Regional maize suitability based on soil water and salt content inversion by integrating machine and transfer learnings in Xinjiang

Soil water content (SWC) and salt content (SSC) are critical factors affecting maize growth. Remote sensing technology has become an effective tool for regional SWC and SSC estimation, but challenges remain in improving estimation accuracy and cross-scale model transfer. In this study, the feature sets were optimized using correlation clustering analysis and full subset selection, and five machine learning models, including the bat-optimized random forest (BA-RF), were compared to estimate SWC and SSC. Further, the inversion model constructed based on UAV features was transferred to satellite scale using transfer component analysis (TCA) and its accuracy was verified. The key findings were as follows: (1) Feature optimization improved estimation accuracy (SWC: R²≥0.541, RMSE≤0.021 cm³ cm^–3; SSC: R²≥0.574, RMSE≤0.816 g kg^–1). (2) The BA-RF model achieved high estimation performance for SWC (R² = 0.705–0.899, RMSE = 0.010–0.020 cm³ cm^–3) and SSC (R² = 0.700–0.897, RMSE = 0.466–0.737 g kg^–1). (3) TCA enabled effective transfer of the BARF-TCA model from UAV to satellite scale, maintaining a high estimation accuracy (SWC: R² ≥ 0.764 RMSE ≤ 0.015cm³ cm^–3, SSC: R² ≥ 0.667, RMSE≤ 0.672 g kg^–1). (4) A water-salinity suitability index was developed to generate dynamic maize suitability maps across growth stages. This study presented an integrated framework for large-scale, high-precision estimation of SWC and SSC, as well as water-salinity-based crop suitability zoning, providing valuable guidance for maize farmland SWC and SSC management in arid and saline-alkaline regions.