WO2021159844A1 - 基于半监督域适应的雨、雪、冰雹分类监测方法 - Google Patents

基于半监督域适应的雨、雪、冰雹分类监测方法 Download PDF

Info

Publication number
WO2021159844A1
WO2021159844A1 PCT/CN2020/136089 CN2020136089W WO2021159844A1 WO 2021159844 A1 WO2021159844 A1 WO 2021159844A1 CN 2020136089 W CN2020136089 W CN 2020136089W WO 2021159844 A1 WO2021159844 A1 WO 2021159844A1
Authority
WO
WIPO (PCT)
Prior art keywords
data set
hail
snow
rain
precipitation
Prior art date
Application number
PCT/CN2020/136089
Other languages
English (en)
French (fr)
Inventor
杨涛
陈志远
郑鑫
师鹏飞
秦友伟
李振亚
Original Assignee
河海大学
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 河海大学 filed Critical 河海大学
Publication of WO2021159844A1 publication Critical patent/WO2021159844A1/zh

Links

Images

Classifications

    • 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
    • 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
    • G01S7/00Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
    • G01S7/02Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
    • G01S7/024Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using polarisation effects
    • 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
    • G01S13/951Radar or analogous systems specially adapted for specific applications for meteorological use ground based
    • 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
    • G01S7/00Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
    • G01S7/003Transmission of data between radar, sonar or lidar systems and remote stations
    • G01S7/006Transmission of data between radar, sonar or lidar systems and remote stations using shared front-end circuitry, e.g. antennas
    • 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
    • G01S7/00Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
    • G01S7/02Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
    • G01S7/35Details of non-pulse systems
    • G01S7/352Receivers
    • G01S7/354Extracting wanted echo-signals
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06F18/2155Generating 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
    • 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 ground weather detection, in particular to a rain, snow, and hail classification monitoring method based on semi-supervised domain adaptation.
  • the microwave communication network has wide coverage, high signal quality, and basically no blind spots. Therefore, the rainfall monitoring and analysis technology of microwave communication network has high promotion and application value in our country.
  • the existing microwave methods for identifying rain, snow, and hail types use traditional machine learning methods and require a large amount of labeled microwave attenuation data. However, in actual scenarios, the amount of microwave attenuation data with precipitation type tags is often not sufficient, so The classification and monitoring schemes for rain, snow, and hail often have the problem of low accuracy.
  • the present invention proposes a rain, snow, and hail classification monitoring method based on semi-supervised domain adaptation.
  • a rain, snow, and hail classification monitoring method based on semi-supervised domain adaptation includes the following steps:
  • S30 Construct a first data set with labels and a second data set without labels according to the preprocessed data, calculate the first covariance matrix of the first data set and the second covariance matrix of the second data set, and calculate the first covariance matrix of the second data set according to the first data set.
  • the covariance matrix determines the first feature subspace, and the second feature subspace is determined according to the second covariance matrix;
  • the radar reflectivity of electromagnetic waves measured by radar under various types of precipitation particle weather includes:
  • the process of determining the differential reflectance includes:
  • A represents the differential reflectance
  • Z h represents the effective horizontal reflectance
  • Z v represents the vertical reflectance
  • acquiring the pre-processed data under various types of precipitation particle weather according to the radar wave reflectivity under various types of precipitation particle weather includes:
  • I ⁇ represents the path attenuation rate of the microwave link with the polarization direction ⁇
  • P ⁇ , 1 represents the microwave frequency at the transmitting end of the microwave link with the polarization direction ⁇
  • P ⁇ , 2 represents the microwave link with the polarization direction ⁇
  • the microwave frequency at the receiving end of the road, L represents the length of the microwave link;
  • S23 Calculate the differential attenuation rate of microwaves on the microwave link according to the path attenuation rates in different polarization directions; the formula for determining the differential attenuation rate includes:
  • O represents the differential attenuation rate of microwave
  • I h represents the vertical polarization attenuation rate of the microwave link
  • I v represents the horizontal polarization attenuation rate of the microwave link
  • S25 Perform steps S22 to S23 for each microwave link under each type of precipitation particle weather to obtain a set of differential attenuation rates corresponding to each type of precipitation particles, and determine according to a set of differential attenuation rates corresponding to each type of precipitation particles. Preprocessed data of various types of precipitation particles.
  • the first feature subspace and the second feature subspace are both subspaces G n ⁇ d with a dimension of n ⁇ d;
  • the determining the kernel function according to the first feature subspace and the second feature subspace includes:
  • S 1 represents the first feature subspace
  • S 2 represents the second feature subspace
  • P 1 is the vertical complement subspace of S 1
  • U 1 and U 2 are the pairs of d ⁇ d and (nd) ⁇ d, respectively
  • the transposed matrix of S 1 , F(v) and E(v) are both d-order diagonal matrices, the diagonal element of F(v) is cos( ⁇ i), and the diagonal element of E(v) is sin ( ⁇ i),0 ⁇ i ⁇ d,i ⁇ 1,2, whil,d ⁇ , ⁇ i geometrically represents the geometric angle of the basis vectors in S 1 and S 2;
  • ⁇ 1 , ⁇ 2 and ⁇ 3 are d ⁇ d diagonal matrices respectively.
  • the process of setting the kernel function includes:
  • K (X i, R j ) represents a j X i and the corresponding R & lt kernel
  • X i represents the source domain D1 i-th sample vector
  • R j represents the j-th target domain D2 sample vectors.
  • training the initial classifier includes:
  • S52 Use the initial classifier to train separately for each type of precipitation particles.
  • the above-mentioned rain, snow, and hail classification monitoring method based on semi-supervised domain adaptation uses radar to measure the radar reflectivity of electromagnetic waves under various types of precipitation particle weather, and obtain various types according to the radar wave reflectivity under various types of precipitation particle weather.
  • Preprocessed data under precipitation particle weather construct a first data set with tags and a second data set without tags based on the preprocessed data, and calculate the first covariance matrix of the first data set and the second data set of the second data set
  • the covariance matrix determines the first feature subspace according to the first covariance matrix, determines the second feature subspace according to the second covariance matrix, and then determines the kernel function according to the first feature subspace and the second feature subspace, so as to The kernel function, taking the first data set as the training sample set, training the initial classifier, and selecting a subset from the second data set to perform unsupervised learning of the initial classifier, so that the selected subset can be used for the initial classifier
  • Provide incremental knowledge to adapt to the target field then obtain the objective function of the initial classifier after unsupervised learning, determine the adjacency graph according to the first data set and the second data set, and optimize the objective function according to the adjacency graph to determine
  • the final classifier uses the final classifier to classify rain
  • FIG. 1 is a schematic flowchart of a rain, snow, and hail classification monitoring method based on semi-supervised domain adaptation according to an embodiment
  • Fig. 2 is a schematic flowchart of another embodiment of a rain, snow, and hail classification monitoring method based on semi-supervised domain adaptation.
  • Fig. 1 is a schematic flowchart of a rain, snow, and hail classification monitoring method based on semi-supervised domain adaptation according to an embodiment, including the following steps:
  • each type of precipitation particle can respectively represent the corresponding radar wave reflectivity by an n-dimensional real number vector.
  • the above preprocessed data can be represented by an n-dimensional real number vector.
  • each type of precipitation particle can be represented by an n-dimensional real number vector to represent the corresponding preprocessed data.
  • the foregoing steps may be specifically preprocessed according to the relevant parameters generated during the acquisition process of radar wave reflectivity under various types of precipitation particle weather, so as to obtain corresponding preprocessed data.
  • S30 Construct a first data set with labels and a second data set without labels according to the preprocessed data, calculate the first covariance matrix of the first data set and the second covariance matrix of the second data set, and calculate the first covariance matrix of the second data set according to the first data set.
  • the covariance matrix determines the first feature subspace, and the second feature subspace is determined according to the second covariance matrix.
  • the above steps may perform spatial transformation on the preprocessed data to obtain the first feature subspace and the second feature subspace.
  • the process of spatial transformation may include:
  • the first data set D1 ⁇ X1, X2,..., Xt ⁇ , as the labeled source domain (the first data set), where the label of Xi is Yi, and the specific label value of Yi depends on the classifier.
  • This step may also include:
  • S40 Determine a kernel function according to the first feature subspace and the second feature subspace.
  • the above steps can perform weighted resampling on the source domain samples, thereby approximating the distribution of the target domain, and realizing the corresponding sample adaptation.
  • the first data set is used as a training sample set to train an initial classifier.
  • the above steps can use the manifold regularization learning method, according to the kernel function, use D1 as the training sample set, train the classifier, and classify the rain, snow, and hail based on the training classifier obtained from the radar data.
  • manifold regularization classification methods are mostly used for binary classification, and the embodiments can be used for the classification of multiple precipitation particle types. Therefore, during training, samples of unprecipitation, rain, snow, and hail are classified into one category in sequence. The other remaining samples are classified into another category, so that the samples of each category construct the corresponding classifier. Take the samples with no precipitation, rain, snow, and hail into one category in turn, and the other remaining samples into another category as an example.
  • the corresponding radar wave reflection data vector is tested using these four trained classifiers, and finally each test has a result f * 1 (x), f * 2 (x), f * 3 (x), f * 4 (x), so the final result is the largest of these four values as the classification result.
  • S60 Select a subset from the second data set to perform unsupervised learning of the initial classifier, so that the selected subset can provide the initial classifier with incremental knowledge to adapt to the target domain.
  • the above steps may include the following processes:
  • F (f(X 1 ),f(X 2 ), whil,f(X t ),f(R 1 ),f(R 2 ), whil,f(R p )),
  • represents the Laplace transform of the adjacency matrix of the data in D1 and D2
  • ⁇ B is a user-defined parameter
  • Xi represents the n-dimensional vector representing radar wave reflection
  • Yi represents the label vector of Xi
  • V(Xi,Yi,f) Represents the cost function
  • t represents the number of labeled radar wave reflectivity vectors
  • p represents the number of vectors in the subset selected from the second data set
  • the adjacency graphs of the data in D1 and D2 are constructed according to the KNN method, where the distance between samples is defined as follows:
  • d (X i, R j ) represents the distance between the X i and R j
  • X i represents the source domain D1 i-th sample vector
  • R j represents the target domain D2 j-th sample vector
  • X j denotes a source
  • R i represents the i-th sample vector in the target domain (the second data set) D2
  • K(X i , R i ) is the result of X i and R i Kernel function
  • K(X j ,R j ) is the kernel function of X j and R j
  • K(X i ,R j ) is the kernel function of X i and R j
  • the adjacency matrix of the data has KNN according to the above formula
  • the method constructs an adjacency matrix M, and adds each row element of the matrix M as the diagonal element of the matrix D.
  • the above-mentioned rain, snow, and hail classification monitoring method based on semi-supervised domain adaptation uses radar to measure the radar reflectivity of electromagnetic waves under various types of precipitation particle weather, and obtain various types according to the radar wave reflectivity under various types of precipitation particle weather.
  • Preprocessed data under precipitation particle weather construct a first data set with tags and a second data set without tags based on the preprocessed data, and calculate the first covariance matrix of the first data set and the second data set of the second data set
  • the covariance matrix determines the first feature subspace according to the first covariance matrix, determines the second feature subspace according to the second covariance matrix, and then determines the kernel function according to the first feature subspace and the second feature subspace, so as to The kernel function, taking the first data set as the training sample set, training the initial classifier, and selecting a subset from the second data set to perform unsupervised learning of the initial classifier, so that the selected subset can be used for the initial classifier
  • Provide incremental knowledge to adapt to the target field then obtain the objective function of the initial classifier after unsupervised learning, determine the adjacency graph according to the first data set and the second data set, and optimize the objective function according to the adjacency graph to determine
  • the final classifier uses the final classifier to classify rain
  • the radar reflectivity of electromagnetic waves measured by radar under various types of precipitation particle weather includes:
  • the above-mentioned radar wave reflectivity can be characterized by an n-dimensional real number vector.
  • the n-dimensional real number vector of a certain type of precipitation particle includes each differential reflectivity corresponding to the type of precipitation particle.
  • the process of determining the differential reflectance includes:
  • A represents the differential reflectance
  • Z h represents the effective horizontal reflectance
  • Z v represents the vertical reflectance
  • acquiring the pre-processed data under various types of precipitation particle weather according to the radar wave reflectivity under various types of precipitation particle weather includes:
  • I ⁇ represents the path attenuation rate of the microwave link with the polarization direction ⁇
  • P ⁇ , 1 represents the microwave frequency at the transmitting end of the microwave link with the polarization direction ⁇
  • P ⁇ , 2 represents the microwave link with the polarization direction ⁇
  • the microwave frequency at the receiving end of the road, L represents the length of the microwave link;
  • S23 Calculate the differential attenuation rate of microwaves on the microwave link according to the path attenuation rates in different polarization directions; the formula for determining the differential attenuation rate includes:
  • O represents the differential attenuation rate of microwave
  • I h represents the vertical polarization attenuation rate of the microwave link
  • I v represents the horizontal polarization attenuation rate of the microwave link
  • S25 Perform steps S22 to S23 for each microwave link under each type of precipitation particle weather to obtain a set of differential attenuation rates corresponding to each type of precipitation particles, and determine according to a set of differential attenuation rates corresponding to each type of precipitation particles. Preprocessed data of various types of precipitation particles.
  • step S22 it may also include
  • S21 Select a dual-polarization microwave link, and transmit a microwave signal at the transmitting end with a selected frequency (for example, the polarization frequency is 40 Hz).
  • the microwave signal is attenuated when it passes through the precipitation area during the propagation process, and is finally received at the receiving end
  • the transmitter power and the receiver power on the horizontal and vertical links are measured, denoted as P h, a , P h, b , P v, a and P v, b respectively .
  • the first feature subspace and the second feature subspace are both subspaces G n ⁇ d with a dimension of n ⁇ d.
  • the determining the kernel function according to the first feature subspace and the second feature subspace includes:
  • S 1 represents the first feature subspace
  • S 2 represents the second feature subspace
  • P 1 is the vertical complement subspace of S 1
  • U 1 and U 2 are the pairs of d ⁇ d and (nd) ⁇ d, respectively
  • the transposed matrix of S 1 , F(v) and E(v) are both d-order diagonal matrices, the diagonal element of F(v) is cos( ⁇ i), and the diagonal element of E(v) is sin ( ⁇ i),0 ⁇ i ⁇ d, i ⁇ 1,2, whil,d ⁇ , ⁇ i is geometrically the angle of the basis vectors in S 1 and S 2 ; at this time, U 1 , U 2 , F( Both v) and E(v) can be calculated, the parameterized function of the curve
  • ⁇ 1 , ⁇ 2 and ⁇ 3 are d ⁇ d diagonal matrices respectively.
  • the process of setting the kernel function includes:
  • K (X i, R j ) represents a j X i and the corresponding R & lt kernel
  • X i represents the source domain D1 i-th sample vector
  • R j represents the j-th target domain D2 sample vectors.
  • step S41 it may further include:
  • g(v) represents the basis of a subspace at each point, and sum the inner product of the infinite-dimensional Hilbert space represented by g(v), that is, the inner product ⁇ g(v), g(v )> is the integration of g(v)'g(v) in the interval [0,1].
  • the derivation process of the positive semi-definite matrix G may include:
  • ⁇ 1 , ⁇ 2 and ⁇ 3 are d ⁇ d diagonal matrices respectively, the diagonal element of ⁇ 1 is ⁇ 1i , the diagonal element of ⁇ 2 is ⁇ 2i , and the diagonal element of ⁇ 3 is ⁇ 3i , where i ⁇ 1,2,ising,d ⁇ , ⁇ 1i , ⁇ 2i , ⁇ 3i are expressed as:
  • G is a positive semi-definite matrix, because it is a kernel matrix, the kernel function can be defined as follows
  • training the initial classifier includes:
  • S52 Use the initial classifier to train separately for each type of precipitation particles.
  • step S51 D1 is normalized, and the classification function is set to f(X), which is a function from the input vector X to a real number, then the popular regularization searches for a classifier in the regenerated Hilbert space,
  • the objective function is:
  • V(Xi,Yi,f) is the cost function
  • this trainer uses the Hinge function, namely:
  • V(X i ,Y i ,f) max(0,1-Y i f(X i )),
  • I the norm of the regenerated Hilbert space induced by the corresponding kernel function
  • ⁇ A is a user-defined parameter
  • Y i represents the label of Xi.
  • the classifier f r * obtained in S51 can be used to adapt the microwave data to the radar data field based on the f * established by the radar data.
  • f * 1 > it means precipitation
  • f * 1 ⁇ 0 means no precipitation.
  • the output results of several other classifiers are similar to it.
  • the one with the largest value is used as the reflection data measured by radar for rain, snow, and hail. The classification result of the classification.
  • the above step S60 can select samples for classifier adjustment from the target domain, so a subset of D2 should be selected for unsupervised learning, so that the selected subset can provide incremental knowledge to the classifier Go to the adaptive target area.
  • This step may specifically include:
  • ⁇ (Xi) represents the high-dimensional characteristic function of the sample, which is an implicit function without a specific expression, and its expression is not used in the subsequent calculations;
  • represents the norm of the Hilbert space; Use the above formula as the basis formula for finding the most similar subset;
  • ⁇ (Xi) represents the high-dimensional feature function of the sample
  • the kernel function has realized the high-dimensional space mapping, so the kernel function corresponding to the norm still uses the kernel function determined by S40, so:
  • the attenuation data of the microwave can be used to obtain the classifier of the precipitation particle type. It has the following advantages:
  • the above-mentioned rain, snow, and hail classification monitoring method based on semi-supervised domain adaptation can also be referred to as shown in Fig. 2.
  • the specific steps are as follows:
  • Step 1 Use radar to measure the reflectivity of electromagnetic waves under different types of precipitation particle weather, and process them to obtain the pre-processing data of electromagnetic wave reflection:
  • A represents the differential reflectivity of radar electromagnetic waves
  • Z h represents the horizontal polarization reflectivity
  • Z v represents the vertical polarization reflectivity
  • multiple groups of radar electromagnetic wave reflection data can be obtained by following steps a and b in the monitoring area (assuming that there are n groups of radar devices in the monitoring area), and multiple groups of differential effective reflectivity can be obtained.
  • An n-dimensional real number vector, that is, X (A 1 , A 2 ,..., A n ) ⁇ R n , where A i is the differential reflectance obtained by the i-th group of radar devices;
  • step c At different times (the selected time should include rain, snow, and hail weather to ensure sufficient representativeness), through step c, multiple n-dimensional real number vectors are obtained;
  • Step 2 Use the microwave link to obtain the attenuation characteristic quantities of different types of precipitation particles in the link microwave, and process them to obtain the pre-processed data of the microwave attenuation:
  • a Select the dual-polarization microwave link, and transmit the microwave signal at the selected frequency at the transmitting end.
  • the microwave signal is attenuated when it passes through the precipitation area during the propagation process, and finally the attenuated signal is received at the receiving end.
  • I ⁇ represents the path attenuation rate of the microwave link with the polarization direction ⁇
  • P ⁇ , 1 represents the microwave frequency at the transmitting end
  • P ⁇ , 2 represents the microwave frequency at the receiving end
  • O ⁇ represents the microwave link with the polarization direction ⁇
  • the total attenuation rate of the path, L is the length of the link, the unit is km;
  • O represents the differential attenuation rate of microwaves
  • O h represents the vertical polarization attenuation rate
  • O v represents the horizontal polarization attenuation rate
  • step d obtain n-dimensional real number vectors of multiple microwave attenuation data
  • Step 3 Perform spatial transformation on the preprocessed data:
  • Combine R1, R2, ..., Rs to form another data set D2 ⁇ R1, R2, ..., Rs ⁇ , as the unlabeled target domain;
  • Weighted resampling is performed on the source domain samples to approximate the distribution of the target domain:
  • S 1 represents the first feature subspace
  • S 2 represents the second feature subspace
  • P 1 is the vertical complement subspace of S 1
  • U 1 and U 2 are d ⁇ d and (nd) ⁇ d, respectively
  • S 1 ' Represents the transposed matrix of S 1
  • F(v) and E(v) are both d-order diagonal matrices
  • the diagonal element of F(v) is cos( ⁇ i)
  • the diagonal element of E(v) is sin( ⁇ i),0 ⁇ i ⁇ d, i ⁇ 1,2, whil,d ⁇ , ⁇ i geometrically represents the geometric angle of the basis vectors in S 1 and S 2 ; in this case, U 1 , U 2 , Both F(v) and E(v) can be calculated, the parameterized function of the curve
  • g(v) represents the basis of a subspace at each point, and sum the inner product of the infinite-dimensional Hilbert space represented by g(v), that is, the inner product ⁇ g(v),g( v)> is the integration of g(v)'g(v) in the interval [0,1] to get:
  • ⁇ 1, ⁇ 2, ⁇ 3 are the diagonal matrix of d ⁇ d, and the diagonal elements are respectively ⁇ 1i, ⁇ 2i, ⁇ 3i, where i ⁇ 1,2,...,d ⁇ , ⁇ 1i , ⁇ 2i
  • ⁇ 3i is:
  • G is a positive semi-definite matrix, because it is a kernel matrix, the kernel function can be defined as follows
  • Step 5 Use the manifold regularization learning method, take the kernel function determined in step 4 as the kernel function, use D1 as the training sample set, train the classifier, and classify the rain, snow, and hail based on the training classifier obtained from the radar data.
  • Commonly used manifold regularization classification methods are mostly used for binary classification, and the present invention is used for the classification of multiple precipitation particle types. Therefore, during training, samples of unprecipitation, rain, snow, and hail are classified into one category in sequence.
  • each test has a result f * 1 (x), f * 2 (x), f * 3 (x), f * 4 (x), so the final result is the largest of these four values as the classification result.
  • the following methods are adopted for the establishment and training of each classifier:
  • V(Xi,yi,f) is the cost function
  • this trainer uses the Hinge function, namely:
  • V(X i ,Y i ,f) max(0,1-Y i f(X i ))
  • I the norm of the regenerated Hilbert space induced by the kernel function K
  • ⁇ A is a user-defined parameter
  • Step 6 Select samples for classifier adjustment from the target domain, so a subset of D2 must be selected for unsupervised learning, so that the selected subset can provide the classifier with incremental knowledge to adapt to the target domain:
  • ⁇ (Xi) represents the high-dimensional characteristic function of the sample, which is an implicit function without a specific expression, and its expression is not used in the subsequent calculations;
  • represents the norm of the Hilbert space; Use the above formula as the basis formula for finding the most similar subset;
  • ⁇ (Xi) represents the high-dimensional feature function of the sample
  • the kernel function determined in step 4 has realized high-dimensional space mapping, so the kernel function corresponding to the norm still uses the kernel function determined in step 4, so :
  • Step 7 Adjust the classifier:
  • the adjacency graphs of the data in D1 and D2 are constructed according to the KNN method, where the distance between samples is defined as follows:
  • d (X i, R j ) represents the distance between the X i and R j
  • X i represents the source domain D1 i-th sample vector
  • R j represents the target domain D2 j-th sample vector
  • X j denotes a source
  • R i represents the i-th sample vector in the target domain (the second data set) D2
  • K(X i , R i ) is the result of X i and R i Kernel function
  • K(X j ,R j ) is the kernel function of X j and R j
  • K(X i ,R j ) is the kernel function of X i and R j
  • the adjacency matrix of the data has KNN according to the above formula
  • the method constructs an adjacency matrix M, and adds each row element of the matrix M as the diagonal element of the matrix D.
  • first ⁇ second ⁇ third involved in the embodiments of this application only distinguishes similar objects, and does not represent a specific order for the objects. Understandably, “first ⁇ second ⁇ third” “Three” can be interchanged in specific order or precedence when permitted. It should be understood that the objects distinguished by “first ⁇ second ⁇ third” can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein.

Landscapes

  • Engineering & Computer Science (AREA)
  • Remote Sensing (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Physics & Mathematics (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Electromagnetism (AREA)
  • Theoretical Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

一种基于半监督域适应的雨、雪、冰雹分类监测方法,包括:采用雷达测得电磁波在各类型降水粒子天气下的雷达波反射率,其中,各类型降水粒子包括雨、雪和冰雹(S10);根据各类型降水粒子天气下的雷达波反射率获取各类型降水粒子天气下的预处理数据(S20);根据预处理数据构建携带标签的第一数据集和未携带标签的第二数据集,计算第一数据集的第一协方差矩阵和第二数据集的第二协方差矩阵,根据第一协方差矩阵确定第一特征子空间,根据第二协方差矩阵确定第二特征子空间(S30);再根据第一特征子空间和第二特征子空间确定核函数(S40);根据核函数,以第一数据集作为训练样本集,训练初始分类器(S50);从第二数据集中选择一个子集来对初始分类器进行无监督学习,使得选择的子集能给初始分类器提供增量知识去自适应目标领域(S60);获取进行无监督学习后的初始分类器的目标函数,根据第一数据集和第二数据集确定邻接图,根据邻接图优化目标函数,以确定最终分类器,采用最终分类器对雨、雪和冰雹进行分类(S70)。该方法可以针对雨、雪、冰雹进行准确分类,使相应的分类监测方案具体更高的准确性。

Description

基于半监督域适应的雨、雪、冰雹分类监测方法 技术领域
本发明涉及地面气象探测技术领域,尤其涉及一种基于半监督域适应的雨、雪、冰雹分类监测方法。
背景技术
对于雨季集中、暴雨多发的地区或者国家,降水的时空分布异常是引发洪涝灾害、山体滑坡、泥石流等自然灾害的重要因素,对于降水的研究早已超过了一个科学研究的范围。在对降水进行测量研究时,首先要区分降水粒子的类型——雨、雪、冰雹等。目前,对雨、雪、冰雹的识别主要根据天气雷达体扫数据和双偏振多普勒雷达偏振参量的方法。其中用天气雷达体扫的方法较为简单,但分辨率低,也忽视了降水粒子的微物理特性;而双偏振多普勒雷达分辨率相较于普通天气雷达高,但容易受到干扰从而误差较大。
目前,微波通信网络覆盖广,信号质量高,基本不存在盲区,因此,微波通信网络降雨监测分析技术在我国有很高的推广应用价值。用微波链路的微波衰减特征来进行反演雨、雪、冰雹的滴谱和粒子形状分布,精准度高,监测盲区小,费用也相比于雷达少,理论上非常适用于识别雨、雪、雹这类特殊天气状况。已有的微波识别雨、雪、冰雹类型方法,利用传统的机器学习方法,需要大量的有标签微波衰减数据,但在实际场景中,有降水类型标签的微波衰减数据量往往不是很充足,故针对雨、雪、冰雹的分类监测方案往往存在准确性低的问题。
发明内容
针对以上问题,本发明提出一种基于半监督域适应的雨、雪、冰雹分类监测方法。
为实现本发明的目的,提供一种基于半监督域适应的雨、雪、冰雹分类监测方法,包括如下步骤:
S10,采用雷达测得电磁波在各类型降水粒子天气下的雷达波反射率;其中,各类型降水粒子包括雨、雪和冰雹;
S20,根据各类型降水粒子天气下的雷达波反射率获取各类型降水粒子天气下的预处理数据;
S30,根据预处理数据构建携带标签的第一数据集和未携带标签的第二数据集,计算第一数据集的第一协方差矩阵和第二数据集的第二协方差矩阵,根据第一协方差矩阵确定第一特征子空间,根据第二协方差矩阵确定第二特征子空间;
S40,根据第一特征子空间和第二特征子空间确定核函数;
S50,根据所述核函数,以第一数据集作为训练样本集,训练初始分类器;
S60,从第二数据集中选择一个子集来对所述初始分类器进行无监督学习,使得选择的子集能给初始分类器提供增量知识去自适应目标领域;
S70,获取进行无监督学习后的初始分类器的目标函数,根据第一数据集和第二数据集确定邻接图,根据邻接图优化所述目标函数,以确定最终分类器,采用最终分类器对雨、雪和冰雹进行分类。
在一个实施例中,采用雷达测得电磁波在各类型降水粒子天气下的雷达波反射率包括:
采用双偏振雷达分别在各类型降水粒子天气下测得降水粒子的多组有效水平反射率Z h和垂直反射率Z v,根据各组有效水平反射率Z h和垂直反射率Z v计算各个差分反射率,根据各类型降水粒子对应的差分反射率确定各类型降水粒子的雷达波反射率。
作为一个实施例,差分反射率的确定过程包括:
Figure PCTCN2020136089-appb-000001
其中,A表示差分反射率,Z h表示有效水平反射率,Z v表示垂直反射率。
作为一个实施例,根据各类型降水粒子天气下的雷达波反射率获取各类型降水粒子天气下的预处理数据包括:
S22,计算各个微波链路在各个偏振方向上的路径衰减率;其中,所述路径衰减率的确定公式包括:
Figure PCTCN2020136089-appb-000002
式中,I θ表示偏振方向为θ的微波链路的路径衰减率,P θ,1表示偏振方向为θ的微波链路的发射端微波频率,P θ,2表示偏振方向为θ的微波链路的接收端微波频率,L表示微波链路的长度;
S23,根据不同偏振方向上的路径衰减率计算出微波链路上微波的差分衰减率;所述差分衰减率的确定公式包括:
Figure PCTCN2020136089-appb-000003
式中,O表示微波的差分衰减率,I h表示微波链路的垂直偏振衰减率,I v表示微波链路的水平偏振衰减率;
S25,在各类型降水粒子天气下针对各个微波链路分别执行步骤S22至步骤S23,获取各类型降水粒子分别对应的一组差分衰减率,根据各类型降水粒子分别对应的一组差分衰减率确定各类型降水粒子的预处理数据。
在一个实施例中,所述第一特征子空间和第二特征子空间均为维度为n×d的子空间G n×d
所述根据第一特征子空间和第二特征子空间确定核函数包括:
S41,在第一特征子空间和第二特征子空间上构建一条从S 1到S 2的曲线;所述曲线的参数化函数包括:
Figure PCTCN2020136089-appb-000004
式中,S 1表示第一特征子空间,S 2表示第二特征子空间,P 1是S 1的垂直补子空间,U 1和U 2分别是d×d和(n-d)×d的对角矩阵,U 1由S 1'S 2=U 1F(v)'通过SVD分解得到,U 2由P 1'S 2=-U 2E(v)'通过SVD分解得到,S 1'表示S 1的转置矩阵,F(v)和E(v)都是d阶对角矩阵,F(v)的对角线元素是cos(αi),E(v)的对角线元素是sin(αi),0<i<d,i∈{1,2,……,d},αi在几何上表示S 1和S 2中基向量几何角度;
S43,获取在每个点上代表一个子空间的基底的基底函数g(v),根据基底函数g(v)计算半正定矩阵G,根据半正定矩阵G设置核函数;所述半正定矩阵G的计算过程包括:
Ω=[S 1U 1,-P 1U 2],
Figure PCTCN2020136089-appb-000005
式中,Λ 1、Λ 2和Λ 3分别为d×d的对角矩阵。
作为一个实施例,所述核函数的设置过程包括:
K(X i,R j)=X i'GR j
式中,K(X i,R j)表示X i和R j对应的核函数,X i表示源域D1中第i个样本向量,R j表示目标域D2第j个样本向量。
在一个实施例中,根据所述核函数,以第一数据集作为训练样本集,训练初始分类器包括:
S51,对第一数据集进行归一化处理,设置输入向量X到实数的分类函数为f(X),为,采用流行规则化在再生希尔伯特空间搜索一个初始分类器,获取初始分类器的目标函数;
S52,采用初始分类器分别针对各类型降水粒子进行训练。
上述基于半监督域适应的雨、雪、冰雹分类监测方法,通过采用雷达测得电磁波在各类型降水粒子天气下的雷达波反射率,根据各类型降水粒子天气下的雷达波反射率获取各类型降水粒子天气下的预处理数据,根据预处理数据构建携带标签的第一数据集和未携带标签的第二数据集,计算第一数据集的第一协方差矩阵和第二数据集的第二协方差矩阵,根据第一协方差矩阵确定第一特征子空间,根据第二协方差矩阵确定第二特征子空间,再根据第一特征子空间和第二特征子空间确定核函数,以根据所述核函数,以第一数据集作为训练样本集,训练初始分类器,从第二数据集中选择一个子集来对所述初始分类器进行无监督学习,使得选择的子集能给初始分类器提供增量知识去自适应目标领域,再获取进行无监督学习后的初始分类器的目标函数,根据第一数据集和第二数据集确定邻接图,根据邻接图优化所述目标函数,以确定最终分类器,采用最终分类器对雨、雪和冰雹进行分类,可以针对雨、雪、冰雹进行准确分类,使相应的分类监测方案具体更高的准确性。
附图说明
图1是一个实施例的基于半监督域适应的雨、雪、冰雹分类监测方法流程示意图;
图2是另一个实施例的基于半监督域适应的雨、雪、冰雹分类监测方法流程示意图。
具体实施方式
为了使本申请的目的、技术方案及优点更加清楚明白,以下结合附图及实施例,对本申请进行进一步详细说明。应当理解,此处描述的具体实施例仅仅用以解释本申请,并不用于限定本申请。
在本文中提及“实施例”意味着,结合实施例描述的特定特征、结构或特性可以包含在本申请的至少一个实施例中。在说明书中的各个位置出现该短语并不一定均是指相同的实施例,也不是与其它实施例互斥的独立的或备选的实施例。本领域技术人员显式地和隐式地理解的是,本文所描述的实施例可以与其它实施例相结合。
参考图1所示,图1为一个实施例的基于半监督域适应的雨、雪、冰雹分类监测方法流程示意图,包括如下步骤:
S10,采用雷达测得电磁波在各类型降水粒子天气下的雷达波反射率;其中,各类型降水粒子包括雨、雪和冰雹。
上述雷达波反射率可以通过n维实数向量表征,比如,各类型降水粒子分别可以通过一个n维实数向量表征相应的雷达波反射率。
S20,根据各类型降水粒子天气下的雷达波反射率获取各类型降水粒子天气下的预处理数据。
上述预处理数据可以通过n维实数向量表征,比如,各类型降水粒子分别可以通过一个n维实数向量表征相应的预处理数据
上述步骤具体可以根据各类型降水粒子天气下的雷达波反射率在获取过程中产生的相关参数进行预处理,以得到相应的预处理数据。
S30,根据预处理数据构建携带标签的第一数据集和未携带标签的第二数据集,计算第一数据集的第一协方差矩阵和第二数据集的第二协方差矩阵,根据第一协方差矩阵确定第一特征子空间,根据第二协方差矩阵确定第二特征子空间。
上述步骤可以对预处理数据进行空间变换,以得到第一特征子空间和第二特征子空间。
在一个实施例中,空间变换的过程可以包括:
S31,在S10和S20中得到的多个n维实数向量(如雷达波反射率及相应的预处理数据)中,对X1、X2、……、Xt等向量标记上降水粒子类型的标签,组成第一数据集D1={X1,X2,……,Xt},作为有标注的源域(第一数据集),其中Xi的标签是Yi,Yi的具体标签取值根据所在分类器的不同有所区别。将R1、R2、……、Rs组成第二数据集D2={R1,R2,……,Rs},作为无标签的目标域(第二数据集);
S32,对D1和D2进行PCA主成分分析,得到特征子空间,该步骤还可以包括:
(3-2-1)计算D1和D2的协方差矩阵H1(第一协方差矩阵)和H2(第二协方差矩阵);
(3-2-2)分别计算H1和H2的特征向量以及特征值,将特征值按从大到小的顺序排序,提取最大的d个特征值,其对应的特征向量为特征子空间的基底,构成特征子空间S 1(第一特征子空间)和S 2(第二特征子空间),其维度为n×d,若记R n的所有d维子空间为G n×d,称为格拉斯曼流型,则S 1和S 2都包含于格拉斯曼流型。
S40,根据第一特征子空间和第二特征子空间确定核函数。
上述步骤可以对源域样本进行加权重采样,从而逼近目标域的分布,实现相应样本自适应。
S50,根据所述核函数,以第一数据集作为训练样本集,训练初始分类器。
上述步骤可以利用流形规则化学习方法,根据核函数,把D1作为训练样本集,训练分类器,根据雷达数据得到的训练分类器进行雨、雪、冰雹分类。常用的流形规则化分类方法多用于二分类法,而实施例可以用于多种降水粒子类型的分类,故在训练的时候依次把未降水、雨、雪、冰雹的样本归为一类,其他剩余的样本归为另一类,这样各个类别的样本就构造出相应分类器。以依次把未降水、雨、雪、冰雹的样本归为一类,其他剩余的样本归为另一类为例。此时第一个分类器中未降水的标签Yi=1,其他三类标签Yi=-1;第二个分类器中降雨的标签Yi=1,其他三类标签Yi=-1;第三个分类器中降雪的标签Yi=1,其他三类标签Yi=-1;第四个分类器中下冰雹的标签Yi=1,其他三类标签Yi=-1。使用这四个分类器分别进行训练,然后的得到四个训练结果。在测试的时候,把对应的雷达波反射数据向量分别利用这四个训练过的分类器进行测试,最后每个测试都有一个结果f * 1(x),f * 2(x),f * 3(x),f * 4(x),于是最终的结果便是这四个值中最大的一个作为分类结果。
S60,从第二数据集中选择一个子集来对所述初始分类器进行无监督学习,使得选择的子集能给初始分类器提供增量知识去自适应目标领域。
S70,获取进行无监督学习后的初始分类器的目标函数,根据第一数据集和第二数据集确定邻接图,根据邻接图优化所述目标函数,以确定最终分类器,采用最终分类器对雨、雪和冰雹进行分类。
在一个示例中,上述步骤可以包括如下过程:
S71,对第二数据集D2进行归一化处理,优化初始分类器,优化的目标函数为:
Figure PCTCN2020136089-appb-000006
式子中,F=(f(X 1),f(X 2),……,f(X t),f(R 1),f(R 2),……,f(R p)),Π表示D1和D2中数据邻接矩阵的拉普拉斯变换,γ B是用户定义的参数;Xi表示代表雷达波反射的n维向量,Yi表示Xi的标签向量,V(Xi,Yi,f)表示代价函数,t表示有标签的雷达波反射率向量的个数,p表示从第二数据集中选择出的子集中向量的个数,
Figure PCTCN2020136089-appb-000007
表示核函数诱导的再生希尔伯特空间的范数;
S72,D1和D2中数据的邻接图按照KNN方法来构造,其中,样本间的距离按照如下定义:
d(X i,R j)=K(X i,R i)+K(X j,R j)-2K(X i,R j)
其中,d(X i,R j)表示X i和R j间的距离,X i表示源域D1中第i个样本向量,R j表示目标域D2中第j个样本向量,X j表示源域(第一数据集)D1中第j个样本向量,R i表示目标域(第二数据集)D2中第i个样本向量,K(X i,R i)为由X i和R i的核函数,K(X j,R j)为由X j和R j的核函数,K(X i,R j)为由X i和R j的核函数,数据的邻接矩阵按照上式有KNN方法构造得到邻接矩阵M,将矩阵M中每行元素相加作为矩阵D的对角元素,D的其他元素为0,则Π=D-M;
S73,利用S72中得到的Π优化f *,得到最终的分类器f r *,r∈{1,2,3,4},设r=1,在区分降水和不降水时,将实时测到的微波衰减数据向量经过处理,最终带入区分降水和不降水分类器的,f * 1>0时表示降水,f * 1<0时表示不降水,其它几个分类器输出结果与其相似,最终在四个分类器输出值中,值最大的那一个分类器结果作为通过微波衰减数据得到的最终分类结果。
上述基于半监督域适应的雨、雪、冰雹分类监测方法,通过采用雷达测得电磁波在各类型降水粒子天气下的雷达波反射率,根据各类型降水粒子天气下的雷达波反射率获取各类型降水粒子天气下的预处理数据,根据预处理数据构建携带标签的第一数据集和未携带标签的第二数据集,计算第一数据集的第一协方差矩阵和第二数据集的第二协方差矩阵,根据第一协方差矩阵确定第一特征子空间,根据第二协方差矩阵确定第二特征 子空间,再根据第一特征子空间和第二特征子空间确定核函数,以根据所述核函数,以第一数据集作为训练样本集,训练初始分类器,从第二数据集中选择一个子集来对所述初始分类器进行无监督学习,使得选择的子集能给初始分类器提供增量知识去自适应目标领域,再获取进行无监督学习后的初始分类器的目标函数,根据第一数据集和第二数据集确定邻接图,根据邻接图优化所述目标函数,以确定最终分类器,采用最终分类器对雨、雪和冰雹进行分类,可以针对雨、雪、冰雹进行准确分类,使相应的分类监测方案具体更高的准确性。
在一个实施例中,采用雷达测得电磁波在各类型降水粒子天气下的雷达波反射率包括:
采用双偏振雷达分别在各类型降水粒子天气下测得降水粒子的多组有效水平反射率Z h和垂直反射率Z v,根据各组有效水平反射率Z h和垂直反射率Z v计算各个差分反射率,根据各类型降水粒子对应的差分反射率确定各类型降水粒子的雷达波反射率。
上述雷达波反射率可以通过n维实数向量表征,比如某类型降水粒子的n维实数向量包括该类型降水粒子对应的各个差分反射率。
作为一个实施例,差分反射率的确定过程包括:
Figure PCTCN2020136089-appb-000008
其中,A表示差分反射率,Z h表示有效水平反射率,Z v表示垂直反射率。
在一个示例中,可以在某个时刻,获取到多组的雷达电磁波反射数据(假设监测区内有n组雷达装置在工作),得到多组差分有效反射率,可以得到一个n维实数向量,即X=(A 1,A 2,……,A n)∈R n,其中A i是第i组雷达装置获取的差分反射率。在不同的时刻(所选取的时刻要包含雨,雪,冰雹天气,以保证足够的代表性),以得到多个n维实数向量。
作为一个实施例,根据各类型降水粒子天气下的雷达波反射率获取各类型降水粒子天气下的预处理数据包括:
S22,计算各个微波链路在各个偏振方向上的路径衰减率;其中,所述路径衰减率的确定公式包括:
Figure PCTCN2020136089-appb-000009
式中,I θ表示偏振方向为θ的微波链路的路径衰减率,P θ,1表示偏振方向为θ的微波链路的发射端微波频率,P θ,2表示偏振方向为θ的微波链路的接收端微波频率,L表示微波链路的长度;
S23,根据不同偏振方向上的路径衰减率计算出微波链路上微波的差分衰减率;所述差分衰减率的确定公式包括:
Figure PCTCN2020136089-appb-000010
式中,O表示微波的差分衰减率,I h表示微波链路的垂直偏振衰减率,I v表示微波链路的水平偏振衰减率;
S25,在各类型降水粒子天气下针对各个微波链路分别执行步骤S22至步骤S23,获取各类型降水粒子分别对应的一组差分衰减率,根据各类型降水粒子分别对应的一组差分衰减率确定各类型降水粒子的预处理数据。
进一步地,在步骤S22之前,还可以包括
S21,选定双偏振微波链路,在发射端用选定好的频率(例如偏振频率为40Hz)发射微波信号,微波信号在传播的过程之中经过降水区域时发生衰减,最终在接收端接收到衰减后的信号,测得水平和垂直链路上的发射端功率以及接收端功率,分别记为P h, a,P h,b,P v,a以及P v,b
在一个实施例中,所述第一特征子空间和第二特征子空间均为维度为n×d的子空间G n×d
所述根据第一特征子空间和第二特征子空间确定核函数包括:
S41,在第一特征子空间和第二特征子空间上构建一条从S 1到S 2的曲线;所述曲线的参数化函数包括:
Figure PCTCN2020136089-appb-000011
式中,S 1表示第一特征子空间,S 2表示第二特征子空间,P 1是S 1的垂直补子空间,U 1和U 2分别是d×d和(n-d)×d的对角矩阵,U 1由S 1'S 2=U 1F(v)'通过SVD分解得到,U 2由P 1'S 2=-U 2E(v)'通过SVD分解得到,S 1'表示S 1的转置矩阵,F(v)和E(v)都是d阶对角矩阵,F(v)的对角线元素是cos(αi),E(v)的对角线元素是sin(αi),0<i<d,i∈{1,2,……,d},αi在几何上S 1和S 2中基向量几何角度;此时,则U 1,U 2,F(v)和E(v)都可以求出,曲线的参数化函数
Figure PCTCN2020136089-appb-000012
也可以由上述公式得到。
S43,获取在每个点上代表一个子空间的基底的基底函数g(v),根据基底函数g(v)计算半正定矩阵G,根据半正定矩阵G设置核函数;所述半正定矩阵G的计算过程包括:
Ω=[S 1U 1,-P 1U 2],
Figure PCTCN2020136089-appb-000013
式中,Λ 1、Λ 2和Λ 3分别为d×d的对角矩阵。
作为一个实施例,所述核函数的设置过程包括:
K(X i,R j)=X i'GR j
式中,K(X i,R j)表示X i和R j对应的核函数,X i表示源域D1中第i个样本向量,R j表示目标域D2第j个样本向量。
具体地,上述步骤S41之后,还可以包括:
S42,令权重函数为w(v)=|1-2v|,其中v∈[0,1],在曲线φ(v)上,当v靠近0或者1时,其对应点表示的子空间越可靠,所以应赋予更高的权重,故用w(t)乘以Φ(t)得到:
g(v)=w(v)φ(v);
上述g(v)在每个点上代表一个子空间的基底,对g(v)表示的无限维的希尔伯特空间的内积求和,即内积<g(v),g(v)>就是g(v)'g(v)在[0,1]区间进行积分。
进一步地,半正定矩阵G的推导过程可以包括:
Figure PCTCN2020136089-appb-000014
由于E(v)和F(v)为对角矩阵,对角元素分别为cos(v.αi)和sin(v.αi),则设:
Ω=[S 1U 1,-P 1U 2],
故有:
Figure PCTCN2020136089-appb-000015
式子中,Λ 1、Λ 2和Λ 3分别为d×d的对角矩阵,Λ 1的对角元素为λ 1i2的对角元素为λ 2i3的对角元素为λ 3i,其中i∈{1,2,……,d},λ 1i2i3i的表达式为:
Figure PCTCN2020136089-appb-000016
Figure PCTCN2020136089-appb-000017
Figure PCTCN2020136089-appb-000018
并且G是一个半正定矩阵,因为是一个核矩阵,其上可以定义核函数如下
K(X i,R j)=X i'GR j
在一个实施例中,根据所述核函数,以第一数据集作为训练样本集,训练初始分类器包括:
S51,对第一数据集进行归一化处理,设置输入向量X到实数的分类函数为f(X),为,采用流行规则化在再生希尔伯特空间搜索一个初始分类器,获取初始分类器的目标函数;
S52,采用初始分类器分别针对各类型降水粒子进行训练。
具体地,上述步骤S51中,对D1进行归一化处理,设分类函数为f(X),为输入向量X到实数的函数,则流行规则化在再生希尔伯特空间搜索一个分类器,其目标函数为:
Figure PCTCN2020136089-appb-000019
其中,V(Xi,Yi,f)为代价函数,本训练器用的是Hinge函数,即:
V(X i,Y i,f)=max(0,1-Y if(X i)),
式中,
Figure PCTCN2020136089-appb-000020
是相应核函数诱导的再生希尔伯特空间的范数,γ A是用户定义的参数,Y i表示Xi的标签。
上述步骤S52中,可以利用S51中得到的分类器f r *,在运用雷达数据建立的f *基础上将微波数据与雷达数据领域自适应可以用微波数据对降水类型分类,r∈{1,2,3,4},设r=1,在区分降水和不降水时,将雷达测到的电磁波反射数据向量经过处理,最终带入该个分类器,f * 1>0时表示降水,f * 1<0时表示不降水,其它几个分类器输出结果与其相似,最终在四个分类器输出值中,值最大的那一个分类器结果作为雷达测得的反射数据进行雨、雪、雹分类的分类结果。
在一个实施例中,上述步骤S60可以从目标域中选择进行分类器调整的样本,故要从D2中选择一个子集来进行无监督学习,使得选择的子集能给分类器提供增量知识去自适应目标领域。该步骤具体可以包括:
S61,分析子集和D1的相似性:DT的任意子集可以用一个0,1的q维向量表示:μ=(μ1,μ2,……,μq),其中μi=1表示Ri在这个子集中,否则不在这个子集中,用最小化均值来表示这个子集的样本与源域D1中样本的相似性,如下:
Figure PCTCN2020136089-appb-000021
式子中,φ(Xi)表示样本的高维特征函数,是一个隐函数,没有具体表达式,后面的计算未用到其表达式;m表示子集的样本个数(m=μ1+μ2+……+μq);|| ||表示希尔伯特空间的范数;
Figure PCTCN2020136089-appb-000022
用上述公式作为寻求最相似子集的依据公式;
S62,φ(Xi)表示样本的高维特征函数,而核函数已经实现了高维空间映射,故设范数对应的核函数仍然用S40确定的核函数,故有:
Figure PCTCN2020136089-appb-000023
再令A=(K(R i,R j)) s×s是D2的样本核矩阵,B=(K(X i,R j)) t×s是D1到D2的样本核矩阵,可将上述寻求最相似子集的依据公式化简为:
Figure PCTCN2020136089-appb-000024
式子中,T=(1,1,……,1)是一个q维向量,上式是一个二次优化问题,其约束条件为0≤αi,且α 12+……+α q=1,因此可用二次优化解出最优的α;
S63,设集合Z={R jj≥τ,R j∈D2},其中,参数τ是自行设定的参数,D2=Z,即在目标域中选定Z中元素,得到集合最终集合还是用D2表示,这个集合被用来作为半监督学习中的无标记样本,计筛选后的D2中样本个数为p个,设D2={R1,R2,……,Rp}。
上述基于半监督域适应的雨、雪、冰雹分类监测方法,通过半监督领域自适应的方法,通过有充足降水类型数据标签的雷达波反射数据建立分类器,用无标签的微波链路优化分类器,最终得到利用微波通过雨、雪、冰雹粒子时发生的衰减率不尽相同的原理,用微波的衰减数据即可以得到降水粒子类型的分类器,有着以下的优异处:
(1)充分利用雨、雪、冰雹粒子对微波造成的不同的衰减影响,对雨、雪、雹粒子进行分类监测;
(2)采用了半监督领域自适应的方法,充分挖掘有标签数据和无标签数据的深层特征,减小对有标签微波衰减数据的需求量;
在建立雨、雪、冰雹粒子分类器时,同时建立多个二分类分类器,使用多个分类器分别进行训练,然后的得到多个个训练结果,选择输出结果最大的分类器分类结果作为最终的分类结果,训练多个分类器同时工作,可加快分类速度。
在一个实施例中,上述基于半监督域适应的雨、雪、冰雹分类监测方法也可以参考图2所示,具体步骤如下:
步骤一:利用雷达测得电磁波在不同类型降水粒子天气下的反射率,加以处理得到电磁波反射的预处理数据:
a、选定双偏振雷达,测得降水时的水平反射率Z h和垂直反射率Z v
b、计算雷达电磁波的差分反射率为:
Figure PCTCN2020136089-appb-000025
公式中,A表示雷达电磁波的差分反射率,Z h表示水平偏振反射率,Z v表示垂直 偏振反射率;
c、在某个时刻,利用监测区内按照步骤a、b可以获取到多组的雷达电磁波反射数据(假设监测区内有n组雷达装置在工作),得到多组差分有效反射率,可以得到一个n维实数向量,即X=(A 1,A 2,……,A n)∈R n,其中A i是第i组雷达装置获取的差分反射率;
d、在不同的时刻(所选取的时刻要包含雨,雪,冰雹天气,以保证足够的代表性),通过步骤c,得到多个n维实数向量;
步骤二:利用微波链路得到不同类型降水粒子在该链路微波中的衰减特征量,加以处理以得到微波衰减的预处理数据:
a、选定双偏振微波链路,在发射端用选定好的频率发射微波信号,微波信号在传播的过程之中经过降水区域时发生衰减,最终在接收端接收到衰减后的信号,测得水平和垂直链路上的发射端功率以及接收端功率,分别记为P h,a,P h,b,P v,a以及P v,b
b、计算出偏振方向上路径衰减率为:
Figure PCTCN2020136089-appb-000026
式中,I θ表示偏振方向为θ的微波链路的路径衰减率,P θ,1表示发射端微波频率,P θ,2表示接收端微波频率,O θ表示偏振方向为θ的微波链路的路径总衰减率,L为链路的长度,单位为km;
c、根据不同偏振方向上的有效衰减率计算出微波链路上微波的差分衰减率为:
Figure PCTCN2020136089-appb-000027
式中,O表示微波的差分衰减率,O h表示垂直偏振衰减率,O v表示水平偏振衰减率;
d、在某个时刻,利用监测区内多条链路(本实施例计区域内有n条链路),按照步骤a、b、c获取微波衰减数据,可以得到一个n维实数向量,即R=(O 1,O 2,……,O n)∈R n,其中O i是第i条链路所测得的差分衰减率;
e、在不同的时刻(所选取的时刻要包含雨,雪,冰雹天气,以保证足够的代表性),通过步骤d,得到多个微波衰减数据的n维实数向量;
步骤三:对预处理后的数据进行空间变换:
a、在步骤一和步骤二中得到的多个n维实数向量中,对X1,X2,……,Xt标记上降 水粒子类型的标签,组成一个数据集D1={X1,X2,……,Xt},作为有标注的源域,其中Xi的标签是Yi,Yi的具体标签取值根据所在分类器的不同有所区别,详见步骤五。将R1,R2,……,Rs组成另一个数据集D2={R1,R2,……,Rs},作为无标签的目标域;
b、对D1和D2进行PCA主成分分析,得到相应的特征子空间:
(b1)计算D1和D2的协方差矩阵H1和H2;
(b2)计算H1和H2的特征向量以及特征值,将特征值按从大到小的顺序排序,提取最大的d个特征值,其对应的特征向量为特征子空间的基底,构成特征子空间S 1和S 2,其维度为n×d,若记R n的所有d维子空间为G n×d,称为格拉斯曼流型,则S1和S2都包含于格拉斯曼流型;
步骤四:样本自适应:
对源域样本进行加权重采样,从而逼近目标域的分布:
a、在Gn×d上构建一条从S 1到S 2的曲线,设曲线的参数化函数
Figure PCTCN2020136089-appb-000028
式子中,S 1表示第一特征子空间,S 2表示第二特征子空间,P 1是S 1的垂直补子空间,U 1和U 2分别是d×d和(n-d)×d的对角矩阵,U 1由S 1'S 2=U 1F(v)'通过SVD分解得到,U 2由P 1'S 2=-U 2E(v)'通过SVD分解得到,S 1'表示S 1的转置矩阵,F(v)和E(v)都是d阶对角矩阵,F(v)的对角线元素是cos(αi),E(v)的对角线元素是sin(αi),0<i<d,i∈{1,2,……,d},αi在几何上表示S 1和S 2中基向量几何角度;此时,则U 1,U 2,F(v)和E(v)都可以求出,曲线的参数化函数
Figure PCTCN2020136089-appb-000029
也可以由上述公式得到;
b、令权重函数为w(v)=|1-2v|,其中v∈[0,1],在曲线φ(v)上,当v靠近0或者1时,其对应点表示的子空间越可靠,所以应赋予更高的权重,故用w(t)乘以Φ(t)得到:
g(v)=w(v)φ(v);
c、g(v)在每个点上代表一个子空间的基底,对g(v)表示的无限维的希尔伯特空间的内积求和,即内积<g(v),g(v)>就是g(v)'g(v)在[0,1]区间进行积分,得到:
Figure PCTCN2020136089-appb-000030
由于E(v)和F(v)为对角矩阵,对角元素分别为cos(v.αi)和sin(v.αi),则设:
Ω=[S 1U 1,-P 1U 2]
故有:
Figure PCTCN2020136089-appb-000031
式子中,Λ1,Λ2,Λ3是d×d的对角矩阵,设其对角元素分别为λ1i,λ2i,λ3i,其中i∈{1,2,……,d},λ 1i2i3i的表达式为:
Figure PCTCN2020136089-appb-000032
Figure PCTCN2020136089-appb-000033
Figure PCTCN2020136089-appb-000034
并且G是一个半正定矩阵,因为是一个核矩阵,其上可以定义核函数如下
K(X i,R j)=X i'GR j
步骤五:利用流形规则化学习方法,以步骤四中确定的核函数作为核函数,把D1作为训练样本集,训练分类器,根据雷达数据得到的训练分类器进行雨、雪、冰雹分类。常用的流形规则化分类方法多用于二分类法,而本发明要用于多种降水粒子类型的分类,故在训练的时候依次把未降水、雨、雪、冰雹的样本归为一类,其他剩余的样本归为另一类,这样4个类别的样本就构造出了4个分类器,第一个分类器中未降水的标签Yi=1,其他三类标签Yi=-1;第二个分类器中降雨的标签Yi=1,其他三类标签Yi=-1;第三个分类器中降雪的标签Yi=1,其他三类标签Yi=-1;第四个分类器中下冰雹的标签Yi=1,其他三类标签Yi=-1。使用这四个分类器分别进行训练,然后的得到四个训练结果。在测试的时候,把对应的雷达波反射数据向量分别利用这四个训练过的分类器进行测试,最后每个测试都有一个结果f * 1(x),f * 2(x),f * 3(x),f * 4(x),于是最终的结果便是这四个值中最大的一个作为分类结果。每个分类器的建立及训练都采取以下方法:
a、对D1进行归一化处理,设分类函数为f(X),为输入向量X到实数的函数,则流行规则化在再生希尔伯特空间搜索一个分类器,其目标函数为:
Figure PCTCN2020136089-appb-000035
其中,V(Xi,yi,f)为代价函数,本训练器用的是Hinge函数,即:
V(X i,Y i,f)=max(0,1-Y if(X i))
式中,
Figure PCTCN2020136089-appb-000036
是核函数K诱导的再生希尔伯特空间的范数,γ A是用户定义的参数;
b、利用a中得到的分类器f r *,r∈{1,2,3,4},设r=1,在区分降水和不降水时,将雷达测到的电磁波反射数据向量经过处理,最终带入该个分类器,f * 1>0时表示降水,f * 1<0时表示不降水,其它几个分类器输出结果与其相似,最终在四个分类器输出值中,值最大的那一个分类器结果作为雷达测得的反射数据进行雨、雪、雹分类的分类结果。
步骤六:从目标域中选择进行分类器调整的样本,故要从D2中选择一个子集来进行无监督学习,使得选择的子集能给分类器提供增量知识去自适应目标领域:
a、分析子集和D1的相似性:DT的任意子集可以用一个0,1的q维向量表示:μ=(μ1,μ2,……,μq),其中μi=1表示Ri在这个子集中,否则不在这个子集中,用最小化均值来表示这个子集的样本与源域D1中样本的相似性,如下:
Figure PCTCN2020136089-appb-000037
式子中,φ(Xi)表示样本的高维特征函数,是一个隐函数,没有具体表达式,后面的计算未用到其表达式;m表示子集的样本个数(m=μ1+μ2+……+μq);|| ||表示希尔伯特空间的范数;
Figure PCTCN2020136089-appb-000038
用上述公式作为寻求最相似子集的依据公式;
b、φ(Xi)表示样本的高维特征函数,而步骤四所确定的核函数已经实现了高维空间映射,故设范数对应的核函数仍然用步骤四中确定的核函数,故有:
Figure PCTCN2020136089-appb-000039
再令A=(K(R i,R j)) s×s是D2的样本核矩阵,B=(K(X i,R j)) t×s是D1到D2的样本核矩阵,可将上述寻求最相似子集的依据公式化简为:
Figure PCTCN2020136089-appb-000040
式子中,T=(1,1,……,1)是一个q维向量,上式是一个二次优化问题,其约束条件为0≤αi,且α 12+……+α q=1,因此可用二次优化解出最优的α;
c、设集合Z={R jj≥τ,R j∈D2},其中,参数τ是自行设定的参数,D2=Z,即在目标域中选定Z中元素,得到集合最终集合还是用D2表示,这个集合被用来作为半监督学习中的无标记样本,计筛选后的D2中样本个数为p个,设D2={R1,R2,……,Rp};
步骤七:调整分类器:
a、对D2进行归一化处理,优化步骤五中得到的雷达数据分类雨、雪、雹的分类函数,优化的目标函数为:
Figure PCTCN2020136089-appb-000041
式子中,F=(f(X 1),f(X 2),……,f(X t),f(R 1),f(R 2),……,f(R p)),Π是D1和D2中数据邻接矩阵的拉普拉斯变换,γ B是用户定义的参数;
b、D1和D2中数据的邻接图按照KNN方法来构造,其中,样本间的距离按照如下定义:
d(X i,R j)=K(X i,R i)+K(X j,R j)-2K(X i,R j)
其中,d(X i,R j)表示X i和R j间的距离,X i表示源域D1中第i个样本向量,R j表示目标域D2中第j个样本向量,X j表示源域(第一数据集)D1中第j个样本向量,R i表示目标域(第二数据集)D2中第i个样本向量,K(X i,R i)为由X i和R i的核函数,K(X j,R j)为由X j和R j的核函数,K(X i,R j)为由X i和R j的核函数,数据的邻接矩阵按照上式有KNN方法构造得到邻接矩阵M,将矩阵M中每行元素相加作为矩阵D的对角元素,D的其他元素为0,则Π=D-M;
c、利用b中得到的L优化f *,得到最终的分类器f r *,r∈{1,2,3,4},设r=1,在区分降水和不降水时,将实时测到的微波衰减数据向量经过处理,最终带入该个分类器,f * 1>0时表示降水,f * 1<0时表示不降水,其它几个分类器输出结果与其相似,最终在四 个分类器输出值中,值最大的那一个分类器结果作为通过微波衰减数据得到的最终分类结果。
以上实施例的各技术特征可以进行任意的组合,为使描述简洁,未对上述实施例中的各个技术特征所有可能的组合都进行描述,然而,只要这些技术特征的组合不存在矛盾,都应当认为是本说明书记载的范围。
需要说明的是,本申请实施例所涉及的术语“第一\第二\第三”仅仅是区别类似的对象,不代表针对对象的特定排序,可以理解地,“第一\第二\第三”在允许的情况下可以互换特定的顺序或先后次序。应该理解“第一\第二\第三”区分的对象在适当情况下可以互换,以使这里描述的本申请的实施例能够以除了在这里图示或描述的那些以外的顺序实施。
本申请实施例的术语“包括”和“具有”以及它们任何变形,意图在于覆盖不排他的包含。例如包含了一系列步骤或模块的过程、方法、装置、产品或设备没有限定于已列出的步骤或模块,而是可选地还包括没有列出的步骤或模块,或可选地还包括对于这些过程、方法、产品或设备固有的其它步骤或模块。
以上所述实施例仅表达了本申请的几种实施方式,其描述较为具体和详细,但并不能因此而理解为对发明专利范围的限制。应当指出的是,对于本领域的普通技术人员来说,在不脱离本申请构思的前提下,还可以做出若干变形和改进,这些都属于本申请的保护范围。因此,本申请专利的保护范围应以所附权利要求为准。

Claims (7)

  1. 一种基于半监督域适应的雨、雪、冰雹分类监测方法,其特征在于,包括如下步骤:
    S10,采用雷达测得电磁波在各类型降水粒子天气下的雷达波反射率;其中,各类型降水粒子包括雨、雪和冰雹;
    S20,根据各类型降水粒子天气下的雷达波反射率获取各类型降水粒子天气下的预处理数据;
    S30,根据预处理数据构建携带标签的第一数据集和未携带标签的第二数据集,计算第一数据集的第一协方差矩阵和第二数据集的第二协方差矩阵,根据第一协方差矩阵确定第一特征子空间,根据第二协方差矩阵确定第二特征子空间;
    S40,根据第一特征子空间和第二特征子空间确定核函数;
    S50,根据所述核函数,以第一数据集作为训练样本集,训练初始分类器;
    S60,从第二数据集中选择一个子集来对所述初始分类器进行无监督学习,使得选择的子集能给初始分类器提供增量知识去自适应目标领域;
    S70,获取进行无监督学习后的初始分类器的目标函数,根据第一数据集和第二数据集确定邻接图,根据邻接图优化所述目标函数,以确定最终分类器,采用最终分类器对雨、雪和冰雹进行分类。
  2. 根据权利要求1所述的基于半监督域适应的雨、雪、冰雹分类监测方法,其特征在于,采用雷达测得电磁波在各类型降水粒子天气下的雷达波反射率包括:
    采用双偏振雷达分别在各类型降水粒子天气下测得降水粒子的多组有效水平反射率Z h和垂直反射率Z v,根据各组有效水平反射率Z h和垂直反射率Z v计算各个差分反射率,根据各类型降水粒子对应的差分反射率确定各类型降水粒子的雷达波反射率。
  3. 根据权利要求2所述的基于半监督域适应的雨、雪、冰雹分类监测方法,其特征在于,差分反射率的确定过程包括:
    Figure PCTCN2020136089-appb-100001
    其中,A表示差分反射率,Z h表示有效水平反射率,Z v表示垂直反射率。
  4. 根据权利要求2所述的基于半监督域适应的雨、雪、冰雹分类监测方法,其特征在于,根据各类型降水粒子天气下的雷达波反射率获取各类型降水粒子天气下的预处理数据包括:
    S22,计算各个微波链路在各个偏振方向上的路径衰减率;其中,所述路径衰减率的确定公式包括:
    Figure PCTCN2020136089-appb-100002
    式中,I θ表示偏振方向为θ的微波链路的路径衰减率,P θ,1表示偏振方向为θ的微波链路的发射端微波频率,P θ,2表示偏振方向为θ的微波链路的接收端微波频率,L表示微波链路的长度;
    S23,根据不同偏振方向上的路径衰减率计算出微波链路上微波的差分衰减率;所述差分衰减率的确定公式包括:
    Figure PCTCN2020136089-appb-100003
    式中,O表示微波的差分衰减率,I h表示微波链路的垂直偏振衰减率,I v表示微波链路的水平偏振衰减率;
    S25,在各类型降水粒子天气下针对各个微波链路分别执行步骤S22至步骤S23,获取各类型降水粒子分别对应的一组差分衰减率,根据各类型降水粒子分别对应的一组差分衰减率确定各类型降水粒子的预处理数据。
  5. 根据权利要求1所述的基于半监督域适应的雨、雪、冰雹分类监测方法,其特征在于,所述第一特征子空间和第二特征子空间均为维度为n×d的子空间G n×d
    所述根据第一特征子空间和第二特征子空间确定核函数包括:
    S41,在第一特征子空间和第二特征子空间上构建一条从S 1到S 2的曲线;所述曲线的参数化函数包括:
    Figure PCTCN2020136089-appb-100004
    式中,S 1表示第一特征子空间,S 2表示第二特征子空间,P 1是S 1的垂直补子空间,U 1和U 2分别是d×d和(n-d)×d的对角矩阵,U 1由S 1'S 2=U 1F(v)'通过SVD分解得到, U 2由P 1'S 2=-U 2E(v)'通过SVD分解得到,S 1'表示S 1的转置矩阵,F(v)和E(v)都是d阶对角矩阵,F(v)的对角线元素是cos(αi),E(v)的对角线元素是sin(αi),0<i<d,i∈{1,2,……,d},αi表示S 1和S 2中基向量几何角度;
    S43,获取在每个点上代表一个子空间的基底的基底函数g(v),根据基底函数g(v)计算半正定矩阵G,根据半正定矩阵G设置核函数;所述半正定矩阵G的计算过程包括:
    Ω=[S 1U 1,-P 1U 2],
    Figure PCTCN2020136089-appb-100005
    式中,Λ 1、Λ 2和Λ 3分别为d×d的对角矩阵。
  6. 根据权利要求5所述的基于半监督域适应的雨、雪、冰雹分类监测方法,其特征在于,所述核函数的设置过程包括:
    K(X i,R j)=X i'GR j
    式中,K(X i,R j)表示X i和R j对应的核函数,X i表示源域D1中第i个样本向量,R j表示目标域D2第j个样本向量。
  7. 根据权利要求1所述的基于半监督域适应的雨、雪、冰雹分类监测方法,其特征在于,根据所述核函数,以第一数据集作为训练样本集,训练初始分类器包括:
    S51,对第一数据集进行归一化处理,设置输入向量X到实数的分类函数为f(X),为,采用流行规则化在再生希尔伯特空间搜索一个初始分类器,获取初始分类器的目标函数;
    S52,采用初始分类器分别针对各类型降水粒子进行训练。
PCT/CN2020/136089 2020-02-12 2020-12-14 基于半监督域适应的雨、雪、冰雹分类监测方法 WO2021159844A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010088477.5A CN111308471B (zh) 2020-02-12 2020-02-12 基于半监督域适应的雨、雪、冰雹分类监测方法
CN202010088477.5 2020-02-12

Publications (1)

Publication Number Publication Date
WO2021159844A1 true WO2021159844A1 (zh) 2021-08-19

Family

ID=71156484

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/136089 WO2021159844A1 (zh) 2020-02-12 2020-12-14 基于半监督域适应的雨、雪、冰雹分类监测方法

Country Status (2)

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

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114706146A (zh) * 2022-03-23 2022-07-05 成都信息工程大学 复杂地形下雹暴过程中雹胚的生长和降雹阶段的预报方法
CN116911082A (zh) * 2023-09-14 2023-10-20 成都信息工程大学 基于降水雷达和同化资料的降水粒子质量和数量估算方法
CN117173350A (zh) * 2023-08-09 2023-12-05 中国科学技术大学 基于主动领域适应学习的地质建模方法、系统及介质

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111308471B (zh) * 2020-02-12 2020-11-24 河海大学 基于半监督域适应的雨、雪、冰雹分类监测方法
CN112733859B (zh) * 2021-01-25 2023-12-19 重庆大学 一种组织病理学图像的深度迁移半监督域自适应分类方法
CN113095442B (zh) * 2021-06-04 2021-09-10 成都信息工程大学 基于半监督学习在多维度雷达数据下的冰雹识别方法
CN113591387B (zh) * 2021-08-05 2023-09-01 安徽省气象台 基于Huber范数约束的卫星资料反演降水方法及系统
CN114814993B (zh) * 2022-03-25 2023-04-18 河海大学 一种基于dcgan和2d-cnn的微波衰减降雪强度监测方法
CN114692692B (zh) * 2022-04-02 2023-05-12 河海大学 一种基于微波衰减信号融合核极限学习机的降雪识别方法
CN114488160A (zh) * 2022-04-02 2022-05-13 南京师范大学 顾及三维风场影响的雷达降雨估算误差校正方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106096627A (zh) * 2016-05-31 2016-11-09 河海大学 顾及特征优化的极化sar图像半监督分类方法
CN108280394A (zh) * 2017-12-15 2018-07-13 西安电子科技大学 基于时频分布奇异值分解的雷达信号识别方法
CN110161480A (zh) * 2019-06-18 2019-08-23 西安电子科技大学 基于半监督深度概率模型的雷达目标识别方法
CN111308471A (zh) * 2020-02-12 2020-06-19 河海大学 基于半监督域适应的雨、雪、冰雹分类监测方法

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10839309B2 (en) * 2015-06-04 2020-11-17 Accusonus, Inc. Data training in multi-sensor setups
CN106599787B (zh) * 2016-11-17 2019-07-02 河海大学 基于半监督子块联合回归的单样本人脸识别方法
US11269058B2 (en) * 2018-06-13 2022-03-08 Metawave Corporation Autoencoder assisted radar for target identification
CN110728210A (zh) * 2019-09-25 2020-01-24 上海交通大学 一种三维点云数据的半监督目标标注方法和系统

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106096627A (zh) * 2016-05-31 2016-11-09 河海大学 顾及特征优化的极化sar图像半监督分类方法
CN108280394A (zh) * 2017-12-15 2018-07-13 西安电子科技大学 基于时频分布奇异值分解的雷达信号识别方法
CN110161480A (zh) * 2019-06-18 2019-08-23 西安电子科技大学 基于半监督深度概率模型的雷达目标识别方法
CN111308471A (zh) * 2020-02-12 2020-06-19 河海大学 基于半监督域适应的雨、雪、冰雹分类监测方法

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
GREICE MARTINS DE FREITAS ; ANA MARIA HEUMINSKI DE ÁVILA ; JOÃO PAULO PAPA: "Rainfall Estimation Using Transductive Learning", IMAGE AND SIGNAL PROCESSING, 2008. CISP '08. CONGRESS ON, IEEE, PISCATAWAY, NJ, USA, 27 May 2008 (2008-05-27), Piscataway, NJ, USA, pages 631 - 634, XP031287107, ISBN: 978-0-7695-3119-9 *
ZHENG, MINWEI: "The Method Study on Automatic Sea Ice Detection with Gao Fen-3 Synthetic Aperture Radar Data in Polar Regions", ACTA OCEANOLOGICA SINICA, vol. 40, no. 9, 1 September 2018 (2018-09-01), pages 1 - 12, XP055836928 *
ZHU, KEFAN: "Semi-supervised Low-resolution Radar Target Recognition Technology Based on Generative Adversarial Network", JOURNAL OF DETECTION & CONTROL, vol. 41, no. 6, 1 December 2019 (2019-12-01), pages 1 - 7, XP055836926 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114706146A (zh) * 2022-03-23 2022-07-05 成都信息工程大学 复杂地形下雹暴过程中雹胚的生长和降雹阶段的预报方法
CN114706146B (zh) * 2022-03-23 2023-11-03 成都信息工程大学 复杂地形下雹暴过程中雹胚的生长和降雹阶段的预报方法
CN117173350A (zh) * 2023-08-09 2023-12-05 中国科学技术大学 基于主动领域适应学习的地质建模方法、系统及介质
CN116911082A (zh) * 2023-09-14 2023-10-20 成都信息工程大学 基于降水雷达和同化资料的降水粒子质量和数量估算方法
CN116911082B (zh) * 2023-09-14 2023-12-05 成都信息工程大学 基于降水雷达和同化资料的降水粒子质量和数量估算方法

Also Published As

Publication number Publication date
CN111308471A (zh) 2020-06-19
CN111308471B (zh) 2020-11-24

Similar Documents

Publication Publication Date Title
WO2021159844A1 (zh) 基于半监督域适应的雨、雪、冰雹分类监测方法
Zhang et al. Classification of polarimetric SAR image based on support vector machine using multiple-component scattering model and texture features
Pasolli et al. Automatic analysis of GPR images: A pattern-recognition approach
CN107103306B (zh) 基于小波分析与支持向量机的冬小麦白粉病遥感监测方法
CN105184309A (zh) 基于cnn和svm的极化sar图像分类
CN110008905A (zh) 基于遥感影像红边波段的区域尺度小麦条锈病监测方法
JP2019211342A (ja) 気象解析装置、気象解析方法、およびプログラム
Hänsch Complex-valued multi-layer perceptrons–an application to polarimetric SAR data
US20210063322A1 (en) Grain mildew detection method and device based on wifi apparatus
CN115496934A (zh) 一种基于孪生神经网络的高光谱图像分类方法
Ma et al. Land cover classification for polarimetric sar image using convolutional neural network and superpixel
CN112884029B (zh) 融合全极化sar与高光谱遥感的协同分类方法
CN115294009A (zh) 一种基于机器学习检测电池极耳焊接缺陷的方法、设备及存储介质
CN116523352B (zh) 一种森林资源信息的管理方法及系统
P. Thampy et al. A convolution neural network approach to Doppler spectra classification of 205 MHz radar
CN106169086B (zh) 导航数据辅助下的高分辨率光学影像损毁道路提取方法
Wei et al. Texture feature analysis in oil spill monitoring by SAR image
Chang et al. Current and future applications of cost-effective smart cameras in agriculture
CN116482618A (zh) 基于多损失特征自校准网络的雷达有源干扰识别方法
Khobragade et al. Optimization of statistical learning algorithm for crop discrimination using remote sensing data
Bai et al. Supervised feature selection for polarimetric SAR classification
DOUTOUM et al. Classification of Guava Leaf Disease using Deep Learning
Majumder et al. Gabor wavelet based fused texture features for identification of mungbean leaf diseases
Tian et al. Entropy based generative adversarial network for PolSAR image classification
Yan et al. Unsupervised classification of polarimetric SAR images using deep embedding network

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: 20918521

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: 20918521

Country of ref document: EP

Kind code of ref document: A1