Denoising Diffusion Probabilistic Models and Transfer Learning for citrus disease diagnosis

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Deep learning for plant disease diagnosis faces challenges due to environmental variations and limited real-world data.
• Accurate training of plant disease diagnostic models is hindered by insufficient sample information.

Purpose of the Study:

• To evaluate Denoising Diffusion Probabilistic Models (DDPM), Swin Transformer, and Transfer Learning for diagnosing citrus diseases with small datasets.
• To assess the feasibility and effectiveness of proposed deep learning methods in addressing data scarcity for plant disease identification.

Main Methods:

• Method 1: DDPM for synthetic image generation, followed by Swin Transformer pre-training and transfer learning fine-tuning on original images.
• Method 2: Transfer learning using a pre-trained Swin Transformer fine-tuned on a dataset augmented with original and DDPM-generated synthetic images.

Main Results:

• Method 1 achieved 96.3% validation accuracy; Method 2 achieved 99.8% validation accuracy.
• Both methods successfully mitigated model overfitting in small-dataset scenarios.
• The proposed methods showed superior performance compared to VGG16, EfficientNet, ShuffleNet, MobileNetV2, and DenseNet121 for citrus disease classification.

Conclusions:

• DDPM-based data augmentation combined with Swin Transformer and transfer learning offers a robust solution for citrus disease diagnosis with limited data.
• The proposed approaches significantly improve diagnostic accuracy and overcome common limitations in agricultural machine learning applications.
• Method 2, utilizing transfer learning on an augmented dataset, demonstrates exceptional effectiveness, achieving near-perfect accuracy.