Enhanced Noise Removal Technique Based on Window Size for SAR Data

Transcription

1 Volume 114 No , ISSN: (printed version); ISSN: (on-line version) url: ijpam.eu Enhanced Noise Removal Technique Based on Window Size for SAR Data G.Siva Krishna 1, N.Prakash 2, 1Research Scholar, Dept of IT, B.S.Abdur Rahman Crescent University, Chennai, India , 2 Associate Professor, Dept of IT, B.S.Abdur Rahman Crescent University, Chennai, India sivakrishna_it_phd_2016@bsauniv.ac.in prakash@bsauniv.ac.in Apr 14,2017 Abstract Synthetic aperture radar (SAR) image feature extraction is becoming a vital technique for remote sensing, and for this, many tools are available. In this research work the researcher proposes a new technique called an Enhance Gamma Map (EGM) technique. In this technique (which depends upon the window size) a subset from the SAR image is collected and this collected subset image should be converted from Slant Range (SR) to the Ground Range (GR). This SR to GR conversion helps to calculate calibrate values and also to find the sigma naught (σ0) value from each pixel. Afterward, based on sigma naught value, the Land use and Land Cover (LULC) facts can be identified. Hence, the EGM technique removes the noise and it helps to calculate the histogram and scatter plots. Then, this calculated 227

2 (histogram and scatter plots) values very effectively identify the SAR image with 72% improvement on ground truth. Key words: SAR, LULC, EGM, SR, GR, Filter Techniques. I. Introduction SAR has two types of radar sensors, namely active sensor and passive sensor. These two types of radars capture the image day and night (24 by 7) and also it provides cloud free data. Today the demand for SAR data is increasing in the real world because the number of images capturing from the radars for analysis purpose. This SAR images can associate with basic textures and it commonly effects by the noise. Therefore, it is very difficult to identify the objects in the SAR image. In order to handle this problem, a popular image processing technology is introduced, that is image feature extraction [1, 2]. SAR image feature extraction is becoming a vital technique for remote sensing [3, 4]. Therefore this research work concentrates more on the speckle noise for removing the speckle. Today plenty of techniques and approaches are proposed by many authors [5, 6] for removing speckle though speckle is not a noise in remote sensing. Because each speckle pixel intensity is an important characteristic to identify LULC categories such as a built-up area, water body, forest etc,. The characteristics of S A R beams have different level of polarization and this polarization has two types of radar system, namely Horizontal (H) and Vertical (V). These H and V are linear polarizations, thus their combinations lead to the following channels: 1) Like-polarized: It means H transmit and H receive (HH) or V transmit and V receive (VV), because, transmit and receive polarizations are same direction[7][8][11]. 228

3 2) Cross-polarized: It means H transmit and V receive (HV) or V transmits and H receives (VH), because transmit and receive polarizations are orthogonal to one another [12]. Hence; each intensity value is relevant to LULC fact value [9] [3] [10]. This research work organizes to follow section II as the proposed technique. The experiments and result explanation in section III, and the conclusion explained in section IV. II. Enhanced noise removal technique The proposed Enhanced Noise Removal Technique is called Enhanced Gamma Map. It proposes two types of phase framework. Phase-I has subset selection from the SAR image and convert into SR to GR, phase-ii, calculate calibrate correction value and also apply various filter techniques for noise removal. This framework explains one by one in detail with figures. The proposed framework is shown in Figure 1. Input SAR Image Phase-I Subset of Convert Phase-II Radiometric correction Calibrate correction Noise Remove Analysis and visualize (ROI, RGB) Figure 1. Framework for the proposed Methodology 229

4 Phase-I: The subset selection is made with an upper pixel axis X, Y and the bottom pixel X, Y or randomly selects some portion of the SAR image. This subset image applies the slant range to ground range for conversion [6] [12] [7].Therefore, this conversion detects an object and it converts only useful information like patch-wise. Phase -II: Here, the conversion image is used to calculate Radiometric Correction and then it returns calibrate coefficient value, and this calibrate coefficient value projects with an equivalent to ground truth values. Here the calculating calibrate value is based on sigma naught (σ0) coefficient. It is shown in the equation below. Normalized Radar Cross-Section (Backscatter Coefficient/ sigma naught) σ0 (db) = 10.Log10 (Energy Ratio) Where Energy Ratio = Received energy by the sensor/energy reflected in an isotropic way, db is the decibel. From sigma naught value, the relevant LULC facts are shown in Table-1 III. Experiment and result B σ Surface a V 0 A Manmade objects (urban), e H b - terrain Rough surface, slopes towards the dense i M 1 - vegetation Medium level (forest) of o 2 vegetation, crops, L B Smooth surfaces, calm o e water, the road, very dry Table 1: identifying the LULC fact values Study Area and Data Collection: Vancouver is part of Canada; a region of British Columbia. This geographic vicinity gives various terrains from a bulky mountain to the flat agricultural lands of the Fraser River Delta in the north of Vancouver. This area information gives the RADARSAT-2, product name is RS2- SLC-FQ28-DES-24-Aug-2011_ PDS_ ; it is 8m Fine Quad-Polarization and some characteristic of RADARSAT-2 sensor specifications is band wavelength 56 mm, frequency 5.4 GHz, the RADARSAT-2 image of the Canadian Space Agency (CSA). T h e SAR satellite 230

5 operates with different frequencies, namely L-band frequency, C- band frequency, X- band frequency, etc with various wavelengths. The RADARSAT-2 has C- band frequency Fine Quad [12] latitude and longitude ( ) values. These data sets (RADARSAT-2) accumulate and ensure the subset image using Next ESA SAR Toolbox (NEST) [11]. IV. Result In this experiment, the result shows that noise can be removed very effectively through EGM technique. The researcher experiments following techniques, namely Lee, Frost and EGM based on window size (3*3, 5*5 and 7*7) to identify the variations among existing techniques and the proposed technique. The experimental results are shown in Figure 3. From this result, one can observe that Lee technique cannot remove noise effectively because of some pixel intensity is almost low, like this, in Frost technique also cannot remove noise effectively because of pixel intensity is almost low like the Lee technique. But the proposed technique (Enhanced Gamma Map technique) can remove the noise more effectively; this can be neatly shown in the Table-3 (each segment σ0 and σ0 mean values are easy to recognize). (a) Before the noise removes (b) Lee (c) Frost (d) Enhance Gamma Map Figure 3: before the noise and after remove the noise using Lee, Frost and Gamma Map techniques Table 3 P P P S S i1 i2 i1 i- i sample 0 data 1 values 4with sigma 1 naught mean. 231

6 Through this technique sigma naught values can be identified very easily because the noise can be very effectively removed in all window size. From this identified values the evaluation of SAR image correlation can be compared between different polarizations. The compared polarization (VH_dB versus HH_dB and VH_dB versus HV_dB) values are visualizing the scatter plot and histogram that are shown in Figure 4. Figure 4: scatter plots and histograms for gamma filter The statistical value of histogram usage of an Enhance Gamma Map technique is very accurate for different polarizations (VH, VV, HH and HV). and these are shown in Table-4. Table 4. Statistical values of histogram usage of an Enhance Gamma Map technique Standard deviation: Coefficient of variation T h e experiment conducted in area of interest (AOI) or the region of Interest (ROI) reveals that the scatter plot and histogram values are more effective than the other existing techniques (outcome). Hence, EGM helps to classify the SAR images easily to identify the LULC facts. For this (LULC) classification purpose the K-Mean clustering method is applied and it brings 72% accuracy in the classified image which is better than other cluster technique like Expectation Maximization, which only gives 56% accuracy for classified classes. This can be shown in Figure 5. Figure 5: K-Mena clustering and classification 232

7 V. Conclusion This experimental research proves that the EGM technique is more effective for reducing the noise of the SAR images because it provides 72% ground truth values in SAR image classification, and also it helps to calculate the statistical values of the area of interest. The limitation of this proposed technique is that the result will be received only on the window size. Therefore, further research can be encouraged to find an improvement in window size. Acknowledgement I would like to thank my supervisor, the teaching and Nonteaching staff from B. S. Abdur Rahman Crescent University for their extraordinary support in this process. References [1]X. Deng and C. Lopez-Martinez, "Analysis of texture distributions of polarimetric SAR data," 2015 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Milan, 2015, pp [2] M. Arii et al., "Theoretical characterization of multi incidence angle and fully Polarimetric SAR data from rice paddies," 2016 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Beijing,2016, pp [3] B.Hou, X. Tang, L. Jiao and S. Wang, SAR image retrieval based on Gaussian Mixture Model classification," nd Asian-Pacific Conference on Synthetic Aperture Radar, Xian, Shanxi, 2009, pp [4]Li Yi-Bo, Zhou Chang and Wang Ning, "A survey on feature extraction of SAR Images," 2010 International Conference on Computer Application and System Modeling (ICCASM 2010), Taiyuan, 2010, pp. V1-312-V [5]D. K. Mahapatra, S. S. Ray and L. P. Roy, "Quantitative measurements on despeckling of SAR clutter amplitude : An experiment on MSTAR data," 2015 International Conference on Microwave, Optical and Communication Engineering (ICMOCE), Bhubaneswar, 2015, pp [6]S. Gupta, S. Kumar, A. Garg, D. Singh and N. S. Rajput, "Class wise optimal feature selection for land cover classification using SAR data," 2016 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Beijing, 2016, pp

8 [7] G. Di Martino, A. Di Simone, A. Iodice and D. Riccio, "Scattering- Based Nonlocal Means SAR Despeckling," in IEEE Transactions on Geoscience and Remote Sensing, vol. 54, no. 6, pp , June [8] J. Singh and M. Datcu, "Use of the second-kind statistics for VHR SAR image retrieval," th International Conference on Communications (COMM), Bucharest, 2012, pp [9] Bin Liu, Chenxian Zhu, Kaizhi Wang, Xingzhao Liu and Wenxian Yu, "A one-class- extraction framework for high resolution SAR image classification," Proceedings of 2011, IEEE CIE International Conference on Radar, Chengdu, 2011, pp [10] Li Xiao-Bing, Chen Yun-Hao, Zhang Yun-Xia, Li Xia and Gong Peng, "Detecting land cover characteristics along with changing scales based on remotely sensed data," IEEE International Geoscience and Remote Sensing Symposium, 2002, pp vol.6 [11] URLhttps://earth.esa.int/web/nest/home, Website Title Home NEST Next ESA SAR Toolbox, Article Title Home NEST Next ESA SAR Toolbox. [12] URLhttp://mdacorporation.com/, Website Title MDA, Article Title MDA 234

9 235

10 236