Identifying long-term winter wheat planting areas using decision tree-derived ensemble learning models and multi-source data

Abstract

[EN] To provide high-quality data support for agricultural policy-making and grain subsidies, this study presents an efficient method for extracting winter wheat planting areas using Landsat series satellite imagery and monthly maximum normalized difference vegetation index (NDVI) stacks. Three ensemble decision-tree algorithms-Random Forest, XGBoost, and CatBoost-were compared. The best model achieved 91% cross-validation accuracy, with a strong statistical validation accuracy in terms of its municipal-scale validation R2 = 0.91 (MAE = 49,650 hm2; RMSE = 64,440 hm(2)) and county-scale R2 = 0.84 (MAE = 7125 hm2, RMSE = 9875 hm2). A spatiotemporal analysis revealed a notable decline in winter wheat area in Shandong Province, which was concentrated in its western and southern regions. The cultivation centre shifted westwards, and then northwards, and the landscape patterns transitioned from large, aggregated patches to small, dispersed patches. This trend of decreasing intensification level for the winter wheat planting regions poses challenges for achieving water-saving and intensive land management goals. These trends are influenced by climate, topography, and socioeconomic factors, as well as agricultural policies. The proposed method offers robust support for attaining large-scale crop monitoring and sustainable agricultural management.