A lightweight neural network framework for sustainable tomato leaf disease detection using explainable artificial intelligence

Diseases to tomato are impacting agricultural productivity and food security across the globe. This issue calls for efficient and interpretable detection systems. In this study, PlantNetLite architecture is proposed that integrates data preprocessing, InsightCAM, and a lightweight convolutional neural network model. InsightCAM serves as a critical component, effectively highlighting the significant parts of an image, which enhances interpretability and transparency in the classification process. The performance of InsightCAM is notable, achieving an Intersection over Union mean of 0.5824 with a standard deviation of 0.149, while requiring an average processing time of 1.7552 s. PlantNetLite model exhibits remarkable efficiency, showcasing very low training times compared to traditional deep learning models, making it accessible for real-time applications. Nonetheless, despite the simplicity of the design, PlantNetLite exhibits competing performance, with an accuracy of 97.76%, precision of 96.15%, and a recall of 96.02%. The proposed PlantNetLite framework with InsightCAM has high accuracy and interpretability. This emphasizes the scientific merit of integrating lightweight models with explainable AI in the context of tomato plant disease identification. Ultimately, this work aims to aid farmers and agronomists in making informed decisions to mitigate the impact of diseases on agricultural yield and food security.