Comparison and Analysis of Publicly Available Bathymetry Models in the East Adriatic Sea

Summary In this paper the latest versions of six publicly available bathymetry models: DTU10bat, EMODnet 2018, ETOPO1, GEBCO 2020, Smith and Sandwell V20.1 and SRTM15+ V.2.1. are compared and evaluated in the area of continental shelf of Croatia settled along well intended east Adriatic coast. Survey data in the area is not available through open access data bases, but publicly accessible in agreement with data holders (hydrographic institute, research centres and industry). These grids provide alternative sources of information about seafl oor topography. Marine researchers should be acquainted with the main characteristics as well as pros and cons of bathymetry models in order to choose the best one for a specifi c purpose. In this paper the most important characteristics and information about grids are presented: resolution, coverage, release date, horizontal and vertical datum, data source, registration method, producer and link to website with an emphasis on the underlying source data. The underlying source data is one of the most important parameters that determine the quality of the bathymetric model. Hypsometry curve that is describing the area distribution of depth is calculated for each bathymetry model over the test area of the east Adriatic. For pixel to pixel comparison, grids were resampled to same one-minute resolution and absolute diff erences between models are calculated in identical points. Absolute diff erences between models show level of mutual compatibility between models as well as areas of highest disagreements that indicate the presence of outliers or systematic errors within models. In order to demonstrate how well publicly available bathymetry models fi t the true topography of the sea fl oor, grids were compared to high-resolution digital bathymetry model interpolated from the multibeam survey in the area of Murter Sea. This paper should assist in the choice of a most suited bathymetry grid in future maritime studies in the Adriatic.


INTRODUCTION / Uvod
Bathymetry refers to the information about depth of the sea, i.e. a vertical distance to the seafl oor relative to the chosen sea level. Digital bathymetry model (DBM) is a digital terrain model that represents topography of seafl oor [32], mostly in the form of regular grid with depth values assigned to grid cells [28]. Since oceans and seas cover more than 71% of the Earth [9], knowing topography of the seafl oor is important on both global and local scale preferably formatted in the form of digital bathymetry model. It is possible to derive seafl oor terrain attributes: slope, orientation, curvature, variability directly from digital bathymetry model [31] or to use it as frame in diff erent scientifi c analyses. Although primary use of bathymetry data is to ensure safety of navigation [27], bathymetry data is a necessary parameter in many studies: geohazard assessment in off shore area [8], reviling marine geomorphology [38], computations in physical geodesy [44], tsunami modelling [30], modelling atmospheric infl uence in off shore area and eff ect of meteotsunami [39,45], modelling ocean currents [15], mapping marine habitats [7,49], etc.
Despite the fact that bathymetry data underpins almost all maritime activities, accessing directly observed bathymetry data is not an easy task for several reasons.
Even though collecting bathymetry data lasts for centuries, less than 20% of ocean fl oor at 30 arc seconds resolution has been directly surveyed by echo sunders [34,48]. It is commonly said that some planets in solar system are better mapped than oceans. It is predicted that seafl oor mapping below 200 m with one ship would last 350 years with a total cost of 3 billion US $. It equals in cost to ne extra-terrestrial mission [19,34].
As compared to deep ocean areas, coastal continental shelf, defi ned as sea area limited to a distance of 200 nautical miles from the baseline of a coastal state [42], are very well surveyed but these data are held private by governments, research institutions or private companies [50].
Alternative bathymetry data sources are publicly available bathymetry grids. Publicly available bathymetry grids can be global or regional. They are calculated using diff erent type of source data: shipboard soundings, nautical chart soundings and contours, satellite derived bathymetry using multispectral images, bathymetry data predicted from gravity, etc. Data is collected in diff erent time periods using diverse technology and interpolation methods. As a result, there is a number of publicly available grids to choose from today. To know which grid to choose for a specifi c purpose, grids should be compared and analysed. As terrestrial global digital terrain models are regularly evaluated on global and regional levels, such studies are rarely done for publicly available bathymetry grids.
Marks and Smith in 2006 evaluated six global publicly available bathymetry grids with a focus on a Woodlark Basin and adjacent Coral sea area east of Papua New Guinea which is an area that exhibits a variety of sea fl oor features, including abyssal hills, seamounts, a plateau, ridges, fracture and a subduction zone [33]. Although evaluation was made 14 years ago and observed models are outdated, they have underlined features and characteristics of digital bathymetry model that aff ect the accuracy: source data, interpolation method, presence of artifacts, etc. Abramova in 2012 followed and broadened the work of Marks and Smith and evaluated grids over specifi c Artic region [1]. In 2019 Florentino et al. compared global bathymetry models in area near Brazil and demonstrated how grids can be updated with more accurate regional data [11]. Analyses of digital bathymetry models in Adriatic have, to some extent, already been made for the purpose of physical geodesy. Bašić and Buble in 2007 evaluated Smith and Sandwell model, incorporated in ETOPO 2 and SRTM30+ models, against contour data derived from nautical map over the Adriatic Sea (scale 1 :1 000 000) [4]. In the same area, Bjelotomić in 2015 mutually compared six bathymetry grids: International Gravimetric Bureau model BGI, DTU10bat, GEBCO, ETOPO, Morelli, SRTM30+, Smith and Sandwell in order to compute regional geoid [5].
The purpose of this study is to analyse and compare latest versions of publicly available bathymetry models in shallow sea area of the east Adriatic Sea. Unlike the west Adriatic coast that is generally regular, sandy and with a gentle slope, east coast is challenging to model because it is irregular, with many islands, and a rocky steeply sloping bathymetry [35]. Six publicly available bathymetry models: DTU10bat, ETOPO1, EMODnet, GEBCO, SRTM15+, Smith and Sandwell are compared and analysed in the continental shelf of the Republic of Croatia settled in east Adriatic Sea (Figure 1a). This work provides mutual comparison between latest versions of grids available at the moment and analyses of underlying source data, general statistics and distribution of depth. Furthermore, digital bathymetry grids are compared with high resolution digital bathymetry model (2 m grid spacing) based on multibeam data in the area of Murter Sea which is an example how data fi t actually terrain (Figure 1b). This analysis should assist in choice of a most suited bathymetry grid in future maritime studies.

DATA AND METHODS / Podaci i metode
In this study six publicly available digital models are tested in the Croatian continental shelf. Following section gives a description of bathymetry grids with an emphasis on underlying source data and methods used to mutually compare them.

Bathymetry models and source data / Batimetrijski modeli i izvor podataka
Basic characteristics of six publicly available grids that are examined are summarised in Table 1. The resolution of digital bathymetry model refers to spatial sampling interval, simply to the size of its grid cell. However, spatial resolution of a sensor, measure of smallest object that can be resolved by sensor can be by far greater [41] or the depth of a particular cell may be interpolated from a relatively distant source data since only a small part of worlds sea areas is measured [34,41], so this can be a misleading value. Thus, it is important to know the source of data upon which cell depth is derived and the information about it should be available for all grid cells. Information about source data that grids are built upon is given below, starting with DTU10BAT model, in alphabetical order.
DTU10 Bathymetry (DTU10BAT) is a Danish global bathymetry model developed at the Danish National Space Institute (DTU Space) [3]. The model is calculated from global gravity model DTU10 and depths from GEBCO One Minute Grid [6] (Figure 2a) using the Smith and Sandwell method to predict depths from gravity [37] with one arc minute grid spacing. In contrast to the fi rst version of the model DNSC08, released in 2008, DTU10 included data from ERS1 and GEOSAT altimetry missions which enhanced recovering of even small seamounts.
EMODnet 2018 bathymetry grid (EMODnet) is a regional bathymetry model that is covering the area of European seas with 1/16 arc minute resolution [10]. It is part of European Marine Observation and Data Network, developed under EU Marine Strategy Framework Directive and the EU Marine Knowledge 2020 agenda and fi nanced by the European Union. Data collection started in 2009 and the fi rst version of the model with 15 arc seconds resolution was released in 2013 [40]. It is constructed from heterogenous bathymetry data sets including survey data, data from nautical chats and composite grids with gaps fi lled with GEBCO 2014 grid. EMODnet bathymetry portal provides source data reference for every cell through metadata with quality indicators and link to data source holder. In test area, part of Kvarner bay and coastal sea area near Dubrovnik are based on ENC (Electronic navigation chart) data with following quality indicators: Vertical (VI): 1 (multisource), Horizontal (HI): 1 (multisource), Usage (UI): 3 (hydrographic survey or compatible with hydrographic standards), part of eastern slope of South Adriatic basin is covered with multibeam data VI: MBES low frequency, HI: 3 (< 20m), UI: 2 (bathymetric/ morphologic survey) and border of test area is interpolated from singlebeam soundings VI:1 (similar than 2+5%d), HI: 3 (< 20m), UI: 2 (bathymetric/morphologic survey). However, most of area is fi lled with GEBCO 2014 data ( Figure 2b). ETOPO 1 is a global terrain model for land and sea areas, generated and published by the American National Geophysical Data Center (NGDC) in 2008 [2]. It is constructed by patching several available regional and global bathymetry models. Bathymetry data is mainly based on GEBCO and Smith and Sandwell (SS) models. ETOPO1 is publicly available in two versions: "Ice Surface" version that includes the Antarctic and Greenland ice sheets, and a "Bedrock" version that delineates the bedrock underneath the two ice sheets. ETOPO 1 only indicates source data for test area to be GEBCO and SS model. GEBCO 2020 grid is a continuous, global topography and bathymetry model with spatial resolution of 15 arc seconds [14]. It is produced through the Nippon Foundation -General Bathymetric Chart of the Oceans (GEBCO) Seabed 2030 Project. Base of the model is SRTM15+ data which is augmented with bathymetry data, mainly based on multibeam surveys, collected through Seabed 2030 project. Compared to the earlier version of the model GEBCO 2019 where new data was added to the model using remove-restore procedure [13], new data is added to the base layer of GEBCO 2020 "as it is", without any blending to avoid the edge eff ect on boundaries between layers. GEBCO 2020 bathymetry grid is accompanied by Type Identifi er Grid (TID). This data set identifi es the type of source data that the corresponding grid cells in the GEBCO grid are based on. As seen in Figure 2c, in test area, data is mainly based on gravity predicted data (SRTM 15+) augmented with soundings from combination of direct measurements methods and diff erent chart data, not referring to the data holder. However, since GEBCO and EMODnet are exchanging data, it is noticeable that diff erent direct measurements in GEBCO TID grid originate from EMODnet database.
SRTM15+V2.1 is a global terrain model for ocean and land with spatial sampling interval of 15 arc seconds [41]. Version The fi rst phase is a construction of the1-minute grid using a combination of V29.1 gravity model (SIO) and ship soundings augmented with depths from GEBCO 1-minute grid [6] following Smith and Sandwell method [37]. The second phase increases the resolution of the grid to 15 arc seconds and upgrades base map using remove-restore procedure where ship soundings are available. As compared to previous versions, new version included more years of altimetry measures and new shipboard soundings covering 10.84% of ocean fl oor in 15 arc resolution, which results in improvement in the spatial resolution (~ 6km) and accuracy of gravity predicted bathymetry. Source Identifi er (SID) Grid supports bathymetry grid and identifi es source of data used to calculate the depth within the grid cell ( Figure 2d). In the test area the data soundings are obtained from NGA, GEBCO and SIO databases, where SIO data is only publicly available.
Smith and Sandwell bathymetry model (SS) released in 1996 was the fi rst bathymetry model derived from satellite gravity data (track spacing 2-4 km) and sparse in situ soundings (hundreds of kilometres between tracks in some areas) [38]. The model was covering ocean area between ±72° latitude with uniform 2-minute grid spacing revealing new geological structures and unknown topography of ocean fl oor with 12.5 km spatial resolution. The mathematical base of the model is Smith-Sandwell method which defi nes correlation between marine gravity anomalies and changes in sea fl oor topography [37]. of Murter sea (Figure 1b) is used for grid comparison. HRDBM is interpolated to a grid with 2m resolution from the multibeam survey, done in 2018 following IHO S44 standard [26] with Kongsberg EM2040 multibeam echo sounder operated using Kongsberg's seafl oor information system (SIS).

Methods / Metode
In this study digital bathymetry models (DBM) are analysed using QGIS software, Open Source Geographic Information System, version 3.10 A Coruna. DBM are analysed over the area of continental shelf of Croatia. General statistics (mean depth ME, median depth MED, maximum depth MAX, standard deviation STDEV) for all models were calculated at their original resolution and projection using the "Zonal Statistics" tool in QGIS. Marine geomorphometry is a science of quantitative analyses of seafl oor focused on characterization of seabed terrain [31]. Hypsometry is widely used geomorphometric parameter that refers to elevation or depth relative to chosen zero level. Hypsometric curve represents the area distribution of depth over specifi ed area. The tool hypsometry curve in QGIS was used to derive the graph for all publicly available digital bathymetry models at 10-meter depth interval. For this operation all grids were transformed and projected to Lambert azimuthal equal area projection with the parameters specifi ed in the European Terrestrial Reference System (ETRS) 1989 recommended by the EU INSPIRE Directive for statistical analyses of data spanning large parts of Europe when true area representations are required [25]. Projection parameters can be found by the European petroleum survey group (EPSG) code 3035 in most GIS software including QGIS.
In order to mutually compare digital bathymetry models of diff erent resolution, it is necessary to resample them to identical grid size [5,43]. All models were resampled to 1-minute grid using bilinear interpolation model in their original datum for two reasons: it is more reliable to convert from higher to coarser resolution and secondly, almost all model to some extent rely on data from SS model that has 1-minute resolution. Absolute diff erences between models are used as statistic in comparison because emphasis is on the magnitude of the diff erences and not their sign. Apart from the most commonly used statistical measures mean (ME) and standard deviation (STDEV), robust statistical methods median (MED) and Mean absolute deviation (MAD) were calculated (Table 3). Publicly available models use data from diff erent sources with unknown quality and robust statistical methods are more reliable for data which may have systematic and gross errors [16].
Bathymetric profi le in the area of Murter sea was generated from high resolution digital bathymetry model HRDBM and compared to profi les between the same points generated from publicly available models using Profi le tool plugin in QGIS.
Methods used for analyses and comparison of digital bathymetry models are graphically presented in Figure 3.

RESULTS AND DISCUSSION / Rezultati i diskusija
Results of mutual comparison of grids as described in fl owchart in Figure 3 are presented and analysed in the next section.

Basic feature / Osnovne značajke
General statistics of tested digital bathymetry models are given in Table 2. Average depth of tested digital bathymetry models in continental shelf of Croatia ranges from 207 m (ETOPO) to 226 m (DTU) and maximum depth ranges from 1181 m (ETOPO) to 1277 m (DTU). It is generally accepted that the deepest point of Adriatic is nearly 1240 m in South Adriatic Pit (SAP) [29,46,47] so diff erences between these values suggest presence of outliers in models. GEBCO, SRTM15+ and Smith and Sandwell (SS) grids are based on the same source data and their compatibility is visible from almost identical statistic values of grids. As compared to those three grids, DTU, EMODnet and ETOPO have diff erent statistic values that are unique for each grid. It was not expected for EMODnet grid to diff er from GEBCO 2020 grid since most of EMODnet data in the area is based on GEBCO 2014 grid. This indicates one can expect noticeable diff erences not only between diff erent bathymetry models but also between diff erent versions of grids.  Figure 3 Flowchart overview of the analyse of digital bathymetry models [5,16,43] Slika 3. Dijagram tijeka analize digitalnih batimetrijskih modela [5,16,43]

Geomorphometry / Geomorfometrija
In order to represent distribution of depth relative to the area they are covering, common geomorphometry function hypsometric curve was derived for all grids with the area calculated for 10-meter depth intervals (Figure 4). Greatest diff erences between models are observed in shallow areas with a depth of up to 250 meters which cover more than 85% of continental shelf of Croatia ( Figure 5) and in the area of South Adriatic Pit (SAP), deeper than 1200 meters. DTU and ETOPO models show high peak in depth range from 0 m to 25 m and cover much more area in this depth range than other models but there is a loss of area in range from 25 meters to 100 meters as compared to others. This can be explained by their source data. DTU is based on GEBCO 1-minute grid which has a depth defi cit in coastal area that is represented with depth of only 5 meters (Figure 2a) [33] noticed that a bias towards gridded digitized depth or contours in source data or fl at sediment area could be seen as spikes in the hypsometric curve. As test area is not fl at, but highly irregular due to well-indented coast, spikes are caused by source data. In depth range from 250 m to 1200 m hypsometric curves are uniform. Observable change happens in deepest area between 1200 m and 1250 m for SRTM15 and SS models. Although base layer for GEBCO 2020 is SRTM15+ model, in deepest part of continental shelf which covers part of South Adriatic Pit, GEBCO is augmented with EMODnet data based on multibeam survey.

Absolute diff erences between models / Apsolutne razlike između modela
Grids were resampled to same grid spacing of 1-minute using bilinear interpolation for pixel to pixel comparison. Absolute diff erences between grids were calculated in all combinations (Table  3). Absolute diff erences were classifi ed in three depth intervals to get better representation of relative coherence. It should be noted that diff erences between models are calculated in shallow test area with 85 percent of depth under 250 meters ( Figure 5). DTU and ETOPO show the highest value of absolute diff erences in combination with all other models. Diff erence between DTU and ETOPO is presented in Figure 6a with biggest discrepancies in South Adriatic Pit as a result of diff erent source   DTU  15  13  31  37  251  61  33  6  DTU  SS  15  13  30  35  236  61  33  5  DTU  EMODNET  14  12  26  30  200  63  34  3  SRTM15+  ETOPO  12  11  34  48  381  64  28  8  DTU  GEBCO  12  11  29  36  261  63  31  6  ETOPO  SS  11  10  34  49  395  65  27  8  ETOPO  GEBCO  10  9  32  48  387  67  24  9  DTU  ETOPO  8  8  25  41  509  73 (Figure 6b), great inconsistency between models happens in coastal area. This is due to diff erence in source data but also to limitations of satellite altimetry near land area and consequently degradation of the accuracy of gravity predicted depths. GEBCO, SRTM15+ and Smith and Sandwell (SS) grid show best coherence with more than 95 percent of diff erences smaller than 25 meters. SRTM15+ and SS model have relatively best fi t with mean diff erence of 5 meters, standard deviation of 8 m and 99 per cent of absolute diff erences smaller than 25 m. SRTM15+ and SS grid share completely identical source data but small diff erence between them is due to the diff erence in interpolation and resolution. GEBCO, whose base layer is SRTM15+, shows slightly greater diff erence as compared to SRTM15+ and SS with 97 per cent of depth diff erences smaller than 25 but this can be explained with EMODnet data incorporated into grid ( Figure 6c). As mentioned in section 3.1, EMODnet bathymetry model surprisingly shows diff erence between GEBCO model with maximum absolute diff erence of 176 m (Figure 6d). In test area EMODnet grid is mostly based on GEBCO 2014, and GEBCO 2020 has merged EMODnet direct soundings into his grid. However, base layer of GEBCO 2014 is GEBCO08 (SRTM30+V.5) and of GEBCO 2020 is SRTM15+V2.1. Discrepancy between models is due to diff erent version of base layer. Bjelotomić [5] in 2015 mutually compared digital bathymetry models in the Adriatic Sea for a specifi c task in physical geodesy, computing a geoid. Among others DTU10BAT, ETOPO 1, GEBCO 2014, Smith and Sandwell Version18.1 and SRTM 30+. Although results cannot be directly compared due to diff erence in test area and version of models, some correlation can be made. Same diff erences between DTU10BAT and ETOPO observed in Figure  6a can be seen in Bjelotomić. Smith and Sandwell and SRTM+ model always show great compatibility due to the same source data. GEBCO 2014 had great shortcoming in shallow coastal area as compared to other models in Bjelotomić, but for GEBCO 2020 this is not the case and it shows great improvement in most recent version. As discussed in earlier studies [20,21,24], diff erences between models are caused by several factors: density and distribution of underlying source data, interpolation method used for grid construction and resolution of grid.

Comparison with high resolution digital bathymetry model (HRDBM) / Usporedba s digitalnim modelom visoke rezolucije
Bathymetry grids were compared with high resolution digital bathymetry model HRDBM with 2 m grid spacing calculated from multibeam data in the area of Murter Sea (Figure 7a). Area of Murter Sea is coastal shallow area with well indented coast and depths in range from 0 m to 180 m. Vertical profi le AA′ was generated for all bathymetry grids. Diff erence in resolution is the fi rst thing to be noticed by visual inspection of digital grids. ETOPO, DTU and Smith and Sandwell (SS) have coarser resolution as compared to other grids and that is evident by visual interpretation. Furthermore, DTU and ETOPO show great defi ciency of depth in the area as seen in Figure 7b and 7d. Both models have a constant depth of about 5 m as can be seen in their vertical profi les (Figure 7h). They do not represent the seafl oor terrain and are inaccurate for the area. As for other models, they all have artifacts but represent sea fl oor far better than DTU and ETOPO (Figure 8a -8d).
Compatibility between GEBCO, SRTM15+ and SS is evident from their vertical profi les ( Figure 7h). As compared to multibeam data it can be noticed that SS has the worst alignment of those three grids and cannot depict small shoals and reefs as GEBCO and SRTM15, but this is probably due to his coarser resolution. It was expected from EMODnet to show the best fi t because it is the grid with highest resolution that should be based on survey data from authoritative sources. Unfortunately, in the area it is based mostly on GEBCO2014 data that is inferior to GEBCO2020 when compared to multibeam data. EMODnet seems much smoother than GEBCO and SRTM15+ and does not depict small reefs and shoals as well as they do, due to source data and interpolation method.

CONCLUSION / Zaključak
Analyses and mutual comparison between six publicly available grids in the area of Croatian continental shelf reveals that for marine researchers GEBCO2020 digital bathymetry model would, at the present moment, be the best choice. Although EMODnet grid has four times higher resolution of 3.75″ and is based on surveys from authoritative sources, gaps between these data are fi lled with GEBCO 2014 grid. Unfortunately, in the test area, there are huge gaps covering most of the area and GEBCO2014 (base layer SRTM30+) bathymetry showed shortcoming of depth in shallow areas near coast. GEBCO2020 is superior to EMODnet digital bathymetry model (DBM) for two reasons, it is the latest, updated version of GEBCO grid based on SRTM15+ model and it is augmented with direct soundings from EMODnet grid in the test area. The fact that surveys from EMODnet data are part of GEBCO grid, especially in the deepest area of South Adriatic Pit makes GEBCO grid more reliable than its base layer SRTM 15+. Smith and Sandwell (SS) one-minute grid diff ers from SRTM15+ digital bathymetry model with 15 seconds grid spacing, only in resolution because they are based on the same source data. DTU and ETOPO grid show great depth defi ciency in shallow coastal area and have the smallest alignment with other models. As digital bathymetry models, especially SS, SRTM+ and EMODnet are regularly updated with new data there are signifi cant diff erences between diff erent versions of models and the latest one should be used. Diff erences between models are due to density and distribution of underlying source data, interpolation method used for grid construction and resolution of grid. For that reason, describing source data is a key element for users of bathymetry data to understand the context in which the data has been acquired, how it has been processed, and its expected quality. Most of publicly available bathymetry grids rarely have quality indicators as their quality is expressed with statistical value of adequacy (tipically Root Mean Square Error) between them and reference data on global level or simply indicate the origin of soundings by accompanied grids. EMODnet is the only grid off ering a detailed description of its quality at the geographical level, providing information about survey method, data quality and link to the data holder through metadata. Regarding the fact that it is almost impossible to obtain information about the quality of source data at present stage, for a quality assessment of publicly available grids and statistically determining which grid is the best fi t in particular marine region, grids should be compared with reliable soundings from nautical charts or survey data.

Aknowledgment / Zahvale
Special thanks to Hydrographic Institute of the Republic of Croatia for the permission to use bathymetry data of Murter Sea for scientifi c and research purposes.