A method for underwater terrain extraction based on neural network and ensemble learning

Abstract

The present invention discloses an underwater terrain extraction method based on neural network and ensemble learning, which includes the following steps: preprocessing remote sensing image data sets by using digital image processing technology combined with visual interpretation assistance, and using the preprocessed remote sensing images and training samples as input data sets; constructing a sub-learner based on BP neural network algorithm, training the BP neural network algorithm, and using the trained neural network model to generate several water depth inversion results; determining the integration strategy based on the minimum outlier method; integrating the water depth inversion result set of the entire study area, and evaluating the accuracy of the integrated underwater terrain map. High-precision shallow sea underwater terrain is achieved; through the combination of neural network and ensemble learning algorithm, the problem of poor robustness of the traditional BP neural network algorithm in the water depth inversion process is solved, and the water depth inversion accuracy and reliability are further improved.

Description

Technical Field

[0001] The invention relates to the field of shallow sea underwater terrain inversion, and in particular to an underwater terrain extraction method based on neural network and ensemble learning. Background Art

[0002] Shallow water depth data is of great significance to shipping management, island and reef development, and ecological protection. At present, the main depth measurement methods are ship-borne sonar and airborne Lidar (laser radar). These on-site depth measurement methods are difficult to use on islands and reefs that are in danger of running aground or are difficult to reach. At the same time, they consume a lot of manpower and material resources, which is not conducive to conducting large-scale depth measurement. In recent years, the depth measurement method based on remote sensing images has the advantages of wide coverage, low measurement cost, and rich data sources, and has received widespread attention.

[0003] Bathymetry research based on remote sensing images originated in the 1970s. With the development of satellite technology and bathymetry theory, a number of classic bathymetry methods based on remote sensing images have been born, such as theoretical analytical method (Lee et al., 1998), Stumpf ratio (Stumpf et al., 2003; Ma et al., 2020), Lyzenga polynomial (Lyzenga, 1978; Manessa et al., 2018), BP neural network algorithm (Sandidge et al., 1998; Liu et al., 2018), etc. Among them, the BP neural network algorithm has shown the best inversion performance in a large number of previous studies (Chu et al., 2019; Ceyhun and Yalcin, 2010; Gholamalifard et al., 2013; Liu et al., 2015), so the BP neural network algorithm is widely used. This method has been applied by many researchers in different types of environments and using a variety of different satellite data, including Landsat-8 (El-Mewafi et al., 2018), Sentinel-2 (Chu et al., 2019), Quickbird (Ceyhun and Yalcin, 2010), Spot-6 (Hussein and Nadaoka, 2017), WorldView-3 (Collin et al., 2017), etc.

[0004] The BP neural network algorithm is a multi-layer feedforward network trained by the error back propagation algorithm and is currently the most widely used neural network model (Li et al., 2012). The BP neural network algorithm was proposed by a scientific team headed by Rumelhart and McCelland in 1986 (Rumelhart and McCelland, 1986). Subsequently, the Naval Research Laboratory of the Stennis Space Center in the United States used the BP neural network algorithm for underwater terrain inversion (Sandidge and Holyer, 1998). Its basic principle is: using spectral reflectance as the input layer and water depth data as the output layer, then using a small amount of measured water depth values ​​to correct the hidden layer weights, and finally using the trained model to invert the water depth of pixels at unknown water depths. The biggest advantage of the BP neural network algorithm is its excellent nonlinear fitting ability (Qiu et al., 2018). The water depth inversion based on remote sensing images is affected by many factors such as phytoplankton, colored soluble organic matter CDOM, suspended particles, bottom sediment, sunlight conditions, etc. Therefore, the water depth inversion process is nonlinear (Ceyhun and Yalcin, 2010), which is also the main reason why the BP algorithm outperforms other classic algorithms in a large number of literatures. Another advantage of the BP neural network algorithm is that it does not need to understand the physical model of water depth inversion, and can directly learn the characteristic laws from the training samples. It is convenient to use and is therefore widely used.

[0005] However, the BP neural network algorithm has the problem of being easily trapped in local minima (Hirose et al., 1991; Deng et al., 2021). The BP neural network algorithm uses the gradient descent method to find the optimal solution. Since the error surface of the BP neural network is uneven, there are some points on the error surface with a gradient of 0. These points belong to local minima, but not necessarily global minima. Therefore, after the BP neural network algorithm falls into these local minimum points, it may mistakenly think that the error is no longer decreasing and stop training, which eventually leads to learning failure (Lee et al., 1993). Selecting appropriate parameters (such as learning rate, hidden layer nodes) and initial value weights can weaken the influence of local minima. The parameter selection can be based on experience or traversal trial method to select appropriate values ​​(Shahjahan and Murase, 2003; Benardos and Vosniakos, 2007). However, the initial value weights are difficult to set manually due to the large number of them, and are usually assigned randomly. Therefore, the BP neural network algorithm is random and difficult to control when it falls into local minima.

[0006] The underwater terrain inversion based on BP neural network algorithm also has the above problems. The local minimum problem will lead to an erroneous underwater terrain inversion result, which is worthy of vigilance. Therefore, a method for underwater terrain extraction based on neural network and ensemble learning is proposed to effectively solve the problem of poor robustness of traditional BP neural network depth inversion and is applied to shallow water underwater terrain detection, which is of great significance.

[0007] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0008] In view of the problems in the related art, the present invention proposes an underwater terrain extraction method based on neural network and ensemble learning to overcome the above-mentioned technical problems existing in the existing related art.

[0009] To this end, the specific technical solution adopted by the present invention is as follows:

[0010] A method for underwater terrain extraction based on neural network and ensemble learning, the method comprising the following steps:

[0011] S1. Preprocessing: Using digital image processing technology and visual interpretation assistance, the remote sensing image data set is preprocessed by atmospheric correction, land-water separation and mean filtering, and the preprocessed remote sensing images and training samples are used as input data sets;

[0012] S2. Constructing a sub-learner: constructing a sub-learner based on the BP neural network algorithm, training the BP neural network algorithm, and using the trained neural network model to generate several water depth inversion results;

[0013] S3, integration strategy: Determine the integration strategy based on the minimum outlier method;

[0014] S4. Integration of water depth inversion results: Integrate the water depth inversion result set of the entire study area and conduct accuracy assessment on the integrated underwater topographic map.

[0015] Furthermore, when performing preprocessing of atmospheric correction on the remote sensing image dataset in S1, the SNAP plug-in of Sentinel image is used to perform atmospheric correction on the remote sensing image.

[0016] Furthermore, the preprocessing of the remote sensing image data set for water-land separation in S1 further includes the following steps:

[0017] The land area is manually determined by combining empirical methods with visual interpretation, and the land area is used as a mask file;

[0018] The remote sensing images of the study area were masked and the land areas in the remote sensing images were removed.

[0019] Furthermore, when performing mean filtering preprocessing on the remote sensing image data set in S1, a 3×3 window is used to perform mean filtering processing on the remote sensing image with the land area removed.

[0020] Furthermore, the formula of the BP neural network in S2 is:

[0021]

[0022] In the formula, I, H, O are the input layer, hidden layer, and output layer vectors, respectively; w and v are the weight vectors of the connection from the input layer to the hidden layer and from the hidden layer to the output layer, respectively; T H and T o are the activation thresholds of neurons in the hidden layer and output layer, respectively.

[0023] Furthermore, the step of training the BP neural network algorithm in S2 and generating a plurality of water depth inversion results using the trained neural network model also includes the following steps:

[0024] The remote sensing images after atmospheric correction, land-water separation and mean filtering are matched with the measured water depth points, and the measured water depth values ​​and the reflectance values ​​of N bands of pixels at the corresponding positions of the measured water depth points constitute training samples, and the BP neural network algorithm is trained through the training samples;

[0025] The reflectivity values ​​of N bands of each pixel in the remote sensing image are input into the trained neural network model, and the water depth value corresponding to each pixel is predicted to achieve the inversion of shallow sea underwater topography.

[0026] Furthermore, the inversion of the shallow sea underwater topography also includes the following steps:

[0027] 300 points are randomly selected from the sonar measured water depth points as the training point set, and the remaining points are used as the test point set;

[0028] ARCGIS software was used to spatially match the training point set and the test point set with the best synthetic image, and the reflectance values ​​of the three bands R, G, and B of each measured water depth point and the corresponding pixel were extracted to form training samples and test samples respectively.

[0029] Select training samples to train the BP neural network model multiple times to generate multiple neural network models;

[0030] Based on the reflectivity values ​​of the R, G, and B bands of the remote sensing image, multiple trained neural network models are input, and the water depth value corresponding to each pixel is predicted to obtain multiple sets of water depth prediction values;

[0031] The underwater topography map is reconstructed according to the outlier degree of all water depth prediction values ​​at each pixel, and the inversion of the shallow underwater topography in the experimental area is completed.

[0032] Furthermore, when the integration strategy is determined based on the minimum outlier method in S3, each sub-learner will generate a water depth inversion result, that is, at each geographical location, a water depth inversion result set S = {d1, d2, ..., d l , …, d L}, where L is the number of sub-learners, d l represents the water depth value of the lth inversion, and the mathematical formula of the minimum outlier degree is:

[0033]

[0034] In the formula, OD l Indicates the water depth d l The outlier degree, OD l The value range of OD is [0, 1]. l The larger the value, the greater the d l The greater the difference from other water depth values, the greater the d l The larger the outlier degree, the more likely it is noise.

[0035] Furthermore, the integration strategy in S3 further includes the following steps:

[0036] Calculate the outlier of each inversion result, and the inversion result S = {d1, d2, d3, ..., d L The outlier degree of} is {OD1, OD2, OD3, …, OD L};

[0037] Find the water depth value corresponding to the minimum outlier degree, and set the minimum outlier degree and its corresponding number p as OD p =min{OD1, OD2, OD3,…, OD L}, p∈{1, 2, 3, …, L}, and the water depth values ​​with the minimum outlier degree are combined into a new result set S'={d p};

[0038] The new result set contains N elements. If N>1, then for S'={d p}Calculate the average value and use the average value as the integrated water depth result;

[0039] If the new result set contains only one element, that is, N = 1, then the new result set d p As the integrated water depth result.

[0040] Furthermore, the S4 further includes the following steps when integrating the water depth inversion result set of the entire study area and evaluating the accuracy of the integrated underwater topographic map:

[0041] Traverse the entire study area and integrate all water depth inversion results to obtain an integrated and high-precision underwater topographic map;

[0042] A water depth inversion experiment was conducted in the study area, and the water depth inversion effect was evaluated based on the RMSE error;

[0043] The smaller the RMSE value is, the higher the accuracy of water depth inversion is.

[0044]

[0045] In the formula, z i is the measured water depth value, is the inverted water depth value, n is the number of test samples, and i is a non-zero natural number.

[0046] The beneficial effects of the present invention are as follows: the present invention constructs a sub-learner based on the BP neural network algorithm, and innovatively proposes an integrated strategy of the minimum outlier method, providing an underwater terrain extraction method based on a neural network and integrated learning, realizing the extraction of high-precision shallow sea underwater terrain, and providing technical support for applications such as island and reef construction, navigation safety, and ecological protection. The present invention solves the problem of poor robustness of the traditional BP neural network algorithm in the process of water depth inversion by combining a neural network with an integrated learning algorithm, further improves the accuracy and reliability of water depth inversion, and overcomes the shortcomings of poor reliability and low accuracy of traditional water depth inversion methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0048] Figure 1 is a flow chart of an underwater terrain extraction method based on neural network and ensemble learning according to an embodiment of the present invention;

[0049] Figure 2 It is an overview map of the case study area;

[0050] Figure 3 is the error distribution diagram of 100 repeated experiments in the case study area;

[0051] Figure 4 This is the underwater terrain inversion result map corresponding to the maximum RMSE error. DETAILED DESCRIPTION

[0052] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention and are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, ordinary technicians in the field should be able to understand other possible implementations and advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0053] According to an embodiment of the present invention, a method for underwater terrain extraction based on neural network and ensemble learning is provided. Compared with traditional methods, the present invention uses BP neural network algorithm as a basic learner to generate multiple water depth inversion results, and proposes an integration strategy based on the minimum outlier method to integrate all water depth inversion results, and finally obtain an integrated high-precision underwater terrain map, which provides a scientific and effective inversion method for shallow sea underwater terrain detection.

[0054] The present invention is further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, according to the underwater terrain extraction method based on neural network and ensemble learning according to an embodiment of the present invention, the method comprises the following steps:

[0055] S1. Preprocessing (preprocessing of remote sensing image dataset): Using digital image processing technology and visual interpretation assistance, the remote sensing image dataset is preprocessed by atmospheric correction, water-land separation and mean filtering, and the preprocessed remote sensing images and training samples are used as input datasets;

[0056] In the above-mentioned S1, when the remote sensing image data set is preprocessed for atmospheric correction, in order to quantitatively extract water depth information, the SNAP plug-in of Sentinel image (Sentinel satellite image) is used to perform atmospheric correction on the remote sensing image, and the corrected remote sensing image is used for subsequent processing.

[0057] The preprocessing of the remote sensing image data set for water-land separation in S1 further includes the following steps:

[0058] The land area is manually determined by combining empirical methods with visual interpretation, and the land area is used as a mask file;

[0059] The remote sensing images of the study area were masked and the land areas in the remote sensing images were removed.

[0060] When performing mean filtering preprocessing on the remote sensing image data set in S1, a 3×3 window is used to perform mean filtering processing on the remote sensing image with the land area removed.

[0061] In this step, the remote sensing image is first preprocessed by using the SNAP plug-in to perform atmospheric correction and other image preprocessing. Then, the land area on the remote sensing image is manually delineated using ARCGIS software, and the land and water are separated by mask processing. Finally, the remote sensing image is mean filtered using a 3×3 window.

[0062] S2, constructing a sub-learner (constructing a sub-learner based on the BP neural network algorithm): using the pre-processed best remote sensing image and training samples as input data sets, constructing a sub-learner based on the BP neural network algorithm, training the BP neural network algorithm, and using the trained neural network model to generate several water depth inversion results;

[0063] The formula of the BP neural network in S2 is:

[0064]

[0065] Where I, H, and O are the input layer (spectral reflectance), hidden layer, and output layer (water depth value) vectors, respectively; w and v are the weight vectors of the connection from the input layer to the hidden layer and from the hidden layer to the output layer, respectively; T H and T o are the activation thresholds of neurons in the hidden layer and output layer, respectively.

[0066] The BP neural network algorithm is trained in S2, and the trained neural network model is used to generate a number of water depth inversion results, which also includes the following steps:

[0067] The remote sensing images after atmospheric correction, land-water separation and mean filtering are matched with the measured water depth points, and the measured water depth values ​​and the reflectance values ​​of N bands of pixels at the corresponding positions of the measured water depth points constitute training samples, and the BP neural network algorithm is trained through the training samples;

[0068] The reflectivity values ​​of N bands of each pixel in the remote sensing image are input into the trained neural network model, and the water depth value corresponding to each pixel is predicted to achieve the inversion of shallow sea underwater topography.

[0069] The inversion of shallow sea underwater topography also includes the following steps:

[0070] 300 points are randomly selected from the sonar measured water depth points as the training point set, and the remaining points are used as the test point set;

[0071] ARCGIS software was used to spatially match the training point set and the test point set with the best synthetic image, and the reflectance values ​​of the three bands R, G, and B of each measured water depth point and the corresponding pixel were extracted to form training samples and test samples respectively.

[0072] The BP neural network model was trained multiple times with 300 training samples to generate multiple neural network models. The model parameters of the BP neural network were set as follows: the maximum number of training times was 1500, the learning rate was 0.05, the momentum factor was 0.9, and the training target error was 1×10 -5 ;

[0073] Based on the reflectivity values ​​of the R, G, and B bands of the remote sensing image, multiple trained neural network models are input, and the water depth value corresponding to each pixel is predicted to obtain multiple sets of water depth prediction values;

[0074] The underwater topography map is reconstructed according to the outlier degree of all water depth prediction values ​​at each pixel, and the inversion of the shallow underwater topography in the experimental area is completed.

[0075] S3, integration strategy (integration strategy based on minimum outlier method): Determine the integration strategy based on the minimum outlier method, integrate all water depth inversion results, and finally obtain the optimal underwater terrain result after integration

[0076] When the integration strategy is determined based on the minimum outlier method in S3, each sub-learner will generate a water depth inversion result, that is, at each geographical location, a water depth inversion result set S = {d1, d2, ..., d l , …, d L}, where L is the number of sub-learners, d l represents the water depth value of the lth inversion, and the mathematical formula of the minimum outlier degree is:

[0077]

[0078] In the formula, OD l Indicates the water depth d l The outlier degree, OD l The value range of OD is [0, 1]. l The larger the value, the greater the d l The greater the difference from other water depth values, the greater the d l The larger the outlier degree, the more likely it is noise.

[0079] The integration strategy in S3 also includes the following steps:

[0080] Calculate the outlier of each inversion result, and the inversion result S = {d1, d2, d3, ..., d L The outlier degree of} is {OD1, OD2, OD3, …, OD L};

[0081] Find the water depth value corresponding to the minimum outlier degree, and set the minimum outlier degree and its corresponding number p as OD p=min{OD1, OD2, OD3,…, OD L}, p∈{1, 2, 3, …, L}, and the water depth values ​​with the minimum outlier degree are combined into a new result set S'={d p}, p can be one value or multiple values;

[0082] The new result set contains N elements. If N>1, then for S'={d p}Calculate the average value and use the average value as the integrated water depth result;

[0083] If the new result set contains only one element, that is, N = 1, then the new result set d p As the integrated water depth result.

[0084] S4, integration of water depth inversion results (integration of all water depth inversion results): Integrate the water depth inversion result set of the entire study area and conduct accuracy assessment on the integrated underwater topographic map;

[0085] Wherein, the step S4 further includes the following steps when integrating the water depth inversion result set of the entire study area and evaluating the accuracy of the integrated underwater topographic map:

[0086] Traverse the entire study area and integrate all water depth inversion results to obtain an integrated and high-precision underwater topographic map;

[0087] The experiment was repeated 100 times in the study area, and the water depth inversion effect was displayed based on the result with the largest RMSE (root mean square error);

[0088] Among them, Figure 4 As shown, the smaller the RMSE value is, the higher the accuracy of water depth inversion is;

[0089]

[0090] In the formula, z i is the measured water depth value, is the inverted water depth value, n is the number of test samples, and i is a non-zero natural number.

[0091] like Figure 2 As shown in the figure, the Anda Reef in the Nansha Islands in the South China Sea is used as the experimental area. Anda Reef is located in the Zhenghe Reefs in the Nansha Islands, with the central longitude and latitude of (10°20′N, 114°42′E), and the reef area is about 30 square kilometers; the reef is shallow in the northeast, with the shallowest point close to the sea level, and deep in the southwest, with a water depth of more than 20 meters, and the average water depth of the reef is about 10 meters. The remote sensing image data in the example uses the Sentinel-2 satellite image of Anda Reef taken on April 9, 2017. In addition, there are 1,315 sonar-measured water depth points in the Anda Reef experimental area.

[0092] The parameters of the underwater terrain extraction algorithm based on neural network and ensemble learning proposed in this paper are set as follows: the number of BP neural network algorithms is 5, the number of input layer nodes is 3 (red, green, and blue), the number of hidden layer nodes is 7, the transfer functions of the hidden layer and the output layer are tansig and purelin, respectively, the training function is trainlm, the maximum number of training times is 1500, the learning rate is 0.05, the momentum factor is 0.9, and the training target error is 1×10 -5 . The experiment randomly selects 300 water depth samples as training samples, and the remaining water depth samples as test samples. Using the underwater terrain extraction algorithm based on neural network and ensemble learning proposed in this invention, the experiment is repeated 100 times on the case study area. It can be seen from the RMSE error distribution results of the repeated experiments that ( Figure 3 ), the error distribution range of the underwater terrain extraction results of this algorithm is concentrated and has strong robustness. Figure 4 This is the case with the largest RMSE error among all the inversion results. It can be seen that even the result with the largest RMSE error in 100 repeated experiments still has a very good water depth inversion effect, with obvious water depth characteristics and clear island and reef contours, highlighting the reliability of the underwater terrain extraction method based on neural network and ensemble learning proposed in the present invention.

[0093] In summary, the present invention constructs a sub-learner based on the BP neural network algorithm, and innovatively proposes an integrated strategy of the minimum outlier method, providing an underwater terrain extraction method based on neural network and integrated learning, realizing the extraction of high-precision shallow sea underwater terrain, and providing technical support for applications such as island and reef construction, navigation safety, and ecological protection. The present invention solves the problem of poor robustness of the traditional BP neural network algorithm in the process of water depth inversion by combining neural network with integrated learning algorithm, further improves the accuracy and reliability of water depth inversion, and overcomes the shortcomings of poor reliability and low accuracy of traditional water depth inversion methods.

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for underwater terrain extraction based on neural network and ensemble learning, characterized in that: The method comprises the following steps: S1. Preprocessing: Using digital image processing technology and visual interpretation assistance, the remote sensing image data set is preprocessed by atmospheric correction, land-water separation and mean filtering, and the preprocessed remote sensing images and training samples are used as input data sets; S2. Constructing a sub-learner: constructing a sub-learner based on the BP neural network algorithm, training the BP neural network algorithm, and using the trained neural network model to generate several water depth inversion results; S3, integration strategy: Determine the integration strategy based on the minimum outlier method; S4. Integration of water depth inversion results: Integrate the water depth inversion result set of the entire study area and conduct accuracy assessment on the integrated underwater topographic map; When the integration strategy is determined based on the minimum outlier method in S3, each sub-learner will generate a water depth inversion result, that is, at each geographical location, a water depth inversion result set S = {d1, d2, ..., d l , …, d L }, where L is the number of sub-learners, d l represents the water depth value of the lth inversion, and the mathematical formula of the minimum outlier degree is: In the formula, OD l Indicates the water depth d l The outlier degree, OD l The value range of OD is [0, 1]. l The larger the value, the greater the d l The greater the difference from other water depth values, the greater the d l The greater the outlier degree of is, the more likely it is noise; The integration strategy in S3 also includes the following steps: Calculate the outlier of each inversion result, and the inversion result S = {d1, d2, d3, ..., d L The outlier degree of} is {OD1, OD2, OD3, …, OD L }; Find the water depth value corresponding to the minimum outlier degree, and set the minimum outlier degree and its corresponding number p as OD p =min{OD1, OD2, OD3,…, OD L }, p∈{1, 2, 3, …, L}, and the water depth values ​​with the minimum outlier degree are combined into a new result set S'={d p }; The new result set contains N elements. If N>1, then for S'={d p }Calculate the average value and use the average value as the integrated water depth result; If the new result set contains only one element, that is, N = 1, then the new result set d p As the integrated water depth result.

2. The underwater terrain extraction method based on neural network and ensemble learning according to claim 1 is characterized in that: When performing preprocessing of atmospheric correction on the remote sensing image dataset in S1, the SNAP plug-in of Sentinel image is used to perform atmospheric correction on the remote sensing image.

3. The underwater terrain extraction method based on neural network and ensemble learning according to claim 1 is characterized in that: The preprocessing of the remote sensing image data set for water-land separation in S1 further includes the following steps: The land area is manually determined by combining empirical methods with visual interpretation, and the land area is used as a mask file; The remote sensing images of the study area were masked and the land areas in the remote sensing images were removed.

4. The underwater terrain extraction method based on neural network and ensemble learning according to claim 1 is characterized in that: When performing mean filtering preprocessing on the remote sensing image data set in S1, a 3×3 window is used to perform mean filtering processing on the remote sensing image with the land area removed.

5. The underwater terrain extraction method based on neural network and ensemble learning according to claim 1 is characterized in that: The formula of the BP neural network in S2 is: In the formula, I, H, O are the input layer, hidden layer, and output layer vectors, respectively; w and v are the weight vectors of the connection from the input layer to the hidden layer and from the hidden layer to the output layer, respectively; T H and T o are the activation thresholds of neurons in the hidden layer and output layer, respectively.

6. The underwater terrain extraction method based on neural network and ensemble learning according to claim 1 is characterized in that: The BP neural network algorithm is trained in S2, and the trained neural network model is used to generate a number of water depth inversion results, which also includes the following steps: The remote sensing images after atmospheric correction, land-water separation and mean filtering are matched with the measured water depth points, and the measured water depth values ​​and the reflectance values ​​of N bands of pixels at the corresponding positions of the measured water depth points constitute training samples, and the BP neural network algorithm is trained through the training samples; The reflectivity values ​​of N bands of each pixel in the remote sensing image are input into the trained neural network model, and the water depth value corresponding to each pixel is predicted to achieve the inversion of shallow sea underwater topography.

7. The underwater terrain extraction method based on neural network and ensemble learning according to claim 6 is characterized in that: The inversion of shallow sea underwater topography also includes the following steps: 300 points are randomly selected from the sonar measured water depth points as the training point set, and the remaining points are used as the test point set; ARCGIS software was used to spatially match the training point set and the test point set with the best synthetic image, and the reflectance values ​​of the three bands R, G, and B of each measured water depth point and the corresponding pixel were extracted to form training samples and test samples respectively. Select training samples to train the BP neural network model multiple times to generate multiple neural network models; Based on the reflectivity values ​​of the R, G, and B bands of the remote sensing image, multiple trained neural network models are input, and the water depth value corresponding to each pixel is predicted to obtain multiple sets of water depth prediction values; The underwater topography map is reconstructed according to the outlier degree of all water depth prediction values ​​at each pixel, and the inversion of the shallow underwater topography in the experimental area is completed.

8. The underwater terrain extraction method based on neural network and ensemble learning according to claim 1 is characterized in that: In S4, the water depth inversion result set of the entire study area is integrated, and the accuracy assessment of the integrated underwater topographic map also includes the following steps: Traverse the entire study area and integrate all water depth inversion results to obtain an integrated and high-precision underwater topographic map; A water depth inversion experiment was conducted in the study area, and the water depth inversion effect was evaluated based on the RMSE error; The smaller the RMSE value is, the higher the accuracy of water depth inversion is. Where zi is the measured water depth, is the inverted water depth value, n is the number of test samples, and i is a non-zero natural number.

Images