CN107632295A - A kind of Radar Echo Extrapolation method based on sequential convolutional neural networks - Google Patents
A kind of Radar Echo Extrapolation method based on sequential convolutional neural networks Download PDFInfo
- Publication number
- CN107632295A CN107632295A CN201710830449.4A CN201710830449A CN107632295A CN 107632295 A CN107632295 A CN 107632295A CN 201710830449 A CN201710830449 A CN 201710830449A CN 107632295 A CN107632295 A CN 107632295A
- Authority
- CN
- China
- Prior art keywords
- radar
- radar return
- convolution
- input
- radar echo
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- 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
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a kind of Radar Echo Extrapolation method based on sequential convolutional neural networks, the following three phases of this method point are achieved.Data processing stage：Radar echo map is resolved into by different frequency range by discrete cosine transform by multiple figures, operation is standardized to exploded view, exploded view is divided into multiple samples, obtains training sample set or test sample collection.The neural metwork training stage：Establish and initialize sequential convolutional neural networks；Neutral net is trained using training sample set, predicted value and counting loss is obtained by network propagated forward, by backpropagation adjusting parameter, restrains neutral net.Neutral net test phase：Handle to obtain test sample collection using the method for data processing stage for radar echo map to be extrapolated, be input to the neutral net specified, obtain predicted value.Predicted value is reduced into the radar return that radar return predicted.This method can overcome conventional method for precipitation particles decay, strengthen the defects of modeling ability is weak.
Description
Technical field
The invention belongs to surface weather observation technical field in Atmospheric Survey, more particularly to one kind to be based on sequential convolutional Neural
The Extrapolation method of network.
Background technology
Convective precipitation amount nowcasting is always an important research problem in weather forecast field, the mesh of this task
Be in certain region interior prediction shorter a period of time（Usually 0 to 6 hour）Precipitation, it gives birth to for many mankind
Production activity plays an important role, the management of operation, water conservancy projects such as airport.Currently existing Forecasting Methodology can be by roughly
It is divided into two kinds：Based on numerical weather prediction method（Numerical Weather Prediction, NWP）With based on radar return
The method of extrapolation figure.The former has very stable effect, but the forecast accuracy for zonule in the case where long-time is predicted
Relatively low, the latter in contrast, therefore is frequently used under the short-term nowcasting scene of precipitation based on Radar Echo Extrapolation figure.
Traditional Radar Echo Extrapolation method is generally basede on centroid tracking method and cross-correlation technique.This kind of method has a pass
The defects of key, their effect with predicted time is elongated when can rapidly be deteriorated, reason is that this kind of method only only accounts for dropping
The translation transformation of water particle, the not decay to them and enhancing is modeled, and as predicted time is elongated, precipitation particles decline
Subtracting and strengthen the effect brought can be superimposed, so that extrapolation effect is drastically deteriorated.The conventional model being limited compared to ability to express,
Have benefited from the operational capability that huge database and hardware device increasingly strengthen, depth model in image and sequential field all
There is extreme influence, and radar echo map is used as a kind of data for being easy to obtain so that deep learning method is solving radar return
There are very big potentiality in extrapolation problem.
Convolutional neural networks are used widely in image processing field as a part important in deep learning, it
Relation between two-dimensional image Plane-point be make use of as priori, improve the efficiency of depth model.Radar echo map
In addition to the feature for having image, also temporal aspect, the present invention pass through the company between convolutional neural networks addition time step
Connect, establish the convolutional neural networks based on sequential.This model make use of simultaneously the two-dimensional space of radar echo map according to
The relation of relying and Temporal dependency relation are as priori, the shared weights between different locus and different time steps,
Parameter amount is reduce further, reduces the risk of model over-fitting.By depth model has stronger ability to express, pass through
Usage history radar return diagram data, model constantly adjust weights, model can Accurate Prediction history radar echo map, it is final to learn
The mechanism of translation, decay and the enhancing of precipitation particles is practised, so as to be extrapolated to new radar echo map.It is worth in addition
It is noted that the spy due to the very different --- --- large scale of the Echo Characteristics evolution over time mode of different scale
Sign is more stable, and the feature of small yardstick easily decays, and this method is extracted the composition under radar echo map different frequency range first,
Use different weights to go to handle these compositions in depth model, network is learnt the radar of different scale under the driving of data
The development law of echo character.
The content of the invention
Goal of the invention：The technical problems to be solved by the invention are for traditional radar echo map Extrapolating model expression energy
Power is limited, and convolutional neural networks are less effective etc. is existing on handling sequence problem the problem of, using radar return historical data,
A kind of Radar Echo Extrapolation method based on sequential convolutional neural networks is proposed, is comprised the following steps：
A kind of 1. Radar Echo Extrapolation method based on sequential convolutional neural networks, it is characterised in that comprise the following steps：
Step 1, processing data：Given radar return data, the radar return data of wherein training are one group when including 15
The radar return graphic sequence of spacer step, the radar return data of test are one group of radar return graphic sequences for including 7 time steps,
Radar echo map is resolved into by two-dimension discrete cosine transform and inverse transformation by several frequency contents first；Calculate in sample set
Average value and standard deviation on all samples are often one-dimensional, z-score is carried out to all samples using their average value and standard deviation
Standardization；The picture segmentation obtained by above step into several samples, composing training sample set or test sample collection；
Step 2, sequential convolutional neural networks are trained：Establish convolutional neural networks, segmentum intercalaris at first 7 of Web vector graphic training sample
Radar echo map on point exports the prediction to radar echo map on rear 8 timing nodes as input；To convolutional neural networks
All parameters initialized；Using training sample set, model predication value is obtained by network propagated forward, uses predicted value
Prediction loss is calculated with actual value；The parameter of sequential convolutional neural networks is updated by backpropagation, repeats this process
Until convergence；
Step 3, test sequence convolutional neural networks：Based on the training network established in step 2, test network is established；To surveying
Radar echo map on probation carries out the processing method of step 1, obtains test sample collection, input test sample set to test network
In, predicted value is obtained by propagated forward；The inverse operation of step 1 is carried out to predicted value, predicted value is synthesized radar echo map.
2. according to the method for claim 1, it is characterised in that step 1 comprises the following steps：
Step 1-1, decompose different frequency contents：For each radar echo map, discrete cosine transform is carried out respectively, it is each
Individual radar return data are all a sequences being made up of radar echo map, and we are in radar return data setIt is individual to include
The size of time step isOKIt is expressed as tensor the data mode of the radar echo map of row, wherein,For the total number of radar return data, have for the radar return data of training, it is right
Have in test radar return data,Regard asIn individual radar return dataThe thunder of individual time step
Up to reflectogramArrangeRadar echo value on row, by rightIn each figure carry out a discrete cosine transform,
ObtainIt is one to be shaped asTensor, claimFor radar echo mapSpectrum signature,'s
Center（Near the 50th row, the 50th row in image）The low-frequency component of corresponding artwork, frequency is higher more outward, next using such as
Gaussian window shown in lower separates the different frequency composition of this spectrogram：
Wherein,,For a series of threshold value,WithSpecifyThe lower frequency border of frequency range and the upper bound residing for individual frequency content, four frequency ranges are rule of thumb selected, that is,, pass
Increase ordered series of numbersThe frequency bound of this four frequency range compositions is represented, Gauss functionHeight and width
It is consistent with spectrogram, window function is applied to by following formula by spectrum signatureIt is central, obtain the spectrum signature under different frequency range：Wherein oepratorValue is Hadamard products, to the spectrum signature under these different frequency rangesInverse discrete cosine transform is carried out, their expressions in the spatial domain can be obtained, claimReturned for former radar
Ripple figureIn a frequency domainComposition in individual frequency range, as long asAndIt is sufficiently large, have here；
Step 1-2, the composition in each frequency range of radar echo map is standardized using z-score standardization：Calculate first each
The average of each position on individual frequency range compositionAnd variance, to each position on each frequency range composition point
Carry out not z-score standardization：
Step 1-3, each frequency range ingredient breakdown after specification into 16 samples：Decomposed using following formula：Wherein, it is possible thereby to each frequency range composition point
Xie ChengTensor, thisIndividual tensor is regarded asIndividual different sample, by each radar return data use with
Upper method is handled, can construct one byThe training sample set that individual sample is formed, each of which radar return number
According to 16 samples in corresponding sample set, hereinafter useInstead of, this has no effect on understanding, in order to
Sample is expressed as the form that convolutional neural networks easily state, in a model using one group of tensor,To expressIndividual sample, the sample of spacer step sequence number expression on time：, in convolution god
Through in network, being commonly referred to asFor image channel sequence number, this title is continued to use here.
3. according to the method for claim 1, it is characterised in that step 2 comprises the following steps：
Step 2-1, establish sequential convolutional neural networks：Sequential convolution layer functions are defined first：
The input of sequential convolution function must is fulfilled for certain condition：WithExported with functionIt it is three by three groups of difference passages
The tensor that the consistent picture of number, line number columns is formed, theirs is shaped as,WithTwo convolution kernels are represented, theirs is shaped as,Biasing is represented, its shape is consistent with function output, it is assumed thatPort number be,Port number be all, function
The port number of output is, thenWithThe shape of the two tensors must be respectively justWith, functionConcrete form hereafter deployed again, we first establish one based on this function here
Model, in order to which easy partial symbols hereafter eliminate sample sequence number, for, define with drag：WhereinBe shaped as,
Be shaped as,Be shaped as,Be shaped as,
The line number core columns of all core is all 3；
For, define with drag：
Wherein、Be shaped as, whereinBe shaped as,'s
It is shaped as, the line number and columns of all core are all 3, symbolRepresent the convolution behaviour of the image of multichannel
Make（The convolution operation has carried out zero padding to input surrounding, to cause input and output image size constant）,For this convolution
Biasing, pay special attention to the above、、、、Assignment is not all obtained also, is here the every of them
Individual element is all entered as 0；
The further convolution operation to more thanAndFunction is defined, for convolution operationWhereinFor input, it is shaped as,For output, it is shaped as, convolution kernelBe shaped as,
Then convolution operation is first to inputSurrounding carry out zero padding, make its wide and it is high all increase by two rows, use symbolMark is filled
The new input arrived, according toWith, calculated by following formulaValue：
HereRepresent tensorMiddle position is set toElement value, each index fromStart to count, for、Together
Reason；
ForFunction, we are defined as follows：
Step 2-2, initialize all parameters：From step 2-1, the shape of each convolution kernel determines that this convolution is grasped
The input channel number and output channel number of work, it is assumed that they are respectivelyWith, then we are using following normal distribution
Each element of convolution kernel is initialized：For all bias terms, we make each of them
Individual element is；
Step 2-3, the sample set of the training obtained using step 1, predicted value is obtained by network propagated forward, uses prediction
Prediction loss is calculated in value and actual value：Training has 15 time steps, for each sample,, make its input as the model defined in step 2-3, obtain exporting tensor,, we use optimization method of the small lot stochastic gradient descent method as model, and this method is sample
The sample of this collection is divided into multiple small lots, and one small lot of training input each time, we define each small lot and includedIndividual sample, following formula is defined as outputCost：Its
InThe starting index concentrated for small lot in total sample；
Step 2-4, the parameter of sequential convolutional neural networks is updated by backpropagation：
In order to describe simplicity, convolution operation is given, define a functionFor
On convolution kernelPartial derivative；Defined functionForInputted on convolutionPartial derivative；
Likewise, given sequential convolution operation, defined function
ForOn convolution kernelPartial derivative；Defined functionForOn inputPartial derivative；It is fixed
Adopted functionForOn inputPartial derivative；Defined functionForClose
In inputPartial derivative；Defined functionForOn biasingPartial derivative；
First by local derviation of the backpropagation counting loss on parameter, the output on some time step is lostLadder
Spend and be：According to gradient chain rule, can obtain：
For time step, loss can be write outPartial derivative on each layer of output：And then obtain：
For time step, have：
By using above-mentioned gradient, lossFor any one parameterGradientIt can obtain, be carried out more using following formula
Newly：Several partial derivative functions of above-mentioned convolution operation and sequential convolution operation are further spread out,
For convolution function,It is defined by the formula：；It is defined by the formula：For
Sequential convolution function, have：Have：
Step 2-5：Above step is repeated until network convergence：Primary parameter is often updated, all to judging to loseWhether than history most
Small value is small, whenWhen continuing 100 iteration and not being less than history minimum value, network convergence is judged, terminate circulation.
4. according to the method for claim 1, it is characterised in that step 3 comprises the following steps：
Step:3-1：According to training network, test network is generated：Based on the training network described in step 2-1, for, we are replaced with following formula：For, replaced with following formula：Other functions
It is constant, and all parameters are inherited from training network, obtain test network；
Step 3-2：For given radar return data to be extrapolated, handled using the method for step 1, obtain test specimens
Test sample collection, is input to test network by this collection, and predicted value is obtained by propagated forward：Radar return data to be extrapolated are
Radar return data are handled, obtained by one radar return graphic sequence being made up of 7 time steps using the method for step 1
To test sample collection, in order to express simplicity, below only displaying there was only the situations of a radar return data, multiple radar return numbers
According to processing method can be regarded as repeating operating, as training sample set, a radar return to be extrapolated
Data are broken down into 16 test samples, can obtain 16 samples for including 7 time steps,
,, useAs the input of the test network defined in step 3-1, propagated forward is carried out, obtains net
Network prediction result,,；
Step 3-3：Neural network forecast result is reduced into Radar Echo Extrapolation result：16 neural network forecast results are merged using following formula
Into a tensor：Wherein, it is equal using what is be calculated in step 1
Value and variance, the inverse transformation of z-score standardization is carried out to output result： Radar Echo Extrapolation figure is representComposition under individual frequency range, Radar Echo Extrapolation figure is obtained by following formula：Here；
The present invention goes fitting history radar return data using depth model, made using in the powerful ability to express of depth model
The mechanism that model learning precipitation particles are mobile, decay and strengthen, so as to have more accurate Radar Echo Extrapolation ability.Pass through
The two dimensional surface characteristic and temporal characteristicses of radar return data are combined, the present invention proposes a kind of sequential convolutional neural networks,
Allow the network to remove learning model with relatively less parameter, there is more preferable generalization ability.
Brief description of the drawings
Fig. 1 is flow chart of the present invention.
Fig. 2 is sequential convolutional neural networks training network.
Fig. 3 is sequential convolutional neural networks test network.
Fig. 4 is picture breakdown schematic diagram.
Embodiment
The present invention is further illustrated underneath with accompanying drawing and specific embodiment.
1. a kind of Radar Echo Extrapolation method based on sequential convolutional neural networks, comprises the following steps：
Step 1, processing data：Given radar return data, the radar return data of wherein training are one group when including 15
The radar return graphic sequence of spacer step, the radar return data of test are one group of radar return graphic sequences for including 7 time steps,
Radar echo map is resolved into by two-dimension discrete cosine transform and inverse transformation by several frequency contents first；Calculate in sample set
Average value and standard deviation on all samples are often one-dimensional, z-score is carried out to all samples using their average value and standard deviation
Standardization；The picture segmentation obtained by above step into several samples, composing training sample set or test sample collection；
Step 2, sequential convolutional neural networks are trained：Establish convolutional neural networks, segmentum intercalaris at first 7 of Web vector graphic training sample
Radar echo map on point exports the prediction to radar echo map on rear 8 timing nodes as input；To convolutional neural networks
All parameters initialized；Using training sample set, model predication value is obtained by network propagated forward, uses predicted value
Prediction loss is calculated with actual value；The parameter of sequential convolutional neural networks is updated by backpropagation, repeats this process
Until convergence；
Step 3, test sequence convolutional neural networks：Based on the training network established in step 2, test network is established；To surveying
Radar echo map on probation carries out the processing method of step 1, obtains test sample collection, input test sample set to test network
In, predicted value is obtained by propagated forward；The inverse operation of step 1 is carried out to predicted value, predicted value is synthesized radar echo map.
2. according to the method for claim 1, it is characterised in that step 1 comprises the following steps：
Step 1-1, decompose different frequency contents：For each radar echo map, discrete cosine transform is carried out respectively, it is each
Individual radar return data are all a sequences being made up of radar echo map, and we are in radar return data setIt is individual to include
The size of time step isOKIt is expressed as tensor the data mode of the radar echo map of row, wherein,For the total number of radar return data, have for the radar return data of training, it is right
Have in test radar return data,Regard asIn individual radar return dataThe thunder of individual time step
Up to reflectogramArrangeRadar echo value on row, by rightIn each figure carry out a discrete cosine transform, obtain
ArriveIt is one to be shaped asTensor, claimFor radar echo mapSpectrum signature,'s
Center（Near the 50th row, the 50th row in image）The low-frequency component of corresponding artwork, frequency is higher more outward, next using such as
Gaussian window shown in lower separates the different frequency composition of this spectrogram：Its
In,,For a series of threshold value,WithSpecifyIt is individual
The lower frequency border of frequency range residing for frequency content and the upper bound, four frequency ranges are rule of thumb selected, that is,, it is incremented by
Ordered series of numbersThe frequency bound of this four frequency range compositions is represented, Gauss functionHeight and width with
Spectrogram is consistent, and window function is applied to spectrum signature by following formulaIt is central, obtain the spectrum signature under different frequency range：Wherein oepratorValue is Hadamard products, special to the frequency spectrum under these different frequency ranges
SignInverse discrete cosine transform is carried out, their expressions in the spatial domain can be obtained, claimFor former radar
ReflectogramIn a frequency domainComposition in individual frequency range, as long asAndIt is sufficiently large, have here；
Step 1-2, the composition in each frequency range of radar echo map is standardized using z-score standardization：Calculate first each
The average of each position on individual frequency range compositionAnd variance, to each position on each frequency range composition point
Carry out not z-score standardization：
Step 1-3, each frequency range ingredient breakdown after specification into 16 samples：Decomposed using following formula：Wherein, it is possible thereby to each frequency range composition point
Xie ChengTensor, thisIndividual tensor is regarded asIndividual different sample, by each radar return data use with
Upper method is handled, can construct one byThe training sample set that individual sample is formed, each of which radar return number
According to 16 samples in corresponding sample set, hereinafter useInstead of, this has no effect on understanding, for handle
Sample is expressed as the form that convolutional neural networks are easily stated, in a model using one group of tensor,
To expressIndividual sample, the sample of spacer step sequence number expression on time：, in convolutional neural networks, commonly referred to as
For image channel sequence number, this title is continued to use here.
3. according to the method for claim 1, it is characterised in that step 2 comprises the following steps：
Step 2-1, establish sequential convolutional neural networks：Sequential convolution layer functions are defined first：
The input of sequential convolution function must is fulfilled for certain condition：WithExported with functionIt it is three by three groups of difference passages
The tensor that the consistent picture of number, line number columns is formed, theirs is shaped as,
WithTwo convolution kernels are represented, theirs is shaped as
,Biasing is represented, its shape is consistent with function output, it is assumed thatPort number be,Port number be all, function
The port number of output is, thenWithThe shape of the two tensors must be respectively justWith, functionConcrete form hereafter deployed again, we first establish one based on this function here
Model, in order to which easy partial symbols hereafter eliminate sample sequence number, for, define with drag：WhereinBe shaped as,Be shaped as,Be shaped as,Be shaped as, institute
The line number core columns of some core is all 3；
For, define with drag：
Wherein、Be shaped as, whereinBe shaped as,'s
It is shaped as, the line number and columns of all core are all 3, symbolRepresent the convolution of the image of multichannel
Operation（The convolution operation has carried out zero padding to input surrounding, to cause input and output image size constant）,Rolled up for this
Long-pending biasing, pays special attention to the above、、、、Assignment is not all obtained also, is here them
Each element be entered as 0；
The further convolution operation to more thanAndFunction is defined, for convolution operationWhereinFor input, it is shaped as,For output, it is shaped as, convolution kernelBe shaped as, then convolution operation is first to inputSurrounding carry out zero padding, make its wide and it is high all increase by two rows,
Use symbolThe new input that mark filling obtains, according toWith, calculated by following formulaValue：HereRepresent tensorMiddle position is set to's
The value of element, each index fromStart to count, for、Similarly；
ForFunction, we are defined as follows：
Step 2-2, initialize all parameters：From step 2-1, the shape of each convolution kernel determines that this convolution is grasped
The input channel number and output channel number of work, it is assumed that they are respectivelyWith, then we are using following normal distribution
Each element of convolution kernel is initialized：For all bias terms, we make their each element
It is；
Step 2-3, the sample set of the training obtained using step 1, predicted value is obtained by network propagated forward, uses prediction
Prediction loss is calculated in value and actual value：Training has 15 time steps, for each sample,, make its input as the model defined in step 2-3, obtain exporting tensor,, we use optimization method of the small lot stochastic gradient descent method as model, and this method is sample
The sample of this collection is divided into multiple small lots, and one small lot of training input each time, we define each small lot and includedIndividual sample, following formula is defined as outputCost：
WhereinThe starting index concentrated for small lot in total sample；
Step 2-4, the parameter of sequential convolutional neural networks is updated by backpropagation：
In order to describe simplicity, convolution operation is given, define a functionFor
On convolution kernelPartial derivative；Defined functionForInputted on convolutionPartial derivative；
Likewise, given sequential convolution operation, defined function
ForOn convolution kernelPartial derivative；Defined functionForOn inputPartial derivative；
Defined functionForOn inputPartial derivative；Defined functionForOn inputPartial derivative；Defined functionForOn biasingPartial derivative；
First by local derviation of the backpropagation counting loss on parameter, the output on some time step is lostGradient
For：According to gradient chain rule, can obtain：
For time step, loss can be write outPartial derivative on each layer of output：And then obtain：For time step, have：By using above-mentioned
Gradient, lossFor any one parameterGradientIt can obtain, be updated using following formula：Several partial derivative functions of above-mentioned convolution operation and sequential convolution operation are further spread out, it is right
In convolution function,It is defined by the formula：’It is defined by the formula：
For sequential convolution function, have：
Have：
Step 2-5：Above step is repeated until network convergence：Primary parameter is often updated, all to judging to loseWhether than history most
Small value is small, whenWhen continuing 100 iteration and not being less than history minimum value, network convergence is judged, terminate circulation.
4. according to the method for claim 1, it is characterised in that step 3 comprises the following steps：
Step:3-1：According to training network, test network is generated：Based on the training network described in step 2-1, for, we are replaced with following formula：For, replaced with following formula：Other
Function is constant, and all parameters are inherited from training network, obtain test network；
Step 3-2：For given radar return data to be extrapolated, handled using the method for step 1, obtain test specimens
Test sample collection, is input to test network by this collection, and predicted value is obtained by propagated forward：Radar return data to be extrapolated are
Radar return data are handled, obtained by one radar return graphic sequence being made up of 7 time steps using the method for step 1
To test sample collection, in order to express simplicity, below only displaying there was only the situations of a radar return data, multiple radar return numbers
According to processing method can be regarded as repeating operating, as training sample set, a radar return to be extrapolated
Data are broken down into 16 test samples, can obtain 16 samples for including 7 time steps,
,, useAs the input of the test network defined in step 3-1, propagated forward is carried out, obtains net
Network prediction result,,；
Step 3-3：Neural network forecast result is reduced into Radar Echo Extrapolation result：16 neural network forecast results are closed using following formula
And into a tensor：Wherein, it is calculated using in step 1
Average and variance, to output result carry out z-score standardization inverse transformation： Radar Echo Extrapolation figure is representComposition under individual frequency range, Radar Echo Extrapolation figure is obtained by following formula：Here。
In summary, the present invention implements to provide a kind of Radar Echo Extrapolation method based on sequential convolutional neural networks,
Larger with traditional method difference, this method employs differentiation of the depth model to radar return and is modeled.In order to locate
Radar return graphic sequence is managed, the present invention proposes the sequential convolutional neural networks with reference to two dimensional surface feature and temporal aspect, base
In spatial domain and the priori of time-domain, specific weight have shared so that the more difficult over-fitting of network.Accompanying drawing above
The simply schematic diagram of a preferred embodiment, the invention described above example number is for illustration only, does not represent the quality of embodiment.
Presently preferred embodiments of the present invention is the foregoing is only, is not intended to limit the invention, within the spirit and principles of the invention,
Any modification, equivalent substitution and improvements made etc., should be included in the scope of the protection.
Claims (4)
- A kind of 1. Radar Echo Extrapolation method based on sequential convolutional neural networks, it is characterised in that comprise the following steps：Step 1, processing data：Given radar return data, the radar return data of wherein training are one group when including 15 The radar return graphic sequence of spacer step, the radar return data of test are one group of radar return graphic sequences for including 7 time steps, Radar echo map is resolved into by two-dimension discrete cosine transform and inverse transformation by several frequency contents first；Calculate in sample set Average value and standard deviation on all samples are often one-dimensional, z-score is carried out to all samples using their average value and standard deviation Standardization；The picture segmentation obtained by above step into several samples, composing training sample set or test sample collection；Step 2, sequential convolutional neural networks are trained：Establish convolutional neural networks, segmentum intercalaris at first 7 of Web vector graphic training sample Radar echo map on point exports the prediction to radar echo map on rear 8 timing nodes as input；To convolutional neural networks All parameters initialized；Using training sample set, model predication value is obtained by network propagated forward, uses predicted value Prediction loss is calculated with actual value；The parameter of sequential convolutional neural networks is updated by backpropagation, repeats this process Until convergence；Step 3, test sequence convolutional neural networks：Based on the training network established in step 2, test network is established；To surveying Radar echo map on probation carries out the processing method of step 1, obtains test sample collection, input test sample set to test network In, predicted value is obtained by propagated forward；The inverse operation of step 1 is carried out to predicted value, predicted value is synthesized radar echo map.
- 2. according to the method for claim 1, it is characterised in that step 1 comprises the following steps：Step 1-1, decompose different frequency contents：For each radar echo map, discrete cosine transform is carried out respectively, it is each Individual radar return data are all a sequences being made up of radar echo map, and we include k-th in radar return data set It is expressed as tensor the data mode for the radar echo map that the size of T time step arranges for 100 rows 100, whereinFor the total number of radar return data, there are T=15 for the radar return data of training, for Test radar return data have T=7,Regard as the radar of t-th of time step in k-th of radar return data Reflectogram xth arranges the radar echo value on y rows, by rightIn each figure carry out a discrete cosine transform, obtain ArriveIt is one to be shaped asTensor, claimFor radar echo mapSpectrum signature,'s Center（Near the 50th row, the 50th row in image）The low-frequency component of corresponding artwork, frequency is higher more outward, next using such as Gaussian window shown in lower separates the different frequency composition of this spectrogram：Wherein,For a series of threshold value,With The lower frequency border of frequency range and the upper bound residing for i-th of frequency content are specified, rule of thumb selects four frequency ranges, that is,Ascending seriesRepresent the frequency bound of this four frequency range compositions, Gaussian window letter NumberHeight and width it is consistent with spectrogram, window function is applied to by following formula by spectrum signatureIt is central, obtain different frequency range Under spectrum signature：Wherein oepratorValue is Hadamard products, to the spectrum signature under these different frequency rangesCarry out discrete remaining String inverse transformation, their expressions in the spatial domain can be obtained, claimFor former radar echo mapIn a frequency domain Composition in i-th of frequency range, as long asAndIt is sufficiently large, have here；Step 1-2, the composition in each frequency range of radar echo map is standardized using z-score standardization：Calculate first each The average of each position on individual frequency range compositionAnd variance, to each position on each frequency range composition Put progress z-score standardization respectively：Step 1-3, each frequency range ingredient breakdown after specification into 16 samples：Decomposed using following formula：Wherein, it is possible thereby to each frequency range ingredient breakdown into 16 tensors, this 16 tensors Regard 16 different samples as, by being handled using above method each radar return data, one can be constructed The individual training sample set being made up of 16N sample, each of which radar return data correspond to 16 samples in sample set, under Replaced in text using k, this has no effect on understanding, easy in order to which sample is expressed as convolutional neural networks The form of statement, in a model using one group of tensorTo express k-th of sample, spacer step sequence on time Number represent sample：, in convolutional neural networks, i is generally referred to as image channel sequence number, here edge With this title.
- 3. according to the method for claim 1, it is characterised in that step 2 comprises the following steps：Step 2-1, establish sequential convolutional neural networks：Sequential convolution layer functions are defined first：The input of sequential convolution function must is fulfilled for certain condition：WithExported with functionIt is three tensors being made up of the consistent picture of three groups of difference port numbers, line number columns, Theirs is shaped as,WithTwo convolution kernels are represented, theirs is shaped as,Biasing is represented, its shape and function are defeated Go out consistent, it is assumed thatPort number be,Port number be all, function output port number be, thenWithThe shape of the two tensors must be respectively justWith, function's Concrete form is hereafter deployed again, and we first establish a model based on this function here, for easy part hereafter Symbol eliminates sample sequence number k, for, define with drag：WhereinBe shaped as,Be shaped as,Be shaped as,Be shaped as, the line number core columns of all core is all 3；For, define with drag：Wherein、Be shaped as, whereinBe shaped as,Be shaped as, the line number and columns of all core are all 3, and symbol * represents the figure of multichannel The convolution operation of picture（The convolution operation has carried out zero padding to input surrounding, to cause input and output image size constant）, For the biasing of this convolution, pay special attention to the aboveAssignment is not all obtained also, Here it is entered as 0 for their each element；Further to more than convolution operation * andFunction is defined, for convolution operationWhereinFor input, it is shaped as,For output, it is shaped as, convolution kernel K shape For, then convolution operation is first to inputSurrounding carry out zero padding, make its wide and it is high all increase by two rows, Use symbolThe new input that mark filling obtains, according toWith, calculated by following formulaValue：HereRepresent tensorMiddle position is set toElement value, each index is counted since 1, right In、Similarly；ForFunction, we are defined as follows：Step 2-2, initialize all parameters：From step 2-1, the shape of each convolution kernel determines that this convolution is grasped The input channel number and output channel number of work, it is assumed that they are respectivelyWith, then we are using following normal distribution Each element of convolution kernel is initialized：For all bias terms, we make their each element be 0.1；Step 2-3, the sample set of the training obtained using step 1, predicted value is obtained by network propagated forward, uses prediction Prediction loss is calculated in value and actual value：Training has 15 time steps, for each sample, make its input as the model defined in step 2-3, obtain exporting tensor, we use optimization method of the small lot stochastic gradient descent method as model, this The sample of sample set is divided into multiple small lots by individual method, and one small lot of training input each time, it is small that we define each Batch includes 128 samples, defines following formula as outputCost：WhereinThe starting index concentrated for small lot in total sample；Step 2-4, the parameter of sequential convolutional neural networks is updated by backpropagation：In order to describe simplicity, convolution operation is given, define a functionFor L On convolution kernelPartial derivative；Define letterInputted for L on convolutionPartial derivative；Likewise, given sequential convolution operation, defined function It is L on convolution kernelPartial derivative；Defined functionBe L on inputPartial derivative；It is fixed Adopted functionBe L on inputPartial derivative；Defined function Be L on inputPartial derivative；Defined functionBe L on biasingPartial derivative；First by local derviation of the backpropagation counting loss on parameter, the output on some time step is lostLadder Spend and be：According to gradient chain rule, can obtain：For time step, partial derivatives of the loss L on each layer of output can be write out：And then obtain：For time step, have：By using above-mentioned gradient, gradients of the loss L for any one parameter WIt can obtain, under use Formula is updated：Several partial derivative functions of above-mentioned convolution operation and sequential convolution operation are further spread out, for convolution function,It is defined by the formula：It is defined by the formula：For sequential convolution function, have：Have：Step 2-5：Above step is repeated until network convergence：Primary parameter is often updated, all to judging to lose whether L compares history Minimum value is small, when L, which continues 100 iteration, is not less than history minimum value, judges network convergence, terminates circulation.
- 4. according to the method for claim 1, it is characterised in that step 3 comprises the following steps：Step:3-1：According to training network, test network is generated：Based on the training network described in step 2-1, for, we are replaced with following formula：For, replaced with following formula：Other functions are constant, and all parameters are inherited from training network, obtain test network；Step 3-2：For given radar return data to be extrapolated, handled using the method for step 1, obtain test specimens Test sample collection, is input to test network by this collection, and predicted value is obtained by propagated forward：Radar return data to be extrapolated are Radar return data are handled, obtained by one radar return graphic sequence being made up of 7 time steps using the method for step 1 To test sample collection, in order to express simplicity, below only displaying there was only the situations of a radar return data, multiple radar return numbers According to processing method can be regarded as repeating operating, as training sample set, a radar return to be extrapolated Data are broken down into 16 test samples, can obtain 16 samples for including 7 time steps, , useAs the input of the test network defined in step 3-1, propagated forward is carried out, is obtained Neural network forecast result,；Step 3-3：Neural network forecast result is reduced into Radar Echo Extrapolation result：16 neural network forecast results are closed using following formula And into a tensor：Wherein, using the average and variance being calculated in step 1, z-score mark is carried out to output result The inverse transformation of standardization：Composition of the Radar Echo Extrapolation figure under i-th of frequency range is represent, Radar Echo Extrapolation is obtained by following formula Figure：Here。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710830449.4A CN107632295A (en) | 2017-09-15 | 2017-09-15 | A kind of Radar Echo Extrapolation method based on sequential convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710830449.4A CN107632295A (en) | 2017-09-15 | 2017-09-15 | A kind of Radar Echo Extrapolation method based on sequential convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107632295A true CN107632295A (en) | 2018-01-26 |
Family
ID=61101292
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710830449.4A Pending CN107632295A (en) | 2017-09-15 | 2017-09-15 | A kind of Radar Echo Extrapolation method based on sequential convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107632295A (en) |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108508505A (en) * | 2018-02-05 | 2018-09-07 | 南京云思创智信息科技有限公司 | Heavy showers and thunderstorm forecasting procedure based on multiple dimensioned convolutional neural networks and system |
| CN108734357A (en) * | 2018-05-29 | 2018-11-02 | 北京佳格天地科技有限公司 | Weather prognosis system and method |
| CN108732550A (en) * | 2018-08-01 | 2018-11-02 | 北京百度网讯科技有限公司 | Method and apparatus for predicting radar return |
| 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 |
| CN108872993A (en) * | 2018-04-28 | 2018-11-23 | 中国人民解放军国防科技大学 | Radar echo extrapolation method based on cyclic dynamic convolution neural network |
| CN109283505A (en) * | 2018-09-03 | 2019-01-29 | 南京信息工程大学 | A method of amendment Radar Echo Extrapolation image Divergent Phenomenon |
| CN109444863A (en) * | 2018-10-23 | 2019-03-08 | 广西民族大学 | A kind of estimation method of the narrowband ultrasonic echo number based on convolutional neural networks |
| CN109640068A (en) * | 2018-10-31 | 2019-04-16 | 百度在线网络技术（北京）有限公司 | Information forecasting method, device, equipment and the storage medium of video frame |
| CN110082738A (en) * | 2019-06-18 | 2019-08-02 | 西安电子科技大学 | Radar target identification method based on Gaussian Mixture and tensor Recognition with Recurrent Neural Network |
| CN110426702A (en) * | 2019-07-24 | 2019-11-08 | 中国人民解放军海军航空大学 | The Sequential filter method and system of system deviation and dbjective state based on more radars |
| CN110456355A (en) * | 2019-08-19 | 2019-11-15 | 河南大学 | A kind of Radar Echo Extrapolation method based on long short-term memory and generation confrontation network |
| CN110717546A (en) * | 2019-10-15 | 2020-01-21 | 上海眼控科技股份有限公司 | Image distribution similarity calculation method and device, computer equipment and storage medium |
| CN110824587A (en) * | 2019-11-01 | 2020-02-21 | 上海眼控科技股份有限公司 | Image prediction method, image prediction 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 |
| CN111369485A (en) * | 2020-03-05 | 2020-07-03 | 上海眼控科技股份有限公司 | Method and equipment for optimizing and predicting radar echo image |
| CN113239614A (en) * | 2021-04-22 | 2021-08-10 | 西北工业大学 | Atmospheric turbulence phase space-time prediction algorithm |
| US11131790B2 (en) | 2019-03-25 | 2021-09-28 | Yandex Europe Ag | Method of and system for generating weather forecast |
| CN113536681A (en) * | 2021-07-21 | 2021-10-22 | 北京航空航天大学 | Electric steering engine health assessment method based on time sequence extrapolation prediction |
| CN113657477A (en) * | 2021-08-10 | 2021-11-16 | 南宁五加五科技有限公司 | Method, device and system for forecasting short-term rainfall |
| CN116380108A (en) * | 2023-06-02 | 2023-07-04 | 山东科技大学 | Track planning method and device based on laser radar |
-
2017
- 2017-09-15 CN CN201710830449.4A patent/CN107632295A/en active Pending
Cited By (27)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108508505A (en) * | 2018-02-05 | 2018-09-07 | 南京云思创智信息科技有限公司 | Heavy showers and thunderstorm forecasting procedure based on multiple dimensioned convolutional neural networks and system |
| 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 |
| CN108872993A (en) * | 2018-04-28 | 2018-11-23 | 中国人民解放军国防科技大学 | Radar echo extrapolation method based on cyclic dynamic convolution neural network |
| CN108734357A (en) * | 2018-05-29 | 2018-11-02 | 北京佳格天地科技有限公司 | Weather prognosis system and method |
| CN108732550A (en) * | 2018-08-01 | 2018-11-02 | 北京百度网讯科技有限公司 | Method and apparatus for predicting radar return |
| CN109283505A (en) * | 2018-09-03 | 2019-01-29 | 南京信息工程大学 | A method of amendment Radar Echo Extrapolation image Divergent Phenomenon |
| CN109444863A (en) * | 2018-10-23 | 2019-03-08 | 广西民族大学 | A kind of estimation method of the narrowband ultrasonic echo number based on convolutional neural networks |
| CN109640068A (en) * | 2018-10-31 | 2019-04-16 | 百度在线网络技术（北京）有限公司 | Information forecasting method, device, equipment and the storage medium of video frame |
| RU2757591C1 (en) * | 2019-03-25 | 2021-10-19 | Общество С Ограниченной Ответственностью «Яндекс» | Method and system for generating weather forecast |
| US11131790B2 (en) | 2019-03-25 | 2021-09-28 | Yandex Europe Ag | Method of and system for generating weather forecast |
| CN110082738B (en) * | 2019-06-18 | 2023-03-28 | 西安电子科技大学 | Radar target identification method based on Gaussian mixture and tensor recurrent neural network |
| CN110082738A (en) * | 2019-06-18 | 2019-08-02 | 西安电子科技大学 | Radar target identification method based on Gaussian Mixture and tensor Recognition with Recurrent Neural Network |
| CN110426702A (en) * | 2019-07-24 | 2019-11-08 | 中国人民解放军海军航空大学 | The Sequential filter method and system of system deviation and dbjective state based on more radars |
| CN110456355B (en) * | 2019-08-19 | 2021-12-24 | 河南大学 | Radar echo extrapolation method based on long-time and short-time memory and generation countermeasure network |
| CN110456355A (en) * | 2019-08-19 | 2019-11-15 | 河南大学 | A kind of Radar Echo Extrapolation method based on long short-term memory and generation confrontation network |
| CN110717546A (en) * | 2019-10-15 | 2020-01-21 | 上海眼控科技股份有限公司 | Image distribution similarity calculation method and device, computer equipment and storage medium |
| CN110824587A (en) * | 2019-11-01 | 2020-02-21 | 上海眼控科技股份有限公司 | Image prediction method, image prediction device, computer equipment and storage medium |
| CN110824587B (en) * | 2019-11-01 | 2021-02-09 | 上海眼控科技股份有限公司 | Image prediction method, image prediction 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 |
| CN111369485A (en) * | 2020-03-05 | 2020-07-03 | 上海眼控科技股份有限公司 | Method and equipment for optimizing and predicting radar echo image |
| CN113239614A (en) * | 2021-04-22 | 2021-08-10 | 西北工业大学 | Atmospheric turbulence phase space-time prediction algorithm |
| CN113536681A (en) * | 2021-07-21 | 2021-10-22 | 北京航空航天大学 | Electric steering engine health assessment method based on time sequence extrapolation prediction |
| CN113536681B (en) * | 2021-07-21 | 2024-04-02 | 北京航空航天大学 | Electric steering engine health assessment method based on time sequence extrapolation prediction |
| CN113657477A (en) * | 2021-08-10 | 2021-11-16 | 南宁五加五科技有限公司 | Method, device and system for forecasting short-term rainfall |
| CN113657477B (en) * | 2021-08-10 | 2022-04-08 | 南宁五加五科技有限公司 | Method, device and system for forecasting short-term rainfall |
| CN116380108A (en) * | 2023-06-02 | 2023-07-04 | 山东科技大学 | Track planning method and device based on laser radar |
| CN116380108B (en) * | 2023-06-02 | 2023-08-11 | 山东科技大学 | Track planning method and device based on laser radar |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107632295A (en) | A kind of Radar Echo Extrapolation method based on sequential convolutional neural networks | |
| CN107358293B (en) | Neural network training method and device | |
| CN105243398B (en) | The method of improvement convolutional neural networks performance based on linear discriminant analysis criterion | |
| CN104217214B (en) | RGB D personage's Activity recognition methods based on configurable convolutional neural networks | |
| CN108182450A (en) | A kind of airborne Ground Penetrating Radar target identification method based on depth convolutional network | |
| Roodposhti et al. | Towards automatic calibration of neighbourhood influence in cellular automata land-use models | |
| CN107909015A (en) | Hyperspectral image classification method based on convolutional neural networks and empty spectrum information fusion | |
| CN107590515A (en) | The hyperspectral image classification method of self-encoding encoder based on entropy rate super-pixel segmentation | |
| CN109360097A (en) | Prediction of Stock Index method, apparatus, equipment and storage medium based on deep learning | |
| CN109118013A (en) | A kind of management data prediction technique, readable storage medium storing program for executing and forecasting system neural network based | |
| CN109711401A (en) | A kind of Method for text detection in natural scene image based on Faster Rcnn | |
| CN107273897A (en) | A kind of character recognition method based on deep learning | |
| CN106600595A (en) | Human body characteristic dimension automatic measuring method based on artificial intelligence algorithm | |
| CN107657243A (en) | Neutral net Radar range profile's target identification method based on genetic algorithm optimization | |
| CN105005789B (en) | A kind of remote sensing images terrain classification method of view-based access control model vocabulary | |
| CN106611423B (en) | SAR image segmentation method based on ridge ripple filter and deconvolution structural model | |
| CN103886216B (en) | A kind of multiple spot Geo-statistic Method based on geology Vector Message | |
| CN106203444B (en) | Classification of Polarimetric SAR Image method based on band wave and convolutional neural networks | |
| CN107705556A (en) | A kind of traffic flow forecasting method combined based on SVMs and BP neural network | |
| CN106022392A (en) | Deep neural network sample automatic accepting and rejecting training method | |
| Zhu et al. | How does the calibration method impact the performance of the air2water model for the forecasting of lake surface water temperatures? | |
| CN106651884B (en) | Mean field variation Bayes's SAR image segmentation method based on sketch structure | |
| CN106683102A (en) | SAR image segmentation method based on ridgelet filters and convolution structure model | |
| CN104732091A (en) | Cellular automaton river bed evolution prediction method based on natural selection ant colony algorithm | |
| CN107292337A (en) | Ultralow order tensor data filling method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180126 |
|
| WD01 | Invention patent application deemed withdrawn after publication |