Explainable transfer learning-based approach for Mango fruit damage detection and cause identification

Abstract

This study presents a deep learning approach to detect mango fruit damage and identify its underlying causes using one-stage and two-stage detectors. Unlike existing methods that focus mainly on detection, this approach integrates XAI techniques to provide interpretable, cause-level insights, and is evaluated on a dataset annotated with anthracnose, bacterial black spot, fruit fly, and other damage classes. RetinaNet achieved the best overall performance, attaining a precision of 0.77, recall of 0.75, and F1-score of 0.76. Per-class accuracy exceeded 0.80 in all categories, except for fruitfly, which remained more challenging to detect. XAI techniques that include Grad-CAM, Grad-CAM++, LIME, SHAP, and LRP were further used to highlight vital regions influencing model decisions. A quantitative evaluation of XAI performance was further conducted using faithfulness, robustness, complexity, localization, and randomization. LRP and SHAP results attained are indicative of more faithful and well-localized explanations, while Grad-CAM based methods exhibited limitations in localization, with LIME producing highly complex explanations that result in reduced interpretability