Abstract: In order to accurately detect citrus infected by Bactrocera dorsalis（Hendel） using visible/near-infrared spectroscopy, this study designed a multi-path non-destructive detection grading system for citrus infected by Bactrocera dorsalis（Hendel） to address the unknown location of infection. Semi-transmissive spectral information from four detection paths was collected and partial least squares discriminant analysis was used to establish and compare classification models for a single detection path and a combination of four detection paths. The results showed that the model combining all detection paths achieved better classification results. Using 47 feature wavelengths selected by the competitive adaptive reweighted sampling method to build a model, the accuracy and specificity of the best prediction set among the four detection paths reached 93.5% and 95.2%, respectively. The PLS-DA hybrid classification model established by using citrus samples from four detection paths, combined with the CARS algorithm for effective feature wavelength variable selection, can improve the accuracy of Bactrocera dorsalis（Hendel） infected citrus classification model and accurately classify Bactrocera dorsalis（Hendel） infected citrus. The research results can provide a reference for online detection of citrus infestation by Bactrocera dorsalis（Hendel）.