Advancing plant disease classification using an attention-based CNN for intra-dataset and cross- dataset training

Abstract

Abstract The precise classification of plant diseases is crucial for ensuring food security for all people and boosting agricultural productivity. Although there has been significant progress in this field using deep learning approaches, cross-dataset training hasn’t drawn as much attention from researchers as intra-dataset training has. Moreover, very few models have successfully blended intra-dataset and cross-dataset training approaches. This paper proposes a novel attention-based Convolutional Neural Network (CNN) to overcome these limitations. The model improves feature extraction and classification accuracy across multiple datasets by using attention mechanisms. It was tested on five datasets (Digipathos, Northern Leaf Blight (NLB), PlantVillage, PlantDoc, and the CD&S dataset) that covered leaf diseases of both corn and potatoes. During intra-dataset training, the model achieved the highest classification accuracy of 99.38% when trained on images of potato leaves from the PlantVillage dataset. During cross-dataset training, the model exhibited the highest average classification accuracy of 82.93% for corn leaf diseases when trained on images from the CD&S dataset with their backgrounds removed. When compared to the techniques taken into consideration in this study under comparable experimental conditions, the results demonstrate improved performance. This study shows how the model may be flexible for both intra- and cross-datasets, offering a flexible way to categorize diseases that affect plants. Because of its ability to generalize across different datasets, it may be helpful in real-world agricultural applications with a wide variety of image quality and situations. This encourages the advancement of precision farming techniques and disease control.

Plant phenotyping relevance

植物葉の画像から病害状態を分類するCNN手法の開発・データセット間検証が中心であり、植物病害の表現型推定に該当する。

Code and data availability

The paper's plant disease classification experiments rely on five publicly available leaf-image datasets, each cited with an explicit public access URL in the reference list: PlantVillage (GitHub), PlantDoc (GitHub), Digipathos (Embrapa), NLB (SciDB), and CD&S (OSF). No author analysis code or trained model checkpoint,