Satellite Altimetry Seabed Topography Correction Method and System Based on VGGNet

Abstract

The present invention discloses a method and system for satellite altimetry bathymetric correction based on VGGNet, which relates to the technical field of deep learning and includes: collecting shipborne multibeam sonar bathymetric data and satellite altimetry bathymetric data as original data; preprocessing the original data to obtain a training data set; constructing a bathymetric correction model based on a convolutional neural network, and inputting the training data set into the bathymetric correction model for model training until the loss function converges to obtain an optimal bathymetric correction model; using the optimal bathymetric correction model to perform water depth correction on satellite altimetry sounding bathymetric data to obtain corrected satellite altimetry water depth data. The present invention uses a deep learning method to correct satellite altimetry water depth data based on multibeam sonar water depth data, providing a new technical means possibility for high-precision and high-real-time bathymetric measurement of the global sea area.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning, and more specifically, to a satellite altimetry-based seabed terrain correction method and system based on VGGNet. Background Art

[0002] Seabed terrain measurement is a basic ocean surveying and mapping task, aiming to obtain the three-dimensional coordinates of seabed terrain points, including information such as measurement location, water depth, water level, sound speed, attitude azimuth, etc. The core is water depth measurement. Shipborne detection is the most direct and primitive method for detecting seabed terrain and landforms. Water depth measurement is the core work of shipborne terrain and landform detection, which has developed from primitive methods such as sounding rods, hammers, and ropes in the early days to current detection methods using sound, light, and electricity. Since light waves and electromagnetic waves attenuate rapidly in water, while sound waves can travel long distances in water, currently, shipborne acoustic detection is still one of the main methods for detecting seabed terrain and landforms. Shipborne water depth measurement has experienced iterative evolution from single-beam to multi-beam in terms of detection methods.

[0003] The multi-beam sounding method has characteristics such as high precision, high efficiency, automation, and digitization, enabling a high-quality leap in the underwater sounding mode from point to line and from line to surface. However, its disadvantages are low efficiency, high detection cost, large consumption of manpower, and long measurement time required, making it difficult to conduct seabed terrain surveys in large areas of the sea. Currently, for the accumulated global ocean sounding data, the coverage of ship-measured sea depth is still very sparse, and the accuracy of early sounding data is not high. The seabed terrain model constructed based on ship-measured water depth has low accuracy and resolution, making it difficult to meet the data conditions required for scientific research.

[0004] Satellite altimetry is a space measurement technology that uses ranging technologies such as radar and laser with artificial satellites as carriers to measure the height from the satellite to the Earth's surface, thereby obtaining the Earth's surface terrain. Through altimetry data, a gravity field model of the ocean can be constructed, and the distribution of gravity anomalies and geoid undulations therein reflects seabed terrain such as seamounts, mid-ocean ridges, and trenches. The emergence of satellite altimetry technology has made seabed terrain measurement no longer limited to traditional shipborne sonars, providing new technical means possibilities for large-scale and high-real-time seabed terrain surveys of the global sea areas. From the analysis of the existing terrain models inverted from satellite altimetry, although their spatial resolution and accuracy have been greatly improved, they still cannot meet the actual needs compared with the multi-beam sounding mode. How to further improve the resolution and accuracy of seabed terrain and achieve fine modeling of the global seabed terrain is a difficult problem that urgently needs to be solved currently.

[0005] In recent years, deep learning has gradually become an important scientific computing tool in various fields. Compared with traditional shallow machine learning algorithms, deep learning adopts a layer-by-layer abstraction method to efficiently and accurately extract low-level attributes from high-level attributes, and has made great contributions in the fields of natural language processing, image and speech recognition. Neural style transfer is an image optimization technique based on deep learning, which is used to obtain three images (content image, style reference image and input image to be styled), transform the basic input image by minimizing the content and style distance (loss) through back propagation, and create an image that matches the content of the content image and the style of the style image. The main problem associated with the algorithm development or model formulation of seabed topography parameter estimation is the complexity of the physical processes involved and the uncertainty associated with them. In this case, advanced computer-based methods such as fuzzy logic, genetic algorithms, artificial neural networks and fractals can be used to derive the required parameters from known influencing parameters. At present, due to the rise of image recognition and artificial intelligence, the research on deep learning is mainly focused on classification, recognition and other aspects, but the research in the field of regression is relatively small. The regression problem involves all aspects of social production and daily life. The research on regression problems has important economic value and social significance. It is a meaningful and feasible research to try to better apply deep learning to the field of regression estimation.

[0006] In summary, there is an urgent need to use multi-source information and methods to establish a global seafloor topography model. The fine modeling of seafloor topography requires breakthroughs in many key theoretical and technical difficulties in the study of fine modeling theory, change characteristics and their mechanisms, and exploration of their mutual connections, spatial distribution and change laws to improve the accuracy and resolution of modeling. It is necessary to make full use of ship-based bathymetry data and combine ship-based bathymetry and satellite altimetry data to perform fine seafloor topography inversion, explore the role and impact of global climate change, material exchange in the earth's spheres, and seafloor plate tectonics, and provide an important basic guarantee for research in geodesy, oceanography, and seafloor plate tectonics.

[0007] Therefore, how to use deep learning methods to correct satellite altimetry and depth data based on multi-beam sonar depth data, and provide new technical means for high-precision and high-real-time seabed topography measurements in global waters is a technical problem that technicians in this field urgently need to solve. Summary of the invention

[0008] In view of this, the present invention provides a method and system for correcting seabed topography based on satellite altimetry based on VGGNet, which uses a deep learning method to correct satellite altimetry depth data based on multi-beam sonar depth data, and can achieve the unification of the spatial resolution of multi-beam data and the temporal resolution of satellite altimetry data, providing new technical means for high-precision and high-real-time seabed topography measurement in global waters.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] A satellite altimetry bathymetric correction method based on VGGNet, comprising the following steps:

[0011] Collect shipborne multibeam sonar bathymetric data and satellite altimetry bathymetric data as raw data;

[0012] Preprocess the raw data to obtain a training dataset;

[0013] Construct a bathymetric correction model based on a convolutional neural network, and input the training dataset into the bathymetric correction model for model training until the loss function converges to obtain an optimal bathymetric correction model;

[0014] Use the optimal bathymetric correction model to perform water depth correction on satellite altimetry bathymetric data to obtain corrected satellite altimetry water depth data.

[0015] The technical effects achieved by the above technical solutions are as follows: Combining the advantages of satellite altimetry and multibeam sonar bathymetry, applying the powerful deep learning method in the field of digital image automation, performing water depth correction on satellite altimetry bathymetric data based on multibeam sonar, providing a new possibility for the accurate modeling of bathymetry.

[0016] Optionally, the preprocessing of the raw data specifically includes the following steps:

[0017] Perform interpolation preprocessing on the shipborne multibeam sonar bathymetric data and output gridded digital elevation model data;

[0018] Obtain satellite altimetry bathymetric data with the same range as the shipborne multibeam sonar bathymetric data, and perform grid resampling on the satellite altimetry bathymetric data according to the resolution of the corresponding shipborne multibeam sonar bathymetric data to unify the resolution of each pair.

[0019] The technical effects achieved by the above technical solutions are as follows: Satellite altimeters are widely used in water depth measurement due to their advantages of large-scale and real-time observation, but there is room for improvement in accuracy; multibeam echo sounding data has high accuracy, but is usually limited to a small coverage area; and this technology can combine the advantages of satellite altimetry and multibeam sonar bathymetry to achieve the unification of the spatial resolution of multibeam data and the temporal resolution of satellite altimetry data, providing a new technical means possibility for high-precision and high-real-time bathymetric measurement of the global sea area.

[0020] Optionally, the construction of the bathymetric correction model based on a convolutional neural network is specifically:

[0021] The VGG-19 model is selected as the basic architecture of the convolutional neural network. The VGG-19 model includes 16 convolutional layers and 3 fully connected layers;

[0022] The same-sized convolutional kernels and max-pooling kernels are selected for the entire network architecture, and ReLU is selected as the activation function to obtain the seabed terrain correction model. Among them, the number of channels is expanded through the convolutional kernels, and the width and height are reduced using the max-pooling kernels.

[0023] The technical effects achieved by the above technical solution are as follows: The main architecture of the seabed terrain correction model is disclosed, which can extract more abstract and deeper-level image features, reduce the number of parameters, while still retaining the same receptive field, and can improve the efficiency and accuracy of image calculation.

[0024] Optionally, obtaining the optimal seabed terrain correction model specifically includes the following steps:

[0025] Define a distance function to describe the difference degree between two input images. Input the shipborne multibeam sonar seabed terrain data and satellite altimetry seabed terrain data covering the same area into the seabed terrain correction model, then return the middle layer of the seabed terrain correction model, and finally output the result of the improved satellite altimetry water depth data. Among them, the used distance function L is expressed as:

[0026]

[0027] Among them, x represents the multibeam sonar water depth image, p represents the satellite altimetry water depth image, and i, j represent the serial numbers of the pixel points of the input image. Let V nn be the pre-trained VGG-19 network, X be any image, then V nn (X) is the feeding network of X; define respectively describe the intermediate feature representations of the network, where the inputs x and p are at the l-th layer of the network;

[0028] Estimate the gap between the model output value and the true value based on the distance function until the distance function converges to obtain the optimal seabed terrain correction model.

[0029] The technical effects achieved by the above technical solution are as follows: Using multibeam data as the content image for matching, inputting and converting satellite altimetry data under the VGG-19 framework to minimize the loss and distance between them, so as to obtain an improved seabed terrain correction model, while integrating the advantages of the high spatial resolution of multibeam data and the high spatial coverage of satellite altimetry data.

[0030] Optionally, the method further includes: Selecting the root mean square error RMSE, the normalized root mean square error NRMSE, and the coefficient of determination R 2As an evaluation metric, it quantifies the differences and relationships between the predicted values and the true values output by the model; where:

[0031]

[0032]

[0033]

[0034] where n represents the number of values in the dataset, i represents the sequence number of the value in the dataset, f represents the predicted value, and y represents the true value.

[0035] The technical effects achieved by the above technical solution are: the normalization of RMSE can make datasets with different numerical ranges easier to compare. The smaller the RMSE and NRMSE, the 2 larger the R, which means the higher the correlation between datasets. Based on this, the model accuracy and correction effect of the constructed model can be verified to promote the refined modeling of the seabed topography.

[0036] The present invention also discloses a satellite altimetry seabed topography correction system based on VGGNet, including: an acquisition module, a preprocessing module, a construction module, a training module, and a correction module;

[0037] The acquisition module is used to acquire shipborne multibeam sonar seabed topography data and satellite altimetry seabed topography data as raw data;

[0038] The preprocessing module is used to preprocess the raw data to obtain a training dataset;

[0039] The construction module is used to construct a seabed topography correction model through a convolutional neural network;

[0040] The training module is used to input the training dataset into the seabed topography correction model for model training until the loss function converges to obtain an optimal seabed topography correction model;

[0041] The correction module is used to perform water depth correction on the satellite altimetry bathymetric data through the optimal seabed topography correction model to obtain corrected satellite altimetry water depth data.

[0042] Optionally, the preprocessing module includes: an interpolation processing sub-module and a resampling sub-module;

[0043] The interpolation processing sub-module is used to perform interpolation preprocessing on the shipborne multibeam sonar seabed topography data and output grid digital elevation model data;

[0044] The resampling sub-module is used to obtain satellite altimetry bathymetric data with the same range as the bathymetric data of the shipborne multibeam sonar, and perform grid resampling on the satellite altimetry bathymetric data according to the resolution of the corresponding shipborne multibeam sonar bathymetric data to unify the resolutions of each pair.

[0045] Optionally, the infrastructure of the bathymetric correction model is the VGG-19 model, and the VGG-19 model includes 16 convolutional layers and 3 fully connected layers;

[0046] The entire network architecture of the bathymetric correction model includes convolutional kernels and max-pooling kernels of the same size as well as ReLU activation functions.

[0047] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a satellite altimetry bathymetric correction method and system based on VGGNet, which has the following beneficial effects:

[0048] (1) The present invention uses a deep learning method to correct satellite altimetry water depth data based on multibeam sonar water depth data, and can more directly and deeply extract key information in the multibeam water depth image through the hierarchical design of the calculation model, correct the satellite altimetry data, and provide new technical means possibilities for high-precision and high-real-time bathymetric surveys of the global sea area;

[0049] (2) Satellite altimeters are widely used in water depth measurement due to their advantages of large-scale and real-time observation, but there is room for improvement in accuracy; multibeam echo sounding data has high accuracy, but is usually limited to a small coverage area; and the present technical solution can combine the advantages of satellite altimetry and multibeam sonar sounding to achieve the unification of the spatial resolution of multibeam data and the temporal resolution of satellite altimetry data;

[0050] (3) The bathymetric correction model constructed by the present invention can extract more abstract and deeper image features, reduce the number of parameters, and still retain the same receptive field, which can improve the efficiency and accuracy of image calculation; using multibeam data as the content image for matching, inputting and transforming satellite altimetry data under the VGG-19 framework to minimize the loss and distance between them, an improved bathymetric correction model can be obtained, while combining the advantages of the high spatial resolution of multibeam data and the high spatial coverage of satellite altimetry data;

[0051] (4) In addition, in the present technical solution, the normalization of RMSE can make datasets with different numerical ranges easier to compare. The smaller the RMSE and NRMSE, and the larger the R 2 the larger, which means the higher the correlation between datasets. Based on this, the model accuracy and correction effect of the constructed model can be verified to promote the refined modeling of the bathymetry. Brief Description of the Drawings

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the provided drawings.

[0053] Figure 1 It is a flowchart of a satellite altimetry sea floor topography correction method based on VGGNet;

[0054] Figure 2 It is a schematic diagram of a deep learning network framework;

[0055] Figure 3 It is a schematic diagram of the structure of the VGG-19 model;

[0056] Figure 4 It is a schematic diagram of the losses of the training set and the validation set in the models of the Western Pacific Ocean, the Southern Ocean, and the Eastern Pacific Ocean;

[0057] Figure 5 It is a schematic diagram of the relationship between water depth and accuracy in the Western Pacific Ocean, the Southern Ocean, and the Eastern Pacific Ocean;

[0058] Figure 6 It is a schematic diagram of the distribution of the differences between the correction values and the true values in the Western Pacific Ocean, the Southern Ocean, and the Eastern Pacific Ocean;

[0059] Figure 7 It is a structural diagram of a satellite altimetry sea floor topography correction system based on VGGNet. Detailed Embodiments

[0060] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0061] Embodiment 1

[0062] Currently, only about 20% of the global sea floor topography has been refinedly modeled, and the rest either lacks data or is not accurate enough to meet actual needs. For this reason, the embodiments of the present invention disclose a satellite altimetry sea floor topography correction method based on VGGNet based on the existing shipborne multibeam sonar sea floor topography data and satellite altimetry sea floor topography data, as Figure 1 shown, including the following steps:

[0063] Collect the multi-beam sonar bathymetric data of the survey ship and the satellite altimetry bathymetric data as the original data;

[0064] Preprocess the original data to obtain a training dataset;

[0065] Build a bathymetric correction model based on a convolutional neural network, and input the training dataset into the bathymetric correction model for model training until the loss function converges to obtain the optimal bathymetric correction model;

[0066] Use the optimal bathymetric correction model to correct the water depth of the satellite altimetry bathymetric data to obtain the corrected satellite altimetry water depth data.

[0067] Next, the technical solution will be elaborated in detail to obtain a clearer and deeper understanding.

[0068] I. Dataset construction

[0069] First, obtain the original shipborne multi-beam sonar bathymetric data from the National Geophysical Data Center, perform interpolation preprocessing on it, and output gridded digital elevation model (DBM) data; at the same time, obtain and extract satellite altimetry bathymetric data from the NGDC's ETOPO11' global terrain model, with the same range as the above-mentioned shipborne multi-beam sonar bathymetric data, and perform grid resampling on the satellite altimetry bathymetric data according to the resolution of the corresponding shipborne multi-beam sonar bathymetric data to unify the resolution of each pair for subsequent operations.

[0070] In this embodiment, three groups of multi-beam sonar - satellite sounding data from the western Pacific Ocean, the southern ocean, and the eastern Pacific Ocean are used for experimental analysis, and the parameters of the data are shown in Table 1.

[0071] Table 1 Water depth data parameters

[0072]

[0073] II. Bathymetric correction model construction

[0074] As Figure 2 shown, the deep learning network structure proposed in this embodiment mainly consists of three parts: 1) Input of the true value and the water depth data to be corrected; 2) Specification of the network model, which requires a pre-trained VGG-19 framework, a loss function, gradient descent, and an optimization loop; 3) Output of the corrected satellite altimetry-inverted sounding data.

[0075] Convolutional neural network (CNN) models have been improved and updated in the past decade for better application in large-scale image recognition. Well-known applications include AlexNet, CaffeNet, LeNet, and VGGNet, etc. In this embodiment, the VGG-19 model is selected as the basic architecture of the convolutional neural network. As a form of VGGNet, compared with most models with 4 - 7 layers derived from CNN, it consists of 19 layers, including 16 convolutional layers and 3 fully connected layers. It can extract more abstract and deeper image features, reduce the number of parameters, and still retain the same receptive field, thus improving the efficiency and accuracy of image computing.

[0076] As Figure 3 shown in the schematic diagram of the VGG-19 architecture, the entire network architecture uses the same-sized convolutional kernel (3×3) and max-pooling kernel (2×2). The combination of several small filter (3×3) convolutional layers is superior to a large filter (5×5 or 7×7) in previous models; the convolutional kernel focuses on expanding the number of channels, and the max-pooling kernel focuses on reducing the width and height. Therefore, the architecture becomes wider, and the increase rate of the computational amount slows down, showing a larger receptive field of the network. At the same time, reducing the network parameters and using ReLU as the activation function multiple times can create more linear transformations and enhance the learning ability.

[0077] III. Model Training

[0078] The principle of the calibration model is to define a distance function to describe the difference degree between two input images. Input the shipborne multibeam sonar bathymetric data and satellite altimetry bathymetric data covering the same area into the bathymetric calibration model, then return the middle layer of the bathymetric calibration model, and finally output the result of the improved satellite altimetry water depth data; among them, the distance function L used is expressed as:

[0079]

[0080] Among them, x represents the multibeam sonar water depth image, p represents the satellite altimetry water depth image, and i, j represent the serial numbers of the pixel points of the input image; let V nn be the pre-trained VGG-19 network, X be any image, then V nn (X) is the feed network of X; define and respectively describe the intermediate feature representations of the network, where the inputs x and p are at the l-th layer of the network; finally, the optimizer update rule is applied to iteratively update the image, so as to minimize the given loss of the input. Estimate the gap between the model output value and the true value based on the distance function until the distance function converges to obtain the optimal bathymetric calibration model.

[0081] The evaluation of calibration accuracy is based on a comparison with previous studies. To quantify the differences and relationships between the predicted values and the true values of the model output, the root mean square error (RMSE), the normalized root mean square error (NRMSE), and the coefficient of determination (R 2 are selected as evaluation metrics; where:

[0082]

[0083]

[0084]

[0085] where n represents the number of values in the dataset, i represents the serial number of the value in the dataset, f represents the predicted value, and y represents the true value. Normalization of RMSE makes it easier to compare datasets with different numerical ranges. The ranges of NRMSE and R 2 are usually from 0 to 1. The smaller the RMSE and NRMSE, and the larger the R 2 , the higher the correlation between the datasets means.

[0086] Using multi-beam data as the content image for matching, input and transform satellite altimetry data under the VGG-19 framework to minimize the loss and distance between them, so as to obtain an improved satellite altimetry seabed terrain model, which can combine the advantages of the high spatial resolution of the former and the high spatial coverage of the latter.

[0087] IV. Result Analysis

[0088] For the VGG-19 model, the input parameter is a pair of multi-beam satellite bathymetry data, and the output parameter is the corrected satellite altimetry data. In the dataset of Table 1, randomly select 50% of them as the training set for initial fitting of the model and updating of parameters, and the remaining 50% is created as the validation set to conduct an unbiased evaluation of the model fitted on the training set, and finally obtain the prediction result.

[0089] The loss function is used to estimate the gap between the model output value and the true value to guide the subsequent optimization steps of the model. The smaller the loss function value, the better the model. The losses of the training set and the test set are as Figure 4 shown. In the three experimental areas, the loss of the model drops sharply to about 0.2 after 20 epochs and starts to gradually decline, especially after 70 epochs. And no obvious overfitting phenomenon is found during the calculation process. Therefore, it can be concluded that machine learning of the VGG-19 model can effectively reduce the loss of the experimental data in the three sea areas.

[0090] The performance parameters of the model were evaluated by testing the multibeam sonar data from the validation set, and the results are shown in Table 2. The R values of the corrected datasets from the Western Pacific, the Southern Ocean, and the Eastern Pacific with respect to the true value dataset are 0.80, 0.81, and 0.77 respectively, indicating excellent goodness of fit. In terms of RMSE and NRMSE, as can be seen from the figure, the correction algorithm produced errors of 267 meters, 102 meters, and 87 meters in the datasets of the Western Pacific, the Southern Ocean, and the Eastern Pacific respectively, and their NRMSEs are 0.031, 0.026, and 0.033 respectively. Compared with previous similar studies, the algorithm of this embodiment can improve the NRMSE of the dataset by more than 19%, demonstrating its potential. In addition, the changing trends of R and RMSE are consistent, both showing that the correction effect of the Southern Ocean data is the best, followed by the Western Pacific, and the Eastern Pacific is the third. 2 Values, which are 0.80, 0.81, and 0.77 respectively, indicating excellent goodness of fit. In terms of RMSE and NRMSE, as can be seen from the figure, the correction algorithm produced errors of 267 meters, 102 meters, and 87 meters in the datasets of the Western Pacific, the Southern Ocean, and the Eastern Pacific respectively, and their NRMSEs are 0.031, 0.026, and 0.033 respectively. Compared with previous similar studies, the algorithm of this embodiment can improve the NRMSE of the dataset by more than 19%, demonstrating its potential. In addition, R 2 and RMSE have the same changing trend, both showing that the correction effect of the Southern Ocean data is the best, followed by the Western Pacific, and the Eastern Pacific is the third.

[0091] Table 2 Satellite altimetry correction accuracy

[0092]

[0093] In the experiment, it can be found that taking R 2 as an example, the correction accuracy varies with water depth, as Figure 5 shown. As can be seen from the figure, the maximum and minimum values in each sea area are basically the same. Generally, the minimum value of R 2 is above 0.2, which appears at the extreme values of water depth, while the maximum value can reach above 0.9. In the Western Pacific data, R 2 is above 0.8 in the range of about -4500 meters to -1900 meters of water depth, showing a strong correlation, and the maximum value appears at about -3200 meters; for the Southern Ocean data, R 2 has a strong correlation at about -500 meters and in the range of about -1800 meters to -2400 meters, and the maximum value appears at about -2200 meters; for the Eastern Pacific data, R 2 has a strong correlation in the range of about -2400 meters to -3600 meters, and the maximum value appears at about -3500 meters.

[0094] According to experience, the accuracy of machine learning is positively correlated with the sample size in the dataset. Without considering other parameters, the larger the sample size, the higher the learning accuracy tends to be, and vice versa. In this experiment, this experience was also verified. Combining with the histogram of water depth values, the depths where the water depth data points are distributed dispersedly and have a large variance are often the areas where R 2 shows a lower value, while R 2Higher depths tend to have a more concentrated distribution with a smaller variance. Specifically, at the maximum and minimum depths in these three sea areas, due to the small amount of sample data, the accuracy of machine learning is also relatively low. The accuracy of the distribution within the depth range with the largest sample size shows a highly correlated relationship. The R 2 curve undulations at a certain depth also reflect to some extent the particularity of the bathymetric value distribution of the local seabed topography. Experiments show that as long as sufficient data is input, the satellite altimetry bathymetry data corrected by the VGG-19 model can achieve a good fitting effect with the multibeam-derived data within a specific depth range.

[0095] By subtracting the true value of the multibeam sonar data from the corrected bathymetric value of the satellite altimetry data, it can be found that the error distribution between the two shows a form of high in the middle and low on both sides, that is, the closer the error is to 0, the more data points there are, and vice versa, as Figure 6 shown. In the data of the Western Pacific, the data point with zero error as the maximum value is isolated and discontinuous with the other parts of the curve, indicating that the number of error-free bathymetric points after the correction of the algorithm in this embodiment has increased significantly. In the other two groups of data, the data curves are relatively continuous, gradually decreasing from the maximum value at the zero point to both sides with increasing absolute value, and the curve of the Southern Ocean data is more convergent near the zero point than that of the Eastern Pacific data, indicating that its correction effect is better.

[0096] To represent it more intuitively, the absolute values of the above results are now used to calculate the percentages of the data within the 2% and 1% ranges in the total depth of each data. Its value represents the error from the true value, as shown in Table 3. As the error range decreases, the number of data points also gradually increases. On average, the data points with an error within 2% of the depth value account for 70.58% of the total, and those within 1% account for 49.21%. Compared with previous studies, the correction accuracy of the deep learning VGG-19 model can be effectively improved by more than 17%.

[0097] In the depth range index, the accuracy of the corrected Southern Ocean data is always better than the other two indexes, and there is a large gap. Within the 2% range, the performance of the Eastern Pacific and Western Pacific data shows almost no difference, and the data of the Eastern Pacific is slightly higher. Under the strictest 1% range standard, the performance results of the two are widened, and the data of the Western Pacific is better. Combining with Table 2, it can be found that the parameter changes in the two tables show relative consistency. The correction effect of the Southern Ocean data is the best, while the data of the Western Pacific and Eastern Pacific are the second and the lowest respectively in most cases.

[0098] Table 3 Proportion of true value correction errors within the 2% and 1% depth ranges

[0099]

[0100] In this technical solution, a pre-trained VGGNet model algorithm based on deep learning is proposed. Using the water depth data inverted by multi-beam sonar as the true value, the water depth data inverted by satellite altimetry is corrected. The core idea of the correction model is to define and minimize the distance (loss) function between the real data and the data to be corrected, and finally output the corresponding corrected satellite altimetry seabed topography. Then, three groups of bathymetric data in the Western Pacific Ocean, the Southern Ocean, and the Eastern Pacific Ocean are used to evaluate the performance of the model. During the test, the data losses of the training set and the validation set are effectively reduced, which proves the effectiveness of the model.

[0101] In addition, three indicators, namely R 2 , RMSE, and its derived NRMSE, are selected to evaluate the correction results of the data, and the results are excellent. The NRMSE indicator is more than 19% higher than that of previous studies. Further, by analyzing the difference in R 2 values at different water depths, it can be found that the correction accuracy of deep learning is positively correlated with the sample size, that is, the more data points, the higher the accuracy of the depth value, and vice versa. Finally, after finding that the difference between the true value and the corrected value gradually decreases from the maximum value at zero to both sides of the number axis, the proportion of the absolute value of the difference in the overall water depth is analyzed, and it is found that the model of this technology can improve the correction accuracy by more than 17% compared with previous studies. Generally speaking, among the three test areas, the correction accuracy of the Southern Ocean data is the highest, followed by the Western Pacific Ocean data, and the Eastern Pacific Ocean data is the last.

[0102] Embodiment 2

[0103] This embodiment discloses a satellite altimetry seabed topography correction system based on VGGNet, as Figure 7 shown, including: a collection module, a preprocessing module, a construction module, a training module, and a correction module;

[0104] The collection module is used to collect ship-measured multi-beam sonar seabed topography data and satellite altimetry seabed topography data as the original data;

[0105] The preprocessing module is used to preprocess the original data to obtain a training data set;

[0106] The construction module is used to construct a seabed topography correction model through a convolutional neural network;

[0107] The training module is used to input the training data set into the seabed topography correction model for model training until the loss function converges to obtain an optimal seabed topography correction model;

[0108] The correction module is used to perform water depth correction on the satellite altimetry bathymetric data through the optimal seabed topography correction model to obtain the corrected satellite altimetry water depth data.

[0109] Further, the preprocessing module includes: an interpolation processing sub-module and a resampling sub-module;

[0110] The interpolation processing sub-module is used to perform interpolation preprocessing on the shipborne multibeam sonar bathymetric data and output gridded digital elevation model data;

[0111] The resampling sub-module is used to obtain satellite altimetry bathymetric data with the same range as the shipborne multibeam sonar bathymetric data, and perform grid resampling on the satellite altimetry bathymetric data according to the resolution of the corresponding shipborne multibeam sonar bathymetric data to unify the resolution of each pair.

[0112] Further, the basic architecture of the bathymetric correction model is a VGG-19 model, and the VGG-19 model includes 16 convolutional layers and 3 fully connected layers;

[0113] The entire network architecture of the bathymetric correction model includes convolutional kernels and max-pooling kernels of the same size as well as ReLU activation functions.

[0114] Based on the existing shipborne multibeam sonar bathymetric data and satellite altimetry bathymetric data, the present invention uses a deep learning method to correct satellite altimetry water depth data based on multibeam sonar water depth data, and can more directly and deeply extract key information in the multibeam water depth image through the hierarchical design of the calculation model, correct the satellite altimetry data, and provide new technical means possibilities for high-precision and high-real-time bathymetric surveys of the global sea area.

[0115] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0116] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for satellite altimetry sea floor topography correction based on VGGNet, characterized in that, Including the following steps: Collecting shipborne multibeam sonar bathymetric data and satellite altimetry bathymetric data as original data; Preprocessing the original data to obtain a training dataset; Constructing a bathymetric correction model based on a convolutional neural network and inputting the training dataset into the bathymetric correction model for model training until the loss function converges to obtain an optimal bathymetric correction model; Using the optimal bathymetric correction model to perform water depth correction on satellite altimetry sounding bathymetric data to obtain corrected satellite altimetry water depth data; The steps for obtaining the optimal bathymetric correction model specifically include the following: Defining a distance function to describe the difference between two input images, inputting shipborne multibeam sonar bathymetric data and satellite altimetry bathymetric data covering the same area into the bathymetric correction model, then returning the middle layer of the bathymetric correction model, and finally outputting the result of the improved satellite altimetry water depth data; where the distance function L used is expressed as: Among them, x represents the multi-beam sonar bathymetric image, p represents the satellite altimetry bathymetric image, and i, j represent the serial numbers of the pixel points of the input image; let V nn be the pre-trained VGG-19 network, X be any image, then V nn (X) is the feeding network of X; define and respectively describe the intermediate feature representations of the network, where the inputs x and p are at the l-th layer of the network; Estimating the gap between the model output value and the true value based on the distance function until the distance function converges to obtain an optimal bathymetric correction model.

2. The satellite altimetry sea floor topography correction method based on VGGNet according to claim 1, characterized in that The preprocessing of the original data specifically includes the following steps: Performing interpolation preprocessing on the shipborne multibeam sonar bathymetric data and outputting gridded digital elevation model data; Obtaining satellite altimetry bathymetric data with the same range as the shipborne multibeam sonar bathymetric data and performing grid resampling on the satellite altimetry bathymetric data according to the resolution of the corresponding shipborne multibeam sonar bathymetric data to unify the resolution of each pair.

3. A satellite altimetry sea floor topography correction method based on VGGNet according to claim 1, characterized in that The constructing of the bathymetric correction model based on a convolutional neural network is specifically: Selecting the VGG-19 model as the basic architecture of the convolutional neural network, and the VGG-19 model includes 16 convolutional layers and 3 fully connected layers; Selecting the same-sized convolutional kernel and max pooling kernel for the entire network architecture and using ReLU as the activation function to obtain the bathymetric correction model; where the number of channels is expanded through the convolutional kernel and the width and height are reduced using the max pooling kernel.

4. A satellite altimetry sea floor topography correction method based on VGGNet according to claim 1, characterized in that The method further includes: selecting root mean square error (RMSE), normalized root mean square error (NRMSE), and coefficient of determination (R) 2 as evaluation metrics to quantify the difference and relationship between the predicted value and the true value output by the model; where: Where n represents the number of values in the dataset, i represents the serial number of the value in the dataset, f represents the predicted value, and y represents the true value.

5. A system for implementing the VGGNet-based satellite altimetry seabed topography correction method according to any one of claims 1-4, characterized in that, Including: A collection module, a preprocessing module, a construction module, a training module, and a correction module; The collection module is used to collect shipborne multibeam sonar bathymetric data and satellite altimetry bathymetric data as original data; The preprocessing module is used to preprocess the original data to obtain a training dataset; The construction module is used to construct a bathymetric correction model through a convolutional neural network; The training module is used to input the training dataset into the bathymetric correction model for model training until the loss function converges to obtain an optimal bathymetric correction model; The correction module is used to perform water depth correction on satellite altimetry sounding bathymetric data through the optimal bathymetric correction model to obtain corrected satellite altimetry water depth data.

6. The satellite altimetry sea floor topography correction system based on VGGNet according to claim 5, wherein, The preprocessing module includes: an interpolation processing sub-module, a resampling sub-module; The interpolation processing sub-module is used to perform interpolation preprocessing on the shipborne multi-beam sonar bathymetric data and output grid digital elevation model data; The resampling sub-module is used to obtain satellite altimetry bathymetric data with the same range as the shipborne multi-beam sonar bathymetric data, and perform grid resampling on the satellite altimetry bathymetric data according to the resolution of the corresponding shipborne multi-beam sonar bathymetric data to unify the resolutions of each pair.

7. The satellite altimetry sea floor topography correction system based on VGGNet according to claim 5, characterized in that, The basic architecture of the bathymetric correction model is the VGG-19 model, and the VGG-19 model includes 16 convolutional layers and 3 fully connected layers; The entire network architecture of the bathymetric correction model includes convolutional kernels and max-pooling kernels of the same size as well as ReLU activation functions.

Images