An Efficient and Robust Few-Shot Information Fusion for Paddy Disease Detection

Abstract

Accurate recognition of paddy leaf diseases is essential for precision agriculture but remains challenging due to limited labeled data, complex lesion patterns, and substantial variability in field environments. Conventional deep learning methods require large annotated datasets, while existing few-shot learning (FSL) approaches often struggle to preserve lesion-discriminative features and maintain stable similarity reasoning under real-world agricultural conditions. To address these limitations, this article proposes FSProtoFusion, a novel information-fusion-driven FSL framework that integrates hybrid attention fusion (HAF), dynamic support adaptation (DSA), and adaptive metric fusion (AMF) for robust paddy leaf disease recognition. Experiments were conducted on a curated multisource paddy leaf disease dataset containing 15 000 images across ten disease classes, compiled from PlantVillage, Mendeley, Kaggle, and GitHub repositories. Under a ten-way few-shot evaluation protocol, FSProtoFusion achieved accuracies of 56.7%, 71.9%, 79.1%, and 85.8% in one-shot, five-shot, ten-shot, and 20-shot settings, respectively. The proposed method outperformed the strongest baseline (DETA++) by 7.8%, 6.7%, 6.7%, and 6.2% across the corresponding shot settings while also achieving superior $F1$ -score, robustness, and cross-domain generalization performance. Additional ablation and efficiency analyses confirm the effectiveness of the proposed components and demonstrate favorable computational cost. The results indicate that FSProtoFusion provides an effective solution for few-shot paddy leaf disease recognition under limited-data and heterogeneous agricultural environments.