Prediction of maize crude fat content based on improved conditional mutual information maximization and SHAP analysis

Traditional conditional mutual information maximization (CMIM) algorithms struggled to capture nonlinear dependencies in continuous near-infrared (NIR) spectral analysis. Therefore, this study proposed a novel framework combining improved CMIM with SHAP for rapid maize crude fat content prediction. Innovatively replacing discretization with continuous variable estimates and two-stage filter-wrapper feature selection. The CMIM_KDE method achieved R²p of 0.7618 and 0.7531 in the PLSR and SVR models, respectively, which was an average improvement of 6.18 % and 4.42 % over other improved strategies. In addition, SHAP revealed the screened feature wavenumbers around 5684 cm^-1 and 4312 cm^-1 were related to the CH group of maize. The validity and generalizability of the method was validated on publicly available datasets. This study solved the applicability bottleneck of traditional CMIM method in NIR spectral analysis. It could provide technical support for the quality detection of other agricultural products as well as low-altitude remote sensing and real-time online prediction.