Algorithms improve synthetic aperture radar coherent change detection performance

Satellites, unmanned air vehicles, and other devices can gather far more imagery on a daily basis than there are available human eyes and minds to evaluate. Algorithms are needed that are sensitive enough to detect significant changes to bring to human attention. Several algorithms exist that can subtract an older image from a more recent one and highlight the changes. Such change detection techniques estimate the similarity of pixel pairs.

We evaluated and compared the detection performance of different change detection algorithms using Gotcha public release synthetic aperture radar (SAR) data.¹ Figure 1 shows non-coherent change estimate (NCCD),^2,3 complex correlation coefficient estimate (CCD),⁴ and maximum likelihood coherence estimate (MLE)⁴ images of the area used for the SAR change detection studies. The area of interest is comprised of several parking lots occupied by numerous parked vehicles. Analysis of the SAR reference and test images of this area were found to contain a total of 33 vehicles that changed during the time interval between the gathering of the reference and test images.

Figure 1. Images developed from Gotcha public release synthetic aperture radar data, using (a) the non-coherent change estimate (NCCD) algorithm, (b) the complex correlation coefficient estimate (CCD) algorithm, and (c) the maximum likelihood coherence estimate (MLE) algorithm. Blue circles show vehicle changes detected.

This set of 33 vehicles consists of either arrivals or departures that occurred during this time interval. The vehicle changes are marked by circles in the NCCD image. The race-track feature emphasized in the NCCD image is also of interest to an image analyst. This feature is visually impossible to detect from the original reference and test SAR images, and it is hardly discernible in the NCCD image. However, the race-track feature is clearly detectible in both coherent change detection images.

The MLE coherent change detection algorithm is presented in Table 1. Let x₁, … , x_N denote complex reflectivities of reference image pixels in some local neighborhood, and x_N+1, … , x_2N denote complex reflectivities of the corresponding pixels in the co-registered test image.^{4, 5} The NCCD non-coherent change detection algorithm presented in the table is derived in a previous paper.²

Figure 2. (a) Target detection receiver operating characteristic (ROC) curves. (b) Per-pixel ROC curves. P_D: Probability of detection. P_FA: Probability of false alarm.

Figure 2(a) shows the probability of detection versus the probability of false alarm (P_D/P_FA) receiver operating characteristic (ROC) curves obtained using a 3×3 box size, when the race-track feature is excluded from the target list. We evaluated detection performance ROCs versus the percent (%) of detected pixels in a target-sized box (25 pixels by 25 pixels). We found that detection performance was not very sensitive to this parameter: the ROC curves obtained for 5–30% are tightly clustered. The curves show that the NCCD provides better performance than either of the coherent change detection algorithms. Furthermore, coherent change detection using the MLE algorithm achieves significantly better performance than the classical CCD algorithm. For example, at 70% P_D, CCD gave ≈700 target-sized FAs, whereas the MLE gave ≈0 FAs.

Most of the false detections with either the CCD or MLE algorithms occur in low-coherence building (and tree) shadows. These low-radar cross section areas, as well as flat asphalt roads, have low coherence due to random phase returns. Since an X-band SAR cannot detect target returns from targets located and masked in shadow areas, coherent change detection performance ROCs could be significantly improved if such areas were masked as ‘don't care’ areas before performing the detection function.

One simple approach for improving the ROC curves is to post-process the standard CCD and MLE images by setting the coherence of the areas that correspond to low-radar cross section areas in both the test and reference images to unity. Change detection is then performed on the change images with low-radar cross section areas masked. These areas can easily be detected as follows. The n-th pixel in the CCD image is declared as belonging to the non-interesting, low-radar cross section area if the average power of the coherent sum and the coherent difference between the corresponding neighborhoods in the test and reference images is below a selected threshold, T:

where N is the number of pixels in a local neighborhood of the n-th pixel, while, as before, x₁, … , x_N are pixels of the reference image and x_N+1, … , x_2N are the corresponding test pixels in the n-th pixel neighborhood.

The per-pixel ROC curves when the race-track is included as a target in Figure 2(b) show that a large gain in coherent change detection performance is achieved after the shadow areas in the SAR image are detected and denoted as ‘don't care’ areas of the scene. The curves show that after shadow removal, CCD performance is significantly improved relative to the original CCD performance. However, the CCD algorithm performance after shadow removal is not as good as the original MLE performance. And MLE performance after shadow removal is the best overall coherent change detection performance result achieved.

In summary, the NCCD algorithm achieved the best overall vehicle detection performance. The MLE version of the coherent change detector performed better than the CCD version of the algorithm. Finally, we presented a simple, robust algorithm for mitigation of false changes in shadow areas (building and tree shadows), thereby improving the performance of both coherent change detection algorithms (MLE and CCD). Our next step will be to transition the algorithms to operational SAR imaging systems such as Global Hawk, Predator, Lynx, and so forth to improve image exploitation and change detection performance.