WO2021218424A1 - 基于rbf神经网络的航海雷达图像反演海面风速方法 - Google Patents

基于rbf神经网络的航海雷达图像反演海面风速方法 Download PDF

Info

Publication number
WO2021218424A1
WO2021218424A1 PCT/CN2021/080049 CN2021080049W WO2021218424A1 WO 2021218424 A1 WO2021218424 A1 WO 2021218424A1 CN 2021080049 W CN2021080049 W CN 2021080049W WO 2021218424 A1 WO2021218424 A1 WO 2021218424A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
sea surface
rbf neural
wind speed
surface wind
Prior art date
Application number
PCT/CN2021/080049
Other languages
English (en)
French (fr)
Inventor
邱海洋
王慧
智鹏飞
朱琬璐
朱志宇
Original Assignee
江苏科技大学
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 江苏科技大学 filed Critical 江苏科技大学
Publication of WO2021218424A1 publication Critical patent/WO2021218424A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/0002Inspection of images, e.g. flaw detection
    • GPHYSICS
    • G01MEASURING; TESTING
    • G01SRADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S13/00Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
    • G01S13/88Radar or analogous systems specially adapted for specific applications
    • G01S13/89Radar or analogous systems specially adapted for specific applications for mapping or imaging
    • GPHYSICS
    • G01MEASURING; TESTING
    • G01SRADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S13/00Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
    • G01S13/88Radar or analogous systems specially adapted for specific applications
    • G01S13/95Radar or analogous systems specially adapted for specific applications for meteorological use
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/23Clustering techniques
    • G06F18/232Non-hierarchical techniques
    • G06F18/2321Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00Image enhancement or restoration
    • G06T5/70Denoising; Smoothing
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/10Image acquisition modality
    • G06T2207/10032Satellite or aerial image; Remote sensing
    • G06T2207/10044Radar image
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20024Filtering details
    • G06T2207/20032Median filtering
    • YGENERAL 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
    • Y02TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02ATECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Definitions

  • the invention relates to the technical field of remote sensing of sea surface wind speed, in particular to a method for inversion of sea surface wind speed based on RBF neural network for navigation radar images.
  • the sea surface wind field is of great significance to the research of navigation operations and ocean dynamics, and is an important means to understand ocean changes and predict risks.
  • the sea surface wind field information mainly includes two aspects of the sea surface wind direction and the sea surface wind speed.
  • the present invention is used to retrieve the sea surface wind speed information.
  • the existing remote sensing observations mainly include scatterometers, airborne or spaceborne synthetic aperture radar (SAR), and satellite altimeters.
  • Scatterometers have the problem of low resolution. Satellite remote sensing has a low re-sampling rate and is interfered by clouds. As a result, the measured data may not be the wind speed information to be detected on the sea surface.
  • X-band radar has the advantages of being unaffected by light, real-time continuous feedback and high resolution, and has become an important means of obtaining wind field information at this stage. At present, X-band navigation radar has been used at home and abroad to realize the research of sea surface waves, currents, rainfall, and sea surface oil spill area.
  • Bueno et al. used the linear integration method to obtain the relationship between the radar echo intensity level and the sea surface wind speed to obtain the sea surface wind speed information.
  • Liu Y et al. proposed the application of hyperbolic line fitting to extract sea surface wind speed information using measured marine radar data.
  • Huang W et al. proposed the use of RCS spectrum analysis algorithm, RCS and sea surface wind speed empirical mode decomposition method for Decca radar, and established a function model to obtain sea surface wind speed.
  • Chen Zhongbiao et al. fitted RCS effective wave height and sea surface wind speed to a linear probability distribution function, thereby obtaining sea surface wind speed information.
  • the model function method does not have universal applicability. It is necessary to model the sea surface wind speed and RCS for different radar models; if the data used in the modeling cannot include various sea conditions, the error can be retrieved for the same radar model under different sea conditions. It will also be very large, restricting the development prospects of this method.
  • Dankert first proposed the neural network method. According to the relationship between the radar cross-sectional area and the wind speed, the sea surface wind direction information and NRCS were used as input, and the sea surface wind speed was inversed based on the BP neural network. In 2006, Dankert considered humidity, temperature, signal-to-noise ratio and other marine factors as the input of BP neural network to improve the applicability of radar data.
  • the present invention discloses a method for retrieving sea surface wind speed from marine radar images based on RBF neural network.
  • the RBF neural network input layer samples are constructed by combining sensor and image information to improve the robustness of the network.
  • the subtractive clustering algorithm uses the subtractive clustering algorithm to select the center of the sample target matrix, terminate the reverse judgment by the cluster density index judgment condition, obtain the basis function parameters and the expansion constant, and use the recursive least squares method to obtain the output layer connection weights, thereby effectively determining RBF neural network model.
  • the experimental data proves the feasibility of the method to extract sea surface wind speed information from the navigation radar image.
  • Step 1 Preprocessing of marine radar image data.
  • Collect the navigation radar image sequence contains N radar images.
  • the actual wind direction, wind speed and sea state information obtained by the corresponding time and position synchronization sensor are collected, and the median filter is performed on each radar image to suppress the same frequency and rain. The effect of snow noise on the image.
  • Step 2 RBF neural network input layer construction. Perform global low-pass filtering on the aerial radar image sequence to obtain the sea surface static feature image, select the image with obvious wind streak characteristics, use the wind streak scale band pass filter to obtain the sea surface wind field energy spectrum, and construct the RBF neural network input layer sample based on the sensor and image information .
  • Step 3 Determine the RBF neural network model. Normalized mapping of the obtained sample data sea surface wind field energy spectrum average value, sea surface wind direction, wind speed information and sea state information, establishes an RBF neural network framework, uses the subtractive clustering algorithm to select the center of the sample target matrix, and the cluster density index The reverse judgment is terminated, the basis function parameters and expansion constants are obtained, and the recursive least square method is used to obtain the connection weights of the output layer, and finally the RBF neural network model is determined.
  • the present invention proposes a method for retrieving sea surface wind speed from marine radar images based on RBF neural network.
  • the step 3 includes the following steps:
  • Step 3.1 normalized mapping is performed on the average value of the energy spectrum of the sea surface wind field, the sea surface wind direction information and the sea state information;
  • i is the number of corresponding radar image sequences.
  • Step 3.2 establish an RBF neural network framework
  • step 3.1 S i, R i, D i P i, and all data is divided into two sections, one for train the RBF neural network empirical model to obtain surface wind speed and one for the test model.
  • the first input layer of the RBF neural network designed by this method is the signal source node.
  • the signal source node is composed of sea surface wind field energy spectrum, measured wind direction, measured wind speed and sea state information, namely:
  • the function of the second hidden layer is a nonlinear basis function.
  • the center of the nonlinear basis function refers to the measured wind speed. According to the characteristics of the nonlinear basis function, it can be known that when the data of the input layer is closer to the measured wind speed, that is, the closer to the center, the hidden The data corresponding to the containing layer will be larger.
  • the third layer is the output layer, which responds to input patterns.
  • the RBF input layer to the hidden layer is a non-linear mapping relationship, which is calculated by multiplying the Euclidean distance between the input value and the center point by a constant. As the center point is determined, the weight is also determined.
  • the connection method from the hidden layer to the output layer uses linear weighting.
  • the nonlinear basis function generally chooses the Gaussian function, and the formula is as follows:
  • x i is the signal source node of the first layer
  • b j refers to the center of the radial basis function, that is, the center vector of the measured wind speed
  • ⁇ i is the expansion constant of the i-th hidden layer unit.
  • the weight vector from the hidden layer to the output layer of the RBF neural network is:
  • the RBF network training process is divided into two steps.
  • the first step is to determine the number of hidden layer units and the center and expansion constant of the basis function;
  • the second step is to determine the connection weight w i from the hidden layer to the output layer.
  • Step 3.3 data point density index calculation
  • This method uses a subtractive clustering algorithm to determine the number of hidden layer units and expansion constants in the design network, and uses a recursive least squares algorithm to determine w i .
  • the calculation formula is as follows:
  • I i is the density value index of the input sample x i
  • ⁇ a is the density index field.
  • the data points outside this field have little or no influence on I i and can be ignored.
  • the point with the largest I i obtained by formula (7) is the first cluster center, and its corresponding data point is marked as x c1 , and the corresponding density index is marked as I c1 , thus, the density of each data point is updated
  • the index formula is as follows:
  • Step 3.4 subtract the clustering algorithm to obtain the node center and expansion constant
  • the cluster center I cs is continuously determined until the clustering judgment condition is met, namely:
  • is a very small value.
  • the finally obtained cluster center is the data node center, and the obtained ⁇ b is the expansion constant.
  • Step 3.5 Determine the connection weight of the output layer in the RBF neural network
  • Application training data S i, R i, D i as the data value, P i is the observed value, obtained using a recursive least squares connection weights w i, namely:
  • e p (i) is the priori error
  • K (i) is the Kalman gain, both the data obtained and observed recursive obtained.
  • Step 4 Extraction of sea surface wind speed information.
  • the average value of the sea surface wind field energy spectrum, the sea surface wind direction information and the sea state information obtained from the test marine radar image sequence are normalized and mapped, and input into the RBF neural network model to obtain the sea surface wind speed information.
  • the present invention proposes a method for retrieving sea surface wind speed from marine radar images based on RBF neural network.
  • An RBF neural network framework is designed to establish the model relationship between marine radar images and sea surface wind fields.
  • the RBF neural network adopts a single hidden layer structure. , Improve the speed of sea surface wind speed inversion;
  • the present invention proposes a method for retrieving sea surface wind speed from marine radar images based on RBF neural network.
  • the designed RBF neural network can approximate nonlinear functions arbitrarily with accuracy, has the ability of global approximation, and does not have the advantage of falling into local optimum , which can improve the applicability of wind speed inversion;
  • the present invention proposes to apply a subtractive clustering algorithm to determine neural network parameters.
  • the idea is to determine the clustering center according to the density index of the input sample, which enhances the training speed of the algorithm in engineering and the robustness of the model;
  • the present invention proposes a method for retrieving sea surface wind direction from marine radar images based on RBF neural network, which uses the normalized information of sea surface wind field energy spectrum, sensors and sea state information to construct RBF neural network input layer samples, which improves the algorithm in engineering The inversion accuracy.
  • FIG. 1 Architecture diagram of RBF neural network
  • FIG. 7 Two types of sea surface wind speed inversion neural network training error distribution diagrams
  • Figure 9 The comparison diagram of the error of two kinds of neural networks extracting sea surface wind speed information
  • FIG. 10 Two kinds of neural networks extract the error distribution range of sea surface wind speed information.
  • Fig. 1 The flow of the specific implementation of the present invention is shown in Fig. 1, which is divided into four major blocks: navigation radar image preprocessing, RBF neural network input layer construction, RBF neural network model determination, and sea surface wind speed information extraction.
  • the specific implementation steps are divided into sixteen steps.
  • the first to second steps are data preprocessing;
  • the third to sixth steps are the construction of the RBF neural network input layer;
  • the seventh to fourteenth steps are the RBF neural network model Confirm;
  • the fifteenth to sixteenth steps are the extraction and analysis of sea surface wind speed information. Specific steps are as follows:
  • the first step is to collect 350 sets of navigation radar image sequences from October to December 2010. Each set of image sequences contains 32 navigation radar images. The total elapsed time is about 1.5 minutes.
  • the sea surface wind direction ⁇ obtained by the anemometer is synchronously recorded.
  • w , wind speed U w , the measured wind direction and wind speed distribution are shown in Figure 2.
  • the second step is to suppress the influence of the same frequency signal on the sea surface wind speed inversion by preprocessing the navigation radar image, and apply a 3 ⁇ 3 template 2D nonlinear smoothing median filter to each radar image in the navigation radar image sequence.
  • the image gray value g'(r, ⁇ ) is:
  • f(s,t) is the echo intensity value at the polar coordinates (s,t) of the radar image; f'(r, ⁇ ) is the gray scale at the polar coordinates (r, ⁇ ) after filtering Value; N (r, ⁇ ) is the pixel point centered at (r, ⁇ ), and (s, t) is the 8 pixel points adjacent to (r, ⁇ ) as the center.
  • the center of the median filter 3 ⁇ 3 template is coincident with the N(r, ⁇ ) center of the polar coordinate image, and the echo intensity values of the surrounding 8 adjacent pixel points (s, t) are compared through (r, ⁇ ) , Select the middle value of echo intensity to update the echo intensity value at N(r, ⁇ ), the template traverses the polar coordinate navigation radar image with a step unit of 1, and finally obtains the navigation radar image after median filtering.
  • the signal-to-noise ratio of each image is calculated as the sea state information.
  • the calculation formula is as follows:
  • I (3) (k, ⁇ ) is the three-dimensional image spectrum of f'(r, ⁇ )
  • F (2) (k) is the two-dimensional ocean wave image spectrum
  • N kx , N ky , and N ⁇ are the ranges of the spectrum
  • ⁇ k x and ⁇ k y are wave number resolution
  • is frequency resolution
  • the obtained signal-to-noise ratio r t of each radar image is normalized with the radar image time series to obtain the normalized value of sea state information R i :
  • the sea surface static feature image containing the wind streak characteristics must be extracted from the marine radar image sequence.
  • the filtered marine radar image sequence is shown in Figure 3.
  • the present invention is implemented by constructing a global low-pass filter, and the construction of the global low-pass filter is as follows:
  • f'( ⁇ ,r,t) is the maritime radar image sequence after median filtering
  • f( ⁇ ,r) is the static feature image of the sea surface in polar coordinates
  • N t is the number of navigation radar images included in the navigation image sequence
  • the static feature image of the sea surface is shown in Figure 4.
  • the closest point interpolation process is performed on f( ⁇ ,r) to obtain the image F(k x , k y ) in Cartesian coordinates.
  • a fifth step the same sea radar site location information on wind anemometer obtained d i, surface wind speed information p i, according to the time-series radar images is normalized, to obtain sea surface wind direction and speed information normalized value D i and P i :
  • a sixth step, a third, four, five steps 350 sets the obtained S i, R i, D i P i and the data is divided into two portions, 175 group of data used to train the RBF neural network empirical model to obtain surface wind speed , The remaining 175 sets of data are used for model testing.
  • the seventh step is based on the correlation between the sea surface wind direction and wind speed and the radar echo image, as shown in the following formula:
  • ⁇ 0 A( ⁇ )u ⁇ ( ⁇ ) (1+B(u, ⁇ )cos ⁇ +C(u, ⁇ )cos2 ⁇ ) (9)
  • ⁇ 0 is the normalized radar cross-sectional area
  • u is the local offshore surface wind speed
  • is the angle between the electromagnetic wave and the wind direction
  • is the radar incidence angle
  • A, B, C and ⁇ are empirical parameters, which are determined by the radar frequency and polarization The way is decided. It can be seen from the formula that the radar NRCS has an exponential function relationship with the sea surface wind speed, and a harmonic function relationship with the angle between the electromagnetic wave and the wind direction.
  • the first input layer of the RBF neural network designed by this method is the signal source node, and the signal source node is composed of sea surface wind field energy spectrum, measured wind direction, measured wind speed and sea state information, namely:
  • the RBF neural framework established by the present invention is shown in FIG. 5.
  • the function of the second hidden layer in the framework is a nonlinear basis function.
  • the center of the nonlinear basis function refers to the measured wind speed. According to the characteristics of the nonlinear basis function, it can be known that the closer the input layer data is to the measured wind speed, that is, the closer to the center. , The data corresponding to the hidden layer will be larger.
  • the third layer is the output layer, which responds to input patterns.
  • the eighth step is to select the RBF neural network model function.
  • the input layer to the hidden layer is a non-linear mapping relationship, which is calculated by multiplying the Euclidean distance between the input value and the center point by a constant. As the center point is determined, the weight is also determined.
  • the connection method from the hidden layer to the output layer uses linear weighting.
  • the nonlinear basis function generally chooses the Gaussian function, and the formula is as follows:
  • x i is the signal source node of the first layer
  • b j refers to the center of the radial basis function, that is, the center vector of the measured wind speed
  • ⁇ i is the expansion constant of the i-th hidden layer unit.
  • the RBF network training process is divided into two steps.
  • the first step is to determine the number of hidden layer units and the center and expansion constant of the basis function;
  • the second step is to determine the connection weight w i from the hidden layer to the output layer.
  • the ninth step is to use the subtractive clustering algorithm to determine the number of hidden layer units in the design network and the expansion constant.
  • the subtractive clustering algorithm flow is shown in Figure 6, and the steps are as follows:
  • the tenth step is to calculate the density index of each input sample x i .
  • the calculation formula is as follows:
  • I i is the density value index of the input sample x i
  • ⁇ a is the density index field. Data points outside this field have little or no impact on I i and can be ignored.
  • the point with the largest I i obtained by formula (15) is the first cluster center, and its corresponding data point is marked as x c1 , and the corresponding density index is marked as I c1 .
  • the density index of each data point is updated, and the calculation formula is as follows:
  • the twelfth step according to the obtained first cluster center and the updated data point density index, obtain the cluster judgment conditions, namely:
  • the cluster center I cs is continuously determined, and the clustering determination is stopped until the clustering condition is met, and the obtained cluster center is the data node center.
  • the thirteenth step is to determine the connection weights of the output layer in the RBF neural network.
  • Application training data S i, R i, D i as the data value, P i is the observed value, obtained using a recursive least squares connection weights w i, namely:
  • e p (i) is the priori error
  • K (i) is the Kalman gain, the data obtained from both recursive and observed.
  • the fourteenth step is to apply the RBF neural network designed by the present invention and the traditional BP neural network to train the 175 sets of training data, and obtain the training error distribution diagram of the sea surface wind speed inversion for the two kinds of neural networks, as shown in Figure 7,
  • (a ) Is the error distribution diagram of the RBF neural network
  • (b) is the error distribution diagram of the traditional BP neural network.
  • a conventional neural network BP 1000 reached steady error, the convergence error of 0.3m / s, the convergence time of 19.4s; RBF neural network training time error of about 300 to stabilize the convergence error of 10 - 3 m/s, the convergence time of MATLAB simulation is 0.8s, and the inversion accuracy and convergence speed can better meet the requirements of engineering applications.
  • the fifteenth step is to normalize the average value of the sea surface wind field energy spectrum, the sea surface wind direction information and the sea state information obtained from the test marine radar image sequence, and input them into the traditional BP neural network and RBF neural network models, two kinds of neural networks
  • the result of extracting sea surface wind speed information is shown in Figure 8. It can be obtained from Figure 8 that the wind speed results retrieved by the RBF neural network fluctuate slightly with the trend of the actual wind speed, but they are closer to the true value of the wind speed.
  • the sea surface wind speed information obtained by the BP neural network has a large fluctuation range with the actual wind speed, especially when the wind speed is 15-20m/s, the fluctuation reaches more than 5m/s.
  • the sixteenth step is to obtain two kinds of neural network inversion wind speed results and actual measured wind speed statistical results as shown in Table 1.
  • the correlation coefficient between the wind speed inversion results of the present invention and the actual measured wind direction reaches 0.87, and the standard deviation is 2.1 m/s.
  • the deviation is -0.08, which fully meets the engineering requirements.
  • the error of the inversion results of the BP neural network and the RBF neural network is compared with the true value.
  • the result is shown in Figure 9, and the error range is shown in Figure 10.
  • Figure 9 the error range of the RBF neural network is generally between -5 and +5m/s, while the error range of the BP neural network is between -10 and +10m/s, indicating the accuracy of the RBF neural network inversion result higher.
  • Figure 10 shows the error distribution range of the sea surface wind speed inversion results of the two algorithms. It can be seen that the error range of the RBF neural network is smaller. 64% of the data errors are mainly concentrated in -2 ⁇ 0m/s and 0 ⁇ 2m/s.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Remote Sensing (AREA)
  • General Physics & Mathematics (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Electromagnetism (AREA)
  • Software Systems (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Evolutionary Biology (AREA)
  • Quality & Reliability (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

一种基于RBF神经网络的航海雷达图像反演海面风速的方法,包含航海雷达图像数据预处理、RBF神经网络输入层构建、RBF神经网络模型确定和海面风速信息提取四个部分。海面风速反演过程是基于单隐层RBF神经网络训练得到的模型完成的,RBF神经网络输入层样本采用海面风场能量谱、传感器信息及海况信息的归一化结果构建;同时提出应用减聚类算法,根据输入样本的密度指标及聚类判断条件确定神经网络确定隐层单元数和基函数的中心及扩展常数,利用递推最小二乘获得网络输出层连接权值。

Description

基于RBF神经网络的航海雷达图像反演海面风速方法 技术领域
本发明涉及属于海面风速遥感技术领域,具体涉及一种基于RBF神经网络的航海雷达图像反演海面风速方法。
背景技术
海面风场对航海作业和海洋动力学研究有重要的意义,且是了解海洋变化与预知风险的重要手段。海面风场信息主要包括海面风向和海面风速两个方面,本发明是用于反演海面风速信息的。
传统的提取海面风速信息的方式主要有两大类:站点式观测和遥感反演。站点式观测通常在船上或岸边使用测风仪,但由于船体结构及船上建筑物的或是岸基等影响而产生的湍流效应,使风速测量的误差较大,不能满足实际应用。若测风仪安装在浮标上,其测量结果虽较为准确,但测得的仅是浮标位置及附近局部区域的风速信息,还易受海上天气或是海上交通的影响,在时间上和空间上缺乏连续性。
现有遥感观测主要有散射计、机载或星载合成孔径雷达(SAR)和卫星高度计等。散射计存在分辨率较低的问题,卫星遥感重复采样率低,且受到云层的干扰,导致测量数据可能不是海表面所要探测的风速信息。X波段雷达具有不受光线影响、实时连续反馈和高分辨率等优点,成为现阶段风场信息获取的重要手段。目前国内外已应用X-band导航雷达实现了海面浪、流,降雨量,海面漏油面积等方面的研究。现有基于航海雷达图像反演海面风速主要算法有两种:一种是模型函数法,一种是神经网络法。Horstmann首次提出应用地球物理模型函数(GMF)得到海面风速信息,但此模型是针对SAR回波强度进行建模,不能准确表达航海雷达回波截面积(RCS)与海面风速的关系,但足已证明海面风场与雷达回波强度具有一定的模型函数关系。Lund等针对FurunoFAR2117BB型号航海雷达,将RCS与海面风速单幅图像进行数据拟合,得到两者之间存在三次多项式非线性关系,并利用计算出风速。Bueno等针对Furuno2117BB型号雷达,利用线性积分法得到雷达回波强度水平与海面风速函数关系获得海面风速信息。Liu Y等针对Decca和Furono两种雷达,提出应用双曲 线拟合应用实测航海雷达数据提取出海面风速信息。Huang W等针对Decca雷达,提出了利用RCS谱分析算法、RCS与海面风速经验模态分解方法,建立函数模型得到海面风速。陈忠彪等针对9.3GHz Furuno雷达,将RCS、有效波高与海面风速拟合成线性概率分布函数,由此获得海面风速信息。但是,模型函数法不具有普遍应用性,需要针对不同雷达型号对海面风速和RCS进行建模;若建模时应用的数据不能包含各种海况条件,对于同一雷达型号在不同海况下反演误差也会很大,制约了该方法的发展前景。2002年Dankert首先提出神经网络法,根据雷达散射截面积与风速之间存在的关系,利用海面风向信息和NRCS作为输入量,基于BP神经网络反演出海面风速。2006年,Dankert考虑了湿度、温度、信噪比等海洋因素作为BP神经网络的输入量,以提高雷达数据的适用性。2007年Horstmann提出利用提取出的雷达图像序列中的风信息和NRCS做为BP神经网络的输入量,以提高海面风速的反演精度。哈尔滨工程大学贾瑞才采用双隐层单极型S函数BP神经网络法反演出海面风速信息,提高了神经网络的收敛速度及网络推广能力。但BP神经网络存在固有的不足之处,主要表现为易限于局部极小值,学习过程收敛速度慢,隐层和隐层节点数难以确定的问题,从而导致网络对于部分雷达数据不适用,反演精度低及运算时间无法达到工程要求的缺陷。
发明内容
针对以上问题,本发明公开了一种基于RBF神经网络的航海雷达图像反演海面风速方法,首先,结合传感器及图像信息构建RBF神经网络输入层样本,提高了网络的鲁棒性。其次,利用减聚类算法对样本目标矩阵选取中心,由聚类密度指数判定条件终止反向判定,得到基函数参数和扩展常数,利用递推最小二乘法获得输出层连接权值,从而有效确定RBF神经网络模型。最后,通过实验数据证明该方法从航海雷达图像中提取出海面风速信息结果的可行性。
具体包括步骤如下:
步骤1,航海雷达图像数据预处理。采集航海雷达图像序列,航海雷达图像序列中包含N幅雷达图像,同时采集对应时间和位置同步传感器获得的实际风向、风速及海况信息,对每幅雷达图像进行中值 滤波,抑制同频、雨雪噪声对图像的影响。
步骤2,RBF神经网络输入层构建。对航雷达图像序列进行全局低通滤波得到海面静态特征图像,选取风条纹特征明显图像,利用风条纹尺度带通滤波器获得海面风场能量谱,结合传感器及图像信息构建RBF神经网络输入层样本。
步骤3,RBF神经网络模型确定。对获取的样本数据海面风场能量谱平均值、海面风向、风速信息及海况信息进行归一化映射,建立RBF神经网络框架,利用减聚类算法对样本目标矩阵选取中心,由聚类密度指数终止反向判定,得到基函数参数和扩展常数,利用递推最小二乘法获得输出层连接权值,最终确定RBF神经网络模型。
本发明提出基于RBF神经网络的航海雷达图像反演海面风速方法,所述步骤3包括以下步骤:
步骤3.1,海面风场能量谱平均值、海面风向信息及海况信息进行归一化映射;
①对获得的海面风场能量谱F(k x,k y)沿x和y轴归一化,得到能量谱均值S i
Figure PCTCN2021080049-appb-000001
其中,i是对应的雷达图像序列数。
②获得每幅雷达图像信噪比r t,以雷达图像时间序列进行归一化,得到海况信息归一化值R i
Figure PCTCN2021080049-appb-000002
其中
Figure PCTCN2021080049-appb-000003
为二维波数谱经校正后的海浪谱,
Figure PCTCN2021080049-appb-000004
为雷达图像海浪信号以外的噪声谱。
③对风力计获得的海面风向信息d i、海面风速信息p i,按雷达图像时间序列进行归一化,得到海面风向、风速信息归一化值D i及 P i
Figure PCTCN2021080049-appb-000005
步骤3.2,建立RBF神经网络框架;
①对步骤3.1获得的S i、R i、D i和P i的所有数据分成两个部分,一部分用于RBF神经网络模型的训练得到海面风速经验模型,另一部分用于模型的测试。
②依据海面风场与相关信息的关联性,建立RBF神经框架;
该方法设计的RBF神经网络第一层输入层为信号源节点。所述的信号源节点由海面风场能量谱、实测风向、实测风速和海况信息组成,即:
x i=[S i D i R i P i],(i=1,2,L,n)    (4)
第二层隐含层的函数是非线性基函数,非线性基函数的中心就是指实测风速,根据非线性基函数的特点可知,当输入层的数据越靠近实测风速,即越靠近中心时,隐含层对应的数据就会越大。第三层是输出层,它随着输入模式的作用而响应。
③RBF输入层到隐含层是一种非线性的映射关系,它按照输入值与中心点的欧式距离乘以一个常数来计算的,随着中心点的确定,权重也随之确定。而隐含层到输出层的连接方式则采用了线性加权。其中,非线性基函数一般选择高斯基函数,公式如下:
Figure PCTCN2021080049-appb-000006
其中,x i为第一层的信号源节点,b j是指径向基函数的中心,即实测风速的中心向量,σ i是第i个隐层单元的扩展常数。
RBF神经网络隐含层到输出层的权值向量为:
w=[w 1,w 2,w 3,L,w p]      (5)
可以得出RBF神经网络的输出G m(x)为:
Figure PCTCN2021080049-appb-000007
RBF网络训练过程分为两步,第一步是确定隐层单元数和基函数的中心及扩展常数;第二步是确定隐层到输出层的连接权值w i
步骤3.3,数据点密度指标计算;
该方法采用减聚类算法确定设计网络中隐层单元数、扩展常数,利用递归最小二乘算法确定w i。为了确定聚类中心,首先要计算输入样本的密集度指标,计算公式如下:
Figure PCTCN2021080049-appb-000008
I i为输入样本x i密度值指标,λ a为密度指标领域,该领域外的数据点对I i的影响很小或者几乎没有,可以忽略,这里λ a定义为一个常数,λ a=18。也就是说I i越高,则在该密度领域λ a的数据点必然较多。
通过公式(7)得到的最大I i的点最为第一个聚类中心,并将其对应的数据点标记为x c1,相对应的密度指标标记为I c1,由此,更新各数据点密度指标公式如下:
Figure PCTCN2021080049-appb-000009
其中,λ b与λ a定义相同,但通常大于λ a,λ b=27。
步骤3.4,减聚类算法得到节点中心和扩展常数;
随着不断更新数据点密度指标,不断的确定聚类中心I cs,直到满足聚类判定条件,即:
Figure PCTCN2021080049-appb-000010
其中,γ为很小的数值。
由此得到最终获得的聚类中心为数据节点中心,得到的λ b为扩展常数。
步骤3.5,RBF神经网络中输出层连接权值的确定;
应用训练样本数据S i、R i、D i作为数据值,P i为观测值,利用递推最小二乘获得网络连接权值w i,即:
w(i)=w(i-1)+K(i)e p *(i)      (10)
其中,e p(i)为先验误差,K(i)为卡尔曼增益,两者均由数据值和观测值递推得到得到。
步骤4,海面风速信息提取。对测试航海雷达图像序列得到的海面风场能量谱平均值,海面风向信息和海况信息进行归一化映射,输入到RBF神经网络模型中,得到海面风速信息。
与传统的BP神经网络反演海面风速方法相比,本发明的优点在于:
1.本发明提出基于RBF神经网络的航海雷达图像反演海面风速方法,设计了一种RBF神经网络框架用于建立航海雷达图像与海面风场的模型关系,RBF神经网络采用单隐层的结构,提高了海面风速反演速度;
2.本发明提出基于RBF神经网络的航海雷达图像反演海面风速方法,设计的RBF神经网络,可以精度上任意的逼近非线性函数,具有全局逼近的能力,不会有陷入局部最优的优点,可以提高风速的反演适用性;
3.本发明提出应用减聚类算法确定神经网络参数,其思想是根据输入样本的密度指标确定聚类中心,增强了算法在工程上的训练速度及模型的鲁棒性;
4.本发明提出基于RBF神经网络的航海雷达图像反演海面风向方法,利用得海面风场能量谱,传感器及海况信息的归一化信息构建RBF神经网络输入层样本,提高了算法在工程中的反演精度。
附图说明:
图1  具体实施方式流程图;
图2  实测海面风向、风速分布图;
图3  航海雷达图像序列;
图4  海面静态特征图像;
图5  RBF神经网络构架图;
图6  减聚类算法流程图;
图7  两种海面风速反演神经网络训练误差分布图;
图8  两种神经网络提取海面风速信息结果图;
图9  两种神经网络提取海面风速信息误差比对图;
图10 两种神经网络提取海面风速信息误差分布范围。
具体实施方式
下面结合附图对本发明提出的一种基于RBF神经网络的航海雷达图像反演海面风速方法作进一步的详细说明。
本发明具体实施方式流程见图1,分为航海雷达图像预处理、RBF神经网络输入层构建、RBF神经网络模型确定和海面风速信息提取这四大块。具体实施步骤共分为十六步,第一步到第二步为数据预处理;第三步到第六步是RBF神经网络输入层构建;第七步到第十四步为RBF神经网络模型确定;第十五步到第十六步为海面风速信息提取及分析。具体步骤如下:
第一步,采集2010年10月至12月350组导航雷达图像序列,每组图像序列中包含32幅航海雷达图像,经历时间总长度约1.5分钟,同步记录由风力计获得的海面风向θ w、风速U w,实测风向、风速分布如图2所示。
第二步,通过对航海雷达图像预处理抑制同频信号对海面风速反演的影响,对航海雷达图像序列中每幅雷达图像都应用3×3模板的2D非线性平滑中值滤波,滤波后图像灰度值g'(r,θ)为:
Figure PCTCN2021080049-appb-000011
式(1)中f(s,t)为雷达图像极坐标(s,t)处的回波强度值;f'(r,θ)为滤波后在极坐标(r,θ)处的灰度值;N(r,θ)为(r,θ)为中心的像元点,(s,t)取以(r,θ)为中心相邻的8个像元点。
将中值滤波器3×3模板中心与极坐标图像的N(r,θ)中心重合,通过(r,θ)与周围8个相邻像元点(s,t)的回波强度值比较,选取回波强度中间值来更新N(r,θ)处的回波强度值,模板以步长单位1遍历极坐标航 海雷达图像,最终获得中值滤波后的航海雷达图像。
第三步,由于信噪比与海面波高信息成正比,计算每幅图像的信噪比作为海况信息,计算公式如下:
Figure PCTCN2021080049-appb-000012
Figure PCTCN2021080049-appb-000013
为二维波数谱经校正后的海浪谱,
Figure PCTCN2021080049-appb-000014
为雷达图像海浪信号以外的噪声谱,计算公式如下:
Figure PCTCN2021080049-appb-000015
Figure PCTCN2021080049-appb-000016
其中,I (3)(k,ω)是f'(r,θ)三维图像谱,F (2)(k)是二维海浪图像谱,N kx、N ky、N ω为谱的范围,Δk x、Δk y为波数分辨率,Δω为频率分辨率,
Figure PCTCN2021080049-appb-000017
表示色散关系带通滤波器。
对获得每幅雷达图像信噪比r t以雷达图像时间序列进行归一化,得到海况信息归一化值R i
Figure PCTCN2021080049-appb-000018
第四步,为了获得海面风场能量谱均值S i,先要从航海雷达图像序列中提取包含风条纹特征的海面静态特征图像,滤波后的航海雷达图像序列如图3所示。本发明通过构建全局低通滤波器来实现,全局低通滤波器的构建如下:
Figure PCTCN2021080049-appb-000019
其中,f'(θ,r,t)为中值滤波后的航海雷达图像序列,f(θ,r)为极坐标海面静态特征图像,N t为导航图像序列中包含导航雷达图像的个数,N t=32。海面静态特征图像如图4所示,对f(θ,r)进行最近点插值处理 得到笛卡尔坐标下图像F(k x,k y)。
对海面风场能量谱F(k x,k y)沿x和y轴归一化,得到能量谱均值S i
Figure PCTCN2021080049-appb-000020
第五步,对雷达同位置的现场风力计获得的海面风向信息d i、海面风速信息p i,按雷达图像时间序列进行归一化,得到海面风向、风速信息归一化值D i及P i
Figure PCTCN2021080049-appb-000021
第六步,对第三、四、五步获得的350组S i、R i、D i和P i的数据分成两个部分,175组数据用于RBF神经网络模型的训练得到海面风速经验模型,剩余的175组数据用于模型的测试。
第七步,依据海面风场海面风向、风速与雷达回波图像的关联性,如下公式所示:
σ 0=A(θ)u γ(θ)(1+B(u,θ)cosΦ+C(u,θ)cos2Φ)    (9)
其中, σ 0为归一化雷达截面积,u为局部近海表面风速,Φ为电磁波与风向夹角,θ为雷达入射角,A,B,C和γ是经验参数,由雷达频率和极化方式决定。从公式可以看出雷达NRCS与海面风速成指数函数关系,与电磁波和风向的夹角成谐波函数关系。
而Wilson提出的深水充分成长风浪风速和有效波高的关系如式所示:
Figure PCTCN2021080049-appb-000022
式中,g表示重力加速度,H sw表示充分成长风浪有效波高,U表示风速。有效波高可以充分反映海况的情况,上式说明海面风场与海况信息具有紧密的相关性。因此,本方法设计的RBF神经网络第一层输入层为信号源节点,所述的信号源节点由海面风场能量谱、实测风向、实测风速和海况信息组成,即:
x i=[S i D i R i P i],(i=1,2,L,n)    (11)
本发明建立RBF神经框架如图5所示。框架中第二层隐含层的函数是非线性基函数,非线性基函数的中心就是指实测风速,根据非线性基函数的特点可知,当输入层的数据越靠近实测风速,即越靠近中心时,隐含层对应的数据就会越大。第三层是输出层,它随着输入模式的作用而响应。
第八步,RBF神经网络模型函数选取。输入层到隐含层是一种非线性的映射关系,它按照输入值与中心点的欧式距离乘以一个常数来计算的,随着中心点的确定,权重也随之确定。而隐含层到输出层的连接方式则采用了线性加权。其中,非线性基函数一般选择高斯基函数,公式如下:
Figure PCTCN2021080049-appb-000023
其中,x i为第一层的信号源节点,b j是指径向基函数的中心,即实测风速的中心向量,σ i是第i个隐层单元的扩展常数。RBF神经网络隐含层到输出层的权值向量为:
w=[w 1,w 2,w 3,L,w p]     (13)因此,可以得出RBF神经网络的输出G m(x)为:
Figure PCTCN2021080049-appb-000024
RBF网络训练过程分为两步,第一步是确定隐层单元数和基函数的中心及扩展常数;第二步是确定隐层到输出层的连接权值w i
第九步,采用减聚类算法确定设计网络中隐层单元数、扩展常数,减聚类算法流程如图6所示,步骤如下:
①先计算每个样本x i数据点的密度指标,确定首个数据点密度指标的聚类中心;②由首个聚类中心计算更新数据点密度指标;
③计算首个聚类中心和更新后的密度指标计算两者比值最大值,判定是否满足聚类判定条件,由此得到最终的聚类中心为数据节点中心,并得到扩展常数。
第十步,计算每个输入样本x i的密集度指标,计算公式如下:
Figure PCTCN2021080049-appb-000025
I i为输入样本x i密度值指标,λ a为密度指标领域,该领域外的数据点对I i的影响很小或者几乎没有,可以忽略,这里λ a定义为一个常数,通过实验选取λ a=18。也就是说I i越高,则在该密度领域λ a的数据点必然较多。
通过公式(15)得到的最大I i的点最为第一个聚类中心,并将其对应的数据点标记为x c1,相对应的密度指标标记为I c1
第十一步,在获得第一个聚类中心后,更新各数据点密度指标,计算公式如下:
Figure PCTCN2021080049-appb-000026
其中,λ b与λ a定义相同,但通常大于λ a,λ b=27,这里得到的λ b为作为RBF神经网络的扩展常数。
第十二步,根据获得的第一个聚类中心和更新的数据点密度指标,得到聚类判定条件,即:
Figure PCTCN2021080049-appb-000027
其中,γ为很小的数值,这里取γ=0.01。随着不断更新数据点密度指标,不断的确定聚类中心I cs,直到满足聚类条件停止聚类判定,获得的聚类中心为数据节点中心。
第十三步,确定RBF神经网络中输出层连接权值。应用训练样本数据S i、R i、D i作为数据值,P i为观测值,利用递推最小二乘获得网络连接权值w i,即:
w(i)=w(i-1)+K(i)e p *(i)    (18)
其中,e p(i)为先验误差,K(i)为卡尔曼增益,两者均由数据值和观测值递推得到。
第十四步,对训练的175组数据分别应用本发明设计的RBF神经网络和传统的BP神经网络进行训练,得到两种神经网络反演海面风速训练误差分布图如图7所示,(a)为RBF神经网络误差分布图,(b)为传统的BP神经网络误差分布图。从图中可以看出传统BP神经网络 在1000次误差达到稳定,收敛误差为0.3m/s,收敛时间为19.4s;RBF神经网络在训练300次左右的时候误差达到稳定,收敛误差为10 -3m/s,应用MATLAB仿真收敛时间为0.8s,反演精度和收敛速度方面更能达到工程应用要求。
第十五步,对测试航海雷达图像序列得到的海面风场能量谱平均值,海面风向信息和海况信息进行归一化映射,输入到传统BP神经网络和RBF神经网络模型中,两种神经网络提取海面风速信息结果如图8所示。从图8中可以得到RBF神经网络反演的风速结果虽然随着实际风速的趋势略有波动,但与风速的真实值是更接近的。而BP神经网络反演得到的海面风速信息与实测风速波动范围较大,尤其在风速15~20m/s时波动达到5m/s以上。
第十六步,通过实验结果计算得到两种神经网络反演风速结果与实测风速统计结果如表1所示,本发明风速反演结果与实测风向相关系数达到0.87,标准差2.1m/s,偏差-0.08,完全达到工程要求。
表1海面风速误差统计
Figure PCTCN2021080049-appb-000028
将BP神经网络与RBF神经网络反演结果与真实值的误差进行对比,其结果如图9所示,误差范围如图10所示。由图9可以看出RBF神经网络的误差范围大体在-5~+5m/s之间,而BP神经网络的误差范围在-10~+10m/s之间,说明RBF神经网络反演结果精度更高。从图10可以看出两种算法海面风速反演结果的误差分布范围,可以看出 RBF神经网络的误差范围更小,64%数据误差主要集中在-2~0m/s和0~2m/s;而BP神经网络误差范围更加分布散,只有35%数据误差集中在-2~0m/s和0~2m/s。可以得出,RBF神经网络比BP神经网络拟合的结果更加稳定。

Claims (4)

  1. 基于RBF神经网络的航海雷达图像反演海面风速方法,其特征在于,包括如下步骤:
    步骤1,航海雷达图像数据预处理,采集航海雷达图像序列,航海雷达图像序列中包含N幅雷达图像,同时采集对应时间和位置同步传感器获得的实际风向、风速及海况信息,对每幅雷达图像进行中值滤波,抑制同频、雨雪噪声对图像的影响;
    步骤2,RBF神经网络输入层构建,对航雷达图像序列进行全局低通滤波得到海面静态特征图像,选取风条纹特征明显图像,利用风条纹尺度带通滤波器获得海面风场能量谱,结合传感器及图像信息构建RBF神经网络输入层样本;
    步骤3,RBF神经网络模型确定,对获取的海面风场能量谱平均值、海面风向、风速信息及海况信息等样本数据进行归一化映射,建立RBF神经网络框架,利用减聚类算法对样本目标矩阵选取中心,由聚类密度指数终止反向判定,得到基函数参数和扩展常数,利用递推最小二乘法获得输出层连接权值,最终确定RBF神经网络模型。
    步骤4,海面风速信息提取,对测试航海雷达图像序列得到的海面风场能量谱平均值,海面风向信息和海况信息进行归一化映射,输入到RBF神经网络模型中,得到海面风速信息。
  2. 根据权利要求1所述的基于RBF神经网络的航海雷达图像反演海面风速方法,其特征在于:
    所述步骤3包括以下步骤:
    步骤3.1,海面风场能量谱平均值、海面风向、风速信息及海况信息进行归一化映射;
    ①对获得的海面风场能量谱I(k x,k y)沿x和y轴归一化,得到能量谱均值S i
    Figure PCTCN2021080049-appb-100001
    式(1)中,i是对应的雷达图像序列数;
    ②获得每幅雷达图像信噪比r t,以雷达图像时间序列进行归一化,得到海况信息归一化值R i
    Figure PCTCN2021080049-appb-100002
    式(2)中
    Figure PCTCN2021080049-appb-100003
    为二维波数谱经校正后的海浪谱,
    Figure PCTCN2021080049-appb-100004
    为雷达图像海浪信号以外的噪声谱;
    ③对风力计获得的海面风向信息d i、海面风速信息p i,按雷达图像时间序列进行归一化,得到海面风向、风速信息归一化值D i及P i
    Figure PCTCN2021080049-appb-100005
    步骤3.2,建立RBF神经网络框架;
    ①对步骤3.1获得的S i、R i、D i和P i的所有数据分成两个部分,一部分用于RBF神经网络模型的训练得到海面风速经验模型,另一部分用于模型的测试;
    ②依据海面风场与相关信息的关联性,建立RBF神经框架,
    该RBF神经网络第一层输入层为信号源节点,第二层隐含层的函数是非线性基函数,非线性基函数的中心就是指实测风速,第三层是输出层,它随着输入模式的作用而响应;
    ③RBF输入层到隐含层是一种非线性的映射关系,它按照输入值与中心点的欧式距离乘以一个常数来计算,随着中心点的确定,权重也随之确定;而隐含层到输出层的连接方式则采用了线性加权;其中,非线性基函数为高斯基函数,公式如下:
    Figure PCTCN2021080049-appb-100006
    式(4)中,x i为第一层的信号源节点,b j是指径向基函数的中心,即实测风速的中心向量,σ i是第i个隐层单元的扩展常数;
    RBF神经网络隐含层到输出层的权值向量为:
    w=[w 1,w 2,w 3,L,w p]  (5)
    可以得出RBF神经网络的输出G m(x)为:
    Figure PCTCN2021080049-appb-100007
    RBF网络训练过程分为两步,第一步是确定隐层单元数和基函数的中心及扩展常数;第二步是确定隐层到输出层的连接权值w i
    步骤3.3,数据点密度指标计算;
    该方法采用减聚类算法确定设计网络中隐层单元数、扩展常数,利用递归最小二乘算法确定w i;为了确定聚类中心,首先要计算输入样本的密集度指标,计算公式如下:
    Figure PCTCN2021080049-appb-100008
    I i为输入样本x i密度值指标,λ a为密度指标领域,该领域外的数据点对I i的影响可以忽略,λ a定义为一个常数;通过公式(7)得到的最大I i的点最为第一个聚类中心,并将其对应的数据点标记为x c1,相对应的密度指标标记为I c1,由此,更新各数据点密度指标公式如下:
    Figure PCTCN2021080049-appb-100009
    式(8)中,λ b与λ a定义相同,但通常大于λ a
    步骤3.4,减聚类算法得到节点中心和扩展常数;
    随着不断更新数据点密度指标,不断的确定聚类中心I cs,直到满足聚类判定条件,即:
    Figure PCTCN2021080049-appb-100010
    其中,γ为很小的数值;
    最终,获得的聚类中心为数据节点中心,得到的λ b为扩展常数;
    步骤3.5,RBF神经网络中输出层连接权值的确定;
    应用训练样本数据S i、R i、D i作为数据值,P i为观测值,利用递推最小二乘获得网络连接权值w i,即:
    w(i)=w(i-1)+K(i)e p *(i)  (10)
    式(10)中,e p(i)为先验误差,K(i)为卡尔曼增益,两者均由数据值和观测值递推得到。
  3. 根据权利要求2所述的基于RBF神经网络的航海雷达图像反演海面风速方法,步骤3.2所述的RBF神经网络第一层为输入层信号源节点,其特征在于,所述的信号源节点由海面风场能量谱、实测风向、实测风速和海况信息组成,即:
    x i=[S i D i R i P i],(i=1,2,L,n)  (11)。
  4. 根据权利要求2所述的基于RBF神经网络的航海雷达图像反演海面风速方法,步骤3.4所述数据点密度指标计算,其特征在于,所述的密度指标领域λ a为18,λ b为27。
PCT/CN2021/080049 2020-04-30 2021-03-10 基于rbf神经网络的航海雷达图像反演海面风速方法 WO2021218424A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010360487.X 2020-04-30
CN202010360487.XA CN111583214B (zh) 2020-04-30 2020-04-30 基于rbf神经网络的航海雷达图像反演海面风速方法

Publications (1)

Publication Number Publication Date
WO2021218424A1 true WO2021218424A1 (zh) 2021-11-04

Family

ID=72114454

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/080049 WO2021218424A1 (zh) 2020-04-30 2021-03-10 基于rbf神经网络的航海雷达图像反演海面风速方法

Country Status (2)

Country Link
CN (1) CN111583214B (zh)
WO (1) WO2021218424A1 (zh)

Cited By (31)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113837185A (zh) * 2021-11-23 2021-12-24 深圳比特微电子科技有限公司 目标检测方法、系统、装置及存储介质
CN114019803A (zh) * 2021-11-09 2022-02-08 青岛杰瑞工控技术有限公司 一种基于环境载荷的拖轮靠离泊动力定位控制算法
CN114065639A (zh) * 2021-11-19 2022-02-18 江苏科技大学 一种挖泥船施工参数的闭环实时反演方法
CN114167076A (zh) * 2021-11-10 2022-03-11 中国地质大学(武汉) 海面风速风向反演模型的建立方法及风速风向的反演方法
CN114199945A (zh) * 2021-12-29 2022-03-18 郑州大学 一种地聚合物砂浆强度及抗氯离子渗透性预测方法
CN114265087A (zh) * 2021-12-24 2022-04-01 青岛镭测创芯科技有限公司 一种海浪参数反演方法及相关组件
CN114398493A (zh) * 2021-12-29 2022-04-26 中国人民解放军92728部队 一种基于模糊聚类与费效价值的无人机型谱构建方法
CN114707248A (zh) * 2022-04-14 2022-07-05 江苏恒力制动器制造有限公司 基于神经网络的ehb压力控制方法
CN115056824A (zh) * 2022-05-06 2022-09-16 北京和利时系统集成有限公司 一种确定控车参数的方法、装置、计算机存储介质及终端
CN115062547A (zh) * 2022-06-24 2022-09-16 复旦大学 风驱海面和舰船尾迹流体微元轨道速度计算方法及系统
CN115356702A (zh) * 2022-08-18 2022-11-18 中国电波传播研究所(中国电子科技集团公司第二十二研究所) 一种基于海上多源雷达回波的蒸发波导深度学习联合反演方法
CN115408301A (zh) * 2022-10-31 2022-11-29 北京千尧新能源科技开发有限公司 一种用于风机仿真的测试集构建方法及系统
CN115438571A (zh) * 2022-08-04 2022-12-06 南方海洋科学与工程广东省实验室(珠海) 一种基于机器学习的地波雷达波浪场计算方法及装置
CN115618732A (zh) * 2022-10-24 2023-01-17 上海交通大学 核反应堆数字孪生关键参数自主优化数据反演方法
CN115711612A (zh) * 2022-11-02 2023-02-24 中国人民解放军国防科技大学 一种波浪有效波高的预测方法
CN115932838A (zh) * 2022-12-12 2023-04-07 中山大学 一种基于神经网络的地波雷达与走航观测的数据校正方法
CN115983141A (zh) * 2023-03-21 2023-04-18 国家海洋局北海预报中心((国家海洋局青岛海洋预报台)(国家海洋局青岛海洋环境监测中心站)) 一种基于深度学习反演海浪波高的方法、介质及系统
CN116068595A (zh) * 2023-04-06 2023-05-05 极诺星空(北京)科技有限公司 海面风速反演方法、装置、电子设备及介质
WO2023103587A1 (zh) * 2021-12-09 2023-06-15 南京邮电大学 短临降水预测方法及装置
CN116629026A (zh) * 2023-07-18 2023-08-22 中国海洋大学 基于bp神经网络的涡核水下最大温度异常反演方法
CN116644658A (zh) * 2023-05-11 2023-08-25 华中科技大学 一种基于物理信息神经网络的气体放电温度反演方法
CN116680994A (zh) * 2023-08-03 2023-09-01 南京信息工程大学 一种基于激光雷达的气溶胶追踪和风场反演方法及系统
CN116822381A (zh) * 2023-08-30 2023-09-29 中国海洋大学 一种基于人工智能的海洋温盐结构反演方法
CN116976091A (zh) * 2023-06-29 2023-10-31 中国空间技术研究院 适用于微波遥感卫星的海浪环境因素影响分析方法
CN116992193A (zh) * 2023-09-28 2023-11-03 中国科学院海洋研究所 一种由海浪信息计算风速的方法、存储介质及设备
CN117115377A (zh) * 2023-10-18 2023-11-24 云南滇能智慧能源有限公司 一种风电场的能源模型创建方法、装置、设备及存储介质
CN117171934A (zh) * 2023-11-03 2023-12-05 成都大学 一种基于pod-anns的架空输电线路舞动响应预测方法
CN117192555A (zh) * 2023-11-02 2023-12-08 中国科学院空天信息创新研究院 Sar图像海面大气锋检测及其冷暖类型判定方法与系统
CN117236448A (zh) * 2023-11-15 2023-12-15 中国人民解放军空军预警学院 一种基于时序知识图谱的雷达意图推理及其模型训练方法
CN117272843A (zh) * 2023-11-22 2023-12-22 中国石油大学(华东) 基于随机森林的gnss-r海面风速反演方法及系统
CN117744893A (zh) * 2024-02-19 2024-03-22 西安热工研究院有限公司 一种储能辅助黑启动的风速预测方法和系统

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111583214B (zh) * 2020-04-30 2023-06-30 江苏科技大学 基于rbf神经网络的航海雷达图像反演海面风速方法
CN113156439B (zh) * 2021-05-08 2022-08-30 中国石油大学(华东) 一种基于数据驱动的sar风场海浪联合反演方法及系统
CN113534085B (zh) * 2021-06-28 2022-03-04 中国科学院国家空间科学中心 一种干涉成像高度计的海面风速和有效波高联合反演方法
CN113534159B (zh) * 2021-07-08 2022-04-26 国家海洋技术中心 一种基于sar卫星遥感数据的海面风速反演方法及系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110172920A1 (en) * 2007-01-12 2011-07-14 Young Paul Yee Process for generating spatially continuous wind profiles from wind profiler measurements
CN102681033A (zh) * 2012-04-27 2012-09-19 哈尔滨工程大学 一种基于x波段航海雷达的海面风场测量方法
CN109636225A (zh) * 2018-12-21 2019-04-16 北京航天泰坦科技股份有限公司 一种智慧海洋大数据监测系统
CN111583214A (zh) * 2020-04-30 2020-08-25 江苏科技大学 基于rbf神经网络的航海雷达图像反演海面风速方法

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE10035921B4 (de) * 2000-07-21 2004-02-05 Gkss Forschungszentrum Verfahren zur Ermittlung von ein in situ Seegangsfeld beschreibenden hydropraphischen Parametern mittels einer Radareinrichtung
CN103941257B (zh) * 2014-04-11 2016-05-04 哈尔滨工程大学 一种基于波数能量谱的导航雷达图像反演海面风向的方法
CN104297753B (zh) * 2014-10-20 2017-03-01 哈尔滨工程大学 一种基于自适应缩减算子的导航雷达图像反演海面风向方法
CN110398738B (zh) * 2019-06-09 2021-08-10 自然资源部第二海洋研究所 一种利用遥感图像反演海面风速的方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110172920A1 (en) * 2007-01-12 2011-07-14 Young Paul Yee Process for generating spatially continuous wind profiles from wind profiler measurements
CN102681033A (zh) * 2012-04-27 2012-09-19 哈尔滨工程大学 一种基于x波段航海雷达的海面风场测量方法
CN109636225A (zh) * 2018-12-21 2019-04-16 北京航天泰坦科技股份有限公司 一种智慧海洋大数据监测系统
CN111583214A (zh) * 2020-04-30 2020-08-25 江苏科技大学 基于rbf神经网络的航海雷达图像反演海面风速方法

Cited By (50)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114019803B (zh) * 2021-11-09 2023-08-15 青岛杰瑞工控技术有限公司 一种基于环境载荷的拖轮靠离泊动力定位控制算法
CN114019803A (zh) * 2021-11-09 2022-02-08 青岛杰瑞工控技术有限公司 一种基于环境载荷的拖轮靠离泊动力定位控制算法
CN114167076A (zh) * 2021-11-10 2022-03-11 中国地质大学(武汉) 海面风速风向反演模型的建立方法及风速风向的反演方法
CN114065639A (zh) * 2021-11-19 2022-02-18 江苏科技大学 一种挖泥船施工参数的闭环实时反演方法
CN114065639B (zh) * 2021-11-19 2024-05-31 江苏科技大学 一种挖泥船施工参数的闭环实时反演方法
CN113837185A (zh) * 2021-11-23 2021-12-24 深圳比特微电子科技有限公司 目标检测方法、系统、装置及存储介质
WO2023103587A1 (zh) * 2021-12-09 2023-06-15 南京邮电大学 短临降水预测方法及装置
CN114265087A (zh) * 2021-12-24 2022-04-01 青岛镭测创芯科技有限公司 一种海浪参数反演方法及相关组件
CN114199945A (zh) * 2021-12-29 2022-03-18 郑州大学 一种地聚合物砂浆强度及抗氯离子渗透性预测方法
CN114398493A (zh) * 2021-12-29 2022-04-26 中国人民解放军92728部队 一种基于模糊聚类与费效价值的无人机型谱构建方法
CN114199945B (zh) * 2021-12-29 2023-09-19 郑州大学 一种地聚合物砂浆强度及抗氯离子渗透性预测方法
CN114707248A (zh) * 2022-04-14 2022-07-05 江苏恒力制动器制造有限公司 基于神经网络的ehb压力控制方法
CN114707248B (zh) * 2022-04-14 2024-04-26 江苏恒力制动器制造有限公司 基于神经网络的ehb压力控制方法
CN115056824A (zh) * 2022-05-06 2022-09-16 北京和利时系统集成有限公司 一种确定控车参数的方法、装置、计算机存储介质及终端
CN115056824B (zh) * 2022-05-06 2023-11-28 北京和利时系统集成有限公司 一种确定控车参数的方法、装置、计算机存储介质及终端
CN115062547A (zh) * 2022-06-24 2022-09-16 复旦大学 风驱海面和舰船尾迹流体微元轨道速度计算方法及系统
CN115438571A (zh) * 2022-08-04 2022-12-06 南方海洋科学与工程广东省实验室(珠海) 一种基于机器学习的地波雷达波浪场计算方法及装置
CN115356702A (zh) * 2022-08-18 2022-11-18 中国电波传播研究所(中国电子科技集团公司第二十二研究所) 一种基于海上多源雷达回波的蒸发波导深度学习联合反演方法
CN115618732A (zh) * 2022-10-24 2023-01-17 上海交通大学 核反应堆数字孪生关键参数自主优化数据反演方法
CN115408301B (zh) * 2022-10-31 2023-01-31 北京千尧新能源科技开发有限公司 一种用于风机仿真的测试集构建方法及系统
CN115408301A (zh) * 2022-10-31 2022-11-29 北京千尧新能源科技开发有限公司 一种用于风机仿真的测试集构建方法及系统
CN115711612B (zh) * 2022-11-02 2024-06-11 中国人民解放军国防科技大学 一种波浪有效波高的预测方法
CN115711612A (zh) * 2022-11-02 2023-02-24 中国人民解放军国防科技大学 一种波浪有效波高的预测方法
CN115932838A (zh) * 2022-12-12 2023-04-07 中山大学 一种基于神经网络的地波雷达与走航观测的数据校正方法
CN115932838B (zh) * 2022-12-12 2023-11-21 中山大学 一种基于神经网络的地波雷达与走航观测的数据校正方法
CN115983141A (zh) * 2023-03-21 2023-04-18 国家海洋局北海预报中心((国家海洋局青岛海洋预报台)(国家海洋局青岛海洋环境监测中心站)) 一种基于深度学习反演海浪波高的方法、介质及系统
CN116068595A (zh) * 2023-04-06 2023-05-05 极诺星空(北京)科技有限公司 海面风速反演方法、装置、电子设备及介质
CN116644658A (zh) * 2023-05-11 2023-08-25 华中科技大学 一种基于物理信息神经网络的气体放电温度反演方法
CN116644658B (zh) * 2023-05-11 2024-03-15 华中科技大学 一种基于物理信息神经网络的气体放电温度反演方法
CN116976091A (zh) * 2023-06-29 2023-10-31 中国空间技术研究院 适用于微波遥感卫星的海浪环境因素影响分析方法
CN116976091B (zh) * 2023-06-29 2024-03-01 中国空间技术研究院 适用于微波遥感卫星的海浪环境因素影响分析方法
CN116629026B (zh) * 2023-07-18 2023-09-26 中国海洋大学 基于bp神经网络的涡核水下最大温度异常反演方法
CN116629026A (zh) * 2023-07-18 2023-08-22 中国海洋大学 基于bp神经网络的涡核水下最大温度异常反演方法
CN116680994A (zh) * 2023-08-03 2023-09-01 南京信息工程大学 一种基于激光雷达的气溶胶追踪和风场反演方法及系统
CN116680994B (zh) * 2023-08-03 2023-09-29 南京信息工程大学 一种基于激光雷达的气溶胶追踪和风场反演方法及系统
CN116822381B (zh) * 2023-08-30 2023-11-21 中国海洋大学 一种基于人工智能的海洋温盐结构反演方法
CN116822381A (zh) * 2023-08-30 2023-09-29 中国海洋大学 一种基于人工智能的海洋温盐结构反演方法
CN116992193A (zh) * 2023-09-28 2023-11-03 中国科学院海洋研究所 一种由海浪信息计算风速的方法、存储介质及设备
CN116992193B (zh) * 2023-09-28 2023-12-19 中国科学院海洋研究所 一种由海浪信息计算风速的方法、存储介质及设备
CN117115377A (zh) * 2023-10-18 2023-11-24 云南滇能智慧能源有限公司 一种风电场的能源模型创建方法、装置、设备及存储介质
CN117192555A (zh) * 2023-11-02 2023-12-08 中国科学院空天信息创新研究院 Sar图像海面大气锋检测及其冷暖类型判定方法与系统
CN117192555B (zh) * 2023-11-02 2024-01-12 中国科学院空天信息创新研究院 Sar图像海面大气锋检测及其冷暖类型判定方法与系统
CN117171934B (zh) * 2023-11-03 2024-01-26 成都大学 一种基于pod-anns的架空输电线路舞动响应预测方法
CN117171934A (zh) * 2023-11-03 2023-12-05 成都大学 一种基于pod-anns的架空输电线路舞动响应预测方法
CN117236448B (zh) * 2023-11-15 2024-02-09 中国人民解放军空军预警学院 一种基于时序知识图谱的雷达意图推理及其模型训练方法
CN117236448A (zh) * 2023-11-15 2023-12-15 中国人民解放军空军预警学院 一种基于时序知识图谱的雷达意图推理及其模型训练方法
CN117272843B (zh) * 2023-11-22 2024-02-02 中国石油大学(华东) 基于随机森林的gnss-r海面风速反演方法及系统
CN117272843A (zh) * 2023-11-22 2023-12-22 中国石油大学(华东) 基于随机森林的gnss-r海面风速反演方法及系统
CN117744893A (zh) * 2024-02-19 2024-03-22 西安热工研究院有限公司 一种储能辅助黑启动的风速预测方法和系统
CN117744893B (zh) * 2024-02-19 2024-05-17 西安热工研究院有限公司 一种储能辅助黑启动的风速预测方法和系统

Also Published As

Publication number Publication date
CN111583214B (zh) 2023-06-30
CN111583214A (zh) 2020-08-25

Similar Documents

Publication Publication Date Title
WO2021218424A1 (zh) 基于rbf神经网络的航海雷达图像反演海面风速方法
WO2022016884A1 (zh) 一种基于K-means聚类算法的海面风速方法
CN109886217B (zh) 一种基于卷积神经网络从近岸海浪视频中检测浪高的方法
CN112711899B (zh) 一种蒸发波导高度的融合预测方法
CN110333554B (zh) Nriet暴雨智能相似分析方法
CN102628944B (zh) 一种基于多普勒雷达资料的层云与对流云自动识别方法
CN110261857B (zh) 一种天气雷达空间插值方法
WO2020207070A1 (zh) 深圳海域水质评价方法及系统
CN111680870B (zh) 目标运动轨迹质量综合评估方法
CN113808174B (zh) 基于全卷积网络和卡尔曼滤波的雷达小目标跟踪方法
CN111428676A (zh) 一种基于稀疏对应和深度神经网络的短临降水预测方法
CN109100723A (zh) 基于多普勒天气雷达数据的高空风反演方法
CN110750516B (zh) 基于雷达图的降雨分析模型构建方法、构建系统及分析方法
CN105891833A (zh) 基于多普勒雷达信息识别暖云降水率的方法
CN108254750B (zh) 一种基于雷达资料的下击暴流智能识别预警方法
Huang et al. Wave height estimation from X-band nautical radar images using temporal convolutional network
CN108663727B (zh) 利用蒸发率在世界海域范围内估算蒸发波导高度的方法
CN115166750A (zh) 一种基于双偏振多普勒雷达数据的定量降水估计方法
Ran et al. Precipitation cloud identification based on faster-RCNN for Doppler weather radar
Beaucage et al. Synthetic aperture radar satellite data for offshore wind assessment: A strategic sampling approach
CN112731564B (zh) 一种基于多普勒天气雷达数据的雷电智能预报方法
CN108596380B (zh) 一种海面台风风场的定量探测方法
CN104297753B (zh) 一种基于自适应缩减算子的导航雷达图像反演海面风向方法
CN115097451A (zh) 一种基于sar卫星数据的海浪多参数反演方法及系统
CN115146564A (zh) 基于垂直分层降尺度技术的城市地面风速精细化模拟方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21795506

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21795506

Country of ref document: EP

Kind code of ref document: A1