Explaining deep learning-based leaf disease identification

Crop diseases adversely affect agricultural productivity and quality. The primary cause of these diseases is the presence of biotic stresses such as fungi, viruses, and bacteria. Detecting these causes at early stages requires constant monitoring by domain experts. Technological advancements in machine learning and deep learning methods have enabled the automated identification of leaf disease-specific symptoms through image analysis. This paper proposes image-based detection of leaf diseases using various deep learning-based models. The experiment was conducted on the PlantVillage dataset, which consists of 54,305 colour leaf images (healthy and diseased) belonging to 11 crop species categorized into 38 classes. The Inception-ResNet-V2-based model achieved a 10-fold cross-validation accuracy of \(0.9991 \pm 0.002\) $0.9991\pm 0.002$ outperforming the other deep neural architectures and surpassing the performance of existing models in recent state-of-the-art works. Each underlined model is validated on an independent cohort. The Inception-ResNet-V2-based model achieved the best 10-fold cross-validation accuracy of \(0.9535 \pm 0.041\) $0.9535\pm 0.041$ and was found statistically significant among other deep learning-based models. However, these deep learning models are considered a black box as their leaf disease predictions are opaque to end users. To address this issue, a local interpretable framework is proposed to mark the superpixels that contribute to identifying leaf disease. These superpixels closely confirmed the annotations of the human expert.