Improving Image-Based Plant Disease Classification With Generative Adversarial Network Under Limited Training Set.

Area of Science:

• Agricultural Science
• Computer Science
• Machine Learning

Background:

• Traditional visual plant disease identification is subjective and inefficient.
• Machine learning, particularly Convolutional Neural Networks (CNNs), shows promise but struggles with limited training data, leading to overfitting.
• Addressing data scarcity is crucial for reliable automated plant disease diagnosis.

Purpose of the Study:

• To improve plant disease recognition accuracy and mitigate overfitting in scenarios with limited training data.
• To evaluate the efficacy of combining Wasserstein generative adversarial network with gradient penalty (WGAN-GP) and label smoothing regularization (LSR).
• To present a robust solution for automated plant disease classification.

Main Methods:

• Utilized Wasserstein generative adversarial network with gradient penalty (WGAN-GP) for data augmentation.
• Integrated label smoothing regularization (LSR) to enhance model generalization.
• Employed a hybrid approach combining WGAN-GP and LSR for training Convolutional Neural Networks (CNNs).

Main Results:

• The proposed WGAN-GP enhanced classification method achieved a 24.4% improvement in plant disease classification accuracy.
• This method outperformed classic data augmentation (20.2% improvement) and synthetic samples without LSR (22% improvement).
• Demonstrated significant reduction in overfitting and improved prediction accuracy under data-limited conditions.

Conclusions:

• The combination of WGAN-GP and LSR effectively addresses the challenge of limited training data in plant disease recognition.
• This approach offers a substantial improvement over existing methods, enhancing the reliability of automated plant disease diagnosis.
• The study provides a valuable contribution to the field of precision agriculture and intelligent crop monitoring.