[Objective] Achieving non-destructive testing of sub-healthy watercore apples. [Methods] First, the logarithmic function method and the power function method proposed by this study were used to correct the sample spectra. Subsequently, the corrected data were converted into different images of gramian angular difference field (GADF), gramian angular summation field (GASF), markov transition field (MTF), recurrence plot (RP), symmetric dot pattern (SDP). Then, the ResNet50 network model with the convolutional block attention module (CBAM) was used to mine the deep image features related to the degree of watercore, which were downscaled by the t-distributed stochastic neighbor embedding (t-SNE) method and analyzed by clustering to determine the most suitable image transformation method. Finally, the most suitable image features were inputted into the improved particle swarm algorithm (IPSO) optimized support vector machine (SVM), extreme learning machine (ELM), k-nearest neighbour (KNN) and random forest (RF) classifier for the three-class classification of watercore apple. [Results] The results showed that the power function method was better than the logarithmic function method in eliminating the effect of diameter on the transmission spectrum. The silhouette coefficient (SC), the calinski harabasz score (CHS), and the davies- bouldin index (DBI) were 0.93, 0.88 and 0.24 after visualization of the image features in the GADF, better than the rest of the image transformation methods. ResNet50-IPSO-ELM achieved the highest classification accuracy of 96.8%. The overall discrimination accuracy of the three watercore classes apples in the test set reached 96.3%, and the stable precision (SP), stable recall (SR), and stable F₁-score (SF) were 87.2%, 95.8%, and 92.3%, respectively. [Conclusion] The model maintains a high classification accuracy for the majority class of apples without watercore and healthy apples with watercore and a high discriminatory ability for the minority class of sub-healthy apples with watercore.