Performance Validation of Surface Filter based on CUBE Algorithm 1 Performance Validation of Surface Filter based on CUBE Algorithm for Eliminating Outlier in Multi Beam Echo Sounding Yosup Park 1 Seom Kyu Jung 1 Nam Do Jang 1 Jung Soo Kim 2 Jun Sik Lee 3 Abstract The purpose of this study is to verify a filtering performance, Combined Uncertainty and Bathymetric Estimator(CUBE) based on Total Propagated Error and statistical estimation of MBES. R/V Badaro2:KM EM3002 multibeam data were collected at National Basic Maps of the Sea Project 2012 and processed using CUBE modeling and discrimination methods, which is actually used at field work. To compare with data processed using CUBE, hydrographic experts of Korea Oceanographic and Hydrographic Association have manually cleaned the same dataset and we evaluated the similarity. The comparison result of 24 datasets showed that the similarity value of them is mean 95%, ranging from 90~99%. CUBE methods would make a optimal water depth s determination easier by providing objectivity and efficiency and can increase efficiency by 30% in the data processing procedure. Kew words : CUBE, TPE, MBES, Surface Filter 1. Introduction Multi Beam Echo Sounders have been changing the paradigm of hydrographic surveys. Existing Single Beam Echo Sounders measured water depth right under the research vessels, interpolated it and presented general outline of bathymetry. Also, because the beam width of sound used by single beam echo sounders relatively wider than that used by multi beam echo sounders, only average outlines of real areas were provided and running direction of the research vessels, movement of the vessels owing to rolling & pitching or sound-ray modeling have huge uncertainty as an severe cause of false measurement. In addition, by measuring water depth at nadir of the research vessels, recalculating horizontal position on each(individual) observed site wasn t needed. But by introducing multi beam echo sounders, the amount of observed value has increased more than hundreds of times. The increased soundings increased the cost and process time and demanded to improve the abilities of computers to process. To exactly remodel the ray of beam scanned back and forth toward the slant range, the observation on physical environment of water column periodically must be corrected and a variety of problems such as variation(difference) due to moving vessels, data quality management of overlapped observation data, etc. Especially, processing huge amount of sounding worked as a bottleneck in the course of all the hydrographic surveys, which became the cause of increasing update cycle of nautical chart and the cost. To foster the efficiency of processing sounding will 1 Maritime Security Research Center, KIOST. 787, Haeanro, Ansan 426-744, KOREA skjung@kiost.ac (Corresponding author) 2 Korea Oceanographic and Hydrographic Association. 505-14, Gasan-dong, Geumcheon-gu, Seoul 153-708, KOREA 3 Hydrography Survey Division, Korea Hydrographic and Oceanographic Administration. 351, Haeyang, Youngdo-gu, Busan 606-806, KOREA
2 Performance Validation of Surface Filter based on CUBE Algorithm make update cycle of whole nautical chart shortened and budget cut down. As the method to read sounding filled out in the nautical chart by checking all the observation value, according to the historic practice of hydrographic surveys, was introduced in the process of multi beam echo sounding, more data processing and resources were demanded than acquiring data in field. To solve this problem, though various discrimination algorithms on outlier data were developed, they didn t secure the quality of level processed by experts(eeg 1995, Debese 2001). Since 2003, CUBE(Combined Uncertainty and Bathymetry Estimator) suggested by Dr. David Brain Calder in the University of New Hampshire, which is statistical surface model and processing method by using the model, has been recognized by experts and has begun to be utilized by being loaded in most of commercial data processing programs(calder, 2003). The representative result of outlier causes & effect analysis research, the basis of CUBE, was Error Budget Analysis for US Naval Oceanographic Office, which was requested to the University of Southern Mississippi by Ocean and Atmospheric Science and Technology Directorate under U.S. Naval Research Laboratory in 2001 and leaded by Rob Hare of Canadian Hydrographic Service. Later, this report was edited and published as an appendix of IHO S-44, Fifth Edition and provided basis of CUBE model of Dr. David Calder. Also, Total Propagated Error(TPE) of Rob Hare provided the base of TPE calculation formula embodied by CARIS(Rob Hare, 2001). Dr. David Calder, based on TPE Model of Rob Hare, suggested CUBE, applying uncertainty of each (individual) beam and statistical methods for topography estimation in Grid Model. CUBE is providing not only bathymetry models but standards of surface filter based on the models and by providing beam uncertainty attributes, is making structural frames of electronic chart data(calder, 2005). After the suggestion of CUBE, NOAA NOS has applied CUBE to real fields and evaluated its value(calder, 2004). CUBE research business of NOS including functional completion of automation filter, time efficiency of the filter compared with experts, methods of grid size selection etc. has been performed for years and applied as standard policies of Field Procedure Manual published NOAA in 2012(NOAA, 2012). CUBE Algorithm has been being consistently developed by CCOM/JHC(Center for Coastal and Ocean Mapping and NOAA/UNH Joint Hydrographic Center) and was embodied by being provided with IVS3D (2003, PFM editor), CARIS HIPS/SIPS (2005, VASE surface technology), QPS (2006, QLoud, a 3D editor), Kongsberg Simrad(2005, SIS), Reson, Triton Imaging International(2005, ISIS)(Brain Calder and David Wells, 2007). Applying CUBE to standard processing procedure of multi beam data, Brazilian Navy Directorate of Hydrography and Navigation(DHN) academically(scientifically) evaluated its efficiency through comparison to Swath Line filtering used those days(lcdr et al., 2009). NetSurvey, which is a representative hydrographic survey company in UK, kept discussing methods of filtering hypothesis produced in CUBE model by using Fledermous of QPS, not HIPS of CARIS(Ducan et al., 2007). This case study referred that if a representative hypothesis autoselected in CUBE doesn t correspond to an actual result, the course that its user reselects a hypothesis is important. Miguel E. Vasquez in Chilean Hydrographic Office(SHOA) provided the result applying special filter by making up of TPE on Atlas FANSWEEP 20(200kHz)and HYDROSWEEP MD2(50kHz) and setting CUBE parameter based on CARIS HIPS6.1(Miguel, 2007). As mentioned above(vide ut supra), it is considered that multi beam echo sounding outlier discrimination based on CUBE algorithm and the resultant digital sounding data processing methods will have a variety of effects on entire process of hydrographic surveys. To understand how CUBE algorithm works and how each parameter affects CUBE algorithm will provide field workers with an important issue. To do so, Chapter 2 shows analysis of CUBE algorithm and Chapter 3 suggests the result analysing performance and efficiency of CUBE algorithm compared to that of
Performance Validation of Surface Filter based on CUBE Algorithm 3 existing processing methods. Chapter 4 summarizes study results and provides future application plans. 2. CUBE(Combined Uncertainty and Bathymetric Estimator) Theory CUBE is the algorithm that estimates grid bathymetry model with sounding data that noises mix. That is, it is the method that calculates uncertainty of beam and estimates specific bathymetry model from the sets of beams with uncertainty attributes. To calculate the location of beam, multi beam echo sounder utilizes not only Multibeam Transducer but also GPS, Gyro Compass, Motion Sensor, observed values of Surface Sound Speed Sensor, Sound Velocity Profiler and Tide gauge. Unlike earth survey on the fixed platform, oceanographic surveys calculate the depth under the datum level of research area by using sensors calibrating the state of in every minute and consistently moving vessels, water temperature structure to estimate the process that sounder beam comes back scattered against seabed and tidal data calibrating movement of surface of the sea in survey area. Each calibration sensor has an error range and false measurement included in calibration affects the final beam computation. The effect that these sensor calibration values have on the beam computation is called Total Propagated Error and Dr. David Calder suggesting CUBE, by using Error Propagation model of Rob Hare in CHS, built CUBE algorithm (Rob Hare, 2001). By introducing TPE, multi beam echo sounding data include vertical & horizontal uncertainty computed based on TPE model. Generally, uncertainty is low near nadir and becomes high at the outskirts where grazing angle becomes reduced. Also, as the distance where each beam echo locates the bottom(surface) becomes far, the uncertainty of each beam becomes increasing. By adopting TPE, the uncertainty of each beam can utilize an index of survey result and TPE filter to use depth less than uncertainty designated by users to the result is embodied in CARIS HIPS. CUBE is based on grid model with a limited area. Grid has two dimensional flat structure and positioning information of grid node doesn t have uncertainty. Thus, uncertainty information of grid node has only attributes of node, namely, has total uncertainty modifying depth and horizontal positioning uncertainty by vertical positioning uncertainty. Traditional interpolation model uses the method that sets up the representative value with the algorithm making up existing grid model through correlation of position between node and each beam such as Kriging, Natural Neighborhood Inverse Weight Distance, etc. Existing method setting up the representative value considers horizontal positioning of observed values but CUBE estimates the representative value, considering not only each positioning but also the range of influence according to horizontal & vertical uncertainty of observed values and the tendency with the surrounding node values. The representative value can be estimated with various hypotheses according to the parameter designated by users. Of various hypotheses, users select an optimal hypothesis through the algorithm, using the tendency with surrounding node and statistical parameter to make up hypotheses. Through TPE analysis model, all the observed depths have estimated errors as attributes and using these attributes and horizontal position of observed area, the observation point is assigned at the grid nod within the radius where the observed depth has an influence. Individual node contains value estimated as depth of the area and the distribution information of observed values as attributes. To clean the noises mixed in observation, the process to monitor if a new observed value corresponds to existing depth data sets is embodied. If the new value exceed the range of existing estimated values, users produce a new estimated value model and keep multiple hypotheses at one estimated dept node. If all observed values are assigned at each node, according to numerical standards, users select an optimal hypothesis at each node. The goal of CUBE lies in estimating real depth by using the given standards including the number of observed depths at each hypothesis, continuity with
4 Performance Validation of Surface Filter based on CUBE Algorithm surrounding(neighboring) hypotheses, consistency between observed values within hypothesis. Through this, estimation depth models of whole target waters can be made up and it is possible to re check the criteria & methods about the process of depth decision by providing statistical background information on an estimated value as depth of relevant grid node. The process to propagate uncertainty of individual observed depth into that of node follows formula (1). (1) This formula is to calculate uncertainty( ) of each node and represents core process of CUBE model. Since each node exists on the coordinate system, horizontal position inaccuracy is zero(0). Therefore, it need not be shown and according to nodespacing designated by users, distance between sounding position and grid centerposition(dist), hes and de, uncertainty of each node, horizontal( )/vertical( ) uncertainty. This process is called Scattering in CUBE. Scattering is the process to assign observed depth to each node. Then, position of sounding, distance between each node and uncertainty that each observed depth has is used. Also, in this process, users can calculate uncertainty of observed depth to significantly affect a relevant node. Two parameters used in the process are Horiz_error_scale(hes) and Distance-exponent(de). To calculate node uncertainty, hes is used as a modulator about horizontal positioning uncertainty of each estimated value within the radius of influence of node. To vertical positioning uncertainty due to distance between node and each observed value of depth, de is used as a factor. Total inaccuracy of each node is calculated by formula(1). Formula(1) represents inaccuracy of node estimated by mixing horizontal inaccuracy and vertical inaccuracy that each observed depth has. Each node in the model doesn t have horizontal positioning error. Grid node only has a vertical positioning error and this vertical positioning error is represented as a mixture of horizontal and vertical positioning error of each sounding. As hes value decreased, an influence of error constituent that has on each node and the probability to make up various estimated hypotheses increases. In other word, if this value becomes decreasing, the influence by distance between node and each observed depth becomes increasing and this means that the reliability of estimated depth becomes high. After calculating total uncertainty of each beam, the process to decide if each depth of node is applied at node calculation is required. This process is called gathering and needs capture distance scale and minimum capture distance given by users. Capture distance scale is to consider detected area by beam which spreads depending on depth and minimum capture distance, in case the depth is too low, is to prevent minimum capture distance by capture distance scale from being smaller(shorter). Through the process of scattering and gathering, multiple observation values are assigned and based on depth and total uncertainty in observed value, the representative value is estimated. Calibrated depth assigned at the node, depending on depth contained in observed value and the rules of creating hypotheses designated by users, can generate multiple and. This process is called intervention in CUBE. Estimated is based on calibrated depth and considering error range that calibrated depth has and separation distance with grid, is an expectation value of set and means uncertainty of i th hypothesis. (2) In the above, and mean node position, is a number of generated hypothesis, is depth and is total uncertainty. As a new observed value is assigned at the node, expectation value and total uncertainty is upgraded and if expectation value exceeds
Performance Validation of Surface Filter based on CUBE Algorithm 5 designated specific distance, another new hypothesis is generated. Users can confirm the distribution state of hypotheses through graphical user interface and through reediting hypothesis selected by algorithm, select the hypothesis estimated by users as the representative hypothesis of grid. This process is called Disambiguation. Selecting the representative hypothesis, users can use the following: (1) the number of samples that a hypothesis consist of, (2) the maximum similarity to neighboring nodes, (3) the mixture of the previous (1)and (2), (4) the method to compare to the similarity to former survey result. As CUBE model of relevant area is generated, assuming the model as a true depth model of object water(sea area), it is embodied that outlier within the observed depths acquired in the research object area can be removed. Estimated values that deviate from the representative value of grid which is the most adjacent to positioning of calibrated data and separation distance error designated by users is judged(considered). In CARIS HIPS 7.0, not going through(using) the existing method that must inspect all observed data and decide good or bad(quality) of them, through generating CUBE model and the resultant filtering, a new method to eliminate outlier is embodied, which can discriminate outlier rapidly. 3. Experimental Method and the Result Research object data were selected depending on topography (geographical features) and depth, from multi beam echo sounding data acquired from Haeyang 2000, Badaro 1 and Badaro 2 which had been sent in National Bathymetry Mapping Program. In Haeyang 2000, Seabat 7125 of Reson is being run and in Badaro 1and Badaro 2, EM3002 system of Konsberg Maritime is being run. Depending on the installment types of equipment of survey vessels and added sensors such as GPS, Compass, Motion Reference Unit, etc., we acquired numerical values to estimate Total Propagated Error(TPE). Table 1 is to summarize the object data used in estimation. Table 1. Dataset for validation of CUBE Algorithm Vessel Badaro1 Multi Beam Echo Sounder KM EM3002 Depth Area range Feature (m) Flat 30 80 Sandwave 50 80 Shoals 50 90 Shipwreck 40 70 Badaro2 KM EM3002 Flat 40 70 HY2000 RESON Seabat 7125 Flat, Sandmining Area 60 80 The features(characteristics)of object data to estimate(appreciate) algorithm is the following. Out of data of Badaro 1, those of flat area, sand wave area, shallow area, area being shipwreck were separately estimated. Out of data of Badaro 2 and Haeyang 2000, those of only flat area were used to estimate algorithm. Predetermined value of TPE to calculate TPE depending on research(exploring) systems was applied referring to guidelines of CARIS. Estimation object data was those that Research Association completed all final inspections and those of field sound velocity correction, tidal height correction was applied pursuant to standards of Association. Estimation on CUBE was performed based on CARIS HIPS 7.1. In order to estimate the features(characteristics) of the method to generate CUBE model, we compared CUBE model made of data from which experts eliminated outlier to that made of raw data not eliminating outlier from. Later, to estimate the performance to process elimination of outlier based on CUBE model, we compared all experts' cleaning dataset to CUBE filtering dataset. In CARIS HIPS, because data which experts decided as outlier didn t delete in file and were processed as flaggings, before application of CUBE, we made all the flaggings initiated and after that applied CUBE algorithm. To compare data filtered by experts to those filtered by CUBE algorithm, all the attribute data( ping number, beam number, sounding, estimation or not, etc.) related to data processing including all the sounding data and flagging or not, etc. were extracted into ASCII file stage by stage(step by
6 Performance Validation of Surface Filter based on CUBE Algorithm step) and to effectively perform all comparison of both data, we utilized comparison script drawn up in PythonXY program based on Python 2.7. Fig. 1 is to schematize the experimental method for the research. Table 2. The Description of role of parameters of CUBE Parameters Default Deep Shallow Estimate Offsets(EO) 4.0 3.0 2.0 Capture Distance Scale(CDS) Capture Distance Minimum(CDM) Horizontal Error Scale(HES) Density Strength Limit(DSL) Local Strength Maximum(LSM) Local Search Radius(LSR) 5.0% 20% 4.0% 0.5m 2.0m 0.4m 2.95 2.95 0.5 2.0 2.0 2.0 2.5 2.5 2.5 1 1 1 Fig. 1. Process Flow of Validation of CUBE. According to Central Limit Theorem, as samples are larger, sample mean becomes nearer to population mean and in that event, it is important how much outlier there is. Hereby decrease in average outlier(error) depends on size(scale) of sample and begins to be stable when size of sample is over 5and 6. When the size is over 30, decrease in average outlier(error) is almost constant(lee, 2011). Grid model is considered as the process to estimate true value, looking upon an actual(real) value of grid as a true value and calibration values as samples. When drawing up grid model, in order to acquire the number of samples up to over 30 recommended by instrumentation engineering, we made up a model with a grid resolution in average depth 5% out of dataset. Though it is known that grid resolutions affect CUBE model, because this digresses from the subject of our thesis, we hope you to refer to Yosup Park s report(2012) to closely discuss the selection of grid resolution model. Parameters necessary to make up models are presented in Table 2. We made and compared models in accordance with default, deep and shallow of CARIS HIPS, which has been already made depending on characteristics of dataset. In addition to estimated representative depths, CUBE model consists of several attribute data including maximum/minimum/median value, reliability and density of data to make up grid node, hypothesis strength, hypothesis count, Standard Deviation(Std_Dev), uncertainty, user nominated depth, etc. In surface filter, after setting the seabed which is standard and designating error range(tolerance) above and below the seabed, if deviating the range, we considered it as outlier. Besides CUBE, data of existing grid models can be utilized as a standard seabed. Hereupon in applying surface filter to eliminate outlier based on CUBE model, altering two discrimination criteria(cube and Standard Deviation) and multiple error range of the criteria, we also performed making up total 24 models and similarity comparison to expert s filtering data in accordance with changes of filter parameters. 3.1 CUBE Model Comparison depending on Filtering Existing study result showed that by introducing CUBE, it got solved that existing filtering process took the longest time and criteria of outlier discrimination depended on workers (observers, researchers) subjective opinions(calder, 2004). First, this study made sure,
Performance Validation of Surface Filter based on CUBE Algorithm 7 in case before using CUBE as the basis of surface filtering, outlier would be contained, how much reliability of the model we would estimate. For this, we compared data with the same grid resolution and same space(surface) range before and after filtering outlier to the similarity to CUBE model. Fig.2(a) is the CUBE model consisting of by experts' cleaning dataset and Fig.2(b) is the CUBE model consisting of raw dataset not filtered. Fig.2(c) is to make up the differences which Fig.2(b) presents when based on Fig.2(a) as a grid model and show them by a (a) CUBE model based on expert cleaned dataset (b) CUBE model based on raw dataset (c) Grid model of gap between CUBE models based on expert cleaned and raw dataset (d) Statistics of difference between CUBE models based on expert cleaned and raw dataset Fig 2. The Case of comparing CUBE models between different datasets (Expert cleaned and raw datasets).
8 Performance Validation of Surface Filter based on CUBE Algorithm diagram. Fig.2(d) is to present statistical values and a histogram for the differences between two models. In most area, very high(much) similarity appears and there are estimated deep depths in red-circled areas. It is shown that the hypothesis selected by algorithm can make the similarity to models by experts high by filtering(editing) hypotheses by users after making up models. When making up CUBE, we built CUBE relevant to both two dataset by utilizing Default Mode and selected hypotheses by utilizing only Density standard of sounding making up the hypotheses. We estimated the difference by applying function of BASE model comparison of CARIS HIPS. When comparing total 46,517 grid nodes, we found the difference max.0.4m and min. -0.17, average 0m, and standard deviation 0.01m. If it is considered that in most areas, we eliminated outlier by Swath Mode filtering, built CUBE model with raw data not performing Swath Mode filtering or fixed histogram interval into 0.1m, we found them same by 99.9%. (a) Statistics of gap between Default and Deep mode of CUBE (b) Statistics of gap between Default and Shallow mode of CUBE (c) Grid model of gap between Default and Deep mode of CUBE (d) Grid model of gap between Default and Shallow mode of CUBE Fig. 3. The Case of Differencing between CUBE model configurations.
Performance Validation of Surface Filter based on CUBE Algorithm 9 3.2 The Result Comparison of Model in accordance with Parameters Change of CUBE Making-up As stated to in Table 2, in order to make up CUBE, 7 parameters are applied and if 4 items(eo, CDS, CDM, HES) are changed, the method to estimate the model is also changed. CARIS HIPS presents users option, which the company has already made up with the most suitable 3 combinations of estimation parameters(default, Deep, Shallow). Users can take them and make up models. For the same dataset, we analyzed the difference among three models made up of with hypothesis selected by algorithm and the differences between depths decided by experts. Figure3(a) presents a histogram and a grid model for deviation between CUBE model applied by Default Mode and CUBE model applied by Deep Mode. This case is to estimate raw dataset (a) Expert Cleaned Set(Outlier: Grey Dot) (b) The Outlier based on Default Mode (c) The Outlier based on Deep Mode (d) The Outlier based on Shallow Mode Fig. 4. Outlier by various CUBE model & Surface filter.
10 Performance Validation of Surface Filter based on CUBE Algorithm used in 3.1. As made sure in Fig..3, deviation between models applied by Deep and Default Mode is very small by 0.04m, showing max.1.76m and min. -1.74m but on the histogram the deviation shows very little appearance ratio. But the deviation between models applied by Default Mode and Shallow Mode shows apparent differences in areas where depth severely changes. The standard deviations between both seem similar but scale of difference in specific nodes is max.3.17m and min. -2.68m. This is because multiple hypotheses arise in areas where depth Table 3. Comparing results between manual cleaned and CUBE based surface filter Mother Ship Data Sets (Total Beams) Model Parameters (Source, Mode, G, CL) MACA (%) MACR (%) MRCA (%) MRCR (%) MACA+ MRCR (%) MRCA+ MRCR (%) Raw, deep, 2.5m, 2 90.51 0.009 9.388 0.091 90.602 9.398 Flat (37,371,506) Raw, default, 2.5m, 2 90.51 0.009 9.388 0.091 90.603 9.397 Raw, shallow, 2.5m, 2 90.31 0.009 9.603 0.074 90.387 9.613 Manual, default, 2.5m, 1 97.43 0.245 1.875 0.445 97.880 2.120 Manual, default, 1m, 2 97.67 0.008 1.914 0.406 98.079 1.921 R/V Badaro 1 (EM3002) Sandwave (5,640,520) Manual, default, 5m, 2 97.66 0.020 1.914 0.405 98.065 1.935 Manual, default, 2.5m, 2 97.66 0.013 1.914 0.406 98.073 1.927 Manual, shallow, 2.5m, 2 97.66 0.017 1.915 0.405 98.068 1.932 Raw, default, 2.5m, 2 97.66 0.013 1.918 0.402 98.069 1.931 Manual, default, 2.5m, 1.5 97.63 0.048 1.902 0.417 98.050 1.950 R/V Badaro 2 (EM3002) R/V HY2000 (Seabat 7125) Shoal (5,379,374) Wreck (5,250,556) Slope (53,582,007) Sandwave (8,918,272) Manual, default, 2.5m 2 95.37 0.182 4.068 0.377 95.750 4.250 Manual, default, 2.5m, 1 94.72 0.831 4.063 0.383 95.106 4.894 Raw, default, 2.5m, 2 95.37 0.182 4.070 0.375 95.748 4.252 Manual, deep, 2.5m 2 93.30 0.469 5.518 0.704 94.013 5.987 Manual, default, 2.5m 2 93.58 0.197 5.508 0.714 94.295 5.705 Manual, shallow, 2.5m, 2 93.30 0.469 5.518 0.704 94.013 5.987 Manual, deep, 2.5m, 2 89.49 0.003 9.954 0.544 90.043 9.957 Manual, default, 2.5m, 2 89.49 0.003 9.954 0.544 90.043 9.957 Manual, shallow, 2.5m, 2 89.49 0.003 9.954 0.544 90.043 9.957 Manual, deep, 2m, 2 88.22 0.019 0.057 11.703 99.923 0.077 Manual, default, 2m, 1 88.12 0.115 0.036 11.724 99.849 0.151 Manual, default, 2m, 2 88.22 0.011 0.061 11.699 99.928 0.072 Manula, default, 2m, 1.5 88.18 0.055 0.044 11.715 99.901 0.099 Manual, shallow, 2m, 2 88.21 0.021 0.058 11.702 99.921 0.079
Performance Validation of Surface Filter based on CUBE Algorithm 11 severely changes and wrong hypothesis out of them is selected. In Defalt Mode to Shallow Mode, because EO becomes narrow by 50% and HES decreases by 6 times, Shallow Mode has more hypotheses than Default Mode and based on the density of candidate hypotheses in the process of selecting an optimal hypothesis, different depth is adopted, which makes a difference. 3.3 Estimation of CUBE Filtering Effect CUBE model is a surface model estimated from calibration data. It is the set of representative depths which copy actual surfaces and is the model which can estimate the reliability of estimated depths, recording vertical uncertainty that the representative depths have. Filtering is the work to assign the attributes that outlier data can t be used in filling in the result by discriminating outlier based on CUBE models. To automatically perform the process of filtering outlier, CARIS HIPS provide the function of Surface Filter. Surface Filter is embodied in the way that setting a standard surface, designating the error range above and below the surface, if values deviate the range, they are considered as outlier. Existing built grid model data, not CUBE models, can be utilized as a standard surface. To estimate the performance of filtering result based on CUBE model, we compared the result to experts handwork filtering result. For comparison, we extracted all individual beam attributes of comparative data, numbering them by using Export function of CARIS HIPS. Because extracted individual beam attributes contain the history of outlier discrimination or not( including flagging or not, in which stage to flag, filter(filtering) name), we performed the result comparison by intercomparing them. We used the script drawn up by utilizing Python2.7 for whole comparison between experts performance and CUBE filtering performance. Fig. 4(a) is to present outlier data adopted by experts on the model. Fig. 4(b), (c) and (d) are to present outlier discrimination data by Surface filter based on Default, Deep and Shallow CUBE, as put in order. We can see that much more outlier were adopted where depth severely changes but we can also make sure that the number was not many for the number of whole outlier samples. Due to limit of space, we summarized the comparison result in various bathymetric areas in Table 3. Table 3 presents data acquired from Haeyang2000, Badaro1 and Badaro2 selected as estimation targets and whole results to perform data processing by combining CUBE and Surface Filter. Assigning data filtered by experts as comparison targets, we estimated all the following: in case CUBE Surface Filter decides calibration data of real depths discrimination by experts as real(actual) depths(maca, Manual Accepted CUBE Accepted), in case CUBE Surface Filter decide calibration data of real depths by experts as outlier(macr, Manual Accepted CUBE Accepted), in case CUBE Surface Filter decides calibration data of outlier as real depths(mrca, Manual Rejected CUBE Accepted) and in case CUBE Surface Filter decides calibration data of outlier discrimination by experts as outlier(mrcr, Manual Rejected CUBE Rejected). When we analyzed the result of Table 3, in setting tolerance to 2 based on CUBE model, the result was similar up to 99.9% to that of experts' cleaning. The beam which brought the result contrary to experts selection formed only 0.07%. And if tolerance range is narrowed by increasing reliability, such values becomes a little low at identical selection rate. This is the expected result, in filter is embodied. Targeting the data acquired in flat, sand wave, shallow and shipwreck areas through Reson Seabat of Haeyang2000, EM3002 system of Badaro1 and Badaro2, we tested the data changing multiple grid size, CUBE generation mode, Surface Filter mode. Similarity to the result of experts' cleaning was max.99.92%, min. 90.3% and mean 95%. What the result suggests, if experts' cleaning focuses on shallow and depth succession area and CUBE Filtering can be used general area, it is expected that one-third in comparison with existing time and cost will be saved(calder, 2004). Also, by introducing CUBE Surface Filter and recording the history of outlier discrimination
12 Performance Validation of Surface Filter based on CUBE Algorithm criteria and process as attributes in individual beam, BAG(Bathymetric Attribute Grid) that S-100, the next generation electronic chart demands will be embodied automatically and we can expect reliability of final performance to enhance. CUBE model follows BAG model of the next generation electronic chart S-100, it isn t necessary to re-draw up attributes in order to make electronic charts. TPE values of each beam calculated to draw up CUBE model will be utilized in BAG model. 4. Conclusion In this validation test, we analyzed the constitution principle of CUBE model and the characteristics of parameters affecting drawing up CUBE model through comparison to experts' cleaning data. The result of this validation test is the following. (1) We made up and compared three rules of CUBE drawing-up(default, Shallow and Deep), with the same data. Deviation among 3 modes isn t large(maximum deviation mean is just 0.12m) but in shallow water, by using Shallow mode, it is necessary to induce the conservative sounding discrimination. (2) When we compared CUBE made up of experts' cleaning data to CUBE madeup of raw data without pre-cleaning, deviation mean was 0m and standard variance of deviation was 0.01m. Though experts may not eliminating outlier by using Swath Editor, in making grid model, we can bring almost same result as filtering-completed grid model with raw data. (3) We estimated performance of Surface Filter based on CUBE model. The similarity between experts' cleaning data and data using Surface Filter was max.99.9%, min.90.3% and mean95%. When the process to eliminating outlier which spends the most time and resources out of data processing is automated, similarity in comparison with handwork was 95%. It is considered that this result will bring various expected effects. (4) If experts focus on principal shallow seas & areas that CUBE can outlier owing to characteristics of CUBE algorithm, most of survey areas becomes assigned in automated process, by dedicating time and expertise to principal depths, data processing time will be shorten and subjectivity of outlier will be excluded. Since Later, to apply data processing technique based on CUBE to fields, error management about field observed data(tide and sound velocity) which are external outlier of survey system will have to be performed and the revision of related guidelines and further research will be needed. Reference Byung-ryul Lee,. Sensor Measurement Engineering, Hongrung Publishing Company, 2011, p.283. Calder, B.R., Automatic Statistical Processing of Multibeam Echosounder Data, Int. Hydro. Review, 4(1), 2003. Calder, B. R. and S. M. Smith. A Time Comparison of Computer-Assisted and Manual Bathymetric Data Processing, Int. Hydro. Review, 5(1), 2004. Calder, B.R., J. S. Byrne, B. Lamey, R. T. Brennan, J. D. Case, D. Fabre, [5] B. Gallagher, R. W. Ladner, F. Moggert, M. Paton. The Open Navigation Surface Project. Int. Hydro. Review, 6(2), 2005. Calder, B.R. and David Wells, CUBE User's Manual, version 1.13, 2007, p. 53. Debese, N., Use of a Robust Estimator for Automatic Detection of Isolated Error Appearing in Bathymetry Data, Int. Hydro. Review, 2(2), 32-44, 2001. Ducan Mallace and Paul Robertson, Alternative use of CUBE; how to fit a square peg in a round hole, US Hydro Conference 2007, www.thsoa.org/hy07/02_02.pdf. Eeg, J., On the Identification of Spikes in Soundings, Int. Hydro. Review, 72(1), 33-41, 1995. LCDR Aluizio Maciel de Oliveira Junior and CDR Izabel King Jeck, Multibeam Processing for
Performance Validation of Surface Filter based on CUBE Algorithm 13 Nautical Charts (Using CUBE and Surface Filter to enhance multibeam processing), International Hydrographic Review Nov. 2009. p.63-73. Miguel E. Vasquez, Tuning the CARIS implementation of CUBE for Patagonian Waters, Mater Thesis of University of New Brunswick, 2007, p.108. NOAA, 2012, Field Procedures Manual, 2012, April, p.284. Rob Hare, 2001, Error Budget Analysis for US Naval Oceanographic Office (NAVOCEANO) Hydrographic Survey Systems, University of Southern Mississippi, Hydrographic Science Research Center (HSRC), Final Report for Task 2, FY 01, p. 153 Yosup, Park., A study on the improvement of multibeam data processing, Report, Korea hydrographic Association, 2012.