COASTLINE EXTRACTION USING ERS SAR INTERFEROMETRY

ABSTRACT

Detection of coastlines in satellite imagery is of interest for a lot of geoscientific applications such as the generation of intertidal DEMs and the monitoring of the coastal morphodynamics. A novel method is presented using the recently developed techniques of interferometric SAR data processing. In SAR interferometry coherence maps are generated that describe the degree of stability of the complex scatterers between the two acquisitions. Since the ocean surface is highly nonstationary as opposed to land surfaces, the coherence information can be exploited to determine the land/sea boundary. A fully automatic procedure is developed detecting the line of highest water level between the two acquisitions (1 day), using a sequence of image processing algorithms to refine the accurate position of the line. The method is tested on two different data sets from the German Bight and the Strait of Messina. For the first one a 50 m topography modal of the intertidal zone provided by the German Federal Institute for Waterway Engineering was used for comparison.

 

Keywords: SAR, interferometry, coastline detection

Introduction

The determination of coastlines is a well-known problem in satellite remote sensing that has been treated in numerous publications [1][2]. Whereas it is a quite simple task with optical data such as Landsat TM, the SAR data inherent properties (e.g. speckle statistics, complexity of the backscatter mechanisms) impede a straight-forward solution of the problem using only the pure amplitude information of the data. Apart from the presence of speckle a lack of contrast between land and sea often is observed in case of wind-roughened or wave-modulated water surfaces. Fig. 1 illustrates the difficulty of shoreline detection in SAR magnitude images.

Fig. 1: ERS-2 SAR magnitude of the coastal zone south of Taormina/Sicily (6.9.95)

In this paper a new technique is proposed to overcome the problem of missing contrast in the SAR magnitude. As will be demonstrated, the use of the interferometric correlation as a measure of the stability of the scatterers between two SAR acquisitions may serve as a criterion to distinguish between the land and sea surfaces. Coherence information has already been exploited successfully in landuse applications [3] or for the purpose of forest/non-forest-discrimination [4]. Forested areas, similar to water, mainly appear decorrelated in repeat-pass interferograms whereas agricultural and urban areas are characterized by medium to high correlation.

In the following we start with a description of the interferometric data processing and the main aspects of calibrated coherence estimation. In section 3 the developed coastline extraction algorithm is outlined. The results of the application of the method on data of two testsites, covering a coastal zone south of Taormina/Sicily and one in the German Bight/Germany are reported in section 4. Section 5 is devoted to a short summary and conclusional remarks.

Interferometric Processing

The generation of calibrated coherence maps is a basic requirement for accurate coastline determination and therefore needs careful interferometric processing of the ERS SAR SLC data. The coherence, defined as

suffers from any kind of spatial phase variation within the averaging window. Subsequently, phase noise as well as systematic phase patterns caused e.g. by the imaging geometry or the ground topography should be avoided.

Regarding the noise contribution, careful interferometric processing was applied including the following features:

• slope adaptive spectral shift filtering in range to eliminate the effect of diverse incidence angles [5];
• band pass filtering in azimuth to remove the frequency shift caused by the antenna squint angle difference [6];
• subpixel fine registration of the data with an accuracy better than 0.1 samples;
• oversampling of the complex data by a factor of 2 in order to avoid aliasing during the interferogram formation process;

The systematic phase pattern was accurately removed by subtracting the (rewrapped) unwrapped phase from the interferogram. This allowed to use large window sizes for a more realistic coherence estimation [7]. Fig. 2 illustrates visually the effect of different averaging sizes, revealing an increasing coherence value homogeneity with increasing sizes at the expense of a decreasing geometric resolution. Note, that the averaging window size may differ significantly from the effective number of averaged independent samples (² number of looks² ) due to the original oversampling of the SLC data and possible subsequent low pass filter operations.

Fig. 2: Coherence maps of the German Bight testsite using the 3 different number of looks 23, 64, and 208, respectively. The common grey value-coding is used (dark = low coherence, bright = high coherence)

Coastline extraction algorithm

The determination of the coastline is purely grounded on an examination of the interferometric correlation. As described in the following, the extraction is performed in a first step individually for each ² resolution level² (i.e. number of averaged samples), followed by a refinement of the individual lines, which are finally merged to yield an accurately positioned shoreline.

Basic coastline determination methodology

The main idea of the algorithm is to detect the positions of large coherence value gradients in the correlation map which should reveal the boundaries between the high coherent land areas and the decorrelated water surface. A series of basic image processing steps are applied to this purpose:

In a first step, the image is median filtered in order to suppress coherence value variance which may be present considerably, yet dependent on the window sizes used for coherence estimation (see Fig. 2). After filtering a threshold is defined, marking all pixels that are supposed to have no correlation at all despite of their actual (non-zero) value. The following gradient estimation in both the x- and y-directions together with the threshold application then leeds to a preliminary, coarse separation of the high and low coherent areas.

Refinement of the individual lines

Each individual line is refined in the following process. First morphological dilation and erosion operators are applied to remove possible holes. The hence ² opened² line structures are afterwards thinned using a standard thinning algorithm [8]. Finally, in order to disregard lines that might be caused by small decorrelated features within the land areas (e.g. small lakes), lines with a total length shorter than 50 image samples are withdrawn from the result.

Merging of the lines

The individual coastlines generated with the different resolution levels differ from each other regarding their shape as well as their location. For instance, small window sizes will lead to a less reliable positioning due to the remaining noise in the coherence map, whereas larger window sizes tend the gradient threshold to be shifted towards the water region. Fig. 3 illustrates the coherence behaviour at the land/sea boundary for the 3 window sizes 23, 46, and 208 looks.

According to Fig. 3 the true location of the coastline is supposed to be at the intersection of the lines of the different resolution levels. Since the individual lines have been generated by gradient estimation the intersection point can be found through a merging of the lines, followed by dilation and again thinning to accurately position the result.

Fig. 3: Coherence behaviour at the land/sea boundary for 3 different number of looks

Results

The developed coastline extraction algorithm was applied on two interferometric data pairs covering the areas south of Taormina/Sicily and the German Bight/Germany, respectively. The relevant scene parameters are reported in Tab. I.

Tab. I: Relevant parameters of the scenes under investigation

Fig. 4 shows the Taormina scene with the extracted shoreline superimposed on the magnitude of the ERS-1 image. A remarkable correspondence can be observed in most of the areas. Compared to the amplitude information contained in the ERS-2 scene (see Fig. 1) the additional use of the coherence indicates a significant improvement regarding land/sea segmentation.

For the German Bight testsite the analogous superimposition can be seen in Fig. 5. This area which is characterized by shallows bordering the land region again underlines the difficulties to extract coastlines solely with SAR magnitude information.

For this scene additional reference data was made available by the German Federal Institute for Waterway Engineering: a topography model of the region (Fig. 6) and information about the average high tide level. Both data sources were used to derive the corresponding contour lines, which subsequently were superimposed on the SAR data. Fig. 7 and Fig. 8 display the result, using the colour coding red = InSAR coastline, green and blue = height contour lines 0 m and 1 m, respectively, yellow = average high tide level. It can be seen that the InSAR coastline is in good correspondence to the 1 m height contour line but seems to be shifted a little towards the water. Having in mind that the InSAR line marks the highest water level that was reached between the two data acquisitions, that level is supposed to lie somewhere between 0 and 1 m.

The coherence measurement reveals additional areas within the shallows that could not have been flooded during the acquisition time interval, hence marking regions with a topographic height of the measured level. This is in contrast to the topography model and may have the reason in the generation date of the model (1992). The shallow areas in the German Bight are known to be subject to intense morphodynamical changes.

Summary and conclusions

A novel fully-automated coastline extraction algorithm was developed which is based on the examination of the interferometric correlation of two coherent SAR images. We showed that an accurate shoreline determination is possible if sufficient data coherence is present in the investigated scenes. The derived coastline of the German Bight was found to be in excellent correspondence with reference data supplied by the German Federal Institute for Waterway Engineering.

We conclude that coherence data contain useful information regarding the detection of land/sea boundaries in SAR data and can be used for the monitoring of coastal morphodynamics. However, some limitations of the methodology have to be kept in mind. In particular, the temporal correlation of the land surface is an important requirement which is often not met, especially if data of larger ERS repeat cycles (e.g. 35 days) are used. Furthermore, the interferometric measurement is restricted to the detection of the line of highest water level between the acquisitions. Nevertheless, the coherence information serves as an additional layer of information even in critical cases, which can support the interpretation of the amplitude information.

Fig. 4: With InSAR extracted coastline (red), superimposed on the ERS-1 magnitude of the coastal zone south of Taormina/Sicily (6.9.95)

Fig. 5: With InSAR extracted coastline (red), superimposed on the ERS-1 magnitude of the German Bight (13.3.96)

Fig. 6: Topography model of the German Bight area (German Federal Institute of Waterway Engineering 1992)

Fig. 7: InSAR-derived coastline (red), height contour line of 0 m (green) and of 1 m (blue), superimposed on the geocoded ERS-1 magnitude of the German Bight scene

Fig. 8: InSAR-derived coastline (red) and average high tide level (yellow), superimposed on the geocoded ERS-1 magnitude of the German Bight scene

Acknowledgement

The work was funded by the German Ministry of Research (BMBF) under contract No. 03F0165C. The data were kindly provided by ESA as part of the ERS AO project AO02.D113.

References

[1] Lee, J. & I. Jurkevich 1990, Coastline Detection and Tracing in SAR Images, IEEE Trans. Geosc. Rem. Sens., Vol. 28, No. 4, 662-668

[2] Zhang, D., L. Van Gool & A. Oosterlinck 1994, Coastline Detection from SAR Images, Proc. IGARSS’94, 2134-2136

[3] Wegmüller, U. 1996, Land Applications Using ERS-1/2 Tandem Data, Proc. FRINGE ‘96, ERS Fringe Workshop Zürich

[4] Askne, J. & J. Hagberg 1993, Potential of Interferometric SAR for Classification of Land Surfaces, Proc. IGARSS’93, Tokyo, 985-987

[5] Gatelli, F., A. Monti Guarnieri, F. Parizzi, P. Pasquali, C. Prati & F. Rocca 1994, The Wavenumber Shift in SAR Interferometry, IEEE Trans. Geosc. Rem. Sens., Vol. 32, No. 4, 855-865

[6] Schwäbisch, M. & D. Geudtner 1995, Improvement of Phase and Coherence Map Quality Using Azimuth Prefiltering: Examples from ERS-1 and X-SAR, Proc. IGARSS’95, Firenze

[7] Touzi, R., A. Lopes, J. Bruniquel & P. Vachon, Unbiased Estimation of the Coherence for SAR Imagery 1996, submitted to IEEE Trans. Geosc. Rem. Sens.

[8] Pavlidis, T. 1982, Algorithms for Graphics and Image Processing, Computer Science Press, Ch. 9.1

Keywords: ESA European Space Agency - Agence spatiale europeenne, observation de la terre, earth observation, satellite remote sensing, teledetection, geophysique, altimetrie, radar, chimique atmospherique, geophysics, altimetry, radar, atmospheric chemistry