Resource Prudent CNN Models for Disease Identification of Rice Crops

Abstract

Agriculture is one of India’s most promising and vital sector, where we secure almost 17% of GDP. Over 60 % of the population is employed in the field of agriculture.Rice is one of the significant staple crops in India. The most prominent hurdle farmer faces are disease identification which affects the crop and reduces yield significantly. Northeast monsoon dropped more than 57% rainfall in Tamil Nadu by 2021, and the Rice crop diseases like Blast, False Smut, Bacterial Leaf Blight, and Brown Spot emerged, which caused a 33% loss in the year 2021. The dataset is collected from the rice field of Chengalpattu district affected by this Northeast monsoon and from Kaggle, UCI Machine learning repository. Early prediction can reduce these losses and increase the quality and yield of the crops. The advancements in computer vision technology can do this with one precise algorithm in deep learning perfected with time: Convolutional Neural Network (CNN). CNN with feature extraction techniques is deployed to classify the diseases early. The different CNN models are trained to classify images into five classifications of diseases i) Blast, ii) Bacterial leaf blight, iii) Brown spot, iv) False smut, v) Tungro, and one classification of healthy crops. When the trained CNN models are deployed in a mobile hand-held device, the size of the CNN model is a significant obstacle. ResNet family CNN models give the highest output accuracy but come at the cost of increased computational resources. In this research, different dense CNN models and MobileNet models are proposed, which give comparable results to ResNet models but take much fewer computational resources.