Labour-driven grading of apples is inconsistent as two labourers will never assign the same rating to one and the same apple. Artificial Intelligence algorithms offer consistency but suffer from another drawback – imbalanced data. The Multi-spectral Apple Dataset includes 32,463 images taken using 8 discrete wavebands. The ripeness dataset includes 29,160 images. The variety dataset includes 1,683 images. Therefore, we have an imbalance ratio of 17:1. In the task of grading by sweetness, the Kinnaur variety has 53% less samples compared to the most represented class – Red Delicious. AppleNet, which is the baseline classifier, demonstrates 87-92% accuracy on balanced tasks and 65% on imbalanced. The major contribution of this work is in developing a diffusion-based approach to solve the class imbalance problem. First, the Denoising Diffusion Probabilistic Model is used to generate new samples corresponding to under-represented classes (Brix and minority varieties). Second, to combine multispectral and RGB images. Third, to develop an edge attention model to extract information about shapes, colors, and defects. Fourth, to use the LIME approach to provide the explanation for the grades assigned to fruits. Preliminary results prove that generated diffusion samples have distributions comparable to real samples. Moreover, edge attention provides ~0.93 precision and ~0.91 recall compared to the state-of-the-art approach AppleNet. Comparisons with popular architectures (MobileNetV3, Swin Transformer, ResNet) are being conducted, which will result in higher robustness against class imbalance and cross-validation tests.