Leveraging moisture elimination and hybrid deep learning models for soil organic carbon mapping with multi-modal remote sensing data

• Leverage multi-modal remote sensing data for soil organic carbon mapping. • Effective moisture elimination strategy by integrating clustering and radiative transfer model. • An advanced hybrid deep learning model, MCCL, was developed. • The novel paradium was verified in both humid and semi-humid regions. Precision management of soil organic carbon (SOC) is crucial for regulating the global carbon cycle and ensuring food security. Currently, SOC prediction remains challenging due to unresolved moisture disturbances, underutilized multimodal remote sensing data, and uncertain model transferability. To address these challenges, a new paradigm integrating moisture elimination with advanced hybrid deep learning has been developed. Firstly, this research employs optimal cluster analysis based on soil moisture spatial characteristics, followed by the application of the Kubelka-Munk radiative transfer model to construct a moisture elimination strategy (MES). Next, a hybrid deep learning model, Multimodal Transformer Mechanism-Convolutional Neural Network-Convolutional Long Short-Term Memory (MT-CNN-ConvLSTM, MCCL), is constructed to enhance predictive accuracy and generalizability. The MCCL model was compared to other machine learning and deep learning models, including LSTM, Random Forest (RF), CNN, Artificial Neural Network (ANN), Support Vector Machine (SVM) and partial least squares regression (PLSR). Results indicate that (1) the proposed paradigm achieves optimal SOC content prediction accuracy in humid regions, with a root mean square error (RMSE) of 3.58 g kg −1 , a coefficient of determination (R 2 ) of 0.76, a ratio of performance to interquartile distance (RPIQ) of 2.26, and a mean absolute error (MAE) of 4.73 g kg −1 . The model shows better performance in semi-humid regions, yielding an RMSE of 3.12 g kg −1 , R 2 of 0.77, RPIQ of 2.27, and MAE of 4.71 g kg −1 , indicating significant spatial transferability. (2) Under MES, multiple models showed improved R 2 using PLSR as the baseline: e.g., MCCL (41.4 %), LSTM (28.3 %), RF (17.2 %), CNN (14.1 %), ANN (8.1 %), and SVM (7.1 %). (3) The MES approach reduces RMSE by 1.06 g kg −1 and MAE by 1.58 g kg −1 , while increasing R 2 by 18.75 %, and RPIQ by 0.82. Using the KM radiative transfer model without cluster partitioning decreases RMSE and MAE by 0.58 g kg −1 and 0.23 g kg −1 , while increasing R 2 and RPIQ by 7.1 % and 0.3, respectively. Specifying the soil moisture gradient in the spectral correction process is crucial. The novel MES-MCCL paradigm proposed in this study is robust and provides promising insights into soil moisture masking’s spectral characterization and the potential of multimodal remote sensing for SOC monitoring.