REDNet: an Explainable Deep Learning Framework for Mango Leaf Disease Classification

Mango leaf diseases need to be accurately identified to ensure high-quality mango production. Conventional testing methods take a lot of time and are at risk of mistakes, especially when identifying diseases that seem to be identical. To solve this problem, we propose REDNet, a deep learning approach for automated mango leaf disease detection. Based on a dataset of 3010 high-resolution images classified into five classes: Die Back, Bacterial Canker, Anthracnose, Healthy, and Gall Midge. We assess eight transfer learning models as a baseline in our study. Global average pooling, batch normalization, Gaussian noise regularization, and feature concatenation are all used in the model’s structure to increase generalization and minimize overfitting. When deploying in situations with limited resources, the model may take into account the additional architectural complexity and computing demands brought about by the integration of various backbone networks. To further increase interpretability and transparency, explainable artificial intelligence (XAI) methods, namely Gradient-weighted Class Activation Mapping (Grad-CAM), Score-weighted Class Activation Mapping (Score-CAM), and Eigen Class Activation Mapping (Eigen-CAM) are used to highlight the most significant areas of leaf images that affect the predictions that the model makes. The REDNet was the best among the tested models, with a classification accuracy of 99.668