Estimating Soil Salinity Based on Time-series Data Mining and Subregional Modeling Strategies

High-precision digital soil salinization mapping is vital for agriculture and ecological sustainability. Despite increasing interest from researchers, satellite-based studies have primarily focused on single- or multi-temporal data, with limited attention to remote sensing time series (RSTS). Numerous studies have highlighted the importance of temporal features in improving prediction accuracy. However, the high dimensionality, intricate temporal dependencies, and substantial redundancy of RSTS data present significant challenges in practical applications. In this study, we proposed a novel soil salinity estimating approach by integrating RSTS data mining with subregional modeling. A three-stage optimization strategy—comprising spatial partitioning, optimal time window selection, and feature optimization—was implemented to support accurate, spatially continuous mapping of soil salinity. The findings reveal that the correlations between RSTS features, and soil salinity display an annual periodic trend, with one optimal time window per year. And the correlation diminishes progressively interval between soil sampling and image acquisition increases. Moreover, marked disparities exist in the optimal temporal windows among subregions, with the vegetation zone spanning July to September and the bare soil zone covering November to January. Without regional partitioning, the model achieves an R² of 0.72 and RMSE of 21.27 g kg⁻¹. After applying subregional modeling, R² increased by 16.66% and RMSE decreased by 18.38%. Incorporating time series features further improved model performance, increasing R² by 2.23% and reducing RMSE by 12.67%. This study offers a new approach for high-precision monitoring and digital mapping of soil salinity in arid regions.