Proceedings of the Institute of Acoustics

 

Laboratory multistatic and multipolar SAR CCD investigation

 

A. Hagelberg, Centre for Electronic Warfare, Information and Cyber, Cranfield University, Defence Academy of the United Kingdom, SN6 8LA, UK
D. Andre, Centre for Electronic Warfare, Information and Cyber, Cranfield University, Defence Academy of the United Kingdom, SN6 8LA, UK
M. Finnis, Centre for Defence Engineering, Cranfield University, Defence Academy of the United Kingdom, SN6 8LA, UK

 

1 INTRODUCTION

 

Synthetic Aperture Radar (SAR) provides reconnaissance and surveillance of the earth’s surface for all weather, during the day and night. SAR Coherent Change Detection (CCD) allows for the detection of very small scene changes between two SAR images, such as vehicle tracks [1,2]. However, good quality high contrast CCD results are reliant on high coherence between the SAR images [3].

 

To reduce the effect of decorrelation from natural processes such as wind and rain, and to monitor developing situations in a timely manner, it is desirable to rapidly collect the required repeated pass SAR collections. Bistatic radar geometries involve having the transmitter and receiver substantially separated, and multistatic geometries are comprised of multiple simultaneous bistatic collections. Multistatic SAR satellite constellations may allow collection of more timely change detection images to be formed using SAR image collections from differing trajectories, thus reducing the repeat pass time [4,5], however the CCD’s may be subject to varying degrees of coherence due to differing baselines. The multistatic SAR data is multidimensional and can be combined in different ways. For high coherence and an improved change detection capability, the repeat pass geometries should be chosen to provide a high degree of spatial frequency (K-space) support overlap [3,6,7], yet it has been found that this alone is insufficient to guaranty high coherence and good CCD contrast [8].

 

The research is aligned to derisking the multistatic SAR satellite constellation being developed by the UK MOD under the name Project Oberon [9], however it is also relevant to multistatic SAR drone swarms, which operate in SAR near field configurations [10,11]. The research is supported by measurements at the Cranfield University Ground Based Synthetic Aperture Radar (GBSAR) laboratory, where antennas are scanned within two vertical rectangular apertures, which allow a great deal of flexibility and fine control of SAR geometries for the scene.

 

This work investigates the levels of coherence between bistatic and multistatic SAR CCD imagery and approaches to improve change detection performance. The multistatic images are formed in a variety of ways including multipass and multi-polarimetric.

 

2 BACKGROUND

 

Multistatic SAR can provide several advantages, such as providing improved cross range resolution [12,13], and Signal to Noise Ratio (SNR) [14], although the specific geometries have a strong influence on these [15]. Polarimetric SAR CCD has also been shown to improve CCD images [2,16]. The coherence value of a CCD image is referred to as the estimated coherence γ ̂, and differs slightly from the true value γ_𝑡𝑟𝑢. The deviation is biased and dependent on calculation features such as the CCD estimation window size [17]. The true coherence can be decomposed into factors which are also in the range 0-1 [3,8,17–21],

 

 

The factors correspond to temporal changes between the two collections γ_{𝑡𝑒𝑚𝑝}; signal to noise ratio (SNR) γ_𝑆𝑁𝑅; signal processing effects γ_{𝑝𝑟𝑜𝑐}; and SAR geometry effects γ_{𝑔𝑒𝑜𝑚}. The signal processing term γ_{𝑝𝑟𝑜𝑐} accounts for decorrelations due to the image formation algorithm accuracy γ_𝑎𝑙𝑔 and image registration accuracy γ_𝑟𝑒𝑔. The geometry term is decomposed into a calculated baseline term γ_{𝑏𝑎𝑠𝑒𝑙𝑖𝑛𝑒} which accounts for K-space overlap; and a Radar Cross Section (RCS) term γ_𝑅𝐶𝑆, which is a function of the polarization [8,21].

 

3 METHODOLOGY

 

3.1 GBSAR laboratory and experimental setup

 

The GBSAR laboratory performs microwave measurements with a Vector Network Analyser (VNA), which generates a stepped frequency waveform. The system was set up for indoor use, with Ultra Wideband horn antennas used here in the range 6.6-10 GHz. The two antenna horns are mounted on rails allowing motion in separate two-dimensional apertures, shown in Figure 1. Due to the high positional accuracy of the rail system, the transmitter trajectory could be repeated precisely, which combined with use of different receiver positions, allowed multistatic data to be collected. The geometry is shown in Figure 1b), with transmit aperture in red, and the six labelled receiver positions in blue, and referred to as R1 through to R6.

 

 

Figure 1: Image (a) shows the GBSAR system in a bistatic configuration, with a gravel rectangle as the scene. Image (b) shows the radar trajectories used in the data collection. The transmitter trajectory is shown in red, the multiple fixed receiver positions in blue and the gravel region is shown in orange. The receivers are numbered 1 to 6 and are referred to as R1 through to R6.

 

A gravel rectangle, approximately 3.5 x 4 m in size, was created at the centre of the scene, allowing speckle SAR image collection suitable for coherence investigation with gravel disturbance tracks drawn in. The letters ‘GBSAR’ were traced onto the scene by gently disturbing the gravel for use in change detection performance evaluation.

 

Three multistatic collections were measured: the first, the Reference collection, comprised the three bistatic geometries with receivers R1, R2, R3; the second and third were the Mission collections, comprising the three bistatic geometries R4, R5, R6 (see Figure 1b). The Mission multistatic geometries were collected at a later time to the Reference, one of them having had a disturbance track applied to the gravel terrain – Mission 2.

 

From the Reference and Mission multistatic collections, three pairs of bistatic geometries were used for CCD creation as follows: R1 and R4, R2 and R5, and R3 and R6 (effectively receiver columns). These allowed the formation of three bistatic CCD images for each of the two missions. The three CCDs were combined to form multistatic CCD images in different ways.

 

3.2 Data processing

 

Both bistatic and multistatic (coherent summation) SAR images were formed using the back projection image formation algorithm (BPA). The SAR images, reference and mission, were then used to form CCD images. Two sets of CCD images were formed, one for each of the two Missions: one where the gravel was undisturbed between SAR images (Mission 1), and another where the gravel had a track traced (Mission 2).

 

Performance metrics such as the Probability of Detection (POD), Critical Success Index (CSI) and the False Alarm Rate (FAR) were used in the evaluation of the CCD images. These were calculated using a confusion matrix. Here T and F mean true and false respectively, and P and N mean positive (change) and negative (no-change) in the CCD image [22],

 

 

Five master images were used to calculate the performance metrics. Four were formed from single polarisation multistatic datasets, and the remaining one was formed using all four polarisations. The fully polarimetric master CCD image is shown in Figure 2 (a) with values ranging from 0 to 1. Figure 2 (b) shows a binary image where pixels with a coherence value below the threshold of 0.87 were detected as a change and set to 0, and those with a coherence above the threshold were set to 1.

 

This shows where detections of a change or no change would be made for the master image. The threshold for the master image was chosen to most accurately replicate the ground truth of the disturbance. The threshold was kept constant for all master images, however it was varied for the comparison images, and this produced a series of curves presented in section 4. Performance metrics (equation 2) can be calculated by comparing the detections in the master and comparison images.

 

Figure 2: Fully polarimetric master images used for the performance metrics. Image (a) is a CCD image formed from the mean of fully polarized multistatic data collections. Image (b) is a binary image formed from applying a threshold to (a).

 

4 RESULTS

 

4.1 Polarimetric variation

 

Different polarisations give rise to different CCD images and coherence. This has been previously demonstrated in [21]. This is shown in Figure 5 and Figure 7, where coherence shows some variance with polarisation.

 

The variation with coherence can also be seen in the performance indices. These are presented in Figure 3 and Figure 4. In these figures individual datapoints are represented by dots. The dotted lines are quadratic curves fitted to the data.

 

In Figure 3 (a) & (b) the FAR is presented, for the bistatic and multistatic case respectively. This graph shows that the HH polarisation had a lower FAR. This is despite the lower coherence. This low value is likely due to a higher number of true positive detections, rather than a reduction in false positives. Figure 4 shows the CSI and POD. In Figure 4 (a) the HH polarisation outperformed the other polarisations with a higher CSI. In Figure 4 (b) the POD is shown, where all polarisations display a similar performance.

 

Analysis of these performance metrics demonstrate that coherence is not the only factor that should be considered when assessing the quality of CCD images. Such analysis is however difficult in non laboratory environments, where environmental decorrelations effect the scene and there are trajectory errors.

 

 

Figure 3: Images showing the FAR for (a) bistatic and (b) multistatic for four polarizations. Datapoints are plotted with dots, and polynomial fitted curves are added to show the data trends.

 

 

Figure 4: Images showing the CSI (a) and POD (b) for the bistatic images across four polarizations. Datapoints are plotted with dots, and polynomial fitted curves are added to show the data trends.

 

4.2 GBSAR track - Qualitative analysis

 

In the tessellated image of bistatic CCD images in Figure 5, parts of the ‘GBSAR’ disturbance can be made out. Different polarisations and position pairs detect different parts of the disturbances. Compared to the master image (Figure 2) the CCD products are of a poorer quality. This is partly due to false detections caused by a low RCS coherence factor as the two SAR images are formed using different radar trajectories. Additionally, there will be a decorrelation across the CCD images due to the k-space non-overlap.

 

Figure 5: Tessellation of bistatic CCD images. Each row contains the CCD images for a particular trajectory pair. These correspond to receiver positions shown in Figure 1(b). Each column corresponds to a polarization.

 

Multistatic CCD images were formed in two manners. The first was to take the maximum, mean and minimum for each pixel across the individual bistatic CCD images formed across the two multistatic collections. Alternatively, coherent summation was applied to the SAR images in each multistatic collection, and then a CCD algorithm was applied across the two multistatic SAR images. The results are shown in Figure 6 for the VV polarisation.

 

 

Figure 6: Tessellation image of multistatic CCD images formed for the VV polarization.

 

In the multistatic images the ‘GBSAR’ track is visually clearer than in the individual bistatic images. Visually, the mean and coherent summation images show some similar performance. For these images the multistatic CCD “minimum” appears to offer the clearest ‘GBSAR’ track representation. It is possible however, that altering the contrast and colour scale of the images could alter the qualitative performance perception, thus a variety of thresholds are used for the performance metrics in the quantitative analysis results shown in section 4.3.

 

4.3 GBSAR track - Quantitative analysis

 

The CCD coherence estimated over the image pixels is presented in Figure 7. The individual bistatic CCD coherence estimates are shown, together with their mean, min and max. Additionally, SAR coherent summation was performed for each of the two multistatic collections, providing two SAR images with finer resolution, from which a CCD was formed and its coherence estimated, designated “coherent sum CCD” in Figure 7.

 

It can be seen that the coherence of some individual bistatic geometries and the mean and coherent sum multistatic CCD images are quite similar. With the maximum and minimum CCD images giving the extreme values. Also, the HH images have a lower spread in coherence values compared with VV (excluding min and max).

 

Figure 7: Estimated coherence of individual bistatic (green) and corresponding multistatic CCD images across four polarizations.

 

The detection performance metrics are compared in Figures 8-10. The green line is the minimum multistatic; the blue the maximum; the red is mean; the black is the coherent sum SAR image multistatic estimated coherence. These metrics are plotted for all four polarizations. These show variation in performance with polarization.

 

The metrics show that the trend lines for the individual bistatic, the mean and coherent sum multistatic are quite close. The minimum multistatic images perform significantly better, and the maximum images perform significantly worse. These behaviors are consistent for the POD, FAR and CSI.

 

 

Figure 8: Graphs showing the POD for different multistatic images compared against bistatic images for four polarizations. Datapoints are plotted with dots, and polynomial fitted curves are added to show the data trends.

 

 

Figure 9: Graphs showing the CSI for different multistatic images compared against bistatic images for four polarizations. Datapoints are plotted with dots, and polynomial fitted curves are added to show the data trends.

 

Figure 10: Graphs showing the FAR for different multistatic images compared against bistatic images for four polarizations. Datapoints are plotted with dots, and polynomial fitted curves are added to show the data trends.

 

4.4 Polarimetric

 

Polarimetric images can be presented as a colour composite of several CCD images represented by RGB values or as a single greyscale CCD image. Two examples are shown in Figure 11. Polarimetric multistatic CCD images can be formed in a similar manner to the approaches taken previously. Maximum, mean and minimum images were compared: each pixel is given the value of the maximum, mean or minimum of 12 images (4 polarisations and 3 position pairs). Compared to the individual polarisation (black), the minimum and mean images resulted in an improved POD but also in increased FAR. The maximum image showed a very low FAR, however this was acompanied by very poor POD and CSI results.
 

 

Figure 11: Fully polarimetric multistatic CCD colour composite image (a) and greyscale multistatic CCD (b). Greyscale CCD was formed as the average of the four polarized multistatic CCD images.

 

Figure 12: Graphs showing the POD, CSI and FAR for different multistatic images. Blue, red and green are formed from all four polarizations and 3 bistatic position pairs. These produced the max, mean or minimum value for each pixel. These are compared with the single polarization multistatic CCD images formed from coherent summation multistatic SAR images. Datapoints are plotted with dots, and polynomial fitted curves are added to show the data trends.

 

5 CONCLUSIONS

 

Qualitative and quantitative methods have been used to assess the performance of bistatic and multistatic CCD images against master images.

 

It has been shown that multidimensional (multistatic and multipolar) CCD can offer improved performance over single bistatic (figures 8-10 & 12), using qualitative and quantitative analysis. It has also been shown that the manner in which the multidimensional CCD image is formed has a significant impact on performance.

 

Additionally, different polarizations result in different levels of coherence and performance at detecting tracks. This variation was however quite minor, but was shown in figures 3, 4(a) and 7-10. Despite the minor variation in performance between polarisations, fully polarimetric CCD images showed a significantly improved POD compared with single polarization bistatic images, albeit with a higher FAR.

 

The value of the performance metrics is dependant on the threshold set for the comparison image, as such multiple thresholds and performance metrics should be used to characterise performance.

 

Future work will explore the impact of polarimetric decompositions on coherence and change detection performance. Additional work will investigate the impact of greater baselines and trajectory differences.

 

6 REFERENCES

 

1. Preiss M, Gray D. A, Stacy N. J. S. Detecting scene changes using synthetic aperture radar interferometry. IEEE Transactions on Geoscience and Remote Sensing. 2006;44(8):2041–2054. doi: 10.1109/TGRS.2006.872910.

2. Preiss M, Stacy N. J. S. Polarimetric SAR coherent change detection. 7th European Conference on Synthetic Aperture Radar. 2008.

3. Jakowatz C. V., Wahl D. E., Eichel P. H., Ghiglia D. C., Thompson P. A. Spotlight-mode synthetic aperture radar: A signal processing approach. Kluwer Academic Publishers; 1996.

4. Keydel W. Present and future airborne and space-borne systems. NATO STO. 2007.

5. Krieger G., Moreira A. Multistatic SAR satellite formations: Potentials and challenges. In: International Geoscience and Remote Sensing Symposium (IGARSS). 2005. p. 2680–2684.

6. Andre D., Blacknell D., Morrison K. Spatially variant incoherence trimming for improved SAR CCD. In: SPIE Defense, Security and Sensing. 2013. p. 153–164.

7. Doerry A. W. SAR data collection and processing requirements for high quality coherent change detection. In: Radar Sensor Technology XII. SPIE; 2008. p. 694706.

8. Hagelberg A., Andre D., Finnis M. Laboratory bistatic SAR CCD investigation. IET Electronics Letters. Submitted and under review.

9. Airbus website. https://www.airbus.com/en/newsroom/press-releases/2019-09-airbus-to-develop-technology-for-ultra-high-resolution-satellites. 2019.

10. Hawkins B., Anderson M., Prager S., Chung S. J., Lavalle M. Experiments with small UAS to support SRA tomographic mission formulation. In: International Geoscience and Remote Sensing Symposium (IGARSS). IEEE; 2021. p. 643–646.

11. Harmer S. W., De Novi G. Distributed antenna in drone swarms: A feasibility study. MDPI Drones. 2023;7(2). doi: 10.3390/drones7020126.

12. Chung C., Kim S., Kim J., Lee U., Lee B., Ka M. H. Analysis on the impact of constellation configuration over multistatic SAR images. In: Proceedings – 2021 7th Asia-Pacific Conference on Synthetic Aperture Radar (APSAR 2021). IEEE; 2021.

13. Renga A., Graziano M. D., Moccia A. Formation flying SAR: Analysis of imaging performance by array theory. IEEE Transactions on Aerospace and Electronic Systems. 2021;57(3):1480–1497. doi: 10.1109/TAES.2020.3043526.

14. Krieger G., Fiedler H., Rodriguez-Cassola M., Hounam D., Moreira A. System concepts for bi- and multi-static SAR missions. In: International Radar Symposium (IRS). 2003.

15. Hagelberg A., Andre D., Finnis M. Non-overlapping trajectory multistatic SAR coherent change detection. In: IET International Conference on Radar Systems. Edinburgh; 2022.

16. Cloude S. R. Polarimetric SAR interferometry. IEEE Transactions on Geoscience and Remote Sensing. 1998;36(5 PART 1):1551–1565. doi: 10.1109/36.718859.

17. Preiss M., Stacy J. Coherent change detection: Theoretical description and experimental results. 2006.

18. Zebker H. A., Villasenor J. Decorrelation in interferometric radar echoes. 1992.

19. Wang Y., Huang Y., Zha Y., Pu W., Yang J. Data collection strategies for high quality multi-pass SAR CCD images. In: International Geoscience and Remote Sensing Symposium (IGARSS). IEEE; 2014. p. 1843–1846.

20. Bickel D. L. SAR image effects on coherence and coherence estimation. 2014.

21. Hagelberg A., Andre D., Finnis M. Polarimetric bistatic SAR image coherence. In: 14th European Conference on Synthetic Aperture Radar. Leipzig, Germany: VDE; 2022. p. 80–85.

22. Wilks D. S. Statistical methods in the atmospheric sciences. 2nd ed. Elsevier; 2006.