Explainable Deep Learning for Coffee Leaf Disease Classification in Smart Agriculture: A Visual Approach

Abstract

In recent years, the application of deep learning techniques for plant disease classification has become increasingly important for smart agriculture. Early classification and treatment of plant diseases are crucial for maintaining the quality and quantity of crops, and deep learning algorithms have the potential to provide accurate and efficient solutions to this problem. In this study, we used two datasets collected from Mendeley for coffee leaf disease classification. The first dataset contained two classes of images, while the second dataset contained the remaining three types of images. The two datasets were combined to create a more robust dataset for training the two-transfer learning and CNN model. We update the pre-trained model MobileNet and RestNet50 by adding a flattened and dense layer to classify the five types of coffee diseases. The model was trained using 100 epochs, and the performance was evaluated using evaluation metrics such as accuracy and loss error for training and validation. The best accuracy was obtained from ResNet50, then MobileNet, and then CNN. The accuracy is 100% for Proposed ResNet50, 98% for CNN, and 99% for MobileNet. To understand the affected areas and help farmers comprehend the reason for the prediction, we used visualization techniques such as Grad-CAM and Grad-CAM++. These techniques generated heat maps of the regions of interest, highlighting the areas that contributed most to the classification decision. Our study highlights the potential of using deep learning techniques for accurate and efficient plant disease detection. By automating the disease classification process, farmers can identify and treat diseases early, leading to better crop yields and quality. The results of our study demonstrate the effectiveness of the proposed method and can be used to develop efficient and cost-effective solutions for plant disease detection.