Enhanced hybrid CNN and transformer network for remote sensing image change detection

Area of Science:

• Geosciences and Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Remote sensing (RS) change detection is crucial but faces challenges with high costs due to false negatives.
• Existing methods struggle to improve precision and focus on specific changes, leading to missed detections and discontinuity.

Purpose of the Study:

• To develop an enhanced hybrid network (EHCTNet) for improved RS change detection.
• To address limitations in focusing on changes of interest, reducing false negatives and discontinuity issues.
• To enhance feature learning and integrate frequency components for boosting recall.

Main Methods:

• Proposed an enhanced hybrid of Convolutional Neural Network (CNN) and Transformer network (EHCTNet).
• Employed a dual branch feature extraction module for multi-scale RS image features.
• Integrated frequency components using refined modules and an enhanced token mining module based on the Kolmogorov-Arnold Network.

Main Results:

• EHCTNet effectively mines complex change information of interest in RS images.
• Demonstrated superior performance over state-of-the-art models in detecting intact and continuous changed areas.
• Achieved more accurate neighboring distinction compared to existing methods.

Conclusions:

• EHCTNet significantly enhances feature learning and frequency component integration for RS change detection.
• The proposed method effectively reduces false negatives and improves the continuity and accuracy of detected changes.
• EHCTNet offers a promising solution for cost-effective and precise remote sensing change analysis.