Denoising Cloud Interference on Landsat Satellite Image Using Discrete Haar Wavelet Transformation
Vol. 10 No. 1 January 2012 International Journal of Computer Science and Information Security Publication January 2012, Volume 10 No. 1 . Copyright � IJCSIS. This is an open access journal distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
(IJCSIS) International Journal of Computer Science and Information Security, Vol. 10, No. 1, January 2012 Denoising Cloud Interference on Landsat Satellite Image Using Discrete Haar Wavelet Transformation Candra Dewi Mega Satya Ciptaningrum Muh Arif Rahman Department of Mathematic Department of Mathematic Department of Mathematic University of Brawijaya University of Brawijaya University of Brawijaya Malang, Indonesia Malang, Indonesia Malang, Indonesia email@example.com firstname.lastname@example.org email@example.com Abstract—Satellite imagery is very useful in information that is performed by Choi and Bindschadler (2004), clouds is acquisition of the earth's surface image, especially the earth's very high reflected in the band 2 (0.52 - 0.60 µm). resources. However, in the process of retrieval information from satellite imagery is often found barriers that can obscure or even The elimination process of noise in the spatial domain can cover the imaging of an area. One of these barriers is a cloud, be applied directly on image pixels. One of the transformation which result the image that covered with lots of noise. Wavelet methods that can be done on the spatial domain is a power-law transformation was usually used to enhance the image or to transformation. While in the frequency domain, the image is eliminate striping noise on satellite image. In this paper is used broken into multiple kernels to be processed by the analysis of Discrete Haar Wavelet transformation to reduce cloud noise on transformation. Transformations that can be done in this Landsat TM image. The process includes the Haar Wavelet domain include Wavelet transformation   . decomposition of image rows and columns. After that, Transformation performed to obtain information and identify thresholding process is also applied for de-noising. Thresholding the original image, by getting its spectrum. Spectrum can be results are then reconstructed using the Inverse Discrete Haar obtained from the image frequency, time, or time-frequency Wavelet. The method is applied to the variation of the band depend on the type of transformation used . image, the type of thresholding (hard and soft), as well as the size of the image convolution. The testing results on the band 1 to It is well known that wavelet transform is a signal band 6 of Landsat TM imagery showed that the lowest error processing technique which can display the signals on in both values are calculated by RMSE (Root Mean Square Error) time and frequency domain. Wavelet transform is superior present in band 1. Image signal to noise ratio in band 1 has the approach to other time-frequency analysis tools because its highest value, which means most high-power image signal to time scale width of the window can be stretched to match the noise. This mean that band 1 has the highest pixel value original signal, especially in image processing studies. similarity between whole testing data. Wavelet transformation can be used to obtain signal both in Keywords; Discrete Haar Wavelet, thresholding, image the frequency domain and time domain. Wavelet time scale convolution, Landsat TM width of the window can be stretched to match the original signal. Wavelet is a conversion function that can be used to break up a function or a signal into different frequency components. These components then can be processed in I. INTRODUCTION accordance with the scale. While the wave is a function of Image of the earth surface recording can be interpreted by moving up and down in space and time periodically the user for the benefit of various fields. In the process of (sinusoidal), wavelet is a limited wave or sometimes is called image acquisition by the satellite, sometimes is found noise that as short wave . can reduce the image quality. This disorder is caused by the Haar transform uses the Haar scaling function and the Haar presence of such clouds or fog that can obscure or even wavelet function. Haar wavelet transformation use the Haar covered the satellite during the imaging process . This noise basis functions that is called a wavelet orthonormal . Haar can interfere the interpretation process therefore the results Wavelet functions can be expressed in matrix form. obtained will not be maximal. In the previous study, wavelet transform is used to sharpen Each pixel in the satellite image has some digital value the cloud-related shadow areas . Beaulieu et al (2003) refine (numeric) in accordance with the band of satellite imagery. For the resolution of a multi-spectral (MS) image using fusion example is Landsat TM image that has 7 bands. Therefore, method and the Stationary Wavelet Transform. In the study each pixel has 7 digital values that are suited to 7 band digital performed by Torres and Infante (2001), wavelet transform is value that is owned. The different characteristic each bands used for denoising stripping noise on satellite imagery. This causes the difference in the ability to detect clouds. In the study 27 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol. 10, No. 1, January 2012 paper applies the Haar wavelet transformation to reduce the III. RESEARCH METHOD noise cloud on Landsat TM imagery. This application was built to reduce noise on Landsat TM satellite image using Haar wavelet transformation method. The II. PREVIOUS RESEARCH limitation of this system includes: The research about the using of wavelet transformation has 1) The image used is a grayscale image of type TIFF been done by some researcher. Torres and Infante (2001) 2) The size of the image used is 256x256 orthonormal present new destriping technique on satellite imagery using Daubechies wavelets of different orders and was tested on a heavily striped Landsat MSS image. Visual inspection and The flowchart of noise reduction process is shown Figure 1. measurement the signal-to-noise ratio showed that the method The inputs of this application consist of satellite imagery with proved produce encouraging results in image quality and clouds noise and image without noise. This input image is performance, overcome some problems commonly found on presented in grayscale values. Some preprocessing was done to traditional destriping techniques and reduce computer time the noise image to reduce the noise. The image without noise is process and storage space. used as a comparison in the testing process. Beaulieu et al, (2003) refine the resolution of a multi- spectral (MS) image by fusion method using a high-resolution panchromatic (PAN) image and the Stationary Wavelet Transform (SWT). They propose to produce high-resolution MS image that has nearly the same statistical properties than the original multi-spectral image with no blocking image artifacts. These algorithms are based on the injection of high- frequency components from the PAN image into the MS image. They prove that pixel-level fusion was a powerful method to refine the spatial resolution of PAN images. Wang et al (2003) present a new approach to eliminate the random image noises inherent in the microarray image processing procedure using stationary wavelet transform (SWT) and applied on analysis of gene expression. The testing result on sample microarray images has shown an enhanced image quality. The results also show that it has a superior performance than conventional discrete wavelet transform and widely used adaptive Wiener filter in this procedure. Elrahman and Elhabiby (2008) developed image sharpening algorithm using wavelet to enhance shadow areas of cloud and tested this algorithm on the panchromatic band of Landsat 7 Figure 1 Flowchart of noise reduction process ETM satellite sub-scenes. The algorithm is applied locally by boosting the image high frequency content in the shadow areas using the defected image de-noised wavelet coefficients. By Firstly, noise image is transformed into the frequency using visual and quantitative analysis was found that the ability domain using the Haar wavelet transform. The quantization to enhance details under shadow areas increased with the process is then performed using a specific threshold value. The increase in the number of wavelet decomposition levels. transformation process is performed to the n level, where N = Beside, were found that enhancing image quality in the shadow 2n and N is the size of the image. At each level, the row areas could be done using only two or three wavelet transformation is done in advance through highpass and decomposition levels. lowpass filters. After that, is done transformation of the column. In these previous studies, the using of wavelets on the satellite image is to sharpen the image and to improve image The next process is tresholding. This process separates resolution. Wang et al (2003) already used wavelet to eliminate pixels based on the degree of grey level values. The wavelet the noise, but is applied to the gene sequence image. In this coefficients which are below the threshold are set to zero and paper will be applied discrete Haar wavelet to reduce noise in than take the other values for purposes of reconstruction of the the form of clouds on satellite images. Although the discrete signal. Threshold used is Hard and Soft Threshold. With ε is wavelet transform has a lower performance of the stationary the threshold value, hard thresh equation is shown in (1). wavelet transform, but its ability to reduce the noise is quite high and does not vary with stationary wavelet transform . x, | x |> ε Thard ( X ) = (1) 0 | x |≤ ε On the hard threshold, all wavelet coefficients with a value below a specified threshold are classified as noise and removed (are set to zero). While the coefficients above the 28 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol. 10, No. 1, January 2012 threshold is classified as signal. In soft thresholding, the B. Peak-to-Signal Noise Ratio (PNSR) wavelet coefficients with a value below the specified threshold PSNR is the comparison between the maximum possible are removed and the wavelet coefficients above the specified signal strength of a digital signal with the power of noise that threshold are reduced by the threshold value. Thus, this method affects on the signal (Alfatwa, 2005). PSNR is defined through reduces the range of wavelet coefficients and signal leveling. the signal-to-noise ratio (SNR) to measure the level of signal Soft threshold was chosen because this procedure does not cause non-continuants at x = ± ε. The equation for the soft quality. Signal quality is directly proportional to the value of threshold is shown in (2). SNR. The larger of the SNR value, the better the quality of the generated signal. PSNR values usually range between 20 and sign( x)(| x | −thresh), | x |≥ ε 40 dB (Alfatwa, 2005). PSNR values can be calculated using Tsoft ( X ) = (2) 0, | x |< ε equation as shown in (6). Value of 255 represents the upper limit value of image pixels. For the determination of threshold values is used equation 255 as in (3). PSNR = 20 log10 (6) RMSE 2σ 2 log(n) t= (3) n V. SOURCE OF DATA Where: t = threshold value that is calculated Image that is used in the testing process is Landsat TM = the variance of data satellite image with each channel has a different sensitivity to n = number of data the wavelength. Landsat TM satellite orbital period for taking pictures of the earth's surface is generally performed at least 6 months. Satellite imagery from two period of taken picture can The equation of variance is shown in (4). be used as a reference on the interpretation process. For example, this study used two images with the same object (the σ2 = ∑ (x i − x) 2 (4) island of Madura) taken in June 2004 and February 2005. In (n − 1) the image taken on 2005 exists cloud covering the particular object and the image taken on 2004 (with the same object) is The last process is the Inverse Haar Wavelet Transform used as reference. (IHWT) which is the process of passing the image through the inverse filter matrix transformations. This process is contrary to Preparation of satellite imagery should be done to obtain the decomposition process. the image that is suited to analysis. The original image is cropped to the size of 256 x 256 pixels and converted into Tif IV. TESTING METHOD extention format. In addition, the original image with 7 bands is separated per-band for used in applications. Details of the data used are as follows: For testing the result is used Root Mean Square Error (RMSE) and Peak-to-Signal Noise Ratio (PNSR). 1) Landsat image of Madura island, dated June 25, 2004 and dated February 4, 2005 A. Root Mean Square Error (RMSE) 2) Landsat image of Java island, dated June 25, 2004 and RMSE is one of the ways to measure the amount of the dated February 4, 2005 difference between the estimated values with actual values by measuring the average of error. RMSE is calculated by Of the two sources of image data was made 2 pieces of comparing the number of errors between the denoising image testing data with each of the data contained six band image data and the original image. The lower the RMSE value the smaller (bands 1 to 6 / 7) with each size is 256 x 256 pixels. the error calculation has been done. RMSE of digital image Data I: latitude 7:7:45.66 S and longitude 113:3:12. 43 E with size NxM could be calculated using equation as shown in (5). Data II: latitude 7:39:41.99 S and longitude 112:56:41.87 E RMSE = ∑ [ f (i, j ) − F (i, j )] 2 (5) VI. RESULT AND DISCUSSION N2 Where: Some examples of images resulted from denoising process are visually displayed in Table I. The first image shows the f(i,j) is pixel value in original image result of denoising on data I (band 1) with convolution 2 (hard F(i,j) is the pixel value on reconstruction image thresholding), the second on data II (band 1) with convolution N2 is an image size (in pixels) 8 (soft thresholding), and the third on data II (band 3) with convolution 8 (hard thresholding). 29 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol. 10, No. 1, January 2012 TABLE I. SAMPLE IMAGE OF TESTING RESULT TABLE III. THE CALCULATION RESULT OF THE RMSE AND PNSR AT CONVOLUTION 8 (DATA II) No Band Input Images Output Images Thresh RMSE PNSR Band Hard Soft Hard Soft Value Threshold Threshold Threshold Threshold 1 1 2,96 7,422 7,041 30,72 31,178 1 (data I) 2 2,46 10,339 9,863 27,841 28,251 3 2,06 19,748 19,069 22,22 22,282 4 2,92 43,214 42,623 15,418 15,538 5 2,69 25,312 24,773 20,064 20,251 1 6/7 1,61 22,551 22,421 21,057 21,118 2 (data II) TABLE IV. THE CALCULATION RESULT OF THE RMSE AND PNSR AT CONVOLUTION 4 (DATA I) RMSE PNSR Thresh Band Hard Soft Hard Soft Value Threshold Threshold Threshold Threshold 3 3 1 3,27 31,306 30,463 18,274 18,455 (data II) 2 2,86 32,353 31,691 17,932 18,112 3 2,86 39,3 38,646 15,243 16,389 4 2,72 43,3 42,702 15,401 15,522 5 2,71 48,174 47,818 14,475 14,539 The RMSE was calculated in the image (the data I and II) which has been transformed with Haar Wavelet. This RMSE 6/7 2,08 46,025 45,744 14,871 14,924 values are calculated against several variations of testing which includes testing of inter-thresholding methods, inter-level TABLE V. THE CALCULATION RESULT OF THE RMSE AND PNSR AT convolution, and inter-band image. Furthermore, the RMSE is CONVOLUTION 4 (DATA II) used as input to the calculation of PNSR to observe the ratio of RMSE PNSR signal strength to noise. Band Thresh Value Hard Soft Hard Soft Based on the results in Table 1 could be known that Threshold Threshold Threshold Threshold visually processes of Haar wavelet denoising did not show the 1 2,93 31,988 31,1 18,031 18,276 significant results, because the cloud noise in each band is 2 2,57 33,509 32,773 17,628 17,82 represented differently. Therefore, an analysis on the basis of testing results on PNSR and RMSE are performed. 3 2,57 40,198 39,393 16,047 16,222 The comparison results of RMSE and PNSR on bands 1 to 4 2,45 43,854 43,025 15,291 15,456 7 with a convolution of 8, 4, and 2 are shown in Table 2 to 5 2,45 49,071 48,754 14,314 14,371 Table 7. The RMSE and PNSR are obtained can be used to find out the best band on Landsat satellite imagery for the cloud 6/7 1,89 46,746 46,469 14,736 14,788 denoising process. TABLE VI. THE CALCULATION RESULT OF THE RMSE AND PNSR AT CONVOLUTION 2 (DATA I) TABLE II. STHE CALCULATION RESULT OF THE RMSE AND PNSR AT CONVOLUTION 8 (DATA I) RMSE PNSR Thresh Band Hard Soft Hard Soft RMSE PNSR Value Thresh Threshold Threshold Threshold Threshold Band Hard Soft Hard Soft Value 1 2,89 7,647 7,362 30,461 30,791 Threshold Threshold Threshold Threshold 1 3,34 30,128 29,633 18,551 18,695 2 2,4 10,745 10,262 27,507 27,906 2 2,93 30,98 30,356 18,309 18,486 3 2,01 20,123 20,141 22,057 22,049 3 2,93 38,246 37,681 16,479 16,608 4 2,86 43,812 43,018 15,299 15,458 4 2,78 42,456 41,998 15,572 15,666 5 2,62 26,025 25,486 19,823 20,005 5 2,77 46,93 46,548 14,702 14,773 6/7 1,57 22,894 22,873 20,936 20,944 6/7 2,12 44,859 44,569 15,094 15,15 30 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol. 10, No. 1, January 2012 TABLE VII. THE CALCULATION RESULT OF THE RMSE AND PNSR AT VII. SUMMARY AND CONCLUDING REMARKS CONVOLUTION 2 (DATA II) RMSE PNSR Thresh In this paper, Discrete Haar Wavelet methods is applied by Band Hard Soft Hard Soft Value Threshold Threshold Threshold Threshold utilize thresholding method to the data testing (Landsat satellite 1 2,59 7,927 7,835 30,148 30,25 image with size 256x256 pixels) to reduce the noise contained in image. The testing result shows that the lowest RMSE value 2 2,15 11,139 10,667 27,194 27,57 is detected on band 1 (29.633) and highest value is on the band 3 1,8 30,564 20,864 21,869 21,75 5 (49.071). As well as the highest PNSR value observe on the band I (18.695 dB) and lowest value is on band 5 (14.314 dB). 4 2,56 44,428 43,333 15,178 15,394 It can be concluded that the best band to perform denoising 5 2,35 26,787 26,279 19,572 19,739 clouds with Haar Discrete Wavelet found on the band I, and worst band found on the band 5. 6/7 1,41 23,313 23,478 20,779 20,178 For further study, is proposed to test the result referable reinforced with a system of classification on Landsat satellite Limitations of different threshold values applied to each imagery. image because the distribution of each image pixels values is different. This research calculates the threshold based on the characteristics of image to obtain the best threshold value. REFERENCES From Table 2, Table 4 and Table 6 can be seen that the band 1 has the smallest RMSE values both in hard thresholding and soft thresholding method (both in the convolution 8, 4, and  A. Abd-Elrahman and M. Elhabiby, “Wavelet Enhancement of Cloud- 2). The lowest RMSE values observed in the convolution 8 Related Shadow Areas in Single Landsat Satellite Imagery”, Beijing: The International Archives of the Photogrammetry, Remote Sensing, and with soft thresholding, which is about 29.633 (Data I) and Spatial Information Science, Vol. XXXVII part B7, p.1247-1252, 2008. 7.041 (data II). Since the highest RMSE value is detected on  H. Choi dan R. Bindschadler, “Cloud Detection in Landsat Imagery of band 5 (using hard thresholding with convolution 2) that is Ice Sheets Using Shadow Matching Technique and Automatic about 49.071 (Data I) and 26.787 (Data II). The quite far Normalized difference Snow Index Threshold Value Decision”, Remote differences of RMSE value is caused by variations in image Sensing of Environment, Vol. 91. p.237-242, 2004. value. Data I is an image with a lot of noise distribution, while  J. Torres and S.O. Infante, “Wavelet Analysis for The Elimination of the data II has less noise in the form of clouds. Base on RMSE Striping Noise In Satellite Images”, Society of Photo-Optical Instrumentation Engineers, DOI: 10.1117/1.1383996, 2001. value can be seen that band 2 has the lowest error values and  M. Beaulieu, M., S. Faucher, and L. Gagnon, « Multi-Spectral Image band 5 has the highest error value. Resolution Refinement Using Stationary Wavelet Transform”, The highest value of PNSR is observed on the band I that is International Geoscience Remote Sensing Symposium, Vol. 6, pp. 4032– 4034, 2003. around 18.695 dB (data I) and 31.178 dB (data II), while the  X.H. Wang, Robert S.H. Istepanian, and Y.H. Song, “Microarray Image lowest value is observed in the band 5 with the value is 14.314 Enhancement By Denoising Using Stationary Wavelet Transform”, dB (data I) and 19.572 dB (data II). It means that the ratio of IEEE Transactions on Nanobioscience, Vol. 2, No.4, 2003. the image signal to noise at a band I higher than the band 5.  D. F. Alfatwa, “Watermarking pada Citra Digital Menggunakan Discrete The Signal strength value at Data II tends to be higher than the Wavelet Transform”, Informatic Study Program, Technoly Institute of data I, because the noise in the form of clouds fewer than on Bandung, 2005. the Data I. It denoted that the highest probability to perform  R. B. Edy Wibowo, “Scattering Problem for A System of Non Linear denoising of cloud can be done on the band 1. On the contrary Klein-Gordon Equations Related to Dirac-Klein-Gordon Equations”, An International Multidisciplinary Journal, Vol. 71, No. 3-4, 2009. the lowest probability is on band 5. These results are quite relevant to the characteristic of a band I with a wavelength of  Gonzales, Rafael C dan Woods, Richard E. 2005. Digital Image Processing, 2nd Edition. New Jersey: Prentice Hall 0.45 to 0.52 µm which serves to increase penetration on water body and humidity. 31 http://sites.google.com/site/ijcsis/ ISSN 1947-5500