Potato is the fourth largest food crop in the world and has rich nutritional value. However, it is easy to be infected by the sickle fungus during storage and transportation, resulting in a huge waste of resources and economic losses. Therefore, it is necessary to quickly and accurately realize the early nondestructive detection of potato dry rot. When pathogenic bacteria infected the samples, they experienced the stages of healthy-gley stage-mild disease-severe disease. The gley stage was difficult to identify, mainly because the disease occurred quickly and no visible disease spots were formed on the surface, similar to the healthy samples. In order to realize the recognition of the gley stage of potato dry rot, this study combined hyperspectral imaging technology and deep learning to carry out early diagnosis of potato dry rot. The hyperspectral images of healthy potatoes and potatoes with different spoilage grades were obtained. Based on the ENVI, healthy parts and spots of samples with different grades of corruption were selected as regions of interest (ROI), and the average spectral value of ROI was calculated as the final spectral information of the sample. The Convolutional Neural Network (CNN) model was established with the spectral data as the input variable and the disease grade as the output variable, and the network structure was optimized. The results of different models were compared and analyzed, and the optimal model was selected as Model_3_3. Based on the optimal structure, the learning rate was optimized, and the Model_0.0001 has the best recognition effect, and its overall accuracy, accuracy, sensitivity and specificity are 99.68%, 99.76%, 98.82% and 99.54%, respectively. In order to further highlight the advantages of CNN in gley stage identification of potato dry rot, LSSVM, RF, KNN and LDAmodels were established. The results showed that the accuracy of the four conventional algorithm models were 90.77%, 92.30%, 93.10% and 92.34%, respectively, and the recognition rate of gley samples was 91.00%, 85.58%, 94.18% and 90.33%, respectively. For overall accuracy, the CNN model improves by 6.58%~8.91% compared with other conventional methods. Compared with conventional methods, the CNN model improved the recognition of gley samples by 5.55%～14.15%. The results show that hyperspectral imaging combined with CNN can effectively recognize the gley stage of potato dry rot, which provides a reference method for improving the intelligence level of early diagnosis of potato disease.