Automatic Classification Service System for Citrus Pest Recognition Based on Deep Learning.

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Pathology

Background:

• Plant diseases significantly reduce agricultural output and cause economic losses.
• Citrus crops are vital globally but susceptible to pests and diseases, impacting yield and quality.
• Early detection of citrus diseases is crucial to prevent pest spread and minimize crop damage.

Purpose of the Study:

• To address the lack of comprehensive datasets and limited pest types in citrus disease research.
• To develop and validate an accurate and efficient automated system for citrus pest and disease detection.
• To create a practical tool for aiding farmers in recognizing and classifying citrus diseases.

Main Methods:

• A dataset of 20,000 citrus pest images (fruits and leaves) was self-collected from cultivation sites.
• Transfer learning with five steps was applied to train, verify, and test models on the dataset.
• The best-performing EfficientNet-b0 model was used to build a web application for disease diagnosis.

Main Results:

• Models achieved an average accuracy of over 97% and an average F1 score of over 96%.
• The developed web application correctly classified diseases using both self-collected and external image samples.
• The EfficientNet-b0 model demonstrated superior performance among the tested transfer learning models.

Conclusions:

• The automated citrus pest and disease diagnosis web system provides a valuable auxiliary tool for disease recognition.
• The system's high accuracy and performance can contribute to improved citrus fruit quality and reduced crop damage.
• This research facilitates the practical application of computer vision in citrus cultivation for disease management.