A multi-source domain feature adaptation network for potato disease recognition in field environment

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate potato disease identification is vital for minimizing crop yield losses.
• Low recognition accuracy often results from domain mismatch due to insufficient data.
• Unsupervised domain adaptation techniques are explored to overcome these data limitations.

Purpose of the Study:

• To evaluate the effectiveness of Multi-Source Unsupervised Domain Adaptation (MUDA) for potato disease identification.
• To propose and validate a novel network, the Multi-Source Domain Feature Adaptation Network (MDFAN), for improved disease recognition.
• To assess the method's robustness against variations in image acquisition, such as lighting conditions.

Main Methods:

• A two-stage alignment strategy is employed within the MDFAN framework.
• Stage one involves aligning source-target domain distributions in multiple feature spaces using multi-representation extraction and subdomain alignment.
• Stage two aligns classifier outputs by leveraging decision boundaries within specific domains.

Main Results:

• The MDFAN achieved high average classification accuracies: 92.11% with two source domains and 93.02% with three source domains.
• The proposed method significantly outperformed existing techniques in transfer tasks.
• MDFAN demonstrated robustness to variations in lighting conditions encountered during field image acquisition.

Conclusions:

• Multi-Source Unsupervised Domain Adaptation (MUDA) is effective for potato disease identification.
• The MDFAN model offers a robust and accurate solution for identifying potato diseases, even with limited or varied data.
• The findings support the practical application of advanced machine learning techniques in agriculture for disease management.