Evaluating Data Augmentation Techniques for Coffee Leaf Disease Classification (2401.05768v1)

Abstract: The detection and classification of diseases in Robusta coffee leaves are essential to ensure that plants are healthy and the crop yield is kept high. However, this job requires extensive botanical knowledge and much wasted time. Therefore, this task and others similar to it have been extensively researched subjects in image classification. Regarding leaf disease classification, most approaches have used the more popular PlantVillage dataset while completely disregarding other datasets, like the Robusta Coffee Leaf (RoCoLe) dataset. As the RoCoLe dataset is imbalanced and does not have many samples, fine-tuning of pre-trained models and multiple augmentation techniques need to be used. The current paper uses the RoCoLe dataset and approaches based on deep learning for classifying coffee leaf diseases from images, incorporating the pix2pix model for segmentation and cycle-generative adversarial network (CycleGAN) for augmentation. Our study demonstrates the effectiveness of Transformer-based models, online augmentations, and CycleGAN augmentation in improving leaf disease classification. While synthetic data has limitations, it complements real data, enhancing model performance. These findings contribute to developing robust techniques for plant disease detection and classification.

• The paper demonstrates that combining offline (CycleGAN) and online (MixUp, CutMix) augmentations enhances performance on the class-imbalanced RoCoLe dataset.
• The methodology leverages pix2pix for leaf segmentation and uses Vision Transformers and ResNet to boost accuracy, precision, recall, and F1-score.
• The findings highlight the potential of generative models to balance scarce agricultural data, offering valuable insights for automated plant disease monitoring.

Evaluating Data Augmentation Techniques for Coffee Leaf Disease Classification

The paper in question tackles the essential task of leaf disease classification, specifically targeting Robusta coffee leaves using the RoCoLe dataset. This dataset presents significant challenges due to its class imbalance and the limited number of samples. The research proposed a robust methodology that deploys data augmentation techniques combined with state-of-the-art deep learning models, including pix2pix, CycleGAN, and various Transformer-based architectures.

Methodological Approach

The primary contributions of the paper involve an in-depth exploration of data augmentation strategies to address the small and imbalanced RoCoLe dataset. The paper introduces both offline and online augmentations. Offline augmentation was achieved using the CycleGAN model to generate synthetic diseased images, thus balancing the dataset by increasing samples in the underrepresented classes. Conversely, online augmentations included various techniques like MixUp, CutMix, and others, applied during the training phase to enrich the dataset diversity dynamically.

For image segmentation, a pix2pix model was employed to mask and preprocess the images, ensuring that classifiers focus solely on the leaf regions, which are of consequential interest when diagnosing leaf disease. The RoCoLe dataset's segmentation, followed by augmentation, set up the framework for training the classification models.

Experimental Findings

Extensive experiments were conducted using Vision Transformers (ViT) and ResNet models to assess the impact of augmentations on classification performance. Notably, augmentation strategies demonstrated substantial improvements in metrics such as accuracy, precision, recall, and F1-score. Specific configurations of Transform-based models showed superior performance over traditional CNN models in handling augmented datasets. The experiments further delineated that training exclusively on synthetic data led to overfitting specific artifacts in the dataset, highlighting the importance of real data inclusion.

ViT-small, configured with dropout and a lower learning rate, achieved better generalization. Furthermore, combination strategies such as the use of rotations and flips, particularly with FMix and CutMix augmentations, showed positive performance implications.

Implications and Future Directions

The results underscore the potential of generative models like CycleGAN in producing meaningful synthetic data that complements scarce datasets, improving model robustness against dataset imbalance. While the research primarily applied these methods to coffee leaf disease classification, the implications extend to other domains where dataset scarcity and class imbalance are prevalent issues.

Future work could explore the deployment of newer generative models like StarGAN, which can handle multi-domain translations, potentially offering a more streamlined solution for multi-class augmentations. As the research showed, alternative GAN frameworks might provide increased stability and efficiency in generating realistic synthetic samples.

The leaf segmentation process, by employing more nuanced semantic segmentation techniques, could further refine disease identification. While this increases computational demands, the expected precision improvements may justify these costs. Finally, experimenting with vision-LLMs could be insightful, given their state-of-the-art image classification capabilities.

In conclusion, the paper represents a comprehensive examination of data augmentation in plant disease classification, particularly under challenging conditions of class imbalance and data scarcity. The approaches discussed lay the groundwork for enhancing the efficacy of automated agricultural monitoring systems, contributing valuable insights into the broader application of deep learning frameworks in data-constrained environments.