CN106886023A

CN106886023A - A kind of Radar Echo Extrapolation method based on dynamic convolutional neural networks

Info

Publication number: CN106886023A
Application number: CN201710110183.6A
Authority: CN
Inventors: 李骞; 施恩; 顾大权
Original assignee: PLA University of Science and Technology
Current assignee: PLA University of Science and Technology
Priority date: 2017-02-27
Filing date: 2017-02-27
Publication date: 2017-06-23
Anticipated expiration: 2037-02-27
Also published as: CN106886023B

Abstract

The invention discloses a kind of Radar Echo Extrapolation method based on dynamic convolutional neural networks, including：Offline convolution neural metwork training：To the training image collection for giving, training sample set is obtained by data prediction, the dynamic convolution neural network model of initialization, and dynamic convolutional neural networks are trained using training sample set, calculating output valve, the process of back-propagating renewal network parameter by network propagated forward restrains dynamic convolutional neural networks.Online Radar Echo Extrapolation：Test chart image set is converted into by test sample collection by data prediction, tested using the trained dynamic convolutional neural networks of test sample set pair, the probability vector phase convolution that will be obtained in last width radar return image in input image sequence and network propagated forward, the Radar Echo Extrapolation image predicted.

Description

A kind of Radar Echo Extrapolation method based on dynamic convolutional neural networks

Technical field

The invention belongs to surface weather observation technical field in Atmospheric Survey, more particularly to one kind is based on dynamic convolutional Neural The Radar Echo Extrapolation method of network.

Background technology

Nowcasting refers mainly to the weather forecast of the high-spatial and temporal resolution of 0~3 hour, and Main Prediction object includes strong drop The diastrous weathers such as water, strong wind, hail.At present, many forecast systems all use Numerical Prediction Models, but due to numerical forecast In the presence of return slow (spin-up) has been forecast, its Nowcasting is limited in one's ability.New Generation Doppler Weather Radar has very high Sensitivity and resolution ratio, the spatial resolution of its data information can reach 200~1000m, and temporal resolution can reach 2 ~15min.Additionally, Doppler radar also has rational mode of operation, comprehensive condition monitoring and fault warning, advanced Real-time Calibration System and abundant Radar meteorology product algorithm, the reliability of Nowcasting can be greatly improved.Nowadays, New Generation Doppler Weather Radar has become one of most effective instrument of nowcasting, is faced using Doppler radar Nearly forecast is based primarily upon Radar Echo Extrapolation technology, i.e., according to current time radar observation result, thus it is speculated that radar return future Position and intensity, to realize the track prediction to strong convection system.

Traditional Radar Echo Extrapolation method is centroid tracking method and the cross-correlation technique based on maximum correlation coefficient (Tracking Radar Echoes by Correlation, TREC), but all there is certain deficiency, barycenter in conventional method Tracing is only applicable to echo compared with strong, the less storm monomer of scope, and the forecast for precipitation on a large scale is unreliable；TREC is general Echo is considered as linear change, and echo change is increasingly complex in reality, while such method is easily by vector field Unordered vector disturbance.Additionally, existing method is low to the utilization rate of Radar Data, and history Radar Data includes local weather system The key character of system change, with researching value very high.

To improve the ageing of Radar Echo Extrapolation, and the change of radar return is studied from substantial amounts of history Radar Data Rule, machine learning method is introduced into Radar Echo Extrapolation.Convolutional neural networks (Convolutional Neural Network, CNN) as the important branch of deep learning, it is widely used in the fields such as image procossing, pattern-recognition.The network is most It is using local connection, weights are shared, the method for down-sampling, deformation, translation and upset tool to input picture the characteristics of big There is stronger adaptability.For the strong temporal correlation existed between radar return image, dynamic convolution of the design based on input Neutral net, the network can dynamically change weighting parameter according to the radar echo map of input, and then predict extrapolated image.Profit Dynamic convolutional neural networks are trained with history Radar Data, network is more fully extracted echo character, study echo change Rule, for improving Radar Echo Extrapolation accuracy, optimization nowcasting effect is significant.

The content of the invention

Goal of the invention：When the technical problems to be solved by the invention are directed to the extrapolation of existing Radar Echo Extrapolation method Effect is short, not enough to Radar Data utilization rate, it is proposed that a kind of Radar Echo Extrapolation method based on dynamic convolutional neural networks, real Now to the plane high such as radar echo intensity display CAPPI (Constant AltitudePlan Position Indicator, CAPPI) outside forecast of image, comprises the following steps：

Step 1, trains offline convolutional neural networks：Input training image collection, data prediction is carried out to training image collection, Training sample set, the dynamic convolution neural network structure of design are obtained, and initializes network training parameter；Assembled for training using training sample Practice dynamic convolutional neural networks, the orderly image sequence of input obtains width prediction by dynamic convolutional neural networks propagated forward Image, calculates the error between prognostic chart picture and control label, updates the weighting parameter of network by backpropagation and biasing is joined Number, repeats this process until reaching training termination condition, obtains convergent dynamic convolutional neural networks；

Step 2, online Radar Echo Extrapolation：Input test image set, data prediction is carried out to test chart image set, is obtained Test sample collection, then will in test sample collection input step 1 obtain dynamic convolutional neural networks in, by before network to biography Calculating probability vector is broadcast, and last width radar return image in input image sequence is mutually rolled up with the probability vector for obtaining Product, the Radar Echo Extrapolation image predicted.

Step 1 of the present invention is comprised the following steps：

Step 1-1, data prediction：Input training image collection, specification is carried out to every piece image that training image is concentrated Change is processed, and 280 × 280 floating number image will be converted into per piece image, floating number image collection is obtained, to floating number image Set is divided, training sample set of the construction comprising TrainsetSize group samples；

Step 1-2, the dynamic convolutional neural networks of initialization：The dynamic convolution neural network structure of design, is configured to generation The sub-network of probability vector, reconstructs the probabilistic forecasting layer for image extrapolation, for the offline neural metwork training stage provides dynamic State convolutional neural networks initialization model；

Step 1-3, the dynamic convolution Neural Network Training Parameter of initialization：E-learning rate λ=0.0001 is made, The sample size BatchSize=10 that training stage is input into every time, most large quantities of frequency of training of training sample setCurrently criticize frequency of training BatchNum=1, the maximum iteration of network training IterationMax=40, current iteration number of times IterationNum=1；

Step 1-4, reads training sample：By the way of batch training, the training sample obtained from step 1-1 is trained every time Concentrate and read BatchSize group training samples, every group of training sample is { x₁,x₂,x₃,x₄, y }, altogether comprising 5 width images, wherein {x₁,x₂,x₃,x₄Used as input image sequence, y is corresponding control label；

Step 1-5, propagated forward：The input image sequence feature that extraction step 1-4 is obtained in a sub-network, obtains level Probability vector HPV and vertical probability vector VPV；Probabilistic forecasting layer in, by the last piece image in input image sequence according to It is secondary with VPV, HPV phase convolution, obtain the output prognostic chart picture of propagated forward；

Step 1-6, backpropagation：The error term of probability vector is reversely tried to achieve in probabilistic forecasting layer, according to probability vector Error term from rear to the preceding error term for successively calculating each Internet in subnet network layers, and then calculate error in each Internet To weighting parameter and the gradient of offset parameter, using the parameter of the gradient updating dynamic convolutional neural networks for obtaining；

Step 1-7, off-line training step control：Overall control is carried out to the offline neural metwork training stage, is divided into following Three kinds of situations：

If training sample is concentrated and still suffers from original training sample, i.e. BatchNum ＜ BatchMax, then step is returned Rapid 1-4 continues to read BatchSize group training samples, carries out network training；

If training sample is concentrated does not exist original training sample, i.e. BatchNum=BatchMax, and current net Network iterations is less than maximum iteration, i.e. IterationNum ＜ IterationMax, then make BatchNum=1, returns Step 1-4 continues to read BatchSize group training samples, carries out network training；

If training sample is concentrated does not exist original training sample, i.e. BatchNum=BatchMax, and network changes Generation number reaches maximum iteration, i.e. IterationNum=IterationMax, then terminate offline neural metwork training rank Section, the dynamic convolution neural network model for being trained.

Step 1-1 data predictions of the present invention are comprised the following steps：

Step 1-1-1, sampling：The image that training image is concentrated is sequentially arranged, and constant duration is distributed, when Between at intervals of 6 minutes, altogether comprising N_TrainWidth image, TrainsetSize is determined by equation below：

Wherein, Mod (N_Train, 4) and represent N_TrainTo 4 modulus, [N] represents the maximum integer for being not more than N, tries to achieve After TrainsetSize, retain training image by sampling and concentrate preceding 4 × TrainsetSize+1 width image, by deleting during sampling Concentrate last image to meet amount of images except training image to require；

Step 1-1-2, normalized images：Image conversion, normalization operation, by original point are carried out to the image that sampling is obtained Resolution is converted into the floating number image that resolution ratio is 280 × 280 for 2000 × 2000 coloured image；

Step 1-1-3, constructs training sample set：

The floating number image set obtained using step 1-1-2 constructs training sample set, by every four in floating number image set Adjacent image, i.e. { 4N+1,4N+2,4N+3,4N+4 } width image as one group of list entries, [4 × (N+1)+1] width figure As by cutting, retain central resolution ratio for 240 × 240 part as correspondence sample control label, for N group samplesIts make is as follows：

In above formula, G_4N+1Represent floating number image set in 4N+1 width images, N is positive integer, and have N ∈ [0, TrainsetSize-1], Crop () represents trimming operation, and the portion that original image center size is for 240 × 240 is retained after cutting Point, finally give the training sample set comprising TrainsetSize group training samples；

Wherein, step 1-1-2 is comprised the following steps：

Step 1-1-2-1, image conversion：To reduce the computation complexity of convolutional neural networks training, step 1-1-1 is adopted The image that sample is obtained is converted into gray level image, by cutting the part for retaining that original image center resolution ratio is for 560 × 560, will Image resolution ratio boil down to 280 × 280 after cutting, obtains the gray-scale map that resolution ratio is 280 × 280；

Step 1-1-2-2, data normalization：By each pixel in step 1-1-2-1 in the gray-scale map of acquisition Value is mapped to [0~1] from [0~255], by obtaining the floating number image that resolution ratio is 280 × 280 after normalization.

Step 1-2 of the present invention is comprised the following steps：

Step 1-2-1, constructs sub-network：

Sub-network is made up of 10 Internets, be followed successively by from front to back convolutional layer C1, down-sampling layer S1, convolutional layer C2, under Sample level S2, convolutional layer C3, down-sampling layer S3, convolutional layer C4, down-sampling layer S5, convolutional layer C5 and grader layer F1；

Step 1-2-2, construction probabilistic forecasting layer：

Dynamic convolutional layer DC1 and dynamic convolutional layer DC2 is constructed in probabilistic forecasting layer, by the vertical probability of sub-network output to Amount VPV as dynamic convolutional layer DC1 convolution kernel, level probability vector HPV as dynamic convolutional layer DC2 convolution kernel；

Wherein, step 1-2-1 is comprised the following steps：

Step 1-2-1-1, constructs convolutional layer：Determine herein below：The output characteristic figure quantity OutputMaps of convolutional layer, Convolution kernel k and offset parameter bias.For convolution kernel, it is thus necessary to determine that the width KernelSize of convolution kernel (convolution kernel point Resolution is KernelSize × KernelSize), the quantity KernelNumber of convolution kernel (value be the convolutional layer be input into it is defeated Go out the product of characteristic pattern quantity), and convolution kernel is constructed according to Xavier initial methods；For offset parameter, its quantity with should The output characteristic figure quantity of layer is identical.For convolutional layer l_C,l_C∈ { C1, C2, C3, C4, C5 }, the output characteristic figure width of this layer ForValue together decided on by the size of its input feature vector figure resolution ratio and convolution kernel, i.e.,Represent convolutional layer l_CLast layer convolutional layer Output characteristic figure width；

For convolutional layer C1, C1 layers of output characteristic figure quantity OutputMaps is made^C1=12, the width of output characteristic figure Degree OutputSize^C1=272, convolution kernel width KernelSize^C1=9, offset parameter bias^C1It is initialized as Zero, C1 layer of convolution kernel k^C1Quantity KernelNumber^C1=48, the initial value of each parameter is in convolution kernelRand () is used to generate random number；

For convolutional layer C2, C2 layers of output characteristic figure quantity OutputMaps is made^C2=32, the width of output characteristic figure OutputSize^C2=128, convolution kernel width KernelSize^C2=9, offset parameter is initialized as zero, C2 layers of convolution kernel k^C2Quantity KernelNumber^C2=384, the initial value of each parameter is in convolution kernel

For convolutional layer C3, C3 layers of output characteristic figure quantity OutputMaps is made^C3=32, the width of output characteristic figure OutputSize^C3=56, convolution kernel width KernelSize^C3=9, offset parameter is initialized as zero, C3 layers of convolution kernel k^C3 Quantity KernelNumber^C3=1024, the initial value of each parameter is in convolution kernel

For convolutional layer C4, C4 layers of output characteristic figure quantity OutputMaps is made^C4=32, the width of output characteristic figure OutputSize^C4=20, convolution kernel width KernelSize^C4=9, offset parameter is initialized as zero, C4 layers of convolution kernel k^C4 Quantity KernelNumber^C4=1024, the initial value of each parameter is in convolution kernel

For convolutional layer C5, this layer of OutputMaps is made^C5=32, OutputSize^C5=4, KernelSize^C5=7, partially Put the convolution kernel k that parameter is initialized as zero, C5 layers^C5Quantity KernelNumber^C5=1024, each ginseng in convolution kernel Several initial values are

Step 1-2-1-2, construction down-sampling layer：Down-sampling layer in not comprising need training parameter, by down-sampling layer S1, The sampling core of S2, S3 and S4 is initialized asFor down-sampling layer l_S,l_S∈ { S1, S2, S3, S4 }, its output Characteristic pattern quantityOutput characteristic figure quantity with the convolutional layer of its last layer is consistent, and output characteristic figure is wide DegreeIt is the 1/2 of the output characteristic figure width of the convolutional layer of its last layer, formula is expressed as follows：

Step 1-2-1-3, structural classification device layer：Grader layer is made up of a full articulamentum F1, F1 layers of weighting parameter It is horizontal weighting parameter matrix W H and vertical weighting parameter matrix W V, size is 41 × 512, makes every in weighting parameter matrix The initial value of one parameter isOffset parameter is Horizontal offset parameter BH and vertical off setting parameter BV, It is initialized as 41 × 1 one-dimensional null vector.

Step 1-5 of the present invention is comprised the following steps：

Step 1-5-1, sub-network calculates probability vector：In a sub-network by convolutional layer and the alternate treatment of down-sampling layer Extract input image sequence characteristic, grader layer in processed by Softmax functions, obtain level probability vector HPV with Vertical probability vector VPV；

Step 1-5-2, calculates probabilistic forecasting layer output image：HPV that step 1-5-1 is obtained and VPV are used as probabilistic forecasting Layer convolution kernel, by the last piece image in input image sequence successively with VPV, HPV phase convolution, obtain the defeated of propagated forward Go out prognostic chart picture.

Step 1-5-1 of the present invention is comprised the following steps：

Step 1-5-1-1, judges network layer type：The Internet in the sub-network being presently in, l initial values are represented with l It is C1, judges the type of Internet l, if l ∈ { C1, C2, C3, C4, C5 }, then l is convolutional layer, step 1-5-1-2 is performed, if l ∈ { S1, S2, S3, S4 }, then l is down-sampling layer, performs step 1-5-1-4；

Step 1-5-1-2, processes convolutional layer：Now there is l=l_C,l_C∈ { C1, C2, C3, C4, C5 }, calculates l first_CLayer J-th output characteristic figureBy l_CThe input feature vector figure convolution nuclear phase convolution corresponding with this layer respectively of layer, convolution results are asked With summed result adds l_CJ-th offset parameter of layerProcessed by ReLU activation primitives again, obtainedCalculate public Formula is as follows：

Wherein,It is l_CI-th input feature vector figure of layer convolution kernel corresponding with j-th output characteristic figure, n is current The output characteristic figure number of the previous down-sampling layer of convolutional layer, Represent l_CI-th input of layer Characteristic pattern, while being also l_C- 1 layer of i-th output characteristic figure, * representing matrix convolution, if l_C=C1, then l_C- 1 layer is input Layer；

All of output characteristic figure is calculated successively, obtains l_CThe output characteristic figure of layerL is updated to l+1, and returns to step Rapid 1-5-1-1 judges network type, carries out the operation of next Internet；

Step 1-5-1-3, treatment down-sampling layer：Now there is l=l_S,l_S∈ { S1, S2, S3, S4 }, step 1-5-1-2 is obtained The output characteristic figure of the convolutional layer for arriving respectively withPhase convolution, then sampled with step-length as 2, sampling obtains l_SLayer Output characteristic figureComputing formula is as follows：

In above formula, Sample () represents the sampling processing that step-length is 2, l_S- 1 previous convolution for representing current down-sampling layer Layer,Represent l_SThe output characteristic figure of layerIn j-th output characteristic figure, obtain l_SThe output characteristic figure of layerAfterwards, by l more It is newly l+1, and return to step 1-5-1-1 judges network type, carries out the operation of next Internet；

Step 1-5-1-4, calculates F1 layers of probability vector：If Internet l is grader layer, i.e. l=F1 is become by matrix Change, be 4 × 4 output characteristic figure with row sequential deployment by the 32 width resolution ratio of C5, obtain that resolution ratio is 512 × 1 F1 layers Output characteristic vector a^F1, difference calculated level weighting parameter matrix W H and a^F1Apposition, vertical weighting parameter matrix W V and a^F1's Apposition, result of calculation is sued for peace with Horizontal offset parameter BH, vertical off setting parameter BV respectively, after being processed through Softmax functions To level probability vector HPV and vertical probability vector VPV, specific formula for calculation is as follows：

By its vertical probability vector VPV transposition, final vertical probability vector is obtained；

Step 1-5-2 is comprised the following steps：

Step 1-5-2-1, predicts DC1 layers of vertical direction：By last width input picture of input layer and vertical probability to Amount VPV phase convolution, obtains the DC1 layers of output characteristic figure a that resolution ratio is 240 × 280^DC1；

Step 1-5-2-2, predicts DC2 layers of vertical direction：By DC1 layers of output characteristic figure a^DC1With vertical probability vector HPV phases Convolution, obtains the output prognostic chart picture of propagated forward, and its resolution ratio is 240 × 240.

Step 1-6 of the present invention is comprised the following steps：

Step 1-6-1, calculates probabilistic forecasting layer error term：The prognostic chart picture that step 1-5-2-2 is obtained and the instruction being input into The control label practiced in sample asks poor, calculates DC2 layers, DC1 layers of error term, finally tries to achieve the error term of level probability vector δ^HPVWith the error term δ of vertical probability vector^VPV；

Step 1-6-2, calculates sub-network error term：According to the error term δ of level probability vector^HPVWith vertical probability vector Error term δ^VPV, from it is rear to preceding successively calculate classification layer F1, convolutional layer C5, C4, C3, C2, C1 and down-sampling layer S4, S3, S2, The error term of S1, the resolution ratio of any layer error term matrix tried to achieve is consistent with the resolution ratio of the output characteristic figure of this layer；

Step 1-6-3, calculates gradient：The error term obtained according to step 1-6-2 calculates the mistake of each Internet of sub-network Difference item is to this layer of weighting parameter and the Grad of offset parameter；

Step 1-6-4, undated parameter：The weighting parameter and the ladder of offset parameter of each Internet that step 1-6-3 is obtained Angle value is multiplied by the learning rate of dynamic convolutional neural networks, the renewal of each Internet weighting parameter and offset parameter is obtained, by original Weighting parameter and offset parameter ask poor with the renewal respectively, weighting parameter and offset parameter after being updated.

Step 1-6-1 of the present invention is comprised the following steps：

Step 1-6-1-1, calculates dynamic convolutional layer DC2 error terms：The prognostic chart picture that step 1-5-2-2 is obtained and the group The control label of sample asks poor, obtains the error term matrix delta that size is 240 × 240^DC2；

Step 1-6-1-2, calculates dynamic convolutional layer DC1 error terms：By zero padding by DC2 layers of error term matrix delta^DC2 It is 240 × 320 to expand, by level probability Vector rotation 180 degree, by the level probability after the error term matrix after expansion and upset Vectorial phase convolution, obtains DC1 layers of error term matrix delta^DC1, its size is 240 × 280, and formula is as follows：

δ^DC1=Expand_Zero (δ^DC2) * rot180 (HPV),

Wherein, Expand_Zero () represents zero extended function, and rot180 () represents the rotation letter that angle is 180 ° Number, 4 × 4 matrix, the matrix after zero expansion, region and former square of the central resolution ratio for 2 × 2 are extended for by 2 × 2 matrix zero Consistent, zero pixel filling of remaining position of battle array；

Step 1-6-1-3, calculates probability vector error term：The error term of calculated level probability vector HPV, by DC1 layers Output characteristic figure and error term matrix delta^DC2Phase convolution, obtains 1 × 41 row vector after convolution, the vector is the error term of HPV δ^HPV, formula is as follows：

δ^HPV=a^DC1*δ^DC2,

The error term of vertical probability vector VPV is calculated, by the input feature vector figure of input layer and error term matrix delta^DC1Mutually roll up Product, obtains 41 × 1 column vector after convolution, the vector is the error term δ of VPV^VPV, formula is as follows：

Wherein,Last piece image in for the input image sequence of training sample；

Step 1-6-2 is comprised the following steps：

Step 1-6-2-1, calculates grader layer F1 error terms：The error term of the probability vector that step 1-6-1-3 is obtained δ^VPVAnd δ^HPVWeighting parameter matrix W V vertical with F1 layers and horizontal weighting parameter matrix W H carry out matrix multiple respectively, then by square The apposition of battle array is sued for peace and is averaged, and obtains F1 layers of error term δ^F1, formula is as follows：

Wherein, × representing matrix apposition, ()^TThe transposition of matrix is represented, the δ for obtaining^F1Size be 512 × 1；

Step 1-6-2-2, calculates convolutional layer C5 error terms：By matrixing, F1 layers for being obtained in step 1-6-2-1 Error term δ^F1It is transformed to the matrix that 32 resolution ratio are 4 × 4Obtain C5 layers of error term δ^C5,Table Show the matrix that the 32nd resolution ratio after conversion is 4 × 4；Matrixing in this time matrixing and step 1-5-1-6 is each other Inverse transformation,

Step 1-6-2-3, judges network layer type：Internet (the l initial values in the sub-network being presently in are represented with l It is S4), judge the type of Internet l, if l ∈ { S4, S3, S2, S1 }, then l is down-sampling layer, step 1-6-2-4 is performed, if l ∈ { C4, C3, C2, C1 }, then l is convolutional layer, performs step 1-6-2-5；

Step 1-6-2-4, calculates down-sampling layer error term：If l layers is down-sampling layer, now there is l=l_S,l_S∈{S1,S2, S3, S4 }, for l_SI-th error term matrix of layerBy zero padding respectively by l_S+ 1 layer each error term matrix Expand to width and beWherein l_S+ 1 layer of latter convolutional layer for representing current down-sampling layer, j represents l_S+ 1 layer J-th error term, and haveAgain by corresponding convolution kernelRotation 180 degree, after then expanding Matrix and upset after convolution nuclear phase convolution, and convolution results are sued for peace, obtain l_SI-th error term matrix of layerIt is public Formula is as follows：

Wherein,

All of error term matrix is calculated successively, obtains l_SThe output characteristic figure of layerL is updated to l-1, and returns to step Rapid 1-6-2-3 judges network type, calculates the error term of a upper Internet；

Step 1-6-2-5, calculates convolutional layer error term：If l layers is convolutional layer, now there is l=l_C,l_C∈{C1,C2,C3, C4, C5 }, because the initial value of l in step 1-6-2-3 is S4, therefore be not in l_CThe situation of=C5, for l_CThe i-th of layer Individual error term matrixFirst to l_CCorresponding i-th error term matrix in+1 layerUp-sampled, will during up-samplingIn each element error entry value average mark to sample area in, obtaining resolution ratio is Up-sampling matrix, then calculate activation primitive in l_CDerivative and the inner product for up-sampling matrix tried to achieve at layer character pair figure, Obtain l_CI-th error term matrix of layerFormula is as follows：

Wherein, representing matrix inner product, ReLU'() derivative of ReLU activation primitives is represented, its form is as follows：

UpSample () represents up-sampling function, one up-sampling of each pixel correspondence after up-sampling in original image Region, in each pixel in original pixel value mean allocation to sample area, calculates all of error term matrix successively, obtains To l_CThe output characteristic figure of layer

Step 1-6-2-6, now l layers is convolutional layer, i.e. l=l_C, two kinds of situations are divided into afterwards：

If l ≠ C1, l is updated to l-1, and return to step 1-6-2-3 judges network type, calculates a upper Internet Error term；

If l=C1, step 1-6-2 sub-networks error term is calculated and terminated；

Step 1-6-3 is comprised the following steps：

Step 1-6-3-1, calculates gradient of the convolutional layer error term to convolution kernel：Use l_CRepresent currently processed convolutional layer, l_C ∈ { C1, C2, C3, C4, C5 }, successively calculates gradient of each convolutional layer error term to convolution kernel, by l since C1 layers_C- 1 layer I-th output characteristic figureWith l_CJ-th error term matrix of layerPhase convolution, convolution results are the ladder of correspondence convolution kernel Angle valueFormula is as follows：

In above formula,WithL is represented respectively_CThe output characteristic figure number of layer and l_C- 1 layer of output characteristic figure number, ▽ k are Grad of the error term to convolution kernel；

Step 1-6-3-2, calculates gradient of each convolutional layer error term to biasing：Use l_CRepresent currently processed convolutional layer, l_C ∈ { C1, C2, C3, C4, C5 }, successively calculates gradient of each convolutional layer error term to biasing, by l since C1 layers_CJ-th of layer Error term matrixIn all elements sued for peace, obtain the Grad of j-th of this layer biasingThe following institute of formula Show：

Wherein, Sum () is represented and all elements of matrix is sued for peace；

Step 1-6-3-3, calculates gradient of the F1 layers of error term to weighting parameter：Respectively calculated level probability vector with it is vertical The error term δ of probability vector^HPV、δ^VPVWith F1 layers of error term δ^F1Inner product, result of calculation be F1 layers of error term to weighting parameter WH, The Grad of WV, formula is as follows：

▽ WH=(δ^HPV)^T×(δ^F1)^T,

▽ WV=δ^VPV×(δ^F1)^T,

In above formula, ▽ WH are Grad of the error term to horizontal weighting parameter, and ▽ WV are error term to vertical weighting parameter Grad；

Step 1-6-3-4, calculates gradient of the F1 layers of error term to offset parameter：By level probability it is vectorial with vertical probability to The error term δ of amount^HPV、δ^VPVRespectively as F1 layers of error term to Horizontal offset parameter BH and the Grad of vertical off setting parameter BV, Formula is as follows：

▽ BH=(δ^HPV)^T,

▽ BV=δ^VPV,

Wherein, ▽ BH are Grad of the error term to Horizontal offset parameter, and ▽ BV are error term to vertical off setting parameter Grad；

Step 1-6-4 is comprised the following steps：

Step 1-6-4-1, updates each convolutional layer weighting parameter：Each convolutional layer error term pair that step 1-6-3-1 is obtained The gradient of convolution kernel is multiplied by the learning rate of dynamic convolutional neural networks, obtains the correction term of convolution kernel, then by former convolution kernel and be somebody's turn to do Correction term asks poor, the convolution kernel for being updatedFormula is as follows：

Wherein, λ is the e-learning rate of determination in step 1-3, λ=0.0001；

Step 1-6-4-2, updates each convolutional layer offset parameter：Each convolutional layer error term pair that step 1-6-3-2 is obtained The gradient of biasing is multiplied by the learning rate of dynamic convolutional neural networks, obtains the correction term of offset parameter, then by former bias term and be somebody's turn to do Correction term asks poor, the bias term for being updatedFormula is as follows：

Step 1-6-4-3, updates F1 layers of weighting parameter：The F1 layers of error term that step 1-6-3-3 is obtained is to weighting parameter The Grad of WH and WV is multiplied by the learning rate of dynamic convolutional neural networks, obtains the correction term of weighting parameter, then former weights are joined Number WH and WV asks poor with the correction term tried to achieve respectively, and the WH and WV for being updated, formula are as follows：

WH=WH- λ ▽ WH,

WV=WV- λ ▽ WV；

Step 1-6-4-4, updates F1 layers of offset parameter：The F1 layers of error term that step 1-6-3-4 is obtained is to offset parameter The Grad of BH and BV is multiplied by the learning rate of dynamic convolutional neural networks, obtains the correction term of offset parameter, then original biasing is joined Number BH and BV asks poor with the correction term tried to achieve respectively, and the BH and BV for being updated, formula are as follows：

BH=BH- λ ▽ BH,

BV=BV- λ ▽ BV.

Step 2 of the present invention is comprised the following steps：

Step 2-1, data prediction：Input test image set, specification is carried out to every piece image that test image is concentrated Change is processed, and 280 × 280 floating number image will be converted into per piece image, then floating number image collection is divided, and is constructed Test sample collection comprising TestsetSize group samples；

Step 2-2, read test sample：The TestsetSize groups test sample input that step 2-1 is obtained is by training Dynamic convolutional neural networks in；

Step 2-3, propagated forward：The image sequence characteristic of input is extracted in a sub-network, obtains level probability vector HPV_testWith vertical probability vector VPV_test；Probabilistic forecasting layer in, by the last piece image in input image sequence successively with VPV_test、HPV_testPhase convolution, obtains the final extrapolated image of dynamic convolutional neural networks.

Step 2-1 of the present invention is comprised the following steps：

Step 2-1-1, sampling：The image that test image is concentrated is sequentially arranged, and constant duration is distributed, when Between at intervals of 6 minutes, altogether comprising N_TestWidth image, TestsetSize is determined by equation below：

If Mod (N_Test, 4)=0

If Mod (N_Test, 4) ≠ 0

After trying to achieve TestsetSize, retain test image by sampling and concentrate preceding 4 × TestsetSize+1 width image, adopt Amount of images is set to meet requirement by deleting the last image of test image concentration during sample；

Step 2-1-2, image normalization：Image conversion, normalization operation, by original point are carried out to the image that sampling is obtained Resolution is converted into the floating number image that resolution ratio is 280 × 280 for 2000 × 2000 coloured image；

Step 2-1-3, constructs test sample collection：The floating number image set obtained using step 2-1-2 constructs test sample Collection, by every four adjacent images in floating number image set, i.e. { 4M+1,4M+2,4M+3,4M+4 } width image is defeated as one group Enter sequence, by cutting, the central resolution ratio of reservation is 240 × 240 part as correspondence sample to [4 × (M+1)+1] width image This control label, wherein being positive integer, and has M ∈ [0, TestsetSize-1] to obtain comprising TestsetSize group test specimens This test sample collection；

Step 2-1-2 is comprised the following steps：

Step 2-1-2-1, image conversion：To reduce the computation complexity of convolutional neural networks training, by colored echo Intensity CAPPI images are converted into gray level image, then by cutting the part for retaining that original image center resolution ratio is for 560 × 560, By the image resolution ratio boil down to 280 × 280 after cutting, the gray-scale map that resolution ratio is 280 × 280 is obtained；

Step 2-1-2-2, data normalization：By each pixel in step 1-1-2-1 in the gray-scale map of acquisition Value is mapped to [0~1] from [0~255], by obtaining the floating number image that resolution ratio is 280 × 280 after normalization；

Step 2-3 is comprised the following steps：

Step 2-3-1, calculates sub-network probability vector：In a sub-network by convolutional layer and the alternate treatment of down-sampling layer The image sequence characteristic of input is extracted, is then processed by Softmax functions in grader layer, obtain level probability vector HPV_testWith vertical probability vector VPV_test；

Step 2-3-2, calculates probabilistic forecasting layer output image：The VPV that step 2-3-1 is obtained_testAnd HPV_testAs probability The convolution kernel of prediction interval, by the last piece image in input image sequence successively with VPV_testAnd HPV_testPhase convolution, is moved The final extrapolated image of state convolutional neural networks；

Step 2-3-1 is comprised the following steps：

Step 2-3-1-1, judges network layer type：The Internet in the sub-network being presently in is represented with p, network is judged The type of layer p, p initial values are C1, if p ∈ { C1, C2, C3, C4, C5 }, then p is convolutional layer, step 2-3-1-2 is performed, if p ∈ { S1, S2, S3, S4 }, then p is down-sampling layer, performs step 2-3-1-4；

Step 2-3-1-2, processes convolutional layer：Now there is p=p_C,p_C∈ { C1, C2, C3, C4, C5 }, calculates p first_CLayer V-th output characteristic figureBy p_CThe input feature vector figure convolution nuclear phase convolution corresponding with this layer respectively of layer, then by convolution knot Fruit is sued for peace, and summed result adds p_CV-th biasing of layerProcessed by ReLU activation primitives again, obtain p_CThe v of layer Individual output characteristic figureComputing formula is as follows：

Wherein,It is p_CU-th input feature vector figure of layer convolution kernel corresponding with v-th output characteristic figure, m is current The output characteristic figure number of the previous down-sampling layer of convolutional layer,Represent p_CU-th of layer is defeated Enter characteristic pattern, while being also p_C- 1 layer of u-th output characteristic figure, * representing matrix convolution works as p_CDuring=C1, then p_C- 1 layer is defeated Enter layer.

All of output characteristic figure is calculated successively, obtains p_CThe output characteristic figure of layerP is updated to p+1, and returns to step Rapid 2-3-1-1 judges network type, carries out the operation of next Internet；

Step 2-3-1-3, treatment down-sampling layer：Now there is p=p_S,p_S∈ { S1, S2, S3, S4 }, step 2-3-1-2 is obtained The output characteristic figure of the convolutional layer for arriving respectively withPhase convolution, then sampled with step-length as 2, sampling obtains p_SLayer Output characteristic figureComputing formula is as follows：

Wherein, Sample () represents the sampling processing that step-length is 2, p_S- 1 previous convolution for representing current down-sampling layer Layer,Represent p_SThe output characteristic figure of layerIn j-th output characteristic figure, obtain p_SThe output characteristic figure of layerAfterwards, by p P+1 is updated to, and return to step 2-3-1-1 judges network type, carries out the operation of next Internet；

Step 2-3-1-4, calculates F1 layers of probability vector：If Internet p is grader layer, i.e. p=F1 is become by matrix Change, be 4 × 4 output characteristic figure with row sequential deployment by the 32 width resolution ratio of C5, obtain that resolution ratio is 512 × 1 F1 layers Output characteristic vectorThen respectively calculated level parameter matrix WH, Vertical Parameters matrix W V withApposition, by calculate tie Fruit is sued for peace with Horizontal offset parameter BH, vertical off setting parameter BV respectively, and summed result obtains level after being processed through Softmax functions Probability vector HPV_test, vertical probability vector VPV_test, computing formula is as follows：

By its vertical probability vector VPV_testTransposition, obtains final vertical probability vector；

Step 2-3-2 is comprised the following steps：

Step 2-3-2-1, predicts DC1 layers of vertical direction：By last width input picture of input layer and vertical probability to Amount VPV_testPhase convolution, obtains the DC1 layers of output characteristic figure that resolution ratio is 240 × 280

Step 2-3-2-2, predicts DC2 layers of vertical direction：Step 2-3-2-1 is obtainedWith level probability vector HPV_testPhase convolution, obtains the final extrapolated image of dynamic convolutional neural networks, and its resolution ratio is 240 × 240, by extrapolated image Compared with the control label of corresponding test sample, judge that dynamic convolutional neural networks carry out the accuracy of Radar Echo Extrapolation.

Beneficial effect：Based on dynamic convolutional neural networks, trained on substantial amounts of radar return image set, and utilize The network for training carries out Radar Echo Extrapolation, effectively increases the degree of accuracy of Radar Echo Extrapolation.

It is specifically of the invention to be had the advantage that compared with existing method：1. the utilization rate of Radar Data is high, with traditional thunder Only compared using a small amount of Radar Data up to echo extrapolation technique, the present invention is trained on substantial amounts of radar return image set, number It is higher according to utilization rate；2. extrapolation timeliness is long, can improve outer by the time interval between sample in adjusting training sample set Push away timeliness.

Brief description of the drawings

The present invention is done with reference to the accompanying drawings and detailed description further is illustrated, it is of the invention above-mentioned or Otherwise advantage will become apparent.

Fig. 1 is flow chart of the present invention.

Fig. 2 is dynamic convolutional neural networks initialization model structure chart.

Fig. 3 is sub-network structural map.

Fig. 4 is probabilistic forecasting layer structural map.

Fig. 5 is that matrix zero expands schematic diagram.

Fig. 6 is the process schematic up-sampled to 2 × 2 matrix.

Specific embodiment

Below in conjunction with the accompanying drawings and embodiment the present invention will be further described.

As shown in figure 1, the present invention is comprised the following steps：

Step 1, as shown in Fig. 2 training offline convolutional neural networks：Input training image collection, is carried out to training image collection Data prediction, obtains training sample set, the dynamic convolution neural network structure of design, and initialize network training parameter；Utilize The dynamic convolutional neural networks of training sample set training, the orderly image sequence of input is by dynamic convolutional neural networks propagated forward A width prognostic chart picture is obtained, the error between prognostic chart picture and control label is calculated, the weights of network are updated by backpropagation Parameter and offset parameter, repeat this process until reaching training termination condition, obtain convergent dynamic convolutional neural networks；

Step 1 of the present invention is comprised the following steps：

Step 1-3, the dynamic convolution Neural Network Training Parameter of initialization：E-learning rate λ=0.0001 is made, is instructed Practice the sample size BatchSize=10 of stage each input, most large quantities of frequency of training of training sample setCurrently criticize frequency of training BatchNum=1, the maximum iteration of network training IterationMax=40, current iteration number of times IterationNum=1；

Step 1-4, reads training sample：By the way of batch training, the training sample obtained from step 1-1 is trained every time Concentrate and read BatchSize group training samples, every group of training sample is { x₁,x₂,x₃,x₄, y }, altogether comprising 5 width images, wherein {x₁,x₂,x₃,x₄As input image sequence, x₁Piece image is represented, y is corresponding control label；

Step 1-1-3, constructs training sample set：

Wherein, step 1-1-2 is comprised the following steps：

Step 1-2 of the present invention is comprised the following steps：

Step 1-2-1, constructs sub-network：Sub-network is configured to structure as shown in Figure 3：

Step 1-2-2, construction probabilistic forecasting layer：Probabilistic forecasting layer is configured to structure as shown in Figure 4：

Wherein, step 1-2-1 is comprised the following steps：

Step 1-2-1-1, constructs convolutional layer：Determine herein below：The output characteristic figure quantity OutputMaps of convolutional layer, Convolution kernel k and offset parameter bias.For convolution kernel, it is thus necessary to determine that the width KernelSize of convolution kernel (convolution kernel point Resolution is KernelSize × KernelSize), the quantity KernelNumber of convolution kernel (value be the convolutional layer be input into it is defeated Go out the product of characteristic pattern quantity), and convolution kernel is constructed according to Xavier initial methods；For offset parameter, its quantity with should The output characteristic figure quantity of layer is identical.For convolutional layer l_C,l_C∈ { C1, C2, C3, C4, C5 }, the output characteristic figure width of this layer ForValue together decided on by the size of its input feature vector figure resolution ratio and convolution kernel, i.e.,

For convolutional layer C1, C1 layers of output characteristic figure quantity OutputMaps is made^C1=12, the width of output characteristic figure Degree OutputSize^C1=272, convolution kernel width KernelSize^C1=9, offset parameter bias^C1Zero is initialized as, C1 layers of convolution kernel k^C1Quantity KernelNumber^C1=48, the initial value of each parameter is in convolution kernelRand () is used to generate random number；

For convolutional layer C2, this layer of OutputMaps is made^C2=32, OutputSize^C2=128, KernelSize^C2=9, Offset parameter is initialized as zero, C2 layers of convolution kernel k^C2Quantity KernelNumber^C2=384, each ginseng in convolution kernel Several initial values are

For convolutional layer C3, this layer of OutputMaps is made^C3=32, OutputSize^C3=56, KernelSize^C3=9, partially Put the convolution kernel k that parameter is initialized as zero, C3 layers^C3Quantity KernelNumber^C3=1024, each ginseng in convolution kernel Several initial values are

For convolutional layer C4, this layer of OutputMaps is made^C4=32, OutputSize^C4=20, KernelSize^C4=9, partially Put the convolution kernel k that parameter is initialized as zero, C4 layers^C4Quantity KernelNumber^C4=1024, each ginseng in convolution kernel Several initial values are

Step 1-5 of the present invention is comprised the following steps：

Step 1-5-1 of the present invention is comprised the following steps：

Wherein,It is l_CI-th input feature vector figure of layer convolution kernel corresponding with j-th output characteristic figure, n is current The output characteristic figure number of the previous down-sampling layer of convolutional layer,Represent l_CI-th input of layer Characteristic pattern, while being also l_C- 1 layer of i-th output characteristic figure, * representing matrix convolution, if l_C=C1, then l_C- 1 layer is input Layer；

Step 1-5-1-3, treatment down-sampling layer：Now there is l=l_S,l_S∈ { S1, S2, S3, S4 }, step 1-5-1-2 is obtained The output characteristic figure of the convolutional layer for arriving respectively withPhase convolution, then sampled with step-length as 2, sampling obtains l_S The output characteristic figure of layerComputing formula is as follows：

Step 1-5-2 is comprised the following steps：

Step 1-6 of the present invention is comprised the following steps：

Step 1-6-1 of the present invention is comprised the following steps：

δ^DC1=Expand_Zero (δ^DC2) * rot180 (HPV),

Wherein, Expand_Zero () represents zero extended function, and rot180 () represents the rotation letter that angle is 180 ° Number, 2 × 2 matrix zero is extended for 4 × 4 process as shown in figure 5, zero expand after matrix, central resolution ratio is 2 × 2 Region is consistent with original matrix, zero pixel filling of remaining position；

δ^HPV=a^DC1*δ^DC2,

Wherein,Last piece image in for the input image sequence of training sample；

Step 1-6-2 is comprised the following steps：

Step 1-6-2-2, calculates convolutional layer C5 error terms：By matrixing, F1 layers for being obtained in step 1-6-2-1 Error term δ^F1It is transformed to the matrix that 32 resolution ratio are 4 × 4Obtain C5 layers of error term Represent the matrix that the 32nd resolution ratio after conversion is 4 × 4；This time matrixing is mutual with matrixing in step 1-5-1-6 It is inverse transformation,

Wherein,

UpSample () represents up-sampling function, and Fig. 6 is the process up-sampled to 2 × 2 matrix, after up-sampling One up-sampling region of each pixel correspondence in original image, each picture in original pixel value mean allocation to sample area In vegetarian refreshments, all of error term matrix is calculated successively, obtain l_CThe output characteristic figure of layer

If l=C1, step 1-6-2 sub-networks error term is calculated and terminated；

Step 1-6-3 is comprised the following steps：

In above formula,WithL is represented respectively_CLayer and l_C- 1 layer of output characteristic figure Number, ▽ k are Grad of the error term to convolution kernel；

Wherein, Sum () is represented and all elements of matrix is sued for peace；

▽ WH=(δ^HPV)^T×(δ^F1)^T,

▽ WV=δ^VPV×(δ^F1)^T,

▽ BH=(δ^HPV)^T,

▽ BV=δ^VPV,

Step 1-6-4 is comprised the following steps：

Wherein, λ is the e-learning rate of determination in step 1-3, λ=0.0001；

WH=WH- λ ▽ WH,

WV=WV- λ ▽ WV；

BH=BH- λ ▽ BH,

BV=BV- λ ▽ BV.

Step 2 of the present invention is comprised the following steps：

Step 2-1 of the present invention is comprised the following steps：

If Mod (N_Test, 4)=0

If Mod (N_Test, 4) ≠ 0

Step 2-1-2 is comprised the following steps：

Step 2-3 is comprised the following steps：

Step 2-3-1 is comprised the following steps：

Step 2-3-1-2, processes convolutional layer：Now there is p=p_C,p_C∈ { C1, C2, C3, C4, C5 }, calculates p first_CLayer V-th output characteristic figureBy p_CThe input feature vector figure convolution nuclear phase convolution corresponding with this layer respectively of layer, then by convolution results Summation, summed result adds p_CV-th biasing of layerProcessed by ReLU activation primitives again, obtain p_CV-th of layer Output characteristic figureComputing formula is as follows：

Step 2-3-1-3, treatment down-sampling layer：Now there is p=p_S,p_S∈ { S1, S2, S3, S4 }, step 2-3-1-2 is obtained The output characteristic figure of the convolutional layer for arriving respectively withPhase convolution, then sampled with step-length as 2, sampling obtains p_S The output characteristic figure of layerComputing formula is as follows：

Step 2-3-2 is comprised the following steps：

The invention provides a kind of Radar Echo Extrapolation method based on dynamic convolutional neural networks, the technology is implemented The method and approach of scheme are a lot, and the above is only the preferred embodiment of the present invention, it is noted that for the art Those of ordinary skill for, under the premise without departing from the principles of the invention, some improvements and modifications can also be made, these change Enter and retouch and also should be regarded as protection scope of the present invention.Each part being not known in the present embodiment can use prior art to add To realize.

Claims

1. a kind of Radar Echo Extrapolation method based on dynamic convolutional neural networks, it is characterised in that comprise the following steps：

Step 1, trains offline convolutional neural networks：Input training image collection, data prediction is carried out to training image collection, is obtained Training sample set, the dynamic convolution neural network structure of design, and initialize network training parameter；It is dynamic using training sample set training State convolutional neural networks, the orderly image sequence of input obtains a width prognostic chart by dynamic convolutional neural networks propagated forward Picture, calculates the error between prognostic chart picture and control label, and the weighting parameter and offset parameter of network are updated by backpropagation, This process is repeated until reaching training termination condition, convergent dynamic convolutional neural networks are obtained；

Step 2, online Radar Echo Extrapolation：Input test image set, data prediction is carried out to test chart image set, is tested Sample set, in the dynamic convolutional neural networks that then will be obtained in test sample collection input step 1, by network propagated forward meter Probability vector is calculated, and by last width radar return image in input image sequence and the probability vector phase convolution for obtaining, is obtained To the Radar Echo Extrapolation image of prediction.

2. method according to claim 1, it is characterised in that step 1 is comprised the following steps：

Step 1-1, data prediction：Input training image collection, is carried out at standardization to every piece image that training image is concentrated Reason, will be converted into 280 × 280 floating number image per piece image, floating number image collection be obtained, to floating number image collection Divided, training sample set of the construction comprising TrainsetSize group samples；

Step 1-2, the dynamic convolutional neural networks of initialization：The dynamic convolution neural network structure of design, is configured to generating probability The sub-network of vector, reconstructs the probabilistic forecasting layer for image extrapolation, for the offline neural metwork training stage provides dynamic volume Product neutral net initialization model；

Step 1-3, the dynamic convolution Neural Network Training Parameter of initialization：E-learning rate λ=0.0001 is made, the training stage is each The sample size BatchSize=10 of input, most large quantities of frequency of training of training sample set Currently criticize frequency of training BatchNum=1, the maximum iteration IterationMax=40 of network training, current iteration number of times IterationNum=1；

Step 1-4, reads training sample：By the way of batch training, the training sample that training is obtained from step 1-1 every time is concentrated BatchSize group training samples are read, every group of training sample is { x₁,x₂,x₃,x₄, y }, altogether comprising 5 width images, wherein { x₁,x₂, x₃,x₄Used as input image sequence, y is corresponding control label；

Step 1-5, propagated forward：The input image sequence feature that extraction step 1-4 is obtained in a sub-network, obtains level probability Vectorial HPV and vertical probability vector VPV；Probabilistic forecasting layer in, by the last piece image in input image sequence successively with VPV, HPV phase convolution, obtain the output prognostic chart picture of propagated forward；

Step 1-6, backpropagation：The error term of probability vector is reversely tried to achieve in probabilistic forecasting layer, according to the mistake of probability vector Difference Xiang Conghou successively calculates the error term of each Internet in subnet network layers to preceding, and then calculates error term pair in each Internet The gradient of weighting parameter and offset parameter, using the parameter of the gradient updating dynamic convolutional neural networks for obtaining；

Step 1-7, off-line training step control：Overall control is carried out to the offline neural metwork training stage, is divided into following three kinds Situation：

If training sample is concentrated and still suffers from original training sample, i.e. BatchNum ＜ BatchMax, then return to step 1-4 Continue to read BatchSize group training samples, carry out network training；

If training sample is concentrated does not exist original training sample, i.e. BatchNum=BatchMax, and current network changes Generation number is less than maximum iteration, i.e. IterationNum ＜ IterationMax, then make BatchNum=1, return to step 1-4 continues to read BatchSize group training samples, carries out network training；

If training sample is concentrated does not exist original training sample, i.e. BatchNum=BatchMax, and network iteration time Number reaches maximum iteration, i.e. IterationNum=IterationMax, then terminate the offline neural metwork training stage, obtains To the dynamic convolution neural network model for training.

3. method according to claim 2, it is characterised in that step 1-1 data predictions are comprised the following steps：

Step 1-1-1, sampling：The image that training image is concentrated is sequentially arranged, and constant duration is distributed, between the time It is divided into 6 minutes, altogether comprising N_TrainWidth image, TrainsetSize is determined by equation below：

Wherein, Mod (N_Train, 4) and represent N_TrainTo 4 modulus,Expression is not more thanMaximum integer, try to achieve After TrainsetSize, retain training image by sampling and concentrate preceding 4 × TrainsetSize+1 width image, by deleting during sampling Concentrate last image to meet amount of images except training image to require；

Step 1-1-2, normalized images：Image conversion, normalization operation, by original resolution are carried out to the image that sampling is obtained Coloured image for 2000 × 2000 is converted into the floating number image that resolution ratio is 280 × 280；

Step 1-1-3, constructs training sample set：The floating number image set obtained using step 1-1-2 constructs training sample set, will Every four adjacent images in floating number image set, i.e., { 4N+1,4N+2,4N+3,4N+4 } width image is used as one group of input sequence Row, by cutting, the central resolution ratio of reservation is 240 × 240 part as correspondence sample to [4 × (N+1)+1] width image Control label, for N group samplesIts make is as follows：

\{\begin{matrix} x_{1}^{N} = G_{4 N + 1}, x_{2}^{N} = G_{4 N + 2}, x_{3}^{N} = G_{4 N + 3}, x_{4}^{N} = G_{4 N + 4} \\ y^{N} = C r o p (G_{4 (N + 1) + 1}) \end{matrix},

Wherein, step 1-1-2 is comprised the following steps：

Step 1-1-2-1, image conversion：The image that step 1-1-1 samplings are obtained is converted into gray level image, is retained by cutting Original image center resolution ratio is 560 × 560 part, and the image resolution ratio boil down to 280 × 280 after cutting is divided Resolution is 280 × 280 gray-scale map；

Step 1-1-2-2, data normalization：By in step 1-1-2-1 obtain gray-scale map in each pixel value from [0~255] is mapped to [0~1], by obtaining the floating number image that resolution ratio is 280 × 280 after normalization.

4. method according to claim 3, it is characterised in that step 1-2 is comprised the following steps：

Step 1-2-1, constructs sub-network：

Sub-network is made up of 10 Internets, and convolutional layer C1, down-sampling layer S1, convolutional layer C2, down-sampling are followed successively by from front to back Layer S2, convolutional layer C3, down-sampling layer S3, convolutional layer C4, down-sampling layer S5, convolutional layer C5 and grader layer F1；

Step 1-2-2, construction probabilistic forecasting layer：

Dynamic convolutional layer DC1 and dynamic convolutional layer DC2, the vertical probability vector that sub-network is exported are constructed in probabilistic forecasting layer VPV as dynamic convolutional layer DC1 convolution kernel, level probability vector HPV as dynamic convolutional layer DC2 convolution kernel；

Wherein, step 1-2-1 is comprised the following steps：

Step 1-2-1-1, constructs convolutional layer：Determine herein below：The output characteristic figure quantity OutputMaps of convolutional layer, convolution Core k and offset parameter bias, for convolution kernel, it is thus necessary to determine that the width KernelSize of convolution kernel, the quantity of convolution kernel KernelNumber, the value is the product of convolutional layer input and output characteristic figure quantity, and the resolution ratio of convolution kernel is KernelSize × KernelSize, and convolution kernel is constructed according to Xavier initial methods；For offset parameter, its quantity with The output characteristic figure quantity of this layer is identical；For convolutional layer l_C,l_C∈ { C1, C2, C3, C4, C5 }, the output characteristic figure of this layer is wide Spend and beValue by convolutional layer l_CInput feature vector figure resolution ratio and convolution kernel widthTogether decide on, i.e., Represent volume Lamination l_CLast layer convolutional layer output characteristic figure width；

For convolutional layer C1, C1 layers of output characteristic figure quantity OutputMaps is made^C1=12, the C1 layer of width of output characteristic figure OutputSize^C1=272, C1 layer of convolution kernel width KernelSize^C1=9, C1 layer of offset parameter bias^C1Zero is initialized as, C1 layers of convolution kernel k^C1Quantity KernelNumber^C1=48, the initial value of each parameter is in convolution kernelRand () is used to generate random number；

For convolutional layer C2, C2 layers of output characteristic figure quantity OutputMaps is made^C2=32, the C2 layer of width of output characteristic figure OutputSize^C2=128, C2 layer of convolution kernel width KernelSize^C2=9, C2 layer of offset parameter is initialized as zero, C2 layers Convolution kernel k^C2Quantity KernelNumber^C2=384, the initial value of each parameter is in convolution kernel

For convolutional layer C3, C3 layers of output characteristic figure quantity OutputMaps is made^C3=32, the C3 layer of width of output characteristic figure OutputSize^C3=56, C3 layer of convolution kernel width KernelSize^C3=9, C3 layer of offset parameter is initialized as zero, C3 layers Convolution kernel k^C3Quantity KernelNumber^C3=1024, the initial value of each parameter is in convolution kernel

For convolutional layer C4, C4 layers of output characteristic figure quantity OutputMaps is made^C4=32, the C4 layer of width of output characteristic figure OutputSize^C4=20, C4 layer of convolution kernel width KernelSize^C4=9, C4 layer of offset parameter is initialized as zero, C4 layers Convolution kernel k^C4Quantity KernelNumber^C4=1024, the initial value of each parameter is in convolution kernel

For convolutional layer C5, C5 layers of output characteristic figure quantity OutputMaps is made^C5=32, the C5 layer of width of output characteristic figure OutputSize^C5=4, C5 layer of convolution kernel width KernelSize^C5=7, C5 layer of offset parameter is initialized as zero, C5 layers Convolution kernel k^C5Quantity KernelNumber^C5=1024, the initial value of each parameter is in convolution kernel

Step 1-2-1-2, construction down-sampling layer：Down-sampling layer in not comprising need training parameter, by down-sampling layer S1, S2, The sampling core of S3 and S4 is initialized asFor down-sampling layer l_S,l_S∈ { S1, S2, S3, S4 }, its output is special Levy figure quantityOutput characteristic figure quantity with the convolutional layer of its last layer is consistent, output characteristic figure widthIt is the 1/2 of the output characteristic figure width of the convolutional layer of its last layer, formula is expressed as follows：

\begin{matrix} {OutputMaps}^{l_{S}} = {OutputMaps}^{l_{S} - 1} \\ {OutputSize}^{l_{S}} = \frac{{OutputSize}^{l_{S} - 1}}{2} \end{matrix};

Step 1-2-1-3, structural classification device layer：Grader layer is made up of a full articulamentum F1, and F1 layers of weighting parameter is water Equal rights value parameter matrix W H and vertical weighting parameter matrix W V, size is 41 × 512, makes each in weighting parameter matrix The initial value of parameter isOffset parameter is Horizontal offset parameter BH and vertical off setting parameter BV, just Beginning turns to 41 × 1 one-dimensional null vector.

5. method according to claim 4, it is characterised in that step 1-5 is comprised the following steps：

Step 1-5-1, sub-network calculates probability vector：Extracted by the alternate treatment of convolutional layer and down-sampling layer in a sub-network The image sequence characteristic of input, grader layer in processed by Softmax functions, obtain level probability vector HPV with vertically Probability vector VPV；

Step 1-5-2, calculates probabilistic forecasting layer output image：HPV that step 1-5-1 is obtained and VPV is used as probabilistic forecasting layer Convolution kernel, by the last piece image in input image sequence successively with VPV, HPV phase convolution, the output for obtaining propagated forward is pre- Altimetric image.

6. method according to claim 5, it is characterised in that step 1-5-1 is comprised the following steps：

Step 1-5-1-1, judges network layer type：The Internet in the sub-network being presently in is represented with l, l initial values are C1, Judge the type of Internet l, if l ∈ { C1, C2, C3, C4, C5 }, then l is convolutional layer, performs step 1-5-1-2, if l ∈ S1, S2, S3, S4 }, then l is down-sampling layer, performs step 1-5-1-4；

Step 1-5-1-2, processes convolutional layer：Now there is l=l_C,l_C∈ { C1, C2, C3, C4, C5 }, calculates l first_CThe jth of layer Individual output characteristic figureBy l_CThe input feature vector figure convolution nuclear phase convolution corresponding with this layer respectively of layer, convolution results are sued for peace, Summed result adds l_CJ-th offset parameter of layerProcessed by ReLU activation primitives again, obtainedComputing formula is such as Shown in lower：

a_{j}^{l_{C}} = Re L U [(Σ_{i = 1}^{n} a_{i}^{l_{C} - 1} * k_{i j}^{l_{C}}) + {bias}_{j}^{l_{C}}], j &Element; [1, {OutputMaps}^{l_{C}}],

Wherein,It is l_CI-th input feature vector figure of layer convolution kernel corresponding with j-th output characteristic figure, n is current convolution The output characteristic figure number of the previous down-sampling layer of layer, Represent l_CI-th input feature vector of layer Figure, while being also l_C- 1 layer of i-th output characteristic figure, * representing matrix convolution, if l_C=C1, then l_C- 1 layer is input layer；

All of output characteristic figure is calculated successively, obtains l_CThe output characteristic figure of layerL is updated to l+1, and return to step 1- 5-1-1 judges network type, carries out the operation of next Internet；

Step 1-5-1-3, treatment down-sampling layer：Now there is l=l_S,l_S∈ { S1, S2, S3, S4 }, step 1-5-1-2 is obtained The output characteristic figure of convolutional layer respectively withPhase convolution, then sampled with step-length as 2, sampling obtains l_SLayer it is defeated Go out characteristic patternComputing formula is as follows：

a_{j}^{l_{S}} = S a m p l e (a_{j}^{l_{S} - 1} * [\begin{matrix} 0.25 & 0.25 \\ 0.25 & 0.25 \end{matrix}]), j &Element; [1, {OutputMaps}^{l_{S}}],

In above formula, Sample () represents the sampling processing that step-length is 2, l_S- 1 previous convolutional layer for representing current down-sampling layer, Represent l_SThe output characteristic figure of layerIn j-th output characteristic figure, obtain l_SThe output characteristic figure of layerAfterwards, l is updated to l + 1, and return to step 1-5-1-1 judges network type, carries out the operation of next Internet；

Step 1-5-1-4, calculates F1 layers of probability vector：If Internet l is grader layer, i.e. l=F1, by matrixing, will The 32 width resolution ratio of C5 are 4 × 4 output characteristic figure with row sequential deployment, obtain F1 layer that resolution ratio is 512 × 1 of output spy Levy vectorial a^F1, difference calculated level weighting parameter matrix W H and a^F1Apposition, vertical weighting parameter matrix W V and a^F1Apposition, Result of calculation is sued for peace with Horizontal offset parameter BH, vertical off setting parameter BV respectively, level is obtained after being processed through Softmax functions Probability vector HPV and vertical probability vector VPV, specific formula for calculation is as follows：

\begin{matrix} H P V = S o f t \max (W H \times a^{F 1} + B H) \\ V P V = S o f t \max (W V \times a^{F 1} + B V) \end{matrix},

Step 1-5-2 is comprised the following steps：

Step 1-5-2-1, predicts DC1 layers of vertical direction：By last width input picture of input layer and vertical probability vector VPV Phase convolution, obtains the DC1 layers of output characteristic figure a that resolution ratio is 240 × 280^DC1；

Step 1-5-2-2, predicts DC2 layers of vertical direction：By DC1 layers of output characteristic figure a^DC1Rolled up with level probability vector HPV phases Product, obtains the output prognostic chart picture of propagated forward, and its resolution ratio is 240 × 240.

7. method according to claim 6, it is characterised in that step 1-6 is comprised the following steps：

Step 1-6-1, calculates probabilistic forecasting layer error term：The prognostic chart picture that step 1-5-2-2 is obtained and the training sample being input into Control label in this asks poor, calculates DC2 layers, DC1 layers of error term, finally tries to achieve the error term δ of level probability vector^HPVWith The error term δ of vertical probability vector^VPV；

Step 1-6-2, calculates sub-network error term：According to the error term δ of level probability vector^HPVWith the mistake of vertical probability vector Difference item δ^VPV, classification layer F1, convolutional layer C5, C4, C3, C2, C1 and down-sampling layer S4, S3, S2, S1 are calculated successively from rear to preceding Error term, the resolution ratio of any layer error term matrix tried to achieve is consistent with the resolution ratio of the output characteristic figure of this layer；

Step 1-6-3, calculates gradient：The error term obtained according to step 1-6-2 calculates the error term of each Internet of sub-network To this layer of weighting parameter and the Grad of offset parameter；

Step 1-6-4, undated parameter：The weighting parameter and the Grad of offset parameter of each Internet that step 1-6-3 is obtained The learning rate of dynamic convolutional neural networks is multiplied by, the renewal of each Internet weighting parameter and offset parameter is obtained, by former weights Parameter and offset parameter ask poor with the renewal respectively, weighting parameter and offset parameter after being updated.

8. method according to claim 7, it is characterised in that step 1-6-1 is comprised the following steps：

Step 1-6-1-1, calculates dynamic convolutional layer DC2 error terms：The prognostic chart picture that step 1-5-2-2 is obtained and this group of sample Control label ask poor, obtain size be 240 × 240 error term matrix delta^DC2；

Step 1-6-1-2, calculates dynamic convolutional layer DC1 error terms：By zero padding by DC2 layers of error term matrix delta^DC2Expand It is 240 × 320, by level probability Vector rotation 180 degree, by the level probability vector after the error term matrix after expansion and upset Phase convolution, obtains DC1 layers of error term matrix delta^DC1, its size is 240 × 280, and formula is as follows：

δ^DC1=Expand_Zero (δ^DC2) * rot180 (HPV),

Wherein, Expand_Zero () represents zero extended function, and rot180 () represents the rotation function that angle is 180 °, by 2 × 2 matrix zero is extended for 4 × 4 matrix, the matrix after zero expansion, region and original matrix phase one of the central resolution ratio for 2 × 2 Cause, zero pixel filling of remaining position；

Step 1-6-1-3, calculates probability vector error term：The error term of calculated level probability vector HPV, by DC1 layers of output Characteristic pattern and error term matrix delta^DC2Phase convolution, obtains 1 × 41 row vector after convolution, the vector is the error term δ of HPV^HPV, it is public Formula is as follows：

δ^HPV=a^DC1*δ^DC2,

The error term of vertical probability vector VPV is calculated, by the input feature vector figure of input layer and error term matrix delta^DC1Phase convolution, volume 41 × 1 column vector is obtained after product, the vector is the error term δ of VPV^VPV, formula is as follows：

δ^{V P V} = a_{4}^{i n p u t} * δ^{D C 1},

Wherein,Last piece image in for the input image sequence of training sample；

Step 1-6-2 is comprised the following steps：

Step 1-6-2-1, calculates grader layer F1 error terms：The error term δ of the probability vector that step 1-6-1-3 is obtained^VPVWith δ^HPVWeighting parameter matrix W V vertical with F1 layers and horizontal weighting parameter matrix W H carry out matrix multiple respectively, then by matrix Apposition is sued for peace and is averaged, and obtains F1 layers of error term δ^F1, formula is as follows：

δ^{F 1} = \frac{1}{2} [{(W V)}^{T} \times δ^{V P V} + {(W H)}^{T} \times {(δ^{H P V})}^{T}],

Step 1-6-2-2, calculates convolutional layer C5 error terms：By matrixing, the mistake of F1 layers for being obtained in step 1-6-2-1 Difference item δ^F1It is transformed to the matrix that 32 resolution ratio are 4 × 4Obtain C5 layers of error term δ^C5,Represent and become The 32nd resolution ratio after changing is 4 × 4 matrix；

Step 1-6-2-3, judges network layer type：The Internet in the sub-network being presently in is represented with l, l initial values are S4, Judge the type of Internet l, if l ∈ { S4, S3, S2, S1 }, then l is down-sampling layer, performs step 1-6-2-4, if l ∈ C4, C3, C2, C1 }, then l is convolutional layer, performs step 1-6-2-5；

Step 1-6-2-4, calculates down-sampling layer error term：If l layers is down-sampling layer, now there is l=l_S,l_S∈{S1,S2,S3, S4 }, for l_SI-th error term matrix of layerBy zero padding respectively by l_S+ 1 layer each error term matrixExpand It is to widthWherein l_S+ 1 layer of latter convolutional layer for representing current down-sampling layer, j represents l_SJ-th of+1 layer Error term, and haveAgain by corresponding convolution kernelRotation 180 degree, then by the matrix after expansion With the convolution nuclear phase convolution after upset, and convolution results are sued for peace, obtain l_SI-th error term matrix of layerFormula is as follows It is shown：

δ_{i}^{l_{S}} = Σ_{j = 1}^{n} [E x p a n d_Z e r o ({δ_{j}}^{l_{S} + 1}) * r o t 180 (k_{i j}^{l_{S} + 1})], i &Element; [1, {OutputMaps}^{l_{S}}],

Wherein,

All of error term matrix is calculated successively, obtains l_SThe output characteristic figure of layerL is updated to l-1, and return to step 1- 6-2-3 judges network type, calculates the error term of a upper Internet；

Step 1-6-2-5, calculates convolutional layer error term：If l layers is convolutional layer, now there is l=l_C,l_C∈{C1,C2,C3,C4, C5 }, because the initial value of l in step 1-6-2-3 is S4, therefore be not in l_CThe situation of=C5, for l_CI-th of layer Error term matrixFirst to l_CCorresponding i-th error term matrix in+1 layerUp-sampled, will during up-sampling In each element error entry value average mark to sample area in, obtaining resolution ratio isIt is upper Sampling matrix, then activation primitive is calculated in l_CDerivative and the inner product of the up-sampling matrix tried to achieve at layer character pair figure, obtain l_C I-th error term matrix of layerFormula is as follows：

δ_{i}^{l_{C}} = U p S a m p l e (δ_{i}^{l_{C} + 1}) \cdot {ReLU}^{'} (a_{i}^{l_{C}}), i &Element; [1, {OutputMaps}^{l_{C}}],

{ReLU}^{'} (x) = \{\begin{matrix} 0 & x < 0 \\ 1 & x &GreaterEqual; 0 \end{matrix},

UpSample () represents up-sampling function, the one up-sampling area of each pixel correspondence after up-sampling in original image Domain, in each pixel in original pixel value mean allocation to sample area, calculates all of error term matrix successively, obtains l_CThe output characteristic figure of layer

If l ≠ C1, l is updated to l-1, and return to step 1-6-2-3 judges network type, calculates the mistake of a upper Internet Difference item；

If l=C1, step 1-6-2 sub-networks error term is calculated and terminated；

Step 1-6-3 is comprised the following steps：

Step 1-6-3-1, calculates gradient of the convolutional layer error term to convolution kernel：Use l_CRepresent currently processed convolutional layer, l_C∈ { C1, C2, C3, C4, C5 }, successively calculates gradient of each convolutional layer error term to convolution kernel, by l since C1 layers_CThe i-th of -1 layer Individual output characteristic figureWith l_CJ-th error term matrix of layerPhase convolution, convolution results are the gradient of correspondence convolution kernel ValueFormula is as follows：

&dtri; k_{i j}^{l_{C}} = a_{i}^{l_{C} - 1} * δ_{j}^{l_{C}}, l_{C} &Element; {C 1, C 2, C 3, C 4, C 5}, i &Element; [1, {OutputMaps}^{l_{C} - 1}], j &Element; [1, {OutputMaps}^{l_{C}}],

In above formula,WithL is represented respectively_CThe output characteristic figure number of layer and l_C- 1 layer Output characteristic figure number, ▽ k are Grad of the error term to convolution kernel；

Step 1-6-3-2, calculates gradient of each convolutional layer error term to biasing：Use l_CRepresent currently processed convolutional layer, l_C∈ { C1, C2, C3, C4, C5 }, successively calculates gradient of each convolutional layer error term to biasing, by l since C1 layers_CJ-th mistake of layer Difference item matrixIn all elements sued for peace, obtain the Grad of j-th of this layer biasingFormula is as follows：

&dtri; {bias}_{j}^{l_{C}} = S u m (δ_{j}^{l_{C}}),

Wherein, Sum () is represented and all elements of matrix is sued for peace；

Step 1-6-3-3, calculates gradient of the F1 layers of error term to weighting parameter：Difference calculated level probability vector and vertical probability The error term δ of vector^HPV、δ^VPVWith F1 layers of error term δ^F1Inner product, result of calculation be F1 layers of error term to weighting parameter WH, WV Grad, formula is as follows：

▽ WH=(δ^HPV)^T×(δ^F1)^T,

▽ WV=δ^VPV×(δ^F1)^T,

In above formula, ▽ WH are Grad of the error term to horizontal weighting parameter, and ▽ WV are ladder of the error term to vertical weighting parameter Angle value；

Step 1-6-3-4, calculates gradient of the F1 layers of error term to offset parameter：Level probability is vectorial with vertical probability vector Error term δ^HPV、δ^VPVRespectively as F1 layers of error term to Horizontal offset parameter BH and the Grad of vertical off setting parameter BV, formula It is as follows：

▽ BH=(δ^HPV)^T,

▽ BV=δ^VPV,

Wherein, ▽ BH are Grad of the error term to Horizontal offset parameter, and ▽ BV are gradient of the error term to vertical off setting parameter Value；

Step 1-6-4 is comprised the following steps：

Step 1-6-4-1, updates each convolutional layer weighting parameter：Each convolutional layer error term that step 1-6-3-1 is obtained is to convolution The gradient of core is multiplied by the learning rate of dynamic convolutional neural networks, obtains the correction term of convolution kernel, then by former convolution kernel and the amendment Item asks poor, the convolution kernel for being updatedFormula is as follows：

k_{i j}^{l_{C}} = k_{i j}^{l_{C}} - λ &dtri; k_{i j}^{l_{C}}, l_{C} &Element; {C 1, C 2, C 3, C 4, C 5};

Wherein, λ is the e-learning rate of determination in step 1-3, λ=0.0001；

Step 1-6-4-2, updates each convolutional layer offset parameter：Each convolutional layer error term that step 1-6-3-2 is obtained is to biasing Gradient be multiplied by the learning rate of dynamic convolutional neural networks, obtain the correction term of offset parameter, then by former bias term and the amendment Item asks poor, the bias term for being updatedFormula is as follows：

{bias}_{j}^{l_{C}} = {bias}_{j}^{l_{C}} - λ &dtri; {bias}_{j}^{l_{C}}, l_{C} &Element; {C 1, C 2, C 3, C 4, C 5};

Step 1-6-4-3, updates F1 layers of weighting parameter：By step 1-6-3-3 obtain F1 layer error term to weighting parameter WH with The Grad of WV is multiplied by the learning rate of dynamic convolutional neural networks, obtains the correction term of weighting parameter, then by former weighting parameter WH Ask poor with the correction term tried to achieve respectively with WV, the WH and WV for being updated, formula are as follows：

WH=WH- λ ▽ WH,

WV=WV- λ ▽ WV；

Step 1-6-4-4, updates F1 layers of offset parameter：By step 1-6-3-4 obtain F1 layer error term to offset parameter BH with The Grad of BV is multiplied by the learning rate of dynamic convolutional neural networks, obtains the correction term of offset parameter, then by former offset parameter BH Ask poor with the correction term tried to achieve respectively with BV, the BH and BV for being updated, formula are as follows：

BH=BH- λ ▽ BH,

BV=BV- λ ▽ BV.

9. method according to claim 8, it is characterised in that characterized in that, step 2 is comprised the following steps：

Step 2-1, data prediction：Input test image set, is carried out at standardization to every piece image that test image is concentrated Reason, will be converted into 280 × 280 floating number image per piece image, then floating number image collection is divided, and construction is included The test sample collection of TestsetSize group samples；

Step 2-2, read test sample：The TestsetSize groups test sample input that step 2-1 is obtained is trained dynamic In state convolutional neural networks；

Step 2-3, propagated forward：The image sequence characteristic of input is extracted in a sub-network, obtains level probability vector HPV_testWith Vertical probability vector VPV_test；Probabilistic forecasting layer in, by the last piece image in input image sequence successively with VPV_test、 HPV_testPhase convolution, obtains the final extrapolated image of dynamic convolutional neural networks.

10. method according to claim 9, it is characterised in that step 2-1 is comprised the following steps：

Step 2-1-1, sampling：The image that test image is concentrated is sequentially arranged, and constant duration is distributed, between the time It is divided into 6 minutes, altogether comprising N_TestWidth image, TestsetSize is determined by equation below：

If Mod (N_Test, 4)=0

If Mod (N_Test, 4) ≠ 0

After trying to achieve TestsetSize, retain test image by sampling and concentrate preceding 4 × TestsetSize+1 width image, during sampling Amount of images is set to meet requirement by deleting the last image of test image concentration；

Step 2-1-2, image normalization：Image conversion, normalization operation, by original resolution are carried out to the image that sampling is obtained Coloured image for 2000 × 2000 is converted into the floating number image that resolution ratio is 280 × 280；

Step 2-1-3, constructs test sample collection：The floating number image set obtained using step 2-1-2 constructs test sample collection, will Every four adjacent images in floating number image set, i.e., { 4M+1,4M+2,4M+3,4M+4 } width image is used as one group of input sequence Row, by cutting, the central resolution ratio of reservation is 240 × 240 part as correspondence sample to [4 × (M+1)+1] width image Control label, wherein being positive integer, and has M ∈ [0, TestsetSize-1] to obtain comprising TestsetSize group test samples Test sample collection；

Step 2-1-2 is comprised the following steps：

Step 2-1-2-1, image conversion：Colored echo strength CAPPI images are converted into gray level image, then are protected by cutting It is 560 × 560 part to stay original image center resolution ratio, and the image resolution ratio boil down to 280 × 280 after cutting is obtained Resolution ratio is 280 × 280 gray-scale map；

Step 2-1-2-2, data normalization：By in step 1-1-2-1 obtain gray-scale map in each pixel value from [0~255] is mapped to [0~1], by obtaining the floating number image that resolution ratio is 280 × 280 after normalization；

Step 2-3 is comprised the following steps：

Step 2-3-1, calculates sub-network probability vector：Extracted by the alternate treatment of convolutional layer and down-sampling layer in a sub-network The image sequence characteristic of input, is then processed in grader layer by Softmax functions, obtains level probability vector HPV_test With vertical probability vector VPV_test；

Step 2-3-2, calculates probabilistic forecasting layer output image：The VPV that step 2-3-1 is obtained_testAnd HPV_testAs probabilistic forecasting Layer convolution kernel, by the last piece image in input image sequence successively with VPV_testAnd HPV_testPhase convolution, obtains dynamic volume The final extrapolated image of product neutral net；

Step 2-3-1 is comprised the following steps：

Step 2-3-1-1, judges network layer type：The Internet in the sub-network being presently in is represented with p, Internet p is judged Type, p initial values be C1, if p ∈ { C1, C2, C3, C4, C5 }, then p be convolutional layer, perform step 2-3-1-2, if p ∈ { S1, S2, S3, S4 }, then p is down-sampling layer, performs step 2-3-1-4；

Step 2-3-1-2, processes convolutional layer：Now there is p=p_C,p_C∈ { C1, C2, C3, C4, C5 }, calculates p first_CThe v of layer Individual output characteristic figureBy p_CThe input feature vector figure convolution nuclear phase convolution corresponding with this layer respectively of layer, then convolution results are asked With summed result adds p_CV-th biasing of layerProcessed by ReLU activation primitives again, obtain p_CV-th of layer is defeated Go out characteristic patternComputing formula is as follows：

a_{v}^{p_{C}} = Re L U [(Σ_{u = 1}^{m} a_{u}^{p_{C} - 1} * k_{u v}^{p_{C}}) + {bias}_{v}^{p_{C}}], v &Element; [1, {OutputMaps}^{p_{C}}],

Wherein,It is p_CU-th input feature vector figure of layer convolution kernel corresponding with v-th output characteristic figure, m is current convolution The output characteristic figure number of the previous down-sampling layer of layer, Represent p_CU-th input feature vector of layer Figure, while being also p_C- 1 layer of u-th output characteristic figure, * representing matrix convolution works as p_CDuring=C1, then p_C- 1 layer is input layer.

All of output characteristic figure is calculated successively, obtains p_CThe output characteristic figure of layerP is updated to p+1, and return to step 2- 3-1-1 judges network type, carries out the operation of next Internet；

Step 2-3-1-3, treatment down-sampling layer：Now there is p=p_S,p_S∈ { S1, S2, S3, S4 }, step 2-3-1-2 is obtained The output characteristic figure of convolutional layer respectively withPhase convolution, then sampled with step-length as 2, sampling obtains p_SLayer Output characteristic figureComputing formula is as follows：

a_{v}^{p_{S}} = S a m p l e (a_{v}^{p_{S} - 1} * [\begin{matrix} 0.25 & 0.25 \\ 0.25 & 0.25 \end{matrix}]), v &Element; [1, {OutputMaps}^{p_{S}}],

Wherein, Sample () represents the sampling processing that step-length is 2, p_S- 1 previous convolutional layer for representing current down-sampling layer, Represent p_SThe output characteristic figure of layerIn j-th output characteristic figure, obtain p_SThe output characteristic figure of layerAfterwards, p is updated to P+1, and return to step 2-3-1-1 judges network type, carries out the operation of next Internet；

Step 2-3-1-4, calculates F1 layers of probability vector：If Internet p is grader layer, i.e. p=F1, by matrixing, will The 32 width resolution ratio of C5 are 4 × 4 output characteristic figure with row sequential deployment, obtain F1 layer that resolution ratio is 512 × 1 of output spy Levy vectorThen respectively calculated level parameter matrix WH, Vertical Parameters matrix W V withApposition, result of calculation is distinguished With Horizontal offset parameter BH, vertical off setting parameter BV summation, summed result through Softmax functions process after obtain level probability to Amount HPV_test, vertical probability vector VPV_test, computing formula is as follows：

{HPV}_{t e s t} = S o f t m a x (W H \times a_{t e s t}^{F 1} + B H),

{VPV}_{t e s t} = S o f t m a x (W V \times a_{t e s t}^{F 1} + B V),

Step 2-3-2 is comprised the following steps：

Step 2-3-2-1, predicts DC1 layers of vertical direction：By last width input picture and vertical probability vector of input layer VPV_testPhase convolution, obtains the DC1 layers of output characteristic figure that resolution ratio is 240 × 280

Step 2-3-2-2, predicts DC2 layers of vertical direction：Step 2-3-2-1 is obtainedWith level probability vector HPV_testPhase Convolution, obtains the final extrapolated image of dynamic convolutional neural networks, and its resolution ratio is 240 × 240.