CN110456355A - A kind of Radar Echo Extrapolation method based on long short-term memory and generation confrontation network - Google Patents
A kind of Radar Echo Extrapolation method based on long short-term memory and generation confrontation network Download PDFInfo
- Publication number
- CN110456355A CN110456355A CN201910763464.0A CN201910763464A CN110456355A CN 110456355 A CN110456355 A CN 110456355A CN 201910763464 A CN201910763464 A CN 201910763464A CN 110456355 A CN110456355 A CN 110456355A
- Authority
- CN
- China
- Prior art keywords
- model
- image
- input
- term memory
- long short
- 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
Links
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
- G01S13/958—Theoretical aspects
-
- 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
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
- G01S7/417—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section involving the use of neural networks
-
- 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
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
- G01S7/418—Theoretical aspects
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- Physics & Mathematics (AREA)
- Computer Networks & Wireless Communication (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Electromagnetism (AREA)
- Radar Systems Or Details Thereof (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of based on long short-term memory and generates the Radar Echo Extrapolation method of confrontation network, successively the following steps are included: A: obtaining radar data collection and obtains greyscale image data set, division obtains training sample set and test sample collection;B: constructing long short-term memory and generates confrontation network model, obtains convergent long short-term memory and generates confrontation network model;C: it by convergent long short-term memory obtained in the test sample collection input step B in step A and generates in confrontation network model, obtains Radar Echo Extrapolation image.The present invention can effectively predict Radar Echo Extrapolation image, provide technical foundation to close on being effectively predicted for weather forecast.
Description
Technical field
The present invention relates to surface weather observation technical field in Atmospheric Survey, more particularly to it is a kind of based on long short-term memory and
Generate the Radar Echo Extrapolation method of confrontation network.
Background technique
Closing on weather forecast is closed on pre- to the high-spatial and temporal resolution forecast in variation faster weather phenomenon 0-2 hours
The object of report is mainly the diastrous weather that the differentiation such as thunderstorm, strong convection, precipitation and sandstorm are rapid, destructiveness is strong.Currently, thunder
It is the technical way for close on weather forecast up to echo extrapolation technique.
Radar Echo Extrapolation refers to the echo data detected according to weather radar, determines the intensity distribution of echo and returns
The movement speed of wave body (such as storm monomer, precipitation area) and direction, it is linear or nonlinear outer by being carried out to echo body
It pushes away, the radar return state after forecasting certain period of time.As China New Generation Doppler radar observation grid is gradually thrown
How enter operation reduces meteorological disaster utmostly using Doppler radar observation grid progress echo Extrapotated prediction,
As a critically important at present job.
Traditional Radar Echo Extrapolation mainly uses monomer centroid method and cross-correlation technique.Monomer centroid method be suitble to big and
Strong target is tracked and is forecast that weather condition and discomfort when merging or dividing occur for or echo more scattered to echo
With.Cross-correlation technique can track stratiform clouds rainfall system, but change cracking strong convective weather for echo, it is difficult to guarantee
The accuracy of tracking.Therefore needing one kind can be to the method that Radar Echo Extrapolation image is effectively predicted, to close on weather
Offer technical foundation is effectively predicted in forecast.
Summary of the invention
The object of the present invention is to provide a kind of based on long short-term memory and generates the Radar Echo Extrapolation method of confrontation network,
Radar Echo Extrapolation image can effectively be predicted, provide technical foundation to close on being effectively predicted for weather forecast.
The present invention adopts the following technical solutions:
It is a kind of based on long short-term memory and generate confrontation network Radar Echo Extrapolation method, successively the following steps are included:
A: obtaining radar data collection, and the every data concentrated to radar data carries out unified size and sequence is handled, then
Every data that radar data is concentrated is converted to greyscale image data and obtains greyscale image data after normalized
Set, finally is divided to obtain training sample set and test sample collection to greyscale image data set;In training set sample set
Every group of image collection in include input label and authentic specimen label;
B: constructing long short-term memory first and generates confrontation network model, and initializes long short-term memory and generate confrontation net
The weight and biasing of network, long short-term memory and generation fight the generation model in network model by growing Memory Neural Networks structure in short-term
At long short-term memory and the discrimination model generated in confrontation network model are made of full Connection Neural Network;It then will be in step A
Obtained training sample set, which is input to, to be generated in model and obtains forecast image;Forecast image and authentic specimen label are inputted again
Into discrimination model, calculates the average absolute overall error between forecast image and authentic specimen label and generate model and sentence
The penalty values of other model, then long short-term memory is updated by backpropagation and generates the weight and biasing of confrontation network, repeat this
Process terminates until training, obtains convergent long short-term memory and generates confrontation network model;
C: by convergent long short-term memory obtained in the test sample collection input step B in step A and confrontation net is generated
In network model, Radar Echo Extrapolation image is obtained.
The step A includes step in detail below:
A1: radar data collection is obtained, and the N data that radar data is concentrated is ranked up according to time incremental order;
A2: carrying out unified size to every data that radar data is concentrated and image converted, by normalization operation by thunder
The greyscale image data after normalization is converted to up to every data in data set, and obtains greyscale image data set;
A3: dividing greyscale image data set, by every four adjacent gray scales in greyscale image data set
Image obtains i group image collection as one group of image collection, first three width i.e. 4i-3 width, 4i-2 in i-th group of image collection
The 4th width i.e. 4i width gray level image of width and 4i-1 width gray level image as one group of input label, in i-th group of image collection
As authentic specimen label, natural number of the i between [1, N/4] is then drawn acquired i group image collection with the ratio of 7:3
It is divided into training sample set and test sample collection.
The step A2 includes step in detail below:
A21: size conversion is carried out to every data that radar data is concentrated, every data size that radar data is concentrated
It is converted into 360 × 480;
A22: gray level image is converted by the data after the conversion of size obtained in step A21, then again to gray level image
Operation is normalized, finally obtains greyscale image data set.
The step B includes step in detail below:
B1: constructing long short-term memory first and generates confrontation network model, and initializes long short-term memory and generate confrontation
The weight and biasing of network;
B2: and then training set sample set obtained in step A3 is input to and is generated in model, in training set sample set
Every group of image collection includes input label input and authentic specimen label true, wherein input label input={ x1, x2,
x3, authentic specimen label true={ x4};x1, x2, x3First three width i.e. 4i- in respectively step A3 in i-th group of image collection
3 width, 4i-2 width and 4i-1 width gray level image, x4Indicate the 4th width i.e. 4i width gray level image in i-th group of image;
B3: forecast image is obtained by generating model, then forecast image and authentic specimen label are sequentially inputted to differentiate
In model, calculates the average absolute overall error between forecast image and authentic specimen label and generate model and discrimination model
Penalty values;
B4: anti-from discrimination model output layer to discrimination model input layer according to the penalty values for the discrimination model being calculated
To propagation, uses learning rate as input parameter, adjust the weight and biasing of every layer network, finally obtain and update weight and biasing
Discrimination model afterwards;
According to the penalty values for generating model are calculated, reversely passed from model output layer is generated to mode input layer is generated
It broadcasts, uses learning rate as input parameter, adjust the weight and biasing of every layer network, finally obtain after updating weight and biasing
Generate model;
B5: repeating step B2 to B4, until reaching maximum number of iterations completes training, finally obtains convergent length and remembers in short-term
Recall and generate confrontation network model.
In the step B1, the network layer for generating model, which is followed successively by, generates mode input layer, the first long short-term memory
Neural net layer, generates the full articulamentum of model and generates model output layer the second long Memory Neural Networks layer in short-term, generates model
Training the number of iterations be 150 times, learning rate 0.001, the size for generating mode input layer is 3 × 360 × 480, and first is long
The hidden layer number of nodes of short-term memory neural net layer and the second long Memory Neural Networks layer in short-term is 128, and it is complete to generate model
The number of nodes of articulamentum is 360 × 480, and the size for generating model output layer is 360 × 480, generates the final output figure of model
Picture size is 360 × 480, that is, the size of the forecast image exported is 360 × 480;
The network layer of discrimination model is followed successively by discrimination model input layer, the full articulamentum of discrimination model first, discrimination model
Second full articulamentum, the full articulamentum of discrimination model third and discrimination model output layer, the training the number of iterations of discrimination model are 150
Secondary, learning rate 0.001, the number of nodes of the full articulamentum of discrimination model first is 256, the section of the full articulamentum of discrimination model second
Points are 128, and the number of nodes of the full articulamentum of discrimination model third is 1, and the size 360 × 480 of discrimination model input layer is sentenced
The size of other model output layer is 1.
In the step B2, the size of the input label input in every group of image collection in training set sample set is 3
×360×480。
In the step B3, the calculating of the average absolute overall error MAE between forecast image and authentic specimen label is public
Formula are as follows:
Wherein, m indicates the pixel number shared in forecast image and authentic specimen label, trueiIt indicates i-th in step A3
Authentic specimen label in group image collection,Indicate authentic specimen label trueiIn j-th of pixel value, oiIt indicates
Forecast image,Indicate the value of j-th of pixel in forecast image;
Successively forecast image and authentic specimen label are input in discrimination model, discrimination model exports two sizes respectively
Scalar D (G (input between [0,1]i)) and D (truei), the two scalar D (G then exported according to discrimination model
(inputi)) and D (truei), calculate separately the penalty values for generating model and discrimination model;
Generate the loss function of model are as follows:
Wherein, V1For the penalty values for generating model, D indicates that discrimination model to be optimized, G indicate generation mould to be optimized
Type,N is training sample set number, and N is the total number of radar data intensive data, and log indicates log-likelihood
Function, inputiFor i-th group of input sample, G (inputi) it is inputiInput generates the forecast image obtained after model G, D (G
(inputi)) indicate to generate differentiation result of the forecast image of the generation of model after discrimination model D differentiation;
The loss function of discrimination model are as follows:
Wherein, V2For the penalty values of discrimination model, trueiIndicate the authentic specimen label in i-th group of image collection, D
(truei) indicate differentiation result of the authentic specimen label after discrimination model differentiates.
In the step C, the size of Radar Echo Extrapolation image is 360 × 480.
The present invention is based on image processing techniques, by converting radar data collection to the greyscale image data collection after normalization
It closes, then constructs based on long short-term memory and generate confrontation network model, pass through backpropagation and learning rate is used to join as input
Several pairs of weights and biasing are adjusted, and fight network model based on long short-term memory and generation after being trained, finally will instruction
Practice sample set and be input to fighting in network model after training based on long short-term memory and generation, obtains forecast image, Neng Gouwei
The accuracy rate for closing on weather forecast provides technical foundation, and diastrous weather can be effectively predicted.
Detailed description of the invention
Fig. 1 is flow chart of the invention.
Specific embodiment
The present invention is made with detailed description below in conjunction with drawings and examples:
As shown in Figure 1, the Radar Echo Extrapolation method of the present invention based on long short-term memory and generation confrontation network,
Successively comprise the steps of:
A: obtaining radar data collection, and the every data concentrated to radar data carries out unified size and sequence is handled, then
Every data that radar data is concentrated is converted to greyscale image data and obtains greyscale image data after normalized
Set, finally is divided to obtain training sample set and test sample collection to greyscale image data set;In training set sample set
Every group of image collection in include input label and authentic specimen label;
Step A includes step in detail below:
A1: radar data collection is obtained, and the N data that radar data is concentrated is ranked up according to time incremental order;
A2: carrying out unified size to every data that radar data is concentrated and image converted, by normalization operation by thunder
The greyscale image data after normalization is converted to up to every data in data set, and obtains greyscale image data set;
The step A2 comprising the following specific steps
A21: size conversion is carried out to every data that radar data is concentrated, every data size that radar data is concentrated
It is converted into 360 × 480;
A22: gray level image is converted by the data after the conversion of size obtained in step A21, then again to gray level image
Operation is normalized, finally obtains greyscale image data set;
A3: dividing greyscale image data set, by every four adjacent gray scales in greyscale image data set
Image obtains i group image collection as one group of image collection, first three width i.e. 4i-3 width, 4i-2 in i-th group of image collection
The 4th width i.e. 4i width gray level image of width and 4i-1 width gray level image as one group of input label, in i-th group of image collection
As authentic specimen label, natural number of the i between [1, N/4] is then drawn acquired i group image collection with the ratio of 7:3
It is divided into training sample set and test sample collection.
In the present embodiment, when being divided to greyscale image data set, by every four in greyscale image data set
Adjacent gray level image obtains i group image collection, i.e. the first width, the second width, third width and the 4th as one group of image collection
Width gray level image is as the 1st group of image collection, wherein the first width, the second width and third width grayscale image in the 1st group of image collection
As being used as one group of input label, the 4th width gray level image in the 1st group of image collection is as authentic specimen label;5th width,
Six width, the 7th width and the 8th width gray level image are as the 2nd group of image collection, wherein the 5th width in the 2nd group of image collection, the 6th
Width and the 7th width gray level image are as one group of input label, and the 8th width gray level image in the 2nd group of image collection is as true sample
This label;And so on;
B: constructing long short-term memory first and generates confrontation network model, and initializes long short-term memory and generate confrontation net
The weight and biasing of network, long short-term memory and generation fight the generation model in network model by growing Memory Neural Networks structure in short-term
At long short-term memory and the discrimination model generated in confrontation network model are made of full Connection Neural Network;It then will be in step A
Obtained training sample set, which is input to, to be generated in model and obtains forecast image;Again simultaneously by forecast image and authentic specimen label
It is input in discrimination model, calculates the average absolute overall error between forecast image and authentic specimen label and generates model
With the penalty values of discrimination model, then by backpropagation update long short-term memory and generate confrontation network weight and biasing, weight
This multiple process terminates until training, obtains convergent long short-term memory and generates confrontation network model;
The step B includes step in detail below:
B1: constructing long short-term memory first and generates confrontation network model, and initializes long short-term memory and generate confrontation
The weight and biasing of network;
The network layer for generating model is to generate mode input layer, the first long Memory Neural Networks layer, the second length in short-term
When Memory Neural Networks layers, generate the full articulamentum of model and generate model output layer, the training the number of iterations for generating model is 150
Secondary, learning rate 0.001, the size for generating mode input layer is 3 × 360 × 480, the first long Memory Neural Networks layer in short-term and
The hidden layer number of nodes of second long Memory Neural Networks layer in short-term is 128, and the number of nodes for generating the full articulamentum of model is 360 ×
480, the size for generating model output layer is 360 × 480, and the final output image size for generating model is 360 × 480, i.e., defeated
The size of forecast image out is 360 × 480;
The network layer of discrimination model is discrimination model input layer, the full articulamentum of discrimination model first, discrimination model second
Full articulamentum, the full articulamentum of discrimination model third and discrimination model output layer, the training the number of iterations of discrimination model are 150 times,
Learning rate is 0.001, and the number of nodes of the full articulamentum of discrimination model first is 256, the node of the full articulamentum of discrimination model second
Number is 128, and the number of nodes of the full articulamentum of discrimination model third is 1, and the size 360 × 480 of discrimination model input layer differentiates
The size of model output layer is 1.
B2: and then training set sample set obtained in step A3 is input to and is generated in model, in training set sample set
Every group of image collection includes input label input and authentic specimen label true, wherein input label input={ x1, x2,
x3, authentic specimen label true={ x4};x1, x2, x3First three width i.e. 4i- in respectively step A3 in i-th group of image collection
3 width, 4i-2 width and 4i-1 width gray level image, x4Indicate the 4th width i.e. 4i width gray level image in i-th group of image;
In step B2, the size of the input label input in every group of image collection in training set sample set is 3 × 360
×480;
B3: forecast image is obtained by generating model, then forecast image and authentic specimen label are input to discrimination model
In, it calculates the average absolute overall error between forecast image and authentic specimen label and generates the damage of model and discrimination model
Mistake value;
Average absolute in step B3 between authentic specimen label obtained in obtained forecast image and step A3 is total
The calculation formula of error MAE are as follows:
Wherein, m indicates the pixel number shared in forecast image and authentic specimen label, trueiIt indicates i-th in step A3
Authentic specimen label in group image collection,Indicate authentic specimen label trueiIn j-th of pixel value, oiIt indicates
Forecast image,Indicate the value of j-th of pixel in forecast image;
Forecast image and authentic specimen label are input in discrimination model, discrimination model exports two sizes respectively and exists
Scalar D (G (input between [0,1]i)) and D (truei), the two scalar D (G then exported according to discrimination model
(inputi)) and D (truei), calculate separately the penalty values for generating model and discrimination model;
Generate the loss function of model are as follows:
V1For the penalty values for generating model, D indicates that discrimination model to be optimized, G indicate generation model to be optimized,N is training sample set number, and N is the total number of radar data intensive data, and log indicates log-likelihood letter
Number, inputiFor i-th group of input sample, G (inputi) it is inputiInput generates the forecast image obtained after model G, D (G
(inputi)) indicate to generate differentiation result of the forecast image of the generation of model after discrimination model D differentiation;
The loss function of discrimination model are as follows:
Wherein, V2For the penalty values of discrimination model, trueiIndicate the authentic specimen label in i-th group of image collection, D
(truei) indicate differentiation result of the authentic specimen label after discrimination model differentiates;
B4: according to the penalty values V for the discrimination model being calculated2, from discrimination model output layer to discrimination model input layer
Backpropagation, use learning rate as input parameter, adjust the weight and biasing of every layer network, finally obtain update weight and partially
The discrimination model postponed;
According to the penalty values V that generation model is calculated1, reversely passed from model output layer is generated to mode input layer is generated
It broadcasts, uses learning rate as input parameter, adjust the weight and biasing of every layer network, finally obtain after updating weight and biasing
Generate model;
To discrimination model and after generating the update that model completes weight and biasing, that is, complete to long short-term memory and generation
Fight the weight of network and the update of biasing.
B5: repeating step B2 to B4, until reaching maximum number of iterations completes training, finally obtains convergent length and remembers in short-term
Recall and generate confrontation network model;
C: by convergent long short-term memory obtained in the test sample collection input step B in step A and confrontation net is generated
In network model, Radar Echo Extrapolation image is obtained.
The size for the input label input in every group of image collection that test sample is concentrated is 3 × 360 × 480;Radar returns
The size of wave extrapolated image is 360 × 480.
Claims (8)
1. a kind of Radar Echo Extrapolation method based on long short-term memory and generation confrontation network, which is characterized in that successively include
Following steps:
A: obtaining radar data collection, and the every data concentrated to radar data carries out unified size and sequence is handled, then by thunder
Up to every data in data set after normalized, is converted to greyscale image data and obtains greyscale image data collection
It closes, finally greyscale image data set is divided to obtain training sample set and test sample collection;In training set sample set
It include input label and authentic specimen label in every group of image collection;
B: constructing long short-term memory first and generates confrontation network model, and initializes long short-term memory and generate confrontation network
Weight and biasing, long short-term memory and the generation model generated in confrontation network model are constituted by growing Memory Neural Networks in short-term,
Long short-term memory and the discrimination model generated in confrontation network model are made of full Connection Neural Network;Then it will be obtained in step A
To training sample set be input to generate model in and obtain forecast image;Forecast image and authentic specimen label are input to again
In discrimination model, calculates the average absolute overall error between forecast image and authentic specimen label and generate model and differentiation
The penalty values of model, then long short-term memory is updated by backpropagation and generates the weight and biasing of confrontation network, repeat this mistake
Journey terminates until training, obtains convergent long short-term memory and generates confrontation network model;
C: by convergent long short-term memory obtained in the test sample collection input step B in step A and confrontation network mould is generated
In type, Radar Echo Extrapolation image is obtained.
2. the Radar Echo Extrapolation method according to claim 1 based on long short-term memory and generation confrontation network, special
Sign is that the step A includes step in detail below:
A1: radar data collection is obtained, and the N data that radar data is concentrated is ranked up according to time incremental order;
A2: carrying out unified size to every data that radar data is concentrated and image converted, by normalization operation by radar number
The greyscale image data after normalization is converted to according to every data of concentration, and obtains greyscale image data set;
A3: dividing greyscale image data set, by every four adjacent gray level images in greyscale image data set
Obtain i group image collection as one group of image collection, first three width i.e. 4i-3 width, 4i-2 width in i-th group of image collection and
Fourth width i.e. 4i width gray level image conduct of the 4i-1 width gray level image as one group of input label, in i-th group of image collection
Then acquired i group image collection is by authentic specimen label, natural number of the i between [1, N/4] with 7: 3 ratio cut partition
Training sample set and test sample collection.
3. the Radar Echo Extrapolation method according to claim 2 based on long short-term memory and generation confrontation network, special
Sign is that the step A2 includes step in detail below:
A21: size conversion is carried out to every data that radar data is concentrated, every data size that radar data is concentrated is turned
Turn to 360 × 480;
A22: gray level image is converted by the data after the conversion of size obtained in step A21, then gray level image is carried out again
Normalization operation finally obtains greyscale image data set.
4. the Radar Echo Extrapolation method according to claim 3 based on long short-term memory and generation confrontation network, special
Sign is: the step B includes step in detail below:
B1: constructing long short-term memory first and generates confrontation network model, and initializes long short-term memory and generate confrontation network
Weight and biasing;
B2: and then training set sample set obtained in step A3 is input to and is generated in model, every group in training set sample set
Image collection includes input label input and authentic specimen label true, wherein input label input={ x1, x2, x3, very
Real sample label true={ x4};x1, x2, x3First three width i.e. 4i-3 width in respectively step A3 in i-th group of image collection,
4i-2 width and 4i-1 width gray level image, x4Indicate the 4th width i.e. 4i width gray level image in i-th group of image;
B3: forecast image is obtained by generating model, then forecast image and authentic specimen label are sequentially inputted to discrimination model
In, it calculates the average absolute overall error between forecast image and authentic specimen label and generates the damage of model and discrimination model
Mistake value;
B4: it according to the penalty values for the discrimination model being calculated, is reversely passed from discrimination model output layer to discrimination model input layer
It broadcasts, uses learning rate as input parameter, adjust the weight and biasing of every layer network, finally obtain after updating weight and biasing
Discrimination model;
According to the penalty values for generating model are calculated, from model output layer is generated to the backpropagation of mode input layer is generated, make
It uses learning rate as input parameter, adjusts the weight and biasing of every layer network, finally obtain the generation after updating weight and biasing
Model;
B5: repeating step B2 to B4, until reach maximum number of iterations complete training, finally obtain convergent long short-term memory and
Generate confrontation network model.
5. the Radar Echo Extrapolation method according to claim 4 based on long short-term memory and generation confrontation network, special
Sign is: in the step B1, the network layer for generating model, which is followed successively by, generates mode input layer, the first long short-term memory mind
It through network layer, the second long Memory Neural Networks layer in short-term, generates the full articulamentum of model and generates model output layer, generate model
Training the number of iterations is 150 times, learning rate 0.001, and the size for generating mode input layer is 3 × 360 × 480, the first length
When Memory Neural Networks layer and the hidden layer number of nodes of the second long Memory Neural Networks layer in short-term be 128, generate model and connect entirely
The number of nodes for connecing layer is 360 × 480, and the size for generating model output layer is 360 × 480, generates the final output image of model
Size is 360 × 480, that is, the size of the forecast image exported is 360 × 480;
The network layer of discrimination model is followed successively by discrimination model input layer, the full articulamentum of discrimination model first, discrimination model second
Full articulamentum, the full articulamentum of discrimination model third and discrimination model output layer, the training the number of iterations of discrimination model are 150 times,
Learning rate is 0.001, and the number of nodes of the full articulamentum of discrimination model first is 256, the node of the full articulamentum of discrimination model second
Number is 128, and the number of nodes of the full articulamentum of discrimination model third is 1, and the size 360 × 480 of discrimination model input layer differentiates
The size of model output layer is 1.
6. the Radar Echo Extrapolation method according to claim 4 based on long short-term memory and generation confrontation network, special
Sign is: in the step B2, the size of the input label input in every group of image collection in training set sample set is 3
×360×480。
7. the Radar Echo Extrapolation method according to claim 4 based on long short-term memory and generation confrontation network, special
Sign is: in the step B3, the calculation formula of the average absolute overall error MAE between forecast image and authentic specimen label
Are as follows:
Wherein, m indicates the pixel number shared in forecast image and authentic specimen label, trueiIndicate the i-th group picture in step A3
Authentic specimen label in image set conjunction,Indicate authentic specimen label trueiIn j-th of pixel value, oiIndicate prediction
Image,Indicate the value of j-th of pixel in forecast image;
Successively forecast image and authentic specimen label are input in discrimination model, discrimination model exports two sizes respectively and exists
Scalar D (G (input between [0,1]i)) and D (truei), the two scalar D (G then exported according to discrimination model
(inputi)) and D (truei), calculate separately the penalty values for generating model and discrimination model;
Generate the loss function of model are as follows:
Wherein, V1For the penalty values for generating model, D indicates that discrimination model to be optimized, G indicate generation model to be optimized,N is training sample set number, and N is the total number of radar data intensive data, and log indicates log-likelihood letter
Number, inputiFor i-th group of input sample, G (inputi) it is inputiInput generates the forecast image obtained after model G, D (G
(inputi)) indicate to generate differentiation result of the forecast image of the generation of model after discrimination model D differentiation;
The loss function of discrimination model are as follows:
Wherein, V2For the penalty values of discrimination model, trueiIndicate the authentic specimen label in i-th group of image collection, D (truei)
Indicate differentiation result of the authentic specimen label after discrimination model differentiates.
8. the Radar Echo Extrapolation method according to claim 1 based on long short-term memory and generation confrontation network, special
Sign is: in the step C, the size of Radar Echo Extrapolation image is 360 × 480.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910763464.0A CN110456355B (en) | 2019-08-19 | 2019-08-19 | Radar echo extrapolation method based on long-time and short-time memory and generation countermeasure network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910763464.0A CN110456355B (en) | 2019-08-19 | 2019-08-19 | Radar echo extrapolation method based on long-time and short-time memory and generation countermeasure network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110456355A true CN110456355A (en) | 2019-11-15 |
CN110456355B CN110456355B (en) | 2021-12-24 |
Family
ID=68487487
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910763464.0A Active CN110456355B (en) | 2019-08-19 | 2019-08-19 | Radar echo extrapolation method based on long-time and short-time memory and generation countermeasure network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110456355B (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111008604A (en) * | 2019-12-09 | 2020-04-14 | 上海眼控科技股份有限公司 | Prediction image acquisition method and device, computer equipment and storage medium |
CN111028260A (en) * | 2019-12-17 | 2020-04-17 | 上海眼控科技股份有限公司 | Image prediction method, image prediction device, computer equipment and storage medium |
CN111915591A (en) * | 2020-08-03 | 2020-11-10 | 中国海洋大学 | Spiral generation network for high-quality image extrapolation |
CN112180375A (en) * | 2020-09-14 | 2021-01-05 | 成都信息工程大学 | Meteorological radar echo extrapolation method based on improved TrajGRU network |
CN114509825A (en) * | 2021-12-31 | 2022-05-17 | 河南大学 | Strong convection weather prediction method and system for improving three-dimensional confrontation generation neural network based on hybrid evolution algorithm |
CN116499532A (en) * | 2023-06-27 | 2023-07-28 | 中建三局集团华南有限公司 | Complex marine environment deep water pile group construction monitoring system constructed based on hydrologic model |
Citations (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH09257951A (en) * | 1996-03-22 | 1997-10-03 | Nippon Telegr & Teleph Corp <Ntt> | Weather forcasting device |
WO2007005328A2 (en) * | 2005-06-30 | 2007-01-11 | Massachusetts Institute Of Technology | Weather radar echo tops forecast generation |
CN107016406A (en) * | 2017-02-24 | 2017-08-04 | 中国科学院合肥物质科学研究院 | The pest and disease damage image generating method of network is resisted based on production |
CN107121679A (en) * | 2017-06-08 | 2017-09-01 | 湖南师范大学 | Recognition with Recurrent Neural Network predicted method and memory unit structure for Radar Echo Extrapolation |
CN107545245A (en) * | 2017-08-14 | 2018-01-05 | 中国科学院半导体研究所 | A kind of age estimation method and equipment |
CN107632295A (en) * | 2017-09-15 | 2018-01-26 | 广东工业大学 | A kind of Radar Echo Extrapolation method based on sequential convolutional neural networks |
CN108846409A (en) * | 2018-04-28 | 2018-11-20 | 中国人民解放军国防科技大学 | Radar echo extrapolation model training method based on cyclic dynamic convolution neural network |
CN109003678A (en) * | 2018-06-12 | 2018-12-14 | 清华大学 | A kind of generation method and system emulating text case history |
WO2019019199A1 (en) * | 2017-07-28 | 2019-01-31 | Shenzhen United Imaging Healthcare Co., Ltd. | System and method for image conversion |
EP3471005A1 (en) * | 2017-10-13 | 2019-04-17 | Nokia Technologies Oy | Artificial neural network |
CN109872346A (en) * | 2019-03-11 | 2019-06-11 | 南京邮电大学 | A kind of method for tracking target for supporting Recognition with Recurrent Neural Network confrontation study |
CN109919032A (en) * | 2019-01-31 | 2019-06-21 | 华南理工大学 | A kind of video anomaly detection method based on action prediction |
CN109948693A (en) * | 2019-03-18 | 2019-06-28 | 西安电子科技大学 | Expand and generate confrontation network hyperspectral image classification method based on super-pixel sample |
-
2019
- 2019-08-19 CN CN201910763464.0A patent/CN110456355B/en active Active
Patent Citations (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH09257951A (en) * | 1996-03-22 | 1997-10-03 | Nippon Telegr & Teleph Corp <Ntt> | Weather forcasting device |
WO2007005328A2 (en) * | 2005-06-30 | 2007-01-11 | Massachusetts Institute Of Technology | Weather radar echo tops forecast generation |
CN107016406A (en) * | 2017-02-24 | 2017-08-04 | 中国科学院合肥物质科学研究院 | The pest and disease damage image generating method of network is resisted based on production |
CN107121679A (en) * | 2017-06-08 | 2017-09-01 | 湖南师范大学 | Recognition with Recurrent Neural Network predicted method and memory unit structure for Radar Echo Extrapolation |
WO2019019199A1 (en) * | 2017-07-28 | 2019-01-31 | Shenzhen United Imaging Healthcare Co., Ltd. | System and method for image conversion |
CN107545245A (en) * | 2017-08-14 | 2018-01-05 | 中国科学院半导体研究所 | A kind of age estimation method and equipment |
CN107632295A (en) * | 2017-09-15 | 2018-01-26 | 广东工业大学 | A kind of Radar Echo Extrapolation method based on sequential convolutional neural networks |
EP3471005A1 (en) * | 2017-10-13 | 2019-04-17 | Nokia Technologies Oy | Artificial neural network |
CN108846409A (en) * | 2018-04-28 | 2018-11-20 | 中国人民解放军国防科技大学 | Radar echo extrapolation model training method based on cyclic dynamic convolution neural network |
CN109003678A (en) * | 2018-06-12 | 2018-12-14 | 清华大学 | A kind of generation method and system emulating text case history |
CN109919032A (en) * | 2019-01-31 | 2019-06-21 | 华南理工大学 | A kind of video anomaly detection method based on action prediction |
CN109872346A (en) * | 2019-03-11 | 2019-06-11 | 南京邮电大学 | A kind of method for tracking target for supporting Recognition with Recurrent Neural Network confrontation study |
CN109948693A (en) * | 2019-03-18 | 2019-06-28 | 西安电子科技大学 | Expand and generate confrontation network hyperspectral image classification method based on super-pixel sample |
Non-Patent Citations (5)
Title |
---|
LIN TIAN 等: "A Generative Adversarial Gated Recurrent Unit Model for Precipitation Nowcasting", 《IEEE GEOSCIENCE AND REMOTE SENSING LETTERS》 * |
YATIAN SHEN 等: "Entity-Dependent Long-Short Time Memory Network for Semantic Relation Extraction", 《2018 5TH IEEE INTERNATIONAL CONFERENCE ON CLOUD COMPUTING AND INTELLIGENCE SYSTEMS (CCIS)》 * |
张玲玲: "基于雷达回波图像的短期降雨预测", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 * |
贾培艳 等: "基于节点分类的概念格同构判定算法", 《计算机应用与软件》 * |
陈元昭 等: "基于生成对抗网络GAN的人工智能临近预报方法研究", 《大气科学学报》 * |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111008604A (en) * | 2019-12-09 | 2020-04-14 | 上海眼控科技股份有限公司 | Prediction image acquisition method and device, computer equipment and storage medium |
CN111028260A (en) * | 2019-12-17 | 2020-04-17 | 上海眼控科技股份有限公司 | Image prediction method, image prediction device, computer equipment and storage medium |
CN111915591A (en) * | 2020-08-03 | 2020-11-10 | 中国海洋大学 | Spiral generation network for high-quality image extrapolation |
CN112180375A (en) * | 2020-09-14 | 2021-01-05 | 成都信息工程大学 | Meteorological radar echo extrapolation method based on improved TrajGRU network |
CN112180375B (en) * | 2020-09-14 | 2022-12-20 | 成都信息工程大学 | Weather radar echo extrapolation method based on improved TrajGRU network |
CN114509825A (en) * | 2021-12-31 | 2022-05-17 | 河南大学 | Strong convection weather prediction method and system for improving three-dimensional confrontation generation neural network based on hybrid evolution algorithm |
CN114509825B (en) * | 2021-12-31 | 2022-11-08 | 河南大学 | Strong convection weather prediction method and system for improving three-dimensional confrontation generation neural network based on hybrid evolution algorithm |
CN116499532A (en) * | 2023-06-27 | 2023-07-28 | 中建三局集团华南有限公司 | Complex marine environment deep water pile group construction monitoring system constructed based on hydrologic model |
CN116499532B (en) * | 2023-06-27 | 2023-09-01 | 中建三局集团华南有限公司 | Complex marine environment deep water pile group construction monitoring system constructed based on hydrologic model |
Also Published As
Publication number | Publication date |
---|---|
CN110456355B (en) | 2021-12-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110456355A (en) | A kind of Radar Echo Extrapolation method based on long short-term memory and generation confrontation network | |
Han et al. | Convolutional neural network for convective storm nowcasting using 3-D Doppler weather radar data | |
Zhou et al. | Forecasting different types of convective weather: A deep learning approach | |
Olatomiwa et al. | Adaptive neuro-fuzzy approach for solar radiation prediction in Nigeria | |
Mahjoobi et al. | Hindcasting of wave parameters using different soft computing methods | |
Litta et al. | Artificial neural network model in prediction of meteorological parameters during premonsoon thunderstorms | |
Akbarifard et al. | Predicting sea wave height using Symbiotic Organisms Search (SOS) algorithm | |
Meng et al. | Forecasting tropical cyclones wave height using bidirectional gated recurrent unit | |
CN111160520A (en) | BP neural network wind speed prediction method based on genetic algorithm optimization | |
CN109001736B (en) | Radar echo extrapolation method based on deep space-time prediction neural network | |
Chaudhuri | Convective energies in forecasting severe thunderstorms with one hidden layer neural net and variable learning rate back propagation algorithm | |
CN113239722B (en) | Deep learning based strong convection extrapolation method and system under multi-scale | |
Li et al. | Decomposition integration and error correction method for photovoltaic power forecasting | |
Elbeltagi et al. | Optimizing hyperparameters of deep hybrid learning for rainfall prediction: a case study of a Mediterranean basin | |
Wang et al. | A quantitative comparison of precipitation forecasts between the storm-scale numerical weather prediction model and auto-nowcast system in Jiangsu, China | |
CN115759445A (en) | Machine learning and cloud model-based classified flood random forecasting method | |
Mar et al. | Optimum neural network architecture for precipitation prediction of Myanmar | |
Zhu et al. | Long lead-time radar rainfall nowcasting method incorporating atmospheric conditions using long short-term memory networks | |
CN115877483A (en) | Typhoon path forecasting method based on random forest and GRU | |
Zhang et al. | A multi-site tide level prediction model based on graph convolutional recurrent networks | |
CN116012618A (en) | Weather identification method, system, equipment and medium based on radar echo diagram | |
Sonkusare et al. | Improved performance of multi-model ensemble through the bias correction based on ANN technique | |
Wang et al. | Data-driven modeling of Bay-Ocean wave spectra at bridge-tunnel crossing of Chesapeake Bay, USA | |
Yuhao et al. | Research on Prediction of Ground Settlement of Deep Foundation Pit Based on Improved PSO-BP Neural Network | |
Wei | Real-time extreme rainfall evaluation system for the construction industry using deep convolutional neural networks |
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 |