Convolutional Rebalancing Network for the Classification of Large Imbalanced Rice Pest and Disease Datasets in the...

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate classification of crop pests and diseases is critical for effective agricultural management and food security.
• Field-collected image datasets often suffer from long-tailed distributions and severe category imbalance, hindering deep learning model performance.
• Existing deep recognition models struggle with the inherent challenges posed by imbalanced datasets in agricultural contexts.

Purpose of the Study:

• To propose a novel convolutional rebalancing network specifically designed for classifying rice pests and diseases from imbalanced field image data.
• To address the significant challenge of category imbalance in agricultural image datasets.
• To enhance the accuracy and robustness of pest and disease classification models in real-world field conditions.

Main Methods:

• Developed a convolutional rebalancing network incorporating a convolutional rebalancing module, an image augmentation module, and a feature fusion module.
• Implemented instance-balanced sampling and reversed sampling within the convolutional rebalancing module to handle category imbalance.
• Utilized an image augmentation module for effective data augmentation and a feature fusion module for comprehensive feature extraction.

Main Results:

• The proposed network achieved a high accuracy of 97.58% on a large-scale imbalanced rice pest and disease dataset (18,391 images).
• Demonstrated superior performance compared to state-of-the-art methods on various public datasets, including plant and pest image datasets.
• Verified the robustness of the network across different datasets, highlighting its generalizability.

Conclusions:

• The novel convolutional rebalancing network effectively addresses the challenge of imbalanced datasets in crop pest and disease classification.
• The proposed method offers a significant advancement for intelligent pest and disease control in agricultural fields.
• This network provides a valuable tool for improving the accuracy and efficiency of automated crop monitoring systems.