Large-Scale Monitoring of Soil Organic Matter and Salinization Using Satellite Hyperspectral Imagery and a Deep Residual Attention Network

Soil degradation threatens environmental health and sustainable land use, with declining soil organic matter (SOM) and increasing salinization being two key manifestations. Accurate assessment of the spatial distribution of SOM and salinity is of great importance for both agricultural productivity and ecological conservation. This paper proposes a novel end-to-end deep learning model—the residual attention regression network (RARNet)—to enable quantitative estimation of soil parameters based on satellite-borne hyperspectral imagery. RARNet combines two-dimensional spectral convolution, residual feature learning, and attention mechanisms to jointly capture local spectral patterns and global spectral dependencies, thereby improving spatial–spectral feature representation. The experimental results obtained over a 17,000 km² black-soil region using a satellite hyperspectral dataset show the improved predictive performance of RARNet. Under the 5-fold cross-validation setting, the validation coefficients of determination (R²) are 0.7291 and 0.8368, the RMSE values are 2.0986 and 2.4806, and the MAE values are 1.7449 and 2.0152 for SOM and soil salinity, respectively. Compared with representative deep learning baselines, including LSTM and CNN with spectral attention, RARNet achieved higher average performance. It also improved R2by approximately 10%–15% compared with the ResNet backbone and by 15%–25% compared with the CARS-XGBoost method. The trained model was applied to the study area, generating large-scale spatial distribution maps of SOM and salinity. The mapping results revealed that the region suffers from low SOM levels and widespread salinization, with over half the area showing a salt content of 1.0 g/kg or higher. Overall, the proposed RARNet provides a feasible framework for large-scale mapping of SOM, salinization, and their spatial patterns in the study area. This demonstrates the potential of combining satellite hyperspectral imagery with deep learning for monitoring of soil properties and land degradation. Although further validation is required for other soil types and climates, the current findings provide a quantitative reference for land restoration and regional ecological risk assessment.