Multi-scale feature progressive fusion network for remote sensing image change detection.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Deep learning, particularly Feature Pyramid Networks (FPNs), is prevalent in change detection (CD).
• Existing FPN-based CD methods struggle with accurately detecting complete change regions and precise boundary localization.
• This limitation hinders the effectiveness of CD in various applications.

Purpose of the Study:

• To introduce a novel Multi-Scale Feature Progressive Fusion Network (MFPF-Net) for improved change detection.
• To address the limitations of existing methods in detecting complete change regions and their boundaries.
• To enhance the accuracy and localization capabilities of deep learning-based CD.

Main Methods:

• Proposed the Multi-Scale Feature Progressive Fusion Network (MFPF-Net) with three key modules: Layer Feature Fusion Module (LFFM), Multi-Scale Feature Aggregation Module (MSFA), and Multi-Scale Feature Distribution Module (MSFD).
• Integrated bi-temporal image features with difference maps at each layer for richer semantic information.
• Employed direct aggregation of change maps and a progressive pyramid-structured fusion strategy for enhanced feature communication and contextual information.

Main Results:

• The proposed MFPF-Net demonstrated superior performance in change detection tasks.
• Experiments on CDD, LEVIR-CD, and WHU-CD datasets confirmed the method's effectiveness.
• The network accurately identified complete change regions and precisely located their boundaries, outperforming comparative methods.

Conclusions:

• MFPF-Net effectively overcomes the limitations of existing FPN-based CD methods.
• The innovative fusion strategy enhances semantic information and contextual understanding for superior change detection.
• The proposed network offers a significant advancement in deep learning-based change detection accuracy and localization.