Ruan, X., Cheng, L., Chu, S., Yan, Z., Zhou, X., Duan, Z., & Li, M. (2020). a New Digital Bathymetric Model of the South China Sea Based on the Subregional Fusion of Seven Global Seafloor Topography Products. Geomorphology, 370, 1

Geomorphology 370 (2020) 107403
Contents lists available at ScienceDirect
Geomorphology
journal homepage: www.elsevier.com/locate/geomorph
A new digital bathymetric model of the South China Sea based on the
subregional fusion of seven global seafloor topography products
Xiaoguang Ruan a,b,c, Liang Cheng a,b,c,d,e,⁎, Sensen Chu a,b,c, Zhaojin Yan a,b,c, Xiao Zhou a,b,c,
Zhixin Duan a,b,c, Manchun Li a,b,c,d,e
a
Jiangsu Provincial Key Laboratory of Geographic Information Science and Technology, Nanjing University, Nanjing 210023, China
b
Collaborative Innovation Center for the South Sea Studies, Nanjing University, Nanjing 210093, China
c
School of Geography and Ocean Science, Nanjing University, Nanjing 210023, China
d
Collaborative Innovation Center of Novel Software Technology and Industrialization, Nanjing University, Nanjing 210023, China
e
Jiangsu Center for Collaborative Innovation in Geographical Information Resource Development and Application, Nanjing 210023, China
a r t i c l e i n f o a b s t r a c t
Article history: An understanding of ocean bathymetry is important in marine planning, navigation, military activities, and envi-
Received 19 October 2019 ronmental monitoring. The fusion of spatial data, such as those from multi-source digital bathymetric models
Received in revised form 1 September 2020 (DBMs), can effectively improve the efficiency of large-scale seafloor topographical research and the accuracy
Accepted 1 September 2020
of the results obtained therein. On the one hand, with the release of 15arc-seconds resolution DBMs, it is now
Available online 6 September 2020
necessary to verify the quality of these new products. On the other hand, in order to generate a new, high-
Keywords:
quality, seamless DBM in the South China Sea (SCS) and adjacent areas, an adaptive subregional spatial-
Bathymetry domain-weighted fusion framework is proposed. First, seven of the most widely used global DBMs are selected,
Digital bathymetric model (DBM) and multi-source subregional measured depth-sounding data undergo data cleaning and other preprocessing.
Seafloor topography Next, based on the homogeneity of the terrain features, an adaptive subregional topographical analysis is
Multi-source data fusion performed, and the subregional data are weight-fused. Finally, the fusion dataset is post-processed via model
Weighted fusion smoothing and other procedures. In addition, the advantages and limitations of the DBMs of the SCS are com-
South China Sea (SCS) pared. The results show that SRTM15_PLUS V2 is the most reliable of the original DBMs. The updated seamless
SCS DBM is void-free and more similar to SRTM15_PLUS V2 with a resolution of 15arc-seconds. The root mean
square error (RMSE) of the new model is 99.60 m. Its accuracy is 13%, 40%, 15%, and 1% higher than those
achieved by the GEBCO_2019, GEBCO_2014, SRTM30_PLUS, and SRTM15_PLUS models, respectively, and its ex-
pression of the topography is more detailed and realistic. The feasibility and limitations of the proposed fusion
framework are demonstrated. The present findings provide a useful reference for the timely reconstruction
and updating of large-scale seafloor topography from multiple datasets.
© 2020 Elsevier B.V. All rights reserved.
1. Introduction geological disaster assessments (Chiocci et al., 2011). By monitoring

changes in the seafloor over time, oceanographers can predict the spread
The world's oceans cover approximately 362 million square kilome- of tsunami waves and other natural disasters (Sepúlveda et al., 2019). The
ters, accounting for 71% of Earth's surface area (Ronov, 1994; Eakins and development of marine infrastructure also requires high-resolution DBM
Sharman, 2010). Many studies employ digital bathymetric model data. data (Ellis et al., 2017). DBM data are helpful for search and rescue oper-
Global seafloor topography provides detailed information regarding ations; for example, in the area where Malaysia Airlines flight MH370
Earth's geological and ocean processes, such as plate tectonics and the crashed, the lack of high-resolution bathymetry data greatly hindered
evolution of the ocean basins (Sandwell and Smith, 1997). Accurate the search and rescue efforts (Picard et al., 2018). An understanding of ba-
ocean circulation modeling increasingly relies on fine bathymetric models thymetry is critical for navigation safety and determining the limits of the
(Gula et al., 2014). Bathymetric data provide the basis for marine continental shelf (Jakobsson et al., 2003). Depth, slope, aspect, curvature,
and terrain variability characteristics greatly influence the distribution of
⁎ Corresponding author at: Jiangsu Provincial Key Laboratory of Geographic Information benthic organisms (Wilson et al., 2007) and are necessary for the protec-
Science and Technology, Nanjing University, Nanjing 210023, China. tion of the marine environment (Rengstorf et al., 2014). Many countries
E-mail addresses: [email protected] (X. Ruan), [email protected] (L. Cheng),
[email protected] (S. Chu), [email protected] (Z. Yan),
and international organizations have realized the importance of establish-
[email protected] (X. Zhou), [email protected] (Z. Duan), [email protected] ing global DBMs and have invested significant effort toward this goal (UN,
(M. Li). 2015; Jakobsson et al., 2017).
https://doi.org/10.1016/j.geomorph.2020.107403
0169-555X/© 2020 Elsevier B.V. All rights reserved.
2 X. Ruan et al. / Geomorphology 370 (2020) 107403
However, a detailed understanding of the bathymetry of most of the datasets must be validated and processed in advance due to possible
global seafloor is still lacking, and only about 6.2% has been mapped at voids and outliers.
fine spatial resolution (≤800 m) using shipboard sonar methods The rest of the paper is organized as follows. Section 2 gives a de-
(Mayer et al., 2018). Indeed, our understanding of other planets, tailed description of the data sources used. In Section 3, we provide
moons, and asteroids exceeds that of the seafloor (Smith et al., 1999; the details of the processing methodology. The results and discussion
Willner et al., 2014). Assembling regional and global gridded DBMs re- are given in Section 4. Finally, Section 5 presents the Conclusions.
mains extremely challenging (Hell and Jakobsson, 2011). Through the
continuous refinement and improvement of bathymetric data from 2. Data sources
ships, marine gravimetric data, etc., several versions of publicly avail-
able global DBM datasets have become available. GEBCO (Jakobsson In this study, we integrate seven global seafloor topography products
et al., 2017), SRTM (Becker et al., 2009), Global Seafloor Topography for the generation of a new digital bathymetric model in the SCS. The to-
(hereinafter referred to as TOPO) (Smith and Sandwell, 1997), ETOPO pography products include the GEBCO, SRTM, SIO, ETOPO1, and DTU10
(Amante and Eakins, 2009), and DTU (Andersen et al., 2008) are the datasets. The measured bathymetry data sources include ship depth-
most widely used datasets. These datasets have proved to be fundamen- sounding data and navigational charts. Our studied geographic extent
tally important in geoscience-related research (see Section 2.2 for a de- ranges from 0.30°S to 23.80°N and 99.17°E to 121.28°E. The specific
tailed discussion). characteristics of the data sources are provided in the following sections.
Similar to the different types of digital elevation models (DEMs),
there are also different types of DBMs, which vary in their accuracy and
originate from a variety of sources (Yue et al., 2015). Many studies are fo- 2.1. Study area and measured bathymetry data sources
cused on generating new DBMs with higher accuracy by supplementing
them with data from other sources, using methods such as inverse 2.1.1. Study area
distance weighting (IDW) (Schaffer et al., 2016), kriging (Merwade, The SCS is a semi-open sea surrounded by mainland China, Taiwan,
2009), spline interpolation (Hell and Jakobsson, 2011), bilinear interpo- Vietnam, Malaysia, Indonesia, and the Philippines. Its seabed topogra-
lation (Ramillien and Cazenave, 1997), and natural neighborhood inter- phy is extremely complex, with continental shelf at the bottom edge
polation (Glenn et al., 2016). In the fusion of DEMs, problems such as and deep-sea basin at the center, and many scattered seamounts, atolls,
noise, data voids, and lack of resolution arise (Yue et al., 2015). Common and islands (Hsiao et al., 2016). A regional seafloor digital terrain model
DEM data fusion methods include frequency domain fusion, spatial do- and the regional seafloor geomorphic features (Harris et al., 2014) near
main fusion, and sparse representation fusion. For example, SRTM and the Spratly Islands and the Paracel Islands are shown in Fig. 1(c)–(d).
ASTER GDEM data have been fused in the frequency domain to fill data We used the SCS and its adjacent areas as the study area (Fig. 1(a)).
voids and improve the overall accuracy of the fused DEM (Karkee et al., The coastline data (including the nearshore harbor boundary and ex-
2008). Multi-scale modeling has been adopted to fill in voids in high- posed islands and reefs) are from the Global Self-consistent, Hierarchi-
resolution DEM data (Jhee et al., 2013). Satellite altimeter data have cal, High-resolution Shoreline (GSHHS) data. The data (https://www.
been used to improve the accuracy of DEM data in a void area, and the ngdc.noaa.gov/mgg/shorelines/gshhs.html) library is maintained by
TIN delta surface method can be used to fill in voids (Yue et al., 2017). the National Centers for Environmental Information (NCEI). The na-
The DEM fusion framework, based on sparse representation, can effec- tional district boundary data are from the Global Map Data maintained
tively suppress improbable local surface geometric abnormalities caused by the United States Geological Survey (USGS) (http://nationalmap.
by errors in the fusion process (Papasaika et al., 2011). Spatial domain fu- gov/small_scale/atlas-ftp-global-map.html).
sion is generally carried out directly at the numerical level (Costantini
et al., 2006). Practice has shown that this method is simple, easy to use, 2.1.2. Ship depth-sounding
and often provides good results (Gruber et al., 2016; Petrasova et al., The ship depth-sounding data for the SCS are primarily from the Ma-
2017). However, it is difficult to determine the input weights needed rine Trackline Geophysics database of the NCEI (bathymetric survey re-
to quantify the influence of the different DEM datasets at each surface lo- cords are available at http://www.ngdc.noaa.gov/mgg/geodas/geodas.
cation (Schindler et al., 2011). In other words, the selection of weights is html). The database contains data on water depth, magnetism, gravity,
even more important than the choice of fusion algorithm. and seismic reflections collected by government vessels and scientific
In general, the accuracy of generating DBM is related to the distribu- vessels during ocean navigation from 1939 to the present. The database
tion of measured bathymetry data (Weatherall et al., 2015). Existing is stored at the US National Geophysical Data Center (NGDC, now NCEI).
studies tend to focus on supplementing a regional DBM with newly MGD7/MGD77T is a data exchange format between the NGDC and
acquired bathymetry data. Few studies have used multiple datasets to re- its international marine agency that contains document headers and
construct and update large-scale reliable seafloor topography in a timely data records. The header records provide metadata for each measure-
manner. The resolution obtained by existing DBMs is from about 30 arc- ment (including SURVEY IDENTIFIER, COUNTRY, CRUISE, and SURVEY
seconds to 1 arc-minute. With the recent release of 15 arc-second reso- DEPARTURE DATE), and the data records contain geophysical data at
lution DBMs, it is now necessary to compare and verify the quality of different times and locations (including SURVEY IDENTIFIER, DATE,
these new products (see Section 2.2 for details). In addition, because TIME, LONGITUDE, LATITUDE, CORRECTED DEPTH, and QUALITY CODE
the accuracy of the DBM datasets may vary with changes in topographic FOR BATHYMETRY) (Wessel and Chandler, 2007). The data used in
conditions (Yang et al., 2018), the DBM can be partitioned and fused this study are primarily single-beam sonar data, but they also include
under the constraint of measured depth-sounding data. We propose an a small number of gridded multi-beam sonar data. We used the
adaptive subregional spatial-domain-weighted fusion framework for un- MGD7/MGD77T tool, based on the Generic Mapping Tools (GMT), to
derwater terrain restoration in the South China Sea (SCS) and its adja- convert the ship depth-sounding data in the MGD77T format into the
cent areas using multiple DBMs. The aim is to integrate multi-source XYZ ASCII data format (Wessel, 2010). The distribution of ship depth-
sounding and DBM datasets into a new, high-quality, seamless SCS sounding data in the study area is shown in Fig. 1(b).
DBM dataset. Our fusion framework includes data editing, recovery of
new measured bathymetry data, weighted fusion, model smoothing, 2.1.3. Navigational charts
and other processing. Following the fusion process, the measured ba- Navigational charts are widely used in navigation, marine engineer-
thymetry data are actually used to enhance global DBM datasets, so as ing construction, marine surveys, marine resource development, mari-
to overcome the insufficiency of the measured bathymetry data. In the time delimitation, and marine military activities. Charts have been
process, the newly released global 15 arc-second GEBCO and SRTM used by many researchers to evaluate navigation data and can be used
X. Ruan et al. / Geomorphology 370 (2020) 107403 3
Fig. 1. Study area and distribution of the measured bathymetry data. (a) Study area; (b) Ship depth-soundings; (c) Regional seafloor digital terrain model near Spratly Islands and Paracel
Islands in the SCS, ranges from 8.13°N to 18.24°N and from 109.90°E to 120.19°E (vertical exaggeration factor 10); (d) Regional seafloor geomorphic features (adapted from Harris et al.,
2014).
as supplementary data when studying underwater terrain in shallow 2.2. Digital bathymetric models
waters (Pe'eri et al., 2014; Liu et al., 2014).
Commonly used chart formats include Electronic Navigational 2.2.1. GEBCO datasets
Charts (ENCs) and Raster Navigational Charts (RNCs). The ENC data The General Bathymetric Chart of the Oceans (GEBCO), overseen by
are collected and compiled by government agencies in strict accor- an international group of cartographic experts and co-sponsored by the
dance with the International Hydrographic Organization (IHO) S- International Hydrographic Organization (IHO) and UNESCO's Intergov-
63 data format (an encrypted S-57 format data) and are accessed ernmental Oceanographic Commission (IOC), contains a range of bathy-
via the Electronic Chart Display and Information System (ECDIS) metric datasets and data products. The GEBCO_2014 grid (Fig. 2(b),
(Wlodarczyk-Sielicka and Stateczny, 2016). RNCs have expressions Table 1), which is a continuous terrain model of the ocean and land
and styles similar to traditional paper charts and are updated weekly with a spatial resolution of 30 arc-seconds, was established in 2014. The
to ensure the accuracy of the chart information. Although raster ocean data were interpolated using ship bathymetric data guided by sat-
maps have some shortcomings, they are a good supplement that ellite gravity data (Weatherall et al., 2015). The existing datasets and data
can be used in conjunction with ENCs. The Navigation Guarantee assembly grid were used in the establishment of the model. Most of the
Department of the Chinese PLA Navy Headquarters (2019) has pub- land data are obtained from the SRTM30 grid digital elevation model.
lished a standard navigational chart of the SCS region at a full range The data for the Arctic region north of latitude 64° are from the Interna-
of scales in accordance with the relevant regulations of the People's tional Bathymetric Chart of the Arctic Ocean (IBCAO), and the data for
Republic of China Surveying and Mapping Law. The series comprises the Southern Ocean south of latitude 60° are obtained from the Interna-
more than 400 charts. tional Bathymetric Chart of the Southern Ocean (IBCSO). The initial
The ENCs and RNCs used in this study were obtained through Beijing GEBCO_08 grid was co-developed by organizations such as GEBCO, the
Situo Ocean Information Technology Co., Ltd. (www.enclive.cn), autho- International Hydrographic Bureau (IHB), the US National geospatial-
rized by the Chinese Naval Hydrographic Survey (www.ngd.gov.cn). intelligence Agency (NGA), and others. When new datasets are imported,
The positions of the objects on the chart are based on the World Geo- the “Remove-Restore” method is used to continuously edit and update
detic System 1984 (WGS-84). The scale is 1:250,000. the model (Smith and Sandwell, 1997). In April of 2019, GEBCO (2019)
All measured bathymetry data were obtained during 2019. released the latest version of the grid, the GEBCO_2019 grid (Fig. 2(a),
Fig. 2. Bathymetry of the SCS at 15 arc-second resolution. (a) GEBCO_2019 Grid; (b) GEBCO_2014 Grid; (c) SRTM15_PLUS V2; (d) SRTM30_PLUS V11; (e) TOPO V19.1; (f) ETOPO1; and
(g) DTU10. Positive values represent land or reef.
Table 1), a model with a grid spatial resolution of 15 arc-seconds jointly 2.2.2. SRTM datasets
developed by the Nippon Foundation-GEBCO Seabed 2030 Project The SRTM30_PLUS dataset was developed by SIO (Scripps Institution
(2019). The data primarily derive from SRTM15_PLUS version 1 (V1), of Oceanography) and has a spatial resolution of 30 arc-seconds. The
and the dataset is compiled by the four Seabed 2030 Regional Centers. land data are based on topographic data with an average resolution of
The editing method is the same as that used in the previous versions. In 1 km generated by the USGS SRTM30 grid DEM data product. The
this study, GEBCO datasets were obtained in 2019. ocean data are based on the latest Cryosat-2 and Jason-1 gravity models
Table 1
Sources of DBMs.
Dataset (DBMs) Update deadline Organization Organizers Resolution Coverage Horizontal/vertical datum
GEBCO_2019 2019 The Nippon Foundation-GEBCO UK-Japan 15″ −179°59′52.5″–179°59′52.5″; WGS84, MSL
−89°59′ 52.5″–89°59′52.5″
GEBCO_2014 2015 GEBCO USA 30″ −179°59′45″–179°59′45″; WGS84, MSL
−89°59′45″–89°59′45″
SRTM15_PLUS V2 2019 SIO USA 15″ −180°–180°; −90°–90° WGS84, MSL
SRTM30_PLUS V11 2014 SIO USA 30″ −180°–180°; −90°–90° WGS84, MSL
TOPO V19.1 2019 SIO USA 1″ −180°–180°; −90°–90° WGS84, MSL
ETOPO1 2009 NCEI USA 1″ −180°–180°; −90°–90° WGS84, MSL
DTU10 2010 DTU Space Denmark 1″ −180°–180°; −90°–90° WGS84, MSL
and are calibrated for the ratio of gravity to topography using 298 mil- In this study, ETOPO1 for 2019 was obtained from: https://www.ngdc.
lion ship-sounding data points. The main contribution of this dataset noaa.gov/mgg/global/global.html.
is the compilation and correction of the original sounding points from
NOAA, SIO, NGA, JAMSTEC, IFREMER, GEBCO, NAVOCEANO, and other 2.2.5. DTU10 dataset
scientific institutions and scientists (Becker et al., 2009). SRTM30_PLUS The Technical University of Denmark released the DTU10 bathyme-
V11.0 (Fig. 2(d), Table 1) is available at ftp://topex.ucsd.edu/pub/ try model in 2010 (Fig. 2(g), Table 1), a global topography model with a
srtm30_plus/. Its spatial resolution is 15 arc-seconds. SRTM15_PLUS resolution of 1 arc-minute based on satellite altimetry data and the
V1.0 is based on an improved SRTM30_PLUS V11.0. Its land data origi- GEBCO model (Andersen et al., 2008; Knudsen et al., 2011). The model
nate from a combination of SRTM, ASTER, and the CryoSat-2 ice sheet uses multi-beam and single-beam sounding data. In some sea areas
topography. The ocean data are the same as those of SRTM30_PLUS where ship survey data are scarce, e.g., data based on the altimetry
and include 494 million ship soundings. The inaccuracies along the data of ERS-1, Geosat, and other altimetry satellites, gravity anomaly
edges have been corrected (Olson et al., 2014). SRTM15_PLUS V2.0 inversion is used to calculate the sea floor topography in the band of
(Fig. 2(c)) was released on March 29, 2019, available at: ftp://topex. 20–120 km. In addition, the bathymetry obtained from bands in
ucsd.edu/pub/srtm15_plus/. In this study, SRTM datasets were obtained the GEBCO model, land topography, and bathymetry shallower than
in 2019. 100 m is used as auxiliary data.
In this study, the DTU10 dataset for 2019 is available at: https://
2.2.3. SIO datasets www.space.dtu.dk/english/research/scientific_data_and_models/
David Sandwell of the Scripps Institution of Oceanography (SIO) and global_bathymetry_model.
Walter HF Smith of the Laboratory for Satellite Altimetry, NOAA, have
long worked to reconstruct ocean bed topography using satellite altim- 3. Methodology
etry and shipboard bathymetry data, such as those from CryoSat-2,
Jason-1, Envisat, Geosat, and ERS-1(Smith and Sandwell, 1997; First, all data sources are preprocessed, including chart vectorization,
Sandwell et al., 2014). The team often updates Global Seafloor Topogra- basic unification, and data cleaning. After preprocessing, the quality of
phy (TOPO) datasets. With the development of the plate theory model the DBMs is compared qualitatively and quantitatively. Finally, an adap-
and the abundance of ship sounding data, the team has released 15 ver- tive subregional spatial-domain-weighted fusion framework is pro-
sions of the Global Seafloor Topography model so far. Each version adds posed. The specific fusion strategy is explained in Section 3.3. The
or optimizes the measurement data, updates the satellite gravity model specific details of the methodology are described below.
data, or optimizes the inversion algorithm parameters. In 2019, the
team released the latest version of TOPO V19.1 (Fig. 2(e), Table 1), 3.1. Preprocessing
which has a spatial resolution of 1 arc-minute. TOPO V19.1 data sources
include SRTM Topography, NGDC, GEODAS, the SIO Multibeam, and 3.1.1. Chart vectorization
others. The predicted bathymetry comes from the V23 global gravity Vectorizing the RNCs first involves digital scanning, and geometric
model, which includes the latest satellite altimetry data from datasets correction is carried out on the scanned image to create point vector
such as Cryosat-2, Jason-1, and Envisat. It is available at ftp://topex. files and mark the geographical location of the soundings. Then, the
ucsd.edu/pub/global_topo_1min/. In this study, TOPO V19.1 was ob- area containing soundings in the chart image is read, the soundings are
tained in 2019. counted, a visual inspection is conducted, and finally the vectorization
of the paper chart is performed. Considering their uncertainty, bathyme-
2.2.4. ETOPO1 dataset try contours were not used for sounding extraction. In this process, we
The latest version of ETOPO1, released by NGDC in 2009, is a global ensured that the soundings of the chart have a uniform geographical
topographic relief model with a resolution of 1 arc-minute (Fig. 2(f), reference and saved them as files in the XYZ ASCII data format.
Table 1) (Amante and Eakins, 2009). It combines land and seafloor to-
pography and is built from global and regional datasets. Its accuracy is 3.1.2. Basis unification
significantly higher than that of the previous ETOPO2 V2 and ETOPO5. Unification of the mathematical bases includes unification of the co-
There are two versions of the model: “Ice Surface” (top of Antarctic ordinate datum reference, vertical datum reference, and the spatial res-
and Greenland Ice sheets) and “base of the Ice sheets,” which are the olution. Safe Software's (http://www.safe.com/) Feature Manipulation
same in all areas except for Antarctica and Greenland. The dataset for Engine (FME) data translation tool package was used to convert the co-
the ETOPO1 global topography includes those from the Japan Oceano- ordinate datum vector bathymetry data reference into the WGS-84 co-
graphic Data Center (JODC), NGDC, the Caspian Environment Pro- ordinate system. It converted the data into the Esri (http://www.esri.
gramme (CEP), and the Mediterranean Science Commission. The com/) ArcGIS shape files for subsequent data quality evaluation and
dataset also includes data from SRTM30 topography, and GLOBE topog- data editing. The seven bathymetric models employed different types
raphy. Datasets from SIO, NGDC, and other institutions that integrate of raster data, such as netCDF, GeoTiff, Esri, and ASCII raster format. All
water depth and topography (such as TOPO and IBCAO) are also used. models were unified into the GeoTiff format. All raster and vector data
were projected into the WGS 1984 Web Mercator Auxiliary Sphere pro- removed according to the IHO Standards for Hydrographic Surveys
jection system using the World Geodetic System datum (WGS-84). (International Hydrographic Bureau, 2008). This process was repeated
China's sounding datum is the Lowest Normal Low Water (LNLW). In for the chart data. Finally, outliers were identified and removed accord-
the SCS and adjacent seas, the maximum difference between Mean Sea ing to the Pauta criterion. In other words, a measured bathymetry was
Level (MSL) and various local chart vertical data is less than 2 m, and considered an outlier if the absolute value of the residual difference
the global average difference is generally less than 1.5 m (Church et al., between it and its predicted depth was greater than three times the
2011). In addition, studies (Peng et al., 2013) have shown that long- standard deviation (SD). Our editing process replicated the method pro-
term global and SCS sea level rise is 1.8 ± 0.4 mm yr−1. It was assumed posed by Becker et al. (2009). The method has two important proper-
that the vertical datum at all water depths is MSL, which only slightly im- ties. The first relates to the outlier test along the navigation direction
pacted the experiment (Becker et al., 2009). However, the accurate cor- (route profile) (Fig. 3(a)–(c)). In the route profile, the Y-axis represents
rection of the tidal range is not the purpose of this study; in addition, the measured bathymetry (black) and the model bathymetry (green) in
there is a lack of data from tidal stations in the high seas. Therefore, we units of meter, and the X-axis represents the number of measured
did not correct the elevation datum, because the ship depth-sounding points. The second feature relates to the testing of the outliers based
data are the same as those from the navigational charts. The GEBCO on the data statistics (Fig. 3(b)–(d)). In the statistical relation graph,
model, which is formed by integrating multi-source heterogeneous data the Y-axis represents the model bathymetry and the X-axis represents
types, was considered as an example of a bathymetric model. Except in the measured bathymetry, in m. The outliers are those data that deviate
local shallow water areas, the vertical datum corresponding to most of from the reference line, which has a slope of 1. We used the bathymetric
its data is MSL. In local shallow water areas, other vertical data, such as grid value from the SRTM30_PLUS model as the predicted depth and
Mean Lower Low Water (MLLW), may be used (Marks, 2016). Similarly, tested the data from all available valid cruises (a total of 118 cruises)
the vertical datum corresponding to the ETOPO1 model regional data, in- in the SCS. Experiments showed that outliers exist in 95 (80.51%) of
cluding MSL, Mean High Water, and Mean Low Water, varies, but the dif- the cruises (Fig. 3(e)), with a total of 19,450 outliers (3.16%). The results
ferences between the data are less than the vertical accuracy of ETOPO1 before and after data cleaning confirm the improved quality of the
(up to 10 m) (Amante and Eakins, 2009). measured points (Table 2).
The best spatial resolution of the bathymetrical models to be edited
and fused is 15 arc-seconds (approximately 450 m). Before fusion, the 3.2. Comparison among models
spatial resolutions of the seven models were unified to 15″. The original
spacing of the ship depth-sounding data is small (the minimum is 1 m), It is important to determine the accuracy and improve the quality of
which indicates that the spatial resolution is high. It was necessary to the DBMs by introducing the source ship-sounding and chart bathyme-
grid the original soundings such that each bathymetric grid (15″ × try data (Amante and Eakins, 2009). Before evaluating the models' qual-
15″) corresponds to only one value. If there were multiple points in a ity, the identities of the seven models were masked and their elevation
grid, the average of the multiple points was calculated. values <0 m were computed. The quality comparison of the DBMs was
carried out both quantitatively and qualitatively. First, the overall qual-
3.1.3. Data cleaning ity of the seven models' SCS results was quantitatively compared using
First, ship sonar data from the 1960s were screened for quality, and the measured data after processing. Next, a quantitative comparison of
invalid and erroneous values (correction_depth = null or ≥ 0 m) were the quality of the data profiles along each cruise was performed. Three
Fig. 3. Results of the tests for outliers. (a)–(b) Outliers from cruise V2109A; (c)–(d) Outliers from cruise VIT33B; (e) Outlier distribution. The red dots represent the outliers.
Table 2 uncertainties of model-predicted bathymetry vary in different locations,

Statistical values for sounding data before and after the elimination of the outliers. such as coastal and deep ocean areas (Tozer et al., 2019). Each of the
Cruise ID Sounding Slope Mean (m) Standard Pearson's seven DBMs has its advantages and disadvantages, and its quality in dif-
number error (m) r ferent regions may vary; thus, it was necessary to integrate and supple-
V2109A Before 247 0.87 −2468.97 1550.39 0.85 ment the models. An adaptive subregional spatial-domain-weighted
After 229 1.01 −2538.76 1538.08 1 fusion framework was then proposed (Fig. 4). First, the SCS terrain
VIT33B Before 445 0.93 −350.99 364.56 0.89 was segmented and classified into seven terrain subregions based on
After 437 1 −337.92 347.63 0.99
four terrain parameters. Next, an optimal spatial-domain-weighted fu-
sion with a minimum error constraint was proposed. Finally, a new
quantitative indices (Eq. (1)) were used to evaluate the vertical accu- DBM in the SCS (SCS DBM) was obtained, the quality of which was de-
racy of each model: Mean Absolute Error (MAE), Standard Deviation termined by the quality of the fusion model using the measured water
(SD), and Root Mean Square Error (RMSE). depth data after the smoothing process.
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u n u n 3.3.1. Terrain segmentation and subregions
1X n u1 X
t
u1 X 2
MAE ¼ j ei j SD ¼ ðei −eÞ RMSE ¼ t
2
e ð1Þ Studies have shown that the accuracy of a bathymetry inversion is
n i¼1 n i¼1 n i¼1 i
highly dependent on the topographic features. For example, the predic-
tion accuracy is greater for a region with a gentle change in slope than
where e is the difference between the corresponding model bathymetry for a mountainous area (Yang et al., 2018; Wan et al., 2019). In order
and the measured bathymetry, n is the number of the measured ba- to consider the contributions of the different terrain units to the
thymetry points tested, and is the mean of the errors. The MAE reflects model fusion, it was necessary to segment the target area.
the robustness of the accuracy without outliers. The SD is used to mea- The process of terrain segmentation and the determination of the
sure the degree of dispersion among a group of values. The RMSE more subregions is shown in Fig. 5. Multi-scale segmentation is one of the
accurately reflects the error distribution in the presence of outliers. In most popular remote sensing image segmentation algorithms (Costa
addition, in areas such as the Spratly Islands and Paracel Islands, the per- et al., 2018; Pont-Tuset et al., 2017). This method begins from the
formance of each model on the detailed underwater terrain was qualita- pixel layer and integrates the image into the larger object, layer by
tively compared by establishing bathymetry contour lines. layer, to produce segmentation results at different segmentation scales.
We utilized local variance (LV) to detect the optimal scale for segmenta-
3.3. Model fusion strategy tion of DBM objects (Drăguţ et al., 2014). The steps were implemented
in the Estimation of Scale Parameters tool of eCognition software. The
Almost all global DBMs are derived from satellite-derived gravity LV of the DBM object homogeneity under different scale parameters
model inversion. Due to the frequency and amplitude variability, the was calculated to determine the optimal segmentation. When the rate
Fig. 4. Flow chart of the proposed fusion strategy.

Fig. 5. General flow of terrain segmentation and determination of subregions.
of change (ROC) of LV reached its peak value, the corresponding seg- results demonstrate that the quality of the DBMs varies in different
mentation scale was deemed to be optimal (Fig. 6). In this case, both in- subregions.
ternal homogeneity and external heterogeneity were maximized. The
shape and compactness were chosen to be 0.1 and 0.5, respectively. 3.3.2. Subregional spatial-domain-weighted fusion
Terrain classification based on an elevation or bathymetry dataset The weighted average of the spatial domain is a simple and practical
has been studied (Drăguţ and Eisank, 2012; Walbridge et al., 2018). fusion algorithm that can suppress the noise in the original data
We established adaptive terrain subregion rules using four terrain pa- (Podobnikar, 2005). For the fusion of multi-source DEM data, the
rameters: bathymetry, terrain slope, terrain roughness, and surface cut weighted average sum is often effective (Schindler et al., 2011). The
depth. Using the terrain segmentation process, the segmented object measured points are considered as a reference to perform a linear
was divided into several terrain subregions. The definitions of the ter- weighted average of the original input data A(xi, yi), B(xj, yj) (Eq. (5),
rain parameters are as follows: Fig. 9).
Sc
Roughness ¼ ð2Þ zðx; yÞ ¼ wi zðxi ; yi Þ þ w j z x j ; y j ð5Þ
Sh
where Sc denotes the curved surface area of a surface element and Sh de- where w represents the weights of the input data.
notes its projected area on a horizontal surface. The fusion method determines the weight values to control the rela-
tive influence of two input DBMs in a fusion location. In this study, ergodic
Cut depth ¼ z−zmax ð3Þ
optimization weights were used for each subregion. The procedure
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi consisted of taking the RMSE minimum of the DBM and the measured
2 2
dz dz value as the constraint and searching for the optimal fusion result. The
Slope ¼ arctan þ ð4Þ
dx dy corresponding weights of each of the seven DBMs was obtained by tra-
versing the weight assignment. A brief flowchart is shown in Fig. 9.
where z denotes the average bathymetry in the neighborhood, zmax rep- First, the weight of DBMi was considered w and that of DBMj was as-
resents the maximum bathymetry in the neighborhood, and denotes sumed to be 1-w, where w was contained in the interval [0,1]. The design
the average bathymetry over all pixels in the neighborhood. calculation step was 0.01. Assuming the weighted coefficient was
In addition, in order to ensure that there were enough measurement between 0 and 1, the fusion optimal solution and weight coefficient
points in each subregion, seven terrain subregions (Zones 1–7) were were obtained via an ergodic process. The relationship between the
established in the research area. The terrain subregions are shown in RMSE and w in Zone 2 is shown in Fig. 10. The X-axis represents the
Fig. 7. The bathymetry varies slightly within each subregion, and the GEBCO_2019 weight w, and the Y-axis unit is m. At less than the optimal
segmentation and subregion results are basically consistent with the weight w = 0.23, the RMSE decreases with an increase in weight, which
benthic topographic features. For each terrain subregion, we subtracted indicates that GEBCO_2019 makes a positive contribution to the DBM
the measured values from the seven DBMs to determine the differences fusion. At values greater than the optimal weight w = 0.23, the RMSE
between them. The quantitative indices are shown in Fig. 8. These increases with increasing weight, e.g., the RMSE minimum = 73.22 m
Fig. 6. Terrain segmentation with the optimal scale value. LV represents local variance, and ROC represents the rate of change of the LV.
Fig. 7. Terrain subregions (Zones 1–7).
when the weight of SRTM15_PLUS is 0.77. This indicates that (3 × 3) was used to smooth the model. In order to avoid the edge effect,
SRTM15_PLUS in Zone 2 makes a greater contribution to the optimal the boundary of the study area was extended outward 30″ to create a
DBM fusion overall, i.e., the overall performance of SRTM15_PLUS is bet- buffer using SRTM15_PLUS V2.0 (Becker et al., 2009).
ter than that of the other models.
4. Results and discussion
3.3.3. Recovery of measured points
The measured data in the research area are primarily single-beam In this study, the performance of seven global seafloor topography
data; thus, the recovery of the actual measured points was not possi- products in the SCS is compared in detail and their advantages and dis-
ble in the classic “Remove–Restore” method (Becker et al., 2009; advantages are analyzed. An updated DBM in the SCS is obtained by fus-
Smith and Wessel, 2002). In order to generate a better fusion ing multi-source data. For a more comprehensive reliability analysis of
model, we developed a method to restore the measured points. The the updated DBM, a series of visual and quantitative evaluations are car-
procedure was to convert the measured points into a regular grid ried out. The feasibility and limitation of the proposed fusion framework
(15″ × 15″ resolution), remove the corresponding grid of the fusion are discussed.
model, and mosaic the two raster datasets to restore the true value.
The flow is shown in Fig. 11(a)–(d). 4.1. Model comparison
3.3.4. Model smoothing 4.1.1. Overall comparison

On the basis of the previous step, the adaptive filtering method of The bathymetries of the seven DBMs of the SCS are shown in Table 3.
neighborhood statistics was introduced to smooth the fusion model It can be seen that the performance of the models varies. GEBCO_2019
(Yue et al., 2017). The flow is shown in Fig. 11(e)–(h). The method has a larger range than GEBCO_2014. GEBCO_2019 extends to a depth
consisted of extracting a neighborhood patch (3 × 3) surrounding of 5408 m, which is deeper than the previous version. SRTM15_PLUS
each center pixel, calculating the difference between the neighborhood V2.0 is slightly different from SRTM30_PLUS V11. There are several
pixel and the center pixel, and classifying any center pixels with differ- voids in SRTM15_PLUS V2.0. This is due to the inclusion of new
ence greater than three times the SD as being outliers. The mean value survey data from around the world, such as that from the Nippon
of the neighborhood pixels was used to replace the outliers, and the Foundation-GEBCO Seabed 2030 Project. However, the increased cover-
empty space in the NoData region (the voids in the original DBMs) age improves the inversion accuracy of the gravity anomaly in the blank
was filled. This process filtered out the outliers. Next, a low-pass filter areas, which improves the spatial resolution of the DBMs.
Fig. 8. Quantitative evaluation of DBMs in different terrain subregions (Zones 1–7).
Fig. 9. Principle of DBM fusion via the weighted sum of multiple datasets.
Fig. 10. General flow of the DBM subregional spatial domain optimal weighted fusion.
Fig. 11. Post-processing steps.
In Table 3, it can be seen that there is little difference in the degree of the most from GEBCO_2019 (about 61% regional difference within
discretization among the first six models (SD = 1503–1509 m). the range −50–50 m, with an average of 14.46 m). TOPO V19.1 and
DTU10's degree of discretization varies from those of the first six models GEBCO_2019 vary the most from one another.
and is higher. Table 4 provides a spatial correlation matrix for the seven The measured data were then introduced, and the differences ob-
DBMs, showing high overall correlation. However, TOPO V19.1 and tained by subtracting the measured values from the seven DBMs were
DTU10 have little correlation with the other models. quantitatively evaluated, as shown in Fig. 13. The differences between
the measured data and GEBCO_2019, GEBCO_2014, SRTM15_PLUS,
4.1.2. Quantitative comparison and SRTM30_PLUS are concentrated within the range (−100–100 m),
The bathymetry differences obtained by subtracting GEBCO_2019 accounting for more than 80% of the domain. The differences between
from the other models reflects the distribution of differences among the measured data and TOPO, ETOPO1, and DTU10 are concentrated
the models (Fig. 12). The differences are primarily concentrated within within the range (−200–200 m), accounting for more than 75%. It can
the range (−100–100 m), accounting for approximately 70–90% of the be seen from the histogram that the differences appear to be normally
domain. As evident from the spatial distribution maps, large differences distributed and the errors tend to be positive, indicating that the ba-
(absolute value >100 m) appear in sea areas with complex topography thymetry of the seven DBMs is shallower than that of the measured
and sparse measurement points, such as over the Spratly and Paracel values. The MAE, SD, and RMSE are smallest for SRTM15_PLUS V2
Islands. The differences appear to be normally distributed. Except for followed by GEBCO_2019, indicating that the overall accuracy of these
those corresponding to DTU10, the relative differences are generally two models is the highest and the most robust. The results show
negative, indicating that GEBCO_2019 is shallower. The model with that the accuracies of GEBCO_2019 and SRTM15_PLUS V2 have been
the smallest difference from GEBCO_2019 is SRTM30_PLUS V11 (about significantly improved: the former by 31% and the latter by 14%. The de-
92% regional difference within the range −50–50 m). DTU10 varies gree of dispersion is also significantly improved. TOPO V19.1 has the
Table 3
Quantitative statistics corresponding to the depth below mean sea level for the seven DBMs.
Dataset (DBMS) GEBCO_2019 GEBCO_2014 SRTM15_PLUS SRTM30_PLUS TOPO ETOPO1 DTU10
MIN (m) −5408 −5190 −5363 −5391 −5347 −5136 −5142

MAX (m) 0 −1 −1 −1 −1 −1 0
MEAN (m) −1249 −1242 −1251 −1245 −1254 −1247 −1239
SD (m) 1505 1505 1503 1503 1504 1509 1478
Table 4
Correlation matrix for the seven DBMs.
Dataset (DBMS) GEBCO_2019 GEBCO_2014 SRTM15_PLUS SRTM30_PLUS TOPO ETOPO1 DTU10
GEBCO_2019 1 0.9957 0.9985 0.9977 0.9909 0.9941 0.9917

GEBCO_2014 0.9957 1 0.9952 0.9979 0.9905 0.9950 0.9904
SRTM15_PLUS 0.9985 0.9952 1 0.9972 0.9912 0.9935 0.9915
SRTM30_PLUS 0.9977 0.9979 0.9972 1 0.9919 0.9942 0.9916
TOPO V19.1 0.9909 0.9905 0.9912 0.9919 1 0.9859 0.9883
ETOPO1 0.9941 0.9950 0.9935 0.9942 0.9859 1 0.9895
DTU10 0.9917 0.9904 0.9915 0.9916 0.9883 0.9895 1
lowest overall accuracy and the most discrete distribution. Although the Y-axis represents bathymetry in units of m, and the X-axis represents
mean of ETOPO1 is small, its degree of dispersion is similar to that of the distance along the cruise in units of km. The black lines (Fig. 14
DTU10. DTU10's MAE is the largest (MAE = 138.04 m), indicating that (c)–(d)) show the largest differences between the measured values
the level of precision provided by this model lacks robustness. Finally, and those of the DBMs along the cruise profile. For Cruise_01, the R2
the R2 of all seven DBMs approaches 1, with that of TOPO_V19 being values of GEBCO_2019, GEBCO_2014, SRTM15_PLUS V2, SRTM30_PLUS
the smallest (R2 = 0.96). V11, TOPO V19.1, ETOPO1, DTU10, and measured values are 0.8977,
0.8753, 0.9087, 0.8937, 0.8995, 0.8349, and 0.8520, respectively. For
4.1.3. Cruise profile comparison Cruise_02, the R2 values are 0.9864, 0.9924, 0.9937, 0.9911, 0.7934,
Fig. 14 shows the bathymetry profiles from Cruise_01 (Cruise ID: 0.9360, and 0.8468. As shown in Fig. 14(c)–(d), these differences have
8491) and Cruise_02 (Cruise ID: 19970254). In the cruise profile, the little correlation with the bathymetry. Most of the differences occur in
Fig. 12. Spatial distribution and statistics of the relative model differences. The bathymetry differences are obtained by subtracting GEBCO_2019 from the other models. Positive values
represent shallower depth than that of GEBCO_2019. (a) GEBCO_2014; (b) SRTM15_PLUS V2; (c) SRTM30_PLUS V11; (d) TOPO V19.1; (e) ETOPO1; (f) DTU10.
Fig. 13. Differences obtained by subtracting the measured values from the DBMs. A positive value indicates that the DBM bathymetry is shallower than that of the measured values.
areas with complex changes in terrain profile. The models show good the predicted bathymetry must be interpolated from 1 arc-minute to
agreement with the measured values in sea areas with gentle changes 15 arc-seconds; hence, the “Remove-Restore” method is adopted to en-
in terrain slope. hance the fine-scale details (Olson et al., 2014). This interpolation pro-
cess may introduce uncertainty and hence, both versions of the
4.1.4. Utility of original DBMs GEBCO datasets and SRTM datasets are considered in this study. The
Gravity anomaly and vertical gravity gradient anomaly data have predicted bathymetry of TOPO V19.1 comes from the V23 global gravity
different bands sensitive to bathymetry inversion. Gravity gradient anomaly model based on the altimetry data of Cryosat-2, Jason-1, and
anomaly is more sensitive to gravity anomaly in the high-frequency Envisat. Each version of the SIO datasets adds new measured data on
part of seafloor topography. In different regions, some errors should the basis of the previous version, optimizes the processing, uses the lat-
be considered, such as the separation error between long and short est gravity data, or optimizes the inversion algorithm (Sandwell et al.,
wavelengths of the gravity signal or the vertical gravity gradient signal, 2014). However, TOPO V19.1 has the worst accuracy among all DBMs
and the error of the density difference constant. These types of errors in the SCS and needs to be further checked for anomalies. Most of the
lead directly to the differing accuracies of the various inversion models bathymetric data in ETOPO1 are derived from SIO datasets (Amante
(Yang et al., 2018; Wan et al., 2019). Consequently, the performance of and Eakins, 2009). ET0P01 provides a clear description of the continen-
the seven global seafloor topography products selected in the SCS is un- tal shelf, ocean basin, trenches, and ridges, which affect the long wave-
known. Therefore, the advantages and limitations of the DBMs are first length information of gravity anomaly, whereas the description of the
compared, and their differing utilities are utilized. Then, adaptive fusion shoals and reefs, pertaining to short wavelength information, is not
of the multi-source model is carried out by referring to the measured very clear, and there are large uncertainties in areas of shallow water.
bathymetry. GEBCO_2019 and SRTM15_PLUS V2 describe the global The DTU10 dataset is based on the altimetric gravity anomaly (ERS-1,
seafloor in more detail than GEBCO_2014 and SRTM30_PLUS V11.0, GEOSAT) and the early GEBCO datasets generated by the contour
owing to the latest regional data compilation and gravity models. model, with statistical deviations (Andersen et al., 2008). The detailed
SRTM15_PLUS V2 has 351 million newly acquired bathymetric data, features of the DTU10 dataset are not clearly delineated and show a dis-
which are sampled at 10.84% at 15 arc-second resolution and 18.29% continuous step-like distribution on the plane.
at 30 arc-second resolution. The predicted bathymetry of GEBCO_2019
is based on satellite-derived gravity data from SRTM15_PLUS V1. How- 4.2. Fusion results and performance
ever, the resolution available from satellite altimetry depends largely on
the regional bathymetry and submarine structure (Weatherall et al., 4.2.1. Overall results
2015). The gravity models used by SRTM15_PLUS V2 are marine Based on the model fusion process presented in Section 3.3, the final
free-air gravity anomaly model (V27.1) and vertical gravity gradient fusion results are shown in Fig. 15 and the corresponding weights of the
anomaly model (V27.1) with resolution of 1 arc-minute. Therefore, seven DBMs for different subregions are shown in Table 5.
Fig. 14. Differences between measured and model values along Cruise_01 (Cruise ID: 8491) and Cruise_02 (Cruise ID: 19970254). (a)–(b) The courses of the cruises with GEBCO_2019 in
the background; (c) Bathymetry profile along Cruise_01; (d) Bathymetry profile along Cruise_02. The Y-axis represents the bathymetry in units of m, and the X-axis represents the distance
along the cruise in units of km. The black dots represent the measured depth. MAX_Difference represents the maximum differences between the DBM values and measured values.
Compared with the original DBMs, the overall accuracy of the up- original DBMs and is concentrated within the range (−100–100 m), ac-
dated SCS DBM is higher, with RMSE of 99.60 m, MAE of 44.03 m, SD counting for about 90%. Its accuracy is 13%, 40%, and 15% higher than
of 99 m, and R2 of 0.98. The range of the differences between the up- those of GEBCO_2019, GRBCO_2017, and SRTM30_PLUS, respectively
dated SCS DBM and measured values is smaller than those of the (Fig. 13). As shown in Fig. 14(c)–(d), for Cruise_01, the R2 value of the
Fig. 15. Fusion results. (a) The updated SCS DBM at 15 arc-second resolution; (b) Accuracy statistics for the updated SCS DBM; positive values indicate that the DBM bathymetry is
shallower than the measured bathymetry.
Table 5
Corresponding weights of the seven DBMs for different subregions.
Dataset (w) Zone 1 Zone 2 Zone 3 Zone 4 Zone 5 Zone 6 Zone 7
DBMi (wi) SRTM15_PLUS (0.85) SRTM15_PLUS (0.77) SRTM15_PLUS (0.50) SRTM15_PLUS (0.73) SRTM15_PLUS (0.91) SRTM15_PLUS (0.78) SRTM15_PLUS (0.80)
DBMj (wj) GEBCO_2019 (0.15) GEBCO_2019 (0.23) SRTM30_PLUS (0.50) GEBCO_2014 (0.27) DTU10 (0.09) GEBCO_2014 (0.22) TOPO V19.1 (0.20)
updated SCS DBM and measured value is 0.9091. For Cruise_02, the R2 in the Paracel Islands region. The results of the two groups of experi-
value is 0.9937. The accuracies of the updated SCS DBM and ments are shown in Figs. 16 and 17, in which the contours (Figs. 16
SRTM15_PLUS (RMSE = 100.74 m) are most similar (the fused SCS (a) and 17(a)) are interpolated by ordinary kriging from the measured
DBM accuracy is 1% higher). This proves that SRTM15_PLUS is the points (from ship depth-soundings and charts). The eight DBMs appear
most accurate, except for the updated SCS DBM. The fusion framework to be consistent with the contour map, but careful inspection reveals
proposed in this paper, including measured data editing, recovery of differences.
measured data, model smoothing, and other processing, can effectively In Fig. 16, it is noted that the updated DBM for the Spratly Islands is
fill voids in the original DBM and improve its quality (Fig. 15). most similar to SRTM15_PLUS and GEBCO_2019. The sizes and locations
of the reefs above sea level are consistent (Fig. 16(b)–(e)). This means
4.2.2. Local detail performance that their elevations above sea level are the same, implying that their
The original DBMs and the fused DBM were compared using the sources are similar. SRTM15_PLUS and GEBCO_2019 are more reliable
cross-validation statistical method (Zhang and Wang, 2010). The exper- than their previous versions in shallow water (Fig. 16(d)–(f)). TOPO
imental locations were the Spratly Islands (9.40°–12.00°N, 113.20°– V19.1 has significant noise where the bathymetry changes (Fig. 16
116.00°E) and Paracel Islands (15.70°–17.25°N, 111.12°–113.12°E) and (g)). In deep water, ETOPO1 is similar to SRTM30_PLUS, while their dif-
their adjacent areas (Fig. 1(c)–(d)). Both areas have large coral reefs ferences above sea level are large (Fig. 16(h)). The inversion depth of
and complex topography with dense height contours. The accuracies DTU10 in deep water areas is generally the smallest, but this model
of the original DBMs and the results obtained after fusion were com- lacks the most detail (Fig. 16(i)). In Fig. 17, the updated DBM for the
pared with the measured data. The measured data were obtained Paracel Islands is seen to have bathymetry that most closely matches
from charts: 2900 points in the Spratly Islands region and 900 points that of SRTM15_PLUS, but the heights of the reefs above sea level are
Fig. 16. DBMs at 15 arc-second resolution in the Spratly Islands. (a) Contour map; (b) The updated DBM; (c) GEBCO_2019; (d) GEBCO_2014; (e) SRTM15_PLUS V2; (f) SRTM30_PLUS V11;
(g) TOPO V19.1; (h) ETOPO1; (i) DTU10. Regions with no data are above sea level.
Fig. 17. DBMs at 15 arc-second resolution in the Paracel Islands. (a) Contour map; (b) The updated DBM; (c) GEBCO_2019; (d) GEBCO_2014; (e) SRTM15_PLUS V2; (f) SRTM30_PLUS V11;
(g) TOPO V19.1; (h) ETOPO1; (i) DTU10. Regions with no data are above sea level.
slightly lower than those from SRTM15_PLUS (Fig. 17(b)–(e)). Simi- updated DBM, followed by GEBCO_2019. The RMSE of the updated
larly, SRTM15_PLUS and GEBCO_2019 are more reliable and more de- DBM is 193.16 m and its accuracy is 1.1% and 6.9% higher than those of
tailed than their previous versions in shallow water (Fig. 17(d)–(f)). SRTM15_PLUS and GEBCO_2019, respectively. The fusion process for
TOPO V19.1 has significant noise in shallow water (Fig. 17(g)). Many the Paracel Islands is less accurate than that for the Spratly Islands be-
details are missing from ETOPO1 and DTU10 (Fig. 17(h)–(i)). This is cause of the sparseness of the measured points and the shallow bathym-
particularly true for DTU10 (Fig. 17(h)–(i)). etry. The results show that the fusion framework most improves the
The MAE, SD, RMSE, R2, and other statistical parameters are given in results when the terrain has a more complicated bathymetry.
Tables 6 and 7. In both experiments, the MAE, SD, and RMSE for the The above visual and quantitative evaluations motivated us to try to un-
updated DBM are smaller than those for the other seven models, derstand the fusion framework by evaluating the spatial distribution of the
and R2 for the fused DBM is the largest among all the models. In errors, as reflected in the error contour map. The error was calculated by
other words, the updated DBM gives the most accurate reconstruc- subtracting the local DBMs from the original DBMs and the updated DBM.
tion of the bathymetry of the two islands, which is consistent with The local DBMs are interpolated by ordinary kriging from the measured
what is seen in Figs. 16 and 17. In the Spratly Islands, DTU10, is second values. The error contours corresponding to Figs. 16 and 17 are shown in
only to the updated DBM, followed by SRTM15_PLUS then GEBCO_2019. Figs. 18 and 19. It can be seen that the error ranges for SRTM15_PLUS
The RMSE of the updated DBM is 358.26 m, with accuracies 6.7%, 7.0%, and GEBCO_2019 in the two regions are larger than that for the updated
and 7.9% higher than those of DTU10, SRTM15_PLUS, and GEBCO_2019, DBM. Compared with those for GEBCO_2019, the errors between
respectively. SRTM15_PLUS in the Paracel Islands is second only to the SRTM15_PLUS and the updated DBM are smaller. The error contours of
Table 6
Quantitative evaluation of DBMs in the Spratly Islands.
Dataset (DBMS) MIN (m) MAX (m) MEAN (m) MAE (m) SD (m) RMSE (m) R2
GEBCO_2019 −4290.0000 2238.3140 57.8757 229.6126 384.8152 389.0775 0.9100

GEBCO_2014 −4241.0000 2437.0000 118.7945 325.9151 484.2331 498.5108 0.8646
SRTM15_PLUS −4294.0000 2157.0000 57.7656 223.7707 380.7697 385.0616 0.9113
SRTM30_PLUS −4286.0000 2318.0000 69.4966 240.2609 397.4885 403.4506 0.9044
TOPO V19.1 −4288.0000 2279.0000 28.1038 245.0346 413.1845 414.0681 0.8975
ETOPO1 −4246.0000 2390.0000 121.2049 341.2172 500.7995 515.1740 0.8515
DTU10 −4062.9700 2145.1150 11.3197 222.6274 379.5405 383.9602 0.9181
Updated DBM −4281.6700 2093.1430 56.1237 220.6726 345.8263 358.2878 0.9620
The bold type indicates the optimal value for each column.
Table 7
Quantitative evaluation of DBMs in the Paracel Islands.
Dataset (DBMS) MIN (m) MAX (m) MEAN (m) MAE (m) SD (m) RMSE (m) R2
GEBCO_2019 −1855.3100 980.9320 68.3935 121.6268 196.0820 207.5694 0.8942

GEBCO_2014 −1987.0000 1041.0000 130.3948 195.6151 260.8552 291.5066 0.8314
SRTM15_PLUS −1851.0000 879.0000 52.2718 110.1555 188.2286 195.2556 0.9025
SRTM30_PLUS −1866.0000 1021.0000 66.4945 121.2286 198.1316 208.8924 0.8919
TOPO V19.1 −1659.0000 1831.0000 10.1636 224.5690 355.4373 355.3941 0.6569
ETOPO1 −1880.0000 1077.0000 151.6461 217.0764 282.0063 320.0621 0.8035
DTU10 −1132.0700 1121.7172 90.1808 162.7742 231.6650 248.4841 0.8513
Updated DBM −1783.7800 898.5560 54.1125 110.1292 185.5026 193.1556 0.9049
The bold type indicates the optimal value for each column.
the updated DBM are all located in low-error regions, which indicates that seven satellite altimeters (Geosat, ERS1, Envisat, Jason-1/2, CryoSat-2,
the updated DBM achieves better results than SRTM15_PLUS and and AltiKa) were used for gravity field recovery, operating in a non-
GEBCO_2019. We note that the error contours corresponding to the Spratly repeat orbital phase. Three of the altimeters (Jason-2, CryoSat-2, and
Islands are more concentrated than those in the Paracel Islands contour AltiKa) continue to acquire data; hence, improvements in gravity will
map, which is related to the density of the coverage by the measured points, continue (Tozer et al., 2019). Therefore, the newly released DBM con-
i.e., there are fewer data points covering the Paracel Islands. Thus, the error tributes more to the fusion results (in Section 4.2.1). However, high ac-
contour maps also illustrate the superiority of the fusion framework. curacy on a global scale does not mean that it is equally accurate locally.
Considering the SRTM Datasets as an example, the new data need to un-
4.3. Feasibility and limitations dergo iterative “Remove-Restore” processing, in which the compilation
of ship depth-soundings relies on a large number of human–machine
4.3.1. Feasibility of the fusion framework interactions (Weatherall et al., 2015). In this progressive process, it is
As described in Section 4.2.1, the updated DBM is much more accu- impossible to eliminate all outliers and void values, especially in areas
rate than GEBCO datasets and so on. The fusion result is more similar to where the seafloor topography changes more dramatically (such as
the high-resolution datasets obtained in recent years, such as the edge of a trough). In addition, traditional satellite altimetry is lim-
GEBCO_2019 and RTM15_PLUS. However, the updated DBM relies ited by the resolution of the trajectory, so it is difficult to obtain high-
heavily on SRTM15_PLUS and is only 1% more accurate (Table 5). This resolution and high-precision data of the marine gravity data. In other
is because the newly released dataset contains the latest measured words, the products of existing satellites cannot fully meet the require-
data and gravity data (in Section 4.1.4). For the SRTM15_PLUS dataset, ments of bathymetry inversion, and new gravity satellites and altimetry
Fig. 18. Error contours maps of several DBMs in the Spratly Islands. (a) GEBCO_2019; (b) SRTM15_PLUS V2; (c) the updated DBM.
Fig. 19. Error contour maps of several DBMs in the Paracel Islands. (a) GEBCO_2019; (b) SRTM15_PLUS V2; (c) the updated DBM.
satellites are needed to provide higher-precision gravity anomaly and results. However, the SCS DBM with a resolution of 15″ obtained by
vertical gravity gradient, which may not be realized in the near term fusion in this paper is not a good choice for monitoring the geomorphol-
(Yang et al., 2018; Wan et al., 2019). Therefore, it is necessary to conduct ogic evolution of coral reefs in shallow water areas. The latest space-
multi-source data fusion in the SCS. The fusion results are comparable to borne lidar, ICESat-2, has been shown to have bathymetric potential in
SRTM15_PLUS and demonstrate the feasibility of the fusion framework, the coral reef areas (Parrish et al., 2019), and passive satellite imagery
which can provide references for data fusion in other regions. The fusion can also be used to retrieve high-resolution shallow underwater terrain
framework in this paper includes data editing, weighted fusion, recov- (Chu et al., 2019).
ery of measured data, model smoothing, and other processing. In our
proposed fusion framework, multiple datasets are used to reconstruct
and update large-scale reliable seafloor topography in a timely manner. 5. Conclusion
In the process, the measured bathymetry data is actually used to control
and enhance the DBMs. To a certain extent, the proposed fusion frame- This study compares and evaluates the performance of seven global
work can overcome the insufficiency of the measured bathymetry data. DBMs most widely used in the SCS. The models are integrated and sup-
When new measured data are added to a local area, the reliability of the plemented by completing their voids and eliminating their outliers, and
DBM can be improved by using the proposed fusion framework. these processes are shown to improve model reliability. An adaptive
The generation accuracy of seafloor terrain model is related to the subregional spatial-domain-weighted fusion framework is proposed
density of bathymetric data (Weatherall et al., 2015). Although the based on these DBMs and several measured datasets.
SCS is used as the study area in this study, the fusion framework can The results show that SRTM15_PLUS V2 and GEBCO_2019 have the
be extended to the surrounding areas. The ship depth-soundings of highest and most robust overall accuracy among the seven datasets.
the SCS area are mainly single-beam data. Therefore, in this work, sub- However, SRTM15_PLUS V2 is shown to have several voids; in addition,
regional fusion of data sets is conducted first, followed by recovery of ac- the number of outliers had to be reduced. The updated seamless SCS
tual measured data. In other words, measured values are used to replace DBM model is void-free and has a resolution of 15 arc-seconds. The
model-predicted bathymetry. For the surrounding areas covered by RMSE, MAE, SD, and R2 of the updated DBM are 99.60 m, 44.03 m, 99
multi-beam or bathymetric lidar data, such as the Malacca Strait m, and 0.98, respectively. The accuracy of the updated DBM is 13%,
(bathymetry between 25 and 150 m), the Karimata Strait (bathymetry 40%, 15%, and 1% higher than those of GEBCO_2019, GEBCO_2014,
between 18 and 37 m), and the Philippine Sea (an average bathymetry SRTM30_PLUS, and SRTM15_PLUS, respectively, and provides a more
of 5000–6000 m), the corresponding datasets can be segmented and detailed and realistic expression of terrain details. The proposed fusion
fused using the proposed fusion framework, and then the classical framework can be extended to the surroundings and other areas.
“Remove-Restore” method can be used to recover the measured values The fusion method and results proposed in this paper still have some
(Becker et al., 2009). The specific approach is employed to first grid the limitations. Due to the lack of sufficient multi-beam data in the SCS, the
new data into a new high-resolution DBM, then compute the difference classical “Remove-Restore” method is no longer applicable. In our fusion
between the new DBM and the fusion DBM, and finally to grid the differ- framework, the measured data are used to replace the gravity-predicted
ence and add it back to the existing fusion DBM. The aim is to achieve a bathymetry. In shallow water areas, especially in the coral reef areas,
smooth transition between the new DBM and the fusion DBM with min- bathymetry shows great uncertainty. The fusion result is more similar
imum of perturbation of the existing fusion DBM. to and relies heavily on the newly released high-resolution datasets,
with only 1% higher accuracy than SRTM15_PLUS. The updated SCS
DBM, with a resolution of 15″, may not be a good choice for monitoring
4.3.2. Effects of temporal variation on fusion results
the geomorphological evolution of the coral reef. However, the full
However, large uncertainties exist in shallow water areas, especially
multi-beam coverage of the global ocean would require approximately
in coral reef areas. It can be seen that the problem of restoring zero
900 years of navigation (Weatherall et al., 2015). More accurate recon-
depth from altimeter and ship data still exists (Hsiao et al., 2016).
struction of shallow-water bathymetry could be obtained using
Over time, coral reefs in the Spratly Islands and Paracel Islands have
remote sensing imagery; this would allow the generation of new
generally deteriorated. This mainly manifests in the significant decline
in living coral coverage and symbiotic zooxanthellae density, which is models/datasets by interpolation. These issues will be addressed in
our future work.
primarily caused by the combined action of the warming climate and
human activities (Zhao et al., 2012; Yu et al., 2006). The reef accretion
rate and its relationship with sea level are two main factors in the geo- Declaration of competing interest
morphologic evolution of coral reefs. Although the growth rate of coral
communities can range from 10 mm yr−1 to 100 mm yr−1, the reef ac- The authors declare that they have no known competing financial
cretion rate is relatively slow at approximately 1–10 mm yr−1 interests or personal relationships that could have appeared to influ-
(Hamylton et al., 2014). At present, the average living coral coverage ence the work reported in this paper.
of the Spratly Islands is about 36.6% (Tkachenko et al., 2020). The reef
accretion rate is seriously affected by thermal anomalies, which is the Acknowledgments
most extensive and threatening factor affecting coral reef geomorphol-
ogic evolution and can even cause coral bleaching. Studies have shown This work was funded by the National Key R&D Plan (Grant No.
that severe coral bleaching mainly occurs in areas with bathymetry of 2017YFB0504205), the National Natural Science Foundation of China
less than 10–15 m (Muir et al., 2017). According to the Fifth Assessment (Grant No. 41622109), and the Guangxi Innovative Development Grand
Report of the Intergovernmental Panel on Climate Change (IPCC), aver- Grant (Grant No. AA18118038). The authors express their sincere
age global sea level rose by 19 cm (17–21 cm) from 1901 to 2010; sea thanks for the comments and contributions of anonymous reviewers
level trends in the SCS are in line with the global average (Huang and and members of the editorial team.
Qiao, 2015). Simulation experiments have shown that different coral
reef landforms, such as reef-front, reef-flat, and lagoon, have different References
responses to sea level rise (Hamylton et al., 2014). Therefore, the addi-
tion of new measured data can more accurately reflect the current situ- Amante, C., Eakins, B.W., 2009. ETOPO1 1 arc-minute global relief model: procedures,
data sources and analysis. NOAA Tech. Mem. NESDIS. https://doi.org/10.1594/
ation of shallow coral reef geomorphology. In practice, as shown in PANGAEA.769615.
Table 5, our weighting results also demonstrate that the newly released Andersen, et al., 2008. The DTU10 global Mean sea surface and Bathymetry. Presented
DBM has more weight and hence, they contribute more to the fusion EGU-2008, Vienna, Austria, April.
Becker, J.J., Sandwell, D.T., Smith, W.H.F., Braud, J., Binder, B., Depner, J., Fabre, D., Factor, J., Olson, C.J., Becker, J.J., Sandwell, D.T., 2014. A new global bathymetry map at 15 arcsecond
Ingalls, S., Kim, S.H., Ladner, R., Marks, K., Nelson, S., Pharaoh, A., Trimmer, R., von resolution for resolving seafloor fabric: SRTM15_PLUS. AGU Fall Meeting Abstracts.
Rosenberg, J., Wallace, G., Weatherall, P., 2009. Global bathymetry and elevation Papasaika, H., Kokiopoulou, E., Baltsavias, E., Schindler, K., Kressner, D., 2011. Fusion of
data at 30 arc seconds resolution: SRTM30_PLUS. Mar. Geod. https://doi.org/ digital elevation models using sparse representations. Lecture Notes in Computer Sci-
10.1080/01490410903297766. ence (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in
Chiocci, F.L., Cattaneo, A., Urgeles, R., 2011. Seafloor mapping for geohazard assessment: Bioinformatics) https://doi.org/10.1007/978-3-642-24393-6_15.
State of the art. Mar. Geophys. Res. https://doi.org/10.1007/s11001-011-9139-8. Parrish, C.E., Magruder, L.A., Neuenschwander, A.L., Forfinski-Sarkozi, N., Alonzo, M.,
Chu, S., Cheng, L., Ruan, X., Zhuang, Q., Zhou, X., Li, M., Shi, Y., 2019. Technical framework Jasinski, M., 2019. Validation of ICESat-2 ATLAS bathymetry and analysis of ATLAS’s
for shallow-water bathymetry with high reliability and no missing data based on bathymetric mapping performance. Remote Sens. https://doi.org/10.3390/
time-series sentinel-2 images. IEEE Trans. Geosci. Remote Sens. https://doi.org/ rs11141634.
10.1109/TGRS.2019.2922724. Pe’eri, S., Parrish, C., Azuike, C., Alexander, L., Armstrong, A., 2014. Satellite remote sensing
Church, J.A., White, N.J., Konikow, L.F., Domingues, C.M., Cogley, J.G., Rignot, E., Gregory, J. as a reconnaissance tool for assessing nautical chart adequacy and completeness.
M., Van Den Broeke, M.R., Monaghan, A.J., Velicogna, I., 2011. Revisiting the Earth’s Mar. Geod. https://doi.org/10.1080/01490419.2014.902880.
sea-level and energy budgets from 1961 to 2008. Geophys. Res. Lett. https://doi. Peng, D., Palanisamy, H., Cazenave, A., Meyssignac, B., 2013. Interannual sea level varia-
org/10.1029/2011GL048794. tions in the South China Sea over 1950-2009. Mar. Geod. https://doi.org/10.1080/
Costa, H., Foody, G.M., Boyd, D.S., 2018. Supervised methods of image segmentation accu- 01490419.2013.771595.
racy assessment in land cover mapping. Remote Sens. Environ. https://doi.org/ Petrasova, A., Mitasova, H., Petras, V., Jeziorska, J., 2017. Fusion of high-resolution DEMs
10.1016/j.rse.2017.11.024. for water flow modeling. Open Geospatial Data, Softw. Stand https://doi.org/
Costantini, M., Malvarosa, F., Minati, F., Zappitelli, E., Seifert, F.M., 2006. A data fusion al- 10.1186/s40965-017-0019-2.
gorithm for DEM mosaicking: building a global DEM with SRTM-X and ERS data. In- Picard, K., Brooke, B.P., Harris, P.T., Siwabessy, P.J.W., Coffin, M.F., Tran, M., Spinoccia, M.,
ternational Geoscience and Remote Sensing Symposium (IGARSS) https://doi.org/ Weales, J., Macmillan-Lawler, M., Sullivan, J., 2018. Malaysia Airlines flight MH370
10.1109/IGARSS.2006.990. search data reveal geomorphology and seafloor processes in the remote southeast
Drăguţ, L., Eisank, C., 2012. Automated object-based classification of topography from Indian Ocean. Mar. Geol. https://doi.org/10.1016/j.margeo.2017.10.014.
SRTM data. Geomorphology. https://doi.org/10.1016/j.geomorph.2011.12.001. Podobnikar, T., 2005. Production of integrated digital terrain model from multiple
Drăguţ, L., Csillik, O., Eisank, C., Tiede, D., 2014. Automated parameterisation for multi- datasets of different quality. Int. J. Geogr. Inf. Sci. https://doi.org/10.1080/
scale image segmentation on multiple layers. ISPRS J. Photogramm. Remote Sens. 13658810412331280130.
https://doi.org/10.1016/j.isprsjprs.2013.11.018. Pont-Tuset, J., Arbelaez, P., Barron, J.T., Marques, F., Malik, J., 2017. Multiscale combinato-
Eakins, B.W., Sharman, G.f., 2010. Volumes of the World’s Oceans From ETOPO1. NOAA rial grouping for image segmentation and object proposal generation. IEEE Trans. Pat-
Natl. Geophys. Data Center, Boulder, CO. tern Anal. Mach. Intell. https://doi.org/10.1109/TPAMI.2016.2537320.
Ellis, J.I., Ellis, J.I., Clark, M.R., Lamarche, G., Rouse, H.L., Lamarche, G., 2017. Environmental Ramillien, G., Cazenave, A., 1997. Global bathymetry derived from altimeter data of the
management frameworks for offshore mining: the New Zealand approach. Mar. Pol- ERS-1 Geodetic Mission. J. Geodyn. https://doi.org/10.1016/S0264-3707(96)00026-9.
icy https://doi.org/10.1016/j.marpol.2017.07.004. Rengstorf, A.M., Mohn, C., Brown, C., Wisz, M.S., Grehan, A.J., 2014. Predicting the distribu-
GEBCO Compilation Group, 2019. GEBCO 2019 Grid. https://doi.org/10.5285/836f016a- tion of deep-sea vulnerable marine ecosystems using high-resolution data: consider-
33be-6ddc-e053-6c86abc0788e. ations and novel approaches. Deep. Res. Part I Oceanogr. Res. Pap. https://doi.org/
Glenn, J., Tonina, D., Morehead, M.D., Fiedler, F., Benjankar, R., 2016. Effect of transect lo- 10.1016/j.dsr.2014.07.007.
cation, transect spacing and interpolation methods on river bathymetry accuracy. Ronov, A.B., 1994. Phanerozoic transgressions and regressions on the continents: a quan-
Earth Surf. Process. Landforms. https://doi.org/10.1002/esp.3891.
titative approach based on areas flooded by the sea and areas of marine and conti-
Gruber, A., Wessel, B., Martone, M., Roth, A., 2016. The TanDEM-X DEM mosaicking: fu- nental deposition. Am. J. Sci. https://doi.org/10.2475/ajs.294.7.777.
sion of multiple acquisitions using InSAR quality parameters. IEEE J. Sel. Top. Appl.
Sandwell, D.T., Smith, W.H.F., 1997. Marine gravity anomaly from Geosat and ERS 1 satel-
Earth Obs. Remote Sens. https://doi.org/10.1109/JSTARS.2015.2421879.
lite altimetry. J. Geophys. Res. B Solid Earth. https://doi.org/10.1029/96JB03223.
Gula, J., Molemaker, M.J., McWilliams, J.C., 2014. Gulf stream dynamics along the South-
Sandwell, D.T., Müller, R.D., Smith, W.H.F., Garcia, E., Francis, R., 2014. New global marine
eastern U.S. Seaboard. J. Phys. Oceanogr https://doi.org/10.1175/jpo-d-14-0154.1.
gravity model from CryoSat-2 and Jason-1 reveals buried tectonic structure. Science.
Hamylton, S.M., Leon, J.X., Saunders, M.I., Woodroffe, C.D., 2014. Simulating reef response
https://doi.org/10.1126/science.1258213.
to sea-level rise at Lizard Island: a geospatial approach. Geomorphology https://doi.
Schaffer, J., Timmermann, R., Erik Arndt, J., Savstrup Kristensen, S., Mayer, C., Morlighem,
org/10.1016/j.geomorph.2014.03.006.
M., Steinhage, D., 2016. A global, high-resolution data set of ice sheet topography,
Harris, P.T., Macmillan-Lawler, M., Rupp, J., Baker, E.K., 2014. Geomorphology of the
cavity geometry, and ocean bathymetry. Earth Syst. Sci. Data. https://doi.org/
oceans. Mar. Geol. https://doi.org/10.1016/j.margeo.2014.01.011.
10.5194/essd-8-543-2016.
Hell, B., Jakobsson, M., 2011. Gridding heterogeneous bathymetric data sets with stacked
Schindler, K., Haris, P., Schütz, S., Baltsavias, E., 2011. Improving wide-area DEMs through
continuous curvature splines in tension. Mar. Geophys. Res. https://doi.org/10.1007/
data fusion – chances and limits. Photogrammetric Week ’11.
s11001-011-9141-1.
Sepúlveda, I., Liu, P.L.F., Grigoriu, M., 2019. Probabilistic tsunami hazard assessment in
Hsiao, Y.S., Hwang, C., Cheng, Y.S., Chen, L.C., Hsu, H.J., Tsai, J.H., Liu, C.L., Wang, C.C., Liu,
South China Sea with consideration of uncertain earthquake characteristics.
Y.C., Kao, Y.C., 2016. High-resolution depth and coastline over major atolls of South
J. Geophys. Res. Solid Earth https://doi.org/10.1029/2018JB016620.
China Sea from satellite altimetry and imagery. Remote Sens. Environ. https://doi.
org/10.1016/j.rse.2016.01.016. Smith, W.H.F., Sandwell, D.T., 1997. Global sea floor topography from satellite altimetry
Huang, C., Qiao, F., 2015. Sea level rise projection in the South China Sea from CMIP5 and ship depth soundings. Science. https://doi.org/10.1126/science.277.5334.1956.
models. Acta Oceanol. Sin. https://doi.org/10.1007/s13131-015-0631-x. Smith, W.H.F., Wessel, P., 2002. Gridding with continuous curvature splines in tension.
International Hydeographic Bureau, 2008. IHO Standards For Hydrographic Surveys, In- Geophysics. https://doi.org/10.1190/1.1442837.
ternational Hydeographic Bureau. https://doi.org/10.1088/1751-8113/44/8/085201. Smith, D.E., Zuber, M.T., Solomon, S.C., Phillips, R.J., Head, J.W., Garvin, J.B., Banerdt, W.B.,
Jakobsson, M., Mayer, L., Armstrong, A., 2003. Analysis of data relevant to establishing Muhleman, D.O., Pettengill, G.H., Neumann, G.A., Lemoine, F.G., Abshire, J.B.,
outer limits of a continental shelf under law of the sea article 76. Int. Hydrogr. Rev. Aharonson, O., Brown, C.D., Hauck, S.A., Ivanov, A.B., McGovern, P.J., Zwally, H.J.,
4 (1), 2–18. Duxbury, T.C., 1999. The global topography of Mars and implications for surface evo-
Jakobsson, M., Allen, G., Carbotte, S.M., Falconer, R., Ferrini, V., Marks, K., et al., 2017. The lution. Science (80-. ) https://doi.org/10.1126/science.284.5419.1495.
Nippon Foundation-GEBCO-Seabed 2030: roadmap for future ocean floor mapping. The Navigation Guarantee Department of the Chinese PLA Navy Headquarters. Available
Availableat:. https://www.gebco.net/documents/seabed_2030_roadmap_v10_low. online: http://hydro.ngd.gov.cn/paperchart.aspx (accessed on 15 January 2019).
pdf. The Nippon Foundation-GEBCO Seabed 2030 Project, Available online: https://sea-
Jhee, H., Cho, H.C., Kahng, H.K., Cheung, S., 2013. Multiscale quadtree model fusion with bed2030.gebco.net/ (accessed on 20 January 2019).
super-resolution for blocky artefact removal. Remote Sens. Lett. https://doi.org/ Tkachenko, K.S., Hoang, D.T., Dang, H.N., 2020. Ecological status of coral reefs in the
10.1080/2150704X.2012.729869. Spratly Islands, South China Sea (East Sea) and its relation to thermal anomalies.
Karkee, M., Steward, B.L., Aziz, S.A., 2008. Improving quality of public domain digital ele- Estuar. Coast. Shelf Sci. https://doi.org/10.1016/j.ecss.2020.106722.
vation models through data fusion. Biosyst. Eng. https://doi.org/10.1016/j. Tozer, B., Sandwell, D.T., Smith, W.H.F., Olson, C., Beale, J.R., Wessel, P., 2019. Global ba-
biosystemseng.2008.09.010. thymetry and topography at 15 arc sec: SRTM15+. Earth Sp. Sci. https://doi.org/
Knudsen, P., Bingham, R., Andersen, O., Rio, M.H., 2011. A global mean dynamic topogra- 10.1029/2019EA000658.
phy and ocean circulation estimation using a preliminary GOCE gravity model. UN, 2015. Transforming our world: the 2030 agenda for sustainable development. A/RES/
J. Geod. https://doi.org/10.1007/s00190-011-0485-8. 70/1, United Nations General Assembly https://doi.org/10.1007/s13398-014-0173-
Liu, T., Zhao, D., Pan, M., 2014. Generating 3D depiction for a future ECDIS based on digital 7.2.
Earth. J. Navig. https://doi.org/10.1017/s0373463314000381. Walbridge, S., Slocum, N., Pobuda, M., Wright, D.J., 2018. Unified geomorphological anal-
Marks, K., 2016. The IHO-IOC GEBCO cook book. IOC Manuals and Guides 63, 429. ysis workflows with benthic terrain modeler. Geosci. https://doi.org/10.3390/
Mayer, L., Jakobsson, M., Allen, G., Dorschel, B., Falconer, R., Ferrini, V., Lamarche, G., geosciences8030094.
Snaith, H., Weatherall, P., 2018. The Nippon Foundation—GEBCO seabed 2030 project: Wan, X., Ran, J., Jin, S., 2019. Sensitivity analysis of gravity anomalies and vertical gravity
the Quest to See the World’s Oceans completely mapped by 2030. Geosciences. gradient data for bathymetry inversion. Mar. Geophys. Res. https://doi.org/10.1007/
https://doi.org/10.3390/geosciences8020063. s11001-018-9361-8.
Merwade, V., 2009. Effect of spatial trends on interpolation of river bathymetry. J. Hydrol. Weatherall, P., Marks, K.M., Jakobsson, M., Schmitt, T., Tani, S., Arndt, J.E., Rovere, M.,
https://doi.org/10.1016/j.jhydrol.2009.03.026. Chayes, D., Ferrini, V., Wigley, R., 2015. A new digital bathymetric model of the
Muir, P.R., Marshall, P.A., Abdulla, A., Aguirre, J.D., 2017. Species identity and depth predict world’s oceans. Earth Sp. Sci. https://doi.org/10.1002/2015EA000107.
bleaching severity in reef-building corals: shall the deep inherit the reef? Proc. R. Soc. Wessel, P., 2010. Tools for analyzing intersecting tracks: the x2sys package. Comput.
B Biol. Sci. https://doi.org/10.1098/rspb.2017.1551. Geosci. https://doi.org/10.1016/j.cageo.2009.05.009.
Wessel, P., Chandler, M.T., 2007. The mgd77 supplement to the generic mapping tools. mortality over the past two centuries. Quat. Geochronol. https://doi.org/10.1016/j.
Comput. Geosci. https://doi.org/10.1016/j.cageo.2006.05.006. quageo.2006.06.005.
Willner, K., Shi, X., Oberst, J., 2014. Phobos’ shape and topography models. Planet. Space Yue, L., Shen, H., Yuan, Q., Zhang, L., 2015. Fusion of multi-scale DEMs using a regu-
Sci. https://doi.org/10.1016/j.pss.2013.12.006. larized super-resolution method. Int. J. Geogr. Inf. Sci. https://doi.org/10.1080/
Wilson, M.F.J., O’Connell, B., Brown, C., Guinan, J.C., Grehan, A.J., 2007. Multiscale terrain 13658816.2015.1063639.
analysis of multibeam bathymetry data for habitat mapping on the continental Yue, L., Shen, H., Zhang, L., Zheng, X., Zhang, F., Yuan, Q., 2017. High-quality seamless DEM
slope. Mar. Geod. https://doi.org/10.1080/01490410701295962. generation blending SRTM-1, ASTER GDEM v2 and ICESat/GLAS observations. ISPRS
Wlodarczyk-Sielicka, M., Stateczny, A., 2016. Clustering bathymetric data for electronic J. Photogramm. Remote Sens. https://doi.org/10.1016/j.isprsjprs.2016.11.002.
navigational charts. J. Navig. https://doi.org/10.1017/s0373463316000035. Zhang, H., Wang, Y., 2010. Kriging and cross-validation for massive spatial data.
Yang, J., Jekeli, C., Liu, L., 2018. Seafloor topography estimation from gravity gradients Environmetrics. https://doi.org/10.1002/env.1023.
using simulated annealing. J. Geophys. Res. Solid Earth https://doi.org/10.1029/ Zhao, M., Yu, K., Zhang, Q., Shi, Q., Price, G.J., 2012. Long-term decline of a fringing coral
reef in the Northern South China Sea. J. Coast. Res. https://doi.org/10.2112/
2018JB015883.
jcoastres-d-10-00172.1.
Yu, K.F., Zhao, J.X., Shi, Q., Chen, T.G., Wang, P.X., Collerson, K.D., Liu, T.S., 2006. U-series
dating of dead Porites corals in the South China Sea: evidence for episodic coral

Ruan, X., Cheng, L., Chu, S., Yan, Z., Zhou, X., Duan, Z., & Li, M. (2020). a New Digital Bathymetric Model of the South China Sea Based on the Subregional Fusion of Seven Global Seafloor Topography Products. Geomorphology, 370, 1

Uploaded by

Copyright:

Available Formats

Ruan, X., Cheng, L., Chu, S., Yan, Z., Zhou, X., Duan, Z., & Li, M. (2020). a New Digital Bathymetric Model of the South China Sea Based on the Subregional Fusion of Seven Global Seafloor Topography Products. Geomorphology, 370, 1

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ruan, X., Cheng, L., Chu, S., Yan, Z., Zhou, X., Duan, Z., & Li, M. (2020). a New Digital Bathymetric Model of the South China Sea Based on the Subregional Fusion of Seven Global Seafloor Topography Products. Geomorphology, 370, 1

Uploaded by

Copyright:

Available Formats

Geomorphology 370 (2020) 107403

Contents lists available at ScienceDirect

journal homepage: www.elsevier.com/locate/geomorph

1. Introduction geological disaster assessments (Chiocci et al., 2011). By monitoring

Table 2 uncertainties of model-predicted bathymetry vary in different locations,

Fig. 4. Flow chart of the proposed fusion strategy.

Fig. 5. General ﬂow of terrain segmentation and determination of subregions.

Fig. 7. Terrain subregions (Zones 1–7).

3.3.4. Model smoothing 4.1.1. Overall comparison

Fig. 8. Quantitative evaluation of DBMs in different terrain subregions (Zones 1–7).

Fig. 11. Post-processing steps.

Dataset (DBMS) GEBCO_2019 GEBCO_2014 SRTM15_PLUS SRTM30_PLUS TOPO ETOPO1 DTU10

MIN (m) −5408 −5190 −5363 −5391 −5347 −5136 −5142

Dataset (DBMS) GEBCO_2019 GEBCO_2014 SRTM15_PLUS SRTM30_PLUS TOPO ETOPO1 DTU10

GEBCO_2019 1 0.9957 0.9985 0.9977 0.9909 0.9941 0.9917

Dataset (w) Zone 1 Zone 2 Zone 3 Zone 4 Zone 5 Zone 6 Zone 7

GEBCO_2019 −4290.0000 2238.3140 57.8757 229.6126 384.8152 389.0775 0.9100

GEBCO_2019 −1855.3100 980.9320 68.3935 121.6268 196.0820 207.5694 0.8942

You might also like