Precise error correction method for NOAA AVHRR image using the same orbital images

Transcription

1 Precise error correction method for NOAA AVHRR image using the same orbital images 127 Precise error correction method for NOAA AVHRR image using the same orbital images An Ngoc Van 1 and Yoshimitsu Aoki 2, Non-members ABSTRACT NOAA images provide very useful information about the Earth and have been widely using. This paper proposes a method that precisely corrects the errors in NOAA images. NOAA images with the same orbit are received at various stations and they overlap each other. The errors in the original image, including missing lines and error lines, are corrected by using reference images, which are received at other stations and overlap the original one. An error information database is used to select the highest quality reference images. After specifying the overlapped area, missing lines are detected by checking time codes, error lines are recognized based on PN codes. Missing lines and error lines are corrected by using the values of the corresponding lines from reference images. As a result, missing data can be fully restored and errors are precisely corrected. This method was used to correct the errors in the NOAA images receiving at Tokyo, Bangkok and Ulaanbaatar. The correction results proved its high precision. The correction time was less than 37 seconds per image. Keywords: NOAA AVHRR, error correction, overlapped area, reference image 1. INTRODUCTION In recent years, AVHRR (Advanced Very High Resolution Radiometer) on the NOAA (National Oceanic and Atmospheric Administration) series of satellites has been an ideal observatory for daily global observation of the Earth. NOAA images provide very useful information about ecosystems, climate, weather and water from all over the world. NOAA images are also widely used in land cover monitoring at global and continental scales [1]. Because of many reasons which occur in scanning, sampling, transmission or recording processes, errors appear in NOAA images. In order to use NOAA images effectively, error correction method is necessary. Manuscript received on March 1, 2007 ; revised on July 3, The author is with the Functional Control Systems Course, Shibaura Institute of Technology, Japan, s: vanngocan@aoki-medialab.org 2 The author is with the Dep. of Information Science and Engineering, Shibaura Institute of Technology, Japan, yaoki@sic.shibaura-it.ac.jp Some methods to correct the errors in NOAA images were proposed. In the conventional correction methods [2], errors are corrected by using only the relations between the lines and pixels in the original image; therefore, a lot of errors could not be recognized and the accuracy of the results was not very high. Recently, an effective method which corrects errors by referring to the reference images that overlap the original one were introduced in [3,4]. According to this method, the overlapped areas between the original image and the reference one are specified by matching the time codes of the lines in both images, and errors are corrected by using data in this overlapped area. Compared to other correction methods that also use overlapped images [5,6], whereas it is quite difficult and complicated to specify the overlapped areas in those methods due to the processes of extracting, comparing or matching the features and templates, it is easy but precise to locate the overlapped areas with the method of [3,4]. However, when an original image has many overlapped images, this method could not select the most suitable reference images to optimize the correction. For this reason, bad quality reference images might be selected. In this case, the overlapped area in the reference images will contain many errors, and not all of the errors in the original images will be removed. This paper improves the method of [3,4] by assessing the quality of the reference images and selecting the best ones to use. First, an error information database is created by collecting the error information in NOAA images. Then, based on the database, a method is proposed in order to find the highest quality reference images among available ones. Hence, the best quality reference images can be selected, and the number of errors in the overlapped area of reference images will be minimized. As a result, almost all of the errors in the original image will be corrected. 2. HRPT FORMAT NOAA images contain lines, and each line consists of pixels. They are received in HRPT (High Resolution Picture Transmission) format. This format includes time code, PN code and AVHRR data Time code AVHRR sensor scans 6 lines per second. Time code is the time when the sensor starts to scan a line. It is recorded for each line, and it includes month,

2 128 ECTI TRANSACTIONS ON ELECTRICAL ENG., ELECTRONICS, AND COMMUNICATIONS VOL.5, NO.2 August 2007 day, hour, minute, second and millisecond. The unit of time code is millisecond. 2.2 PN code PN code is a pseudo-random sequence provided by a feedback shift register. There are three PN codes in HRPT format. The first PN code is 60 bits long and it can be used to specify the beginning data of a line. The second PN code includes 1270 bits and the third one includes 1000 bits. The total length of all PN codes is 2330 bits. These bits have pre-fixed values for each line of a NOAA image. When a line contains error bits, its PN codes will differ from the fixed value. As a result, PN codes can be used to detect the error lines. The error rate of each line is the ratio of PN code to A line will contain error when its error rate is greater than AVHRR data The AVHRR data that sensor obtains from the Earth is stored in the lines of image. Each line includes five data channels. Channels 1 and 2 are used to monitor land surface processes, whereas channels 3, 4 and 5 are used for sea surface temperature determination and cloud mapping [1]. Every channel in a line contains 2048 pixels. A pixel is coded in 10 bits. 3. Errors in NOAA image can be due to the errors in the scanning or sampling equipment, in the transmission or recording of data, or in the reproduction of the media containing the data [1]. In this study, all errors in NOAA images are divided into missing lines and error lines Missing lines Normally, the time code of each line in NOAA images is recorded. However, when the sensor cannot get the data of a line, its time code will not be recorded, and it is considered as a missing line. Because the sensor scans 6 lines per second, a missing line can be detected by checking the time codes of other lines. The missing time code can also be inferred Error lines A line in NOAA image might contain error pixels. When a line contains one or more error pixels, it is an error line. The error line s PN codes will differ from its fixed PN codes. Therefore, an error line can be detected by checking its PN codes. 3.3 Error areas in NOAA images The statistical results of [2] have shown that errors appear only at the top and bottom parts of NOAA images. The reason is the receiving systems of NOAA data. At the time when the data of the top or bottom part in NOAA images are being received, the distance from satellite to receiving station is long; hence, errors will occur. 4. CURRENT CORRECTION METHODS 4. 1 Conventional method As was shown in [2], this method corrects errors by using only the relation of the lines and pixels in the original image. First, missing lines are detected by finding the missing time codes in the image. A blank line is then inserted into image to replace the missing line. For this reason, all missing data is lost. Then, error pixels are recognized by comparing their values with their neighbors. According to the statistical results [2], the difference in the values between an error pixel and its neighbors is about 512, 256, 128, 64 or 32. Once being detected, an error pixel is corrected by adding or subtracting this different value. Therefore, an error pixel is probably assigned a wrong value. Furthermore, conventional method also could not detect all the error pixels in the original images. Table 1 shows the number of error pixels detected by this method of the image AH from Tokyo, and the actual number of error pixels detected by comparing this image with the overlapped image AH from Bangkok. The result of this table is calculated on 107 error lines in the overlapped area of the image from Tokyo. The corresponding lines in the image from Bangkok are correct lines (they are neither missing lines nor error lines). Clearly, the number of error pixels detected by this method is much smaller than the actual number (12978/36017). Table 1: Error pixels in 107 lines Errors All errors Detected Actual Method using reference images Another correction method using reference images was introduced in [3,4]. NOAA images with the same orbit are received at various stations and they overlap each other. The data in the overlapped area of the reference images can be used to correct the errors in the corresponding area of the original image. First, the reference image which overlaps the original one is selected. Next, the overlapped area in both images is specified by matching the time codes. Then, missing lines are detected by checking the time codes of the original image. They are corrected by assigning the values of the corresponding lines from the reference image. If the corresponding line is also a

3 Precise error correction method for NOAA AVHRR image using the same orbital images 129 missing line, a blank line is inserted to the original image. Final, the error lines, whose error rates are greater than zero, are detected by checking the PN codes. They are corrected by assigning the values of the corresponding lines in the reference image. If the corresponding lines are also error lines, the error pixels in the error lines of the original image are corrected by using the conventional method. With this method, the missing lines are restored only when their corresponding lines are correct lines. Similarly, the error lines are precisely corrected only when the corresponding lines are not missing lines and their error rates are zero. In the case that the corresponding lines are missing lines or error lines, errors are corrected by using the method of [2]. Thus, the missing data in the missing lines will be lost, the error pixels might be assigned wrong values and many error pixels still cannot be recognized. Furthermore, when an original image has many reference images, this method could not select the highest quality reference images. The bad quality image with a lot of errors in the overlapped area might be used; hence, not all errors in the overlapped area of the original image could be corrected. the first line. For example, the NOAA image AH was scanned at the time 19:10:54 on January, 25th, Every NOAA image has less than 6500 lines. Because the sensor scans 6 lines per second, the total time to scan a whole NOAA image is less than 1100 (6500/6) seconds. Therefore, two images with the same orbital data may overlap each other if the absolute value of the difference in the time extracting from their names is less than 1100 seconds. The time difference between the original and reference image is an algebraic value. It will be a positive number when the original image is received at later time than the reference one is and vice versa. Figure 2 shows an example of which the image AH from Tokyo overlapped and image AH from Bangkok. The image from Tokyo is received at 05:43:08 and the one from Bangkok is at 05:39:03 on June, 17th, The time difference between them is +245 seconds. 5. PROPOSED METHOD In order to improve the method using reference images, a new method is proposed. Figure 1 is the steps of new method. Fig.2: Overlapped images Fig.1: Steps of proposed method NOAA images have been receiving at many stations. Therefore, an original image may have many reference images. The new method will find all available reference images from receiving stations. Their quality will be assessed based on the error information database. The reference image with highest quality will be selected first and the overlapped area will be specified. Data in this area will be used to correct errors. If there are still errors in the overlapped area of the original image after correction, reference images with lower quality will be used Finding available reference images A NOAA image is saved as a file whose name includes the time when the sensor started to scan 5. 2 Error information database After available reference images are found, they need to be analyzed to detect errors and specify the overlapped areas before use. If the analyzing is applied to all reference images, it will take a long time compared to the total correction time. Moreover, in many cases, only one or two reference images are enough to correct all errors in the original image and it is not necessary to analyze the rest ones. For this reason, statistics are carried out with NOAA images. The error information in NOAA images is collected, and it will be used to assess the quality of the reference image. Furthermore, after collecting error information, the results will be used not only for the images belonging to the statistical process but also for other images. As a result, the information to assess reference images will be available without spending a

4 130 ECTI TRANSACTIONS ON ELECTRICAL ENG., ELECTRONICS, AND COMMUNICATIONS VOL.5, NO.2 August 2007 long time to analyze all of them. Figure 3 is a NOAA image with errors at the top and bottom parts. As was shown in [2], errors in NOAA images appear at the top and the bottom parts. In this image, the Error Top and Error Bottom area are the error area at the top and bottom part, respectively. In the direction from top to bottom, the final line of Error Top area and the first line of Error Bottom area are missing or error lines. The rest area is correct (do not contains any missing or error line). When the NOAA image in figure 3 is used as reference image, the overlapped area should locate within the correct area to minimize the number of errors. For this reason, the total lines, the number of lines in the Error Top and Error Bottom area are collected when doing statistics. The information of each image using in statistical process are stored in the error information database. Therefore, if a reference image belongs to the database, its information will be available immediately. In addition, to improve the correction result for the images which do not belongs to the database, the information will be calculated from the database with a cumulative parameter. For example, when a NOAA image from Bangkok which does not belong to the database is used as a reference image, and the cumulative parameter is 90%, the information getting from the database will be ET=669, EB = 836, T L = 4723, which means that the 90% images from Bangkok in the database have ET <= 669 lines, EB <= 836 lines and T L <= 4723 lines. The cumulative parameter is necessary because if only the maximum value (the cumulative parameter is 100%) in the database is used, the overlapped area will be too small (contains not all errors in the original image); similarly, if only the minimum value (the cumulative parameter is approximately 0%) is used, the overlapped are will be too large (contains a lot of errors in the reference image); and if the average value is used, the overlapped in the original image may still contain errors with high probability. Table 3: Statistical result with images from Tokyo Fig.3: Statistical information The statistics are carried out with the NOAA images receiving at Tokyo (Japan), Bangkok (Thailand) and Ulaanbaatar (Mongolia), from June to October in In each month, about 90% number of images is used for statistics and the rest 10% is for testing. Table 2 is the number of NOAA images using for statistics (upper row for each station) and testing (lower row for each station). Table 2: Error pixels in 107 lines Station Jun Jul Aug Sep Oct Total Tokyo Bangkok Ulaanbaatar Information Jun Jul Aug Sep Oct Ave. Max ET Min Ave Max EB Min Ave Max T L Min Ave Table 4: Statistical result with images from Bangkok Information Jun Jul Aug Sep Oct Ave. Max ET Min Ave Max EB Min Ave Max T L Min Ave The statistical results for NOAA images from Tokyo, Bangkok and Ulaanbaatar are shown in Table 3, 4 and 5. In these tables and the figures below, ET and EB is number of the lines in the Error Top and Error Bottom area of the image, respectively; T L is total lines in the image. The information of the Ave. columns is the average values calculating from the information in the tables 3, 4, 5 and the number of images in the table Assessing and selecting reference images An original image may have some reference images, but not all of them are always used to correct errors. Therefore, the available reference images should be assessed so that the most suitable one will be used first, and then, if it is necessary, the less suitable ones will be taken into account. Since the data in

5 Precise error correction method for NOAA AVHRR image using the same orbital images 131 Statistical result with images from Ulaan- Table 5: baatar Information Jun Jul Aug Sep Oct Ave. Max ET Min Ave Max EB Min Ave Max T L Min Ave the overlapped area of reference image will be used to correct errors, available reference images will be assessed based on the quality of this area. The criteria to assess are: the better quality a reference image is, the less error appears and the more lines it has in the overlapped area. The statistical error information from the database is used to assess the quality. Suppose that the original image O have n reference images R i, i = 1..n. The algebraic time difference between O and R i is d i, and the number of correct lines in the overlapped area between O and R i is l i. Figure 4 is an original image and its i th reference image. In this case, the original image is received later than the reference one; thus, d i > 0 and the overlapped area in the reference image is used to correct the errors in the top part of the original one. In order to correct as many errors in the top part of the original image as possible, the error areas in the reference image should not be located in the overlapped area. Therefore, d i should be in the range of d imin and d imax, which are calculated as follow: d imin d i d imax d imin = ET i d imin = T L i (ET o + EB i ) (1) The number of correct lines in the overlapped area is: l i = T L i (d i + ET o + EB i ) (2) In the case of figure 5, the reference image is received later than the original one; therefore, d i < 0, and the overlapped area is used to correct the errors in the bottom part of the original image. d i should be in the range of d imin and d imax, which are calculated as follow: d imin d i d imax d imin = T L o (EB o + ET i ) d imin = 0 if T L o (T L i EB i ) d imin = T L o (T L i + EB i ) othwise (3) The number of correct lines in the overlapped area is: l i = T L o (d i + EB o + ET i ) (4) Fig.4: Original image is earlier than reference im- Fig.5: age Original image is later than reference image Those images R i whose d i > 0 and d i satisfies (1) will be sorted by l i calculating by (2). They can be used as reference images in the descending order of l i to correct the errors in the top part of original image O. Similarly, those images R i whose d i < 0 and d i satisfies (3) will be also sorted by l i calculating by (4). They can be used as reference images in the descending order of li to correct the errors in the bottom part of the original image O Correcting errors After selecting reference images, the data in the overlapped area of the reference image will be used to correct the errors in the original image. The correction steps are shown in figure 6. First, both original and reference images are analyzed. The time codes will be checked to detect the missing lines. Because the errors appear at the top and bottom parts of the image [2], the time codes of the lines in the middle of image are accurate. Moreover, sensor scans 6 lines per second; consequently,

6 132 ECTI TRANSACTIONS ON ELECTRICAL ENG., ELECTRONICS, AND COMMUNICATIONS VOL.5, NO.2 August 2007 Fig.6: Correction steps based on the time code of a correct line in the middle of the image, the accurate time codes for the other lines are inferred and saved into a time code table. If a value in the time code table cannot be found in the image, a missing line is detected, and a blank line will be inserted to the appropriate position. Next, to find out the overlapped area, the time codes of the lines in the original and reference images are matched. Two lines in two images will contain the same data if they have the same time code. Consequently, each line in the overlapped area of the original image has a corresponding line in the reference images and they have same time code. Once the overlapped area is specified, the errors in this area of the original image will be corrected. In order to correct the errors in the top part of the original image, the reference image with d i > 0 are used; similarly, the reference image with d i < 0 are used to correct the errors in the bottom part of the original image. When a missing line or an error line in the original image is detected, if the corresponding line in the reference image is correct line (not a missing line or error line), its value will be assigned to the missing line or the error line. As a result, the missing data can be restored and the error lines will be corrected. In the case that the corresponding line is also a missing or error line, the second reference image with smaller l i will be analyzed and used. This process will continue until all errors in the original image are corrected. 6. TESTING AND EVALUATION The proposed method is applied to correct the errors in the images receiving at Tokyo, Bangkok and Ulaanbaatar. Four tests are implemented with sample images. In the first test, the sample images are selected from the images belonging to the database. In the second test, the sample images is selected from the images not belonging to the database but were received from June to October in In the third test, the sample images are selected from the image not receiving in Table 6, 7 and 8 are the results of these tests. In the final test, all types of sample images (belong to the database or not, were received from June to October in 2002 or not) are used and the value of cumulative parameter is changed. Table 9 is the result of this test. In each test, 20 sample images are used and the numbers of errors (including missing lines and error lines) in the original image, before and after correction, are recorded. The cumulative parameter is 95%. The tests are implemented on the Sun Ultra 45 Workstation with 1.6 GHz Sun UltraSPARC IIIi processor and 1GB RAM. The results show that the number of errors remaining after correction is quite small. All most all of the errors are corrected. In some cases, there is still one error line in the corrected image. The reason is that, in these cases, there is only one reference image is available to correct the errors in the top or bottom part of the original image;therefore, if the error appears in the overlapped area of the reference image, the program cannot find any other reference image to replace, and it will try to correct as many error lines as possible. Another reason is that the error occurs at the acquisition process; as a result, all available references images will contain the same error line as the original image. Table 6: Correction result of the first test Image s Before correction After correction Time Index ET EB ET EB (sec) Total Ave. Errors remaining 0.05% 0.06% 27.35

7 Precise error correction method for NOAA AVHRR image using the same orbital images 133 Table 7: Correction result of the second test Table 9: Cumulative parameter evaluation Cumulative Error remaining (%) Parameter (%) ET EB Table 8: Correction result of the third test The percent of errors remaining after correction is bigger when the sample images are not in the database. The reason is: when sample images belong to database, the information to assess reference images is real information, which is calculated by analyzing them; when the sample images do not belong to database, the information to assess reference images is statistical information, which may differ from real information. Thus, the reference images will be assessed more accurately when they belong to the database, and the number errors remaining will be smaller. The results of table 7 and 8 are not so different. It proves that this correction method can be effectively applied to the images received at other months or years. In the first test, the information about the original and reference images is read directly from the database;therefore, the correction time is smaller than the rest testes. In general, the correction time is less than 37 seconds. Compared to the size of a NOAA image, which is about 100MB per image, this is an acceptable correction time. Table 9 shows that when the cumulative parameter is around 95%, the percent of errors remaining is the smallest. As mentioned before, when the cumulative parameter is 100%, the overlapped area will be too small, it will not contain all errors in the original image; when the cumulative parameter is smaller (such as 70, 60 or 50) the overlapped area will be larger, it will contain more errors in the reference image Figure 7, 8 and 9 are examples of the first, second and third test. In figure 7, the original image from Ulaanbaatar has two reference images from Bangkok and Tokyo; both of them can be used to correct the errors in the bottom part of the original one. Because these images belong to the database, their error information has been exactly known. Whereas the overlapped area of the image from Bangkok is small and its Error Top area overlaps the Error Bottom area of the original image, the overlapped area of the image from Tokyo is large and its Error Top area does not overlap the Error Bottom area of the original image. In the previous methods, the reference image from Bangkok might be selected and not all the errors in the original image are corrected. With proposed method, in this case, the image from Tokyo is selected as the most suitable reference image. Therefore, all the errors in the bottom part of the original image are corrected. In figure 8, the original image from Bangkok has also two reference images from Tokyo and Ulaanbaatar; both of them can be used to correct the errors in the top part of the original one. The statistical error information is used because these images do not

8 134 ECTI TRANSACTIONS ON ELECTRICAL ENG., ELECTRONICS, AND COMMUNICATIONS VOL.5, NO.2 August 2007 belong to the database. In the previous methods, the image from Ulaanbaatar with smaller number of the correct lines in the overlapped area might be selected and probably not all of errors are corrected. With proposed method, based on the statistical information, because the number of correct lines in the overlapped area of the image from Tokyo is greater, it is selected to make sure that more errors will be corrected. After correction, all the errors in the top part of the original image are corrected. The real error information of the images from Bangkok and Tokyo is then added to the database. In figure 9, the original image from Tokyo has two reference images from Bangkok and Ulaanbaatar. Because these images do not belong to the database, the statistical error information is used. Based on this information, the image from Bangkok is selected to correct the error at the bottom part of the original image; the image from Ulaanbaatar is selected to correct the error at the top part of the original image. After correction, all errors are corrected. The real error information of these images is also added to the database. Compared to the previous methods, the processing time of proposed method is shorter because it does not need to analyze the error information of the reference images before selection. References [1] Paul M. Mather, Computer Processing of Remotely-Sensed Images, John Wiley and Sons, Inc., England, Third Edition, 2004, ch. 2. [2] K. Yamauchi, M. Takagi, Quality inspection and error correction of NOAA data, and its application to the generation of Asian mosaic, Proceedings of International Symposium on Remote Sensing 2000, pp , [3] A. N. Van, Y. Aoki, Error correction for NOAA AHVRR data using reference data, Proceedings of Asian Conference on Remote Sensing 2006 (CDROM, I-11), Ulaanbaatar, Mongolia, [4] A. N. Van, Y. Aoki, Error correction for NOAA AVHRR data with reference to the same orbital data, Proceedings of International Workshop on Advanced Image Technology, Bangkok, Thailand, pp , [5] A. Ardeshir Goshtasby, 2-D and 3-D Image Registeration for Medical, Remote Sensing, and Industrial Applications, John Wiley and Sons, Inc., Hoboken, New Jersey, 2005, ch. 4. [6] Bernardo Esteves Pires, Perdo M. Q. Aguiar, Registration of Images with Small Overlap, IEEE 6th Workshop on Multimedia Signal Processing, pp , CONCLUSIONS A new method to correct missing lines and error lines in NOAA image is proposed. This method corrects errors by using the reference images that overlap the original one. The quality of the reference images is assessed based on the error information database; therefore, the number errors in the overlapped area of reference images are minimized. The information in the database also helps to save the analyzing time for un-used reference images. The best quality reference image is used first to correct errors. After correction, the other ones will be used if there are still errors in the original image. Missing and error lines are corrected by assigning the correct values of the corresponding lines from reference images. Consequently, missing data could be fully restored and error lines could be precisely corrected. The results proved that all most all of the errors were corrected within a quite short time. At this time, the database is quite small because it contains only the information of the image three receiving stations in the short duration of time. In the future, the information of the images from other receiving stations will be added to the database. At that time, an original image will have more reference images and the result will be better. The cumulative parameter also has an important role. When more data is added to the database, more values of this parameter will be tested and its best value will be found more accurately. An Ngoc Van received the BS and MS degrees on Information Technology and Communication from Hanoi University of Technology, Vietnam, in 2000 and 2003, respectively. He is currently a Ph.D. student at Shibaura Institute of Technology, Japan. His research interests include speech signal processing and satellite image processing. Yoshimitsu Aoki received the BS, MS and Ph.D. degrees on Applied Physics from Waseda University, Japan in 1996, 1998 and 2001, respectively. He is currently Associate Professor of Department of Information Science and Engineering, Faculty of Engineering, Shibaura Institute of Technology, Japan. He is involved in researches related to image processing and vision systems that integrate images and sensory information. He is a member of the Institute of Electronics, Information and Communication Engineers (IEICE) and Information Processing Society of Japan (IPSJ).

9 Precise error correction method for NOAA AVHRR image using the same orbital images 135 Fig.7: The first test: All images belong to the database. Fig.8: The second test: All images do not belong to the database but are received from June to October,

10 136 ECTI TRANSACTIONS ON ELECTRICAL ENG., ELECTRONICS, AND COMMUNICATIONS VOL.5, NO.2 August 2007 Fig.9: The third test: All images do not belong to the database and are received in 1998