate and dense three-dimensional point clouds, which enable us to analyze the mechanism of tree bending. We present a novel algorithm to automatically detect wind load using the curvature on the bending tree trunks. The curvature is captured from center points of tree trunks through the partition of the three-dimensional point cloud into traverse sections for each tree height. Then, the curvature rate of tree trunks is measured by tracking the center points. For the fieldwork, the transect sampling plots were placed perpendicular to the coastline to collect TLS scanning data and several forest sites were chosen from the coastal plantation forest of Japanese black pine (Pinus thunbergii) along Boso peninsula of Chiba Prefecture, Japan. The accuracy of TLS based curvature measurement is validated with field data collected by the total station. The results of the estimated wind load data from the curvatures are validated with meteorological stations closest to each research site. The characterizing bending trees can indicate constant wind movement in coastal forests. The results showed that the trees closer to the coastline are more susceptible to bend relative to the inland trees and the directions to bend is related with the aspect of wind to the coastline.

The influence of climate change on forests is evident, but there is a lack of detailed understanding how stress affects structural and biochemical characteristics. Tree physiology has the ability to identify specific traits serving as indicators for vitality and presence of stressors. Recent developments of advanced optical and active remote sensing sensors in combination with UAS (unoccupied aerial systems) makes long-term observation of vitality patterns feasible. But to what extent biochemical leaf/needle characteristics are detectable with biophysical-spectral methods and provide plausible correlations is still unclear.

Our study site, situated in a TERENO network in North-Germany, comprises about one hectare of mature, mixed forest (oak, beech, larch, spruce and douglas fir). At its centre a crane platform was established (45 m above ground) that enables 3-dimensional crown volume access. This installation allows sampling of biomarkers and spectral measurements at different tree vitality states under controlled conditions with simultaneous remote sensing recordings. It is part of the interdisciplinary, five-year project FeMoPhys, aiming to develop a remote sensing-based monitoring procedure based on a physiologically-founded vitality assessment of tree species in mixed stands.

Here, we focus on the analysis and comparison of GNSS-referenced UAS- and terrestrial laser scanning, UAS-based multi- and hyperspectral data with the in-situ measurements, e.g. handheld spectroscopy, stomatal conductance, leaf area index. Our data basis will be the vegetation periods of 2022/2023, including leaf-on and -off forest conditions. LiDAR point clouds are co-registered and processed, where standard LiDAR products of forest structure are derived.

Further, multi- and hyperspectral vegetation indices at single-tree, branch and leaf level are extracted. This multi-source dataset is evaluated against the first in-situ results of stress and vitality indicators. With the knowledge gained, we will contribute to a better understanding of forest ecosystems under a changing climate and improve forest management strategies.

Tropical forests vary widely in their structure and thus in their biomass carbon stocks, and this variation is of great basic and applied interest. However, we lack widespread high-quality data on tropical forest structure and biomass because traditional estimates based on tree censuses have inherently large errors due to variability in tree allometry. Mobile laser scanners offer the potential for relatively rapid collection of high-quality three-dimensional data on forest structure and biomass. We quantified the performance of the GeoSLAM ZEB Horizon mobile laser scanner in a moist tropical forest in Panama, evaluating different methods for collecting and processing the data, and comparing results with terrestrial laser scanning collected with a Riegl VZ-400 in six 0.36 ha plots. We tested various walking routes, obtaining good performance with a route that included parallel transects 5 m apart, connected every 5 m to “close loops”. With this design, field data collection in a well-marked plot took one person 15-25 minutes per 20 x 20 m quadrat. In comparison with terrestrial laser scanning data, mobile laser scanning data were considerably noisier and benefited from the application of a denoising filter. Nonetheless, with careful choice of data collection and processing methods, the mobile laser scanning data were able to capture the intricate details of the forest understory and enable precise measurements of tree locations, trunk diameters, and understory wood volume, although data quality degraded towards the upper canopy. We present statistics on systematic and random errors of the mobile laser scanner in measuring trunk diameters, stand basal area, canopy height, and wood volume relative to terrestrial laser scanning measurements. We conclude that mobile laser scanners are a promising tool for rapid collection of understory forest structure data in dense tropical forests, but results are highly sensitive to the details of data collection and processing.

European beech (Fagus sylvatica L.), the predominant tree species in Europe's natural forest vegetation, regionally shows partial to complete canopy dieback due to severe droughts in 2018/19. This suggests that vulnerability to extreme droughts is much higher than previously thought in this ecologically and economically important timber tree species. However, a clear pattern has not emerged, as a high heterogeneity exists within forest sites, with individuals showing strong drought effects alongside vital trees. There are numerous interacting abiotic and biotic factors affecting trees on small scale. We studied more than 20 forest sites along a climatic gradient in Bavaria, Germany, focusing on the position of trees in relation to relief, neighborhood and their structure. We hypothesize that high competitive pressure negatively affects tree vitality during drought periods. At the same time, we assume that larger canopy gaps to the south of the target trees and the location on southern slopes may lead to additional drought stress. Tree competition is seen as one of the most important factors as it can also be influenced by silvicultural measures. Therefore, traditional approaches based on measurements of tree height and distance to neighbors are used along with forest specific metrics derived from airborne LiDAR data, further combined with high-resolution mobile laser scanning data. Using these datasets, we aim to quantify competition for light based on actual tree shapes, also considering their topographic position. The data is additionally used to quantify structural complexity of the trees via the box-dimension (D_b). Morphometric and hydrographic indices were also used to describe the microsite-effects on tree vitality. In addition, the proportion of canopy gaps around each target tree is quantified for each cardinal direction to examine the influence of gap size in specific directions. The correlations between different indices, box-dimension and tree vitality (leaf loss) will be presented.

GEDI data sample forest structures at very high density over large portion of the Earth surface and could contribute to the development of more precise and accurate estimation of forest attributes. One key issue with the use of GEDI data is the difficulty to get spatially matching field reference data for both the calibration and validation of models.

A solution to bridge the gap between GEDI and National Forest Inventory (NFI) data consists in using continuous auxiliary information. Here we propose a solution based on a k-nearest neighbour (kNN) bagging approach. Bagging consists in sampling without replacement, with a sample size defined as a function of the population size.

The approach was tested over the Vosges Mountains in France using Sentinel 2 images and a digital elevation model (DEM) as continuous data aggregated at 30 m resolution. The data vector included 10 GEDI RH values, Sentinel mean spectral bands and DEM derived metrics (i.e. mean altitude, slope, and aspect). The bagging experiment was repeated 1000 times with a sample size equal to the squared root of the population size and with kNN set with Euclidean distance and k = 1. For each 30 m cell, the GEDI footprint with the Rh profile best fitting the mean bagging profile was selected.

Following the spatial matching, a conventional modelling approach was applied to predict volume using NFI plots. To evaluate the potential of this approach, a preliminary model was trained on 80% of the plots and tested on the remaining 20%. The model explained 57% of the variance of the field measured volume. The mean absolute error of the model was 90 m³.ha^-1 or 31.6% on the test data.

The Global Ecosystem Dynamics Investigation (GEDI) was installed on-board the International Space Station (ISS) in 2018, with the aim of providing improved characterization of forest structure. The mission, which was due to be retired earlier this year, has recently been granted an extension, with data collection resuming in 2024, potentially until 2030.

Plant area index (PAI) is one of several retrieved quantities in the official GEDI Level 2B (L2B) product suite. Validation is necessary to ensure fitness-for-purpose, but since release, few validation studies have been carried out.Using in situ reference measurements available through the Copernicus Ground Based Observations for Validation (GBOV) service, we provide an initial validation of PAI estimates from GEDI’s L2B product.

Our results show that GEDI L2B PAI retrievals represent a nearly unbiased estimate of effective PAI (PAI_e, i.e. not accounting for clumping) (RMSD < 1.0, NRMSD < 40%, bias < 0.02), but systematically underestimate PAI (RMSD > 1.4, NRMSD > 40%, bias > 0.9). We attribute this to an assumed random distribution of plant material in the retrieval algorithm. Having provided an initial assessment, further work is now needed to validate the product over additional locations and time periods.

The ForestScan project is an ESA-funded initiative to develop data collection frameworks and new reference data to generate above ground biomass (AGB) estimates at multi-ha scales. Here, we present a comprehensive package of updated TLS field data acquisition protocols which aims to facilitate the planning of TLS field campaigns and contains both theoretical and practical guidance.

This contribution reports on TLS campaign experiences from a range of tropical forest plots in Gabon, French Guiana and Ghana as well as temperate forest plots in Germany, France and the United Kingdom. These campaigns built on previous protocols but implemented new features in order to speed-up plot-scale TLS acquisitions. First, we present a guide for setting up a TLS chain sampling grid with target-less registration technologies for both the RIEGL VZ400 and REIGL VZ400i TLS scanner series which reduces the time for target placement in between scans. Second, the use of Global Navigation Satellite System (GNSS) Real Time Kinematic (RTK) positioning that enables real-time scan auto-registration was evaluated. A step-by-step guide for integrating GNSS RTK protocols to the TLS chain sampling is provided. GNSS was also used to geo-reference the TLS point clouds in global coordinate reference systems. In particular post-processing options for the GNSS data were explored. Third, we provide best practice protocols for integrating tree census with TLS data. The integration of TLS point clouds with traditional forest inventories adds species information to the structural data and thereby improves AGB estimations with species-specific wood density estimates. Finally, we provide recommendations on field campaign logistics, field team and equipment requirements that can increase efficiency in the field.

While research on interpreting airborne LiDAR points cloud is expanding, there are still many gaps in post-processing. This research uses multiple study areas (New Forest, UK, River Red Gum, AU, and Amazon Forest, BR) for investigating: the uneven footprint density of scanning patterns, inconsistency in intensities, and occlusions of laser beams that results in lower canopy being dismissed. To overcome the uneven footprint density, the open source software DASOS orthorectificates the data into a 3D space and it creates a 3D grayscale image. It then extract information from that space (previous work names this method voxelisation, but to avoid confusion with voxelisation that refers to buckets of points orthorectification is a more appropriate name). We showed that intensities values vary significantly from flightline to flightline within the same study area. Further, information for converting intensities to physical values were missing from all the LAS files and the emitted pulse was not recorded. With this information and considering that the emitted intensity changes according to the returned intensity, calibration was not possible. DASOS though normalises the intensities across the voxels and when full-waveform LiDAR are used this results to equal pulse width per voxel. Even though this is an improvement, it does not solve the variations in the emitted pulse intensities. Finally, an approach for preserving lower occluded intensities was proposed. The sum intensity of each waveform is calculated and percentages are used to improve the intensities of each waveform sample. This results into preserving returns closes to the ground that would have been discarded during lower filtering otherwise. Metrics exported from DASOS has been used for tree species classifications in a Brazilian Atlantic forest and it was shown to outperform metrics exported by lidR package. It is suspected that the post-processing and its unique metrics contributed to the improved predictions.

In Finland, forest growth potential is characterized by six understory vegetation classes, which are defined by the plant species present at the forest floor. In the most fertile classes, the understory consists primarily of herbs. As the fertility decreases, dwarf shrubs become more dominant, and at the poorest soils the fraction of lichens increases (Figure 1). Unfortunately, remote sensing of understory vegetation can be difficult due to the occlusion by the canopy layer. In this study, we applied Optech Titan multispectral airborne laser scanning (ALS) data and Landsat 8 imagery to classify n = 150 national forest inventory field plots into vegetation classes 2 (Oxalis-Myrtillus type), 3 (Myrtillus type), and 4 (Vaccinium type). We hypothesized that vegetation indices computed from multispectral intensities of single ground echoes and structural variables computed using only understory echoes with heights ≤ 4 m above ground level would be especially useful as predictors of understory vegetation types. The classifications were made using ordinal regression analysis and leave-one-out cross validation. Our first model that utilized only ground intensities and understory structure reached a modest weighted kappa coefficient = 0.28 and overall accuracy = 0.59. In the second model, also ALS variables from the canopy layer and spectral bands from Landsat 8 were used. This model reached a slightly better weighted kappa = 0.43 and overall accuracy = 0.68, but the accuracy was still not sufficient for practical use. Although there are results that various ground vegetation types such as lichens can be recognized using ALS intensities, our results demonstrated that in practice the classification of understory vegetation types can be difficult. The vegetation classes can overlap with each other, and their appearance can vary depending on canopy height and density, which makes accurate classification challenging.

Close-range remote sensing applications have gained widespread popularity over the last decades. This is true especially for light detection and ranging (LiDAR) systems. The new technology is espacially useful to model three dimensional forest structures and has been used in many scientific fields including biodiversity studies. Nonetheless a systematic overview of published articles approaching this topic still lacks. It is therefore necessary to fill this gap to uncover research trends and knowledge gaps which may help to guide future work.

The presented review investigates all articles of the last decade listed in the 'Web of Knowledge' database which use close-range remote sensing methods (especially LiDAR) to determine forest structures and correlate it to biodiversity research questions. According to the PRISMA scheme for systematic literature reviews all articles found are first evaluated based on their title, then on their abstract and at last on their full text and non-fitting ones are excluded. The remaining papers are statistically evaluated with focus on the used remote sensing method, the forest type, its structural parameters, the species group investigated and the related biodiversity variables.

The review is still in work but there are two things that may already be highlighted during this early phase. First most articles that used LiDAR to identify forest structure relied on aerial laser scanning and were therefore expelled from the literature list. Although it does not fall under the scope of this review it shows that most work did indeed not center around close-range remote sensing. Second it appears that by far the most investigated species groups are birds probably because the relationship beween structural richness of the habitats and bird abundance is known for a long time. But other species are underrepresented. A statistical evaluation of all included articles is nonetheless still pending.

High resolution remote sensing technologies have huge promise to improve our understanding of forest functioning, including tree health, ecosystem resilience, carbon estimation and assessment of habitat quality and biodiversity. Drones are a particularly attractive platform for challenging environments, because of their high spatial coverage and rapid data collection ability. However, the identification of individual trees in drone- generated remote sensing datasets remains a major barrier to large scale uptake of their use, because whilst individual tree monitoring is key to understanding many aspects of forest functioning, their delineation in 2D or 3D imagery remains challenging.

This study aims to test the ability of artificial intelligence (AI) to automate the detection of individual trees within data collected from drone platforms in forest plots in Europe. By using co-located UAV LiDAR and UAV photogrammetry, alongside high-resolution terrestrial laser scanning of individual trees as a ground-truth, the accuracy of the AI-based methods used can be fully quantified. Specifically, machine learning methods for 3D analysis will be used to perform instance segmentation of trees from the drone-generated 3D point clouds. By accurately detecting and monitoring individual trees in drone-generated remote sensing data with AI-based methods, we can enhance our ability to understand forest dynamics in the face of global change.

In recent years, the combined use of unmanned aerial vehicles (UAVs) together with light detection and ranging (LiDAR) sensor systems have shown high potential for the remote sensing based assessment of forest structures and the retrieval of tree attributes. However, target coverage needs to meet task-specific requirements and is highly dependent on the UAV and LiDAR system capabilities, as well as forest structure and flight parameters. The latter are often chosen based on empirical values, lacking a scientific framework. First studies target this issue by testing platform set-ups with respect to canopy occlusion effects, yet their results are mostly sensor and region specific and not necessarily transferable for the monitoring of central European forests and the use of low-cost LiDAR systems. This study investigates the effects of multiple flight and sensing parameters on the data completeness of vertical tree structure in a mixed forest in Germany, using a DJI Matrice 300 RTK UAV and a low-cost DJI Zenmuse L1 LiDAR sensor. The goal is to provide a scientific framework of UAV LiDAR survey parameters for structural monitoring in central European forests. Therefore, multiple data acquisitions with different combinations of flight speed, flight lines, sensor scanning geometries and sensor tilt angles were conducted and their influence on volume exploration evaluated. First results of an occlusion analysis targeting geometric shading effects are presented here.

The Above Ground Biomass (AGB) of savanna ecosystems varies spatially and temporally as a result of complex interactions of both abiotic and biotic drivers. Given the complex nature of savanna ecosystems, it is important to efficiently monitor and quantify the AGB in savannas. We implemented a non-destructive approach based on TLS and quantitative structure models (QSMs). Leaf-off multi scan TLS point clouds were acquired during the dry season in September 2015 and October 2019 respectively around the Skukuza flux tower, using a Riegl VZ1000 TLS in Kruger National Park, South Africa. The 3D data covered an area of 6.25 ha with an average point density of 315.3 points / m². Individual tree segmentation was applied on the two clouds using the comparative shortest-path algorithm in LiDAR 360 software. We reconstructed optimized QSMs for trees and the computed tree volume was converted to AGB using a wood density value of 0.9. The AGB change was derived with the TLS data for 2015 and 2019. The results of our study showed that volume reconstructions algorithms such as TreeQSMs together with high resolution TLS datasets are key to non-destructively assess biomass and its change over time. The results of this study are key in understanding savanna ecology given its complex and dynamic nature and to accurately quantify the carbon contribution of savanna ecosystems to the global carbon pool.

Understanding forest structure is critical to sound forest management practices. Forest structural complexity can be broadly defined as the diversity or heterogeneity of physical and spatial attributes within a forest ecosystem, including, but not limited to, the arrangement of tree crowns, understory vegetation, and dead wood components. Over the last few decades, forest structural complexity, along with the numerous methodologies of measuring it, has become a topic of particular interest to foresters due to its connection with forest health, among other desirable management outcomes. LiDAR can be an essential tool for quantifying complexity due to its ability to measure heights on top of and within forest canopies accurately. However, despite its widespread use, there is not yet a clear understanding of how scale impacts complexity metrics using LiDAR. In studies involving LiDAR-derived metrics of complexity, choosing an appropriate plot size is essential to balance the need to recognize fine-scale details while still being representative of the greater stand. This study compares some commonly used LiDAR-derived measurements of structural complexity across a range of plot sizes using 180 circular plots in the White Mountain National Forest and eight different radii for each plot. Through an analysis of coefficients of variation and pairwise linear regressions, it was found that the degree of variance seen in commonly used metrics of complexity plateaus somewhere between a plot radius of 10 and 20 meters, and the effects of scaling vary between metrics. This has important implications for the scale-dependency of LiDAR-derived metrics of structural complexity.

Forests worldwide are undergoing large-scale and unprecedented changes in terms of structure and species composition due to anthropogenic disturbances, climate change and other global change drivers. Climate, disturbances and forest structure are all closely linked: changes in climate can lead directly to physical changes in forest structure and vice versa or to an anticipated increase in forest disturbances. However, it is still uncertain how forest structure is impacted by disturbances (locally) and how we can detect and monitor various levels of disturbance regimes using spaceborne satellite data (globally).

Here we will introduce the framework of the SPACETWIN project. This project will focus on the impact of drought, fire and logging disturbances across a range of tropical and temperate forest ecosystems. It will lead to a step-change in our ability to observe, quantify and understand forest disturbances and recovery by using time series of the most detailed structural and radiometric 3D forest models ever built: 'digital twin' forests. The key innovations will be: (1) the establishment of an unprecedented 4D dataset across 57 disturbed sites using terrestrial laser scanning (~11,500 individual trees); (2) the development of next generation methods to enable big data science of forest point clouds; (3) the identification of key axes of variation of disturbed tree and forest structure; (4) the first ever implementation of digital twins for optical and microwave radiative transfer modelling; (5) the near-real time inversion of remote sensing of forest disturbances using emulation; and (6) the embedding of forest structure in the global observation process to understand the uncertainties in monitoring disturbances.

European countries conduct national and regional forest inventories (NFI/RFI) to monitor their forests and assess wood volume, forest structure, and diversity-related indexes. With field measurement and allometric equations they obtain information such as diameter at breast height (DBH), Tree height or wood volume. Even though the methods used are widely spread and accepted, there exist some limitations to them. Tree height measured on the field often shows positive or negative biases depending on the species (Stereńczak et al. 2019) and wrongly used allometric equation lead to imprecision in the estimated wood volume (Duncanson et al. 2015).

Mobile laser scanners offer a valuable addition to NFI, allowing precise volume estimations and field measurements (Vandendaele et al. 2022, Wu et al. 2013) as well as an opportunity to develop new indices and metrics (J.Donager et al. 2021).

During this study, we sampled 60 circular plots of 0.1 ha from Wallonia’s RFI where the DBH of every tree above 20 cm was collected manually for a little more than 4500 stems in total. The height of 15 randomly chosen trees per plot was also measured. All plots were scanned with the Zeb horizon (GEOSLAM) and 27 of them additionally with aerial unmanned laser scanner (ULS). These plots were sampled in hardwood (33) and softwood (27) forest during leaf-off period.

The point cloud produced by the MLS scan has been processed to extract DBH, total height, crown projected area, crown volume and branch and trunk wood volume.

We compared DBH estimations with field measurements, height and crown estimations with ULS measurements, and wood volumes with Dagnelie's allometric equations at both the stem and plot levels. (Dagnelie et al. 1999).

The possibility to include metrics based on the crown spatial feature and light competition was also explored.

Various commercial and research Terrestrial Lidar Scanning (TLS) instruments have been deployed over the past two decades in the mixed conifer and deciduous experimental forests of Harvard Forest, (Petersham, Massachusetts, USA) to monitor forest structure and assess change due to weather, climate, and invasive pests. Harvard Forest, with an extensive historical record of forest management, several flux towers, a NEON tower, and a periodically inventoried ForestGEO site, is an ideal location to compare and evaluate TLS instruments and to test lidar scanning techniques and strategies. Much effort over the past few years has been placed on assessing the impact of the Hemlock Woolly Adelgid (HWA - Adelges tsugae) infestation, which has decimated foundational hemlock stands across the southern and eastern US. HWA affects the lower branches first, complicating efforts to monitor the early spread of the pest by airborne and spaceborne techniques. More recently, the high-end commercial TLS, the Riegl VZ400i, has been deployed at Harvard Forest to provide baseline 3D reconstructions of mature Ash stands that are now threatened by the invasive Emerald Ash Borer (EAB - Agrilus planipennis). The scans are focused on several monitored Ash field plots that have been set up at Harvard Forest near the ForestGEO Megaplot to track the EAB infestation over time. While EAB has not yet significantly damaged these Ash trees, the infestation is widespread in the region and effects should soon become evident, necessitating repeat TLS scanning.

In addition to the invasive HWA and EAB lidar monitoring efforts, 3D lidar reconstructions of the venerable Harvard Forest Witness Tree and the multi-stemmed Feldman Oak have also been obtained to provide estimates of canopy and trunk structure for historical context, and the educational aims and outreach efforts of Harvard Forest.

Precision Forestry and its applications in sustainable forest management (SFM) are gaining importance with the growing concerns over climate change and environmental sustainability. The harvesting of precision (PH) is a notable instance of this approach that aims to maximize efficiency while minimizing the impact on the environment. Optimal allocation of assortments is a crucial component of PH, and stem taper equations are used to accurately merchandize trees based on log length and diameter. There are several taper model formulations and methods of parameter estimation in use, with the selection of the appropriate model being more important than the fitting method. Terrestrial Laser Scanning (TLS) technology provides an unprecedented opportunity to acquire high-accuracy data with reduced time compared to conventional methods. Pinus nigra, a widely used and ecologically flexible conifer species in Europe, is an important resource for wood industries. In this study, we developed a tool specific to Pinus nigra species to provide more accurate estimates of volume and enable better optimization of assortments for industrial purposes using data collected with a multi-scan approach with TLS. We evaluated the performance of three taper models from different class categories in predicting diameter over bark (dob) and stem volume specifically.

Arctic and boreal permafrost ecosystems are rapidly changing due to climate change and resulting environmental disturbances, vegetation shifts and permafrost thaw, which can be detected and characterized using remotely sensed point cloud datasets. The increasing availability of point clouds acquired by various acquisition methods such as LiDAR (airborne - ALS, UAS-borne - ULS, mobile - MLS, terrestrial - TLS) and SfM (airborne, UAS-borne) is enabling the study of these changes at unprecedented detail. However, manually segmenting and classifying point clouds in basic components such as terrain, low vegetation, high vegetation and individual trees is time-consuming and not feasible given the extensive number of large datasets. It becomes even more challenging when considering the distinct characteristics of the diverse point clouds, such as point density and sampling footprint, which are influenced by the sensor, method, and acquisition specifications.

Previous studies have demonstrated the potential of machine learning-based semantic segmentation for point cloud types with similar levels of detail, such as TLS and ULS. In this analysis, we aim to develop a deep learning-based segmentation model capable of accurately segmenting a diverse range of point cloud types, including those generated by TLS, MLS, ULS, ALS, and SfM using (multi)spectral UAS and airborne image data from the Arctic to boreal region. Our approach relies on transfer learning to reduce the time-consuming labelling process while creating a robust model that can be applied across various point cloud characteristics and high-latitude regions.

The results of this study will facilitate the application of point clouds for a broad spectrum of research spanning from quantifying erosion or subsidence to analysing forestry or ecosystem changes.

The anthropic intervention in the tropical forest landscapes can reduce aboveground biomass and carbon stock at rates close to deforestation. The Global Ecosystem Dynamics Investigation (GEDI) sensor is the first spaceborne lidar deigned for global forest biomass mapping and could be potentially used to assess the effects of degradation in carbon stock at large scales. However, the characteristics regarding the spatial and temporal resolution might limit the use of GEDI for some applications such as for monitoring small forest patches in fragmented landscapes. High-resolution images, both from aerial and spaceborne platforms, have been an alternative dataset to map forest characteristics. Differently form GEDI data, image-based data can be available wall-to-wall at a higher temporal resolution. Fusing both GEDI and image dataset might therefore enhance forest attribute estimates. Nevertheless, such data fusion has not been explored in structurally complex tropical forests. The objective of this study is investigating the potential synergies on fusing GEDI and high-resolution images to map forest attributes. Airborne lidar and field inventory data were collected in four fragments of the Atlantic Forest in the state of Minas Gerais, in the southeast region of Brazil. Available stereo images derived from high-resolution spaceborne sensors were retrieved for the study sites. These datasets will be used to derive models to predict vegetation characteristics using GEDI L2A, L2B and L4A footprint level products as reference – representing the relative height, vertical profile, and aboveground biomass density metrics, respectively. We expect that the results demonstrate the potential of integrating large footprint full-waveform lidar with high resolution images and their products to monitor degraded and fragmented tropical forests.

Accurate mapping of Above-ground Biomass (AGB) is crucial for forest management, carbon cycle sciences, carbon accounting, and natural capital industries. AGB map accuracy heavily relies on ground-based calibration datasets. In countries like Australia, where frequent fire events occur, the temporal mismatch between ground measurements before and after such events needs to be assessed. This study was conducted in a 100 x 100m plot at Tumbarumba, a wet eucalypt long-term monitoring site impacted by a high-intensity ground and canopy fire in January 2020. The study aims to (i) use Terrestrial Laser Scanning (TLS) to compare vegetation structural parameters (volume, height, and diameter at breast height) before and after the fire event, and (ii) develop models to predict structural parameters from Unoccupied Aerial System-based lidar (ULS) data taken before the fire (and calibrated with the TLS datasets). TLS data was collected in 2016 and 2021, while the ULS data in 2019. We show that although volume and height were significantly altered by the fire, the diameter at breast height remained unchanged. The ULS model captured these differences, with more accurate predictions when using the calibration dataset from vegetation structural parameters measured by TLS before the fire. We conclude that the use of TLS data can improve AGB estimation accuracy in post-disturbed sites, given that the diameter at breast height values did not change, which are used to retrieve AGB using allometric equations. To upscale AGB ground measurements to the drone or satellite scale in post-disturbed sites, it is critical to avoid temporal mismatch of calibration. Our study highlights the importance of accurate mapping of AGB and the role of ground calibrations in improving AGB estimation accuracy, especially in post-disturbed sites.

This paper explores the use of Terrestrial Laser Scanning (TLS) data to estimate tree dendrometric characteristics, namely diameter at breast height (DBH) and height in a monitoring network for spruce. The study area is located in Romania, in the western part of the Southern Carpathians, in the Țarcu Mountains. Reference data (tree positions, DBH, heights) were taken with high precision by field measurements. Tree segmentation from TLS data acquired in survey areas with a large number of trees and the presence of shrubs brings modeling problems, resulting a low estimation accuracy, therefore Hough transform was applied in DBH estimation to fit circles to existing cross sections. Thus, even when only a fragment of the trunk is available, DBH can be estimated from the curvature of the section. The processing and analysis of the TLS data sets obtained through field inventories and those based on TLS data was carried out using the FORTLS, lidR, TreeLS, packages of the R software. Parameters used in the segmentation algorithm set the condition that in a radius of 40 cm around the identified trunk does not exist another segmented trunk. The results obtained show that there is a strong significant correlation (r= 0,992) between the DBH values measured in the field and the estimated ones, with a root mean square error (RMSE) value of 1.52 cm, for the same trees, based on point clouds. Moreover, between the values of tree heights measured in the field and the heights estimated for the same trees there was obtained a strong significant correlation (r=0,945) and an RMSE value of 1.42 m. The results of this study indicate that using this methodology it was possible to estimate 98.5% of the total volume of trees identified by point cloud segmentation, respectively 91.4% of the total volume measured in the field.

The growing accessibility of Very High Resolution (VHR) commercial data affords the opportunity to gather increasingly detailed Earth observations from space. One important application of this development is in characterizing vertical structure patterns using remotely sensed data. When combined with airborne Lidar such as NASA’s Ice, Cloud, and land Elevation Satellite (ICESat-2), the addition of optical image data with high resolution (2-5 m GSD) is critical for obtaining precise vertical structure estimates, particularly in areas with highly variable canopy height gradients. Here we present methodology for the application of a regression-based convolutional neural network using ICESat-2 and WorldView Multispectral data from Maxar to estimate canopy height measurements in Senegal.

Climate crisis is threating forest ecosystems while it is increasingly becoming more severe. Ecologists are using climate scenarios to predict the future distribution of forests and identify vulnerable ecosystems. Forest vulnerability to climate change is often based on models that utilise detailed forest plot data collected in the field. Large-scale inventory plot data of forests are crucial in comprehending the interactions among species, forest structure and demography. Even though plot data are very detailed and include information such as the specie and height of individual trees, they are not easily scalable, not spatially continuous, temporally constrained, and are very time-consuming to collect. There is, therefore, a need for technological solution. While LiDAR can provide accurate 3D information, they are expensive to collect, they have still limited scalability and lack temporal resolution. The technological advancement in Earth Observation, the availability of open data and ability to manage huge amounts of data efficiently in the cloud, bring new opportunities in forest ecology that have not been exploited yet. Here, we present a new software that can manage thousands of plot locations around the globe and extract local information from various Earth Observation datasets. The data are exported into csv files for easy interpretation in statistical software and merged with plot data for training and evaluating machine learning models. The software currently supports Sentinel-1 and Sentinel-2 data, but the implementation is flexible to support more collections in the future. Information extracted include standard indices (e.g., NDVI, RVI), as well as new phenological related metrics. We further use machine learning approaches to understand which bands and indices extracted are the most important for species distributions analysis, forest patterns and temporal trends. The software will be released as open source enhancing this way the usage of Earth Observation data to the wider forest ecology community.

Natural disturbances like hurricanes can cause extensive disorder in forest structure, composition, and succession. Consequently, ecological, social, and economic alterations may occur. Terrestrial laser scanning (TLS) and deep learning have been used for estimating forest attributes with high accuracy, but to date, no study has combined both TLS and deep learning for assessing the impact of hurricane disturbance at the individual tree level. Here, we aim to assess the capability of TLS and convolutional neural networks (CNNs) combined for classifying post-Hurricane Michael damage severity at the individual tree level in a pine-dominated forest ecosystem in the Florida Panhandle, Southern U.S. We assessed the combined impact of using either binary-color or multicolored-by-height TLS-derived 2D images along with six CNN architectures (Densenet201, EfficientNet_b7, Inception_v3, Res-net152v2, VGG16, and a simple CNN). The confusion matrices used for assessing the overall accuracy were symmetric in all six CNNs and 2D image variants tested with overall accuracy ranging from 73% to 92%. We found higher F-1 scores when classifying trees with damage severity varying from extremely leaning, trunk snapped, stem breakage, and uprooted compared to trees that were undamaged or slightly leaning (<45°). Moreover, we found higher accuracies when using VGG16 combined with multicolored-by-height TLS-derived 2D images compared with other methods. Our findings demonstrate the high capability of combining TLS with CNNs for classifying post-hurricane damage severity at the individual tree level in pine forest ecosystems. As part of this work, we developed a new open-source R package (rTLsDeep) and implemented all methods tested herein. We hope that the promising results and the rTLsDeep R package developed in this study for classifying post-hurricane damage severity at the individual tree level will stimulate further research and applications not just in pine forests but in other forest types in hurricane-prone regions.

Management practice of Central European broad-leaf production forests has led to large-scale predominance of structurally homogeneous forest landscapes. This, in turn, contributes to low landscape-scale beta diversity and beta ecosystem multifunctionality.

To test whether artificially increased structural heterogeneity can enhance beta diversity and beta multifunctionality in forest landscapes, a large-scale field experiment was established in beech-dominated production forests throughout Germany, where forest structure was altered through ESBC (Enhancement of structural beta complexity) interventions. The applied ESBC interventions combine the creation of different standing and lying deadwood structures by tree removal with two variations of spatial arrangement of the intervention (aggregated and distributed), leading to a total of 15 treatments. We compare the heterogeneity of forest structure in Enhanced forest districts, where all plots received the same treatment, with Control forest districts, where all plots received a different treatment. To assess forest structure, we measured pre- and post-intervention stand structural complexity (SSCI), canopy openness, and understory complexity (UCI) with Terrestrial Laser Scanning in a yearly time series starting in 2018.

The ESBC interventions severely decreased SSCI and increased canopy openness especially in the gap-like aggregated treatments. As time went on, natural rejuvenation emerged quickly where the interventions had increased light availability, which led to a rapid increase of SSCI and UCI. On the scale of forest districts, applying different treatments to different forest patches could significantly increase the structural heterogeneity of Enhanced districts, which might have promising implications for beta diversity and beta ecosystem multifunctionality.

Protected forests provide essential resources for climate-related, ecological and social functions. To ensure the continued protection and improvement of their functional status, classification and monitoring of these areas are of high importance. In addition, legal requirements and international agreements mandate the classification and monitoring of protected areas with significant importance. Different biodiversity classification schemes (like Natura 2000) are often characterized by a great variety of factors like species compositions, vegetation structure, water availability, soil and bedrock, topography and elevation and differ amongst classes. Combining remote sensing data of different spatial, temporal and spectral resolution, e.g. airborne laser scanning (ALS), image-based point clouds (IM), Sentinel 1 (S1) and Sentinel 2 (S2) and train machine learning models for classification and monitoring shows potential for small to medium scale areas on a pixel level (Iglseder et al., 2023).

However, pixel-based classification is facing inevitable challenges with class boundaries, different spatial resolution of input data and classification uncertainties. One way to overcome these challenges is to create segments as primary observation units upstream of classification and monitoring. The presented research focuses on gathering, testing and adapting segmentation strategies based on combinations of various remote sensing data (ALS, IM, S1, S2) regarding their applicability to designate homogeneous areas for further assessment in the context of biodiversity analysis. Therefore, it is of particular interest to find remote sensing data-based features that make it possible to delineate these areas, whose internal homogeneity is defined on the basis of different factors. In this contribution, results from a study in East-Austria are presented and discussed. Preliminary results show that distance water bodies, aspect, slope, nDSM (ALS) and various S2 bands are features of great relevance.

Iglseder et al., 2023. The potential of combining satellite and airborne remote sensing data for habitat classification and monitoring in forest landscapes. 10.1016/j.jag.2022.103131

Managers seeking to treat forests to increase their resilience to wildfire and drought are increasingly factoring in the spatial arrangement of trees (horizontal heterogeneity). Measuring these patterns of tree clumps and openings is challenging, expensive, and time consuming, and planning treatments around these concepts is even harder. We have created two tools that use airborne lidar data to assist managers in these tasks. Airborne lidar data provides a powerful tool for assessing these patterns because it enables a high resolution census of overstory trees over a large spatial extent. However, processing lidar data to access information about clumps and openings and simulating post-treatment patterns still requires a high level of technical expertise. These tools are intended to bridge that gap. They analyze tree clump and opening patterns by comparing study sites against reference sites that exhibit desired conditions, simulating potential treatments to understand the impact to resilience, and allowing efficient and streamlined testing of different potential paths forward for management. RxGaming is a stand based tool that emphasizes flexibility and exploration of different alternatives and allows for rapid prototyping of treatment options. It focuses on visualization of these patterns and modeled treatments for communication with stakeholders or implementation in an existing workflow. Licosim takes the concepts of RxGaming to a landscape scale and applies treatment simulation across a project area as large as the user supplies. This facilitates understanding landscape conditions, such as where treatment might be most effective or where to reinforce existing patterns. It is also particularly useful during planning, such as alternative development in a formal environmental evaluation process (e.g. the US NEPA). We are actively seeking partners for beta testing in advance of public release.

Advances in terrestrial laser-scanning technology (TLS) and associated data processing algorithms now allow for comprehensive measurements of the structural complexity of trees, including the “box-dimension” (Db). Tree structural complexity unfolds slowly over their long lifespans, through successive iterations of stem ramification, following a fractal-like geometry. Trees seek to maximize their surface area for light capture and gas exchange, while minimizing the cost of constructing the volumetric network of stems and branches that supports the leaves. As such, the limits of tree structural complexity may approach the dimensions of the Menger Sponge, a theoretical object that maximizes surface to volume ratio through fractal geometry. Here, we used TLS to test the limits of surface to volume ratios of trees of different sizes and species, growing with various levels of restriction on light availability, in Germany and the USA, including some of the biggest trees in the state of Michigan, USA. We found that the crown surface area (CSA) to woody volume (WV) ratio of trees increases as a saturating exponential function of Db, though falling far short of the theoretical maximum box-dimension (Db = 2.72) of the Menger Sponge. The results suggest that open-grown trees (growing without neighbors) rapidly reach their maximum structural complexity at small sizes and are more structurally complex than forest-grown ones at a given size, because shading from neighboring trees reduces their structural complexity. Shade tolerant species appear to maintain a higher CSA/WV ratio at a given Db, suggesting that shade-intolerant species might approach the limits of their own structural complexity more rapidly. Also discussed are the differences in estimation of Db, CSA and WV when scanning trees with and without leaves on them.

A major challenge with Lidar data from forests is the semantic segmentation of the point cloud into stems, branches, leaves, ground, etc. In this research, we use a deep learning framework of 3D convolutions, U-NET, to segment the tree point clouds into branch and stem using supervised learning. A main challenge in fulfilling this task is to generate a large amount of data for training so that the point clouds have correct class labels for each point. There are effective algorithms for tree isolation and stem-branch segmentation that can be efficiently used to accurately label high-quality 3D point clouds.

A labelled forest point cloud can then be dynamically and systematically sampled to produce a lot of training data. The samples are drawn dynamically to keep a balance among each class, and systematically from vertical layers such as the base of the stem and crown during the whole training process. Moreover, the labelled data can be made more challenging by introducing artificial occlusions while training.

A discretization-based method is applied to convert point clouds into volumetric voxels needed in the U-NET for voxel-wise segmentation. All points within a voxel are then assigned the same semantic label as the voxel. The input for the U-NET has a limited resolution, e.g., 40x40x40 voxels of limited size, e.g., 5x5x5 cm3. As a result, the input data size is limited as compared to the whole tree, especially for the whole forest plot. To recover proper location and other information lost due the limitations and discretization, we introduce multiple feature channels to the input in addition to the occupation. These channels can include features such as height from the ground, distance from the closest stem or statistical features of points inside the voxels independent from the neighbouring voxels.

Southern pine forests are a significant source of timber for the US and are essential for sequestering carbon. Terrestrial Laser Scanning (TLS), which enables the collection of high-density, three-dimensional point cloud data to calculate tree characteristics and estimate biomass without destructive sampling, has the potential to become an important tool for effective forest management. To reconstruct the branch structure with the points from the wood component or compute the leaf area index with the leaf points, we need to efficiently and precisely segment single tree LiDAR point cloud data. Algorithms for classifying these components may be divided into geometric-based, radiometric-based, or a combination of the two. They have largely been tested on broadleaf trees with more recognized shapes. In this work, the performance of several geometric feature-based approaches for classifying wood and leaves was investigated using a TLS dataset at a site dominated by longleaf pine in the Austin Cary Forest, Florida, US. A Riegl VZ 400i scanner was used to collect TLS data, and point cloud pre-processing and manual segmentation were performed. Applying overall accuracy (OA) to assess classification accuracy, three open-source methods, including TLSeparation and two methods from the lidUrb package, have been evaluated. TLSeparation's OA is 74.6%, lidUrb-graph's is 82.7%, and lidUrb-dbscan is 83.03%. The confusion matrices demonstrate that TLSeparation can detect either leaf or wood points equally (true positive 73.7% and true negative 75.6%), while lidUrb-graph is better at recognizing leaves (true positive 87.3%) and lidUrb-dbscan is better at detecting wood points (true negative 90.7%). The results do not compare favorably to earlier studies in broadleaf forests. To improve TLS separation of wood and leaf components in southern pine forests, we will expand the sample size and look at the capacity of deep learning algorithms to classify leaf and woody components using more TLS data.

Canopy structure (leaf area index, clumping index) is a crucial driver of the forest radiation regime, and it is essential for many ecosystem monitoring tasks. We developed a method to measure canopy structure with TLS point clouds, based on existing voxel-based approaches which estimate leaf area density with ray tracing. Our method uses a voxel grid of leaf area density to estimate clumping index and a newly developed variable that summarizes canopy structure, called the forest silhouette to total area ratio (STAR_f). STAR_f quantifies the degree of self shading in a canopy with respect to its density.

We measured clumping index and STAR_f in 38 forest stands in Finland, Estonia and the Czech Republic and analyzed their relationships with forest inventory variables and Landsat 8 OLI surface reflectance. Stand level clumping, in conifer stands, was closely correlated to reflectance, which indicated that although increasingly clumped canopies have lower reflectance, they also have increased gaps which increases the contribution of the forest floor to surface reflectance. In our data, the latter effect appeared to be stronger, and showed that clumped forests are brighter overall than randomly distributed forests because more of the forest floor becomes visible from a nadir perspective.

From the perspective of radiation regime modeling, STAR_f is a directly useful variable, as it is closely related to the photon recollision probability, a concept used to model the forest radiation regime. We offer a thorough introduction to the definition of STAR_f, based on its origins as a shoot clumping factor, and we discuss how it can be interpreted, and how forests with very different appearance may, from a radiative transfer perspective, be very similar.

Plant photosynthesis mediates the exchange of water and carbon dioxide (CO₂) between vegetation and the atmosphere. Quantifying the rates of these exchange processes is highly challenging since they depend on various biotic and abiotic variables specific to the underlying vegetation type and environmental state. Detailed knowledge of forest gas exchange, however, is essential to advance understanding and quantifying the consequences of climate change on these fragile and important ecosystems. An established approach to estimating forest gas exchange is the eddy covariance technique. With vertically distributed sensors, atmospheric variation of CO₂ and water vapor are measured as proxies for the underlying gas exchange on a local scale. However, scaling up these local observations to larger scales requires several assumptions on the spatial distribution of biotic and abiotic factors. Remote sensing (RS) is a complementary approach frequently suggested for large-scale assessments of gas exchange. Particularly satellite RS offers the advantage of high spatial coverage while being non-destructive and providing harmonized observations over large regions. A notable disadvantage of RS, however, is its inherent top-of-canopy perspective that limits the sensitivity of RS data for vertical canopy heterogeneity. This study aims to gain insight into how common RS estimates of abiotic and biotic factors are related to the 3D variability of forest ecosystems. We construct 3D virtual scenes of two contrasting Swiss forests (i.e., the mixed deciduous forest Laegern and the coniferous forest Seehornwald in Davos) using LiDAR and optical RS data. The radiative transfer model DART (Discrete Anisotropic Radiative Transfer) was combined with these 3D virtual scenes to simulate abiotic and biotic factors including absorbed photosynthetic active radiation (APAR) and plant area density. We then quantify the 3D distribution of these factors and evaluate differences between the RS derived estimates of these factors with their real 3D dynamics for both structurally contrasting forests.

Predictive Ecosystem Mapping (PEM) products in British Columbia (BC), Canada infer ecological function at a site level but do not characterize existing vegetation. Thus, PEM is insensitive to increasingly frequent and severe disturbance events that disrupt or permanently alter water/energy/nutrient fluxes, species composition, and vegetation structure. These unpredictable shifts in ecological function underscores the need for PEM data to be explicitly connected to on-the-ground conditions to monitor how these shifts are manifesting across the landscape.

Research has shown that three-dimensional vegetation structure is a determinant and product of ecological function and is strongly associated with ecological indicators including biodiversity, disturbance regime, and productivity. Data from Airborne Laser Scanning (ALS) can penetrate the forest canopy, providing an invaluable structural “snapshot” capturing stand conditions. ALS data is therefore suitable for creating structural baselines by which to assess site-level ecosystem condition for monitoring purposes.

We combine an ALS dataset (density: ~40 points/m²) and 5m resolution PEM map spanning an ecologically-diverse region in BC. Using this data, we develop a novel methodology to: (a) generate cluster-based structural baselines using ALS-derived metrics; (b) identify stands (pixels) exhibiting typical/atypical structure relative to ecosystem class; and (c) assess overlap of pixels within universal cluster space. First, we generate wall-to-wall rasters for 25 structural metrics. Within ecosystem strata of subzones and site types, we apply a two-step clustering approach to the metrics to generate clusters and classify pixels using a Maximum Likelihood classifier. The resulting probability vectors are used to classify each pixel as typical/atypical within their strata. Finally, we re-apply the clustering approach on the unstratified dataset and evaluate the distribution of ecosystem classes and typical/atypical pixels within this universal cluster space. This methodology demonstrates the utility of ALS data for contextualizing site-level ecosystem function and ultimately to inform proactive forest management in an uncertain climate future.

In forestry, coloured spray paint is used for different purposes, such as for marking trees that were chosen for the next harvesting or that will be promoted by future silvicultural activities, or for demarcation of skid/cable roads. For the acquisition of tree parameters - diameters, height, and position – in a forest inventory and monitoring context, both systems Terrestrial Laser Scanners (TLS) as well as Personal Laser Scanners (PLS) have been successfully tested. When a Laser Scanner is combined with a camera, the colour information can be mapped onto the laser point cloud, allowing for an extra analysis of spectral information.

In this current project, we have tested the analysis of joint information from imagery and 3D point clouds in two different settings: (i) PLS point cloud data from a GeoSLAM ZEB Horizon was complemented with spectral information from images collected with a NCTech iSTAR Pulsar 360-degree-camera, (ii) TLS point cloud data from a RIEGL VZ-400i was combined with imagery from a DSLR camera. Major objective was the automatic classification trees that were marked with colour sprayed.

Data was collected in a forest stand with a total of 566 trees, of which 146 trees were randomly selected and marked with streaks (48 trees), dots (45 trees), and random two and three digit numbers (53 trees). The forest stand was separately scanned with the PLS and the TLS system. The classification was performed by a machine learning algorithm that used the grey value and the RGB- and hsv-values as input information. The preliminary results showed an overall accuracy of 97,7 % for the TLS data and 87,8 % for the PLS data.

The utilization of Light Detection And Ranging (LiDAR) provides new insights into forest inventories due to its ability to efficiently and accurately capture the 3D structure of forests. Accurate interpretation of the massive LiDAR point cloud and segmentation of the point cloud into key forest components such as foliage, stem, ground, and lower objects allows accurate estimation of forest attributes. These attributes include total wood volume, tree hydraulics, leaf density, and leaf angle and arrangement, etc. How to interpret the massive point cloud data remains a challenging question. Traditional rule-based or machine learning methods often suffer from poor generalization. Recent breakthroughs in supervised deep learning for 3D scene understanding provide potential solutions for addressing this issue. However, many studies have focused only on forest environments using their own data at small spatiotemporal scales and are restricted to specific ecosystems and sensor types.

A major obstacle to the development of effective deep-learning-based point cloud semantic segmentation methods in the forest environment is the lack of a dedicated large-scale public dataset across a wide range of forest scenes. To fill this gap, we constructed perhaps the first large-scale forest semantic segmentation dataset. It includes over 100 tiles and covers over 40,000 m² of diverse typical ecosystems and forest types and was built upon point clouds collected by various types of sensors (ULS, TLS, ALS, BLS, etc.). Careful manual labelling was carried out by a group of trained students and experts. We thoroughly evaluated the performance of state-of-the-art deep learning semantic segmentation networks on the dataset. The results confirmed its effectiveness in training deep learning models for segmenting points into Foliage, Stem, Ground and Lower objects. We will gradually open the dataset to the entire community to promote the development of 3D point cloud deep learning for forestry studies.

Foliage has a significant role in the interaction between a tree and its environment, but the measurement of the leaves remains a challenging task. Traditionally leaf measurements have been done manually, which is labor intensive and destructive. Remote sensing of trees with laser scanners has proved to be useful in determining foliage characteristics.

The foliage of a tree can be characterized by leaf area density, leaf orientation and leaf size distributions. Leaf area density dictates the positioning of leaves on the tree. Leaf orientation and leaf size distribution define the directions of leaf normal and the sizes of individual leaves. A significant deficiency in methods determining leaf distributions from remote sensing methods, such as laser scanning measurements, is the lack of validation data. Simulating laser scans to produce validation data presents a faster alternative, but the challenge is a generation of tree models with accurately defined foliage distributions.

We have developed an algorithm for generating foliage on quantitative structure models (QSMs) of trees by sampling the individual leaves from user defined distributions. The method builds on the previously published QSM-FaNNI method by Åkerblom et al. 2018 (Interface Focus). The distributions depend on structural variables of the QSM and are characterized with parametric functions, such as the beta-distribution. We have set the leaf area density distribution dependent on structural variables such as vertical position, position on branch and azimuthal direction seen from the stem. Additionally, the leaf orientation function can vary vertically. An example of a generated foliage is presented in the figure. The tree-leaf model produced by our algorithm can be used in a laser scanning simulation and the resulting point cloud can be exploited for multiple purposes, e.g., validating the accuracy of methods for remote sensing of leaves or developing more accurate inversion methods to infer leaf distributions.

Structural information about forests (e.g., tree heights, tree diameter, and tree density) is important for biomass and carbon stock quantification and managing forest ecosystems under current and future climate and land use change scenarios. From late 2018 NASA's Global Ecosystem Dynamics Investigation mission (GEDI mission) characterized canopy heights worldwide. Whilst the GEDI mission has provided a wealth of data so far, the satellite's footprints do not provide high-resolution spatial data on tree heights. Herein, we present an approach for 10 meters canopy height mapping through a combination of GEDI data with Sentinel-1, Sentinel-2, and topographic variables (elevation, slope, and aspect). Three machine learning approaches were tested: Random Forest (RF), Gradient Boosting (GB), and Classification and Regression Trees (CART). The study area was in two Mediterranean Italian forests characterized by a conifer stand and a multilayer mixed-species stand. The complete GEDI data acquired over the study area was considered for this study. GEDI data were divided into training (70%) and testing (30%) datasets to evaluate the prediction accuracy. We compared the agreement of the GEDI canopy height map to Airborne Laser Scanning (ALS) through a pixel-by-pixel analysis. The accuracy of the algorithms was higher for RF (adjusted R-squared: adj. R2 = 0.90; root mean squared error: RMSE= 3.09) than GB (adj. R2 = 0.83; RMSE= 3.93) and CART (adj. R2 = 0.84; RMSE= 3.16). Considering the results of all algorithms, the comparison among GEDI canopy heights and ALS data revealed moderate values of adj. R2 that ranged from 0.27 (RMSE = 5.23) to 0.48 (RMSE = 4.58) for pure stands and weak values of adj. R2 that ranged from 0.06 (RMSE = 6.87) to 0.14 (RMSE = 5.59) for multilayer mixed-species stand. We expect the proposed approach can contribute to support forest monitoring initiatives, particularly for inaccessible or remote forests worldwide.

GEDI's geolocation uncertainty may limit further science applications of its footprint-level products. Here we developed a means to rapidly evaluate and mitigate the impact of systematic geolocation error on the performance of GEDI's forest height estimates in the US. By integrating nationwide high-resolution airborne lidar data collected through the 3D Elevation Program of the USGS, we provided optimal geolocation adjustments of GEDI at per beam level and tracked their performances over time. Our results suggest that the first release of GEDI product (R01) can have large systematic geolocation errors at beam level (i.e., 50.5% of beams with an error > 20 m). Its impact on canopy height measurement could drastically vary in space and time, which in turn also offers a separate indirect method to evaluate and track geolocation performance. The second release of GEDI data (R02) has achieved a much-improved systematic geolocation accuracy which is shown to meet the mission requirement (0.2% beams >20 m and 80.8% beams <10 m) and should be able to meet requirements from many practical science applications tolerant to moderate geolocation errors.

Instead of a traditional point cloud produced by laser scanning, we use a so-called fuzzy cloud consisting of 3D Gaussian distributions. An example of such a fuzzy cloud is presented in the submitted figure. Each discrete point measurement is used as the expected value of a distribution and the variance results primarily from the beam’s divergence at the object’s range and its incidence angle on the object’s surface. Our aim is to leverage the extra information contained in the continuous distributions over a discrete point cloud, as the uncertainty magnitude which due to the incidence angle and multiple scanning positions, can vary significantly between individual points. We however note that the approach introduces a challenging cyclical relation, as the geometry is fitted to a fuzzy cloud whose uncertainty depends on the geometry itself.

For a discrete point cloud, minimising the distance between the geometry and data akin to least-squares is a logical objective for finding the optimal geometry. We extend this notion to a continuous fuzzy cloud by using the statistical expected distance. Instead of using the Euclidean distance, the squared Mahalanobis distance is used as it is unitless, scale-independent and incorporates the anisotropic uncertainty expected in laser scanning data. The geometry is approximated locally as a tangent-line for each distribution to enable an analytical solution and avoiding laborious sampling. The tangent approximation also has the benefit of being easily adaptable to other cross-sectional shapes such as ellipses or rectangles.

Urban trees play an important role in cities, providing a range of ecosystem services. Climate change, causing more frequent droughts and extreme wind events, is potentially leading to declining tree health. These changes have implications for tree stability and therefore for public safety. Regular monitoring and maintenance is therefore essential to ensure their healthy condition. 3D data capture methods such as laser scanning enable rapid acquisition of detailed models of tree architecture, which can be valuable for assessment of condition, but high costs might limit widespread adoption. This study explores the use of terrestrial close-range stereo-photogrammetry as a low-cost alternative for three-dimensional capture of tree structure for the purpose of tree stability monitoring. While previous studies focused on building models from multiple viewpoints, our approach looks at the simple scenario of two low-cost cameras installed on a single rig with a variable baseline. The aim of the reported research is to determine how complete and accurate a tree structure can be obtained from this simple setup, at a range of baselines and distances from the target tree. Imagery is processed using a structure-from-motion pipeline. The resulting point clouds are then compared with terrestrial laser scans, acquired with a Leica RTC-360 scanner, to assess completeness and accuracy. A number of tree species in the leaf-off phase were selected for the study, to compare the quality of reconstruction for a range of tree crown structures. The impact of light and background on the reconstruction was also explored. Results show that even with a low-cost camera setup, the tree trunks and major branches of leaf-off trees can be reconstructed from image pairs and the 3D models can be further improved using skeletonization. This approach is ideally suited for long-term deployment and monitoring, particularly of high-value and high-risk trees in an urban setting.

Lidar is the optimum technology for measuring bare Earth elevation through vegetation and the structure of vegetation. The current generation of satellite lidars have proven the utility of spaceborne lidar for measuring forests, the atmosphere and ice. However, the energy requirements of lasers result in limited coverage compared to other satellite technologies. This sparse sampling causes uncertainty in data products, preventing robust change detection whilst also making the data unusable in many common applications of airborne lidar, such as flood modelling and urban mapping.

For currently in orbit technology, a 30 m resolution continuous global map could be made within 5 years using 12 ICESat-2 sized lidar satellites. Whilst this would provide unique data, it is likely to be prohibitively expensive. In order to realise this more cost-effectively, potential ways to increase the lidar coverage per unit cost through technological developments have been explored, including:

• Instrument: Laser and detector efficiencies improved with new photonics
• Platform and optics: Maximise payload power and telescope area per unit cost
• Signal processing: Reduce energy requirements with signal processing
• Optimum sampling: Determine the optimum sampling in space and time for carbon change monitoring

A satellite lidar simulator was used to investigate the use of novel lidar modes. This has shown that diode laser lidar could be used from space and can provide greater coverage than currently in orbit system. In addition, studying the impact of different spatial and temporal sampling densities on carbon flux estimation reveals that accurate carbon flux mapping could be achieved with 20% spatial sampling, with at least annual temporal sampling being preferred. It is concluded that a Global Lidar System with wall-to-wall coverage is possible with current technology, but that developments in photonics and deployable optics will allow such a system to be realised more cost-effectively.

Remote sensing (RS) data has been extensively used as auxiliary data using model-based inferences for estimating forest variables. This study aims at estimating forest volume (VOL) using a two-phase approach that combines terrestrial laser scanning (TLS) and very high resolution (VHR) airborne laser scanning (ALS) data, and to estimate the uncertainty within hybrid-inference framework. The study area covers around 50,000ha located in mid-Sweden. The VHR ALS data with average point density of 593 points/m² was acquired during September 2019 in 4 systematic strips of 11km length and 0.1km width each, over the study area (design-phase). Field inventory was done in 2019 for 32 plots with 10m radius, spread randomly within these 4 strips. Individual trees were measured within each plot for height and diameter at breast height (dbh) (over bark). TLS scans were done from the center of each plot using single-scan mode to provide dbh estimates of trees. A tree crown segmentation algorithm was implemented for detecting trees in these 4 strips and ALS metrics (height percentiles) were derived for each detected tree segment. The TLS and field inventoried trees were linked plot-wise with the ALS segments based on dbh values, transformed geolocations and proximal tree top distances with 10m search radius. The VOL of the linked TLS trees () were estimated using Brandel’s VOL functions. The model, VOL_TLS=β₀+β₁X₁+…+β_pX_p, where [β₀,β_p] are the model coefficients for p ALS metrics selected as explanatory variables (model-phase), was used to relate the tree-level VOL_TLS and ALS metrics and then to predict the VOL for the entire area based on ratio-estimates from the 4 VHR strips. Finally, the uncertainty was estimated from both the phases. Therefore, this study demonstrates a workflow replacing manual field inventory data required for forest VOL estimations implementing hybrid-inference, making it less time and labour intensive and more precise.

Forest plays a major role in the ecological transition, and there is consequently an increasing need for the reliable and more continuous observation of its attributes. In order to meet these growing requirements, National Forest Inventories (NFI) increasingly rely on the complementary and supplementary information from remote sensing measurements. An example of such a trend is this research endeavor where we spatiotemporally “generalize” the valuable information about the forest structure coming from the GEDI mission by simultaneously using other remote sensing sensors and NFI measurements.

In this contribution we present a model combining clustering, classification and regression techniques, used to establish a relation between GEDI lidar profiles of the forest on one side and the spatially matching Sentinel-1 (S1) and Sentinel-2 (S1) imagery and the digital terrain model (DTM) on the other side. A supervised clustering is applied on the sample of GEDI relative height (RH) profiles creating the notion of the GEDI profile class, discriminative in terms of both the total height and the form of the RH curve. In the classification part we use a multilayer perceptron in order to be able to predict the previously defined GEDI profile classes outside of GEDI footprints, where S1, S2 and DTM are available.The regression step is as well based on using multilayer perceptrons, which are now allowing to synthesize entire GEDI RH profiles outside of footprints, before subsequently pairing them inside a class with measured RH profiles for the purpose of obtaining the predicted RH profile.

The presented contribution allows the potential utilization of the predicted class as the NFI post-stratification criterion, as well as the calibration of the transformative model linking the predicted RH profiles to a forest attribute of interest measured at the plot, whose performances will be illustrated at the conference.

Climate change is affecting the disturbance regime of forests. In the North-Eastern part of France, the increased winter temperatures favoured the development of bark-beetle populations. Together with recurring droughts, which have weakened forest health, the conifers species, mainly spruce are declining since 2018. Evaluating the amount of resources impacted by bark-beetle is challenging.

Here we proposed a solution based on the combination of two approaches. First, forest dieback areas were detected using times series of Sentinel 2 data and the ForDead method. The method consists in fitting a harmonic model on time-series of SWIR continuum removal indices starting before the epidemic and detecting anomalies resulting from it over time. Second, the amount of resource impacted is evaluated by intersecting the detected areas with a resource map derived from National Forest Inventory (NFI) data and lidar point cloud. Only the areas detected after the lidar data acquisition were evaluated, and the total and variance of volume was estimated using a model-based inferential framework.

For the year 2021, the amount of area impacted by bark beetle was evaluated to 741 ha. The amount of volume impacted was estimated to 327,359 m³ with an error evaluated at 2.4%. With the renewal of 3D point s every 3 years in France via photogrammetry, the approach could be used to develop forest monitoring systems.

NASA’s ICESat-2 with a Photon Counting LiDAR Sensor was successfully launched in September 2018. The sensor uses an advanced detection system called the Advanced Topographic Laser Altimeter System (ATLAS). The ATLAS sensor detects signal photons at high speed and is highly sensitive. However, the sensor also extracts a large amount of background photon noise coming from the atmosphere, ground, sun, or other radiation. This condition is particularly evident in forest areas. Therefore, the method of filtering noise is of great significance for any use of the data. Without human intervention, automatic machine learning can form a set of processes needed for classification, namely feature selection, model selection, and model evaluation. This method offers convenient calculation, transferability, applicability, and interpretability.

We used only 10% of the sample points for training on five datasets in the forest region and compared the performance of the classifiers. First, we conclude that the integrated learning performance generally outperforms single models, and the accuracies of all tests are approximately 90%. Then, compared to distinguishing noise photons, the optimal classifiers did better classifying signal photons from noise. The prediction performance for low signal-to-noise (SNR) datasets is better than that for high SNR datasets. Crucially, the classifiers could correct misclassified labels in the reference datasets and show good stability in
different conditions.

A new method for separation of forest signal from noise has been demonstrated, which uses only a very limited number of sample points for training, ensuring operational efficiency and training accuracy. The method would be largely unaffected by differences in topography, noise distribution, and SNR. Moreover, the classifiers demonstrated the ability to correctly identify signals considered as noise photons in the ICESat-2 Level-3A ATL08 product. Overall, our method gave improved performance to the official ATL08 product over the regions tested, and can further improve data availability.

Field inventoried data are often used as references (ground truth) in forest remote sensing studies. However, the reference values are affected by various kinds of errors, which tend to make the reported accuracies of the remote sensing-based predictions worse than they are (if error free references were used). Our work addressed the impact of uncertainties in field reference data due to measurement errors, model errors, and position errors when evaluating the accuracy of biomass predictions from airborne laser scanning at plot level. We present novel theoretical analysis methods that take the interactions of the error sources into account and allow us to quantify the impact of the different error sources. Furthermore, we show how the assessment of prediction accuracy can be corrected when field references contain errors, e.g., how root mean square error (RMSE) estimates can be adjusted to better reflect the remote sensing contribution of the estimates.

Based on data from two Swedish test sites, we show that the field reference errors have an impact on the remote sensing-based predictions. By accounting for these errors the RMSE of the remote sensing-based predictions was reduced by 6–18%. The most influential sources of error in the field references were found to be the residual errors of the allometric biomass model and the field plot position errors. Together, these two sources accounted for 97% of the variance while measurement errors and biomass model parameter uncertainties were negligible in our study.

Woody features such as trees outside forest (TOF), hedges and shrubs are important landscape components given their ecological, economic and recreational relevance as well as their contribution to carbon storage and sequestration. Aerial laser scanning (ALS) data offers opportunities to detect and characterise these woody elements at landscape scales, given sufficient point density. We take advantage of Switzerland’s nation-wide program for ALS capture, flown over a 6-year period, with planned 6-yearly repeat cover. Point density is 5 pts/m2 minimum, averaging 15-20 pts/m2. We first exclude forest area using the National Forest Inventory forest mask, and differentiate trees from other vegetation using height criteria. We model the above ground biomass (AGB) of TOF across all Swiss non-forest areas on a 25m resolution raster using a cross-validated linear regression model. The explanatory variable is per-plot (500m2) green volume calculated from ALS-derived 50cm vegetation voxels. The model is currently trained with AGB estimates derived from field inventory data (tree height and diameter at breast height) of 5589 non-forest trees on 911 plots from all bioregions and relevant land use types across Switzerland. The resulting model has a RMSE value of 2.3 T/plot, R2 of 0.70, and a 12.5% error over the full study area. Separately, additional ecological function information for vegetation features is determined using ALS derived height and structural characteristics and return intensity, as well as the normalised difference vegetation index from high resolution (10-25cm) aerial imagery. ALS vegetation points are extracted, mapped as vector objects and classified according to e.g. structural complexity and ‘greenness’, relevant for ecological functional. These classified vector objects can be integrated into Swiss-wide habitat mapping, an invaluable data source for ecological assessments, planning for biodiversity management, and understanding ecological networks. This work demonstrates the nationwide application of ALS data for multi-functional assessments of landscapes and landscape elements.

Stem frequency distributions of forest attributes have been a key source of information in tactical and long-term forest planning. We compared four inventory methods for predicting stand-level distributions of forest attributes using harvester data as reference data and predictor variables computed from airborne laser scanner (ALS) data. We predicted distributions of stem diameter, tree height, volume, and sawn wood volume using k-nearest neighbor (kNN) and random forest (RF) imputation. We compared four ALS-based forest inventory methods: (1) individual tree crown (ITC), semi-ITC, area-based approach (ABA) and enhanced ABA (EABA). We assessed the accuracies of predicted distributions using a variant of the Reynold’s error index, obtaining greatest accuracies of 0.15, 0.13, 0.12 and 0.21 for stem diameter, tree height, volume, and sawn wood volume, respectively. Accuracies obtained using the semi-ITC, ABA and EABA inventory methods were significantly greater than accuracies obtained using the ITC method. The kNN and RF techniques performed equally well for most forest attributes and inventory methods. Predicted distributions of sawn wood volume were significantly less accurate than predicted distributions of stem diameter, tree height and volume. The inventory method, stand size, and the delay between ALS and harvester data acquisition had significant effects on the accuracies of predicted distributions. This study highlights the utility of harvester and ALS data for predicting distributions of forest stand attributes. Based on the results, the inventory methods semi-ITC, ABA and EABA can equally well be recommended.

National forest inventories (NFIs) are an important source of information to assess the state and dynamics of forest ecosystems. Traditional statistical forest inventory procedures are labour intensive, while the actual coverage of evaluated plot area is limited. Many forest parameters rely on assessments based on expert knowledge. These expert assessments can be subject to observer bias and hence may lack in robustness. Close range remote sensing shows large potential for quantitative forest assessment. In recent studies, mobile handheld laser scanning (MLS) showed good performance both in terms of acquisition time and level of accuracy for the derivation of forest inventory relevant parameters (e.g. tree position, diameter at breast height (DBH)). However, extensive analysis on the robustness of data acquisition and feature extraction is needed to evaluate their suitability for an operational inclusion within the framework of an NFI.

Here, we assess the robustness of an MLS device (GeoSLAM ZebHorizon) for the operational inclusion within the framework of the Swiss NFI. On 26 NFI plots (50x50 m²) distributed all over Switzerland, covering a large range of the forest characteristics and topographical variation to be encountered in Switzerland, the handling and performance of the MLS device was tested. Additionally, the influence of different sensor operators (i.e. slightly varying trajectories) as well as the phenology, on the extracted point-cloud and retrieved forest parameters are evaluated.

Initial results show high robustness of the tested MLS technique to produce a 3D point cloud with high level of detail, even in very challenging environments. Paired with a robust, automatic feature extraction, an inclusion of the MLS technique within an operational NFI could become possible, opening new possibilities for assessing the state and dynamics of forest ecosystems.

Carbon sequestration of forest trees is essential for mitigating climate change, and careful planning of forestry-related management activities has a key role in this context. The objective should be a compromise of both timber production targets as well as carbon storage goals, and this requires a holistic planning approach on a wider landscape level. In some regions, however, the average size of holdings is small in relation to the landscape, which poses challenges when both benefits and responsibilities of forest ownership should be equally shared. This also sets high expectations for the correctness and timeliness of forest resource data to enable knowledge-based and efficient planning. In this context, the single planning unit should be a forest stand, but the most essential issue is to have unbiased forest resource estimates over a larger landscape scale.

This study focuses on the errors with respect to the chosen landscape size and auxiliary data used in forest resource prediction, presenting some of the first steps achieved in a larger project dedicated to this theme. The main interest is in the added value of 3D (ALS, photogrammetric point clouds) data in addition to the satellite images but viewing this from the perspective of the landscape size, i.e., via the gain of adding 3D information to decrease the extents of the predicted area while keeping the prediction errors similar. Additionally, we test the importance of variable point cloud densities and discuss the results from the viewpoint of forest holdings’ sizes to find out the practical significance of the potential improvement. Given that the study is only in its initial phase, the emphasis of the presentation will be on premises and early tests.

In the framework of the project 4Map4Health, multi-spectral LiDAR (MS-LiDAR) is adopted to enhance the ability of information extraction in forest regions. One of the core ideas of using MS-LiDAR is tree species classification at the individual tree level. Hence, multiple MS-LiDAR campaigns have been carried out in several European countries, including Austria. The scanning equipment is mounted under a helicopter, consisting of three laser scanners in different wavelengths. This work is to present the initial view of this MS-LiDAR data set, as well as the first result of tree species classification using this data set. Firstly, we combine three wavelengths of information by aggregating point clouds near the canopy surface. Secondly, basic multi-spectral products, e.g., NDVI, are derived and help interpret the difference between each tree species. Furthermore, point clouds within the area of the single tree crown are also involved to classify tree species by analyzing spectral information in the vertical and horizontal distribution. The current results show the potential of MS-LiDAR in this task. Compared to traditional laser scanning with single wavelength information, various behaviors of tree species are already observable in raster products of MS-LiDAR. In future investigations, more features will be discovered to facilitate and improve the accuracy of tree species classification.

Storms represent one of the most serious disaster for German forests, and are prone to happen more often with climate change. Tree movement is correlated with the passage of wind gusts of small spatial scale. Previous studies showed that to explain the sensitivity to storm of a forest stand, timber removal and selective thinning were more important than soil and site conditions or topographic variables.

This suggests that a finer scale of observation is necessary to understand the predisposition of a tree to storm damage. In particular, a tree can damp its own oscillation by friction with its neighbors or by transferring the energy to branches and twigs higher frequency agitation.

From 16^th to 21^st of February 2022, North-Western Europe and in particular the Northern half of Germany was affected by a series of powerful winter-storm (Dudley – Eunice – Franklin). In some days, hurricane force winds were recorded, and some stations measured wind gusts of 100 km.h-1.

Here, we observe the effects of the storm on trees, and in particular the role of structural traits of individual tree and neighbors in the response to this disturbance. We focus on two conifer species, Pseudotsuga menziesii, and Picea abies. The observations were made at four locations in Lower Saxony (Germany), each of them composed of four stand type: Pure P. menziesii, pure P. abies, and mixtures of P. menziesii or P. abies with native broadleaved Fagus sylvatica. In March 2021 and March 2022, all the sites were scanned with mobile laser scanner (LiDAR). The resulting point clouds were segmented into individual trees. For each tree, we compare cloud overlap, the changes in crown volume, number of branches and stem leaning. Indeed, even if there not thrown by the storm, changes in tree structure could affect their vitality on the long term.

Southern forests represent roughly 41% of the timberland area in the U.S., contributing over 50% of the roundwood products in the country. Given this importance, managers need to understand the characteristics that affect these forests’ productivity, including age. Considering ages is fundamental because this will affect the management plans, and it can be used to establish growth models. Conventional age data collection demands time, effort, and can cause tree damage. LiDAR (light detection and ranging) tools such as Terrestrial Laser Scanning (TLS) are currently used for the estimation of forest parameters and can possibly be used to address age as well. Herein, we aimed to assess the feasibility of using TLS-derived individual tree-level metrics for the modeling of ages of longleaf pine (Pinus palustris Mill.) trees in the Southern US. Field and TLS data were collected within 15 plots at the Escambia Experimental Forest in Alabama. To train and validate the models, we collected age data of 147 trees using increment borers at 1.37m height above the ground. Data processing was done in the R environment. From the LiDAR data, we excluded highly correlated metrics (|0.9| < r). We used the regsubsets function to find the best models containing between one and seven variables. The quality of each model was assessed based on the corrected Akaike information criterion. We evaluated the models using R²-adj and both absolute and relative (%) RMSE and Bias. Additionally, we conducted a leave-one-out cross-validation procedure. The final model selected contained only two TLS-derived metrics, diameter at breast height (DBH) and height, the R²-adj was 83%, RMSE = 10.6, RMSE (%) = 22.07, and bias and relative bias -0.53, and -1.10% respectively. This study shows the effectiveness of using TLS data for age prediction of longleaf pine trees in the Southern US.

[Background]China TECIS(“Goumang”) was successfully launched in August 2022 (Zhang et al., 2022). As the first forestry user satellite combining active and passive carbon monitoring, data collection has been carried out globally. It is necessary and urgent to estimate and analyze the forest parameters based on TECIS's 5-beam LiDAR waveform. [Data and method] As shown in Figure B, we obtained TECIS waveform data and ALS point cloud data from Genhe on September 2, 2022 and August 20 to September 13, 2022, respectively. This study first conducted waveform validity screening on TECIS data in larch forestry area, followed by waveform metrics calculation using waveform recognition and Gaussian decomposition results. Finally, the consistency between the maximum ALS CHM and the waveform metrics tree height estimation results was analyzed at the footprint level (Hofton et al., 2020; Li et al., 2023). [Result and conclusion] The consistency results are shown in Figures A and C below; Preliminary studies have shown that the overall tree height estimation of TECIS (N=367) has achieved good results. Using the screened rH metrics (rH5, rH15, rH30, rH60, rH95), TE, LE, the consistency results between TECIS metrics and ALS CHM maximum are: R²=0.72, RMSE=2.32 m. When the slope is below 20°, R² is 0.74; When the SNR is greater than 15, R² is 0.79. The best consistency results shown in the footprints with slope below 20° and SNR greater than 15(R²=0.81, RMSE=1.88 m); meanwhile, the consistency is poor in the conditions of large slope and low SNR. Subsequent research will continue to conduct forest parameter estimation analysis on TECIS waveform data from aspects such as waveform data collection conditions, waveform metrics, and model development of forest parameter estimation.

Close-range LiDAR devices are considered one of the tools with greater potential to enhance forest inventories (IFs) estimates. However, this statement is still under discussion, especially when the target is focused on relatively large spatial scales. This work explores the potential of terrestrial laser scanner (TLS) as sampling instrument to improve efficiency in stand-level estimates. The main hypothesis behind this work is that TLS can affordably capture variability of the target population, thus reducing errors in estimates and costs. For that purpose, we compared the performance of model-assisted and hierarchical model-based by using three sources of information inference techniques, and with field-based simple random sampling (SRS) approach. In this sense, specific data sources considered were field measurements and TLS single-scans for model-assisted inference, and UAV LiDAR coverage as the third source of data in the case of hierarchical model-based inference approach. The study case was conducted in an experimental plot of 17.2 hectares dominated by Pinus radiata and Pinus pinaster, which is a common productive forest in the study area (NW Spain). The study area covered the different states of development and silvicultural treatments for these forest systems, and we focused on stand volume and biomass estimates. Our findings showed how model-assisted techniques give smoother estimates and more reduced errors for small sample sizes. In the case of hierarchical model-based approaches, errors were significantly lower compared to those based on SRS approach, with the main advantage of generating a wall-to-wall volume and biomass map. These findings suggest that close-range LiDAR devices have an enormous potential as sampling instruments, with the added value of incrementing spatial and time scales without increasing costs proportionally. Therefore, the transfer of these techniques to realistic situations may play an important role in operationalizing the use of close-range LiDAR devices in FIs.

Forest inventories using systematic sample designs are common globally for landscape-level management-oriented inventories. A common challenge with forest inventories — regardless of sampling methodology — is remeasurement and the often-high financial costs associated with maintaining up-to-date sample data. Exploring methods to effectively balance financial costs with desired estimate accuracies for remeasurement cycles are deserving of further research. Airborne laser scanning (ALS) data coverages are becoming increasingly common over forested landscapes and their existing sample networks. Prevalence of ALS data paired with the increasing need for up-to-date inventory information merits pursuing opportunities to explore how structural metrics could be integrated to guide informed and representative inventory resampling approaches. In this study we used ALS structural metrics and three sampling methods to sub-sample an existing systematic inventory (n = 4378) at the Berchtesgaden National Park in South-Eastern Germany. Sub-sampling was performed using a Monte Carlo simulation where Latin hypercube sampling, balanced sampling, and simple random sampling methods were iterated 500 times for sample sizes from 100 to 1500 sample units. Sample statistics were then calculated for each iteration and compared to population parameters derived from the full systematic sample. Simulation outcomes indicated that Latin hypercube and balanced sampling approaches consistently delivered more precise forest attribute estimates than a randomly acquired sample, especially at smaller sample sizes. We found that incorporation of ALS metrics into sub-sampling approaches provided a means to target fewer sample units for remeasurement while maintaining acceptable accuracy and precision of attribute estimates. We indicate that structurally guided sub-sampling of systematic inventories could provide accurate and unbiased attribute estimates, thus potentially aiding in limiting financial costs associated with resampling efforts.

To guarantee the production of high-quality timber, it is crucial to apply silvicultural tending and thinning treatments in seedling stands. However, allocating the right time and place to carry out these silvicultural treatments is difficult. In this research, we examined and assessed two methodological changes to the standard area-based approach (ABA_Ordinary) that could be applied to airborne laser scanning-based forest inventories, particularly seedling stand characterization. We hypothesized that ABA with additional individual tree detection-derived features (ABA_ITD) or with edge-tree effects corrected (ABA_Edge) would be improved for estimating the mean tree height and tree density of seedling stands. The hypothesis was evaluated using 89 sample plots of single-photon laser (SPL) and linear-mode laser (LML) scanning data.

The results confirmed the hypothesis because the methodological changes enhanced the characterization of seedling stands. When we used ABA_ITD, the relative bias in tree density estimation decreased from 17.2% to 10.1% when compared to the results of ABAordinary. When we used ABA_EdgeITD, the relative root mean square error for the mean height estimation dropped from 19.5% to 16.3%. When compared to ordinary LML technology, the SPL technology offered essentially equivalent or, in some cases, improved performance in terms of characterizing seedling stands. The results indicate that when ALS-based inventories supporting forest management and silvicultural decision-making are further developed, careful consideration should be given to the tested methodological improvements.

Full paper is published in ISPRS Journal of Photogrammetry and Remote Sensing: https://doi.org/10.1016/j.isprsjprs.2022.07.005

In the past years, many studies were presented that show the opportunities of LiDAR for forest inventories or measuring forest structural parameters. The availability of different LiDAR platforms (TLS, MLS or UAV-LiDAR) provides the opportunity to choose the acquisition method that optimizes the trade-off between data quality and acquisition efficiency.It is crucial to understand how parameter estimations differ between the scanners, which will give insights into which (temporal) changes are caused by the measurement setup and which changes are true representations of vegetation dynamics. Therefore, it is important to determine whether datasets acquired by different LiDAR sensors can be inter-operable.

We present two case studies which give insight into the magnitude of differences in derived forest metrics when using different scanning systems. The first case concerns temperate pine forest, where TLS data were collected of a 50x50m plot with a 10 year time interval and MLS data were collected coincident with the 2^nd TLS data acquisition. Analysis shows that variation in point density between the TLS acquisitions leads to large differences in detected trees/ha and total plot wood volume, whereas such a magnitude in change would be unexpected given the site conditions. Normalizing point densities of the different time steps results in more stable results. The second case concerns two German NFI-like plots that were scanned with TLS, MLS and UAV-LiDAR. Data are co-aligned and biomass of the plots and individual tree parameters (height, DBH, volume) are calculated using different methods . The derived metrics will be compared for consistency and sensitivity of the methods for resampling/smoothing/upsampling of the pointclouds will be discussed. This study showcases the differences between structural parameter estimation that is due to different scanners, and proposes measures to counteract these biases. Overall, this work supports long-term forest monitoring and detection of change as required by NFIs.

Forests in the Himalayan region play a crucial role in maintaining ecological balance, conserving biodiversity, and supporting livelihoods. However, due to limited accessibility, scientific knowledge on the interplay between forest management, abiotic and biotic conditions, and forest structure is still sparse for this biome. Here, insights into relevant processes determining forest structural complexity could be acquired by assessing the structural complexity via mobile laser scanning (MLS) and identifying its key biotic, abiotic, and anthropogenic drivers. In Nepal's Annapurna conservation area, a stratified sampling technique was employed to select 69 sample plots based on altitude and precipitation. The 3D mobile laser scanner ZEB-Horizon scanned the sixty-nine 25-meter diameter plots. Point cloud data obtained were pre-processed using GeoSLAM Hub V6.1 software and post-processed in LiDAR360 V5.0 software. Wolfram Mathematica software was used to calculate the box dimension (D_b), an index that quantifies forest structural complexity. Potential influencing variables were extracted from forest inventory data (tree number, tree diversity), climatic databases (precipitation), the google earth platform (distance from closest settlement), and digital elevation models (altitude, aspect). Multiple regression was used to test for the dependency of D_b on those predictors. Overall, the model explained 53.61% of the variation in D_b. Tree numbers, north-facing aspect, and tree diversity impacted D_b positively, whereas settlement distance and altitude had a negative effect. Surprisingly, precipitation did not affect structural complexity, possibly due to the impact of other human-induced disturbances. This method provides a promising approach to obtaining comprehensive and accurate information on the structural complexity of forests, especially for rapid assessment in hard-to-access regions, which can inform effective management and conservation.

Assessing forest habitats is crucial for plant and animal species conservation and management. Fine-grained three-dimensional forest data are considered a good source of predictors for such assessments. Light detection and ranging (LiDAR) technology from unoccupied aerial vehicles (UAVs) is effective in obtaining forest structural information for characterizing forest habitats. In this study, we aimed to demonstrate the complementarity of the information provided by UAV-LiDAR metrics compared to field surveys for forest habitat assessment. Firstly, we acquired a point cloud using UAV-LiDAR technology over a 4 ha subalpine forest, primarily composed of Norway spruce (Picea abies (L.) Karst.), European larch (Larix decidua Mill.), silver fir (Abies alba Mill.), and Swiss stone pine (Pinus cembra L.), in the Latemar region (Italian Eastern Alps). Secondly, in the field, we directly surveyed all trees within the area of interest and recorded their biometric traits. Thirdly, to simulate forest habitats suitable for animal species with a limited home range (e.g., many lizard species), we extracted field (as for a classical forest inventory) and LiDAR metrics (identified through a literature review) from a hundred 20 m radius randomly distributed plots. We tested the metrics for multi-collinearity using the Pearson correlation coefficient and variance inflation factor (VIF). Then we performed clustering on both sets of metrics and compared and described the results using a principal component analysis (PCA). Our analysis showed that UAV-LiDAR metrics provide complementary information to field surveys allowing comprehensive capture of the forest canopy structure and topography that are difficult to assess in the field. In conclusion, our study highlighted quantitatively the potential of UAV-LiDAR technology for forest habitat assessment, which can provide a valuable tool for conservation and management purposes. Further research is needed to explore the potential of such technology and develop robust and standardized protocols for investigating larger forest areas.

Airborne Laser Scanning (ALS) is one of the most appealing technique for precisely monitoring forest above-ground biomass (AGB) over large areas. However, the utilization of variable selection methods for filtering ALS metrics for AGB estimation is remain a challenge due to the presence of high-dimensional, strongly correlated, and multicollinear ALS metrics. In this study, we proposed a new variable selection method termed SPV, which combines sure independence screening, the Pearson correlation coefficient, and variance inflation factor. We compared the effectiveness of SPV with stepwise feature selection (SFS) in four different types of forests using field data and corresponding ALS data collected from 1002 sample plots distributed throughout Guangxi province, southern China. Five different regression techniques for estimating AGB were constructed for each forest type. The results indicated that the variables selected by combining SPV and GAM were less collinear, more interpretable, and more comprehensive in comparison to the variables chosen by SFS and GAM in combination. The AGB estimation model using SPV and GAM in combination supported the following findings: 1) The canopy density metrics in higher percentiles showed strong effect on AGB in coniferous forests, while it had no effect on AGB in broad-leaved forests. 2) The random effect of city had significant impact on AGB in broadleaved forest, while it had no significant impact on AGB in coniferous forest. 3) Incorporating the time interval between ALS data acquisition and field measurements in GAM model significantly improved the accuracy of AGB estimation for E. grandis forest, while the time interval did not affect the results of other forests. By introducing a novel variable selection method for ALS metrics, this study has facilitated species-specific AGB estimation in southern China, providing valuable insights into monitoring AGB over large areas.

Data recorded during harvesting operations can be used as references for remote sensing applications (Maltamo et al., 2019). In this study, we used harvester head coordinates, measurements of stem diameter (DBH), estimates of stem volume, and species recorded during harvesting to train and evaluate remote sensing predictions. The study area was composed of 18 stands (ranging between 3.6 and 17 hectares) located in central Sweden (59.7°-60.0°N 14.4°-14.7°E). The area was scanned using two Riegl instruments mounted on a helicopter, generating a dual-wavelength airborne laser scanner (ALS) dataset with more than 300 points per m². Individual tree crown segments were derived from the point cloud using an algorithm developed by Holmgren et al. (2022), and the segments were subsequently linked to the harvested trees. We used k-MSN modelling (Moeur & Stage, 1995) to impute species, DBH, and stem volumes of the linked trees using 18 predictors extracted from the point cloud and a leave-one-stand out approach. Our estimates for species-specific DBH distributions and stem volumes were evaluated at the stand level for all the segments linked to harvested trees. DBH distribution errors are generally quantified using error indices such as those proposed by Reynolds et al. (1988) and Packalén & Maltamo (2008). Here, we propose the use of an index based on distribution overlap. This index has the advantage of being visually interpretable and its use is not dependent on the selection of a particular bin size. Overall, our results show that data recorded during harvester operations can provide a large amount of low-cost references that reduce the need for field plots.

Most methods for estimating forest structure parameters rely on high-quality digital elevation models (DEMs). The accuracy and precision of these estimates depend on the quality of the DEM. Airborne LiDAR is commonly used to obtain high-precision DEMs of forest understory, but obtaining high-resolution elevation data in areas lacking airborne LiDAR is a challenging task. Although optical satellite data can provide DSM for such areas, it is difficult to obtain dense understory elevation data due to forest canopy cover. In contrast, spaceborne LiDAR can penetrate the forest canopy to provide footprint level ground elevation data. Therefore, we propose a method to extract dense forest understory DEM from satellite stereo imagery DSM by combining the ground point data obtained from satellite-based LiDAR. The proposed method was validated in two study areas at six study sites with different elevation, slope, vegetation height, and vegetation cover. In one study area, the proposed method achieved a correlation coefficient of 0.92-0.96, root mean square error of 3.99-4.66 m, and bias of 1.04-1.56 m, while in the other study area, it achieved a correlation coefficient of 0.97-1, root mean square error of 7.64-8.88 m, and bias of 0.06-1.49 m. The proposed method demonstrated its applicability under high vegetation cover and high topographic relief conditions.

Terrestrial laser scanning has revolutionised forest ecology by facilitating the collection of data that best represents the three dimensional space that trees grow, compete and ultimately die within. At the core of forest dynamics, competition for light encompasses all of these dynamics and plays a key role in defining the overall 3D structure of a canopy. Until now, measures of competition between trees have been restricted to 1D and 2D proxies but we show how lightweight ray tracing applied to TLS point clouds can potentially unlock key interactions between individuals that vary within a day, between seasons and across biomes in European Forests with drastically different environmental constraints. I will showcase the potential of these data and complexities in distilling this rich information into single holistic measures at the individual level.

Emergence of close-range sensing technologies have revolutionized forest mensuration by enabling the characterization of trees at an unprecedented level of detail. However, generating knowledge out of point clouds requires computational methodologies capable of extracting the structures of interest and providing quantification of their characteristics. While methodologies for characterizing tree stems have been developed since terrestrial laser scanning (TLS) was first introduced for tree measurements, the characterization of trees’ branching structure has generally remained less explored. In general, the branching structure refers to a set of features such as the location, branching angle, length, and diameter of individual branches that, all together, are important when exploring the eco-physiological functioning of trees. The aim of this study was to develop a point cloud processing method for identifying branch locations as the intersections between stem surface and branches, and segmenting individual branches based on connected components to be used in further analysis of their characteristics. The methodological finding was to apply cartesian-to-cylinder coordinate transformation to enhance stem-branch separation (see Fig. 1.). The method, implemented in MATLAB and provided openly available, was validated with TLS point clouds covering a 0.4-ha mature and managed southern boreal forest stand. Performance assessment was based on visual interpretation of three randomly-sampled stem sections from each 100 randomly-sampled Scots pine (Pinus sylvestris (L.)) trees, including 1882 individual branches in total. The results showed a recall of 79%, a precision of 93% and an F1-score of 0.85 in the detection of branches located below 70% relative tree height. Based on our study, it appears that in a managed Scots pine stand, most of the branches can be detected with the developed algorithm up to the height where the stem surface can still be reconstructed, highlighting the capacity of using close-range laser scanning to characterize tree and crown structures.

Coastal forests and other ecosystem types are increasingly under pressure from human land use decisions, sea level rise, salt water intrusion, erosion, hurricanes and other forces. Recent research has successfully detected “ghost forests” using 30m Landsat imagery (Ury et al. 2021). However, the transition from forests to salt-tolerant wetlands can be a slow process that spans decades (Williams et al. 1999, Taillie et al. 2019). Over the period of transition, some areas show signs of stress and reduced productivity (Spalding and Hester 2007) that can be observed from multispectral imagery (Hamzeh et al. 2016). Our analysis confirms this, but also shows areas that have enough understory vegetation such that time series analysis of satellite multispectral reflectance is difficult to interpret. We explore forest locations on the Delmarva Peninsula, on the eastern U.S. coast, using Landsat and Sentinel-2 combined with multi-temporal lidar analysis from three historic acquisitions (2010, 2015, 2018). Low-lying areas in this region show evidence of forest degradation (including ghost forests), agricultural degradation from salt intrusion and invasive phragmites, and transitions in marshland areas. For some degraded areas, confirmed through field visits and analysis of high-resolution imagery, the signal is obvious from the multitemporal Landsat or Sentinel-2 analysis alone. Other degraded forested areas with significant understory vegetation do not show a trend in the satellite time series. However, changes in canopy height and density over time can easily be seen in the very high resolution lidar time series (e.g., maximum height of 19.8 m, 13.3 m, and 10.2 m in spring of 2010, 2015, and 2018 respectively at one site that transitioned to a ghost forest). Our results highlight the value of high-resolution structural information from multitemporal lidar for both the quantification of degradation and its detection in areas that may not be visible from satellite time series.

In boreal environments, old deciduous trees, particularly European Aspen, contribute to a rich and resilient ecosystem, providing unique ecological niches that support various wildlife, including cavity-nesting birds, insects, and mammals, and promoting overall forest health. Due to a low economic value and relatively sparse and scattered occurrence of aspen in boreal forests, there is a lack of information on the spatial and temporal distribution of aspen, which hampers efficient planning and implementation of sustainable forest management practices and conservation efforts. Additionally, standing dead trees play a significant role in maintaining biodiversity in a boreal forest, offer essential habitats for numerous species and serve as indicators of forest health. Accurate identification of tree species and standing dead trees detection is thus essential not only in forest resource mapping but also for monitoring of biodiversity in boreal forests.

New unmanned aerial systems (UAS) based remote sensing proved to be very efficient in providing ultra-high spatial and temporal resolution imagery for detailed forest properties assessment at reasonable costs. Flexibility, accessibility and customizable sensor payload enable rapid and cost-efficient data acquisition in forested regions that may be challenging using manned aircraft.

In this study, our objective was to assess the accuracy of different UAS-based sensors and their combinations in classifying Scots pine (Pinus sylvestris), Norway spruce (Picea albies), birches (Betula pendula and Betula pubescens), European aspen (Populus tremula L.) and standing dead trees at tree level using spectral and structural features derived from LiDAR, high-resolution RGB and multispectral photogrammetric point clouds. We selected 1,250 field-measured trees (250 trees per class) for our analysis, dividing 70% of them for model training and the remaining 30% for validating the model's performance.

Trees outside forests are recognized as being vitally important as an additive component of natural climate solutions. When coupled with state-of-the-art trees outside forest mapping efforts, spaceborne lidars offer tremendous opportunity to characterize structural differences between these tree communities, with eventual implications for the mitigative facets of states' nationally determined contributions. We are characterizing trees outside forests throughout India using spaceborne lidars as part of a current NASA land cover and land use change project. As a pilot study, we intersected ICESat-2 transect segments (relative heights and related variables from the ATL08 product suite) and GEDI footprints (relative heights and related variables from the L2a product) with a high-quality publicly-available map differentiating trees outside forests from natural forests in Andhra Pradesh, India. The average and median relative heights from the two sensors differed, with the relative heights reported by GEDI substantially lower in both forest community types than those reported by ICESat-2. However, data from both sensors revealed that natural forests have relative heights higher at every reported percentile than trees outside forests. The ICESat-2 ATL08 median relative height for the 95th percentile (after anomalous or poor quality data were removed) is 20.7 m for natural forests (n = 34,399) and 13.2 m for trees outside forests (n = 25,164). Trees outside forests are not only shorter but less heterogeneous, with implications for biodiversity. The ICESat-2 ATL08 median canopy openness proxy (the standard deviation of relative heights for canopy photons within the segment) is 5.5 m for natural forests and 3.7 m for trees outside forests. Expansion to other states in India is underway. While sensor-specific differences in structural metrics are being further evaluated using reference data streams, it is clear that spaceborne lidars will provide important data to characterize trees outside forests and their role in natural climate solutions.

Characterizing forest disturbances has typically focused on disturbance scale, frequency, intensity, severity. As a result, disturbance regimes often differentiate between large, stand-replacing disturbances that substantially alter initial forest characteristics, and subtler, spatially inconsistent, and temporally pervasive non-stand replacing (NSR) disturbances. The latter, trigger fine-scale alterations in forest structure that are difficult to detect and differentiate from natural mortality, competition dynamics, and growth.

Despite efforts to characterize NSR disturbances through analysis of optical remote sensing passive imagery, there remains gaps in our understanding of how forest structure can modify the severity of NSR disturbances at fine spatial scales. The availability of small footprint aerial laser scanner (ALS) products before and after NSR events offers an opportunity to more accurately account for changes in forest structure and their role in mediating or exacerbating disturbance severity.

In this research, a new approach for forest structural change detection, due to NSR disturbances, is proposed using a harmonized bitemporal ALS dataset of managed boreal forest in Quebec, Canada. Pixel-level ALS metrics are extracted describing point vertical distribution, return intensity, and a variance-covariance matrix eigendecomposition is applied to reduce variability and noise in point density and distribution. Next, ALS metrics are compared over time and used as a proxy of change. Finally, detected change pixels are attributed to specific NSR disturbances derived from conventional aerial interpretation and field plot approaches.

Ultimately, this research provides a set of sensitive ALS metrics with which to examine NSR disturbance-induced change and offers an approach to forest managers and planners for fine-scale NSR disturbance characterization when bitemporal ALS acquisitions are available.

Boreal forests are currently subject to the most rapid warming on Earth from climate change. Since the last ice-age, they have sequestered atmospheric carbon, but their current fate to act as a net sink or source is uncertain due to climatological feedback processes. Here, we estimate young boreal forest vertical growth as an indicator of resilience to climate warming by combing two remote sensing datasets. Using ~49.7 million 20 m hcan height profile geo-segments from the Advanced Topographic Laser Altimeter System onboard the Ice Cloud and Elevation Satellite combined with Landsat stand age data from 1984 – 2020 in a space-for-time substitution approach we found strong patterns of vertical-growth when calculated over a 0.5°×0.5° grid. These empirically derived estimates provide a biome wide gridded reference to constrain modeled estimates of boreal forest vertical growth which is important to provide a current reference when estimating woody above ground biomass accumulation.

LIDAR technology is able to replace inaccurate and time-consuming manual measurements of tree dimensions in forest inventory fieldwork assessment. Additional data-collection (quality of stems, regeneration) is needed and should be preferably performed while collecting the point-cloud by a backpack SLAM device.

This study introduces a method that combines voice input and LiDAR scanning technology for optimized forest inventory. LiDAR technology can extract single tree parameters, but tools for tree species recognition and damage assessment are still missing. The study uses the Stonex X120go backpack scanner with a rotating Hesai Pandar 32-channel sensor to collect point-clouds, which are less noisy compared to similar SLAM-devices.

The method combines speech input and laser scanning technology to determine tree species and other properties such as damage survey and tree quality. The user can identify tree species by speaking while walking through the forest with the scanner, and the voice recording is done by a mobile phone or tablet with real-time conversion into database entries. After post-processing the point-cloud, the high-resolution trajectory can be synchronized with the voice-based database entries.

The data collected with the laser scanner can be processed by AI technology to automatically identify tree species by considering the voice input and extracting quantitative tree parameters. Different approaches will be investigated to ensure accuracy in assigning spoken quantitative tree information.

The results of the DBH evaluation of the scans, as well as the coordinates from the trajectory, are joined with the speech input data in a database using a script that utilizes the timestamp or GPS coordinates. The combination of speech input and laser scanning technology provides an efficient and precise method for forest inventory. Overall, the study concludes that the combination of voice input and LiDAR scanning technology provides an operational and optimized forest inventory process.

Site index (SI) is a proxy for forest productivity dependent on dominant height and age, and an important variable in forest management. SI can be determined without age information if there are data on dominant heights from two points in time, along with the time interval between measurements. Previous studies have used this approach in an area-based inventory using bitemporal airborne laser scanning (ALS) data, by predicting dominant heights at two points in time. However, a time series of predicted dominant heights may provide a more accurate representation of dominant height development, and, consequently, a better estimate of SI compared to using data from only two points in time. Few studies have explored the use of time series of remotely sensed data for SI determination. With the increasing availability of time series of ALS data, there is a growing need to investigate and develop these approaches to SI determination. We demonstrate a method whereby SI is derived from a time series of three ALS datasets spanning over 23 years. By predicting the dominant height at each of the three points in time, the SI curve that matches the predicted dominant height development can be selected by minimizing residuals. The proposed method shows potential for SI mapping over larger areas using remotely sensed data.

In recent years, the interest in low-cost sensor solutions for forest applications has increased thanks to their undemanding nature. Such sensors are lightweight and user-friendly, making them suitable for robust data acquisition that can be performed even by non-experts. In particular, low-end spherical cameras have shown great potential for video acquisition in forest environments providing a full visual field coverage of the forest scene. So far, the spherical video acquisition has been limited to using a single camera. To this end, most videogrammetric methods rely on detecting visual targets to scale a point cloud built from a monocular video. Such visual targets need to be placed throughout the forest site which can heavily increase acquisition time. Moreover, the accuracy of the generated point cloud may depend on expert target distribution during acquisition and adequate target detection in the captured video.

Here, we propose a novel approach for dense point cloud generation in forests based on stereo videos. Our stereo system consists of two spherical cameras with a known baseline distance. We exploit the stereo information to build and scale a point cloud without the need for visual targets. To assess the performance of the stereo videogrammetric approach, we conducted a study on three forest plots (50mx50m) with different levels of forest stand complexity. Spherical videos acquired using our stereo equipment were imported into videogrammetry software and dense point clouds were successfully built. The resulting point clouds were compared to 1) points clouds generated from a single video, and 2) terrestrial laser-scanning data.

The first analysis reveals the high potential of our approach and its advantages over the monocular videogrammetry for point cloud generation of forest plots. The great overlap between the two spherical cameras in the stereo system allows for robust point cloud generation and tree position extraction (Figure 1).

Airborne laser scanning (ALS) can be used for accurate estimation of forest leaf area index (LAI). Canopy height, intensity and density metrics have all been applied in empirical models for LAI estimation. Here, we test if metrics derived from polar transformed ALS point clouds can increase the accuracy of ALS based LAI estimation compared to the conventional metrics. Polar transformed ALS data can be analysed similarly to hemispherical photographs to obtain multi-angular canopy gap fractions and fractions of between- and within- crown gaps (Figure 1). Empirical models were built for clumping-corrected LAI (LAI_c) and canopy clumping index (Ω_E) using ordinary least squares models and 73 field plots from southern Finland. In addition to polar metrics, we also used height- and density- based metrics as well canopy penetration indices as predictors. The predictors were selected using an exhaustive search that reports the set of predictors producing the lowest RMSE values, and the models were validated following the leave-one-out cross validation. The results showed that LAI_c and Ω_E were predicted with good accuracy (RMSE%: 20.6% and 4.3% respectively) when polar metrics were included in the pool of predictors. When they were excluded from the predictor search, the LAI_c model’s RMSE% increased to 21.8% while RMSE% of the Ω_E model remained similar (4.1%). When modelling with only polar metrics, the RMSE% values further increased to 24.9% and 6.0%. Polar metrics therefore may provide empirical models with canopy information that is not captured by the commonly used ALS metrics, but don’t seem to work as stand-alone predictors. However, ALS data with pulse density greater than 11 m^-2 employed in this study might further increase the utility of polar metrics, as the polar images can then be constructed with more details.

The utilization of terrestrial laser scanning (TLS) data in forest inventory has become increasingly popular in the past two decades. The purpose of this study was to compare the performance of circle, ellipse, and spline fits for cross-section area modeling, evaluating the influence of different modeling parameters on the cross-section area and thus volume estimation. For this purpose, 20 trees were scanned with FARO Focus3D X330 and felled to collect stem disks at different heights. The contours of the disks were digitized to provide reference data for the evaluation of the TLS-based cross-section modeling. The most accurate estimate of the cross-section area when compared to the reference cross-section area was achieved using the spline model fit, with an RMSD (Root Mean Square Deviation) and bias of only 3.66% and 0.17%, respectively, and a ratio between intersection and union of modeled and reference cross-section area of 88.69%. In comparison, contour fits with ellipses and circles yielded higher RMSD (5.28% and 10.08%, respectively) and bias (1.96% and 3.27%, respectively). The circle fit was especially robust concerning varying parameter settings, but provided exact estimates only for regular-shaped stem disks, such as those from the upper parts of the stem. Spline-based models of the cross-section at breast height were further used to examine the influence of caliper orientation on volume estimation. Simulated caliper measures of the DBH showed an RMSD of 3.99% and a bias of 1.73% when compared to the reference DBH, which was calculated via the reference cross-section area, revealing statistically significant deviations from the reference. These findings cast doubt on the customary utilization of manually calipered diameters as reference data for the evaluation of TLS data, as TLS-based estimates have reached an accuracy level surpassing traditional caliper measures.

The 3D reconstruction of woody structures in forests has become increasingly important for various fields, including forest management, remote sensing, ecology, and carbon accounting. Although numerous methods exist for reconstructing individual trees from Terrestrial Laser Scanning (TLS) data, a key challenge is developing efficient and accurate techniques for reconstructing entire forest stands.

In this study, we introduce the use of RayCloudTools software as a novel method for efficient and automated woody structure reconstruction without prior leaf point classification nor individual tree segmentation. We validate the individual tree woody volume estimations derived from these reconstructions against a set of destructively harvested trees with coincident TLS data. Furthermore, we compare biomass estimates to those obtained using TreeQSM. Preliminary results indicate comparable accuracy (RayCloudTools RMSE: 191 kg, TreeQSM RMSE: 170kg) for the Rushworth (VIC, Aus) reference dataset.

To provide additional validation of volume estimates, we use simulated LiDAR datasets generated with the open-source LiDAR simulator Helios++ of existing 3D tree models. We employ variable simulation configurations to evaluate the scalability of our reconstruction method for aerial LiDAR platforms.

We also show typical run times for tree extraction and reconstruction on 1-hectare TLS datasets. Preliminary results show runtimes for a 1-hectare plot range from 15 minutes for a tropical savanna to 45 minutes for a tall eucalypt forest.

Overall, our results indicate that RayCloudTools is an efficient and accurate method for automated forest stand reconstruction. Significantly faster than any previously developed method, RayCloudTools is particularly well-suited for large-scale forest reconstruction tasks.

Aerial lidar data are increasingly used to assess forest conditions and plan their management. However, for many potential users, processing raw lidar point clouds into usable layers of forest metrics can be a time consuming or intimidating task. We’ve developed a free, open-source tool, named Lapis, designed to reproduce the suites of forest metrics produced by other tools such as FUSION, LAStools, and lidR. It differs from these other tools in several ways. First, as a new implementation, it can take advantage of memory and processor capabilities commonly found on advanced desktop computers and workstations. On a decent workstation, Lapis can process a terabyte of lidar data in twelve hours, allowing all but the largest of lidar datasets to be processed in under a day. Second, it is designed to simplify the processing workflow. For example, pre-processing steps are minimized, or eliminated entirely on many datasets. Lapis automatically handles units and coordinate reference systems, so a mixture of data from different sources can be used seamlessly, and the output can be in any projection, to match other layers in your dataset. Lapis produces a metadata file with every run, documenting both the parameters used and a description of each output layer. Third, Lapis is designed with a simple menu driven interface to allow ease of use by novice or occasional users but allows more experienced users to customize their products. As an open source product, expert users can add their own capabilities, or verify the methods used to calculate any layer. Current outputs include common forest structure metrics (e.g., height percentiles, canopy cover), tree segmentation, intensity data, and topographic metrics. We are actively seeking users and testers for Lapis, as well as contributors to the project.

Individual tree models provide detailed information on stand growth, facilitating spatial explicit predictions for accurate silvicultural interventions. Estimating potential individual tree height, defined as the maximum individual tree height at a specific time (most commonly the end of the rotation), provides information for silviculture decisions and harvesting schemes, as well as indicating the potential economic value per tree. Tree competition plays an important role in the forest growth dynamics, however, it is rarely taken into account over large areas because obtaining the spatial distribution of individual trees and estimating their competition is both expensive and time consuming. In this presentation we demonstrate a method for generating potential tree height models in Pinus radiata D. Don even-age plantations in south-central Chile. To accomplish this we used a single survey of airbourne laser scanning data, and implemented an individual tree detection approach to first extract tree attributes, calculate competition indices, and model potential tree height. To do so, we developed an individual tree height growth model using a Chapman–Richards model form and a chronosequence of tree heights derived from the point cloud data. The developed models included the effect of intertree competition and stand-level top height (TH) on the upper asymptote of tree height growth. The results showed that using chronosequence of heights, competition, and TH resulted in accurate predictions of potential tree height (root mean square error = 2.9 m; mean absolute percentage error = 0.154%). We concluded that individual tree height growth is significantly influenced by competition, with increased competition values associated with reductions in potential height growth by 22.2% at 30 years (Figure 1).

Figure 1: Tree height growth under different competition values (CI) ranging from 1 to 6.

Due to the risks associated with the combinations of biodiversity loss and climate change especially to the predominant homogeneous forest ecosystems associated with the clearcut system, reassessments of the current forest management practices are needed. Continuous-cover forestry (CCF) is considered one of the key regimes in the ecosystem-based forest management toolbox. Research of CCF is lacking especially from the point-of-view of individual tree characteristics, and efficient methods to measure and monitor them in the complex forests are urgently needed.

Stem taper models are widely used in forestry, e.g. in forest growth calculations, modelling and simulations, forest management planning, forest carbon balance calculations. In our research we evaluated how the stem taper in mature Norway spruces (Picea Abies (L.) H. Karst.), in stands managed with single-tree selection differed from those in even-aged management. We approached the stem taper modelling with terrestrial laser scanning (TLS) data on 49 and 51 mature individuals from three canopy classes (dominant, intermediate and suppressed) in 7 selection and 14 even-aged forest stands in southern Finland, respectively. We used these data to re-parametrize the most used stem taper function in Finland (Laasasenaho 1982) to each treatment, stand, and canopy class.

Our preliminary results showed that there was no statistically significant difference in stem taper or form in the mature trees between the selective and even-aged stands. The taper function parameters established by Laasasenaho (1982) were valid in both data sets, and no significant improvement to taper and volume calculations was gained through reparametrized models using the TLS data at the treatment- or stand-levels.

Forests across the globe display a tremendous structural diversity, from open woodlands to dense, multi-layered tropical forests. The dimensions of canopies and how trees are arranged within them shape micro-environments and regional climates, carbon stocks and fluxes, and the resilience of forest ecosystems to disturbance. Here, we investigate how forests vary in their structural characteristics through a large global database of airborne laser scanning (ALS). The ALS database covers all biogeographic realms and biomes and is processed based on a standardized and openly available processing pipeline. We show that a few simple metrics of vertical and horizontal structure are sufficient to capture differences between forest types and summarize most of the variation in forest canopies across the globe. We also analyse how these metrics relate to each other and how they are shaped by environmental determinants, topographic gradients, and disturbance regimes. Our analysis provides an important complement to space-born missions and a unique look at fine-scale forest structure at global scales.

The accurate estimation of forest biomass is crucial for various applications, e.g., carbon stock assessment, forest management, and climate change mitigation. Terrestrial LiDAR scanners (TLS) have become increasingly popular for obtaining precise measurements of forest parameters at plot levels, including biomass. Obtaining biomass estimates from TLS data requires the use of leaf-filtering algorithms. However, the performance of available algorithms for various data and site conditions is unknown. Therefore, the aim of this study is to analyse the performance of different algorithms.

Using 95 trees from five different site conditions (Cameroon, Guyana, Indonesia, Peru, and Germany), we compare four commonly used leaf-filtering algorithms: LeWoS, TLSeparation, CANUPO, and Random Forest (RF). We compare the algorithms in terms of computational efficiency, point-wise classification accuracy, and QSM-based volume extraction. We also investigate the influence of varying point cloud densities on achievable point cloud classification accuracy and the tree volume derived from the classified wood points.

Our results show that the RF model outperforms other leaf-filtering algorithms in terms of pointwise classification accuracy, volume comparison, and resistance to point cloud density changes. TLSeparation had relatively the lowest point-wise classification accuracy, while LeWoS performed poorly in volume comparison and was most sensitive to variations in point cloud density. The performance of all algorithms decreased with decreasing point cloud density, demonstrating the importance of collecting data with a higher point cloud density. Furthermore, TLSeparation is the slowest in terms of processing time, while LeWoS is the fastest, followed by CANUPO and RF.

All the investigated algorithms only use geometric features. It is expected that the additional use of radiometric features could further improve the accuracies. Our research gives an extensive overview of the investigated leaf-filtering algorithms for various site and data conditions and is an excellent decision-making aid in the selection of the appropriate algorithms.

One of the important methods used in forestry to assess genetic variability within the same tree species are provenance experiments in which different geographic origins of trees are tested. The aim of the presented work was to indicate the usefulness of ULS for automatic acquisition of selected biometric parameters of larches (4 species) and verification of existing traditional measurements. One of the larch IUFRO provenance experiments was tested on 94 plots (established in 1949 with 5-year-old larches representing 36 origins). Test site "Kolanow" is located in the Brzesko (RDLP Kraków, Poland). In April 2021, heights, DBH were measured and the health condition of 1254 larches was assessed (reference). ULS clouds were obtained from the 80m AGL using the GS-260P (Geosun) based on the LIVOX_AVIA and DJIM_300. The ULS point cloud was processed in gAirHawk (matching, georeferencing) and TerraScan (Terrasolid) software. The average ULS cloud density was 2450 pts/m2. 1206 larches were detected using the tree segmentation method. Calculations were carried out on 1160 trees, grouped into 3 provenance groups: AD (28 plots, 8 provenances, 387 specimens), AG (18 plots, 7 provenances, 192 specimens) and PSK (30 plots, 11 provenances, 403 specimens). The results of ULS_LiDAR analyses indicate an underestimation of tree height by 2.06m (2.52m RMSE). The average height of the trees reached 34.39m. The tallest larch (H=41.35m) belongs to provenance Neumuenster (Schleswig, Germany; PSK). Assuming the correct heights from the ULS LIDAR and DBH from the reference, leads to an increase in total volume by 106.77 m³. The tree with the largest volume of 6.52m³ comes from Alto Adige (Val Venosta; Italy); DBH=75.25cm; H= 34.40m. The lowest SD of H had the provenance from Lower Austria (Semmering; SD=1.33m) and the largest from the White Carpathians (Czechia; SD=3.35m). All trees showed mean SD=2.36m in height (coefficient of variation 6.86%).

The project compares the accuracy of point clouds generated using: UAV_RGB imageries: NADIR_OFF (20Mpx; DJI_P4_March_2021 and SONY_RX1_42Mpx VTOL TRINITY F90+), oblique images (SHARE_102v2; 120Mpx; DJI_M300); ULS GS-260P (Geosun; LIVOX_AVIA; DJI_M300) and ALS_LIDAR (RIEGL VQ-780i; 18.11.2020; density>400 pts/m2 – reference data-set). UAS missions were carried out at different phenological seasons to determine the effect of the presence of foliage (deciduous trees) and needles (Larch) on the accuracy of 3D point cloud generation for stands in different age classes (young and mature) on a test plot in the Niepołomice Forest District (RDLP Kraków, Poland). The use of IPC_SONY_RX (May 2020; GSD 1.3cm; 150m AGL) showed an underestimation of height for young and mature Scots pine stands, respectively: 38cm and 9cm. For young Norway spruce stands, the under-estimation was 68cm, and for European Larch, it was 44cm. Young beech trees that had not shed their leaves for the winter showed an undershoot of 7cm as did black alder of 59cm. "Overestimation" of the IPC was obtained for trees that quickly lost their leaves in autumn 2020, i.e. for young oaks 21cm, and old (> 120 years) oaks as much as 39cm. Young leafless maples showed a higher height on the IPC by 116cm than on the dense ALS cloud. IPC_P4_March_2021 and comparison to ULS_2021 showed an underestimation of the height of young Scots pines by 10cm and mature trees by 205cm. Oak stands deprived of leaves in spring showed the most significant errors (young 660cm; mature 2150cm). For beech, black alder and maple, the underestimation was respectively: 138cm; 431cm and 577cm. IPC_SHARE clouds underestimated during LEAF-ON for all analysed species: old Scots pine 58 cm; mature oak 62cm; larch 32cm, beech 23cm; maple 26cm, spruce 64cm. Carrying out ULS_2021_May and ULS_2022_July missions (approximately 2-growing seasons) allowed the height increment to be precisely determined.

Plant allometry is key for laser scanning methods to determine the amount of carbon stored in forests from proxy measurements such as tree diameters or heights, and it is thus crucial to accurately determine the role of forests in global carbon cycles. Authors typically employ allometric models optimised by their statistical performance. On the other hand, theory based approaches, such as Metabolic Scaling Theory (MST), bear a promise for mechanistic basis of allometry that may allow global generalisation of allometric models and release the need for calibration in laser scanning methods. We developed MST-based allometric models from the postulate that MST scaling is applicable only to the proportion of plant tissue with supportive functionality. Our models are a set of generalized MST (gMST) relationships allowing for variable rates of conduit coalescence and taper and a transition between transport and diffusive domains. These postulates reconcile previous MST-based models and derive an inverse relationship between stem volume taper and conduit widening. The impact of this derived relationship is in that it contradicts a rule set by Leonardo da Vinci and underpinned MST development: that the sapwood area remains constant across all branching levels. Based on this, we deduced that the allometry of trees progressed into more complex relationships as plants evolved tissues of specialized conductive functionality. Our generalised MST (gMST) models were thus created by considering conductive lumen as unsupportive area, consequentially crucially departing from the original MST 2/3 scaling. According to this principle, we deducted generalized gMST relationships with mechanistically deducted coefficients, using data from Chave et al. (2014). Our gMST relationships outperformed the models originally adjusted with that same dataset. These results indicate that the further development of generalised allometric models bear a promise for direct derivation of forest carbon fluxes from proxies derived from airborne scanning.

Canopy-dominant trees (CDT) play a critical role in shaping the 3D structure of forests and contribute disproportionally to aboveground carbon storage in these ecosystems. Consequently, understanding why CDT can vary so dramatically in their crown size and shape is critical if we are to shed light on the processes that constrain the structure and function of the world's forests. However, because CDT only makes up a small proportion of stems in a forest and because their crowns are inherently challenging to measure from the ground, we continue to lack a clear picture of what drives the enormous variation in crown architecture we observe in nature. Here we take a novel approach to address this knowledge gap by using airborne laser scanning (ALS) to directly measure the height and crown area of CDT from above. Specifically, we compiled co-located ALS and RGB imagery at 28 sites spanning all major forest types and used these to manually delineate the crowns of >20,000 CDT. Using this unique dataset, we explored how crown area–tree height scaling relationships and crown symmetry vary within and between forest types in relation to climate, wind exposure, topography, local competitive environment, and biogeographic history. We find that aridity plays a key role in driving broad-scale differences in crown architecture across forest types, but that local factors related to wind exposure, topographic position and neighbourhood density are just as important in explaining variation among individual trees. Crucially, we also show that the crown architecture of these CDT is poorly predicted using existing allometric databases compiled from field data, as these are systematically biased towards smaller trees. Our study takes a key step towards better representing the spectrum of crown architectures that characterises the world's CDT, with important implications for integrating forest monitoring programs with remote sensing and forest models.

Restoring degraded tropical forests is widely considered to be one of the most promising avenues for mitigating climate change and safeguarding biodiversity. However, the effectiveness of different restoration interventions designed to boost the recovery of degraded tropical forests, including enrichment tree planting and liana cutting, remains unclear. Here we take a novel approach to addressing this question by using repeat airborne LiDAR to quantify the effects of tree planting and liana cutting on rates of aboveground carbon density (ACD) accumulation and canopy gap closure across the Sabah Biodiversity Experiment (SBE) in Malaysian Borneo. SBE is landscape-scale restoration experiment covering 500-ha of logged forests in Sabah established in 2002, where 124 plots were assigned to one of three treatments: natural regeneration, tree planting using different combinations of 16 native tree species, or a combination of enrichment tree planting and liana cutting. Using this unique experimental setup, we tested three hypotheses: (1) planting diverse tree mixtures leads to faster recovery of ACD and canopy gap closure; (2) removal of lianas further accelerates canopy recovery; and (3) rates of recovery vary predictably across the landscape in relation to topography. Surprisingly, we found that tree planting by itself had almost no effect on rates of ACD recovery, although it did speed up canopy gap closure. By contrast, liana cutting had a profound impact on rates of ACD accumulation, as it more than doubled the rate of canopy height growth. However, rates of ACD accumulation also varied substantially across the landscape in relation to topography, being considerably faster in low-lying areas near rivers compared to more exposed ridges. Our results have major implications for forest restoration practices aiming to speed up the recovery of logged tropical forests, and highlight the importance of taking a landscape-scale view when setting baselines against which to benchmark success.

Tropical forest disturbance rates are expected to increase with climate change, potentially reducing the amount of carbon these forests can store. However, disturbance rates are hard to measure because the vast majority of disturbances are small (95% are less than 0.1 ha) and therefore cannot be reliably detected from satellites. We use repeat airborne laser scanning data to detect canopy disturbances at high spatial resolution and compare disturbance rates in Malaysian Borneo with those in the Brazil and French Guiana. Contrary to our expectations, we found higher disturbance rates in valleys compared to hill tops and ridges. We also found higher disturbance rates in taller forests (particularly the tall dipterocarp forests in Malaysia). I will discuss the implications of these results and the need for regular disturbance monitoring to help determine the main drivers.

Semantic segmentation of leaves and wood from three-dimensional point clouds is a crucial step in characterising many biophysical properties of individual trees and entire forests such as leaf area and biomass. These methods also show great promise in improving individual tree segmentation from TLS point clouds. Until recently, approaches have mostly relied upon thresholding of simple geometric features that characterise the arrangement of points in space, but the advent of deep learning running on GPUs has led to the development of new algorithms that learn from labelled data, necessitating fewer user defined parameters. We took an existing open-source algorithm and made modifications to improve segmentation in European forests, expanding its capacity in classifying at finer scales and improving its computational efficiency across forest types. We also created a unique training dataset using TLS data from the across Europe and leveraged high performance GPU computing facilities to train a new openly available model.

Forest floor vegetation forms an important component of forest ecosystems, and the spectral properties of forest floor are connected to its vegetation composition. Mapping the spectral properties of forest floor has been mainly performed with multi-angular satellite data that have coarse spatial resolution. Here we tested the use of a forest reflectance model based on photon recollision probability to retrieve forest floor reflectance from medium spatial resolution near-nadir multi- and hyperspectral data (Sentinel-2 MSI, PRISMA). The method was complemented with airborne LiDAR, which provided parameters describing forest canopy structure in the model (i.e., canopy interception of solar radiation, photon recollision probability). We also used openly available forest inventory maps, which provided the tree species fractions. The method was tested in boreal, hemiboreal, and temperate forest sites in Europe (48°–62° N). Airborne LiDAR had comparable performance to in situ hemispherical photos in providing the forest canopy structure for the forest reflectance model. There was no large difference between the computationally efficient area-based approach (utilizing height distribution metrics of the LiDAR returns) and synthetic hemispherical photographs (obtained with ray tracing). With the help of LiDAR data and the tree species maps, we derived wall-to-wall maps of forest floor reflectance from Sentinel-2 MSI and PRISMA satellite images for a boreal forest site. The performance of the retrieval of forest floor reflectance was fairly good in sparse forests (effective plant area index less than 2). In dense forests, the retrievals were less accurate, because the contribution of forest floor to the remote sensing signal was small. We showed that the retrieved forest floor reflectance can be used, for example, in monitoring the recovery of forest floor vegetation after a thinning disturbance. Overall, the results demonstrate the synergies between airborne LiDAR data sets and the data from existing and forthcoming multi- and hyperspectral satellite missions.

We assembled a unique terrestrial and airborne lidar dataset covering 50 kHa of primary and secondary miombo woodlands in Mozambique, and generated highly-accurate above-ground biomass estimates (AGB) via direct 3D measurements of forest structure; the first time such estimates have been generated independent of allometrics. We found that 1.5 to 2.2 times more AGB was stored across this region than estimated by conventional counterparts. This difference is in part owing to the systematic underestimation of large trees by allometry. Were these results observed across the world’s miombo woodlands, their total carbon stock would potentially require an upward revision of ca. 4.0 PgC, implying we currently underestimate their carbon sequestration and emissions potential, and undervalue and disincentivise their protection and restoration.

Bamboo forest is one of the most distinctive forest types in China. Estimates of bamboo forest area for administrative regions are among the most useful information for adjusting local economic policies as well as promoting national long-term storage of carbon in bamboo forest plants and soils. Parametric regressions have been proposed to infer population parameters taking remote sensing data as predictor variables. However, surface feature spectral reflectance captured by remote sensors has measurement errors due to the variation in atmosphere and illumination, which cause bias in estimated regression coefficients, and thus biasing the estimated population parameters. The objectives of the study are threefold: (1) to demonstrate the effects of measurement errors in remote sensing observations using Landsat 8 images covering Fujian Province in China; (2) to reduce the bias in area estimates using nonlinear errors-in-variables (EIV) modeling; (3) to compare the EIV approach with a conventional model fitted using ordinary least squares, in terms of estimates of regression coefficients, estimates of population parameters and their variances. Three conclusions are drawn. First, measurement errors in Landsat predictors biased the estimated regression coefficients to the direction of zero. Second, nonlinear EIV modeling had effects on reducing both bias and variance of the estimated bamboo forest area, but not necessarily for variance. Third, it is still worthy to apply the nonlinear EIV modeling though variance covariance of the estimated regression coefficients increased slightly, because the decrease in bias at either the modeling stage or the inference stage facilitated the eventual decrease in the overall uncertainty revealed by the mean squared error.

Segmentation of individual tree instances from forest point clouds remains a key challenge in the transformation of raw Terrestrial laser scanning (TLS) point cloud data to accurate digital tree models . The strong variation within and across forests in terms of, tree density, species composition, understory, etc. makes this task particularly hard to automate. Current tools employ geometric, cluster-based or topological techniques.

The state of the art in instance segmentation for indoor and urban settings are deep learning methods. Recent advances like transformers, the building block behind revolutionary technology like ChatGPT, have shown promising results. Applying transformers to point clouds is an ongoing research direction, as the complexity of large-scale point clouds pose efficiency problems not found in the corresponding language processing tasks.

Whereas deep learning has been applied successfully to forest point clouds in the semantic segmentation task, no instance segmentation method based on deep learning has been proposed. We believe such a method could be more robust to the variability of forest settings. Furthermore, while training such a model is time-consuming and computationally expensive, inference on new data is generally fast.

The irregular characteristics of forest point clouds make this setting more challenging than indoor and urban environments, and the applicability of deep learning models to this type of setting remains to be seen. Furthermore, the scarceness of high quality, manually segmented data might prove to be a bottleneck for training such a model. However, we believe that a computationally efficient deep learning-based method can greatly speed up and improve tree segmentation, enabling the generation of a large library of digital twins and improving biomass estimation. Here we will demonstrate the first results of this approach on the 2015 Wytham Woods dataset, consisting of 876 manually segmented trees.