ABSTRACT
Many solutions have been proposed to classify plant diseases in remote areas using traditional machine learning methods, however, the accuracy of their output was relatively poor. Due to the fact that the majority of the early efforts were based on small, self-curated datasets, their performances were not comparable. The scope of this problem was significantly reduced by using the “Plant Village” collection, which has 54,309 images of 26 diseases and 14 different types of crops. In order to effectively remove the background of the image, U2Net architecture was used. This algorithm helps to attain high resolution with less memory. Convolutional neural networks (CNN), a recently developed deep learning-based method, is a powerful tool for any classification task. Thus, selecting a suitable CNN model for plant disease identification was the final objective. We have evaluated the accuracy of four distinct CNN models, which include traditional CNN, VGG16, ResNet, and MobileNet. According to our findings we found that among the models ResNet had an average accuracy of 99.6% and a loss of less than 0.1%. When the diseases were identified using deep learning algorithms, appropriate treatments for each ailment were also suggested.
