Tomatose: An Explainable Deep Learning Framework for Tomato Leaf Disease Detection

Abstract

Diseases that affect tomato leaves are a big danger to food security and agricultural output around the world. To manage crops well, we need reliable and timely ways to find them. This work outlines a thorough methodology for automated identification of tomato leaf diseases, employing deep learning approaches integrated with explainable AI. We present the Tomatose dataset, an innovative hybrid compilation featuring nine unique classes of tomato leaf states, encompassing eight disease categories and healthy specimens. This dataset is created by merging five existing comprehensive datasets to provide diversity and robustness for model training. We use a carefully tailored transfer learning architecture that uses many pre-trained convolutional neural networks as feature extractors. A thorough study of several topologies showed that DenseNet121 performed better than the others, with 97 % accuracy and strong generalization abilities without overfitting. To enhance model interpretability and agricultural practitioner trust, we integrated explainable AI techniques using Grad-CAM and Grad-CAM++ visualizations, which successfully highlighted critical diseaseaffected regions on tomato leaves, providing transparent insights into model decision-making processes and demonstrating the model's ability to focus on pathologically relevant features while minimizing attention to irrelevant leaf structures. This study enhances precision agriculture by offering a reliable, interpretable, and realistically implementable approach for detecting tomato leaf disease, tackling significant issues of model transparency and real-world applicability in agricultural contexts.

Keywords