An interpretable crop leaf disease and pest identification model based on prototypical part network and contrastive...

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Computer vision algorithms aid in crop disease and pest identification from images.
• Current deep learning models lack interpretability, acting as "black boxes".

Purpose of the Study:

• To develop an intrinsically interpretable model for crop leaf disease and pest identification.
• To address the "black box" nature of existing deep learning approaches.

Main Methods:

• Proposed the Contrastive Prototypical Part Network (CPNet) model.
• Utilized similarity calculations between feature maps and prototype representations for decision basis.
• Employed supervised contrastive learning to enhance feature representation with limited data.

Main Results:

• CPNet demonstrated improved performance over baseline methods on four public datasets.
• The model successfully provided interpretable evidence for its identification decisions.
• Achieved accurate identification of disease and pest categories on crop leaves.

Conclusions:

• CPNet offers an interpretable alternative to traditional black-box models for crop disease and pest recognition.
• The approach enhances identification accuracy and provides visual insights into decision-making processes.
• Supervised contrastive learning effectively addresses data limitations in this domain.