High-Resolution Mapping of Forest Parameters in Tropical Rainforests Through AutoML Integration of GEDI With Sentinel-1/2, Landsat 8, and ALOS-2 Data

Forests are vital carbon sinks, with tree height and biomass critical for carbon research. NASA's GEDI spaceborne LiDAR enhances vegetation monitoring through 3D structure analysis. This study established relationships between GEDI products and Sentinel-1/2, Landsat 8, ALOS, and GLO-30 features using the AutoML method. We constructed a total of 432 features, primarily from 14 types of earth observation features. In a 95.56% forest-covered area, AutoML improved canopy height (FCH) and biomass (AGBD) accuracy by up to 5.25 m and 32.18 Mg/ha over other methods. Polarization interference features, especially phase, explain 20% of forest parameters, showing high stability. Wavelet and Fourier-based texture features also demonstrate strong potential. Two mapping methods are proposed: 10 m resolution (FCH $R^{2}$ = 0.53, RMSE = 11.49 m; AGBD $R^{2}$ = 0.53, RMSE = 133.56 Mg/ha) and 500 m resolution (FCH $R^{2}$ = 0.64, RMSE = 10.06 m; AGBD $R^{2}$ = 0.66, RMSE = 114.25 Mg/ha). Compared to existing maps (AGBD: $R$ $<$ 0.1, RMSE $>$ 180 Mg/ha; FCH: $R <$ 0.2, RMSE $>$ 15 m), our method (AGBD $R$ = 0.74, RMSE = 131.39 Mg/ha; FCH $R$ = 0.73, RMSE = 11.30 m) significantly improves accuracy. The approach shows minimal saturation effects and broad applicability for forest parameter estimation.