Non-Gaussian Scattering Models for Persistent Scatterer Detection

Abstract:

Rapidly decorrelating surfaces, which include forested regions, vegetated hillsides, and much natural terrain, present a fundamental challenge to many promising potential applications of InSAR imagery. Persistent scatterer (PS) techniques offer one solution for analysis of these decorrelated datasets by identifying and utilizing only the most phase-stable points in the images. Because the application of PS techniques depends on the quality of the selected points, any method of identifying PS must therefore be reliable for any ensuing analysis of the surface to be meaningful.

To date, PS detection theory has largely been based on Gaussian-derived models, where the returns from either one or both of the dominant and distributed scatters in a single resolution element (resel) are modeled as complex circular Gaussian random variables. While such models are convenient in their elegance and simplicity, much previous work in radar backscattering has shown they are unable to accurately describe returns from high-resolution SAR imagery, particularly over inhomogeneous surfaces. This disconnect could significantly inhibit the effectiveness of PS techniques in high-resolution datasets over complex terrain, as the use of Gaussian detectors to estimate quantities that are inherently non-Gaussian results in sub-optimal performance.

We describe extensions to the scattering model for PS detection reflecting the non-Gaussian scattering behavior observed in high-resolution imagery. This behavior results from a breakdown in the assumptions required for the Central Limit Theorem, primarily those of scatterer inhomogeneity, and incorporate realistic speckle distributions. We present simulation results that incorporate fluctuations observed in actual high-resolution radar backscatter in PS detection algorithms. Then, we compare our new distribution models with returns from both PS and non-PS pixels in Sentinel-1A data over several types of natural terrain, and finally discuss the implications for expected performance of PS detection.