Plant Disease Identification on Real-World Data using Deep Learning: A Comparative Study

Abstract

Correct and timely identification of disease in plants, especially for a country like India, where agriculture continues to serve as a cornerstone of its economy, is an indispensable tool demanding our solicitousness. Recently, researchers have started to employ autonomous real-time systems involving deep learning techniques for this purpose. However, the wide va-riety of heterogeneous diseases affecting crop yield continues to prove itself a mammoth task for farmers and stymies the researchers. The majority of the current state-of-the-art models utilize datasets like Plant Village, consisting of leaf images taken in a controlled lab environment, which do not serve as accurate representative data of the real-world scenario. Moreover, the effectiveness of state-of-the-art models like EfficientNetLite, which provided notable improvements in accuracy for similar deep learning applications, remains untested on plant disease datasets. Hence, an exhaustive study on the performance of various state-of-the-art models with varied training conditions on real-time datasets like PlantDoc is imperative to further this area of research. In this paper, we have explored state-of-the-art CNNs like InceptionResNet, EfficientNetLite_0, and VGG-19, under various parameter settings, image augmentations techniques, and loss functions on the real-time PlantDoc dataset. We have evaluated and presented a comparative analysis of the exhaustive combinations and performances gauged by accuracy, top-5 accuracy, F1 scores, and other inferences drawn from training and testing all these networks on 27 different classes of crop diseases. We infer that EfficientNetLite proved to be a most effective architecture, especially given its relatively smaller size. EfficientNetLite coupled with the focal loss function and Albumentaions augmentation library yielded the best results with an accuracy of 70.71%, mean F1-score as 0.7, and 95.57% top-5 accuracy, on a test set congruent with real-world relevance.