CN111308471A - Rain, snow and hail classification monitoring method based on semi-supervised domain adaptation - Google Patents

Rain, snow and hail classification monitoring method based on semi-supervised domain adaptation Download PDF

Info

Publication number
CN111308471A
CN111308471A CN202010088477.5A CN202010088477A CN111308471A CN 111308471 A CN111308471 A CN 111308471A CN 202010088477 A CN202010088477 A CN 202010088477A CN 111308471 A CN111308471 A CN 111308471A
Authority
CN
China
Prior art keywords
snow
hail
rain
reflectivity
data set
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.)
Granted
Application number
CN202010088477.5A
Other languages
Chinese (zh)
Other versions
CN111308471B (en
Inventor
杨涛
陈志远
郑鑫
师鹏飞
秦友伟
李振亚
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.)
Hohai University HHU
Original Assignee
Hohai University HHU
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 Hohai University HHU filed Critical Hohai University HHU
Priority to CN202010088477.5A priority Critical patent/CN111308471B/en
Publication of CN111308471A publication Critical patent/CN111308471A/en
Application granted granted Critical
Publication of CN111308471B publication Critical patent/CN111308471B/en
Priority to PCT/CN2020/136089 priority patent/WO2021159844A1/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO 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
    • G01S13/00—Systems 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
    • G01S13/88—Radar or analogous systems specially adapted for specific applications
    • G01S13/95—Radar or analogous systems specially adapted for specific applications for meteorological use
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
    • 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
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Abstract

The invention discloses a rain, snow and hail classification monitoring method based on semi-supervised domain adaptation, which comprises the steps of obtaining preprocessing data under various types of precipitation particle weather in radar wave reflectivity under various types of precipitation particle weather, constructing a first data set carrying a label and a second data set not carrying the label, calculating a first covariance matrix and a second covariance matrix, determining a first characteristic subspace, determining a second characteristic subspace, determining a kernel function according to the first characteristic subspace and the second characteristic subspace, training an initial classifier by taking the first data set as a training sample set according to the kernel function, carrying out unsupervised learning on the initial classifier, enabling the initial classifier to be in a self-adaptive target field, obtaining and determining an adjacent map, optimizing a target function to determine a final classifier, and classifying rain, snow and hail by adopting the final classifier, the rain, snow and hail can be accurately classified, and the corresponding classification monitoring scheme is higher in accuracy.

Description

Rain, snow and hail classification monitoring method based on semi-supervised domain adaptation
Technical Field
The invention relates to the technical field of ground meteorological detection, in particular to a rain, snow and hail classification monitoring method based on semi-supervised domain adaptation.
Background
For areas or countries with concentrated rainy seasons and frequent rainstorms, abnormal space-time distribution of rainfall is an important factor for causing natural disasters such as flood disasters, landslides, debris flows and the like, and research on rainfall is beyond a scientific research range. When measuring and researching precipitation, the types of precipitation particles, such as rain, snow, hail and the like, are firstly distinguished. At present, rain, snow and hail are mainly identified according to weather radar volume sweep data and a method of dual-polarization Doppler radar polarization parameters. The method of body scanning by the weather radar is simple, but the resolution is low, and the micro-physical characteristics of precipitation particles are ignored; the dual-polarization Doppler radar has higher resolution than the common weather radar, but is easily interfered, so that the error is larger.
At present, the microwave communication network has wide coverage, high signal quality and basically no blind zone, so the microwave communication network rainfall monitoring and analyzing technology has high popularization and application values in China. The method uses the microwave attenuation characteristics of the microwave link to invert the drop spectrum and the particle shape distribution of rain, snow and hail, has high precision, small monitoring blind area and less cost compared with radar, and is very suitable for identifying special weather conditions such as rain, snow and hail in theory. The existing method for identifying rain, snow and hail types by microwaves needs a large amount of labeled microwave attenuation data by utilizing a traditional machine learning method, but in an actual scene, the amount of the labeled microwave attenuation data of the rainfall type is often insufficient, so that the classification monitoring scheme for rain, snow and hail often has the problem of low accuracy.
Disclosure of Invention
Aiming at the problems, the invention provides a rain, snow and hail classification monitoring method based on semi-supervised domain adaptation.
In order to realize the purpose of the invention, the invention provides a rain, snow and hail classification monitoring method based on semi-supervised domain adaptation, which comprises the following steps:
s10, radar wave reflectivity of the electromagnetic waves in various precipitation particle weather is measured by adopting a radar; wherein the types of precipitation particles include rain, snow and hail;
s20, acquiring preprocessing data under the weather of various types of precipitation particles according to the radar wave reflectivity under the weather of various types of precipitation particles;
s30, constructing a first data set carrying labels and a second data set not carrying labels according to the preprocessed data, calculating a first covariance matrix of the first data set and a second covariance matrix of the second data set, determining a first feature subspace according to the first covariance matrix, and determining a second feature subspace according to the second covariance matrix;
s40, determining a kernel function according to the first characteristic subspace and the second characteristic subspace;
s50, training an initial classifier by taking the first data set as a training sample set according to the kernel function;
s60, selecting a subset from the second data set to perform unsupervised learning on the initial classifier, so that the selected subset can provide incremental knowledge for the initial classifier to adapt to the target field;
s70, obtaining an objective function of the initial classifier after unsupervised learning, determining an adjacency graph according to the first data set and the second data set, optimizing the objective function according to the adjacency graph to determine a final classifier, and classifying rain, snow and hail by using the final classifier.
In one embodiment, the radar wave reflectivity of the electromagnetic wave in various types of precipitation particle weather is measured by adopting a radar, and the radar comprises the following steps:
measuring multiple groups of effective horizontal reflectivity Z of precipitation particles under various precipitation particle weather by adopting dual-polarization radarhAnd vertical reflectivity ZvAccording to each group of effective horizontal reflectivity ZhAnd vertical reflectivity ZvAnd calculating the differential reflectivity, and determining the radar wave reflectivity of each type of precipitation particle according to the differential reflectivity corresponding to each type of precipitation particle.
As one embodiment, the differential reflectivity determination process includes:
Figure BDA0002382893760000021
wherein A represents a differential reflectance, ZhRepresenting the effective horizontal reflectivity, ZvIndicating the vertical reflectivity.
As an embodiment, the acquiring the preprocessed data for each type of precipitation particle weather according to the radar wave reflectivity for each type of precipitation particle weather includes:
s22, calculating path attenuation rates of each microwave link in each polarization direction; wherein the determination formula of the path attenuation rate comprises:
Figure BDA0002382893760000022
in the formula IθRepresents the path attenuation ratio, P, of a microwave link with a polarization direction of thetaθ，1Representing the transmitting-end microwave frequency, P, of a microwave link with a polarization direction of thetaθ，2The microwave frequency of a receiving end of the microwave link with the polarization direction theta is represented, and L represents the length of the microwave link;
s23, calculating the differential attenuation rate of the microwave on the microwave link according to the path attenuation rates in different polarization directions; the determination formula of the differential attenuation rate comprises the following steps:
Figure BDA0002382893760000031
wherein O represents a differential attenuation ratio of microwave, IhRepresents the vertical polarization attenuation ratio, I, of the microwave linkvRepresenting the horizontal polarization attenuation rate of the microwave link;
and S25, respectively executing the steps S22 to S23 for each microwave link under the weather of each type of precipitation particles, acquiring a group of differential attenuation rates corresponding to each type of precipitation particles, and determining the preprocessing data of each type of precipitation particles according to the group of differential attenuation rates corresponding to each type of precipitation particles.
In one embodiment, the first feature subspace and the second feature subspace are both subspaces G with dimensions of n × dn×d；
The determining a kernel function from the first feature subspace and the second feature subspace comprises:
s41, constructing a slave S on the first feature subspace and the second feature subspace1To S2The curve of (d); the parameterized function of the curve includes:
Figure BDA0002382893760000032
in the formula, S1Representing a first feature subspace, S2Representing a second feature subspace, P1Is S1Vertical complement space of, U1And U2Diagonal matrices, U, of dxd and (n-d) xd, respectively1From S1'S2＝U1F (v)' is obtained by SVD decomposition, U2From P1'S2＝-U2E (v)' is obtained by SVD decomposition, S1' represents S1The transposed matrix of (1), F (v) and E (v) are diagonal matrices of order d, the diagonal elements of F (v) are cos (α i), the diagonal elements of E (v) are sin (α i),0<i<d, i ∈ {1,2, … …, d }, α i geometrically denote S1And S2A medium base vector geometric angle;
s43, acquiring a basis function G (v) of a basis representing a subspace at each point, calculating a semi-positive definite matrix G according to the basis function G (v), and setting a kernel function according to the semi-positive definite matrix G; the calculation process of the semi-positive definite matrix G comprises the following steps:
Ω＝[S1U1,-P1U2]，
Figure BDA0002382893760000033
in the formula, Λ1、Λ2And Λ3Respectively, a diagonal matrix of d x d.
As an embodiment, the setting process of the kernel function includes:
K(Xi,Rj)＝Xi'GRj，
in the formula, K (X)i,Rj) Represents XiAnd RjCorresponding kernel function, XiRepresenting the i-th sample vector, R, in the source domain D1jRepresenting the jth sample vector of the target domain D2.
In one embodiment, training the initial classifier according to the kernel function with the first data set as a training sample set comprises:
s51, normalizing the first data set, setting a classification function from an input vector X to a real number as f (X), and searching an initial classifier in a regenerative Hilbert space by adopting popular regularization to obtain a target function of the initial classifier;
and S52, training each type of precipitation particle by adopting an initial classifier.
The rain, snow and hail classification monitoring method based on semi-supervised domain adaptation comprises the steps of measuring radar wave reflectivity of electromagnetic waves in various types of precipitation particle weather by adopting a radar, obtaining preprocessing data in various types of precipitation particle weather according to the radar wave reflectivity in various types of precipitation particle weather, constructing a first data set carrying a label and a second data set not carrying the label according to the preprocessing data, calculating a first covariance matrix of the first data set and a second covariance matrix of the second data set, determining a first characteristic subspace according to the first covariance matrix, determining a second characteristic subspace according to the second covariance matrix, determining a kernel function according to the first characteristic subspace and the second characteristic subspace, training an initial classifier by taking the first data set as a training sample set according to the kernel function, and selecting a subset from the second data set to perform unsupervised learning on the initial classifier, the selected subset can provide incremental knowledge for the initial classifier to adapt to the target field, then an objective function of the initial classifier after unsupervised learning is obtained, an adjacency graph is determined according to the first data set and the second data set, the objective function is optimized according to the adjacency graph to determine a final classifier, the final classifier is adopted to classify rain, snow and hail, accurate classification can be carried out on the rain, snow and hail, and the corresponding classification monitoring scheme is higher in accuracy.
Drawings
FIG. 1 is a schematic flow chart of a rain, snow and hail classification monitoring method based on semi-supervised domain adaptation according to an embodiment;
fig. 2 is a schematic flow chart of a rain, snow and hail classification monitoring method based on semi-supervised domain adaptation according to another embodiment.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.
Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. It is explicitly and implicitly understood by one skilled in the art that the embodiments described herein can be combined with other embodiments.
Referring to fig. 1, fig. 1 is a schematic flow chart of a rain, snow and hail classification monitoring method based on semi-supervised domain adaptation according to an embodiment, including the following steps:
s10, radar wave reflectivity of the electromagnetic waves in various precipitation particle weather is measured by adopting a radar; wherein the types of precipitation particles include rain, snow and hail.
The radar wave reflectivity can be represented by n-dimensional real number vectors, for example, each type of precipitation particle can represent the corresponding radar wave reflectivity by one n-dimensional real number vector.
And S20, acquiring preprocessing data of the precipitation particle weather according to the radar wave reflectivity of the precipitation particle weather.
The preprocessed data may be characterized by n-dimensional real number vectors, for example, each type of precipitation particle may be characterized by an n-dimensional real number vector
The steps can be specifically carried out with preprocessing according to relevant parameters generated in the acquisition process of radar wave reflectivity under various types of precipitation particle weather, so as to obtain corresponding preprocessing data.
S30, constructing a first data set carrying labels and a second data set not carrying labels according to the preprocessed data, calculating a first covariance matrix of the first data set and a second covariance matrix of the second data set, determining a first feature subspace according to the first covariance matrix, and determining a second feature subspace according to the second covariance matrix.
The above steps may perform a spatial transformation on the preprocessed data to obtain a first feature subspace and a second feature subspace.
In one embodiment, the process of spatial transformation may include:
s31, in the n-dimensional real vectors (e.g. radar wave reflectivity and corresponding preprocessed data) obtained in S10 and S20, labels of the types of precipitation particles are marked on vectors such as X1, X2, … …, Xt, etc., to form a first data set D1 ═ { X1, X2, … …, Xt }, which is used as an annotated source domain (first data set), where the label of Xi is Yi, and the specific label value of Yi is different according to the different classifiers. Forming R1, R2, … … and Rs into a second data set D2 ═ { R1, R2, … …, Rs }, which serves as an unlabeled target domain (second data set);
s32, performing PCA principal component analysis on D1 and D2 to obtain a feature subspace, which may further include:
(3-2-1) calculating covariance matrices H1 (first covariance matrix) and H2 (second covariance matrix) of D1 and D2;
(3-2-2) respectively calculating eigenvectors and eigenvalues of H1 and H2, sorting the eigenvalues in descending order, extracting the largest d eigenvalues, and forming an eigen subspace S by taking the corresponding eigenvector as the base of the eigen subspace1(first feature subspace) and S2(second feature subspace) with dimension n × d, if R is takennAll d-dimensional subspaces of Gn×dCalled the Grassmann flow pattern, then S1And S2Are included in the glasmann flow pattern.
S40, determining a kernel function according to the first feature subspace and the second feature subspace.
The steps can carry out weighted resampling on the source domain samples, thereby approximating the distribution of the target domain and realizing the self-adaption of the corresponding samples.
And S50, training an initial classifier by taking the first data set as a training sample set according to the kernel function.
The steps can utilize a manifold rule learning method, train a classifier by taking D1 as a training sample set according to a kernel function, and classify rain, snow and hail according to the training classifier obtained by radar data. The conventional manifold regularization classification method is mainly used for a two-classification method, and the embodiment can be used for classifying various precipitation particle types, so that samples without precipitation, rain, snow and hail are classified into one class in sequence during training, other remaining samples are classified into another class, and the corresponding classifier is constructed by the samples of each class. For example, samples of non-precipitation, rain, snow and hail are classified into one class, and other remaining samples are classified into another class. At the moment, the label Yi of the first classifier which is not precipitated is 1, and the other three types of labels Yi are-1; the label Yi of rainfall in the second classifier is 1, and the other three types of labels Yi are-1; the label Yi of snowfall in the third classifier is 1, and other three types of labels Yi are-1; the label Yi of hail in the fourth classifier is 1, and the other three types of labels Yi are-1. The four classifiers are used for training respectively, and then four training results are obtained. During testing, corresponding radar wave reflection data vectors are tested by the four trained classifiers respectively, and finally, each test has a result f* 1(x)，f* 2(x)，f* 3(x)，f* 4(x) The final result is then the largest of these four values as the classification result.
S60, selecting a subset from the second data set to perform unsupervised learning on the initial classifier, such that the selected subset provides incremental knowledge to the initial classifier to adapt to the target domain.
S70, obtaining an objective function of the initial classifier after unsupervised learning, determining an adjacency graph according to the first data set and the second data set, optimizing the objective function according to the adjacency graph to determine a final classifier, and classifying rain, snow and hail by using the final classifier.
In one example, the above steps may include the following processes:
s71, carrying out normalization processing on the second data set D2, and optimizing an initial classifier, wherein the optimized objective function is as follows:
Figure BDA0002382893760000071
wherein F ═ F (X)1),f(X2),……,f(Xt),f(R1),f(R2),……,f(Rp) Π denotes the laplace transformation, γ, of the data adjacency matrices in D1 and D2BIs a user-defined parameter; xi represents an n-dimensional vector representing radar wave reflection, Yi represents a label vector of Xi, V (Xi, Yi, f) represents a cost function, t represents the number of radar wave reflectivity vectors with labels, p represents the number of vectors in a subset selected from the second data set, | | f | y |HKA norm representing a kernel-function-induced regenerative Hilbert space;
the adjacency graph of data in S72, D1, and D2 is constructed according to the KNN method, wherein the distance between samples is defined as follows:
d(Xi,Rj)＝K(Xi,Ri)+K(Xj,Rj)-2K(Xi,Rj)
wherein d (X)i,Rj) Represents XiAnd RjDistance between, XiRepresenting the i-th sample vector, R, in the source domain D1jRepresents the jth sample vector, X, in the target domain D2jIn the representation source domain (first data set) D1J-th sample vector, RiRepresents the ith sample vector, K (X), in the target domain (second data set) D2i,Ri) Is composed of XiAnd RiKernel function of, K (X)j,Rj) Is composed of XjAnd RjKernel function of, K (X)i,Rj) Is composed of XiAnd RjThe adjacency matrix of the data is constructed by a KNN method according to the formula to obtain an adjacency matrix M, elements of each row in the matrix M are added to be used as diagonal elements of a matrix D, and if other elements of D are 0, pi is equal to D-M;
s73, using pi optimization f obtained in S72*To obtain the final classifier fr *R belongs to {1,2,3,4}, and r is 1, when distinguishing precipitation from non-precipitation, the microwave attenuation data vector measured in real time is processed and finally brought into a classifier for distinguishing precipitation from non-precipitation, f* 1>0 time denotes precipitation, f* 1<When the value is 0, the water is not reduced, the output results of other classifiers are similar to the water, and finally, the classifier result with the largest value is used as the final classification result obtained through microwave attenuation data in the four classifier output values.
The rain, snow and hail classification monitoring method based on semi-supervised domain adaptation comprises the steps of measuring radar wave reflectivity of electromagnetic waves in various types of precipitation particle weather by adopting a radar, obtaining preprocessing data in various types of precipitation particle weather according to the radar wave reflectivity in various types of precipitation particle weather, constructing a first data set carrying a label and a second data set not carrying the label according to the preprocessing data, calculating a first covariance matrix of the first data set and a second covariance matrix of the second data set, determining a first characteristic subspace according to the first covariance matrix, determining a second characteristic subspace according to the second covariance matrix, determining a kernel function according to the first characteristic subspace and the second characteristic subspace, training an initial classifier by taking the first data set as a training sample set according to the kernel function, and selecting a subset from the second data set to perform unsupervised learning on the initial classifier, the selected subset can provide incremental knowledge for the initial classifier to adapt to the target field, then an objective function of the initial classifier after unsupervised learning is obtained, an adjacency graph is determined according to the first data set and the second data set, the objective function is optimized according to the adjacency graph to determine a final classifier, the final classifier is adopted to classify rain, snow and hail, accurate classification can be carried out on the rain, snow and hail, and the corresponding classification monitoring scheme is higher in accuracy.
In one embodiment, the radar wave reflectivity of the electromagnetic wave in various types of precipitation particle weather is measured by adopting a radar, and the radar comprises the following steps:
measuring multiple groups of effective horizontal reflectivity Z of precipitation particles under various precipitation particle weather by adopting dual-polarization radarhAnd vertical reflectivity ZvAccording to each group of effective horizontal reflectivity ZhAnd vertical reflectivity ZvAnd calculating the differential reflectivity, and determining the radar wave reflectivity of each type of precipitation particle according to the differential reflectivity corresponding to each type of precipitation particle.
The radar wave reflectivity may be characterized by an n-dimensional real number vector, for example, the n-dimensional real number vector of a certain type of precipitation particle includes respective differential reflectivities corresponding to the type of precipitation particle.
As one embodiment, the differential reflectivity determination process includes:
Figure BDA0002382893760000081
wherein A represents a differential reflectance, ZhRepresenting the effective horizontal reflectivity, ZvIndicating the vertical reflectivity.
In one example, multiple sets of radar electromagnetic wave reflection data may be acquired at a certain time (assuming that n sets of radar devices are operating in a monitoring area), multiple sets of differential effective reflectivities may be obtained, and one n-dimensional real number vector may be obtained, that is, X ═ a (a ═ X1，A2，……,An)∈RnWherein A isiIs the differential reflectivity obtained by the i-th group of radar devices. At different times (chosen to include rain, snow, hail weather to ensure adequate representativeness), to obtain a plurality of nA vector of real numbers.
As an embodiment, the acquiring the preprocessed data for each type of precipitation particle weather according to the radar wave reflectivity for each type of precipitation particle weather includes:
s22, calculating path attenuation rates of each microwave link in each polarization direction; wherein the determination formula of the path attenuation rate comprises:
Figure BDA0002382893760000091
in the formula IθRepresents the path attenuation ratio, P, of a microwave link with a polarization direction of thetaθ，1Representing the transmitting-end microwave frequency, P, of a microwave link with a polarization direction of thetaθ，2The microwave frequency of a receiving end of the microwave link with the polarization direction theta is represented, and L represents the length of the microwave link;
s23, calculating the differential attenuation rate of the microwave on the microwave link according to the path attenuation rates in different polarization directions; the determination formula of the differential attenuation rate comprises the following steps:
Figure BDA0002382893760000092
wherein O represents a differential attenuation ratio of microwave, IhRepresents the vertical polarization attenuation ratio, I, of the microwave linkvRepresenting the horizontal polarization attenuation rate of the microwave link;
and S25, respectively executing the steps S22 to S23 for each microwave link under the weather of each type of precipitation particles, acquiring a group of differential attenuation rates corresponding to each type of precipitation particles, and determining the preprocessing data of each type of precipitation particles according to the group of differential attenuation rates corresponding to each type of precipitation particles.
Further, before step S22, a method may further include
S21, selecting a dual-polarization microwave link, transmitting a microwave signal at a transmitting end by using a selected frequency (for example, the polarization frequency is 40Hz), attenuating the microwave signal when the microwave signal passes through a water-reducing area in the process of propagation, and finally receiving the attenuation at a receiving endThe power of the transmitting end and the power of the receiving end on the horizontal link and the vertical link are measured and are respectively marked as Ph，a，Ph，b，Pv，aAnd Pv，b。
In one embodiment, the first feature subspace and the second feature subspace are both subspaces G with dimensions of n × dn×d。
The determining a kernel function from the first feature subspace and the second feature subspace comprises:
s41, constructing a slave S on the first feature subspace and the second feature subspace1To S2The curve of (d); the parameterized function of the curve includes:
Figure BDA0002382893760000093
in the formula, S1Representing a first feature subspace, S2Representing a second feature subspace, P1Is S1Vertical complement space of, U1And U2Diagonal matrices, U, of dxd and (n-d) xd, respectively1From S1'S2＝U1F (v)' is obtained by SVD decomposition, U2From P1'S2＝-U2E (v)' is obtained by SVD decomposition, S1' represents S1The transposed matrix of (1), F (v) and E (v) are diagonal matrices of order d, the diagonal elements of F (v) are cos (α i), the diagonal elements of E (v) are sin (α i),0<i<d, i ∈ {1,2, … …, d }, α i geometrically S1And S2A medium base vector geometric angle; at this time, then U1,U2F (v) and E (v) can both be calculated as parameterized functions of the curve
Figure BDA0002382893760000102
And can also be obtained from the above formula.
S43, acquiring a basis function G (v) of a basis representing a subspace at each point, calculating a semi-positive definite matrix G according to the basis function G (v), and setting a kernel function according to the semi-positive definite matrix G; the calculation process of the semi-positive definite matrix G comprises the following steps:
Ω＝[S1U1,-P1U2]，
Figure BDA0002382893760000101
in the formula, Λ1、Λ2And Λ3Respectively, a diagonal matrix of d x d.
As an embodiment, the setting process of the kernel function includes:
K(Xi,Rj)＝Xi'GRj，
in the formula, K (X)i,Rj) Represents XiAnd RjCorresponding kernel function, XiRepresenting the i-th sample vector, R, in the source domain D1jRepresenting the jth sample vector of the target domain D2.
Specifically, after the step S41, the method may further include:
s42, let w (v) ═ 1-2v |, where v ∈ [0,1], in the curve Φ (v), when v is close to 0 or 1, the subspace represented by the corresponding point is more reliable, so a higher weight should be given, so it is obtained by multiplying w (t) by Φ (t):
g(v)＝w(v)φ(v)；
the above g (v) represents the base of a subspace at each point, and the sum of the inner products of the infinite dimension Hilbert space represented by g (v) is that g (v) is integrated in the interval of [0,1], i.e., < g (v) > is g (v)' g (v).
Further, the derivation process of the semi-positive definite matrix G may include:
Figure BDA0002382893760000111
since e (v) and f (v) are diagonal matrices, the diagonal elements are cos (v. α i) and sin (v. α i), respectively, then:
Ω＝[S1U1,-P1U2]，
therefore, the method comprises the following steps:
Figure BDA0002382893760000112
in the formula, Λ1、Λ2And Λ3Diagonal matrices of dxd, Λ, respectively1Has a diagonal element of λ1i,Λ2Has a diagonal element of λ2i,Λ3Has a diagonal element of λ3iWhere i ∈ {1,2, … …, d }, λ1i,λ2i,λ3iThe expression of (a) is:
Figure BDA0002382893760000113
Figure BDA0002382893760000114
Figure BDA0002382893760000115
and G is a semi-positive definite matrix because it is a kernel matrix on which kernel functions can be defined as follows
K(Xi,Rj)＝Xi'GRj。
In one embodiment, training the initial classifier according to the kernel function with the first data set as a training sample set comprises:
s51, normalizing the first data set, setting a classification function from an input vector X to a real number as f (X), and searching an initial classifier in a regenerative Hilbert space by adopting popular regularization to obtain a target function of the initial classifier;
and S52, training each type of precipitation particle by adopting an initial classifier.
Specifically, in step S51, the D1 is normalized, and the classification function is f (X), which is a function from the input vector X to a real number, so that the popularity of regularization searches for a classifier in the regenerated hilbert space, and the objective function is:
Figure BDA0002382893760000116
where V (Xi, Yi, f) is a cost function, the present trainer uses a Hinge function, namely:
V(Xi,Yi,f)＝max(0,1-Yif(Xi))，
in the formula (I), the compound is shown in the specification,
Figure BDA0002382893760000121
is the norm, gamma, of the corresponding kernel-function-induced regenerated Hilbert spaceAIs a user-defined parameter, YiA label representing Xi.
In the above step S52, the classifier f obtained in S51 can be usedr *At f established using radar data*On the basis, the microwave data and radar data field self-adaption can be used for classifying precipitation types, r belongs to {1,2,3 and 4}, and r is set to be 1, when precipitation and non-precipitation are distinguished, electromagnetic wave reflection data vectors measured by a radar are processed and finally brought into the classifier, and f* 1>0 time denotes precipitation, f* 1<When the value is 0, the rain is not reduced, the output results of other classifiers are similar to the output results of the rain, snow and hail, and finally, the classifier result with the largest value is used as the classification result of classifying the rain, snow and hail by using the reflection data measured by the radar.
In one embodiment, the step S60 can select the sample for the classifier adjustment from the target domain, so a subset is selected from D2 for unsupervised learning, so that the selected subset can provide the classifier with incremental knowledge to adapt to the target domain. The step may specifically include:
s61, analysis of the subset and D1 similarity: any subset of DTs can be represented by a q-dimensional vector of 0, 1: μ ═ 1(μ 1, μ 2, … …, μ q), where μ i ═ 1 denotes that Ri is in this subset, and otherwise not in this subset, the similarity of the samples of this subset to the samples in the source domain D1 is expressed in a minimized mean, as follows:
Figure BDA0002382893760000122
in the formula, phi (Xi) represents a high-dimensional characteristic function of a sample, is a hidden function and has no specific expression, and the expression is not used in the following calculation; m represents the number of samples of the subset (m ═ μ 1+ μ 2+ … … + μ q); | | represents the norm of the hilbert space;
Figure BDA0002382893760000123
using the above formula as a formula for finding the most similar subset;
s62, phi (Xi) represents the high-dimensional feature function of the sample, and the kernel function has implemented the high-dimensional spatial mapping, so the kernel function corresponding to the norm is still the kernel function determined in S40, so there are:
Figure BDA0002382893760000124
then let A be (K (R)i,Rj))s×sIs the sample kernel matrix of D2, B ═ K (X)i,Rj))t×sThe sample kernel matrices of D1 through D2, the above equation for finding the most similar subset can be simplified as:
Figure BDA0002382893760000131
where T ═ 1,1, … …,1 is a q-dimensional vector, and the above formula is a quadratic optimization problem with constraints of 0 ≦ α i, and α1+α2+……+αq1, so the optimal α can be solved with two sub-optimizations;
s63, set Z ═ Rj|αj≥τ，RjE.g. D2, where the parameter τ is a self-set parameter, D2 ═ Z, i.e. selecting elements in Z in the target domain, resulting in a final set or represented by D2, this set is used as unlabeled samples in semi-supervised learning, and the number of samples in screened D2 is p, let D2 ═ { R1, R2, … …, Rp }.
The rain, snow and hail classification monitoring method based on semi-supervised domain adaptation establishes a classifier through radar wave reflection data with sufficient rainfall type data labels by a semi-supervised domain adaptive method, optimizes the classifier by using a non-labeled microwave link, finally obtains the classifier of the rainfall particle type by using the principle that attenuation rates of microwaves passing through rain, snow and hail particles are different, and can obtain the classifier of the rainfall particle type by using attenuation data of the microwaves, wherein the method has the following advantages:
(1) the method fully utilizes different attenuation influences of rain, snow and hail particles on microwaves to classify and monitor the rain, snow and hail particles;
(2) by adopting a semi-supervised field self-adaptive method, deep features of labeled data and unlabeled data are fully mined, and the demand for labeled microwave attenuation data is reduced;
when the rain, snow and hail particle classifiers are established, a plurality of two-classification classifiers are established simultaneously, the plurality of classifiers are used for training respectively, then a plurality of training results are obtained, the classifier classification result with the largest output result is selected as the final classification result, the plurality of classifiers are trained to work simultaneously, and the classification speed can be accelerated.
In an embodiment, the rain, snow and hail classification monitoring method based on semi-supervised domain adaptation can also be referred to as shown in fig. 2, and includes the following specific steps:
the method comprises the following steps: the reflectivity of the electromagnetic wave in different types of precipitation particle weather is measured by using a radar, and the reflectivity is processed to obtain the preprocessing data of the electromagnetic wave reflection:
a. selecting dual-polarization radar, and measuring horizontal reflectivity Z of rainfallhAnd vertical reflectivity Zv；
b. Calculating the differential reflectivity of the radar electromagnetic wave as follows:
Figure BDA0002382893760000132
in the formula, A represents the differential reflectivity of the radar electromagnetic wave, ZhRepresenting the horizontal polarization reflectivity, ZvRepresents the vertical polarization reflectivity;
c. at a certain moment, multiple groups of radar electromagnetic wave reflection data can be acquired according to the steps a and b in the monitoring area (assuming that n groups of radar devices work in the monitoring area), multiple groups of differential effective reflectivity are obtained, and one n-dimensional real number vector can be obtained, namely X ═ A1，A2，……,An)∈RnWherein A isiIs the differential reflectivity obtained by the ith group of radar devices;
d. obtaining a plurality of n-dimensional real number vectors at different moments (the selected moments include rain, snow and hail weather to ensure sufficient representativeness) through the step c;
step two: obtaining attenuation characteristic quantities of different types of precipitation particles in microwaves of the microwave link by using the microwave link, and processing the attenuation characteristic quantities to obtain preprocessing data of microwave attenuation:
a. selecting a dual-polarization microwave link, transmitting a microwave signal at a transmitting end by using a selected frequency, attenuating the microwave signal when the microwave signal passes through a water-reducing area in the process of propagation, finally receiving the attenuated signal at a receiving end, and measuring transmitting end power and receiving end power on a horizontal link and a vertical link, which are respectively marked as Ph，a，Ph，b，Pv，aAnd Pv，b；
b. Calculating the path attenuation ratio in the polarization direction as follows:
Figure BDA0002382893760000141
in the formula IθRepresents the path attenuation ratio, P, of a microwave link with a polarization direction of thetaθ，1Representing the transmitting end microwave frequency, Pθ，2Indicating the receiving end microwave frequency, OθExpressing the total path attenuation rate of the microwave link with the polarization direction theta, wherein L is the length of the link and has the unit of km;
c. calculating the differential attenuation rate of the microwaves on the microwave link according to the effective attenuation rates in different polarization directions as follows:
Figure BDA0002382893760000142
wherein O represents a differential attenuation ratio of the microwave, OhRepresents the attenuation ratio of vertical polarization, OvRepresents the horizontal polarization attenuation rate;
d. at a certain moment, a plurality of links in the monitoring area (n links in the area of the embodiment) are utilized to obtain microwave attenuation data according to the steps a, b and c, and an n-dimensional real number vector can be obtained, namely R ═ O1，O2，……,On)∈RnIn which O isiIs the differential attenuation rate measured by the ith link;
e. at different moments (the selected moments include rain, snow and hail weather to ensure sufficient representativeness), obtaining n-dimensional real number vectors of a plurality of microwave attenuation data through the step d;
step three: and (3) carrying out spatial transformation on the preprocessed data:
a. in the n-dimensional real number vectors obtained in the first step and the second step, labels of the types of the precipitation particles are marked on X1, X2, … … and Xt to form a data set D1 ═ X1, X2, … … and Xt }, which serves as a labeled source domain, wherein the label of Xi is Yi, and the specific label value of Yi is distinguished according to the difference of the classifiers in which the label is located, which is detailed in the fifth step. Forming another data set of D2 ═ R1, R2, … …, Rs from R1, R2, … …, Rs as unlabeled target domains;
b. PCA principal component analysis was performed on D1 and D2 to obtain the corresponding feature subspace:
(b1) calculating covariance matrixes H1 and H2 of D1 and D2;
(b2) calculating eigenvectors and eigenvalues of H1 and H2, sorting the eigenvalues in descending order, extracting the largest d eigenvalues, and forming an eigen subspace S by taking the corresponding eigenvector as the base of the eigen subspace1And S2Dimension is n × d, if R is notednAll d-dimensional subspaces of Gn×dCalled the grassmann flow pattern, both S1 and S2 are included in the grassmann flow pattern;
step four: sample adaptation:
weighted resampling of source domain samples to approximate the distribution of the target domain:
a. construction of a Slave S on Gn × d1To S2Curve of (2), setting a parameterized function of the curve
Figure BDA0002382893760000151
In the formula, S1Representing a first feature subspace, S2Representing a second feature subspace, P1Is S1Vertical complement space of, U1And U2Diagonal matrices, U, of dxd and (n-d) xd, respectively1From S1'S2＝U1F (v)' is obtained by SVD decomposition, U2From P1'S2＝-U2E (v)' is obtained by SVD decomposition, S1' represents S1The transposed matrix of (1), F (v) and E (v) are diagonal matrices of order d, the diagonal elements of F (v) are cos (α i), the diagonal elements of E (v) are sin (α i),0<i<d, i ∈ {1,2, … …, d }, α i geometrically denote S1And S2A medium base vector geometric angle; at this time, then U1,U2F (v) and E (v) can both be calculated as parameterized functions of the curve
Figure BDA0002382893760000152
Can also be obtained by the above formula;
b. let the weight function be w (v) |1-2v |, where v ∈ [0,1], and on the curve Φ (v), when v is close to 0 or 1, the subspace represented by the corresponding point is more reliable, so a higher weight should be given, so multiplying w (t) by Φ (t) yields:
g(v)＝w(v)φ(v)；
c. g (v) at each point representing the base of a subspace, and summing the inner products of the infinite dimension Hilbert space represented by g (v), i.e. the inner product < g (v), g (v) > is g (v)' g (v) integrated in the interval [0,1], to obtain:
Figure BDA0002382893760000161
since e (v) and f (v) are diagonal matrices, the diagonal elements are cos (v. α i) and sin (v. α i), respectively, then:
Ω＝[S1U1,-P1U2]
therefore, the method comprises the following steps:
Figure BDA0002382893760000162
in the formula, Λ 1, Λ 2, Λ 3 is a diagonal matrix of d × d, and its diagonal elements are respectively λ 1i, λ 2i, λ 3i, where i ∈ {1,2, … …, d }, λ1i,λ2i,λ3iThe expression of (a) is:
Figure BDA0002382893760000163
Figure BDA0002382893760000164
Figure BDA0002382893760000165
and G is a semi-positive definite matrix because it is a kernel matrix on which kernel functions can be defined as follows
K(Xi,Rj)＝Xi'GRj；
Step five: and (3) training a classifier by using a manifold rule learning method and taking the kernel function determined in the fourth step as a kernel function and D1 as a training sample set, and classifying rain, snow and hail according to the training classifier obtained from radar data. The invention is used for classifying a plurality of precipitation particle types, so that samples without precipitation, rain, snow and hail are classified into one type in sequence during training, other rest samples are classified into another type, 4 classifiers are constructed by the 4 types of samples, the label Yi of the first classifier is 1, and the labels Yi of the other three types are-1; the label Yi of rainfall in the second classifier is 1, and the other three types of labels Yi are-1; first, theThe label Yi of snowfall in the three classifiers is 1, and the labels Yi of other three classes are-1; the label Yi of hail in the fourth classifier is 1, and the other three types of labels Yi are-1. The four classifiers are used for training respectively, and then four training results are obtained. During testing, corresponding radar wave reflection data vectors are tested by the four trained classifiers respectively, and finally, each test has a result f* 1(x),f* 2(x),f* 3(x),f* 4(x) The final result is then the largest of these four values as the classification result. The establishment and training of each classifier adopt the following methods:
a. d1 is normalized by taking the classification function as f (X) which is a function from the input vector X to the real number, and popular regularization searches a classifier in the regenerated hilbert space, wherein the objective function is:
Figure BDA0002382893760000171
where V (Xi, yi, f) is a cost function, the present trainer uses a Hinge function, namely:
V(Xi,Yi,f)＝max(0,1-Yif(Xi))
in the formula (I), the compound is shown in the specification,
Figure BDA0002382893760000172
is the norm, gamma, of the kernel function K induced regenerated Hilbert spaceAIs a user-defined parameter;
b. using the classifier f obtained in ar *R belongs to {1,2,3,4}, and r is 1, when distinguishing precipitation from non-precipitation, electromagnetic wave reflection data vector measured by radar is processed and finally brought into the classifier, f* 1>0 time denotes precipitation, f* 1<When 0, the result is not precipitation, the output results of other classifiers are similar to the result, and finally, the classifier result with the largest value is used as the reflection data measured by the radar to classify rain, snow and hail in the four classifier output valuesThe classification result of (1).
Step six: the samples from the target domain are selected for classifier adaptation, so a subset from D2 is selected for unsupervised learning, such that the selected subset provides the classifier with incremental knowledge to adapt to the target domain:
a. subset and D1 were analyzed for similarity: any subset of DTs can be represented by a q-dimensional vector of 0, 1: μ ═ 1(μ 1, μ 2, … …, μ q), where μ i ═ 1 denotes that Ri is in this subset, and otherwise not in this subset, the similarity of the samples of this subset to the samples in the source domain D1 is expressed in a minimized mean, as follows:
Figure BDA0002382893760000173
in the formula, phi (Xi) represents a high-dimensional characteristic function of a sample, is a hidden function and has no specific expression, and the expression is not used in the following calculation; m represents the number of samples of the subset (m ═ μ 1+ μ 2+ … … + μ q); | | represents the norm of the hilbert space;
Figure BDA0002382893760000174
using the above formula as a formula for finding the most similar subset;
b. phi (Xi) represents a high-dimensional characteristic function of the sample, and the kernel function determined in the step four realizes high-dimensional space mapping, so the kernel function corresponding to the set norm still uses the kernel function determined in the step four, so that:
Figure BDA0002382893760000175
then let A be (K (R)i,Rj))s×sIs the sample kernel matrix of D2, B ═ K (X)i,Rj))t×sThe sample kernel matrices of D1 through D2, the above equation for finding the most similar subset can be simplified as:
Figure BDA0002382893760000181
where T ═ 1,1, … …,1 is a q-dimensional vector, and the above formula is a quadratic optimization problem with constraints of 0 ≦ α i, and α1+α2+……+αq1, so the optimal α can be solved with two sub-optimizations;
c. let set Z ═ Rj|αj≥τ，RjE.g. D2, where the parameter τ is a self-set parameter, D2 ═ Z, i.e. an element in Z is selected in the target domain, and a final set is obtained or represented by D2, this set is used as an unlabeled sample in semi-supervised learning, the number of samples in screened D2 is p, and D2 ═ { R1, R2, … …, Rp };
step seven: and (3) adjusting the classifier:
a. d2 is normalized, the classification function of the radar data classification rain, snow and hail obtained in the step five is optimized, and the optimized objective function is as follows:
Figure BDA0002382893760000182
wherein F ═ F (X)1),f(X2),……,f(Xt),f(R1),f(R2),……,f(Rp) Π is the laplace transform, γ, of the adjacency matrix of data in D1 and D2BIs a user-defined parameter;
b. the adjacency graph of data in D1 and D2 is constructed according to the KNN method, wherein the distance between samples is defined as follows:
d(Xi,Rj)＝K(Xi,Ri)+K(Xj,Rj)-2K(Xi,Rj)
wherein d (X)i,Rj) Represents XiAnd RjDistance between, XiRepresenting the i-th sample vector, R, in the source domain D1jRepresents the jth sample vector, X, in the target domain D2jRepresents the j-th sample vector, R, in the source domain (first data set) D1iRepresents the ith sample vector, K (X), in the target domain (second data set) D2i,Ri) Is composed of XiAnd RiKernel function of, K (X)j,Rj) Is composed of XjAnd RjKernel function of, K (X)i,Rj) Is composed of XiAnd RjThe adjacency matrix of the data is constructed by a KNN method according to the formula to obtain an adjacency matrix M, elements of each row in the matrix M are added to be used as diagonal elements of a matrix D, and if other elements of D are 0, pi is equal to D-M;
c. optimization of f using L obtained in b*To obtain the final classifier fr *R belongs to {1,2,3,4}, and r is 1, when the precipitation is distinguished from the non-precipitation, the microwave attenuation data vector measured in real time is processed and finally brought into the classifier, and f is f* 1>0 time denotes precipitation, f* 1<When the value is 0, the water is not reduced, the output results of other classifiers are similar to the water, and finally, the classifier result with the largest value is used as the final classification result obtained through microwave attenuation data in the four classifier output values.
The technical features of the above embodiments can be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the above embodiments are not described, but should be considered as the scope of the present specification as long as there is no contradiction between the combinations of the technical features.
It should be noted that the terms "first \ second \ third" referred to in the embodiments of the present application merely distinguish similar objects, and do not represent a specific ordering for the objects, and it should be understood that "first \ second \ third" may exchange a specific order or sequence when allowed. It should be understood that "first \ second \ third" distinct objects may be interchanged under appropriate circumstances such that the embodiments of the application described herein may be implemented in an order other than those illustrated or described herein.
The terms "comprising" and "having" and any variations thereof in the embodiments of the present application are intended to cover non-exclusive inclusions. For example, a process, method, apparatus, product, or device that comprises a list of steps or modules is not limited to the listed steps or modules but may alternatively include other steps or modules not listed or inherent to such process, method, product, or device.
The above-mentioned embodiments only express several embodiments of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present patent shall be subject to the appended claims.

Claims (7)

1. A rain, snow and hail classification monitoring method based on semi-supervised domain adaptation is characterized by comprising the following steps:
s10, radar wave reflectivity of the electromagnetic waves in various precipitation particle weather is measured by adopting a radar; wherein the types of precipitation particles include rain, snow and hail;
s20, acquiring preprocessing data under the weather of various types of precipitation particles according to the radar wave reflectivity under the weather of various types of precipitation particles;
s30, constructing a first data set carrying labels and a second data set not carrying labels according to the preprocessed data, calculating a first covariance matrix of the first data set and a second covariance matrix of the second data set, determining a first feature subspace according to the first covariance matrix, and determining a second feature subspace according to the second covariance matrix;
s40, determining a kernel function according to the first characteristic subspace and the second characteristic subspace;
s50, training an initial classifier by taking the first data set as a training sample set according to the kernel function;
s60, selecting a subset from the second data set to perform unsupervised learning on the initial classifier, so that the selected subset can provide incremental knowledge for the initial classifier to adapt to the target field;
s70, obtaining an objective function of the initial classifier after unsupervised learning, determining an adjacency graph according to the first data set and the second data set, optimizing the objective function according to the adjacency graph to determine a final classifier, and classifying rain, snow and hail by using the final classifier.
2. The method for classifying and monitoring rain, snow and hail based on semi-supervised domain adaptation according to claim 1, wherein the step of measuring the radar wave reflectivity of the electromagnetic waves in various types of precipitation particle weather by using radar comprises the following steps:
measuring multiple groups of effective horizontal reflectivity Z of precipitation particles under various precipitation particle weather by adopting dual-polarization radarhAnd vertical reflectivity ZvAccording to each group of effective horizontal reflectivity ZhAnd vertical reflectivity ZvAnd calculating the differential reflectivity, and determining the radar wave reflectivity of each type of precipitation particle according to the differential reflectivity corresponding to each type of precipitation particle.
3. The method for classifying and monitoring rain, snow and hail based on semi-supervised domain adaptation according to claim 2, wherein the differential reflectivity determination process comprises:
Figure FDA0002382893750000011
wherein A represents a differential reflectance, ZhRepresenting the effective horizontal reflectivity, ZvIndicating the vertical reflectivity.
4. The method for classifying and monitoring rain, snow and hail based on semi-supervised domain adaptation according to claim 2, wherein the obtaining of the preprocessed data of each type of precipitation particle weather according to the radar wave reflectivity of each type of precipitation particle weather comprises:
s22, calculating path attenuation rates of each microwave link in each polarization direction; wherein the determination formula of the path attenuation rate comprises:
Figure FDA0002382893750000021
in the formula IθRepresents the path attenuation ratio, P, of a microwave link with a polarization direction of thetaθ，1Representing the transmitting-end microwave frequency, P, of a microwave link with a polarization direction of thetaθ，2The microwave frequency of a receiving end of the microwave link with the polarization direction theta is represented, and L represents the length of the microwave link;
s23, calculating the differential attenuation rate of the microwave on the microwave link according to the path attenuation rates in different polarization directions; the determination formula of the differential attenuation rate comprises the following steps:
Figure FDA0002382893750000022
wherein O represents a differential attenuation ratio of microwave, IhRepresents the vertical polarization attenuation ratio, I, of the microwave linkvRepresenting the horizontal polarization attenuation rate of the microwave link;
and S25, respectively executing the steps S22 to S23 for each microwave link under the weather of each type of precipitation particles, acquiring a group of differential attenuation rates corresponding to each type of precipitation particles, and determining the preprocessing data of each type of precipitation particles according to the group of differential attenuation rates corresponding to each type of precipitation particles.
5. The semi-supervised domain adaptation-based rain, snow and hail classification monitoring method according to claim 1, wherein the first and second feature subspaces are both a subspace G with dimension n x dn×d；
The determining a kernel function from the first feature subspace and the second feature subspace comprises:
s41, constructing a slave S on the first feature subspace and the second feature subspace1To S2The curve of (d); the parameterized function of the curve includes:
Figure FDA0002382893750000023
in the formula, S1Represents the first featureSymbol space, S2Representing a second feature subspace, P1Is S1Vertical complement space of, U1And U2Diagonal matrices, U, of dxd and (n-d) xd, respectively1From S1'S2＝U1F (v)' is obtained by SVD decomposition, U2From P1'S2＝-U2E (v)' is obtained by SVD decomposition, S1' represents S1The transposed matrix of (1), F (v) and E (v) are diagonal matrices of order d, the diagonal elements of F (v) are cos (α i), the diagonal elements of E (v) are sin (α i),0<i<d, i ∈ {1,2, … …, d }, α i denotes S1And S2A medium base vector geometric angle;
s43, acquiring a basis function G (v) of a basis representing a subspace at each point, calculating a semi-positive definite matrix G according to the basis function G (v), and setting a kernel function according to the semi-positive definite matrix G; the calculation process of the semi-positive definite matrix G comprises the following steps:
Ω＝[S1U1,-P1U2]，
Figure FDA0002382893750000031
in the formula, Λ1、Λ2And Λ3Respectively, a diagonal matrix of d x d.
6. The method for classifying and monitoring rain, snow and hail based on semi-supervised domain adaptation according to claim 5, wherein the setting process of the kernel function comprises:
K(Xi,Rj)＝Xi'GRj，
in the formula, K (X)i,Rj) Represents XiAnd RjCorresponding kernel function, XiRepresenting the i-th sample vector, R, in the source domain D1jRepresenting the jth sample vector of the target domain D2.
7. The method of claim 1, wherein training an initial classifier based on the kernel function with a first data set as a training sample set comprises:
s51, normalizing the first data set, setting a classification function from an input vector X to a real number as f (X), and searching an initial classifier in a regenerative Hilbert space by adopting popular regularization to obtain a target function of the initial classifier;
and S52, training each type of precipitation particle by adopting an initial classifier.
CN202010088477.5A 2020-02-12 2020-02-12 Rain, snow and hail classification monitoring method based on semi-supervised domain adaptation Active CN111308471B (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN202010088477.5A CN111308471B (en) 2020-02-12 2020-02-12 Rain, snow and hail classification monitoring method based on semi-supervised domain adaptation
PCT/CN2020/136089 WO2021159844A1 (en) 2020-02-12 2020-12-14 Rain, snow, and hail classification monitoring method based on semi-supervised domain adaptation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010088477.5A CN111308471B (en) 2020-02-12 2020-02-12 Rain, snow and hail classification monitoring method based on semi-supervised domain adaptation

Publications (2)

Publication Number Publication Date
CN111308471A true CN111308471A (en) 2020-06-19
CN111308471B CN111308471B (en) 2020-11-24

Family

ID=71156484

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010088477.5A Active CN111308471B (en) 2020-02-12 2020-02-12 Rain, snow and hail classification monitoring method based on semi-supervised domain adaptation

Country Status (2)

Country Link
CN (1) CN111308471B (en)
WO (1) WO2021159844A1 (en)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112733859A (en) * 2021-01-25 2021-04-30 重庆大学 Depth migration semi-supervised domain self-adaptive classification method for histopathology image
CN113095442A (en) * 2021-06-04 2021-07-09 成都信息工程大学 Hail identification method based on semi-supervised learning under multi-dimensional radar data
WO2021159844A1 (en) * 2020-02-12 2021-08-19 河海大学 Rain, snow, and hail classification monitoring method based on semi-supervised domain adaptation
CN113591387A (en) * 2021-08-05 2021-11-02 安徽省气象台 Huber norm constraint-based satellite data inversion precipitation method and system
CN114488160A (en) * 2022-04-02 2022-05-13 南京师范大学 Radar rainfall estimation error correction method considering influence of three-dimensional wind field
CN114692692A (en) * 2022-04-02 2022-07-01 河海大学 Snowfall identification method based on microwave attenuation signal fusion kernel extreme learning machine
CN114814993A (en) * 2022-03-25 2022-07-29 河海大学 Microwave attenuation snowfall intensity monitoring method based on DCGAN and 2D-CNN

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114706146B (en) * 2022-03-23 2023-11-03 成都信息工程大学 Method for forecasting growth of hail embryo and hail-down stage in hail-down storm process of complex terrain
CN116911082B (en) * 2023-09-14 2023-12-05 成都信息工程大学 Precipitation particle quality and quantity estimation method based on precipitation radar and assimilation data

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160358107A1 (en) * 2015-06-04 2016-12-08 Accusonus, Inc. Data training in multi-sensor setups
CN106599787A (en) * 2016-11-17 2017-04-26 河海大学 Single sample face recognition method based on semi-supervised block joint regression
CN110161480A (en) * 2019-06-18 2019-08-23 西安电子科技大学 Radar target identification method based on semi-supervised depth probabilistic model
US20190383904A1 (en) * 2018-06-13 2019-12-19 Metawave Corporation Autoencoder assisted radar for target identification
CN110728210A (en) * 2019-09-25 2020-01-24 上海交通大学 Semi-supervised target labeling method and system for three-dimensional point cloud data

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106096627A (en) * 2016-05-31 2016-11-09 河海大学 The Polarimetric SAR Image semisupervised classification method that considering feature optimizes
CN108280394A (en) * 2017-12-15 2018-07-13 西安电子科技大学 Radar Signal Recognition method based on time-frequency distributions singular value decomposition
CN111308471B (en) * 2020-02-12 2020-11-24 河海大学 Rain, snow and hail classification monitoring method based on semi-supervised domain adaptation

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160358107A1 (en) * 2015-06-04 2016-12-08 Accusonus, Inc. Data training in multi-sensor setups
CN106599787A (en) * 2016-11-17 2017-04-26 河海大学 Single sample face recognition method based on semi-supervised block joint regression
US20190383904A1 (en) * 2018-06-13 2019-12-19 Metawave Corporation Autoencoder assisted radar for target identification
CN110161480A (en) * 2019-06-18 2019-08-23 西安电子科技大学 Radar target identification method based on semi-supervised depth probabilistic model
CN110728210A (en) * 2019-09-25 2020-01-24 上海交通大学 Semi-supervised target labeling method and system for three-dimensional point cloud data

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021159844A1 (en) * 2020-02-12 2021-08-19 河海大学 Rain, snow, and hail classification monitoring method based on semi-supervised domain adaptation
CN112733859A (en) * 2021-01-25 2021-04-30 重庆大学 Depth migration semi-supervised domain self-adaptive classification method for histopathology image
CN112733859B (en) * 2021-01-25 2023-12-19 重庆大学 Depth migration semi-supervised domain self-adaptive classification method for histopathological image
CN113095442A (en) * 2021-06-04 2021-07-09 成都信息工程大学 Hail identification method based on semi-supervised learning under multi-dimensional radar data
CN113095442B (en) * 2021-06-04 2021-09-10 成都信息工程大学 Hail identification method based on semi-supervised learning under multi-dimensional radar data
CN113591387A (en) * 2021-08-05 2021-11-02 安徽省气象台 Huber norm constraint-based satellite data inversion precipitation method and system
CN113591387B (en) * 2021-08-05 2023-09-01 安徽省气象台 Satellite data inversion precipitation method and system based on Huber norm constraint
CN114814993A (en) * 2022-03-25 2022-07-29 河海大学 Microwave attenuation snowfall intensity monitoring method based on DCGAN and 2D-CNN
CN114488160A (en) * 2022-04-02 2022-05-13 南京师范大学 Radar rainfall estimation error correction method considering influence of three-dimensional wind field
CN114692692A (en) * 2022-04-02 2022-07-01 河海大学 Snowfall identification method based on microwave attenuation signal fusion kernel extreme learning machine
CN114692692B (en) * 2022-04-02 2023-05-12 河海大学 Snowfall recognition method based on microwave attenuation signal fusion kernel extreme learning machine

Also Published As

Publication number Publication date
WO2021159844A1 (en) 2021-08-19
CN111308471B (en) 2020-11-24

Similar Documents

Publication Publication Date Title
CN111308471B (en) Rain, snow and hail classification monitoring method based on semi-supervised domain adaptation
CN108388927B (en) Small sample polarization SAR terrain classification method based on deep convolution twin network
US20230252761A1 (en) Method for classifying hyperspectral images on basis of adaptive multi-scale feature extraction model
Pasolli et al. Automatic analysis of GPR images: A pattern-recognition approach
CN108846426B (en) Polarization SAR classification method based on deep bidirectional LSTM twin network
CN104504393B (en) Polarimetric SAR Image semisupervised classification method based on integrated study
CN111160176B (en) Fusion feature-based ground radar target classification method for one-dimensional convolutional neural network
CN108133232A (en) A kind of Radar High Range Resolution target identification method based on statistics dictionary learning
CN112434643A (en) Classification and identification method for low-slow small targets
CN109901130B (en) Rotor unmanned aerial vehicle detection and identification method based on Radon transformation and improved 2DPCA
CN104732244A (en) Wavelet transform, multi-strategy PSO (particle swarm optimization) and SVM (support vector machine) integrated based remote sensing image classification method
CN111639587B (en) Hyperspectral image classification method based on multi-scale spectrum space convolution neural network
CN104809471B (en) A kind of high spectrum image residual error integrated classification method based on spatial spectral information
CN106096506A (en) Based on the SAR target identification method differentiating doubledictionary between subclass class
CN107358214A (en) Polarization SAR terrain classification method based on convolutional neural networks
CN113962883A (en) Method for enhancing ship target detection SAR image data
CN114742102B (en) NLOS signal identification method and system
CN110427878A (en) A kind of sudden and violent signal recognition method of Rapid Radio and system
CN111580058A (en) Radar HRRP target identification method based on multi-scale convolution neural network
CN111563528B (en) SAR image classification method based on multi-scale feature learning network and bilateral filtering
CN110703221A (en) Urban low-altitude small target classification and identification system based on polarization characteristics
CN105160351A (en) Semi-monitoring high-spectral classification method based on anchor point sparse graph
CN114740441A (en) Low-slow small-target radar echo identification method based on small samples
CN116482618B (en) Radar active interference identification method based on multi-loss characteristic self-calibration network
P. Thampy et al. A convolution neural network approach to Doppler spectra classification of 205 MHz radar

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
GR01 Patent grant
GR01 Patent grant