Remote Sensing Change Detection With Bitemporal and Differential Feature Interactive Perception.

Area of Science:

• Remote Sensing (RS)
• Geospatial analysis
• Computer Vision

Background:

• Transformers have shown success in remote sensing change detection (CD) due to their long-distance modeling.
• Current methods often use computationally expensive stacked transformers in UNet-style architectures.
• Existing approaches inadequately utilize ground semantic information by treating bitemporal or differential features separately.

Purpose of the Study:

• To develop a computationally efficient and effective network for remote sensing change detection.
• To improve the utilization of semantic information by integrating bitemporal and differential features.
• To address the high computational costs associated with transformer-based CD methods.

Main Methods:

• Proposed the Multiscale Dual-Space Interactive Perception Network (MDIPNet).
• Simplified stacked multi-head transformer blocks into single-layer, single-head attention modules.
• Introduced the Lightweight Parallel Fusion Module (LPFM) for efficient information integration.
• Developed the Cross-Space Perception Module (CSPM) to connect bitemporal and differential feature spaces.

Main Results:

• MDIPNet achieved state-of-the-art (SOTA) performance on three challenging datasets and an urban expansion scene.
• The proposed method effectively suppressed pseudo-changes and mined richer semantic consistency of the change of interest (CoI).
• MDIPNet demonstrated controlled computational costs compared to mainstream CD methods.

Conclusions:

• MDIPNet offers a significant advancement in remote sensing change detection.
• The network provides a balance between high performance and computational efficiency.
• The approach enhances semantic understanding and reduces false positives in change detection tasks.