A Deep Learning Approach for Real-Time Detection and Classification of Crop Leaf Diseases to Support Sustainable Farming Practices

Abstract

Early and correct diagnosis of the crop leaf diseases is essential to guarantee agricultural output, reduce yield damages and to sustain agriculture. Traditional diagnostic techniques or disease diagnostics requiring manual examination are manual, subjective and cannot be applied in large-scale or real-time agricultural monitoring. Despite the recent progress in the field of deep learning, which has shown that high classification rates can be achieved with the help of deep learning in the field of plant disease identification, most of the present methods are only able to perform image-level classification, are not able to localize the disease, and cannot be deployed in real-time in the field. This study introduces a deep learning system of real-time detection and classification of crop leaf diseases that combines effective object detection and disease classification in a single system. The strategy proposed uses a one-stage detection model and an optimized convolutional backbone, data augmentation, and transfer learning to balance the accuracy, robustness, and computational efficiency with the proposed strategy. With standard performance metrics and real-time inference analysis the framework is tested on a curated dataset of about 6,500 samples of crop leaf images of five representative classes including healthy and diseased ones. Experimental data indicate good and consistent performance in terms of disease-wise and a false positive rate of 95.6 and F1-score of 95.4 respectively. The normalized confusion which is depicted in the normalized confusion matrix is highly dominant on the diagonal meaning that there is no inter-class confusion and the sensitivity is certain in all the categories of the disease. The presence of the correct localization of the symptomatic area of the leaves in various visual conditions with the help of qualitative detection is proved. The unified detection classification design proves to be effective as verified by comparative and ablation studies, and real-time assessment demonstrates an inference rate of 26.3 FPS, which is appropriate to be used in edge-based and in-field implementation. All in all the proposed framework will help in closing the gap between laboratory models of high accuracy and deployable real-time agricultural solutions. The method allows localizing the disease and diagnosing it within a short time, which contributes to the development of sustainable and precision farming systems, facilitates early intervention, specific treatment, minimizes the use of chemicals, and enhances crop management.