Synergistic Multi-Source Remote Sensing and Machine Learning for Regional-Scale Soil Salinity Retrieval in the Hyper-Arid Tarim River Basin

Saline–alkali land is known colloquially as the “ringworm of the Earth” because of its wide distribution, difficulty of management, and the serious threat it poses to agricultural production and ecological security in arid regions. The Tarim River Basin is the primary distribution area of saline–alkali land in the arid region of northwest China. Therefore, accurate assessment of its distribution and degree of saline–alkali development is important for the comprehensive management and maintenance of regional ecological security in this region. However, the distribution and characteristics of regional saline–alkali land are unclear because of the limitations of traditional land survey methods. To address this, a field campaign was conducted across the Tarim River Basin, yielding 147 topsoil specimens subsequently subjected to electrical conductivity determinations, and it used a retrieval model for saline–alkali soils by integrating remote sensing techniques and machine learning algorithms. The results show that a Random Forest (RF) model integrating all multi-source features was optimal for soil salinity retrieval, achieving the highest accuracy (R2 = 0.73, RMSE = 3.58 dS m−1, MAE = 1.89 dS m−1), with superior robustness and extreme value prediction, while Extreme Gradient Boosting (XGBoost) complemented saline vs. non-saline boundary extraction. Macro-climatic variables (MAT, MAP) dominated with a 33.2% combined contribution, consistent with the dominant climatic control on soil salinity in hyper-arid regions. Multi-source synergies including optical indices, Sentinel-1 SAR (12.0%), and environmental covariates enhanced physical interpretability. Spatially, soil salinity exhibited strong heterogeneity (CV = 113.04%), being concentrated in the middle–lower plains and irrigation districts affected by anthropogenic activities, contrasting with low-salinity upper mountainous regions. Our results demonstrate that incorporating land survey, remote sensing, and machine learning supports screening-level retrieval of the extent of saline–alkali land in the hyper-arid Tarim River Basin, providing a foundation for its systematic improvement.