Ground Control Point Generation from Simulated SAR Image Derived from Digital Terrain Model and its Application to Texture Feature Extraction

Ground Control Point: GCP generation from simulated topographic map derived from Digital Terrain Model: DTM is proposed. Also, texture feature extraction is attempted from the simulated image. In this study, simulated image is derived from elevation data only, under assumptions of a simple scattering model without consideration of complex dielectric constant of the targets of interest. The performance of the acquired GCPs was evaluated by using several measures with texture features of GCP chip images. This paper describes the details about proposed method for acquisition of GCPs and simulated results on relationship between texture features and GCP matching success rate corresponding to the cross correlation between reference and distorted GCP chip images. Keywords—Ground Control Point: GCP; Digital Terrain Model: DTM; scattering model; complex dielectric constant; texture feature; matching success rate; GCP chip


I. INTRODUCTION
Various researches have been conducted on the method of extracting [1]- [5] Ground Control Point: GCP from image data obtained by optical sensors. In particular, these days, the technology in optical sensors can be effectively applied to image data by SAR (synthetic aperture radar) [6]. Have been shown in a study by Dr. Guindon et al [1]. They have proposed an automatic matching method between real and simulated SAR images with distortion due to terrain effects.
Essentially, space-based SAR imagery data has distortions due to layover, foreshortening, shadowing and so on. Therefore, geometric fidelity of the SAR imagery data is still key issues. One of the problems of geometric fidelity improvement of the space-based SAR imagery data is GCP extraction from the SAR imagery data. GCP extraction is not so easy. Also, another problem is matching accuracy improvement. By using GCPs, geometric distortions have to be removed through GCP matching. This is not so easy either. In order to improve these GCP extraction and GCP matching processes, GCP extraction from digital elevation model, and chip image based GCP matching are proposed here.
In this study, simulated a SAR image is generated using digital terrain model (DTM) and tried to extract GCP. In addition, as an evaluation of the extracted GCP, matching was performed between the GCP chip image with some distortion and the original image, and the accuracy was examined.
Furthermore, in order to evaluate the matching accuracy of the GCP chip using only the chip image, the correlation between the texture features and the accuracy was examined [7].
In this paper, the author first describes the simulation method for the SAR image adopted in this study, and then introduce the features introduced as indices for GCP evaluation [8]. Finally, the extracted GCPs are evaluated using these characteristic quantities, and the conditions for the optimal GCP are examined.
In the following section, related research works are described. Then, the proposed method is described followed by experimental set-up together with experimental results. After that, concluding remarks and some discussions are described.

II. RELATED RESEARCH WORKS
Automated matching of real and simulated SAR image as a tool for GCP acquisition is proposed [1]. Image-scale and look direction effect on the detectability of lineaments in radar images is studied and well reported [2]. Studies on GCP matching of remote sensing image data is conducted [3]. Accuracy of digital elevation data by various interpolation methods is assessed [4].
A method for textural analysis of synthetic aperture radar images by Grey Level Co-occurrence Matrix: GLCM method is proposed [5]. Also, Synthetic Aperture Radar: SAR is well overviewed [6]. Landsat Thematic Mapper: TM image classification using seasonal change of texture information is proposed [7]. Outline of national land numerical information is well organized [8].
Some studies on GCP selection process and its accuracy is investigated [9] together with effects on GCP success rate [10]. On the other hand, experiment on GCP matching is conducted and well reported [11] together with studies on GCP matching of remote sensing imagery data [12].
GCP acquisition using simulated SAR derived from Digital Elevation Model: DEM is attempted [13] together with GCP acquisition using simulated SAR and evaluation of GCP matching accuracy with texture features [14].
Effect of planimetric correction with a few GCPs on terrain height estimation with stereo pair is evaluated [15].
Meanwhile, speckle noise removal of SAR images with DEM is proposed [16]. Then, a method of speckle noise reduction for SAR data is proposed and tested [17]. After that, a new method for SAR speckle noise reduction (Chi Square Filter) is proposed and evaluated its effectiveness [18].
Decomposition of SAR polarization signatures by means of eigen-space representation is proposed [19]. Also, SAR image classification based on Maximum Likelihood: MLH decision rule with texture features taking into account a fitness to the probability density function is well reported [20]. Meanwhile, polarimetric SAR image classification with maximum curvature of the trajectory in eigen space domain on the polarization signature is proposed [21] together with polarimetric SAR image classification with high frequency component derived from wavelet multi resolution analysis: MRA [22].
Comparative study of polarimetric SAR classification methods including proposed method with maximum curvature of trajectory of backscattering cross section in ellipticity and orientation angle space is conducted [23]. Multi-Resolution Analysis: MRA and its application to polarimetric SAR classification is proposed [24]. Recently, Sentinel 1A SAR data analysis for disaster mitigation in Kyushu is reported. [25]. Comparison of geometric features and color features for face recognition is conducted and discussed [26].

A. Generation of SAR Image
The proposed method for simulation of SAR image is based on DTM or DEM. There are some sources of the DTM. It allows slope calculation in accordance with its mesh size or spatial resolution. Without consideration of geometric distortions derived from foreshortening, shadowing, layover, speckle noise, SAR image can be simulated using just slope calculated with DTM, satellite altitude and off-nadir angle (look angle).
In the proposed model, the backscattering coefficient is calculated according to the following procedure.

1) Calculate the slope between each point from DTM.
2) Calculate the angle of incidence of the radar wave from the positional relationship between the satellite and the observation point, and calculate the distribution of the backscattering cross section 6a assuming equation (1).
where Ir, In, and θ are the scattering direction unit vector, the scattering plane normal vector, and the radar wave incident angle, respectively.

3)
Since the observation point is distorted in the distance between pixels due to the terrain effect, Lissasoprisog is performed at right angles using the following equation.
where P(i): i-th pixel value, Li: i-th pixel value before resampling, x: distance between pixels, δxi: Pixel shift due to terrain effect.
Pixel shift δxi due to the effect of layover, foreshortening, and shadowing was calculated from topographic information, and linear interpolation was performed by equation (2). However, the addition of a few terms in equation (2) was performed after judging whether or not x / 2-Δxi contributed to the i-th pixel value.

4)
Convert the obtained values of Pi (X, Y) X, Y = 1 to 320 into 8-bit gray level image data of 0 to 255.

B. Extraction of GCP
In order to extract GCP chip images from the simulated SAR image, the SAR image is divided with small size of subimages as candidates of GCP. Then, variance of the candidates of GCP chip image is calculated. It is assumed that GCP chip image has a relatively large variance. The image obtained in this way was divided into chips of 32 × 32 pixels, and the variance of the pixel value of each chip was calculated.
The actual SAR image has a geometric distortion caused by the deviation of the position and attitude of the satellite or their estimation errors such as skew and rotation. In order to simulate them, a skew or rotation is applied to the original image obtained by extracting the GCP chip as geometric distortion, and the area correlation between the resulting distortion image and the original GCP chip image is calculated. Here, the skew and the rotation were given as the center of rotation of the GCP chip image as shown in Fig. 1.

C. GCP Matching
Detailed steps for matching are as follows,

1)
A point corresponding to the center point of each GCP chip in the original image is defined as GCP, and skew distortion and rotation distortion are given around this point.
2) A search window of 48 × 48 pixels is provided on the original image side centering on GCP.
3) Shift the GCP chip by one pixel in the search window, calculate the area correlation of the overlapping part, and obtain the peak of the correlation coefficient.
Here, the peak of the correlation coefficient was obtained by extracting nine correlation coefficients centered on the pixel of interest and interpolating independently in the X and Y directions by a quadratic equation.

D. Texture Feature Extraction
The author proposes a method for quantitatively evaluating the matching accuracy of GCP chips based on texture information. If the correlation between the matching accuracy and some feature amount of the chip image is strong, the GCP candidate with high automatic matching accuracy can be easily selected by evaluating the feature amount. In this study, feature extraction was performed using GLCM (Gray Level Co-occurrence Matrix) proposed by Haralick et. al (1973).
where P is the normalized co-occurrence probability matrix, and P is the co-occurrence probability. The rate matrix (GLCM), d is the distance between pixels, and Nq is the number of gray levels.
Next, the following texture feature values were calculated from GLCM. Here, P (i, j, d, θ) is abbreviated as Pij.
i) Angular Second Moment ASM = ΣiΣjPji 2 (5) ii) Homogeneity iii) Contrast iv) Dissimilarity vii) Chi-square where, Pij is the average and standard deviation of the pixel values in the co-occurrence probability matrix direction, and Px (i) Py (j) is the appearance probability in the matrix direction, respectively.

A. Generation of SAR Image
The DTM used for the simulation is a magnetic tape file (KS110) produced by the Geospatial Information Authority of Japan. In this file, altitude data is stored in m units according to a format called a standard area mesh system. The primary area division mesh included in this data has a size of 80km by 80km, which is equivalent to one terrain map of 20 planes, and is large enough as an East area. The section of code No.5338 where all the data of the grid points were prepared was selected from these. This is the area that includes Mt. Fuji in the lower right as shown in Fig. 2. Fig. 3 shows the simulated SAR image (Full scene) derived from the DTM.
The data is divided into 320 × 320 primary partitions, and arranged in a complicated manner according to the mesh code system, and is converted into a two-dimensional image. Since the data is divided into primary partitions of 320 × 320 and arranged in a complicated manner according to the mesh code system, it is difficult to process it as a two-dimensional image. In addition, in this data, since there are several abnormal grid points without elevation value as inland waters, they were replaced with the average value of two points before and after. Fig. 4 shows an example of a sub scene of the simulated SAR image.
Furthermore, the altitude values from 0 to 3776 m were requantized to 255 gray levels for processing as monochromatic images. The grid point interval of this mesh is 250 m on the actual ground, but it is coarse compared to the resolution, so the upper left corner was cut out as a subscene (Northern Akaishi Mountains), and interpolation was performed at 50 m intervals with a cubic spline function.

B. GCP Extraction
Since the observation area is a mountain area covered mostly by trees, it is considered to be a perfect diffusion surface, and it is assumed that the physical backscattering coefficient is uniform, and the reflected power obeys Rampart's cosine law. As the observation conditions from the satellite, as shown in Fig. 5.
It is assumed that observation is performed from west of the test site at an altitude of 570 km and an off-nadir angle of 35 ° while flying north-south.
The images obtained in this way are divided into chips of 32 x 32 pixels, and the results of calculating the variance of the pixel values of each chip are shown in Fig. 6, 7. Of these, 14 chips shown in Fig. 6, 7 with a tether were randomly selected as GCP candidates. Figure 8 shows three-dimensional images of these chips. Table I shows the matching accuracy of each chip when the skew and rotation were changed every 1° from 1° to 4°, respectively. The matching accuracy is the Euclidean distance between the estimated GCP indicated by the peak of the correlation coefficient and the true position. Skew and rotation can occur simultaneously in the real image, respectively, but each case is treated independently to separate and evaluate. Miss-identification is defined as the sum up of distances, between GCP chip center and functional correlation peak point, for which skew or rotation distortions from 1 to 4 degrees with 1 deg. step are added to the original image. Here, R 1 and R 2 correspond to, respectively, skew and rotation. Table II shows the relationship between GCP matching accuracy and each feature value by correlation coefficient.

C. Texture Feature of GCP Chip Image
As an example, Fig. 9 shows the relationship between the magnitude of GCP mismatch due to skew distortion and GLCM CON. Here, the sizes of the mismatches from 1 to 4 were added and evaluated. Clearly a negative correlation is observed. Almost the same results were obtained for the other feature values. When the texture features in four directions were adjusted, it was found that there was a difference in the texture features in the 90° direction compared to the other three directions, as shown in Table III.       Table III shows that in this image, the 90 direction has less change than the other three directions. This is because the observation direction is 180. It can be seen that these features are highly dependent on the beam irradiation direction in the SAR image.
2) Skew distortion has a higher correlation between each feature and matching success rate than rotation distortion. This is considered to be due to the fact that the rotation has more distortion (larger loss (change) of the original information) than the skew that deforms only in one direction even at the same angle.
3) Variance and chi-square values have low correlation with GCP matching accuracy. This is because they do not always correspond to the complexity of terrain change (BUSYNESS).
4) The contrast, dissimilarity, entropy, angular second moment and homogeneity show relatively strong negative or positive correlation with GCP mismatch degree. Therefore, a regression equation can be obtained from these correlations to predict the degree of GCP mismatch. 5) HOM, COM and DIS have cross-correlation, and ASM and ENT have high cross-correlation. Furthermore, there is a negative correlation between these groups. Therefore, it is desirable to evaluate GCP by selecting an appropriate one.

VI. CONCLUSION
A method for extracting chip images suitable for GCP using simulated SAR images created from DTM was proposed. The followings are major results, 1) As a result of examining the relationship between the matching accuracy and the texture feature of the chip, the correlation with the index indicating the feature of GLCM of