A deep geostatistical fusion framework for forest canopy height retrieval in complex tropical terrain

Forest canopy height (FCH) is a fundamental parameter for assessing forest biomass, carbon stocks, and ecosystem health, yet accurate large-scale estimation remains challenging, particularly in complex terrains where remote sensing signals are susceptible to geometric distortions and spatial discontinuity. This study develops a novel framework that integrates an Adaptive Deep Kriging Network (ADKN) for spatial interpolation of discrete spaceborne LiDAR data with a Deep Gaussian Mixture Copula Regression (DGCR) model for FCH estimation. Using UAV-LiDAR-derived CHM as reference observations, and integrating ICESat-2/ATLAS and GEDI LiDAR footprints with Sentinel-1 SAR, Sentinel-2 optical imagery, and topographic variables as primary predictor inputs, we evaluated the proposed methods in a topographically complex tropical forest region. ADKN effectively generated continuous 10 m surfaces of spaceborne LiDAR-derived variables, eliminating the striping artifacts and oversmoothing inherent to traditional interpolation methods. DGCR demonstrated superior performance by explicitly modeling complex multivariate dependencies among heterogeneous remote sensing features, achieving the highest accuracy (R 2 = 0.809, RMSE = 1.991 m) when integrating all data sources together with topographic factors, substantially outperforming other machine learning models. Multi-source data fusion progressively improved estimation accuracy, compensating for individual sensor limitations, while slope was identified as a key topographic factor controlling model performance, with accuracy declining notably under extreme terrain conditions. The proposed ADKN-DGCR framework provides a robust and scalable solution for wall-to-wall FCH mapping in complex environments, with significant implications for carbon cycle modeling and forest resource management.