Quantitative comparison of the performance of SAR segmentation algorithms

Area of Science:

• Remote Sensing
• Image Processing
• Geospatial Analysis

Background:

• Synthetic Aperture Radar (SAR) imagery presents unique challenges for image segmentation due to coherent speckle noise.
• Accurate segmentation is crucial for extracting meaningful information from SAR data for various applications.
• Existing segmentation algorithms for SAR data require robust performance evaluation methods.

Purpose of the Study:

• To develop and apply quantitative, statistically based methods for evaluating SAR image segmentation algorithm performance.
• To compare the performance of two distinct SAR segmentation algorithms: iterative edge detection/segment growing and Maximum A Posteriori (MAP) estimation.
• To assess algorithm performance on both simulated and real SAR data.

Main Methods:

• Development of local and global homogeneity measures based on SAR speckle properties and a scene model with abrupt edges.
• Application of these measures to segmentations generated by iterative and MAP-based algorithms.
• Quantitative comparison of algorithm outputs using the derived performance metrics and visual assessment.

Main Results:

• Quantitative measures aligned with visual assessments of segmentation quality.
• On simulated data, both algorithms performed well, with MAP estimation showing superior visual and measurable results.
• On real SAR data, MAP estimation maintained performance, while the iterative algorithm's performance significantly degraded.

Conclusions:

• MAP estimation using simulated annealing is a more robust and effective method for SAR image segmentation compared to iterative approaches.
• Current performance measures require refinement with more realistic scene models to address oversegmentation.
• Further development is needed for improved SAR image segmentation evaluation techniques.