The purpose of this article
The devastating major earthquake which occured on December 23, 2003 with a moment magnitude of 6.6, that caused a huge number of human casualties, as well as major destructions in Bam and the surrounding Kerman province in Iran, had been the topic of several geophysical studies in the past 10 years. Furtunately, and due to the scientific effort of the seismological societey, the basic processes that induced the earthquake are well known by now. This short article should provide some useful information for processing spaceborne Radar data for the detection of earthquakes and faults. For this purpose I have chosen ASARAdvanced Synthetic Aperture Radar data from the ENVISAT mission before and after the event to determine the structural differences and rupture zones induced by the earthquake.
The data sets provided by ESA are:
subset_9_of_ENVISAT-ASA_IMS_1PXPDE20040211_061300_000000142024_00120_10194_0013_ML.dim
subset_10_of_ENVISAT-ASA_IMS_1PNUPA200331203_061259_000000162022_00120_09192_0099_ML.dim
Which processing steps were applied to the product?
As post-processing is very important for Radar data, several corrections and image processing algorithms have to be applied on the data. According to the product name the Single-Look Complex (SLC) image data is a high-resolution, narrow swath products based on data acquired at one of seven subswaths. It is intended for SAR image quality assessment, calibration and interferometric or wind/wave applications. A small number of corrections and interpolations are performed like the preprocessing (for all data the same; i.e. I/Q bias removal, I/Q gain imbalance correction, I/Q non-orthogonality correction), a Doppler Centroid Estimation and a Range Doppler Image Formation. To obtain the Doppler Centroid Estimation for the high-resolution Image Mode of the product, the Multi-Look Cross Correlation Method is used.
NEST Multilook Operator
The original SAR images with inherent speckle noise were processed with the Multilook Operator to reduce the speckle noise by incoherently combining several images as if they corresponded to different looks of the same scene and to produce an image with nominal image pixel size (azimuth to range ratio of 5:1, reprocessed to 1:1). As a result, the final images approximately have square pixel spacing.
The speckle effect is a result of the interference of many waves of the same frequency, having different phases and amplitudes, which add together to give a resultant wave whose amplitude, and therefore intensity, varies randomly. When a surface is illuminated by a light wave, according to diffraction theory, each point on an illuminated surface acts as a source of secondary spherical waves. The light at any point in the scattered light field is made up of waves which have been scattered from each point on the surface.
Range and Azimuth Direction obtained from the metadata
The original intensity image has num_output_lines=5000, num_samples_per_line=1200 with an azimuth_looks=1.0 and range_looks=1.0. The processed image has num_output_lines=1000 with azimuth_looks=5.0. Therefore the azimuth direction is from north to south whereas the range direction is from west to east.
Why is the automatic coregistration necessary?
The process accurately co-register the master image (11th February 2004) with the slave image (3rd December 2003). For this purpose the master image is used as a reference image and the slave image is overlayed in a way that each pixel of the slave image corresponds to the same geoposition pixel in the master image. Due to the fact that both images are complex images and that the scene described by the images is in a flat area the build-in algorithm ensures a matching accuracy of better than 0.05 pixels. For complex images the registration is achieved by maximizing the coherence between master and slave images at a series of imagettes defined across the image. After the registration, a warp function is created which maps the master-slave Ground Control Points (GCP) of the slave image into pixels in the master image.
What information can be obtained out of the residual file?
The Create Stack Operator allows collocating two spatially overlapping products. Collocating two products implies that the pixel values of one product (the slave) are resampled into the geographical raster of the other (the master). When two products are collocated, the band data of the slave product is resampled into the geographical raster of the master product. In order to establish a mapping between the samples in the master and the slave rasters, the geographical position of a master sample is used to find the corresponding sample in the slave raster. The calculation of the new pixel value is performed by weighting the 36 (for the chosen settings) surrounding pixels.
RMS mean = 1.2889993312800754
RMS std = 3.7586558379704043
Parse 1
RMS mean = 0.02553238345095844
RMS std = 0.022476339664258316
Parse 2
RMS mean = 0.0241239868209786
RMS std = 0.02009919619001828
For the parse 0, every valid GCP is used to resample the projected output product (by Cubic Convolution) of the slave image. The respective RMS errors in x- (row) and y- (column) direction are listed in the residual file. All GCP-pairs of the master and slave image that exceed the RMS threshold defined previously are excluded from the 2nd parse of the same routine. As a result, the RMS mean gets better by each use of the Cubic Convolution. As the RMS values below show, the accuracy increases significantly from Parse 0 to Parse 1, but only very poor from Parse 1 to Parse 2. As the desired accuracy of 0.05 was already reached in Parse 1, the process could have been stopped after the first parse.
The definition of coherence is the similarity of the co-registered GCP-pairs (by Coarse Registration and Fine Registration by performing a cross-correlation of master and slave imagettes) and a computation of the best sub-pixel shift of GCP such that the slave imagette at the new GCP position gives the maximum coherence with the master imagette according to Powell’s method. Assuming I1 and I2 be RxC imagettes and I*2 is the complex conjugate of I2, coherence is computed by:
Method:
The coherence in this scene is computed with a 10×2 (azimuth x range due to the 5:1 looks ratio) sliding window in two steps:
- First for each pixel in the imagette, a 10×2 window centered at the pixel is determined for both master and slave imagettes, and coherence is computed for the two windows using equation above.
- Average coherences computed for all pixels in the imagette to get the final coherence for the imagette.
The low coherence areas (dark spots) correspond mainly to building structures like houses and roads as well as vegetation, whereas the high coherence areas (bright spots) correspond to the surrounding desert in the scene. Due to the earthquake, a phase-shift occured on the building structures (destroyed buildings) and the vegetation differs naturally with each image. Therefore the coherence value is very low for the city area whereas the surrounding desert isn’t effected that much, because the phase-shift is quite low (except for the new crack) due to a lack of horizontal or vertical displacement and therefore low differentiation.
The low coherence areas (dark spots) in Figure 3 correspond mainly to building structures like houses and roads as well as vegetation, whereas the high coherence areas (bright spots) correspond to the surrounding desert in the scene. Due to the earthquake, a phase-shift occured on the building structures (destroyed buildings) and the vegetation differs naturally with each image. Therefore the coherence value is very low for the city area whereas the surrounding desert isn’t effected that much, because the phase-shift is quite low (except for the new crack) due to a lack of horizontal or vertical displacement and therefore low differentiation. See Figure 4 for a more natural looking image of the scene.
Why do we have to remove the topographic phase?
InSAR is normally applied for estimating topography. Here we are interesting in the displacement only, meaning that we need to subtract the phase difference caused by topography. That can be done if a DEM is available or by means of a third SAR image. The topography around Bam is rather flat, and therefore the interferogram doesn’t change much.
The interferogram and estimate of the displacement
This interferogram (Figure 5) represents contours of ground displacements towards and away from the Envisat satellite that occurred during the earthquake (each fridge corresponds to half the wavelength of the radar = 2.8cm). There is a two-quadrant pattern consistent (in this descending orbit image) with a near-vertical strike-slip fault oriented north-south (Figure 6). The ground moved approximately 30cm towards the satellite in the south-east quadrant, and approximately 15cm away from the satellite in the north-east quadrant.
Advanced Synthetic Aperture Radar
More details about the earthquake