UAV-BASED SOYBEAN DISEASE CLASSIFICATION & ANOMALY DETECTION

Abstract

Accurate and early detection of plant diseases is critical for sustainable agriculture and crop yield optimization. This study presents a comprehensive framework for soybean disease classification and anomaly detection using high-resolution UAV-based aerial imagery. We investigate two complementary deep learning approaches tailored to different aspects of disease detection. First, we implement a Vision Transformer (ViT)-based model for image-level classification, exploiting

its global attention mechanism to capture subtle disease patterns across complex canopy structures. Second, we deploy a Memory-Augmented Autoencoder (MemAE) for anomaly detection, which reconstructs healthy samples and flags deviations indicative of disease presence, offering a robust approach for scenarios with limited labeled data. The proposed multi-perspective methodology is designed to address key challenges in real-world agricultural monitoring, including label scarcity, intra-class variability, and the spatial complexity of field environments. The ViT-based classifier achieves strong performance on disease identification, while the MemAE highlights abnormal regions that diverge from learned healthy patterns, providing complementary insight. Extensive experiments demonstrate that the integration of these models facilitates robust, scalable, and interpretable disease monitoring from aerial data, data, establishing a powerful toolkit for precision agriculture applications.