Marginal land identification and grain production capacity prediction of the coverage area of western route of China’s South-to-North Water Diversion Project

The western route of the South-to-North Water Diversion Project (SNWDP) provides opportunities to improve agricultural production by altering regional water availability. This study identifies and evaluates marginal land—defined as undeveloped reserve cultivated land and low-quality and inefficiently-utilized farmland—within provinces along the SNWDP route. Using ecological, topographic, climatic, and soil indicators, we identified 145,062 km 2 of marginal land, including 3,626 km 2 of reserve cultivated land and 141,436 km 2 of low-quality and inefficiently-utilized farmland, mainly concentrated in northwestern Xinjiang, with Qinghai having the least. To assess the grain production potential of these lands, we used maize and wheat as representative crops. Three modeling approaches—random forest regression, gradient boosted regression trees, and two-point machine learning (TPML)—were compared for their predictive accuracy. The TPML model showed the best performance. For maize, the model yielded a root mean square error (RMSE) of 48.94, a mean absolute error (MAE) of 34.01, and a mean absolute percentage error (MAPE) of 7.65%. For wheat, the RMSE was 23.92, MAE 17.67, and MAPE 6.31%. Results reveal that maize has a higher production capacity than wheat, and that grain yields are higher in the west and lower in the east, with Xinjiang showing the highest average yields on marginal land. These findings provide a scientific basis for optimizing land use, improving food self-sufficiency, and supporting regional sustainable development and national food security.