Performance Assessment of Change Detection Based on Robust PCA for Wavelength Resolution SAR Images Using Nonidentical Flight Passes

TL;DR

This paper shows that robust PCA can detect changes in SAR images even when the radar images come from very different flight paths.

Contribution

The paper demonstrates that robust PCA is effective for SAR change detection using nonidentical flight passes, relaxing prior constraints.

Findings

RPCA achieves a false alarm rate of one per square kilometer with a detection probability above 90% for 5° flight pass differences.

A detection probability of 97.5% is achieved with a false alarm rate of 0.917 per square kilometer for nonidentical flight passes.

Wavelength-resolution SAR images allow change detection with flight pass differences up to 95°.

Abstract

One of the main challenges in Synthetic Aperture Radar (SAR) change detection involves using SAR images from different flight passes. Depending on the flight pass, objects have different specular reflections since the radar cross-sections of these objects can be totally different between passes. Then, it is common knowledge that the flight passes must be close to identical for conventional SAR change detection. Wavelength-resolution SAR refers to a SAR system with a spatial resolution approximately equal to the wavelength. This high relative resolution helps to stabilize the ground clutter in the SAR images. Consequently, the restricted requirement about identical flight passes for SAR change detection can be relaxed, and SAR change detection becomes possible with nonidentical passes. This paper shows that robust principal component analysis (RPCA) is efficient for change detection even using wavelength-resolution SAR images acquired with very different flight passes. It presents several SAR change detection experimental results using flight pass differences up to 95°. For slightly different passes, e.g., 5°, our method reached a false alarm rate (FAR) of approximately one false alarm per square kilometer for a probability of detection (PD) above 90%. In a particular setting, it achieves a PD of 97.5% for a FAR of 0.917 false alarms per square kilometer, even using SAR images acquired with nonidentical passes.

Figures