Cross-modal integration framework for soil organic matter estimation using proximal and satellite spectral data: Modeling optimization with particle size effects and spatial similarity

• A cross-modal spectral reconstruction framework transfers proximal spectra to satellite imagery for SOM mapping. • Reconstructed hyperspectral images outperformed Sentinel-2 and Zhuhai-1 in SOM estimation accuracy. • The spatial similarity-based Random Forest local strategy improved R 2 by 7.64 %. • Hyperspectral modeling benefits from soil samples processed to finer particle sizes. Remote sensing (RS) technology enables the rapid and accurate acquisition of soil organic matter (SOM) content, which is crucial for ensuring food security and promoting precision agriculture. Multispectral imagery is widely used for large-scale SOM mapping, but its limited spectral resolution substantially constrains estimation accuracy. While proximal hyperspectral data provide detailed spectral information, their point-based observations limit scalability across large regions. To overcome these limitations, a cross-modal modeling framework integrating proximal hyperspectral and satellite spectral data was proposed. Hyperspectral reconstruction technology was applied to enhance satellite spectral resolution and to extend proximal hyperspectral observations into spatially continuous imagery, achieving a balance between spectral accuracy and spatial continuity. To address SOM spatial heterogeneity, a spatial similarity-based random forest (SS-RF) local modeling strategy was introduced. Furthermore, the study systematically evaluated the impact of different soil particle size levels on spectral reflectance and SOM estimation accuracy. This study was conducted in a typical black soil region located in Northeast China. A multimodal dataset was constructed for SOM modeling, including in-situ and laboratory hyperspectral data with multiple particle size treatments, as well as satellite imagery from Zhuhai-1 and Sentinel-2A. The results indicated that the proposed cross-modal fusion and SS-RF framework demonstrated superior SOM estimation performance. The reconstructed hyperspectral imagery effectively integrated proximal and satellite spectral data, thereby preserving spectral integrity and enhancing their correlation with SOM. Among these, the reconstructed imagery based on finer particle sizes (100 mesh, ≤0.15 mm) exhibited the best performance (R 2 = 0.874, LCCC = 0.756, RMSE = 2.871 g·kg −1 , and RPIQ = 2.159), while the reconstruction using 50 mesh particles (≤0.35 mm) also achieved comparatively good accuracy (R 2 = 0.864). In contrast, the model constructed using field in-situ hyperspectral reconstructed imagery produced the lowest accuracy (R 2 = 0.730). The estimation accuracy based on the reconstructed imagery was significantly higher than that achieved using Sentinel-2A (R 2 = 0.712) and Zhuhai-1 (R 2 = 0.759). Compared to traditional global models, the proposed SS-RF local strategy improved accuracy, increasing R 2 by 7.64 %. This synergistic optimization approach, which combines spectral reconstruction, local modeling, and particle size standardization provides new insights and technical support for high-precision SOM estimation at the regional scale.