PaddyGuard: An Uncertainty-Aware Swin Transformer with Deterministic Test-Time Augmentation for In-Field Detection of Multiple Diseases of a Paddy Plant

Paddy diseases threaten rice production across Asia, and deep learning models can identify them from field imagery with high accuracy yet most report accuracy alone, without calibrated probabilities or per prediction uncertainty, which limits their trustworthy use in the field. We present PaddyGuard, an uncertainty aware framework for in-field detection of paddy diseases built on the Swin Transformer. A Swin-S backbone pre trained on ImageNet-22K is fine-tuned on the Paddy Doctor benchmark 10,407 field-collected images covering nine diseases and healthy plants using Mixup, Cutmix, and label smoothing. At inference, deterministic test-time augmentation (DTTA) applies a fixed set of eight geometric transforms and interprets the variance across their predictions as a reproducible uncertainty estimate, avoiding the stochastic sampling of MC Dropout; probabilities are further calibrated by temperature scaling. PaddyGuard attains 98.4% validation accuracy and 98.16% test accuracy, outperforming ResNet-50 / EfficientNet/ViT baselines by 9.4 points while requiring under 350ms per image. Predictions are mapped to plant organs, aggregated into a severity-weighted health score, and explained with Grad-CAM, delivered through a web interface and REST API. The results show that hierarchical shifted window attention combined with deterministic uncertainty estimation offers a practical recipe for trustworthy, deployable plant-disease detection.