Prediction of Diseases in Paddy Crop Using Machine Learning and Deep Learning

Abstract

Abstract Paddy crop diseases have a significant impact on the production of rice across the globe with immense losses in crop production as well as food security. Conventional disease diagnosis systems depend on visual inspection where the visual system is time-consuming, subjective and in many cases inaccurate when in the field. The use of artificial intelligence or, specifically, machine learning, and deep learning offers potent plant disease detection instruments in recent developments. The paper provides a detailed disease prediction model of paddy crop based on machine learning and deep learning. An acquired dataset in the form of field was used with images of the leaves of healthy and diseased paddy, which included rice blast, bacterial leaf blight, brown spot, and sheath blight. In the case of handcrafted machine learning models, color, texture, and shape attributes were obtained and categorized with the help of Support Vector Machine, Random Forest, k-Nearest Neighbors, and Logistic Regression algorithms. Custom Convolutional Neural Network and transfer-based architectures (ResNet50 and EfficientNet-B0) were both trained to serve as deep learning models. Experiment scores prove that deep learning models are much better than traditional machine learning classifiers where EfficientNet-B0 model with highest classification accuracy of 97.4% was made. The confusion matrix and learning curve results indicate good generalization that has been achieved by the models in realistic field conditions. The results indicate deep learning as a powerful tool to diagnose paddy disease automatically and as being applicable to smart farming to predict disease early and avoid losses caused by disease.

Plant phenotyping relevance

イネ葉画像から病害状態を自動分類する機械学習・深層学習モデルが研究の中心であり、植物病害という状態の推定手法を開発・評価している。

Code and data availability

The paper describes a ~5,000-image paddy leaf disease dataset and ML/DL models, but provides no public dataset URL, code repository, model checkpoint, or availability statement. No qualifying paper-specific public assets are present, and no allowed URLs were supplied.

No evidence-backed public reproduction asset is currently recorded.