A Cross-Layer Feature Fusion Framework with Hierarchical Interaction for Remote Sensing Change Detection.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Change detection (CD) in remote sensing (RS) is crucial but challenged by VHR imagery complexities.
• Existing hybrid Transformer-CNN models struggle with appearance variations, clutter, and illumination fluctuations.
• Limitations include difficulty modeling fine textures, feature continuity, and suppressing false positives.

Purpose of the Study:

• To propose a Cross-layer Feature Fusion Framework (CLFF) for accurate and stable CD in VHR imagery.
• To enhance the collaborative fusion of multi-layer features for improved change detection.
• To address limitations of traditional CNNs in multi-layer contextual modeling.

Main Methods:

• Introduced the Multi-level Interaction Perception Block (MP-Block) for semantic feature interaction.
• Utilized a Multi-branch Interaction Fusion Mechanism (MIFM) with parallel reconstruction and recalibration branches (RFRB, ACGF).
• Incorporated a lightweight position-aware attention module to refine spatial responses and suppress background noise.

Main Results:

• CLFF demonstrated significant performance improvements across four benchmark datasets (LEVIR, WHU, SYSU, HRCUS).
• IoU metric improvements ranged from 1.35% to 4.85% compared to baseline models.
• The framework effectively mitigated issues like limited receptive fields and insufficient multi-layer feature interaction.

Conclusions:

• The proposed CLFF framework significantly enhances change detection accuracy and stability for VHR imagery.
• Effective multi-layer feature fusion and interaction are key to overcoming challenges in complex remote sensing data.
• The method offers a promising advancement for reliable change detection applications.