By 2024, all Parties adhering to the Paris Agreement must report on greenhouse gas emissions and absorptions every two years. According to Article 5 of the Paris Agreement, forests play a fundamental role in limiting greenhouse gas emissions, which recognizes the importance of forest conservation in the global fight against climate change. The quality of forest data is critical in this context. Methods for mapping above-ground forest biomass (AGB) need to be improved. In this study, AGB estimates were compared with different Discrete (D) and Full-waveform (FW) laser scanning (LS) data sets using Ordinary Least Squares (OLS) and Bayesian approaches (Support Vector Machines, SVM). Airborne Laser Scanning (ALS_D), Unmanned Laser Scanning (ULS_D), and Space Laser Scanning (SLS_FW) were used to extract forest metrics as independent variables. We applied variable selection, tuned SVM regression and cross validation with machine learning approach. As a result of the forest inventory of ten permanent plots located in the Santa Rosa National Park Environmental Monitoring Super Site, province of Guanacaste, Costa Rica, the dependent AGB variables were calculated. The observed values for AGB range from 25 to 167 Mg·ha-1. Based on our findings, important variables are associated with tree heights in estimating AGB using ALS_D and ULS_D. The leaf area index, canopy coverage and height, terrain elevation, and full-waveform signal energy are the most important variables for SLS_FW. The SVM regressions with radial kernel for all laser scanning systems exhibited biomass errors ranging from 8.08% to 8.93%. No finding significant difference in AGB estimation between the three sensors.

The ForestScan project is an ESA-funded initiative to develop frameworks and new reference data to generate above ground biomass (AGB) estimates at multi-ha scales. ForestScan explores methods to improve AGB estimation, by using and combining new data from terrestrial (TLS), UAV (UAV-LS) and airborne (ALS) laser scanning and census observations. This activity is designed to improve the use of new earth observation (EO) estimates of AGB such as GEDI, BIOMASS, NISAR. ForestScan data collection is focused on tropical sites in 3 continents (Paracou, French Guiana; Gabon, W. Africa; Sepilok, Malaysian Borneo) that capture gradients of AGB and diversity across tropical forests. A core aim of ForestScan is to quantify uncertainty in TLS-, UAV-LS-, ALS- and census-derived height and AGB estimates, and their dependence on collection and processing; we will also assess what is possible given varying levels of resources, expertise and historical census data in these diverse tropical regions. Project outcomes will include protocols for data collection and analysis, open-source tools for analysing multi-scale lidar data in a consistent way and estimates of AGB to compare to allometric and EO-derived estimates. ForestScan is closely aligned with other international activities, particularly the CEOS WGCV AGB cal/val protocols, as well as GEO-TREES, a new GEO initiative aimed at establishing a network of forest biomass reference measurement (FBRM) sites. ForestScan is the first demonstration of what could be done more widely under GEO-TREES, which would in turn, significantly expand and improve the use of EO-derived AGB estimates.

Precise predictions of forest above-ground biomass (AGB) are critical to understanding changes in forest carbon stocks. Spatially explicit uncertainties in AGB predictions for population units are underestimated if spatial structure in the form of residual spatial autocorrelation and heteroscedasticity is ignored. Methods that consider the spatial structure of biomass model residuals are needed to comprehensively estimate, as well as to effectively reduce, the uncertainty in biomass predictions. The objectives of the study were to demonstrate a spatial data assimilation (DA) procedure that harnesses small-footprint airborne LiDAR, the best linear unbiased predictor (BLUP) and the spatial structure of biomass model residuals to reduce prediction variances of individual tree biomass and plot-level biomass density; to derive a variance estimator that decomposes the variance into components associated with corresponding error sources; and to compare prediction variances for three methods used to calibrate a height-based allometric model for tree biomass: ordinary least squares (OLS), generalized least squares (GLS), and spatial DA using the BLUP. We found notable assimilation effects on both individual tree biomass and biomass density predictions, variances of which were reduced respectively, due to DA decreasing both residual variance and covariance. Second, OLS, which assumed no spatial structure, underestimated prediction variance for LiDAR-predicted biomass density by 48%. Third, from the perspective of prediction accuracy, DA reduced the RMSE for individual tree biomass predictions by 11% and 14% and reduced the RMSE for biomass density predictions by 28% and 33% relative to OLS and GLS. Fourth, the omission/commission difference model was effective for correcting the systematic prediction error in the LiDAR-predicted biomass density. Overall, the proposed spatial DA procedure demonstrated great potential for reducing the uncertainty in forest biomass predictions, thereby facilitating more efficient biomass inventories.

Fires are among the most significant disturbances reducing forest biomass, and monitoring the effects of fire on forest biomass and carbon stock is important in terms of climate change mitigation. Forest fires can be analyzed, for example, according to their severity. Burn severity describes the effects of fire on vegetation and soil, such as changes in organic matter. Satellite imagery provides a means for analyzing burn severity. The estimation is based on changes in the reflectance of different regions of electromagnetic spectrum before and after the fire. Near infrared and short-wave infrared spectral regions are especially sensitive to fire effects, and normalized burn ratio (NBR), which combines these bands, is a commonly used index for assessing burn severity.

However, the difference of pre- and post-fire NBR-values (dNBR) and the relative values (RdNBR) provide only a rough estimate of the effects of fire on the forest floor. More detailed information about the actual changes under the canopy can be obtained with terrestrial laser scanning. TLS provides accurately positioned 3D point clouds of trees and ground vegetation, and with measurements before and after the fire, even the smallest changes in biomass can be detected. The aim of this study is to link satellite-based burn severity estimates with bitemporal TLS measurements.

Nine one-hectare study sites were established in areas of controlled burning across southern Finland from which time series of Sentinel-2 imagery were collected and pre- and post-fire TLS measurements taken. Both absolute and relative volume changes in ground vegetation were extracted from TLS point clouds and these values were compared to NBR-derivatives with a spatial resolution of 10 m. The purpose is to find out whether the fire-induced changes on the forest floor correlate with NBR-values and whether the stand characteristics (e.g., canopy cover) affect the linkage between TLS and Sentinel-2 data.

Despite the widespread adoption of Deep Learning (DL) for 2D image processing, these methods are not widely used in lidar modeling for forestry. Most current studies that apply DL to lidar data use rasterized or tabular lidar metrics, which do not take full advantage of the spatial resolution inherent to lidar. Moreover, the effectiveness of applying DL directly to airborne laser scanning (ALS) point cloud data for forest attribute modeling remains unclear. Given that forest aboveground biomass (AGB) knowledge is critical for forest management, we selected AGB estimation to evaluate the effectiveness of lidar DL for forest attribute modelling.

To assess the accuracy of point cloud deep learning for AGB estimation, we employed the Dynamic Graph Convoluted Neural Network (DGCNN), a versatile and high-performing method for 3D modeling tasks, to estimate plot-level forest AGB directly from ALS point clouds. We used field data from the New Brunswick Provincial Forest Inventory, consisting of 2448 plots (radius = 11.28m). ALS point clouds were down sampled using furthest point sampling such that each plot had 7168 points (18 pts/m²). Data augmentation (rotation and random noise addition), was applied to increase training dataset size and diversity.

DGCNN predicted plot-level AGB with R² = 0.65, RMSE = 39.84 Mg/ha, MAPE = 3.87% for the unseen test dataset. These results provide evidence for the potential of point cloud DL for forest AGB estimation and can be compared to a small number of prior studies that adopted an alternative voxel DL approach. Overall, the ability of DGCNN to learn abstract representations from a lidar point cloud demonstrates a more detail-oriented application of lidar compared to traditional area-based approaches. Future research may investigate the fusion of additional datasets in a DL model such as multispectral imagery to improve model performance.