A Lightweight Dual-Stage Conv2D–UNet Framework for Real-Time Chili Leaf Disease Detection and Severity Estimation

A Computationally Efficient Framework for Real-Time Plant Disease Identification and Localization

Abstract

Timely crop disease detection is important to increase the productivity and reduce the economic loss in farming. For automated chilli leaf disease diagnosis, a Conv2D-based classification model and a UNet based segmentation model are combined for lightweight dual-stage deep learning framework in the proposed study. The severity of the disease was estimated at the pixel level by the Segmentation model, which was able to transgress the disease area from the sensors. The training set was augmented to improve the generalization, and the training, validation, and testing sets were divided into 80:10:10: The dataset contains 3,780 chilli leaf images. Results: Our Conv2D classifier attains an accuracy of 96.5% with F1-score of 0.96 and recall of 0.95, obtaining superior results over popular models like VGG16 and EfficientNet-B0 and is also lightweight with only 28 MB size and 2.4 million parameters. Results:The UNet model obtains a mean mIoU of 0.89 with the maximal IoU of 0.92 and the dice coefficient is 0.91, corresponding to a pixel accuracy of 94.8% for the Bacterial Spot lesions. The framework is lightweight, allowing on-line disease classification and severity prediction in precision agriculture