Waveshift Augmentation: A Physics-Driven Strategy in Fine-Grained Plant Disease Classification

Recent advancements in computer vision and machine learning enable automatic and precise plant disease diagnosis. However, the necessity for meticulous data segmentation by origin and the demand for extensive data collection lead to performance degradation in even state-of-the-art models. These issues arise because symptom-identifying features are often weak and diverse, and domain differences can overshadow these distinctions. To address these challenges, this study introduces Waveshift augmentation, a novel online augmentation technique inspired by the natural phenomenon of light propagation. The method aims to enhance model generalization by simulating light dynamics to improve adaptability to varying image conditions. Our methodology involves applying Waveshift augmentation to a large-scale fine-grained private plant dataset (tomato, strawberry, cucumber, and eggplant) and several public datasets, complemented by geometric transformations and contemporary techniques such as CLAHE, AugMix, and RandAugment. The results demonstrate that Waveshift augmentation achieves significant improvements in diagnostic performance, with notable macro F1 score gains across different backbone architectures including EfficientNetV2 (1.37 points), Swin Transformer (1.35 points), and ConvNeXt (0.80 points). Moreover, our method maintains competitive computational efficiency and supports real-time applications. This study highlights the effectiveness of Waveshift augmentation in overcoming generalization challenges and opens avenues for its application in various image-processing disciplines. Future work will focus on refining the hyperparameter optimization and exploring the impact of varying image shapes and asymmetric propagators.