A multi-source stacking framework with inverse-transformation bias correction for mapping soil salinity in arid oases: A case study from the Yarkant River Oasis

The Yarkant River Oasis, Xinjiang, Northwest China, a representative arid inland oasis system. We developed a reproducible multi-source data-fusion and stacking framework with inverse-transformation bias correction to improve soil salinity mapping in arid oases, where strong spatial heterogeneity and water-salt interactions challenge conventional approaches. Applied to the study region, the RF+XGBoost stacking model achieved stable predictive performance (test R 2 ≥0.70) and improved extrapolation reliability after bias correction. By integrating multi-source remote sensing, meteorological, and topographic variables at 300 m resolution, the framework captures both surface spectral responses and underlying hydro-environmental controls on salinity distribution. The incorporation of inverse-transformation bias correction further enhances calibration across different salinity levels, enabling more reliable identification of spatial salinity gradients and hotspots relevant to irrigation-drainage processes. Mean total soil salt content (TS) remained stable (2.55–2.63 g/kg) during 2021–2025 with an increase of about 1.18% in five years, indicating no dominant basin-scale trend during the spring sowing period. However, pronounced spatial divergence persists: downstream areas remain salinity hotspots, whereas several midstream and upstream counties show increasing trends (>10%), reflecting spatial differences in groundwater conditions and irrigation-drainage regimes. This pattern suggests coexistence of remediation and risk, highlighting the need for zone-specific monitoring and management in arid oasis systems. • We propose a multi-source stacking framework to map soil surface salinity (TS). • RF+XGBoost stacking yields robust TS predictions with test R 2 ≥ 0.70. • Perbin_train_smear improves local calibration across salinity intervals. • Oof_smear enhances external generalization; perbin_oof_smear can be unstable. • TS maps reveal spatial heterogeneity and stable temporal patterns in the Oasis.