Explainable AI with SHAP and Grad-CAM: A Novel Approach to Visualizing Cassava Leaf Health using CIVE, ECI Indices, and an Advanced UNet Framework

This paper presents an in-depth study on the classification and segmentation of both healthy and diseased Cassava leaves by visualizing traditional image processing techniques with deep learning approaches, and introducing explainability to enhance model transparency. Agricultural monitoring typically relies on expensive multispectral sensors to assess plant health, but this research explores the use of affordable RGB cameras. Two vegetation indices, the Elliptical Color Index (ECI) and Color Index for Vegetation Extraction (CIVE), are applied to extract vegetation features from RGB images, with Otsu thresholding used to create binary masks for segmentation. While CIVE consistently outperforms ECI in segmenting leaf regions, both methods struggle to distinguish between healthy and diseased leaves effectively. To address these limitations, a UNet deep learning model was trained on the binary masks generated by the indices, achieving significantly improved segmentation accuracy through techniques like data augmentation and optimized hyperparameters. UNet outperformed traditional methods, reaching 97.36% accuracy, 93.14% precision, 94.27% recall, an F1-score of 96.11%, an IoU of 91.08%, and a Dice coefficient of 92.32%. A novel aspect of this study is the integration and comparison of Explainable AI (XAI) techniques, including Grad-CAM and SHAP, to provide visual insights into how the UNet model distinguishes between healthy and diseased regions. These tools offer a deeper understanding of the model’s decision-making process, with SHAP proving particularly effective in highlighting critical areas of focus. This comprehensive evaluation of traditional image processing, deep learning, and XAI methods contributes to a more accurate and comprehensive solution for plant health monitoring, addressing both performance and transparency challenges in agricultural applications.