Combining Single‐Date and Seasonal Composite Features for Cropland Soil Salinity Mapping in the Yellow River Delta

ABSTRACT

Accurate monitoring of soil salinity is crucial for crop productivity and agricultural management. Although remote sensing has been widely applied, the complementary value of features derived from single‐date and seasonal composite imagery in soil salinity prediction has not been systematically evaluated. This study evaluated soil salinity prediction models based on single‐date and seasonal‐scale multi‐temporal Sentinel‐2 imagery using 214 soil electrical conductivity (EC) samples in Kenli District, China. Four Random Forest (RF) models were developed using (1) single‐date features extracted from images acquired on the sampling dates; (2) seasonal composite features derived from mean‐composited images within each sampling season; (3) integrated multi‐temporal features combining both single‐date and seasonal composite features; and (4) an optimal subset of integrated multi‐temporal and environmental features identified via SHAP‐based feature selection. Results showed that the single‐date model performed poorly ( R ²  = 0.34, RMSE = 726.24 μS/cm), while the seasonal composite model offered only slight improvement ( R ²  = 0.30, RMSE = 724.63 μS/cm). Integrating both substantially enhanced prediction accuracy ( R ²  = 0.44, RMSE = 646.19 μS/cm), yielding an increase of 0.10–0.14 in R ² and a reduction of 78–80 μS/cm in RMSE. Further a refined model achieved the best performance ( R ²  = 0.56, RMSE = 556.94 μS/cm). Soil salinity exhibited a southwest–northeast gradient, decreasing in June due to irrigation and rainfall and increasing in October after harvest under higher evaporation. This study provides a framework for seasonal soil salinity monitoring and guidance for cropland salinisation management.