Explainable deep learning for accurate multi-crop leaf disease classification in intelligent agricultural systems

Abstract

Identifying crop leaf diseases accurately and in an automated fashion is pivotal for advancing agricultural productivity and tackling global food challenges. This study utilizes an explainable deep learning approach for multi-crop leaf disease classification incorporating five benchmark architectures VGG16, ResNet, EfficientNet, Inception, and Vision Transformer (ViT). The models' performance is evaluated using an extensive dataset of 55,824 images, covering 28 classes of diseases from five fruit crops. For each image, resizing, normalization, and augmentation were applied as preprocessing strategies to promote generalization. Among the models, the VGG16 network attained the highest classification accuracy of 97.68% and the highest F1-score of 0.977, surpassing ResNet (96.2%) and EfficientNet (96.5%). In comparison, ViT and Inception relied more on data and were lower performing at 80.06% and 51.29% accuracy, respectively. The explainable aspects of the framework were validated using grad-CAM techniques which localized the affected disease areas, assuring agricultural practitioners of the framework's reliability. These findings explain why less sophisticated convolutional networks were able to deliver better performance than complex transformer networks for tasks involving agricultural images. The proposed explainable model offers a scalable and interpretable solution for early disease detection, supporting precision farming initiatives.