A team of researchers led by Professor XIE Chengjun and Associate Professor ZHANG Jie from the Hefei Institutes of Physical Science of the Chinese Academy of Sciences, inspired by causal inference, developed an innovative Decoupled Feature Learning (DFL) framework to address the challenge of distribution bias in crop pest recognition.

The research results were published in Pest Management Science.

Accurate pest recognition is crucial for smart agriculture, as it ensures crop health, yield, and quality while maintaining ecological balance. Despite advancements in deep learning for pest recognition, existing techniques struggle with distribution bias in training datasets for deep learning models, which often leads to over-reliance on background features rather than key pest characteristics.

To address this challenge, the team proposed an innovative DFL framework. It applies causal inference techniques to mitigate training data bias by constructing diverse training domains and employs the Center Triplet Loss to enhance the deep learning models' ability to capture core pest features across different domains.

The researchers tested DFL on three different datasets: the Li dataset, Dong's Few-Shot Pest Dataset, and the large-scale IP102 dataset. These datasets are collections of images used to train and evaluate the accuracy of pest recognition models.

Results showed that DFL significantly improved the performance of the deep learning models, achieving high recognition accuracies of 95.33%, 92.59%, and 74.86% on the three datasets, respectively.

Furthermore, visualizations of the results confirmed that DFL helped the models focus on key characteristics of pests, allowing them to maintain high accuracy even when the distribution of test data changed.

"This research represents a significant advancement in addressing data distribution bias and enhancing the reliability of deep learning in agricultural applications," said Professor XIE Chengjun, corresponding author of the work.

Figure 1. Decoupled Feature Learning (DFL) Framework. (Image by XIE Chengjun)

Figure 2. The results image from DFSPD. The last row showcases the attention area under the DFL framework. (Image by XIE Chengjun)