SREM-CNN: A Near-Infrared Spectroscopy-Based method for Determining lignocellulosic composition in crop straw through spectral Reorganization and Cross-Region coupling learning

Near-infrared (NIR) quantitative analysis of complex lignocellulosic matrices is limited by the conventional treatment of spectra as undifferentiated one-dimensional sequences, which obscures physically related information distributed across overtone and combination-band regions. Here, we propose a Spectral Reorganization Enhancement Method (SREM) for spectroscopy-guided representation learning. Rather than selecting or reducing variables, SREM retains the full measured spectrum and reorganizes it into three channels corresponding to the combination band region, first overtone region, and second overtone region. After independent normalization and alignment, the reorganized spectrum is used as the input of a convolutional neural network to learn both intra-region local features and inter-region coupling patterns. The SREM-CNN model was evaluated using 550 straw samples from wheat, maize, rice, rapeseed, and cotton collected across multiple provinces and two years for simultaneous determination of cellulose, hemicellulose, and lignin. Compared with representative linear, kernel-based, ensemble, and deep learning baselines, SREM-CNN reduced independent-test RMSE by 13.64 %, 13.91 %, and 9.87 % for cellulose, hemicellulose, and lignin, respectively. SHAP-based wavelength attribution and inclusion–exclusion interaction analysis further showed that the strongest non-additive coupling occurred between the first overtone and combination-band regions. These results demonstrate that SREM improves NIR quantitative modeling by exposing cross-region overtone-combination band synergy rather than by variable reduction, providing a physically interpretable representation strategy for molecular spectrometric analysis.