Convolutional Neural Networks-Based Evaluation of Disease Severity in Grape Plant Using Colored and Grayscale Leaf Images

The primary objective of smart farming is to create innovative solutions for ensuring the long-term sustainability of the human population. One of the significant obstacles to achieving food security is the protection of crops from both biological as well as non-biological factors, with plant diseases posing a critical challenge. These diseases not only result in the destruction of crops and a decline in their quality but also necessitate the use of agrochemicals, which contaminate the soil over time, rendering it unsuitable for future cultivation. In order to optimize the agricultural output, the use of such chemicals should be minimized. This can be achieved, if the severity stage of a disease is known, so that a recommended amount of pesticide is applied, under the observation of expert agronomists and botanists. In this research work, two standard convolutional neural networks (CNN) have been implemented, namely SqueezeNet and ResNet50, for the evaluation of disease severity (early, middle, end) in two grape plant diseases, black rot, and leaf blight. Images of both colored and grayscale versions of diseased leaves were manually chosen from the PlantVillage dataset for each severity stage and used for training the models separately. SqueezeNet achieved the highest accuracy of 89.49% for black rot with colored images and 81.54% for grayscale images. Similar performance was observed with leaf blight images.