Laser scanners have been successfully employed for forest variables estimation. Nevertheless, the regular monochromatic LiDAR systems cannot capture enough information for tree species classification. Therefore, optical multispectral (MS) imagery or their integration with airborne LiDAR has been studied for classifying forest tree species. Modern multispectral airborne LiDAR has improved tree species identification accuracy compared to single-channel systems, by incorporating both structural and spectral information. Hence, the characterization of tree species is one of the primary and most popular potential applications of multi-wavelength laser scanning attracted attention in recent years. State-of-the-art deep learning (DL) algorithms have recently demonstrated promising potential in boosting the automation and accuracy of geospatial data processing. However, the effectiveness of DL models has never been explored in tree species classification using MS-LiDAR data, and the conducted studies have relied only on conventional handcrafted features. Therefore, the aim of this paper is to report our investigations on the capability of automatic DL-based features for unsupervised classification of tree species utilizing MS airborne LiDAR data. First, the point clouds of three missions with Green, NIR, and SWIR wavelengths were pre-processed and merged to create MS LiDAR datasets. Subsequently, individual trees were delineated using SLIC superpixel segmentation. A Convolutional Auto-Encoder (CAE) was used for extracting deep features from the MS LiDAR data. Then, several statistical features were extracted for each tree segment from deep ones. Finally, the Mini Batch K-means was used to cluster the resulting features and identify the tree species. The achieved results indicate the superiority of DL-based features over the manual ones for tree species classification, improving the overall accuracy and kappa coefficient by 5.63% and 8.32%, respectively. To the best of our knowledge, this is the first study exploring unsupervised DL for forest mapping using MS LiDAR and confirms its potential as a single-data source solution.

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

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.

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.

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.

Deadwood is an important component of the forest ecosystem and represents a central input for carbon storage studies. In addition to information on the presence of deadwood, the volume or biomass as well as the degree of decay are of interest. Usually, deadwood is surveyed in the context of forest inventories at plot level, whereby the lying as well as the standing deadwood is recorded by means of diameter at breast height, and tree height and stem length respectively. Another method is the line-intersect sampling (LIS), where all lying trunks intersecting with a line are recorded. Finally, statistical methods can be used to estimate the total amount of deadwood in a defined area.

Due to the rapid developments in the field of UAV-LiDAR, it is now possible to acquire 3D point clouds with a very high point density even for larger survey areas at favourable prices. This makes it possible to determine the deadwood for large areas.

The lying deadwood is detected and its volume is estimated from UAV-LiDAR and image data (RIEGL-VUX-120, PhaseOne-iXM100) data with a point density of >4000 pts/m² for the Rohrach natural forest reserve with an area of 48 ha. From the aerial images a true-orthopoto is calculated with a pixel size of 5 cm. Voxel-based approaches are used which take into account not only the geometrical properties but also the radiometric properties of the backscattering objects. The results are compared with reference surveys of 48 plots (437 trees) and with LIS (174 trees).

The results show a very high completeness of the detected deadwood. The validation with the plot-based deadwood volumes shows high accuracies. The total deadwood volume for the entire area show some deviations which have to be analysed in more detail in the next months. This study is funded via the Austrian Waldfonds.

During SilviLaser 2021, a benchmark was organized to demonstrate the capability of different 3D data acquisition techniques for capturing various forest parameters. In total, 9 groups of participants revealed their advanced techniques and setups to acquire point clouds in the designated sites. Based on the applied equipment and platform, all approaches are categorized into 3 groups: mobile laser scanning, terrestrial laser scanning, and photogrammetry (plus others). In order to efficiently and accurately extract basic forest parameters (e.g., stem position, tree species, DBH) from this data set, understanding the data behavior of different approaches is the main key to maximizing their strength. This study aims to characterize each method in terms of spatial distribution and coverage of point clouds, extra attributes, as well as pros and cons for different usage purposes.
The output of this study is valuable for selecting an adequate method to fulfill the requirements of user-specific forestry applications. Also, it is beneficial for solving the existing or future problems of multi-source point cloud processing in forests, e.g., co-registration of various data sources. This study is done within the framework of the project 4Map4Health.

Multiple studies have analyzed the error budget of biomass predictions from airborne laser scanning (ALS) data utilising the area-based approach (ABA). While some error sources, such as high-incidence scan angles, can be excluded from analyses, the main error sources of (a) geolocation errors, (b) quality of the reference data, and (c) residual modelling errors remain. In this work, we aim to focus on the contribution of inherent uncertainty of the trees themselves, e.g., caused by wind during the ALS data acquisition. At Petawawa Research Forest in Ontario, Canada, we analysed overlapping pairs of ALS flight lines and trained a Random Forest to estimate biomass from 223 forest inventory plots, both on the merged dataset and separated by flight strips using a set of state-of-the-art LiDAR metrics as independent variables. Quantifying inference performance on withheld validation data, we compared single-strip predictions with estimates from the merged dataset to examine how much of model error is due to the model’s capacity and reference data quality, and how much is caused by different representations from the separated strips. Our results show that after excluding problematic scan angles, the RMSE differences between merged and separated strips amount to approximately 50-80% of the predictor RMSE. These errors tend to be larger for taller trees, especially when close to clearings, where effects of wind on the point cloud metrics are larger. Furthermore, we analysed multiple data acquisitions acquired from several sensors. Overall, increased footprint size and increased point density corresponded to a decrease in biomass RMSE. Considering current developments towards small-footprint (UAV-based) laser scanning, uncertainty caused by movement of tree crowns may be an increasingly limiting factor for accurate biomass- and other forest-related metric extraction from laser scanning. We therefore strongly suggest that this uncertainty should be quantified when analysing laser scanning data of forests.