Soil erosion is a dominant geomorphic process and a critical determinant of ecological productivity in mountainous regions. Assessing its rate, spatial pattern, and controlling factors is essential for understanding and managing its impact. This study assesses soil erosion susceptibility in the Lidder watershed of the Greater Himalayas by combining the Revised Universal Soil Loss Equation (RUSLE) with six machine and deep learning (MDL) models and presents the first application of Shapley Additive Explanations (SHAP) based Game Theory for erosion modeling in the Himalayan region. The erosion map generated using RUSLE estimated annual soil loss (A) of 58.81 t ha−1 yr−1. It was used to delineate non-erosion zones (A < 1st quartile = 0.74 t ha−1 yr−1) and erosion zones (A > 3rd quartile = 97.20 t ha−1 yr−1) in the watershed. These zones formed inventory datasets for applying the MDL models with twenty-four soil erosion conditioning factors, including topographic, hydroclimatic, lithological, geomorphological, soil physiochemical, and landcover factors. Machine learning models outperformed deep learning-based models in soil erosion susceptibility assessment. Random Forest performed the best, with area under the receiver operating characteristic curve of 0.96, area under the precision-recall curve of 0.96, and Cohen’s Kappa (<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="477_2025_3042_Article_IEq1.gif" Format="GIF" Height="10" Rendition="HTML" Resolution="72" Type="Linedraw" Width="13" /> </InlineMediaObject> <EquationSource Format="TEX">\(\kappa \)</EquationSource> </InlineEquation>) index of 0.76. This performance is attributed to its efficiency in learning from limited datasets and its ability to handle heterogeneous data effectively. The results reveal that 7.51%, 5.77%, 9.26%, 16.60%, 40.14%, and 20.72% of the watershed area exhibit negligible, minimal, moderate, high, severe, and extreme erosion potential, respectively. SHAP-based Game Theory revealed that slope, elevation, surface temperature, stream density, and bare ground were the strongest contributors to erosion risk, while vegetation cover, built-up, and agricultural terracing mitigated erosion. The study demonstrates that these anthropogenic activities counter the negative impacts of slope steepness on soil erosion in the mountainous areas.

Application of explainable artificial intelligence to decode water-induced soil erosion in Lidder watershed of the Greater Himalayas

期刊论文
发表日期：

2025年7月12日

查看全文

Syed Irtiza Majid,
Manish Kumar,
Sourav Bhadwal

摘要

Soil erosion is a dominant geomorphic process and a critical determinant of ecological productivity in mountainous regions. Assessing its rate, spatial pattern, and controlling factors is essential for understanding and managing its impact. This study assesses soil erosion susceptibility in the Lidder watershed of the Greater Himalayas by combining the Revised Universal Soil Loss Equation (RUSLE) with six machine and deep learning (MDL) models and presents the first application of Shapley Additive Explanations (SHAP) based Game Theory for erosion modeling in the Himalayan region. The erosion map generated using RUSLE estimated annual soil loss (A) of 58.81 t ha⁻¹ yr⁻¹. It was used to delineate non-erosion zones (A < 1st quartile = 0.74 t ha⁻¹ yr⁻¹) and erosion zones (A > 3rd quartile = 97.20 t ha⁻¹ yr⁻¹) in the watershed. These zones formed inventory datasets for applying the MDL models with twenty-four soil erosion conditioning factors, including topographic, hydroclimatic, lithological, geomorphological, soil physiochemical, and landcover factors. Machine learning models outperformed deep learning-based models in soil erosion susceptibility assessment. Random Forest performed the best, with area under the receiver operating characteristic curve of 0.96, area under the precision-recall curve of 0.96, and Cohen’s Kappa ( \(\kappa \) ) index of 0.76. This performance is attributed to its efficiency in learning from limited datasets and its ability to handle heterogeneous data effectively. The results reveal that 7.51%, 5.77%, 9.26%, 16.60%, 40.14%, and 20.72% of the watershed area exhibit negligible, minimal, moderate, high, severe, and extreme erosion potential, respectively. SHAP-based Game Theory revealed that slope, elevation, surface temperature, stream density, and bare ground were the strongest contributors to erosion risk, while vegetation cover, built-up, and agricultural terracing mitigated erosion. The study demonstrates that these anthropogenic activities counter the negative impacts of slope steepness on soil erosion in the mountainous areas.