Abstract: Aiming at the limitations of existing foodborne pathogen detection methods, such as complex and time-consuming sample pretreatment and high reagent dependency, this study utilizes hyperspectral imaging (HSI) combined with the minimum redundancy maximum relevance - gradient boosting decision tree (mRMR-GBDT) series feature selection methods to conduct qualitative detection research on Escherichia coli, Salmonella, and Staphylococcus aureus in a mutton-derived mixed bacterial system. Firstly, by analyzing the composition structure of the microbial community in fresh mutton, the symbiotic bacterial structure of mutton origin was elucidated, and a multi-bacterial mixed system consistent with actual detection conditions was constructed. Secondly, HSI data of the mixed system composed of mutton-derived symbiotic bacteria and foodborne pathogens were acquired. One-dimensional spectra were extracted through mask generation and morphological methods. Then, the mRMR-GBDT series methods (including mRMR-XGBoost, mRMR-LightGBM, and mRMR-CatBoost) were used to perform feature wavelength screening related to pathogens on the preprocessed spectral data. Finally, the performance of different feature selection methods and classification models was systematically compared to determine the optimal feature extraction method and classification model. The research results showed that: the total bacterial colony count in 90.5% of mutton samples was distributed in the range of 10²～10⁴ CFU/g, with the relative abundances of the core dominant microbiota being 35.67% (Pseudomonas) and 35.56% (Acinetobacter), respectively. The mRMR-GBDT series methods obtained 17, 21, and 12 feature wavelengths, respectively. Among them, the 12 feature wavelengths selected by the mRMR-CatBoost method possessed optimal spectral relevance and minimal redundancy. Classification models (SVM, LightGBM, and BPNN) were constructed based on this feature subset. Through comparative analysis, the BPNN model performed best, achieving validation set and test set accuracies of 0.9763 and 0.9655, respectively. The results indicate that HSI combined with mRMR-CatBoost-BPNN enables efficient detection of foodborne pathogens in a mixed bacterial system derived from authentic mutton. This research can provide a theoretical basis and technical reference for the efficient and accurate detection of pathogenic bacteria in mutton and other animal-derived foods.