Cropland soil salinity retrieval using a spectral-spatial cross-attention deep learning framework with environmental interpretability

This is the authors' abstract. We don't add key points for this paper.

Cropland soil salinization severely constrains crop growth and leads to substantial yield reductions, thereby constituting a critical ecological threat to national food security. Owing to its extensive spatial coverage and high temporal efficiency, remote sensing has become an indispensable tool for large-scale soil salinity monitoring. However, conventional remote sensing-based retrieval approaches are challenged by weak salinity-related spectral signals, complex coupling between soil salinity and environmental factors, and limited model generalization capability, which collectively hinder accurate characterization of cropland salinity patterns. To address these limitations, this study develops a multimodal deep learning framework, termed Spectral and Spatial Soil Net (SS-SoilNet), to retrieve soil salinity in coastal croplands of the Yellow River Delta. To enhance retrieval accuracy and capture the nonlinear interactions between soil salinity and the geo-environment, the model integrates multi-source inputs, including remote sensing observations, topographic features, and crop growth parameters. Specifically, SS-SoilNet incorporates a spectral sequence convolution module to extract dynamic spectral characteristics, a spatial attention module to account for surface heterogeneity and environmental features, and a multi-head cross-attention mechanism in which spatial features serve as queries to selectively guide spectral and vegetation index representations. Experimental results demonstrate that: (1) Compared with conventional machine-learning models such as Random Forest and Support Vector Regression, SS-SoilNet exhibits improved and more stable predictive performance. On the validation dataset, SS-SoilNet achieves an RMSE of approximately 3.6 g kg −1 , an MAE of about 1.7 g kg −1 , and an R 2 of around 0.68. (2) The integration of multimodal deep learning with multi-source covariates, including hydrological conditions, crop vigor indicators, and topographic factors enhances retrieval robustness relative to spectral-only approaches. (3) Interpretability analysis of covariate contributions reveals pronounced coupling effects among soil salinity, crop growth, and terrain. Overall, SS-SoilNet not only improves the accuracy and interpretability of soil salinity retrieval in coastal croplands but also provides quantitative evidence for understanding the environmental processes underlying soil salinization.