Transfer learning for versatile plant disease recognition with limited data.

Area of Science:

• Computer Vision
• Agricultural Science
• Machine Learning

Background:

• Deep learning models for plant disease recognition require large datasets, which are costly and time-consuming to collect.
• Limited data availability is a major challenge for achieving high accuracy in plant disease recognition.
• Existing transfer learning methods often focus on limited datasets, hindering versatile application.

Purpose of the Study:

• To propose a novel transfer learning strategy for high-performance, versatile plant disease recognition across multiple datasets.
• To address the challenge of limited data in plant disease identification using advanced deep learning techniques.
• To develop a robust model applicable to various plant health monitoring tasks.

Main Methods:

• Utilized the large-scale PlantCLEF2022 dataset for pre-training a Vision Transformer (ViT) model.
• Implemented a dual transfer learning approach: pre-training on ImageNet with self-supervised loss and on PlantCLEF2022 with supervised loss.
• Applied the ViT model with the proposed transfer learning strategy to 12 diverse plant disease datasets.

Main Results:

• Achieved a mean testing accuracy of 86.29% across 12 datasets in a 20-shot scenario, surpassing the state-of-the-art by 12.76%.
• Demonstrated superior performance compared to popular methods across different dataset settings.
• Outperformed other methods in plant growth stage prediction and weed recognition tasks.

Conclusions:

• The proposed transfer learning strategy offers a significant advancement for versatile plant disease recognition.
• The approach effectively mitigates data limitations and enhances model generalizability.
• The public release of codes and pre-trained models aims to foster community development and applications.