Real-time jute leaf disease classification using an explainable lightweight CNN via a supervised and semi-supervised self-training approach.

Abstract

Introduction Timely detection of jute leaf diseases is vital for sustaining crop health and farmer livelihoods. Existing deep learning approaches often rely on large, annotated datasets, which are costly and time-consuming to produce. Methods and results To address this challenge, a lightweight convolutional neural network integrated with a semi-supervised learning self-training framework was proposed to enable accurate classification with minimal labeled data. The model combines modified depthwise separable convolutions, an enhanced squeeze-and-excite block, and a modified mobile inverted bottleneck convolution block, achieving strong representational power with only 2.24M parameters (8.54 MB). On a self-collected dataset of jute leaf images across three classes (Cescospora leaf spot, golden mosaic, and healthy leaf), the proposed model achieved a best accuracy of 98.95% under the supervised training with training, testing and validation split of 80:10:10. Remarkably, the model also attained a best accuracy of 97.89% in the semi-supervised learning (SSL) setting with only 10% labeled and 90% unlabeled data, demonstrating that near-supervised performance can be maintained while substantially reducing the dependency on costly labeled datasets. The application of explainable AI method such as Grad-CAM provided interpretable visualizations of diseased regions, and deployment as a Flask-based web application demonstrated practical, real-time usability in resource-constrained agricultural environments. Conclusion These results highlight the novelty of combining SSL with a lightweight CNN to deliver near-supervised performance, improved interpretability, and real-world applicability while substantially reducing the dependence on expert-labeled data.

Plant phenotyping relevance

葉画像から病害状態を推定する軽量CNNと半教師あり学習を開発・評価しており、植物病害フェノタイピング手法が研究の中心である。

Code and data availability

The paper's self-collected jute leaf disease dataset (920 images, three classes) is publicly available on Kaggle via the authors' data availability statement.