ResHAN-GAM: A novel model for the inversion and prediction of soil organic matter content

Soil organic matter (SOM) is crucial in determining soil health, improving crop production, and enabling sustainability in agriculture. Precise determination of SOM content is thus crucial for land management as well as for maintaining ecological equilibrium. This research introduces a new hierarchical attention mechanism that unifies residual networks with GAM attention. Through data smoothing and discretization in terms of fractions, the model is equipped to effectively repress noise as it enhances primary spectral features related to SOM, thus enhancing the robustness as well as explainability of the model. Hyperspectral reflectance data were recorded in the visible to near-infrared (Vis-NIR) range (350–2500 nm) with a high spatial-resolution sensor. The dataset is made available with samples from lands under cultivation for soybean as well as corn in the fertile black soil region. Experimental results indicate that the proposed method achieves an R 2 value of 0.945, an RMSE of 0.117% and RPD of 4.26 on the validation set. Furthermore, the model’s generalization ability was validated using the Land Use/Cover Area Frame Survey (LUCAS) dataset, a large-scale European soil database, where similarly high performance was achieved. These results highlight the effectiveness and transferability of the proposed method in estimating SOM content, offering a reliable, non-destructive tool for large-scale soil monitoring and environmental protection applications. • An innovative hierarchical attention-based model is used for SOM prediction with hyperspectral data. • Grouped training of spectral data achieves high accuracy, with R 2 up to 0.945. • Model generalization was validated using the LUCAS dataset, showing strong robustness. • Deep learning enables non-contact soil analysis for sustainable agriculture and soil health.