MustardFSL-Binary: A Few-Shot Learning Framework for Binary Mustard Leaf Disease Detection Using Vision Transformer and CBAM Attention

Abstract

Mustard (Brassica juncea) is an important oilseed crop in South Asia, contributing almost 28% of India’s edible oil supply. Fungal, bacterial and viral diseases, together known by the term “unhealthy” leaf conditions, can decimate crop yield by 20-70% unless detected early. Here, we provide a comprehensive research framework for binary mustard leaf disease detection using Few-Shot Learning (FSL), built upon a personalised real-world dataset organised into train/healthy, val/healthy and val/unhealthy splits. Traditional deep learning requires many thousands of labelled samples per class. Yet, this constraint does not always hold and posing a challenge when dealing with new diseases or in areas with low geographical diversification. We introduce a new framework called MustardFSL-Binary, that leverages a pre-trained Vision Transformer (ViT-B/16) backbone augmented with a Convolutional Block Attention Module (CBAM) and Prototypical Network-based meta-classification. The training method uses the plentiful healthy class to form a high-level “normal” embedding anchor, it treats the unhealthy class as a novel few-shot target, just right for the asymmetries of the train/val structure of the dataset. Three complementary FSL strategies are tested: Prototypical Networks, One-Class SVM with ViT features and MAML fine-tuning in 1-shot and 5-shot settings on the validation split. In the 5-shot scenario, ideal settings achieve 89.4% binary accuracy, with realistic deployment opportunities towards smartphone-based field usage.