Integrated Prediction and Analysis of Soil Erosion and Soil Organic Carbon Loss in the Dayekou Watershed, Qilian Mountains

ABSTRACT Soil erosion poses a critical threat to ecological security in mountain watersheds worldwide. Predicting its dynamics and associated soil organic carbon (SOC) loss remains an ongoing challenge during the 21st century due to complex interactions among climate, land use, and topography. This study developed an integrated modeling framework by combining Long Short‐Term Memory (LSTM) neural networks, Cellular Automata‐Markov (CA‐Markov), and the InVEST Sediment Delivery Ratio module to predict soil erosion and SOC loss in the Dayekou Watershed in northwest China's Qilian Mountains. LSTM accurately forecasted 48 months of rainfall (RMSE: 177.9–211.7) using 300 months of historical data (2000–2024), while CA‐Markov simulated land use patterns (2024–2028) with high precision (Kappa = 0.85). Results revealed that severe erosion areas (< 9% of watershed) contribute > 97% of total soil loss, concentrated above 3750 m elevation of the watershed. Land use type dominated erosion spatial distribution ( q = 78.8%), with glacier/snow and bare land exhibiting the highest erosion modulus. Average SOC loss increased from 9.877 to 21.524 t/(km 2 ·a) (2000–2023), with forested areas experiencing disproportionate carbon depletion despite lower erosion rates. The coupled framework demonstrates its robustness in predicting dynamic erosion in complex mountain systems, providing strong evidence‐based conservation support for protected areas in northwest China.