Research on Soil Organic Matter Content Prediction Model Based on Deep Learning

(1) Background: Traditional methods for determining soil organic matter (SOM) are expensive and current models also have high dimensional spectral data challenges. In this paper, we proposed a deep learning model called SAE-GS-TabNet.(2) Method: The study used 206 Guangxi soil samples and preprocessed hyperspectral data (400–2400 nm) with SG-FD-MSC. Key features were extracted with a sparse autoencoder (SAE) and TabNet hyperparameters were optimized dynamically using GSPSO algorithm.(3) Results: The SAE-GS-TabNet model demonstrated strong performance in the test phase, achieving an R2 value of 0.9547, an RMSE of 2.6321 g/kg, and an RPD of 4.6991.(4) Conclusion: Comparing DNN and TCN model results showed that SAE-GS-TabNet is more precise, stable and efficient. It is cost-efficient and non-invasive approach for soil analysis in southern red soil regions, potentially resulting in significantly improved soil properties.