Multi-source remote sensing retrieval and spatiotemporal distribution characteristics of soil moisture content in typical karst farmlands of southwestern China

Southwest karst region,China In karst regions characterized by high vegetation coverage and complex topography, indirect prediction methods utilizing natural environmental factors—including vegetation, topography, and meteorology—are essential for estimating farmland soil moisture content (SMC). Identifying optimal combinations of environmental factors and machine learning algorithms is fundamental to developing effective SMC prediction models. Focusing on Qiannan Prefecture in Guizhou Province—a core karst area of southwestern China—this study established an optimal selection mechanism for input variables and machine learning algorithms through environmental factor importance analysis and optimization experiments. This study developed a methodology for constructing optimal soil moisture content (SMC) prediction models through systematic evaluation of environmental factors and machine learning algorithms. Applied in Qiannan Prefecture, the approach produced a farmland SMC20 prediction model integrating eight key environmental variables with XGBoost algorithm. Validation demonstrated high accuracy within the study area ( R² =0.85, RMSE =0.057, Bia s = -0.013, KGE =0.887). Compared to conventional models using fixed algorithms and variables, this method ensures relative optimality through scientific screening. Transferability tests in Guangxi (China) and Baena (Spain) revealed decreasing accuracy with increasing distance from the study area, prompting discussion on model generalization. Temporally, farmland SMC20 peaked in June and reached its lowest in January. Spatially, after excluding transitional zones, SMC20 showed gradual decrease with distance from rivers in karst basins. These findings provide quantitative data and technical support for mechanistic studies, temporal forecasting, and drought warnings in karst farmland systems. • A Construction Method for SMC Prediction Models in Karst Regions Using ML. • The combined model of topography, meteorological factors, and XGBoost is optimal. • Farmland SMC initially increases then gradually decreases with water system distance. • In karst core areas, farmland SMC peaks in June and reaches its lowest in January.