Mango Leaf Disease Classification with Transfer Learning, Feature Localization, and Visual Explanations.

Growing mangoes is an important part of life in southern India and a major economic driver for the area. Nevertheless, several leaf diseases often impede mango tree development and production, substantially affecting harvest output and quality. Detecting and identifying mango leaf diseases early can be challenging due to the diverse crop varieties, climatic circumstances, and numerous disease signs. While deep-learning methods have been developed to address this problem, they generally need help to detect illnesses across geographies and crop types. To tackle this difficulty, this research offers a transfer learning model that uses Explainable Artificial Intelligence (XAI) characteristics to identify and categorize leaf diseases. This research proposes MLTNet (Mango Leaf Disease Classification with Transfer Learning, Feature Localization, and Visual Explanations) in this study. Our study utilized a dataset from Southern India comprising 1,275 high-quality images of mango leaves affected by diseases like rust and powdery mildew, augmented to 11,480 images across 14 classes to enhance model training and robustness. This novel model utilizes Explainable Artificial Intelligence (XAI) techniques such that leaf disease detection and categorization may achieve higher levels of accuracy. The research work lies in the development of the MLTNet model, which integrates Explainable Artificial Intelligence (XAI) techniques with the ResNet50 architecture to enhance classification accuracy in mango leaf disease detection. This model uniquely employs advanced data pre-processing methods like Error Level Analysis and incorporates Grad-CAM for feature localization and visual explanations. We compared MLTNet's performance with state-of-the-art models like ResNet-50, VGG-16, and InceptionV3, focusing on accuracy, interpretability, and computational efficiency. MLTNet demonstrated superior performance, achieving a training accuracy of 94.3% and a test accuracy of 86.3%, which notably surpasses other models under similar conditions. This success is attributed to the model's ability to leverage complex features from the augmented dataset and the added interpretability provided by XAI techniques. [ABSTRACT FROM AUTHOR]