High Carbon Emission Simulated in the Permafrost Degradation Regions of the Qinghai–Tibet Plateau by Remote Sensing and Deep Learning Modules

The Qinghai-Tibet Plateau (QTP) stores a significant amount of organic carbon in permafrost regions, and the temporal dynamic changes under permafrost degradation remain uncertain. In this study, integrating on-site and multi-source remote sensing data, we proposed a dual-input small-sample deep learning framework for estimating the soil organic carbon (SOC) density and storage at the depth of 0-3m in permafrost regions based on attention mechanisms and deep learning (DL) methods. Our model achieved an improvement of 10.6% and 22.9% in the accuracy of SOC estimation compared to previous studies in the shallow (0-30 cm) and deep (0-100 cm) layers of permafrost regions, respectively. The SOC storage over permafrost regions to the depth of 3m were 14.27 ± 4.38 Pg and 12.26 ± 2.02 Pg in 2005 and 2020 respectively, suggesting a release of 2.01 ± 0.48 Pg of SOC over the past 15 years. The carbon release intensities per unit area in thermokarst lakes and retrogressive thaw slumps are approximately 24.6 times and 21.0 times higher than the mean value of whole permafrost regions, respectively. The factor analysis revealed that precipitation, NDVI (Normalized Difference Vegetation Index), and MAGT (Mean Annual Ground Temperature) are the primary controlling factors when estimating the shallow layers (0-30 cm). In contrast, at deeper layers, the spatial distribution of shallower SOC, soil water content, and DEM possess greater weight. This finding is crucial for modeling SOC storage and its dynamics in the permafrost regions on the QTP.