Segmentation-Guided Hybrid Transformer with Prototype-Calibrated Learning for Field-Ready Plant Leaf Disease Diagnosis

Abstract

Plant leaf diseases compromise yield, quality, and farmer income, while late or inaccurate diagnosis drives excess pesticide use and production losses. Automated, image-based detection offers a scalable alternative to manual scouting, yet field images suffer from illumination shifts, background clutter, and class imbalance, which degrade model reliability. This work proposes an end-to-end pipeline for robust plant leaf disease recognition that combines physics-guided preprocessing with hybrid deep representations and class-aware learning. First, a preprocessing module applies illumination normalization (Retinex-inspired color constancy) and haze suppression, followed by Multi-Stage Attention Leaf Segmentation (MSALS) to isolate lamina and lesions from complex backgrounds. Next, for feature extraction, we introduce a Hybrid Shifted Vision Transformer (HS- ViT) that fuses a lightweight CNN stem (local texture cues) with cross-scale window-shifted transformer blocks (global lesion geometry), augmented by channel-spatial attention to emphasize symptomatic regions. Finally, classification uses a prototype-aware focal objective with temperature scaling to handle imbalance and sharpen decision boundaries; an uncertainty-weighted ensemble stabilizes predictions under domain shift. To improve minority classes, an Adaptive Augmentation module synthesizes realistic variations in lesion color, size, and spread. We evaluate on PlantVillage (~54k images, 38 classes) and a field collection from Telangana (chilli and sugarcane leaves; expert-annotated). The proposed system attains 99.0% top-1 accuracy on PlantVillage and 95.3% balanced accuracy in field conditions, exceeding recent CNN/ViT baselines by 2–6 percentage points and improving macro- F1 on rare diseases. The results indicate a practical path to deployable, edge-friendly diagnosis that can reduce chemical inputs and support precision agronomy at scale.

Plant phenotyping relevance

植物葉の病徴・病斑を画像から分離し、病害状態を推定する画像ベースの表現型解析パイプラインを開発・評価しており、フェノタイピング手法が研究の中心である。

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.