An Explainable Machine Learning and Cloud‐Based Remote Sensing Framework for Monitoring Terrace Degradation and Restoration in Mountain Landscapes

Abstract

Terrace systems play a critical role in mitigating soil erosion, stabilizing slopes, and supporting agricultural production in mountainous regions. Despite their importance, the degradation and restoration of terraces remain insufficiently quantified, particularly in ecologically sensitive and fragmented landscapes such as China's Yellow River Basin (YRB). This study investigates terrace abandonment patterns in the Zulihe River Basin—a representative watershed in the YRB—by integrating explainable machine learning with multi‐source remote sensing and cloud‐based processing. We employed Sentinel‐1 and Sentinel‐2 imagery, DEM‐derived topographic features, and land cover products, supported by high‐resolution Google Earth reference data. A Google Earth Engine–Colab workflow enabled efficient data integration and classification. An ensemble feature selection (EFS) approach combining Gini importance and recursive feature elimination (RFECV) was implemented for optimal feature selection, followed by classification using Random Forest and LightGBM models. The SHAP (SHapley Additive exPlanations) algorithm was applied to enhance interpretability, revealing geomorphic (37.2%) and SAR texture (27.6%) features as dominant predictors. Model performance exceeded 91%, with F1‐scores > 0.96. Our results indicate that agricultural terraces declined by over 20% between 2015 and 2024, with the most significant losses occurring between 2018 and 2021. These areas consistently exhibited low C‐band SAR backscatter (VV/VH), reflecting structural degradation and fluctuations in vegetation cover. In contrast, grassy terraces expanded on convex slopes, suggesting partial ecological restoration. This study introduces a robust and scalable framework for terrace monitoring, offering interpretable insights into land degradation dynamics. The approach can support sustainable land management strategies and inform policy responses in erosion‐prone regions.