Causality-inspired crop pest recognition based on Decoupled Feature Learning.

Area of Science:

• Agricultural Science
• Computer Science

Background:

• Crop pest recognition is vital for global agriculture and ecological balance.
• Deep learning methods show potential but suffer from biased training data, limiting accuracy.
• Existing models struggle with pest images from environments not represented in training datasets.

Purpose of the Study:

• To introduce a novel framework to overcome limitations in deep learning-based crop pest recognition caused by biased training data.
• To enhance the accuracy and reliability of automated pest identification systems in agriculture.

Main Methods:

• Developed the Decoupled Feature Learning (DFL) framework utilizing causal inference techniques.
• Manipulated training data based on classification confidence to create diverse training domains.
• Employed center triplet loss for learning discriminative class-core features.

Main Results:

• The DFL framework significantly improved baseline model performance.
• Achieved high recognition accuracies: 95.33% on Li, 92.59% on DFSPD, and 74.86% on IP102 datasets.
• Demonstrated superiority over standard baseline models in pest recognition tasks.

Conclusions:

• DFL effectively mitigates data distribution bias in pest recognition.
• The framework encourages models to focus on essential class-core features, improving generalization.
• DFL represents a significant advancement for reliable deep learning applications in agriculture.