<scp>RS</scp> ‐ <scp>EnvRFE</scp> ‐ <scp>PGR</scp> : A Novel Framework for High‐Precision Soil Organic Matter Mapping in Heterogeneous Black Soil Regions

ABSTRACT The Songnen Plain (SNP) and the Sanjiang Plain (SJP) are typical core black soil regions in northeastern China, where strong cross‐regional heterogeneity poses challenges for accurate soil organic matter (SOM) mapping. To address this issue, we compiled 493 soil samples from SNP and SJP and integrated multi‐temporal Landsat‐8 bare soil imagery (April–May, 2014–2022) with climatic and topographic covariates. A novel framework, termed remote sensing–environmental covariates–recursive feature elimination–plain‐based global regression (RS‐EnvRFE‐PGR), was developed to enhance cross‐regional SOM prediction. Results showed that: (1) the optimal modeling periods for SNP and SJP were April and May, respectively; (2) locally regressed models, constructed using environmental factors and feature selection, significantly outperform traditional global models in both prediction accuracy and stability; (3) key variables selected by RFE, including spectral, climatic, and terrain factors, highlight the dominant contribution of remote sensing data in SOM modeling, with precipitation showing stable performance across all models; (4) SOM in SNP exhibited a northeast–southwest decreasing gradient, while SJP showed a low‐center, high‐edge pattern. Further comparison with an advanced prior‐knowledge‐based hybrid mapping framework integrating attention‐based convolutional neural networks and convolutional long short‐term memory networks (A‐CNN‐ConvLSTM+PHM) confirmed the superior performance of the proposed RS‐EnvRFE‐PGR framework. Overall, this framework enhances the accuracy and adaptability of cross‐regional SOM mapping and provides methodological support for land quality regulation, carbon stock assessment, and sustainable agricultural management.