Unsupervised Anomaly Detection in Plant Disease Localization

Early and accurate detection of plant diseases is essential for integrated pest management (IPM) and sustainable agriculture, yet most current deep learning methods rely heavily on annotated data and fail to generalize across crops and different stages of disease manifestation. In this work, we propose an unsupervised anomaly detection framework based on Attention U-Net to identify and localize plant diseases without requiring any manual annotations. The model is trained to reconstruct healthy leaf images from synthetically augmented inputs containing spatially constrained, realistic anomalies. We use a composite loss function combining mean squared error (MSE), structural similarity index (SSIM), and perceptual components to guide learning. Evaluation across a wide variety of crops demonstrates the model's ability to highlight disease-affected regions accurately while maintaining structural fidelity in healthy regions. The simplicity, scalability, and interpretability of our approach make it a promising direction for real-time, label-efficient disease monitoring in precision agriculture.