An explainable AI-based cotton leaf disease classification using EfficientNet, Grad-Cam, Lime & Shap

The quality of the cotton crop is reduced in terms of fibre, yield quality, and economic standard, especially due to leaf infections and disease symptoms, which spread in all field areas. The automated visualisation of disease screening from the infected images of leaves, due to limitations of the black box nature of the deep learning models, creates the less agriculture developments. This research provides a framework with integration of the deep learning and explainable AI-based approach for the detection and classification with the EfficientNet approach for the cotton crop disease, along with an explanation by adopting the Grad-CAM and SHAP discussion and explanation. The experimentation is performed using 800 Images with labelled leaf images of cotton crop disease obtained from Kaggle. The data preprocessing is used for the image resized as 224 x 224 pixels, augmented and normalised, with spilt into validation, training, and testing data subsets. The results show that after 20 epochs, the EfficientNet Model provides subtle results and is stable with 92% accuracy for image disease detection. The confusion matrix shows the 45 correctly healthy classified images, the disease leaves 43, false positives 5, and false negatives 7 by providing the 88.00% yielding test accuracy and 89.58% the precision of the disease class, recall 86.00%, f1 score 87.76%. The spatial heatmap, highlighted by the Grad-CAM, provides the symptoms of the leaf region. Whereas the pixel-level explanation is obtained by the LIME and summarises the visual contextual explanation of the feature from the image. The predictive performance is focused in this framework with transparency, reliability, and interpretability for the cotton crop diseases.