ForestCarbonNet: integrating terrain-corrected GEDI, Landsat, and PALSAR2 for enhanced forest aboveground carbon density estimation

Accurate quantification of forest biomass carbon (C) stocks is essential for regional carbon accounting and climate change mitigation. However, high-resolution mapping remained challenging as contemporary satellite LiDAR-based methods faced an inherent trade-off between spatial resolution and prediction accuracy due to their limited spatial coverage. In this study, we proposed ForestCarbonNet, a scatter-supervised image-to-image U-Net framework that integrated in-situ measurements with terrain corrected GEDI LiDAR observations, Landsat imagery, and ALOS-2 PALSAR-2 radar backscatter data to overcome this resolution-accuracy compromise. We generated high-resolution (30-m) forest aboveground C density maps and achieved robust model performance (R2 = 0.83, %RMSE = 24.02%) across temperate forests in northeast China for 2020. Our approach offered three key advantages: (1) the locally calibrated allometric model effectively bridged in-situ measurements with terrain-corrected GEDI observations (3.69 million samples), enabling robust C density estimation at GEDI footprint scale; (2) GEDI expanded C density sampling across climatic gradients while mitigating saturation effects in high-biomass forests, reducing RMSE by 24.8%; and (3) ForestCarbonNet preserved spatial contextual information and outperformed traditional pixel-based methods, reducing RMSE by 12.3%. We estimated the total forest aboveground biomass carbon stock to be 2.74 ± 0.38 PgC, with projected carbon sink peak at 4.21–5.08 TgC year−1 in the 2060s but decline to only 32% of peak levels by 2100s. The proposed method enables accurate quantification of forest carbon stocks, providing critical data that strengthens evidence-based forest management strategies and informs climate policy decisions.