Digital mapping of soil salinity with time-windows features optimization and ensemble learning model

Abstract

Soil salinization poses considerable global environmental and ecological risks. Remote-sensing time-series data enable more accurate monitoring and prediction of soil salinity levels, offering a refined approach to soil salinization assessment. However, the current limitations of time-series data analysis—particularly in terms of timely and effective information extraction—hinder high-precision soil salinity assessments. This study proposes a data mining approach using Sentinel-1 time-series data, integrating time-series decomposition and feature selection to capture seasonal and trend components correlated with soil salinity, and determine optimal time windows and effective time spans. An advanced feature-selection algorithm was then applied to refine the model-relevant features, and the transferability of the method across different regions was validated through empirical testing. The results revealed a 12 month periodicity in the correlation between Sentinel-1 time-series features and soil salinity, with an annual decay rate of 0.0019. In the study area, the optimal time window was from July to September, with the maximum effective years ranging from 19 to 21. Recursive feature elimination has shown a gradually increasing trend in the importance of SAR features from single-temporal to multi-temporal to time-series data. The time-series analysis combined with feature selection not only significantly reduced data volumes, but also improved the prediction accuracy of the model—increased R² of the prediction set was improved by 0.11, and a reduced root mean square error of 3.08 g kg⁻¹, compared to single-temporal data. Furthermore, the results demonstrate that the ensemble model outperforms the individual models in terms of inversion accuracy, whereas the time-series mining method exhibits generalizability across diverse study areas and metrics. The combination of the time-series mining method with the ensemble model helps achieve a higher accuracy in digital soil mapping.