Ensemble Learning-Based Assessment of Soil Salinization at the Agricultural Parcel Scale in Arid Regions: A Case Study of Delingha City in the Qaidam Basin, China

Soil salinization is a major constraint on sustainable agricultural development in arid regions, yet soil salinity dynamics are commonly assessed using pixel-based remote sensing products that are difficult to relate to agricultural management units. To solve this problem, we developed an agricultural parcel-scale framework for soil salinity monitoring and mitigation assessment in Delingha City, Qinghai Province, China. Cropland parcels were extracted using a Recurrent Residual U-Net (R2U-Net) model, and soil salinity inversion for April during 2021–2025 was conducted by integrating Sentinel-1/2 imagery with a stacking ensemble learning model. The model incorporated Random Forest Regression (RFR), Gradient Boosting Regression (GBR), Multi-Layer Perceptron (MLP), and Support Vector Regression (SVR) to improve prediction accuracy and robustness. The selected features included vegetation indices, salinity indices, and SAR backscatter parameters. Among them, the Soil Adjusted Vegetation Index (SAVI) showed the strongest correlation with soil salinity, while Salinity Index 2 (SI2) exhibited the highest contribution to model performance. The R2U-Net achieved an F1-score of 0.8574 for parcel extraction. The ensemble model produced the best inversion results with an R2 of 0.52 and reduced prediction errors compared with individual models. Results indicated an overall decline in soil salinity from 2021 to 2025, suggesting an improvement in soil salinity conditions during the study period. Parcel-scale aggregation reduced spatial noise; improved temporal stability; and revealed heterogeneous field responses, including salinity-declining, fluctuating, and increasing trends. The proposed framework enhances the interpretability and management relevance of soil salinity monitoring and provides practical support for precision agricultural management in arid regions.