Multiscale segmentation and anomaly enhancement of SAR imagery

Area of Science:

• Remote Sensing
• Image Processing
• Statistical Modeling

Background:

• Synthetic Aperture Radar (SAR) imagery presents challenges in segmenting natural clutter and detecting anomalies due to radar speckle.
• Terrain types like grass and forest exhibit distinct statistical properties that vary with scale, influenced by SAR sensor characteristics.

Purpose of the Study:

• To develop efficient multiscale approaches for segmenting natural clutter (grass, forest) in SAR imagery.
• To enhance the detection and pinpointing of anomalies within SAR data.
• To exploit the scale-dependent statistical differences caused by radar speckle for improved analysis.

Main Methods:

• Utilized multiscale stochastic processes, specifically scale-autoregressive models, for analyzing SAR imagery.
• Developed terrain-specific models (grass, forest) to guide pixel classification and segmentation.
• Employed model residuals to identify and enhance anomalous regions by analyzing their correlation across scales.

Main Results:

• Achieved efficient pixel classification and grass-forest boundary estimation using scale-autoregressive models.
• Demonstrated effective anomaly enhancement with minimal computational overhead.
• Validated the techniques on high-resolution (0.3-m) SAR data, showing robust performance.

Conclusions:

• The proposed multiscale stochastic processes offer an efficient framework for SAR image analysis, including segmentation and anomaly detection.
• The methods effectively leverage the unique statistical properties of SAR data across different scales.
• The approach provides a computationally efficient means to enhance and pinpoint anomalies in SAR imagery.