Abstract
<p>Accurate mapping of forest aboveground biomass (AGB) is critical for carbon budget accounting, sustainable forest management as well as for understanding the role of forest ecosystem in the climate change mitigation.</p><p>In this study, spaceborne Global Ecosystem Dynamics Investigation (GEDI) LiDAR data were used in combination with Sentinel-1 synthetic-aperture radar (SAR) and Sentinel-2 multispectral imagery as well as elevation data to produce a wall-to-wall AGB map of Australia that is more accurate and with higher spatial and temporal resolution than what is possible with any one data source alone. Specifically, the AGB density map was produced that covers the whole extent of Australia at 200m spatial resolution for the Austral winter (June-August) of 2020. To produce this map Copernicus Sentinel-1 and Sentinel-2 composites and ALOS World 3D Digital Surface Model (DSM) were trained with samples from the GEDI Level 4A product.</p><p>From GEDI Level 4A data available within Australia between June – August 2020, all measurements not meeting the requirements of L4A product quality, and those with degraded state of pointing or positioning information and an estimated relative standard error in GEDI-derived AGB exceeding 50% were rejected. Mean Sentinel-1 composite was generated using thermal noise corrected, radiometrically calibrated and terrain corrected VV- and VH-polarization backscatter imagery. Similarly, median Sentinel-2 composite was generated using cloud and cloud-shadow free Level-2A imagery, and was further used to calculate Normalized Difference Spectral Indices (NDSIs) from all spectral bands. Finally, aspect and slope were calculated from the DSM.</p><p>The boosting tree machine learning model was applied to predict wall-to-wall AGB density map. For each 200m × 200m cell the number of available GEDI measurements was calculated and models were built based on average AGB density of cells containing > 5 GEDI measurements.</p><p>Up to ≈62000 cells, each 200m × 200m, were used to train predictive machine learning models of AGB density. The predictive performance of models based on Sentinel-2 imagery only (single-data source) and a fusion of Sentinel-2 with Sentinel-1 imagery and elevation data (multi-data source) was compared. Bayesian hyperparameter optimization was used to identify the most accurate Light Gradient Boosting Machine (LightGBM) model using 5-fold cross-validation. </p><p>The single-data source analysis based on Sentinel-2 imagery resulted in AGB density predicted with the coefficient of determination (<em>R</em><em>2</em>) of 0.74-0.81, root-mean-square error (<em>RMSE</em>) of 40-44 Mg/ha and root-mean-square percentage error (<em>RMSPE</em>) of 45-55%.Model performance improved only marginally with the addition of Sentinel-1 and DSM information: AGB density prediction with <em>R</em><em>2</em> of 0.75-0.82, <em>RMSE</em> of 36-41 Mg/ha and <em>RMSPE</em> of 44-48%. Using a SHapley Additive exPlanations (<em>SHAP</em>) approach to explain the output of LightGBM models it was found that Sentinel-2 derived <em>NDSIs</em> using Red Edge and Short-wave Infrared bands were the most important in predicting seasonal AGB density. </p><p>Similar model performance is expected for annual prediction of AGB density at a finer resolution (e.g. 100m) due to higher density of GEDI measurements. This research highlights methodological opportunities for combining GEDI measurements with satellite imagery and other environmental data toward seasonal AGB mapping at the regional scale through data fusion.</p>
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have