Advancing Provincial Cropland Soil Mapping With Temporal Satellite Data Integration

ABSTRACT Understanding the spatiotemporal distribution of soil constraints, such as soil acidification and degradation, is critical for evaluating soil productivity and safeguarding ecosystem health. Digital soil mapping (DSM), which integrates environmental covariates, offers significant advantages in providing detailed soil information across scales. However, existing DSM approaches often fail to adequately represent agricultural management practices on cropland, limiting their ability to accurately characterise soil properties. In this study, we employed random forest (RF), Cubist, XGBoost and Artificial Neural Network (ANN) models within a DSM framework, incorporating crop growth dynamics, planting patterns and satellite time series spectral data to predict soil organic matter (SOM) and pH. The results showed that the RF model, which integrated temporal information, outperformed the other models for SOM prediction, achieving an R 2 of 0.42 and an RMSE of 9.67 g kg −1 , representing a 13.51% improvement in R 2 and a 3.21% reduction in RMSE compared to the baseline model without temporal information. For soil pH prediction, the XGBoost model, which incorporated temporal information, delivered the best performance with an R 2 of 0.61 and an RMSE of 0.63, resulting in a 12.96% improvement in R 2 and a 5.97% reduction in RMSE relative to the baseline model. This study demonstrates that integrating temporal information, such as crop growth and management dynamics, substantially enhances the accuracy of soil property predictions. These findings provide a promising approach for high‐resolution soil mapping with broad applicability in agricultural and environmental management.