A Hybrid Window Transformer for accurate and generalizable soil salinity prediction across China

Accurate mapping of soil salinity is essential for sustainable land management and mitigating land degradation in arid and semi-arid regions. Conventional approaches frequently encounter difficulties in capturing nonlinear environmental interactions and achieving reliable generalization to sparsely sampled areas, due to inherent sampling biases, data heterogeneity, and spatial non-stationarity. To address these challenges, we developed the Hybrid Window Transformer (HW-Transformer), a deep learning framework for predicting soil electrical conductivity (EC) at 100 m resolution across China. The model integrates multi-source remote sensing data with auxiliary environmental variables (climate, topography, and soil properties) through four complementary modules: (1) a multimodal feature embedding (MFE) layer for aligned fusion of heterogeneous inputs, (2) hierarchical window attention (HWA) to capture multiscale spatial dependencies, (3) cross-level feature interaction (CFI) to approximate empirical patterns consistent with known environmental drivers, and (4) an MLP-based output head for nonlinear regression. The HW-Transformer achieves a coefficient of determination (R2) of 0.68 under leave-one-region-out cross-validation, 0.75 under ten-fold cross-validation, and 0.62 on independent validation. These results consistently outperform standard Transformer and other baseline models, underscoring the framework's effectiveness in mitigating sampling imbalances through multimodal integration and attention mechanisms, thereby enhancing predictive consistency across heterogeneous bioclimatic zones. Finally, we applied the model to generate a 2024 national soil EC map, which captures detailed salinity patterns and gradients in key salinization hotspots. This approach provides a practical, scalable tool for monitoring land salinization, supporting precision agriculture, irrigation optimization, and land degradation assessment in vulnerable arid and semi-arid environments.