Spatial prioritization of terrace construction on the Chinese Loess Plateau using interpretable machine learning

Terraces play a critical role in improving land productivity and controlling soil erosion in mountainous regions worldwide. On the Chinese Loess Plateau (CLP), decades of large-scale terrace construction have created one of the world’s most extensive terraced landscapes. However, many of these terraces have been abandoned or degraded due to inadequate planning. Given the continuing need for terrace expansion, identifying suitable areas for future construction remains a challenge. This study introduces a data-driven framework for terrace site selection on the CLP, based on extensive field investigations and 22 features encompassing hydrology, topography, soil, watershed morphology, and socio-economic conditions. We compared four machine learning algorithms and applied the best-performing one to classify approximately 711 million pixels across the CLP, thereby identifying areas suitable for future terrace construction. To understand how features influence suitability and to quantify their marginal effects, the interpretable machine learning technique SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDP) were applied. The results indicated that XGBoost outperformed other models, achieving an overall accuracy of 89.1% and high scores across various evaluation metrics. The XGBoost-SHAP framework further revealed that terrace construction suitability is primarily governed by water availability, terrain stability, and socio-economic conditions. The analysis identified approximately 5.78 million hectares as highly suitable for terracing, of which 2.47 million hectares remain undeveloped—primarily consisting of grassland (58.6%) and cropland (41.2%). These findings and the data-driven framework provide valuable guidance for terrace planning and sustainable watershed management on the CLP and in similar regions worldwide.