Multi-features deep learning framework for aboveground biomass mapping in semi-arid forests by integrating LiDAR with Sentinel-1 and Sentinel-2 time series

Semi-arid forests harbour major carbon stocks. However, they have received little attention and are disappearing rapidly due to the global expansion of agriculture. The high-precision, large-scale estimation of semi-arid forest aboveground biomass (AGB) presents a major challenge to remote sensing because of the high spatiotemporal heterogeneity and structural complexity of semi-arid forests. This study developed a multi-features fusion transformer temporal-spatial model (MFF-TTSM) for mapping the AGB of semiarid forests by integrating time series Sentinel-1 (S-1) and Sentinel-2 (S-2) data with Global Ecosystem Dynamics Investigation (GEDI) data. An AGB reference map was constructed by combining the measured AGB and variables from airborne LiDAR data by utilising the Random Forest (RF) method, thus providing precise AGB estimates and high-accuracy sample data for subsequent modelling (R 2 = 0.833, RMSE = 21.926 Mg/ha and RMSE r = 14.785 %). Various combinations of S-1, S-2 and GEDI data, including S-1 data (ASA), S-2 data (AOP), S-1 and S-2 data (AOP + ASA), S-2 and GEDI data (AOP + AGE), S-1 and S-2 and GEDI data (AOP + ASA + AGE) and the top 15 % of all data (the TOP15%), were compared. AOP + ASA + AGE exhibited the optimal estimation performance on the basis of the MFF-TTSM model, with an R 2 of 0.884, RMSE of 18.224 Mg/ha, and RMSE r of 12.382 %. Compared with other advanced models, the MFF-TTSM model exhibited the highest accuracy under three different combinations (AOP + ASA + AGE, TOP15%, and AOP + ASA). The TOP15% experiment with the MFF-TTSM model also gave promising results (R 2 = 0.848, RMSE = 20.839 Mg/ha, RMSE r = 14.159 %), yielding an AGB map for the research region. This study thus provides an advanced deep learning method for the high-precision mapping of AGB in semi-arid forests. Such an approach is critical for the large-scale sustainable management and carbon stock monitoring of semi-arid forests. • Mapping AGB in semi-arid forests is crucial for carbon stock monitoring. • Generation of AGB reference map provides accurate samples for subsequent modelling. • The proposed MFF-TTSM model enhances the precision of biomass retrieval. • Multi-source remote sensing data integration enhances the accuracy of AGB estimation.