1 A VGGNet-based Correction for Satellite Altimetry-derived Gravity

2 Anomaliesy to Improve the Accuracy of Bathymetry Down to Depths

3 of 6500 meters
4
5 Abstract
6 Understanding the topographic patterns of the seafloor is a very important part of
7 understanding our planet. Although the science involvedof in bathymetricy surveying
8 has advanced much over the decades, less than 20% of the seafloor topography has
9 been preciselyfinely modeled up to nowdate, and there iswith an urgent need to
10 address the improvement theof accuracy and reducetion theof uncertainty of
11 underwater survey data. In this study, we introduce a pre-trainedpretrained VGGNet
12 method based on deep learning. To apply this method, weby inputting gravity
13 anomaly data derived from ship-measurement measurements and satellite altimetry
14 into the model, and correcting the latter, which haswith a largerhigh spatial coverage,
15 based on the former, which is considered, i.e., the trueth value and, iswith morehigh
16 accuratecy. After obtaining the corrected high-precision gravity model, it is inverted
17 to the corresponding bathymetric model by applying the gravity-depth correlation. We
18 choose four data pairs collected from different environments, i.e.,in the Southern
19 Ocean, Pacific Ocean, Atlantic Ocean and Caribbean Sea,, respectively, to evaluate
20 the topographic correction results of the model. The experiments show that the R 2 is
21 reachesup to 0.834 among the results of the four experimental groups, signifying a
22 high correlation. StandardThe standard deviation (SD) and normalized root mean
23 square error (NRMSE) are also evaluated, andwith their accuracy of their
24 performance improved by up to 24.2%, compared with similar research done in recent
25 years. The evaluation of the R2 valuesperformance at different water depths shows
26 that our model can achieve a performance results above 0.90 inat certain water depths,
27 and at the same time can also achieve a significantly improvement results fromin the
28 mid-water depths when compared to previous research. Finally, the bathymetry
29 corrected by our model is able to show an accuracy improvement level of more than
30 21% within 1% of the total water depths, which is sufficient to prove that the
31 VGGNet-based method has the ability to perform a gravity-bathymetry correction and
32 achievewith outstanding results.
33
34 Key words: gravity anomaly, bathymetry inversion, VGGNet, multibeam sonar,
35 satellite altimetry.
36

37 1 Introduction

38 Submarine topographicy survey issurveys are a fundamentalbasic aspect of marine

1
 1 mapping task, the purpose of which is to obtain the three-dimensional coordinates of
2 submarine topographic points, including information such as position, water depth,
3 water level, sound velocity and attitude bearing;, the corecentral component of
4 thesewhich is bathymetry (Zhao et al., 2017). Shipborne sounding surveys areis the
5 most direct and primitive method used to determinedetect the topography of the
6 seafloor, and the methods have undergone an iterative evolution from single-beam to
7 multibeam multibeam echosoundersin terms of detection methods. and at At present,
8 multibeam echosounder sonar system issystems are still one of the main methods for
9 mappingdetecting the topography of the seafloor (Wu et al., 2021). Multibeam sonar
10 method has the advantage of high spatial accuracy, which has enabled the collection
11 of underwater sounding datamode to make a largehigh- leap in quality leap from the
12 acquisition of point measurements to continuous lines and from lines to surfaces (Li,
13 1999). However, the disadvantages of low efficiency, high cost and the long
14 measurement time required make it difficult to conduct frequent submarine surveys
15 over a wide range of marinesea regionsareas (Colbo et al., 2014; Hughes Clarke,
16 2018). As a result, the area coveredage byof shipborne surveys remains very sparse at
17 present. It is estimated that only less than 20% of the global ocean floorsubmarine
18 area hasis been surveyedcovered by shipborne systemssurvey, and a significant
19 proportion of this area, especially in deep-sea areas, has relatively low accuracy
20 (Coley, 2022).
21
22 Satellite altimetry is a spatial measurement technique that uses artificial satellites as
23 carriers to measure the height of satellites relative to the Earth's surface using radar,
24 lasers and other ranging techniques to obtain the topography of the Earth's surface
25 (Zwally et al., 2002). Satellites carrying radar altimeters measure the shape of the
26 global sea surface along their orbits, from which maps of the ocean gravity field can
27 be produced. The satellite gravity field and existing depth measurements are used to
28 determine the relationshipcorrelation between the gravity field and seafloor
29 topography, and by applying thisit relationship to the gravity field, the seafloor
30 topography can be predicted within certain radar bandbands (Smith and Sandwell,
31 1994; Smith and Sandwell, 1997). In most cases, the accuracy of the seafloor
32 topographicy model obtained via gravity inversion is accurate to within ±100 m by
33 gravity inversion. Although this model is a predicted dataset with errors, it is still the
34 best estimate currently available for the large-scale and even global submarine
35 topography available currently.
36
37 Parker (1972) derived a detailed expression for gravity in the frequency domain and
38 proposed a material interface to describe variations in the gravity anomaly model
39 caused by fluctuations that could be applied tofor the frequency domain model of
40 gravity anomaly variation caused by fluctuations, which laid the foundation for the
41 development of submarine inversion topographicy inversion. Ibrahim et al. (1972)
42 proposed the Gravity Geological Methodgravity geological method (GGM) tofor
43 measure rock heights measurement under submarine terrestrial sediments, which has
44 since been widely used (and continuouslypurposefully improved) forin the
2

1
 1 applications of inversion of seafloor topography in various seas (Oldenburg, 1974;
2 Braitenberg, 2006; Ouyang et al., 2014; Wei et al., 2021; An et al., 2022). The
3 Admittance Theoryadmittance theory of gravity anomalies is based on the spectral
4 relationship between the oceanic gravity field and the seafloor topography and was
5 proposed in conjunction with the equilibrium assumption of Airy (1855) and Parker
6 (1972) to estimate the seafloor topography., Thisand was applied toin the study of
7 modeling the submarine mountains (Watts et al., 2006; Gong et al., 2021) and to the
8 inversion of the seafloor topography (Ouyang et al., 2015; Fan et al., 2020). Over the
9 years, researchers have used these methods to produce a variety of global seafloor
10 topographicy models, such as the ETOPO, GEBCO, DTU and SIO series.
11
12 DespiteWith the increasing need for accuracy in marine scientific research in recent
13 years, there is limited room for improving the accuracy of submarine topographicy
14 models obtained by the traditional gravity anomaly inversion of gravity anomalies
15 derived from satellite altimetry;, andwhile shipborne multibeam sonar and gravimetry,
16 which have high accuracy characteristics, are often confined to areas of relatively
17 limited coverage (Yale et al., 1998; Shi et al., 2017). Therefore, there is an urgent need
18 to combine the advantages of their respective spatial coverages and spatial resolutions
19 to developachieve a submarine topographic inversion method that covers a large
20 range, is highly accuratecy and has high flexibility of submarine topographic
21 inversion.
22
23 In recent years, deep learning has gradually become an important scientific computing
24 tool in the field of digital image processing. Deep learning uses a layer-by-layer
25 abstraction approach to efficiently and accurately achieve the extraction of bottom- to
26 top-level attributes, and has made great contributions in the fields of natural language
27 processing (Otter et al., 2020), image and speech recognition (Wu et al., 2015), and
28 remote sensing data fusion (Benedetti et al., 2018; Yuan et al., 2020). Jena et al.
29 (2012) developed an artificial neural network (ANN) model based on the radial basis
30 function (RBF) to predict undetected seafloor topographic features using satellite
31 altimetry-derived gravity data, with the results showing better accuracy than global
32 bathymetry models. Moran (2020) provideds an experiment-based evaluation of a
33 variety of existing machine learning methods applied toin global bathymetric
34 inversions and gaveives recommendations for model selection based on different
35 regional conditions. Fan et al. (2020) introduced a method for predicting seafloor
36 topography that involveding vertical gravity gradients and useding a nonlinear least
37 squares algorithm, with the accuracy of its results demonstrating the feasibility of this
38 algorithm compared with global seafloor topographicy models. Zhu et al. (2021) used
39 a multilayer perception (MLP)-based neural network to correct satellite altimeter
40 gravity by shipborne data. A comparative analysis was found to achieve an effective
41 and large-scale refinement of the accuracy of the satellite-derived gravity anomaly
42 model. Seoane et al. (2022) used an extended Kalman filter (EKF) method that
43 iteratively integrateds continuous input gravity anomaly and geoid height data to
44 simulate regional topographic models of the seafloor in areas lacking ship-measured
3

1
 1 bathymetry. Annan and Wan (2022) introduced the signal of vertical deflection s in
2 bathymetric inversion, which was combinedncatenated with gravity anomalies and
3 vertical gravity gradients by an outperforming CNN, demonstrating the usability of
4 vertical deflections in bathymetric inversion. Chen et al. (2022) applied a pretrained
5 VGGNet model to applyperform corrections to the satellite-derived bathymetry based
6 on multibeam sonar data, and the results showed that this method could achieve
7 significantly enhancement to the accuracy of the satellite data.
8
9 Previous studies have often pointed out noted that the main problem associated with
10 the algorithm development or model formulation for the estimation of seafloor
11 topographic parameters is the complexity of the physical processes involved and the
12 uncertainties associated with them. Therefore, the selection of a suitable model is a
13 very critical step before starting the inversion work. Gatys et al. (2015; 2017)
14 proposed a method based on VGGNet, a convolutional neural network (CNN)
15 commonly used for high-precision visual recognition, to computationally extract the
16 style and content information of a set of input images and output a new fused image
17 with both style and content information fromof the input ones. Inspired by itthis, we
18 can treat the shipborne gravity data and satellite altimeter-derived gravity data as the
19 input of style and content information, respectively, in order to synthesis a to
20 synthesize new image data through the VGGNet.
21
22 In this study, a CNN-based optimization algorithm is proposed to correct the satellite
23 altimetry-derived gravity data that containswith errors through the useing of a pre-
24 trainedpretrained VGGNet network with shipborne marine gravity data applied as the
25 correcttruth value, logically speaking. Structurally, the input of the model input is the
26 data to be corrected, and the output is the data with the true values data;, andwith the
27 aim is toof identifyobtaining the high -order image features in the intermediate layer.
28 The role of the intermediate layer is to establish thea complexdeep abstraction
29 relationship between the input and output values, i.e., to deriveconvert the original
30 image pixels into a complex understanding of the image features from the original
31 image pixels and to construct a representation of the internal mapping. After
32 calibrating the satellite-derived gravity and inverting it to the seafloor topographicy
33 model of the corresponding area from it, a more accurate inversion of the bathymetry
34 can be obtained.
35
36 The main contributions of this study are as follows.
37
38 1. A method is proposed that uses deep learning to correct the gravity anomaly
39 data derived from satellite altimetry based on deep learning with shipborne
40 data used as the trueth value, and to derive a high-precision seafloor
41 topographicy model accordingly.
42 2. The CNN-based VGGNet model is first introduced to generate a satellite-
43 derived data with improvedcorrected accuracy by calculating and minimizing
44 the distance (loss) between the shipborne trueth data and the satellite data
4

1
 1 (which containswith errors). The new fused data will havepossess the
2 advantages of both the high accuracy of the former and the high coverage of
3 the latter.
4 3. The performance of the model is validated by experiments on datasets from
5 the Southern Ocean, Pacific Ocean and Atlantic Ocean, with the results
6 showing that this method is effective forin improving the computational
7 accuracy of the inversion of the seafloor topographicy inversion.

8 2 Methodology and Data

9 The research structure of this paper is shown in Figure 1. Overall, relying on gravity
10 data pairs derived from ship measurements and satellite altimetry, the pre-
11 trainedpretrained VGGNet model is able to implement the correction theof satellite
12 data with ship measurements as (the trueth value in the framework of a neural
13 network). After obtaining the corrected satellite altimetry-derived gravity data, the
14 corresponding bathymetric maps can be obtained by inversion, and the accuracy can
15 be verified with the help of shipborne multibeam sonar bathymetry data, which can be
16 considered as measured values.

1
 1

2 Fig. 1. The research structure.

3

4 2.1 Framework of VGGNet

5 In recent years, convolutional neural networks (CNNs) have made great progress in
6 the field of patternintelligent recognition inof digital images, and some excellent
7 algorithmic models, such as AlexNet (Krizhevsky et al., 2012), CaffeNet (Jia et al.,
8 2014) and VGGNet (Simonyan et al., 2015), have emerged. VGGNet applies a very
9 small field of viewperception ratherinstead thanof a large area field, and
10 considersusing 3 × 3 pixels with a step size of 1. The decision function is more
11 discriminative because there are 3 ReLU units. There are also fewer parameters, with
12 27 times the number of channels. Without modifying the perceptual field, VGGNet
13 uses a 1 × 1 convolutional layer to make the decision function more nonlinear. Due to
14 the tiny size of the convolutional filter, the VGGNet model can have a considerable
15 number of weight layers, which means better performance potential. There are two
16 general forms of VGGNet, VGG-16 and VGG-19, which are not fundamentally
17 different, but only differ in the depth of the network, with VGG-19 having three more
6

1
 1 convolution layers. Considering the experimental requirements and data parameters,
2 VGG-19, which can offer deeper information, is chosen in the experiment. Compared
3 with most previous CNN-originated models that have 4–7 layers, VGG-19 is
4 constructed with 19 layers, including 16 convolutional layers and 3 fully connected
5 layers, enabling it to extract the more abstract and more profound image features and
6 reduce the number of parameters while being able to retain the same receptive field.
7 Thus, it has improved the efficiency and accuracy of image computing (Huo et al.,
8 2020; Islam et al., 2020; Schulz et al., 2020).
9
10 The structure of VGG-19 is displayed in Figure 2. The entire network uses the exact
11 size of the convolution kernels (3 × 3) and the maximum pooling kernels (2 × 2). The
12 combination of several small filter (3 × 3) convolutional layers is better than a large
13 onefilter (5 × 5 or 7 × 7), as in the previous models. Since the convolution kernel
14 focuses on expanding the number of channels and the pooling kernel focuses on
15 reducing the width and height, the architecture is more profound and broader.
16 AdditionallyAt the same time, the increase in the number of calculations slows the
17 process down, showing the network a larger receptive field. FurthermoreAt the same
18 time, the network parameters are reduced, and the ReLU (Rectified Linear
19 Unitrectified linear unit (ReLU) activation function is used in convolutional layers
20 and fully connected layers to create more linear transformations in order to enhance
21 the learning ability.
22
23

24
25 Fig. 2. Architecture of the VGG-19 model. The boxes represent the size of each layer.
26

27 2.2 Model Training Steps

28 We perform the correction of ocean gravity data using the model of VGG-19 model.
7

1
 1 The structural framework is to take the satellite altimetric gravity data to be corrected
2 as the input values and the shipborne gravity data to be treated as trueth values as
3 output. These procedures allow the machine to train a rationalization model in the
4 intermediate layer that minimizes the loss (distance) of this pair of datasets. The
5 principle of the correction model is to define a distance function that describes how
6 different the two input images are. The gridded shipborne data image and the satellite
7 altimetry-derived data image that coversing the same area are passed to the model,
8 which is supposed to return the intermediate layer outputs from the model. The
9 distance function L that we use is shown below:
Ll ( x , p )=∑ ( F lij ( x )−Plij ( p ) )
2
(1)
i, j

10 where x stands for the shipborne gravity image, p stands for the satellite altimetry-
11 derived gravity image, and i∧ j standstands for the serial number of pixel points of
12 the input images. Let V nn be a pre-trainedpretrained VGG-19 network and X be any
13 image,; then, is the network fed by . Let l ∈ and l ∈
V nn ( X ) X F ij ( x ) V nn (x) Pij ( p ) V nn ( p)
14 describe the respective intermediate feature representation of the network with the
15 inputs x and p at layer l . At lastFinally, optimizers’ rules are applied to iteratively
16 update the images, which minimizes a given loss to the inputs.
17
18 The evaluation of the correctedion precision is based on comparisons with thea
19 previous study. To quantify the differences and connections between the predicted
20 value and trueth value, here we choose four evaluation measurements, root-mean-
21 square error (RMSE), normalized RMSE (NRMSE), coefficient of determination (R2),
22 and standard deviation (SD), as follows, respectively:

√
n

∑ ( f^i− y i )
2

i=1
(2)
RMSE=
n
RMSE
NRMSE= (3)
y max − y min
n

∑ ( y i−f i )2
R2=1− i =1
n (4)
∑ ( y i− y ) 2

i=1

√
n

∑ ( f i−f )2 (5)
i =1
SD=
n
23 where n represents the number of the values from the dataset, i represents the serial
24 number of the value from the dataset, f represents the predicted values, and y
25 represents the trueth values. As a standardization process, NRMSE is able to limit the
26 value of RMSE to a certain range, eliminate the influence of the scale between
8

1
 1 indicators, and solve the comparability among data indicators. NRMSE and R2 usually
2 range from 0 to 1. TheA smaller RMSE , NRMSE , SD and biggerlarger R2 mean that
3 there is a higher correlation and more stable distribution amongof the datasets.
4
5 The default hyperparameters used in the VGG-19 training model are shown in Table
6 1. In the experiments, the weights between the content layer, style layer, content, style
7 and total variance loss are set to maintain approximately the same order of magnitude.
8 The learning rate is set so that the iterations occur in a convergent process and that the
9 loss curves decrease smoothly.
10
11 Table 1. The default hyperparameters of the training model.
Hyperparameters Settings
Content layer ‘conv4_2’
‘conv1_1’, ‘conv2_1’, ‘conv3_1’, ‘conv4_1’,
Style layers
‘conv5_1’
Weight of loss at content
1
layer
Weights of loss at style
1, 1, 1, 1
layers
Weights among content,
1 × 10-4, 1, 1 × 10-5
style, and total variation loss
starts at 10 and, linear decays over 100 iterations to
Learning rate
1
12

13 2.3 Bathymetric Inversion of Bathymetry

14 According to the flexural isostasy theory proposed and summarized by Watts (1978;
15 2001), a certain degree of flexural isostasy (also known as equilibrium compensation)
16 r ( x ) occurs in the Moho belowsurface of the corresponding oceanic crust whenunder
17 the action of the topographic undulations h(x ) are applied toof the seafloor. Thus, the
18 main anomalous field sources that generate the sea surface gravity anomaly are the
19 seafloor topography and its equilibrium compensation material. The gravity anomaly
20 produced by a certain seafloor topography can be expressed as:
∞ ∞
21 Δ G ( k )=2 πG ( ρc −ρ w ) e−kd ∑ H n ( k ) +2 πG ( ρm −ρc ) e−k (t +d ) ∑ R n ( k )
n=1 n =1
(6 )
22 where Δ G ( k ) is the Fourier transform of the gravity anomaly, G stands for the gravity
23 constant, of gravity, and ρc , ρw and ρm represent the mantle, crust and seawater
2π
24 densitydensities, respectively. k = , where λ represents the topographic wavelength.
λ
25 d stands for represents the average water depth, while t represents the average oceanic
26 crust thickness. R(k) and H (k) stand for represent the Fourier transformtransforms of
9

1
 1 r ( x ) and h( x ), respectively, and n is the exponential power, taken as a positive integer
2 (Parker, 1972). While in In the frequency domain:,
ρc − ρw '
3 R ( k ) =−H ( k ) Φ (k )
ρm −ρc e
(7)
4 where ' is called the flexural response equation and expressed as:
Φe ( k )

[ ]
−1
' D k4
5 Φ (k )=
e +1
( ρ m− ρ c ) g
( 8)
3
6 where D represents the lithosphere flexural stiffness and D= ETe . E represents
2
12(1−ν )
7 the Young’s modulus, T e represents the effective elastic thickness, ν represents the
8 Poisson's ratio, and g stands for gravitational acceleration (Walcott, 1970).
9 Empirically, Eq. (6) converges quickly, and when only the linear term of this equation
10 (n = 1) is taken, taking into account, Eq. (7) and Eq. (8) at the same time, the gravity
11 anomaly generated by the seafloor topography of the seafloor can be expressed as:

12 Δ G ( k )=2 πG ( ρc −ρ w ) e−kd H ( k ) [ 1−Φ 'e ( k ) e−kt ]

(9 )
13 thatThat is, the admittance function Z(k ) from the seafloor topography to the sea
14 surface gravity anomaly can be expressed as:

15 Z ( k )=2 πG ( ρ c − ρw ) e
−kd
[ 1−Φ 'e ( k ) e−kt ]
( 10 )
16
17 The theoretical admittance function indicates that the sea surface gravity anomaly is
18 strongly correlated with the seafloor topography only in the short and medium
19 wavelengths due to the effects of equilibrium compensation and water depth extension
20 (Hu et al., 2020). Therefore, based on the flexural isostasy theory, gravity anomalies
21 are usually used only for the inversion of seafloor topography in the short and
22 medium wavelengths (10–20 – 200 km), while shipborne bathymetry is used to
23 construct long wavelength (>200 km) seafloor topography.
24
25 From the above, the flexural response equation part in Eq. (9) can be neglected in the
26 short wavelength band (<200 km). The inversion of the seafloor topography from the
27 sea surface gravity anomaly is given by:
ekd
28 H ( k )= Δ G(k )
2 πG ( ρ c −ρw )
( 11 )
10

1
 1
2 For the obtained corrected satellite-derived gravity data, we invertinverted them using
3 the filter method (Smith and Sandwell, 1994) to obtain the corresponding bathymetric
4 data. The filter method combines flexural isostasy theory, admittance theory and the
5 gravity noise signal ratio to propose a low-pass and band-passbandpass filter-based
6 method to predict bathymetry using gravity data.
7
8 The short- and mid-wavelength bathymetry is calculated using a scaling factor S(x ),
9 which is determined empirically by regression and reflects the correlation between the
10 gravity data and the bathymetry data in the corresponding band. Eventually, the
11 seafloor topographicy model will consist of two parts: (i) a long-wave
12 (>200kmlongwave (>200 km) model constructed by shipborne bathymetry;
hlong ( x )

13 and (ii) a short-shortwave and mid-wavemedium wave (10–20 – 200 km) model
14 inversed by band-limited gravity anomaly δg ( x ), as shown in Eq. 12, with the
15 processing flow shown in Figure 3.
h ( x )=hlong ( x ) + S ( x ) ∙ δg ( x )
16
( 12 )
17
18

19
20 Fig. 3. The calculation and processing flow of the filter method.

21 2.4 Experimental Data

22 In the experiment, the satellite altimeter-derived gravity anomaly data used are
23 acquired fromat the Bureau Gravimetrique International (BGI) (Bonvalot et al., 2012).
24 The shipborne gravity anomaly data and multibeam sonar bathymetry data used are
25 acquired fromat the National Geophysical Data Center (NGDC) of the National
26 Oceanic and Atmospheric Administration (NOAA) (NOAA National Centers for
27 Environmental Information, 2004; 2015). For data fromof different sources, rasters
28 were resampleding is applied to the sameunify their resolution for subsequent
29 analysis.
30
31 A total of four pairs of shipborne gravity-satellite gravity-multibeam bathymetry
32 datasets from the Southern Ocean, Pacific Ocean, Atlantic Ocean, and Caribbean Sea,
33 respectively, were selected, with detailed information listed in Table 2. We have not
11

1
 1 included the temporal resolution of these data in the table, taking into account the
2 temporal scale of ocean variability.
3
4 Table 2. The parameters of the datasets.
Center Spatial Data Size Area Bathymetry
Point Resolution (km2) Depth
Coordinates (m) Range (m)
Southern 71°S, 173°E 93 5,097,104 43,700 －4077~－
Ocean 211
Pacific 9°S, 140°W 93 33,048,000 283,337 －4992~－
Ocean 113
Atlantic 32°N, 65°W 93 3,240,000 27,778 －4920~－
Ocean 58
Caribbea 18°N, 82°W 123 10,614,363 150,310 －6580~－1
n Sea
5
6 In the correction process of correction, we randomly split theeach gravity datasets
7 fromin each region in two, with 50% used as the training set to debug the parameters
8 in the network and the remaining 50% used as the validation set to tune the
9 hyperparameters of the model and to conduct a preliminary evaluation of the
10 performance of the model.

11 3 Experimental Results

12 3.1 Model Performance Evaluation

13 A loss function is applied to estimate the degree of inconsistency between the

14 predicted (corrected) and truthtrue values of the model, as shown in Figure 4. It is a
15 non-negativenonnegative real-valued function, with the smaller the loss function, the
16 better the robustness of the model, and vice versa. As can be seen from the results, in
17 the experiments conductconducted in the four regions, a sudden drop in loss occurred
18 in the first 20 epochs. After dropping to between 0.1 and 0.2, the loss starts to decline
19 gently, with this phenomenon being more pronounced after 70 epochs. No significant
20 overfitting is found from the curves. This performance of the loss function
21 demonstrates the usefulness of the VGG-19 model for gravity correction.
22

12

1
 1
2 Fig. 4. The training and validation loss from the experiments in theof (a) Southern Ocean, (b)
3 Pacific Ocean, (c) Atlantic Ocean, and (d) Caribbean Sea.

4 3.2 Overall Accuracy Analysis

5 The output of the model results forin a corrected gravity anomaly, with its verification
6 of accuracy listed in Table 3. As can be seen from the Tabletable, the R2 of the
7 network-corrected gravity anomaly model can generally reach a high level of 0.90 or
8 more, indicating that it is an excellent matches with the trueth values fromdata of the
9 shipborne measurements quite well. The SD is similar to the RMSE in terms of
10 performance. The results of the normalized RMSE can provide the opportunity to
11 conduct a comparison with previous research, and it is found that the accuracy of the
12 output fromby our method can reach an level of accuracy that is indistinguishable
13 level from resultsthe accuracy obtained when using the Multi-Layer
14 Perceptronmultilayer perceptron (MLP) method, with respective performances (Zhu
15 et al., 2021). These results provide a solid basis for the next step of bathymetry
16 inversion step.
17
18 Table 3. The overall Overall accuracy of the gravity correction.
R2 SD (mGal) RMSE (mGal) NRMSE
13

1
 Southern Ocean 0.902 18.333 13.630 0.113
Pacific Ocean 0.955 17.892 10.050 0.118
Atlantic Ocean 0.930 19.567 21.549 0.114
Caribbean Sea 0.919 21.051 16.485 0.113
1
2 To visualize the correction results, we arbitrarily select three sites with visual
3 differences toand representativeness in each of the four experimental areas for
4 comparative display, as shown in Figure 5. As can be seen from the Figurefigure,
5 generally speaking, the texture of the corrected data texture is much finer and more
6 consistent with the actual gravity field distribution on the seafloor. This provides a
7 foundation for the following work ofon bathymetry inversion.
8

14

1
 1
2 Fig. 5. Comparison of local details of gravity anomalyanomalies before and after correction from
3 the (a) Southern Ocean, (b) Pacific Ocean, (c) Atlantic Ocean, and (d) Caribbean Sea.
4
5 The overall accuracy of the bathymetry correction process has been verified with the
6 multibeam sonar bathymetry used as the trueth value, and the performance of the
7 parameters is presented in Table 4. In terms of R2, tThe four datasets all haveperform
8 similarly, R2 values ofall aroundapproximately 0.80, indicating a high correlation
9 between the corrected bathymetry and the true values. There is some degree of
10 positive correlation between the differences in the performance of the SD and RMSE
11 for the four datasets, with values relating to the range of water depths in the
12 experimental area. A comparison of NRMSE performance usingagainst the same
15

1
 1 criteria shows that the Atlantic data have the best correction accuracy, followed by the
2 Southern Ocean and Pacific Ocean data, while the Caribbean Sea onesdata rank the
3 last. In comparison with similar studies in recent years, we have found that this
4 method is effective forin improving accuracy by up to 15.4% (gravity-bathymetry
5 correction) or 24.2% (bathymetry-only correction), demonstrating that this VGGNet
6 method enables a new level of accuracy in seafloor topography correction (Annan and
7 Wan, 2022; Chen et al., 2022).
8
9 Table 4. The overall accuracy of the bathymetry correction.
R2 SD (m) RMSE (m) NRMSE
Southern Ocean 0.822 104.790 107.024 0.027
Pacific Ocean 0.834 117.630 126.366 0.026
Atlantic Ocean 0.833 124.847 136.622 0.028
Caribbean Sea 0.783 139.583 164.475 0.025
10
11 By comparing the variation in R2 values at different water depths, we plot the
12 relationship between the two parameters, as shown in Figure 6. It can be seen that for
13 For the results of all four experimental areas, the minimum values of R 2 are above 0.2
14 and, all occurring at the extremes of the water depth. The maximum values of R2 are
15 all above 0.9 and, whose occurring depth values where those values are found vary
16 from one area to another, indicating that the correction model is able to invert
17 bathymetry with high accuracy over a range of water depths.

18 3.3 Inter-ComparisonIntercomparison of Accuracy

19 Looking at the data separately for each region, we see that in the Southern Ocean, the
20 accuracy reachesclimbs to a maximum valueplateau in the water depth range around
21 of approximately -2000 m to -1600 m and then begins to slowly decline, with a steep
22 drop occurring around at approximately -200 m. In the Pacific Oceandata, the
23 accuracy increasesclimbs more rapidly from the beginning, reaching an extreme value
24 near -4000 m, and then begins a gentle decline to near 0.5. The curve for the Atlantic
25 data generally behaves generally similarly to the former dataset, with
26 accuracaccuraciesy rising rapidly to a maximum near -4000 m beforeand starting ato
27 gentle decline to aroundapproximately 0.4 gently, however, there iswith a large
28 downward trend in the last 100 m of the gradient. Finally, in the Caribbean Sea data,
29 the curve shows a different trend, with more isolated points at the water depth
30 extremes and a more rugged overall curve trend. The fit lines show that the accuracy
31 of the data in this region has local peaks at -5000 m, -4000 m to -3000 m, and -2500
32 m, with the last one being the highest, and the changes at the depth extremes are both
33 steeper.
34
35 As a rule of thumb, the accuracy of deep learning models tends to be positively
36 correlated with the sample size. With other parameters held constant, the larger the
37 data sample is, the higher the training accuracy tends to be, and vice versa. In the
16

1
 1 experiment, the analysis combined with the histogram of the data pointspoint
2 distribution shows that the lowest R2 tends to be foundexhibited where the number of
3 data points is the lowest or sparsest, such as at the extremes of water depth extremes.
4 In contrast, the depths where higher levels of R 2 occur tend to be the depths with the
5 largest and most concentrated distribution of data points. Then, the underwater
6 morphological specificity in the experimental areas can also have some impact on the
7 training accuracy. For example, in the Caribbean Sea dataset, the large and
8 fragmented distribution of local islands can lead to greater irregularity in the gravity
9 and bathymetric undulations within some regions, which may become a factor
10 affecting the output accuracy. In addition, since the spatial resolution of the Caribbean
11 Sea data is aboutapproximately a quarter the sizelower than that of the other three, it
12 may also be one of the reasons thatfor the less satisfactory accuracy
13 resultsperformance fromof this dataset were less satisfactory.
14

15
16 Fig. 6. Performance of R2 values at different water depths and its comparedison with the method
17 of bathymetry-only correction methos infrom the (a) Southern Ocean, (b) Pacific Ocean, (c)
18 Atlantic Ocean, and (d) Caribbean Sea.
19
20 To make the accuracy validation more comparable, we reworked each of the four
17

1
 1 datasets according to the bathymetry refinement method presented in Chen et al.
2 (2022) and overlay the fitted trend lines of the accuracy results shown in Figure 6 for
3 comparison. It can be summarized that, in general, our gravity-bathymetry correction
4 method can effectively improve the accuracy compared to the bathymetry-only
5 method. ForWhen considering the maximum performance of the accuracy withinof
6 the datasetsdata set fromin each experimental area, we see that there is not much
7 difference between the two methods. However, the gravity-bathymetry method is able
8 to improve the accuracy mainly in the area of intermediate water depth ranges.
9 Specifically, we see when wein the compareison of the performance results of the
10 Southern Ocean dataset, it can be observed that the R2 valueperformance of the
11 bathymetry-only method has startsed to decline starting near -2500 m, while the R2
12 values performance fromof our method beginhas just started to surpassovertake those
13 values at that depth and is preparing to reaches a stable plateau until
14 approximatelywithin -1000 m, which is when the two methods alignhave their
15 respective leads. In the comparison of the Pacific dataset, theour performance of our
16 results exceeds the other set of resultsleads from the extreme maximum water depths
17 until around -4000mapproximately -4000 m. Between approximately -3000 m and -
18 1000 m, our method also appears to performbe significantly betterhigher performing.
19 However, in shallow waters up to -1000 m, our method underperforms the
20 bathymetry-only method. In the Atlantic dataset, the performance of our method
21 performs better leads for a long time within the depth range of approximately -4500 m
22 to -1000 m, but is again overtaken in shallow waters. In the Caribbean Sea dataset, the
23 two methods perform similarlyclosest to each other, with alternating leads occurring
24 from time to time. Our method outperforms the bathymetry-only method mostly
25 between depths belowin the range of extreme depths to near -5000 m, between
26 aboutapproximately -4500 m andto -3500 m and between -2000 m andto -1000 m.
27
28 It can be discovered that the The gravity-bathymetry correction method has a more
29 excellent better R2 valuesperformance in the intermediate water depthsdepth range
30 and can improve the overall accuracy as well,; however, if we focus only on depthsthe
31 range aboveup to approximately -1000 m, theits performance of it, on the contrary, is
32 inferior to that of the bathymetry-only method. ThisIt is presumed to be related to the
33 band-limited correlation of the satellite gravity-bathymetry signals within the
34 corresponding water depths (Smith and Sandwell, 1994; 1997; Sandwell et al.et al,
35 2022).
36

18

1
 1
2 Fig.Figure 7. Percentage distributions of NRMSE (×10-3) performance in datasetdatasets from the
3 (a) Southern Ocean, (b) Pacific Ocean, (c) Atlantic Ocean, and (d) Caribbean Sea.
4
5 We also apply NRMSE, the normalized metric of RMSE, for the comparative
6 evaluation of the correction effect in the four experimental areas, as plotted in Figure
7 7 (the values on the horizontal axis are magnified by a factor of 1000 for simplicity
8 ease of display). From the figure, it can be found that the NRMSE performance curves
9 of the four regions almost all follow the form of high in the middle - low at the edges ;
10 thus, thusit can be concluded that the majority of the data points are distributed in the
11 interval of moderate performance. The extreme values of the performance of the four
12 regions are almost identical and are distributed between 0.010 and 0.035. No obvious
13 prominent peak areas can be seen in the performance curves of the Southern Ocean
14 data,; instead, a relatively stable plateau distribution with localized ups and downs is
15 observed. The peak in the Pacific data occurs aroundat approximately 0.017 and then
16 begins to decline, but there is a small local peak aroundat approximately 0.032,
17 indicating a less desirable sudden change in accuracy at certain data points. The
18 Atlantic dataset behaves similarly to that of the Pacific, with the difference that it has
19 a bimodal structure with two local peaks between 0.015 and 0.020 to 0.025, with the
20 former of which being higher than the latter. It is worth mentioning that the
21 percentage of data points corresponding to the maximum peak in the curve of the
22 Atlantic Ocean is the highest among the four data setsdatasets, which proves the more
23 concentrated distribution of its high-precision data points. Compared with the right-
24 skewed distribution exhibited by the upper two datasets, the curve of the Caribbean
25 Sea dataset shows a clear left-skewed distribution, i.e., a larger distribution of lower
26 precision data points, symbolizing a relatively unsatisfactory precision, despite its
19

1
 1 better-than-average peak share. It is also evident from Table 4 that the accuracy of the
2 Caribbean data is relatively the lowest.
3
4 Table 5. Proportion of corrected errors from trueth values within 2% and 1% of the depth range.
1% of depth (%) 2% of depth (%)
Southern Ocean 68.05 75.69
Pacific Ocean 61.53 77.52
Atlantic Ocean 57.58 68.69
Caribbean Sea 51.03 64.58
5
6 To demonstrate the effect of the correction more intuitively, we then subtract the
7 corrected bathymetric values from the truthtrue values. For the The absolute values of
8 the obtained differences, they are calculated as a percentage of the total water depth,
9 with the results shown in Table 5. As can be seen from the table, the number of data
10 points decreases as the margin of error narrows. The mean percentage of data points
11 in the four experimental areas is aboutapproximately 71.62% within a 2% water depth
12 range error, while about and approximately 59.55% within a 1% error. A comparison
13 with the results of the bathymetry-only correction method proposed in Chen et al.
14 (2022) shows that our gravimetry-bathymetry combined correction method, while
15 improving the accuracy by a mere 1.47% in the 2% water depth range error, improves
16 by a larger 21.01% in the 1% error range. In this comparison, it can be concluded that
17 the accuracy improvement provided byof our novel method for accuracy lies to a
18 large extent within the 1% water depth range error, despite of its lack of significance
19 in the 1% to 2% interval.
20
21 In summary, among these evaluation indicators, the correction performance of this
22 correction method in the Southern Ocean is relatively better than in the other
23 basinsthe most outstanding, while the Pacific Ocean and the Atlantic Ocean
24 performances are similar, and each has its own leading edge, and the performance of
25 the method in Caribbean Sea ranks thelowest fourth amongstin most cases.

26 4 Conclusions

27 In this study, we propose a deep learning-based inversion solution to address the

28 current problem of the lack of high-precision topographic seafloor modeling of
29 seafloor topography. We output the refinesment of the satellite-derived gravity
30 anomaly model, which was based on a pre-trainedpretrained VGGNet model, by
31 training the shipborne gravity anomaly data while usingby the inspiration of a neural
32 style transfer algorithm to minimize the distance (loss) between the two datasets. For
33 the obtained corrected gravity anomaly model, we use the conventional filtering
34 method to invert it and derive what can be considered as a corrected bathymetry
35 model.
36

20

1
 1 Four gravity-bathymetry data pairs from the Southern Ocean, Pacific Ocean, Atlantic
2 Ocean, and Caribbean Sea are selected to evaluate the accuracy of the corrected
3 bathymetry model. Through the evaluation based on R2, SD, RMSE and NRMSE
4 metrics, we find that it can improve the accuracy by up to over 15%, compared with
5 the previous gravity-bathymetry correction method, or over 24%, compared with the
6 previous bathymetry-only correction method. WithBy plotting the water depths-R2
7 and RMSE proportion as a function of water depthplotting, we conclude that this
8 method can effectively improve the fineness of the bathymetry. More quantitatively,
9 our method offershas the best correction for the top 1% of the total depth range of
10 data points, which is more than 21% better compared to than previous research. This
11 demonstrates that the proposed model is useful and efficient as a novel approach to
12 modeling submarine topography on a large scale of refinement.
13
14 As for For the factors affecting the differences in accuracy in different experimental
15 areas, as mentioned before, the specificity of the seafloor topography can have an
16 important effect. For instance, the scale of the water depth, the complexity of the
17 geomorphology and the interference of land (islands) factors, are all possibilities that
18 can cause errors in the signals of the ship measurement instruments (Smith and
19 Sandwell, 1994; 1997; Koh et al., 2022). For signals transmitted by satellites, the
20 latitude of the operational orbit is a factor that affects their performance. In addition,
21 the band limitations of the gravity-depth signal for altimetry and its relation to the
22 geological environment also exert an influence (Scharroo and Visser, 1998).
23 Combining the R2 and NRMSE in Tables 3 and 4, we can reason that there is some
24 degree of accuracy loss occurring during the operation of the inversion of the gravity
25 model to the bathymetry model. The causes and mechanisms of this loss may require
26 detailed analysisdiscussion of the control variables. These systematic errors are all
27 factors that need tomust be taken into account.
28
29 As forFor the factors of the model itself, the generalization can have an impact on its
30 ability to predict new unknown data. Currently, only the Adam optimizer is applied in
31 this pre-trainedpretrained model. A comparison of different types of optimizers should
32 be discussed before future experiments are undertaken in order to select the one that
33 can balance the loss and generalization performance (Reddi et al., 2018). Furthermore,
34 with larger data involved, instead of dividing the training and validation sets within
35 the dataset itself, it is also worth discussing whether selecting a completely new
36 dataset as the validation set to be applied in the training process will improve the
37 correction accuracy of the model.
38

21