An open set model for pest identification.

Area of Science:

• Agricultural Science
• Computer Science
• Entomology

Background:

• Accurate agricultural pest identification is crucial for crop production.
• Traditional methods are time-consuming and require expert knowledge.
• Existing image-based machine learning models often need large, curated datasets and struggle with unknown pests.

Purpose of the Study:

• To develop an open set pest identification approach capable of rejecting irrelevant inputs.
• To create a practical and effective tool for identifying harmful tephritid fruit flies.
• To enable pest identification beyond trained datasets in open-world scenarios.

Main Methods:

• Collected and filtered tephritid fruit fly images using Inception-V3 and k-means clustering.
• Developed an EfficientNet-B2 model for closed-set identification of four major genera.
• Adapted the model for open-set recognition to handle unknown classes.

Main Results:

• The closed-set model achieved 89.65% accuracy in classifying four tephritid genera.
• The open-set model attained 86.48% overall accuracy and a 94.44% macro F1-score, including an unknown class.
• The model demonstrated capability in rejecting irrelevant inputs.

Conclusions:

• The proposed open set model is a practical and effective tool for identifying harmful fruit flies.
• The model can be easily integrated into existing agricultural pest control systems.
• This approach enhances pest identification capabilities in open-world agricultural settings.