A hybrid deep fuzzy clustering framework with fusion-based learning for robust SAR change detection

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Computer Vision

Background:

• Multi-temporal Synthetic Aperture Radar (SAR) data are crucial for environmental monitoring, urban analysis, and disaster assessment.
• Speckle noise, radiometric variations, and terrain heterogeneity significantly challenge accurate change detection in SAR imagery.
• Existing methods struggle with subtle variations and false alarms, necessitating improved detection frameworks.

Purpose of the Study:

• To develop a robust hybrid change detection framework for multi-temporal SAR data.
• To enhance the sensitivity and discriminability of changed regions despite imaging artifacts.
• To reduce false alarms and improve the overall accuracy and reliability of SAR change detection.

Main Methods:

• Integration of Fusion-based Difference Imaging (FDI) to enhance subtle backscatter variations.
• Combination of complementary difference representations for improved region discriminability.
• Application of a deep feature-guided clustering method for enhanced separability and noise reduction.

Main Results:

• The proposed framework achieved high detection accuracy, with Overall Accuracy (PCC) up to 99.97%, Kappa coefficient up to 98.30%, and F1-score up to 99.59%.
• The approach demonstrated consistent superior performance compared to existing methods across diverse datasets and speckle conditions.
• Fusion-based representation learning with uncertainty-aware clustering proved effective in challenging SAR imaging scenarios.

Conclusions:

• The hybrid FDI and deep clustering framework offers a scalable, efficient, and flexible solution for robust SAR change detection.
• The method effectively addresses limitations posed by speckle noise and complex terrain features.
• This approach significantly advances the reliability of environmental monitoring and disaster assessment using multi-temporal SAR data.