Swin-BNN-RF: Hierarchical Attention-Based Probabilistic Framework for Mustard Leaf Disease Detection Using Swin Transformer Bayesian Neural Networks and Ensemble Learning

Abstract

Brassica juncea (Mustard) is one of the most economic seed vegetable crops of the world, playing a major role in the production of world edible oil and the agricultural economy. The third-largest producer is India, which had an area under cultivation of about 8.6 million hectares of Mustard in 2021 22, and annual revenue of over USD 5 billion in 2021 22. Nevertheless, the presence of diseases like Alternaria Leaf Spot, White Rust, Powdery Mildew, and Septoria Leaf Spot threatens yield and quality by up to an estimated 2070% loss every year, based on the severity of the disease, and thus economic losses are estimated at over USD 1.5 billion per year in India alone. Traditional diagnostic systems are based on a manual examination of an agronomist trained to look at the sample and make a judgment, which is time-consuming, subjective, and subject to human error. Current deep learning methods of automated disease detection, promising as they are, are prone to inaccuracies on complex disease patterns, poor uncertainty estimation that is essential in real-world implementation, and poor generalizability to different field conditions. In response to these drawbacks, Swin-BNN-RF, a hybrid framework that combines Swin Transformer as a hierarchical attention-based feature extractor, Bayesian Neural Network (BNN) with symmetrized posterior as a probabilistic learner, and a Random Forest (RF) as an ensemble classifier, is proposed in this study. The Swin Transformer also harnesses local and global spatial biases with its shifted window self-attention network, and it is able to extract better features on leaf images of high-resolution. The uncertainty estimates of the BNN component are trusted, and unambiguous predictions are highlighted to get the opinion of the human expert. Random Forest classifier uses the bagging and boosting ensemble methods to improve stability and the robustness of the classification. A large dataset was experimented with; it consisted of more than 10,000 samples per category of disease in four diseases. In the case of binary classification, the proposed model was 98.32% accurate, 98.52% precise, 98.70% recall, and 98.36% F1. On multi-class classification, it obtained 97.50, 97.82, 98.51, and 97.46 accuracy, precision, recall, and F1 score, respectively, which showed consistent performance in comparison with state-of-the-art models such as EfficientNet, MobileNet, and Residual Networks. The contribution of each component is verified by the Ablation studies and statistical analysis of significance (p

Plant phenotyping relevance

マスタード葉の画像から病害状態を推定する深層学習フレームワークを開発・評価しており、植物病害表現型の取得・分類手法が中心である。

Code and data availability

The paper states its image data came from two public Kaggle repositories (900 Mustard Leaf Dataset and Mustard Leaf Dataset by Bintang Pratomo) plus field-collected images, but no URLs, identifiers, or deposit links for those datasets are provided in the supplied blocks. No author analysis code, trained models, or data

No evidence-backed public reproduction asset is currently recorded.