A Multimodal Deep Learning Framework Fusing 1D Spectra, 2DCOS Maps, and Environmental Covariates for Soil Organic Carbon Prediction in Disturbed Mining Areas

Conventional soil organic carbon (SOC) measurements are laborious and destructive, constraining their suitability for large‐scale monitoring. Pedotransfer functions (PTFs) based on soil spectral information provide a rapid and nondestructive alternative, but their accuracy is often reduced in heterogeneous soils and across diverse regions. To address these challenges, this study introduces a novel multimodal deep learning framework that integrates environmental covariates, one‐dimensional spectra, and 2DCOS images to enhance SOC prediction. The fusion of these diverse data sources improves the extraction of complementary features, representing a significant advancement over conventional unimodal models. The framework was applied to 408 surface soil samples (0–20 cm) collected from disturbed mining areas in central Huaibei, Anhui Province, China, which are characterized by multiple subsidence patterns. The fusion of 2DCOS, spectral data, and environmental covariates achieved the best performance, with the coefficient of determination of 0.898 and the root mean square error of 1.17. Compared with classical PTFs and widely used machine‐learning models (e.g., XGBoost, GBDT, Random Forest), the framework improved the SOC predictive performance, highlighting its potential for SOC monitoring in mining‐disturbed soil.