Forest Canopy Height Mapping Using GEDI-SAWA and Sentinel-1/2 with Deep Learning Model in Northeast China

Forest canopy height (FCH) is a key indicator of forest carbon storage. The discrete relative height (RH) metrics produced by Global Ecosystem Dynamics Investigation (GEDI) data has been widely used for continuous FCH mapping with machine learning. However, the RH metrics will be overestimated by the influence of terrain effects, which leads to inaccurate measurements of canopy height. To address this, this study utilizes the terrain-insensitivity waveform metrics, GEDI slope-adaptive waveform metrics (GEDI-SAWA), as label to train and validate U-Net model with the help of Sentinel-1, Sentiniel-2 data and produce the FCH map at 10m resolution in a mountainous region of northeastern China. Finally, aerial LiDAR data is used to assess the generated FCH map. The results demonstrate that using GEDI-SAWA as reference data helps reduce the overestimation of FCH with bias decreasing by 2m and yield more accurate and precise results with RMSE reduced by 1.5m compared to GEDI L2A.