InSAR Data Coherence Estimation Using 2D Fast Fourier Transform Andrey V. Sosnovsky1 , Viktor G. Kobernichenko1 , Nina S. Vinogradova1 , Odhuu Tsogtbaatar1,2 1 Ural Federal University, Yekaterinburg, Russia and 2 Ulan-Baator Technical Unisersity, Ulan-Baator, Mongolia sav83@e1.ru Abstract. Interferometric coherence is an important indicator of relia- bility for interferograms obtained by interferometric synthetic aperture radar (Interferometric SAR, InSAR). Areas with low coherence values are unsuitable for interferometric data processing. Also, the coherence may be used as a classification indicator for various coverage types. Coher- ence magnitude can be calculated as an absolute value of the correlation coefficient between two complex SAR images with averaging in a local window. The problem of coherence estimation is in its dependence on the phase slope caused by relief topography (topographic phase). A co- herence estimation algorithm is proposed that is based on the 2-FFT peak height assessment without calculation of the correlation coefficient. It is shown that such estimate has significantly less dependence on the topographic phase slope and provides satisfactory results in InSAR data quality assessment. Keywords: Synthetic aperture radar images, InSAR systems, coherence estimation 1 Introduction Interferometric data processing (InSAR) for extraction of information about the Earth terrain and its changes becomes one of the general guidelines in develop- ment of contemporary space-based radar systems together with the implemen- tation modes of ultra-high spatial resolution (1–3 meters) and full-polarimetric processing [1–3]. The InSAR processing for building the digital elevation models (DEM) includes the following steps: synthesis of the pair of complex synthetic aperture radar (SAR) images and their coregistration; forming the interferogram by the element-wise complex multiplication of these SAR images; compensation of the phase slope caused by side-looking imaging geometry; multilooking (non- coherent summation); phase noise supression; phase unwrapping (the elimina- tion of phase ambiguities); and conversion of the absolute phase interferogram in elevation grid data and its projection. Interferometric coherence is an important indicator of suitability of the data scene obtained by a radar remote sensing system for the further processing and 99 solving the final problem, i.e. the DEM generation or terrain changes mapping. The coherence factor is calculated as an absolute value γ0 of the correlation co- efficient between samples of two complex SAR images (single-look data complex, SLC) taken in the local windows P | ż1 (m, n) · z̄2 (m, n)| γˆ0 = pP P , (1) |ż1 (m, n)|2 · |z̄2 (m, n)|2 where z1(2) (m, n) are the SLC samples (z̄1(2) (m, n) are the complex-conjugate samples) [1–4, 6], γˆ0 takes values in interval [0,1], and near-zero values corre- spond to areas of high or full decorrelation, which are not suitable for interfero- metric data processing. An intrinsic radar signal decorrelation is caused by the radar looking angle difference (geometric decorrelation) and by the Earth sur- face variability (temporal decorrelation). Strong decorrelation occurs due to loss of echo-signal (typical for water surfaces), volume scattering (forest vegetation), signal layover, and shadowing, etc. Commonly, the areas with coherence lower than 0.2–0.3 are unsuitable for conversion into DEM. So, the coherence value may be used as an interferometric phase quality indi- cator, which gives an opportunity to reject the areas, where the phase is unstable and is not related to the Earth topography. Also, rejection of the decorrelated areas before phase unwrapping simplifies this processing step because some un- wrapping algorithms become slower and unstable while processing such areas. The other way of the interferometric coherence utilization implies its usage as a parameter in adaptive phase noise filters. A commonly used Goldstein-Baran adaptive frequency filter adapts a frequency response in dependence on local coherence [5] F (k, l) = |S(k, l)|1−γ̂ · S(k, l), (2) where S(k, l) and F (k, l) are the spectra of raw and filtered interferograms rel- atively, taken in a local window; γ̂ is the local coherence estimate. So, the lower coherence leads to extension of filter’s bandwidth and vice versa. For this reason, the coherence map is usually calculated before the phase noise supression stage in the InSAR processing chain. However, this approach entails some problems because, in fact, a random variable is estimated here, but not a random process. So, any phase slopes caused by both natural topography variability and by point-of-view geometry (remote sensing radar systems have a side-looking configuration) lead to the degradation of estimate (1). Its value depends on the slope and tends towards the value γ0 ≈ 0.2...0.3, i.e. a bias of the estimate for independent Gaussian values of the correlation coefficient [3], which in practice takes the value about 0.1–0.3 (Fig 1). Thus, coherence loses its properties as measure of the interferogram quality because its value becomes independent on the relation between topographic and fluctuating components of the phase. The problem of coherence estimate degradation was described by [3, 4, 7], and some approaches for its correction were proposed. The basic approach involves a 100 Fig. 1. Dependencies of γ̂0 coherence estimate on the phase slope angle for different simulated correlation coefficients (a) and local window (sample) sizes (b). reference digital elevation map (with lower spatial sampling frequency, as a rule) for phase compensation prior to the coherence estimation. But such reference DEM with a sufficient sampling factor is often unavailable for a specific terri- tory (especially for the Northern ones). Another aprroach implies elimination of the phase components from estimate [3] andcalculation of an estimate based on images magnitude. But as it is shown [8], such estimates have a large bias and they can not be used for detection of low-coherent areas. So, a reasonable approach may lie in an adaptive phase slope estimation and its compensation, as it offered in [3, 9], but some additional measures should be considered to improve the efficiency of such approach. 2 A 2-FFT interferometric coherence estimate To eliminate the effect of coherence estimates degradation, we construct the estimate in a way, which makes it immunable to phase slope value. Since a constant phase slope can be considered as the interferogram spatial frequency modulation, it is reasonable to use a 2D fast Fourier transform (2-FFT) to determination of modulation frequencies, which can be found as [ωx , ωy ] = argmax{F2 [ż1 · z̄2 ]}, (3) where ωx and ωy are the spatial frequencies in both spatial dimensions, which are estimated in an interferogram sample of M × N size. A common idea for 2-FFT coherence estimation is that further spatial frequency demodulation is 101 performed according to the rule: ωx ωy z2c = z2 · exp[−j( + )], (4) M N that eliminates influence of the phase slope on the estimate value. After substi- tution of (4) into (1), instead of z2 the correlation coefficient would give a correct estimate for coherence magnitude [3, 9]. Here, the main problem is that such way has a low computational efficiency: the advantages of the FFT application are negated by the need of a correlation coefficient computation, and, so, a computational time for this estimate exceeds the same one for the usual estimate. From the other side, the images’ magnitude information gives minor contribution to the estimate value, as it is shown in [8]. Moreover, radar brilliant points significantly degrade the estimate in their neibourhood. So, it is reasonable to normalize images magnitudes taking |z1 | = 1, |z2 | = 1 (5) and calculate a peak value of the discrete spectrum for a normalized interfero- gram (in the local window) as follows M = max{F2 [z1 · z2 ]}. (6) In this case, taking into account the Parseval’s identity, for a scene sample M × N with a constant phase slope, one can get that the following: 1) for fully coherent scenes a 2-FFT spectrum has a single peak P of an M × N height in a position matching to a spatial frequency value; 2) fully incoherent scenes may be considered as a discrete white noise with the spectrum fluctuating near 1. Dependence of the spectrum peak value P and scene decorrelation γ may be retrieved by simulation of Gaussian homogenous scenes with different correla- tions (Fig. 2). As can be seen from Fig. 2, the dependence between a normalized spectrum peak value P/(M × N ) and a simulated correlation is slightly biased. So, a possible way for coherence estimation is to recalculate a peak value (or peak to mean value ratio) into a coherence magnitude in the following way: P γ̂2 = . (7) M ×N Estimate γ̂2 is less dependent of the phase slope value than a standard one γ̂0 , which affects only peak location within a spectrum. Figure 3 demonstrates the estimate behaviour by the simulation of homogenous scenes with different simulated correlation coefficients. As can be seen, that for high correlations, the estimate tends to fall down near the slopes of π radians to the level about 0.71 of γ̂2 , which is related to sampling effects. 102 Fig. 2. Depedencies of a normalized spectrum peak value of the 2-FFT spectrum on the simulated correlation coefficients of different local window (sample) sizes. 3 Experimental results An experimental assesment of the proposed method of interferometric coherence estimation was carried out within application of a reference DEM. Since the coherence value indicates, first of all, the fluctuations level of the absolute phase around the topographic phase, and the latter is closely related with the reference DEM heights [2], it is obviously to use a statistics of phase deviations to assess the quality of the coherence estimates. Such assesment was performed in the following way: 1) the reference DEM was reprojected into a flight vehicle coordinate system (range-azimuth), and then the heights were recalculated to reference absolute phase values using an InSAR height ambiguity factor; the reminder shift was fixed through the cross-correlation between the reprojected DEM and the scene absolute phase obtained after phase unwrapping stage; 2) coherence estimates γ̂0 and γ̂2 were calculated with the window size 19×19; 3) coherence values interval [0,1] were splitted into subintervals with a fixed step 0.05; 4) for each subinterval i the standard deviation of the absolute phases and reference absolute phases σψi were counted in corresponding image points. The obtained dependence of σψ on γ̂ acts as a quality indicator for the inter- ferometric coherence estimate. The better estimates should give a less-fluctuating 103 Fig. 3. Dependencies of γ̂2 coherence estimate on the phase slope angle for different simulated correlation coefficients. decreasing dependence at least for low- and medium-valued coherences (except extreme-low values) with a possible wider range for both coherence and standard deviation values. This means that a better estimate is more sensitive to the level of phase noise fluctuations. The reference DEM covered a territory 8 × 5 km, which contained average hills and river valleys and had the vertical height accuracy about 1.7 m, which is, at least, triple times better than the expected ALOS PALSAR DEM accuracy (about 7–8 m [10]. An interferogram has a size of 2000 × 750 elements. The obtained dependecies σpsi (γ̂) are shown in Fig. 4. The range of values for γ̂0 was about 0.25–0.75, and 0.2–0.6 for γ̂2 . Extremely high values (γ̂0 > 0.7 and γ̂2 > 0.55) were excluded because they were, as a rule, corresponded to terrain elements that are not joined with the Earth relief (buildings, facilities, roads, engineer communications, etc). As it is seen from Fig. 4, γ̂2 estimate has more wide band for the standard deviation than γ̂0 and it has a linear decreasing section for 0.2-0.45 coherence values, so it better reflects the quality of the InSAR data; γ̂0 has an abnormal behaviour, the standard deviation increases within the coherence value. Both estimates have an abnormal behaviour for extremely low and extremely high coherences. 4 Conclusions A coherence estimation algorithm is proposed based on 2-FFT normalized peak height assessment. It is shown that such estimate has a significantly less depen- 104 Fig. 4. Dependency of the absolute phase standard deviation on the coherence esti- mates. dence on the topographic phase slope. A method of coherence estimates quality assesment is proposed, based on calculation of InSAR absolute phase deviation from a reference DEM. It is also shown, that proposed coherence estimate has a quazi-linear decreasing section on the σψ (γ̂) dependency that correctly charac- terizes it as a quality measure for InSAR data. Acknowledgments. The work was supported by Act 211 Government of the Russian Federation, contract 02.A03.21.0006. References 1. Elizavetin, I. V., Ksenofontov, E. A.: Rezultaty experimentalnogo issledovaniya vozmozhnosti pretsizionnogo izmereniya reliefa Zemli interferentsionnym metodom po dannym kosmicheskogo RSA [The results of experimental research of precious Earth relief measurement by interferometric method with space-based SAR]. Issle- dovaniya Zemli iz kosmosa. 1, 75-90 (1996) 2. Hanssen R. F.: Radar Interferometry. Data Interpretation and Error Analysis. Dordretch. Kluwer (2001) 3. Touzi, R., Lopes, A., Bruniquel, J., Vachon, P. W.: Coherence estimation for SAR imagery. IEEE Trans. Geosci. Rem. Sens. 37, 135-149 (1999) 4. Lopez-Martinez, ., Pottier, E.: Coherence estimation in synthetic aperture radar data based on speckle noise modeling. Applied Optics. vol. 46, no. 4, 544-558 (2007) 5. Baran, I., Stewart, M. P., Kampes, B. M., Persky, Z., Lilli P.: A Modification to the Goldstein Radar Interferogram Filter. IEEE Trans. Geosci. Rem. Sens. vol. 41, no. 9, 2114-2118 (2003) 105 6. Zebker, H. Chen, K.: Accurate Estimation of Correlation in InSAR Observations. IEEE Geoscience and Remote Sensing Letters, vol. 2, no. 2, 124-127 (2005) 7. Payne, T. M.: The effect of scene elevation on the coherence of wide-angle crossing- node SAR pairs. IEEE Trans. Geosci. Rem. Sens. vol. 42, no. 3, 520-528 (2004) 8. Lee, H.; Liu, J. G.: Topographic phase corrected coherence estimation using multi- pass differential SAR interferometry: differential coherence. IGARSS 2000. vol.2, 776-778 (2000) 9. Kobernichenko, V. G., Sosnovsky, A. V.: Coherence estimation in interferometric SAR data processing. 2014 24th Int. Crimean Conf. Microwave & Telecommuni- cation Technology (CriMiCo2014), Sevastopol, pp. 1155-1156. SevGU (2014) 10. Sosnovsky, A., Kobernichenko, V.: An accuracy estimation of digital elevation models obtained by interferometic synthetic aperture radars. XXII international conferece Radiolokatsiya, navigatsiya, svyaz’ (RLNC’2016), Voronezh, Russia, pp. 1074-1081. NPF SAKVOEE (2016)