Real-CD: Change Detection Under Real-World Complex Interference via Dynamic Distribution Correction.

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Change detection (CD) is vital for monitoring Earth's surface dynamics.
• Existing CD models struggle with seasonal variations and sensor interference.
• Current CD datasets lack diversity, limiting model generalization.

Purpose of the Study:

• To address the performance degradation of CD models in real-world conditions.
• To develop a robust CD method capable of handling distribution shifts.
• To create a comprehensive benchmark for evaluating CD models under diverse perturbations.

Main Methods:

• Constructed a benchmark with 75 interference-rich scenarios by adding perturbations (fog, snow, blur, noise) to existing datasets and including a seasonal dataset.
• Proposed Real-CD, a novel method leveraging bi-temporal correlations for adaptive distribution alignment.
• Introduced the Distribution Shifts Alleviation Module (DSAM) and Distribution Consistency Optimization Strategy (DCOS) to correct distribution shifts and stabilize training.

Main Results:

• Perturbations caused significant distribution shifts, leading to performance degradation in existing CD models.
• Real-CD demonstrated superior performance compared to state-of-the-art methods on the proposed benchmark.
• The proposed method effectively handles complex bi-temporal disturbances and heterogeneous real-world scenarios.

Conclusions:

• The developed benchmark highlights the limitations of current CD approaches in real-world settings.
• Real-CD offers a robust solution for change detection by addressing distribution shifts.
• The proposed method enhances the practical applicability of CD techniques for Earth surface monitoring.