Identification of Paddy Blast Disease Field Images Using Multi-layer CNN Models

Abstract

Farmers and agricultural experts can take action on many areas of paddy crop handling and management practices with the use of actionable information from the in-field diagnosis of paddy blast disease. To successfully diagnose the blast disease affecting fifteen different paddy crop varieties, three transfer learning multi-layer convolutional neural network (CNN) models, such as, CapsNet, EfficientNet-B7, and ResNet-50 are presented in this paper. The field images of blast disease are captured and classified based on disease severity levels, such as low, medium, high, and severe. The study employing the CapsNet model with dataset consisting a total of 20,000 labeled images demonstrate the significant results with the testing efficiency of 90.79% and validation efficiency of 93.29%. The ResNet-50 and EfficientNet-B7 models have yielded the average testing efficiencies of 85.10% and 88.72%, respectively. On the held out blast disease affected paddy field image dataset, the CapsNet model outperformed the EfficientNet-B7 and ResNet-50 CNN models related to both classification efficiency and computational efficiency.

Plant phenotyping relevance

圃場画像からイネいもち病の重症度を推定し、複数CNNモデルを比較・検証することが研究の中心であるため、植物病害フェノタイピング手法に該当する。

Code and data availability

The paper's paddy blast field-image dataset (3,987 captured images augmented to 20,000 labeled images across four severity levels) is explicitly declared publicly available on Kaggle by the authors in the Data availability statement. No code or trained model deposit is stated.