EXPLAINABLE EDGE AI FOR EARLY DETECTION OF PLANT DISEASES IN SMART AGRICULTURE USING LIGHTWEIGHT MOBILENETV2 WITH GRADIENT SALIENCY

Abstract

Plant diseases are a significant global threat to food security, resulting in annual crop losses that account for 20-40% of total agricultural products. Early and accurate plant disease detection is of primary importance for timely intervention and sustainable agricultural practices. In this paper, a new explainable edge artificial intelligence (XAI-Edge) framework for early plant disease detection in smart agriculture environments is proposed. The proposed system utilizes a light-weight MobileNetV2 architecture, where the network's width multiplier "alpha" is set to 0.75, and the model is trained on the New Plant Diseases Dataset, comprising 87,867 images from 38 plant disease classes. The proposed system focuses on a specific set of 10 plant disease classes, comprising 19,189 training images, covering early and late stages of plant diseases, including tomato, potato, corn, and apple crops. The proposed system achieves outstanding classification performance, yielding 98.54% accuracy, 98.54% F1-score, 98.56% precision, and 98.54% recall on the test set. A two-phase transfer learning approach, where the final 30 layers of the network are fine-tuned, along with a lower learning rate of 1e-5, is utilized for optimal performance. The proposed system utilizes gradient-based saliency mapping for model explainability, allowing farmers to visualize the regions of plant diseases.