Using a Resnet50 with a Kernel Attention Mechanism for Rice Disease Diagnosis.

Abstract

The domestication of animals and the cultivation of crops have been essential to human development throughout history, with the agricultural sector playing a pivotal role. Insufficient nutrition often leads to plant diseases, such as those affecting rice crops, resulting in yield losses of 20-40% of total production. These losses carry significant global economic consequences. Timely disease diagnosis is critical for implementing effective treatments and mitigating financial losses. However, despite technological advancements, rice disease diagnosis primarily depends on manual methods. In this study, we present a novel self-attention network (SANET) based on the ResNet50 architecture, incorporating a kernel attention mechanism for accurate AI-assisted rice disease classification. We employ attention modules to extract contextual dependencies within images, focusing on essential features for disease identification. Using a publicly available rice disease dataset comprising four classes (three disease types and healthy leaves), we conducted cross-validated classification experiments to evaluate our proposed model. The results reveal that the attention-based mechanism effectively guides the convolutional neural network (CNN) in learning valuable features, resulting in accurate image classification and reduced performance variation compared to state-of-the-art methods. Our SANET model achieved a test set accuracy of 98.71%, surpassing that of current leading models. These findings highlight the potential for widespread AI adoption in agricultural disease diagnosis and management, ultimately enhancing efficiency and effectiveness within the sector.

Plant phenotyping relevance

イネ葉の病徴という植物状態を画像から分類する新規深層学習手法を開発・交差検証しており、植物病害表現型の取得・推定が中心である。

Code and data availability

The paper's phenotyping input is a public rice disease leaf image dataset (Kaggle, HuyDo 2019) used to train and evaluate the SANET model; the dataset is explicitly cited with a public URL. No author analysis code or trained model release is mentioned.