Long-term spatiotemporal prediction and changing patterns of salt-affected soils at continental scale using machine learning methods

Soil salinization exhibits significant spatiotemporal variability and threatens global food security. Asia, which has the largest area of salt-affected soils, faces particularly urgent challenges. Existing studies have primarily focused on static spatial patterns, with limited insight into the long-term changes and underlying drivers of continental-scale soil salinization. In the research, we developed a machine learning spatiotemporal modeling framework that integrated both static and dynamic predictors with explicit temporal sampling to map salt-affected soils in Asia from 1980 to 2018. This approach enabled continuous long-term mapping of soil salinization indicators, including saturated paste extract electrical conductivity (ECe), pH, and exchangeable sodium percentage (ESP). The number of measured data used for these three indicators is 11,979, 15,353, and 6148, respectively. The spatial and temporal evolution of soil salinization along with its drivers was comprehensively analyzed. To identify the optimal spatiotemporal prediction model, five machine learning models were compared, including multiple linear regression, cubist, random forest, quantile regression forest, and quantile regression neural network. Among them, the quantile regression forest performed optimally, with a mean absolute percentage error mostly below 48.9%, coefficient of determination values of 0.52∼0.88, and Nash-Sutcliffe efficiency values of 0.52∼0.88. Spatially, soil salinity and sodicity are concentrated in arid and semi-arid regions, as well as coastal areas. Temporal analysis revealed a worsening trend in soil salinization, with the area of salt-affected soils increasing by an average rate of 10.61 million hectares every five years. Climatic and topographic factors are the primary drivers of the spatial distribution of salt-affected soils, while vegetation and climate variables are critical for explaining temporal dynamics. This study offers comprehensive insights into large-scale long-term prediction and management of salt-affected soils.