Deep learning-based semantic segmentation of tomato leaf diseases using U-Net and classification of blight using ResNet

Abstract

Effective disease control requires the early identification and diagnosis of plant diseases, especially those affecting tomato leaves. A crucial stage in this process is segmenting images of diseased leaves, but this can be difficult because of the uneven shapes, varied sizes, vibrant colors, and frequently blurry borders of the affected portions, in addition to the messy backgrounds. We propose a deep learning-based strategy based on the U-Net architecture for addressing these issues, enabling precise segmentation and timely identification of tomato leaf diseases. With a DICE score of 0.93 and an accuracy of 93% in identifying healthy from diseased locations, this technique shows promise in helping farmers carry out focused interventions. Furthermore, the ResNet18 model has good levels of specificity, sensitivity, and accuracy when used to classify early and late blight. These outcomes highlight the way our suggested models perform in actual agricultural environments. Subsequent research endeavors will center on augmenting the model's generalizability in various agricultural settings and investigating multi-modal imaging methodologies. It is also intended for the advancement of mobile applications and edge computing to enable real-time disease monitoring and sustainable farming methods worldwide.