CN114926727A - Underwater terrain extraction method based on neural network and ensemble learning - Google Patents

Underwater terrain extraction method based on neural network and ensemble learning Download PDF

Info

Publication number
CN114926727A
CN114926727A CN202210391456.XA CN202210391456A CN114926727A CN 114926727 A CN114926727 A CN 114926727A CN 202210391456 A CN202210391456 A CN 202210391456A CN 114926727 A CN114926727 A CN 114926727A
Authority
CN
China
Prior art keywords
water depth
neural network
inversion
remote sensing
sensing image
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210391456.XA
Other languages
Chinese (zh)
Inventor
程亮
楚森森
程俭
张雪东
吴洁
刘东阁
左潇懿
薛清仁
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University
Original Assignee
Nanjing University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University filed Critical Nanjing University
Priority to CN202210391456.XA priority Critical patent/CN114926727A/en
Publication of CN114926727A publication Critical patent/CN114926727A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/05—Underwater scenes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
    • G06T17/05—Geographic models
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/10—Image enhancement or restoration by non-spatial domain filtering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10032—Satellite or aerial image; Remote sensing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/30—Assessment of water resources

Abstract

The invention discloses an underwater terrain extraction method based on a neural network and ensemble learning, which comprises the following steps: preprocessing a remote sensing image data set by adopting a digital image processing technology and combining a visual interpretation auxiliary mode, and taking a preprocessed remote sensing image and a training sample as an input data set; constructing a sub-learner based on a BP neural network algorithm, training the BP neural network algorithm, and generating a plurality of water depth inversion results by using the trained neural network model; determining an integration strategy based on a minimum outlier method; and integrating the water depth inversion result sets of the whole research area, and carrying out precision evaluation on the integrated underwater topography. Realizing high-precision shallow sea underwater topography; by combining the neural network and the integrated 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 precision and reliability are further improved.

Description

Underwater terrain extraction method based on neural network and ensemble learning
Technical Field
The invention relates to the field of shallow sea underwater topography inversion, in particular to an underwater topography extraction method based on a neural network and ensemble learning.
Background
The shallow water depth data has important significance for shipping management, island reef development, ecological protection and the like. Present sounding means is mainly with shipborne sonar and airborne Lidar (laser radar), and these on-the-spot sounding means are difficult to be used for having the island reef that touches the reef dangerous or be difficult to reach, can consume a large amount of manpower and materials simultaneously, are unfavorable for developing the depth of water measurement on a large scale. In recent years, a depth measurement method based on remote sensing images has the advantages of wide coverage range, low measurement cost, abundant data sources and the like, and is receiving wide attention.
Depth measurement research based on remote sensing images originated from 1970s, and with the development of satellite technology and depth measurement theory, a group of classical remote sensing image-based depth measurement methods are currently produced, such as theoretical analytic methods (Lee et al, 1998), stmlpf ratios (stmpf et al, 2003; Ma et al, 2020), Lyzenga polynomials (Lyzenga, 1978; Manessa et al, 2018), BP neural network algorithms (Sandidge et al, 20151998; Liu et al, 2018), and the like, wherein the BP neural network algorithms show optimal inversion performance in a large number of past studies (Chu et al, 2019; Ceyhun and yalin, 2010; ghanalhard et al, 2013; Liu et al, 2018), and thus the application of the BP neural network algorithms is wide. This method has been applied by many researchers in different types of environments and uses a variety of different satellite data, including Landsat-8(El-Mewafi et al, 2018), Sentinil-2 (Chu et al, 2019), Quickbird (Ceyhun and Yalcin, 2010), Spot-6(Hussein and Nadaoka, 2017), WorldView-3(Collin et al, 2017), and the like.
The BP neural network algorithm is a multi-layer feedforward network trained according to an error inverse propagation algorithm, and is the most widely used neural network model at present (Li et al, 2012). The BP neural network algorithm was proposed in 1986 by a scientific group, including Rumelhart and mccleland (1986). Subsequently, the american stannis space center navy research laboratory used the BP neural network algorithm for underwater terrain inversion (Sandidge and holier, 1998), whose rationale was: and (3) taking the spectral reflectivity as an input layer and the water depth data as an output layer, correcting the weight of the hidden layer by using a small amount of actually measured water depth values, and finally inverting the water depth of the pixel at the unknown water depth by using the trained model. The greatest advantage of the BP neural network algorithm is the excellent nonlinear fitting capability (Qiu et al, 2018). The water depth inversion based on the remote sensing image is influenced by numerous factors such as phytoplankton, colored soluble organic matter (CDOM), suspended particulate matters, substrate, solar illumination conditions and the like, so that the water depth inversion process is nonlinear (Ceyhun and Yalcin, 2010), which is also a main reason that the BP algorithm is superior to other classical algorithms in a large number of documents. The BP neural network algorithm has the other advantage that a physical model for water depth inversion does not need to be understood, the characteristic rule can be directly learned from a training sample, and the BP neural network algorithm is convenient to use and is widely applied.
However, BP neural network algorithms suffer from the problem of being prone to local minima (Hirose et al, 1991; Deng et al, 2021). The BP neural network algorithm works out an optimal solution through a gradient descent method, and because the error curved surface of the BP neural network is uneven, points with the gradient of 0 exist on the error curved surface, the points belong to local minimum value points but are not necessarily global minimum value points, and therefore the BP neural network algorithm can mistakenly consider that the error is not reduced any more and stop training after falling into the local minimum value points of the error, and finally lead to learning failure (Lee et al, 1993). The influence of local minimum values can be weakened by selecting appropriate parameters (such as learning rate and hidden layer nodes) and initial value weights, wherein the parameter selection can be based on experience or selecting appropriate values by a traversal trial method (Shahjahan and Murase, 2003; Benardos and Vosnikos, 2007), but the initial value weights are difficult to manually set due to large number, and a random assignment mode is usually adopted, so that the BP neural network algorithm has randomness when falling into the local minimum values and is difficult to manually control.
The underwater topography inversion based on the BP neural network algorithm also has the problems, and the local minimum value problem can cause an erroneous underwater topography inversion result, which is vigilant. Therefore, the underwater topography extraction method based on the neural network and the ensemble learning is provided, the problem of poor robustness of water depth inversion of the traditional BP neural network is effectively solved, and the method is applied to shallow sea underwater topography detection and has important significance.
An effective solution to the problems in the related art has not been proposed yet.
Disclosure of Invention
Aiming at the problems in the related art, the invention provides an underwater terrain extraction method based on a neural network and ensemble learning, so as to overcome the technical problems in the prior related art.
Therefore, the invention adopts the following specific technical scheme:
an underwater terrain extraction method based on neural network and ensemble learning comprises the following steps:
s1, pretreatment: the method comprises the steps that a digital image processing technology is adopted, a visual interpretation auxiliary mode is combined, pretreatment of atmospheric correction, land and water separation and mean value filtering is conducted on a remote sensing image data set, and a preprocessed remote sensing image and a training sample are used as input data sets;
s2, constructing a sub-learner: constructing a sub-learner based on a BP neural network algorithm, training the BP neural network algorithm, and generating a plurality of water depth inversion results by using a trained neural network model;
s3, integration strategy: determining an integration strategy based on a minimum outlier method;
s4, integrating water depth inversion results: and integrating the water depth inversion result sets of the whole research area, and carrying out precision evaluation on the integrated underwater topography.
Further, when the preprocessing of atmospheric correction is performed on the remote sensing image data set in S1, the remote sensing image is subjected to atmospheric correction by using the SNAP plug-in of the Sentinel image.
Further, the preprocessing of performing water-land separation on the remote sensing image data set in S1 further includes the following steps:
determining a land range manually by an empirical method in combination with visual interpretation, and taking the land range as a mask file;
and carrying out mask processing on the remote sensing image of the research area, and removing the land area in the remote sensing image.
Further, when the preprocessing of mean filtering is performed on the remote sensing image data set in S1, a window with a size of 3 × 3 is used to perform mean filtering processing on the remote sensing image without the land area.
Further, the formula of the BP neural network in S2 is as follows:
Figure BDA0003597092640000031
wherein I, H, O represents the vectors of input layer, hidden layer and output layer, w and v represent the weight vectors from input layer to hidden layer and from hidden layer to output layer, respectively, T H And T o Activation thresholds for hidden layer and output layer neurons, respectively.
Further, the training of the BP neural network algorithm in S2, and the generating of the inversion results for a plurality of water depths by using the trained neural network model further include the following steps:
matching the remote sensing image after atmospheric correction, land-water separation and mean filtering with the actual water depth point, forming a training sample by the actual water depth value and the reflectivity values of N wave bands of pixels at the corresponding positions of the actual water depth point, and training a BP neural network algorithm through the training sample;
and inputting the reflectivity value of N wave bands of each pixel in the remote sensing image into the trained neural network model, and predicting the water depth value corresponding to each pixel to realize inversion of shallow sea underwater topography.
Further, the implementation of the inversion of the shallow sea underwater topography further comprises the following steps:
randomly selecting 300 points from sonar actual measurement water depth points as a training point set, and taking the rest points as a test point set;
respectively carrying out spatial matching on the training point set and the test point set and the optimal synthetic image by using ARCGIS software, extracting the reflectivity values of each actually measured water depth point and three wave bands of the corresponding pixel R, G, B, and respectively forming a training sample and a test sample;
selecting training samples to train the BP neural network model for multiple times to generate a plurality of neural network models;
inputting a plurality of trained neural network models based on the reflectivity value of the R, G, B waveband of the remote sensing image, and predicting the water depth value corresponding to each pixel to obtain a plurality of sets of water depth predicted values;
and reconstructing an underwater topography map according to the outliers of all the water depth predicted values at each pixel to finish the inversion of the shallow water underwater topography in the experimental area.
Further, when the integration strategy is determined based on the minimum outlier method in S3, each sub-learner generates a water depth inversion result, that is, a water depth inversion result set S ═ d { d } is generated at each geographic location 1 ，d 2 ，…，d l ，…，d L L is the number of sub-learners, d l And expressing the water depth value of the ith inversion, wherein the minimum outlier mathematical formula is as follows:
Figure BDA0003597092640000041
in the formula (I), the compound is shown in the specification,OD l indicating the depth of water d l Degree of outlier, OD l Has a value range of [0, 1 ]]，OD l Larger, indicates d l The greater the difference from other water depth values, i.e. d is represented l The greater the degree of outlier(s) of (a), the more likely it is noise.
Further, the integration strategy in S3 further includes the following steps:
calculating the outlier of each inversion result, and obtaining S ═ d 1 ，d 2 ，d 3 ，…，d L The degree of outlier is { OD } 1 ，OD 2 ，OD 3 ，…，OD L }；
Finding out the water depth value corresponding to the minimum outlier, and setting the minimum outlier and the corresponding number p as OD p ＝min{OD 1 ，OD 2 ，OD 3 ，…，OD L Is determined as p ∈ {1, 2, 3, …, L }, while the water depth value with the smallest degree of outlier is formed into a new result set S' ═ { d } p }；
The new result set contains N number of elements, if N>1, then, for S' { d } p Averaging, and taking the average value as an integrated water depth result;
if the new result set contains only 1 element, i.e. N equals 1, the new result set d is added p As a result of the integrated water depth.
Further, in S4, the integrating the water depth inversion result set of the entire research area, and the performing accuracy evaluation on the integrated underwater topography further includes the following steps:
traversing the whole research area and integrating all water depth inversion results to obtain an integrated and high-precision underwater topographic map;
performing a water depth inversion experiment on the research area, and evaluating a water depth inversion effect based on the RMSE error;
wherein, the smaller the RMSE value is, the higher the accuracy of water depth inversion is;
Figure BDA0003597092640000051
in the formula, z i In order to actually measure the water depth value,
Figure BDA0003597092640000052
in order to invert the water depth value, n is the number of test samples, and i is a non-zero natural number.
The beneficial effects of the invention are as follows: the invention constructs a sub-learner based on a BP neural network algorithm, innovatively provides an integration strategy of a minimum outlier method, provides an underwater terrain extraction method based on a neural network and integrated learning, realizes high-precision shallow sea underwater terrain extraction, and provides technical support for island and reef construction, navigation safety, ecological protection and other applications. According to the invention, through the combination of the neural network and the integrated learning algorithm, the problem of poor robustness of the traditional BP neural network algorithm in the water depth inversion process is solved, the water depth inversion accuracy and reliability are further improved, and the defects of poor reliability and low accuracy of the traditional water depth inversion method are overcome.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.
Fig. 1 is a flowchart of an underwater topography extraction method based on neural network and ensemble learning according to an embodiment of the present invention;
FIG. 2 is an example study area profile;
FIG. 3 is a graph of error distribution for 100 replicates of an example study area;
FIG. 4 is a graph of the inversion results for the maximum RMSE error for underwater topography.
Detailed Description
For further explanation of the various embodiments, the drawings which form a part of the disclosure and which are incorporated in and constitute a part of this specification, illustrate embodiments and, together with the description, serve to explain the principles of operation of the embodiments, and to enable others of ordinary skill in the art to understand the various embodiments and advantages of the invention, and, by reference to these figures, reference is made to the accompanying drawings, which are not to scale and wherein like reference numerals generally refer to like elements.
According to the embodiment of the invention, a neural network and ensemble learning-based underwater topography extraction method is provided, compared with the traditional method, the method takes a BP neural network algorithm as a basic learner, generates a plurality of water depth inversion results, provides an integration strategy based on a minimum outlier method, integrates all the water depth inversion results, finally obtains an integrated high-precision underwater topography map, and provides a scientific and effective inversion method for shallow sea underwater topography detection.
Referring to the drawings and the detailed description, the invention will be further explained, as shown in fig. 1, in an embodiment of the invention, a method for extracting underwater topography based on neural network and ensemble learning, the method includes the following steps:
s1, preprocessing (remote sensing image data set preprocessing): the method comprises the following steps of carrying out pretreatment of atmospheric correction, land-water separation and mean value filtering on a remote sensing image data set by adopting a digital image processing technology and combining a visual interpretation auxiliary mode, and taking a pretreated remote sensing image and a training sample as an input data set;
when preprocessing of atmospheric correction is performed on the remote sensing image data set in S1, in order to perform quantitative extraction of water depth information, an SNAP plug of a 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.
The preprocessing of water-land separation of the remote sensing image data set in S1 further includes the steps of:
manually determining a land range by using an empirical method and visual interpretation, and taking the land range as a mask file;
and carrying out mask processing on the remote sensing image of the research area, and removing the land area in the remote sensing image.
When preprocessing of mean filtering is performed on the remote sensing image data set in S1, a window with a size of 3 × 3 is used to perform mean filtering processing on the remote sensing image from which the land area is removed.
In the step, firstly, image preprocessing such as atmospheric correction is carried out on the remote sensing image by using an SNAP plug-in, then, land areas on the remote sensing image are manually drawn by using ARCGIS software, and land and water separation is realized by mask processing; and finally, carrying out mean filtering on the remote sensing image by using a window with the size of 3 multiplied by 3.
S2, constructing a sub-learner (the sub-learner based on the BP neural network algorithm is constructed): taking the preprocessed optimal remote sensing image and the training sample as input data sets, constructing a sub-learner based on a BP neural network algorithm, training the BP neural network algorithm, and generating a plurality of water depth inversion results by using the trained neural network model;
wherein, the formula of the BP neural network in S2 is:
Figure BDA0003597092640000071
wherein I, H, O are the vectors of input layer (spectral reflectance), hidden layer and output layer (depth value), w and v are the weight vectors of input layer to hidden layer and hidden layer to output layer, respectively, T H And T o Activation thresholds for hidden layer and output layer neurons, respectively.
In S2, training the BP neural network algorithm, and generating inversion results for a plurality of water depths by using the trained neural network model further includes the following steps:
matching the remote sensing image after atmospheric correction, land-water separation and mean filtering with the actual water depth point, forming a training sample by the actual water depth value and the reflectivity values of N wave bands of pixels at the corresponding positions of the actual water depth point, and training a BP neural network algorithm through the training sample;
and inputting the reflectivity values of N wave bands of each pixel in the remote sensing image into the trained neural network model, and predicting the water depth value corresponding to each pixel to realize inversion of the shallow-sea underwater topography.
The method for realizing the inversion of the shallow sea underwater topography also comprises the following steps:
randomly selecting 300 points from sonar actual measurement water depth points as a training point set, and taking the rest points as a test point set;
respectively carrying out spatial matching on the training point set and the testing point set with the optimal synthetic image by using ARCGIS software, extracting reflectivity values of each actually measured water depth point and three wave bands of the corresponding pixel R, G, B, and respectively forming a training sample and a testing sample;
selecting training samples of 300 points to train the BP neural network model for multiple times to generate a plurality of neural network models; setting model parameters of the BP neural network as follows: the maximum training frequency is 1500, the learning rate is 0.05, the momentum factor is 0.9, and the training target error is 1 × 10 -5 ；
Inputting a plurality of trained neural network models based on the reflectivity value of the R, G, B waveband of the remote sensing image, and predicting the water depth value corresponding to each pixel to obtain a plurality of sets of water depth predicted values;
and reconstructing an underwater topography map according to the outliers of all the water depth predicted values at each pixel to finish the inversion of the shallow water underwater topography in the experimental area.
S3, integration strategy (integration strategy based on minimum outlier method): determining an integration strategy based on a minimum outlier method, integrating all water depth inversion results, and finally obtaining an integrated optimal underwater topography result
When the integration strategy is determined based on the minimum outlier method in S3, each sub-learner generates a water depth inversion result, that is, a water depth inversion result set S ═ d is generated at each geographic location 1 ，d 2 ，…，d l ，…，d L Where L is the number of sub-learners, d l And expressing the water depth value of the ith inversion, wherein the minimum outlier mathematical formula is as follows:
Figure BDA0003597092640000081
in the formula, OD l Indicating the depth of water d l Degree of outlier, OD l Has a value range of [0, 1 ]]，OD l Larger, indicates d l The greater the difference from other water depth values, i.e. d is represented l The greater the degree of outlier(s), the more likely it is to be noise.
The integration strategy in S3 further includes the steps of:
calculating the degree of outlier of each inversion result, and obtaining S ═ d 1 ，d 2 ，d 3 ，…，d L The degree of outlier is { OD } 1 ，OD 2 ，OD 3 ，…，OD L }；
Finding out the water depth value corresponding to the minimum outlier, and setting the minimum outlier and the corresponding number p as OD p ＝min{OD 1 ，OD 2 ，OD 3 ，…，OD L Is determined as p ∈ {1, 2, 3, …, L }, while the water depth value with the smallest degree of outlier is formed into a new result set S' ═ { d } p P may be one value or a plurality of values;
the new result set contains N number of elements, if N>1, then, for S' { d } p Averaging, and taking the average value as a water depth result after integration;
if the new result set contains only 1 element, i.e. N equals 1, the new result set d is added p As a result of the integrated water depth.
S4, integrating the water depth inversion results (integrating all the water depth inversion results): integrating a water depth inversion result set of the whole research area, and carrying out precision evaluation on the integrated underwater topography;
wherein, in S4, integrating the water depth inversion result sets of the whole research area, and performing accuracy evaluation on the integrated underwater topography map further includes the following steps:
traversing the whole research area and integrating all water depth inversion results to obtain an integrated and high-precision underwater topography;
performing 100 repeated experiments on the research area, and displaying a water depth inversion effect based on a result with the maximum RMSE (root mean square error) error;
as shown in fig. 4, the smaller the RMSE value is, the higher the accuracy of water depth inversion is;
Figure BDA0003597092640000091
in the formula, z i In order to measure the water depth value actually,
Figure BDA0003597092640000092
for inverting the water depth value, n is the number of test samples, and i is a non-zero natural number.
As shown in fig. 2, the example takes the ada reef in the south sea and south sand archipelago as the experimental area. The Anda reef is located at Zheng and reef cluster of Nansha, the longitude and latitude of the center are (10 degrees 20 'N, 114 degrees 42' E), and the area of the reef disc is about 30 square kilometers; the reef disc is shallow in the northeast, the shallowest position is close to the sea level, the southwest position is deep, the water depth is more than 20 meters, and the average water depth of the reef disc is about 10 meters. In the example, remote sensing image data are acquired from a Sentinel-2 satellite image of the Anda reef, which is shot from 2017-04-09, and in addition, 1315 sonar actual measurement water depth points are arranged in the Anda reef experimental area.
The parameters of the underwater terrain extraction algorithm based on the neural network and the ensemble learning are set as follows: the number of BP neural network algorithms is 5, the number of nodes of an input layer is 3 (red, green and blue), the number of nodes of an implicit layer is 7, transfer functions of the implicit layer and an output layer are respectively tandig and purelin, a training function is rainlm, the maximum training time is 1500, the learning rate is 0.05, a momentum factor is 0.9, and the error of a training target is 1 multiplied by 10 -5 . The experiment randomly selects 300 water depth samples as training samples, and the residual water depth samples as testing samples. The experiment is repeated for 100 times for the example research area by using the underwater terrain extraction algorithm based on the neural network and the ensemble learning provided by the invention. As can be seen from the RMSE error distribution results of repeated experiments (figure 3), the error distribution range of the underwater terrain extraction result of the algorithm is concentrated, and the robustness is strong. FIG. 4 shows the maximum RMSE error among all inversion results, and it can be seen that even the maximum RMSE error among 100 repeated experiments, the water depth inversion effect is still good, the water depth characteristics are obvious, and the island is formedThe reef disc has clear outline, and the reliability of the underwater terrain extraction method based on the neural network and the integrated learning is highlighted.
In conclusion, the invention constructs the sub-learner based on the BP neural network algorithm, innovatively provides the integration strategy of the minimum outlier method, provides the underwater terrain extraction method based on the neural network and the integrated learning, realizes the extraction of high-precision shallow sea underwater terrain, and provides technical support for the applications of island reef construction, navigation safety, ecological protection and the like. According to the invention, through the combination of the neural network and the integrated learning algorithm, the problem of poor robustness of the traditional BP neural network algorithm in the water depth inversion process is solved, the water depth inversion precision and reliability are further improved, and the defects of poor reliability and low precision of the traditional water depth inversion method are overcome.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.

Claims (10)

1. An underwater terrain extraction method based on a neural network and ensemble learning is characterized by comprising the following steps:
s1, pretreatment: the method comprises the following steps of carrying out pretreatment of atmospheric correction, land-water separation and mean value filtering on a remote sensing image data set by adopting a digital image processing technology and combining a visual interpretation auxiliary mode, and taking a pretreated remote sensing image and a training sample as an input data set;
s2, constructing a sub-learner: constructing a sub-learner based on a BP neural network algorithm, training the BP neural network algorithm, and generating a plurality of water depth inversion results by using the trained neural network model;
s3, integration strategy: determining an integration strategy based on a minimum outlier method;
s4, integrating water depth inversion results: and integrating the water depth inversion result sets of the whole research area, and carrying out precision evaluation on the integrated underwater topography.
2. The method for extracting the underwater topography based on the neural network and the ensemble learning of claim 1, wherein when the preprocessing of the atmospheric correction is performed on the data set of the remote sensing image in the S1, the SNAP plug-in of the Sentinel image is used for performing the atmospheric correction on the remote sensing image.
3. The method for extracting underwater topography based on neural network and ensemble learning of claim 1, wherein the preprocessing of water-land separation for the remote sensing image data set in S1 further comprises the following steps:
manually determining a land range by using an empirical method and visual interpretation, and taking the land range as a mask file;
and carrying out mask processing on the remote sensing image of the research area, and removing the land area in the remote sensing image.
4. The method for extracting underwater topography based on neural network and ensemble learning of claim 1, wherein in the preprocessing of mean filtering the data set of the remote sensing image in S1, a window with a size of 3 × 3 is used to perform mean filtering processing on the remote sensing image without land area.
5. The method for extracting underwater topography based on neural network and ensemble learning of claim 1, wherein the formula of the BP neural network in S2 is:
Figure FDA0003597092630000011
wherein I, H, O are input layer, hidden layer and output layer vectors, w and v are weight vectors from input layer to hidden layer and from hidden layer to output layer, respectively, T H And T o The activation thresholds for the hidden layer and output layer neurons, respectively.
6. The underwater topography extraction method based on neural network and ensemble learning of claim 1, wherein the training of the BP neural network algorithm in S2 and the generation of the inversion results for a plurality of water depths by using the trained neural network model further comprises the following steps:
matching the remote sensing image after atmospheric correction, land-water separation and mean filtering with the actual water depth point, forming a training sample by the actual water depth value and the reflectivity values of N wave bands of pixels at the corresponding positions of the actual water depth point, and training a BP neural network algorithm through the training sample;
and inputting the reflectivity values of N wave bands of each pixel in the remote sensing image into the trained neural network model, and predicting the water depth value corresponding to each pixel to realize inversion of the shallow-sea underwater topography.
7. The method for extracting underwater topography based on neural network and ensemble learning of claim 6, wherein said implementing the inversion of the shallow sea underwater topography further comprises the steps of:
randomly selecting 300 points from sonar actual measurement water depth points as a training point set, and taking the rest points as a test point set;
respectively carrying out spatial matching on the training point set and the test point set and the optimal synthetic image by using ARCGIS software, extracting the reflectivity values of each actually measured water depth point and three wave bands of the corresponding pixel R, G, B, and respectively forming a training sample and a test sample;
selecting training samples to train the BP neural network model for multiple times to generate a plurality of neural network models;
inputting a plurality of trained neural network models based on the reflectivity value of the R, G, B waveband of the remote sensing image, and predicting the water depth value corresponding to each pixel to obtain a plurality of sets of water depth predicted values;
and reconstructing an underwater topography according to the outliers of all the water depth predicted values at each pixel, and completing inversion of shallow water underwater topography in the experimental area.
8. The method of claim 1, wherein when the integration strategy is determined based on the minimum outlier method in S3, each sub-learner generates a water depth inversion result, that is, a set of water depth inversion results S { d } is generated at each geographic location 1 ，d 2 ，…，d l ，…，d L Where L is the number of sub-learners, d l The water depth value of the l-th inversion is represented, and the minimum outlier mathematical formula is as follows:
Figure FDA0003597092630000021
in the formula, OD l Indicating the depth of water d l Degree of outlier, OD l Has a value range of [0, 1 ]]，OD l Larger, indicates d l The greater the difference from other water depth values, i.e. d is represented l The greater the degree of outlier(s) of (a), the more likely it is noise.
9. The method for extracting underwater topography based on neural network and ensemble learning of claim 8, wherein said integration strategy in S3 further comprises the steps of:
calculating the outlier of each inversion result, and obtaining S ═ d 1 ，d 2 ，d 3 ，…，d L The degree of outlier is { OD } 1 ，OD 2 ，OD 3 ，…，OD L }；
Finding out the water depth value corresponding to the minimum outlier, and setting the minimum outlier and the corresponding number p as OD p ＝min{OD 1 ，OD 2 ，OD 3 ，…，OD L Is determined as p ∈ {1, 2, 3, …, L }, while the water depth value with the smallest degree of outlier is formed into a new result set S' ═ { d } p }；
The new result set contains N number of elements, if N>1, then, for S' { d } p Averaging, and taking the average value as an integrated water depth result;
if the new result set contains only 1 element, i.e. N equals 1, the new result set d is added p As a result of the integrated water depth.
10. The method for extracting underwater topography based on neural network and ensemble learning of claim 1, wherein in S4, the method further comprises the following steps when integrating the water depth inversion result set of the whole research area and performing precision evaluation on the integrated underwater topography:
traversing the whole research area and integrating all water depth inversion results to obtain an integrated and high-precision underwater topographic map;
performing a water depth inversion experiment on the research area, and evaluating a water depth inversion effect based on RMSE errors;
wherein, the smaller the RMSE value is, the higher the water depth inversion precision is;
Figure FDA0003597092630000031
in the formula, z i In order to actually measure the water depth value,
Figure FDA0003597092630000032
for inverting the water depth value, n is the number of test samples, and i is a non-zero natural number.
CN202210391456.XA 2022-04-14 2022-04-14 Underwater terrain extraction method based on neural network and ensemble learning Pending CN114926727A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210391456.XA CN114926727A (en) 2022-04-14 2022-04-14 Underwater terrain extraction method based on neural network and ensemble learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210391456.XA CN114926727A (en) 2022-04-14 2022-04-14 Underwater terrain extraction method based on neural network and ensemble learning

Publications (1)

Publication Number Publication Date
CN114926727A true CN114926727A (en) 2022-08-19

Family

ID=82806729

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210391456.XA Pending CN114926727A (en) 2022-04-14 2022-04-14 Underwater terrain extraction method based on neural network and ensemble learning

Country Status (1)

Country Link
CN (1) CN114926727A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115494013A (en) * 2022-11-17 2022-12-20 河北先河环保科技股份有限公司 Method and apparatus for detecting water quality abnormality and storage medium

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115494013A (en) * 2022-11-17 2022-12-20 河北先河环保科技股份有限公司 Method and apparatus for detecting water quality abnormality and storage medium

Similar Documents

Publication Publication Date Title
Hodúl et al. Satellite derived photogrammetric bathymetry
Hsu et al. A semi-empirical scheme for bathymetric mapping in shallow water by ICESat-2 and Sentinel-2: A case study in the South China Sea
Calder et al. Automatic processing of high‐rate, high‐density multibeam echosounder data
CN102176001B (en) Permeable band ratio factor-based water depth inversion method
CN104613944A (en) Distributed water depth prediction method based on GWR (geographically weighted regression) and BP (back propagation) neural network
CN105303030B (en) A kind of MODIS remote sensing estimation method of eutrophic lake algae Gauss vertical characteristics structural parameters
WO2014043764A1 (en) On-ground or near-ground discrete object detection method and system
CN115077656B (en) Reservoir water reserve retrieval method and device
CN104680151B (en) A kind of panchromatic remote sensing image variation detection method of high-resolution for taking snow covering influence into account
CN113050091B (en) Wind speed and direction joint inversion method and system for satellite-borne synthetic aperture radar
CN112013822A (en) Multispectral remote sensing water depth inversion method based on improved GWR model
CN114201732A (en) Sentinel-2A image-based shallow sea water depth inversion method
CN115512247A (en) Regional building damage grade assessment method based on image multi-parameter extraction
Park et al. Delineation of high resolution climate regions over the Korean Peninsula using machine learning approaches
CN114926727A (en) Underwater terrain extraction method based on neural network and ensemble learning
CN110132246B (en) Water depth remote sensing detection method based on residual partitioning
Derin et al. Modeling level 2 passive microwave precipitation retrieval error over complex terrain using a nonparametric statistical technique
Xie et al. Satellite-derived bathymetry combined with Sentinel-2 and ICESat-2 datasets using machine learning
CN111951204B (en) Sea surface wind speed inversion method for Tiangong No. two detection data based on deep learning
CN112966656A (en) Data processing method and device
Sharma et al. Semi-automated workflow for mapping the extent and elevation profile of intertidal zone of parts of Gulf of Kutch, India, using landsat time series data
Ishalina et al. Recognition of crevasses with high‐resolution digital elevation models: Application of geomorphometric modeling and texture analysis
Liu et al. Bathymetric ability of SPOT-5 multi-spectral image in shallow coastal water
Tournadre et al. High-resolution imaging of the ocean surface backscatter by inversion of altimeter waveforms
CN115060656A (en) Satellite remote sensing water depth inversion method based on sparse prior actual measurement points

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination