Attention-guided CNN-SM: Sentinel-1/2 fusion in bare soil season improves soil organic matter prediction in cultivated black soils of Northeast China

Aim A novel two-dimensional convolutional neural network model integrating attention mechanisms, designated as CNN-SM, was developed to advance spatial prediction of soil organic matter (SOM) in cultivated soils under complex environmental conditions. Methods This study incorporated 154 georeferenced soil samples from Lishu County, Jilin Province, China, synergistically fused with: (1) spectral indices and derivative features from Sentinel-2 multispectral data, (2) Sentinel-1 SAR textural metrics acquired during the bare soil season, and (3) topographic derivatives, vegetation proxies, and climatic covariates. Feature selection was optimized through Out-of-Bag (OOB) estimation, with model performance rigorously evaluated across 11 feature combination scenarios. Results Results demonstrated that bare soil-season Sentinel-2 near-infrared (NIR) and shortwave infrared (SWIR) band reflectance, combined with Sentinel-1 polarization-enhanced texture features significantly outperformed conventional vegetation indices in contributing to the SOM prediction model. The CNN-SM framework achieved optimal performance (R² = 0.695, MAE = 2.800 g kg ^-1 , RMSE = 3.064 g kg ^-1 ) through attention-driven prioritization of microwave-optical feature synergies. Conclusions This multi-sensor digital soil mapping approach provides a paradigm shift in precision SOM estimation, offering actionable insights for spatially explicit soil health management in intensively cultivated black soil regions.