CN102495943B - Modeling method for geophysical model - Google Patents

Abstract

The invention relates to the field of marine informatization, in particular to a modeling method for a geophysical model. The method includes obtaining synchronously measured samples of a radiometer SSM/I and a scatterometer QuickSCAT; selecting a training sample for artificial neural network training; determining a function of the geophysical model; and determining values of unknown coefficients in the function of the geophysical model by the aid of the selected training sample. By the aid of the modeling method for the geophysical model, the high-quality geophysical model can be created.

Description

Modeling method for geophysical model

Technical field

The present invention relates to field of marine informatization, be specifically related to a kind of modeling method for geophysical model.

Background technology

Ocean surface wind field is the basic parameter enlivening Summing Factor ocean dynamics affecting wave, ocean current, water body, to the monitoring of global ocean wind field, preventing and reducing natural disasters in coastland, marine environment guarantee, and promote in the related science research of ocean significant.Satellite scatterometer is with its round-the-clock, round-the-clock, high-spatial and temporal resolution, and the features such as large coverage, have become the of paramount importance observation method of Global ocean wind field.Satellite scatterometer is a kind of radar through calibration, and it to sea active emitting electromagnetic wave, and receives the echoed signal through sea modulation.By the process to radar echo signal, only relevant with sea condition normalization backscattering coefficient (NRCS can be drawn, or σ 0), the σ 0 recorded from scatterometer can extract Ocean Wind-field further, and the information extraction process of Ocean Wind-field is called wind vector retrieval.

The backscattering from ocean surfaces coefficient inverting sea surface wind vector recorded from scatterometer needs the problem of solution three aspects: set up geophysical model, wind vector derivation algorithm, fuzzy solution removes algorithm, wherein, geophysical model is a kind of functional form describing electromagnetic wave and sea interaction mechanism and quantitative relationship thereof, can be divided into physical model and empirical model according to the mode that it is set up.The backscattering coefficient that scatterometer is measured needs just can be finally inversed by sea surface wind vector by concrete geophysics modular function, and therefore the precision of geophysics mould directly decides the quality of Wind-field Retrieval.

The physical geography module function that current operationization is commonly used in running mostly is semiempirical model.Semiempirical model utilizes sea NRCS measured data and synchronous sea surface wind vector field measurement data or Numerical weather forecasting model predictions wind field to set up simultaneous observation sample more, and show that backscattering coefficient is to the statistical relationship between sea surface wind vector by the mode of empirical fit.As the CMOD4 model of C-band, SASS-2, NSCAT-1/2, QSCAT-1 model of Ku wave band.The semiempirical geophysical model set up of this modeling pattern in low wind speed (< 20m/s) and when occurring without rainfall, the sea surface wind vector precision that inverting obtains is: wind speed ± 2m/s, wind direction ± 20 °.For high wind speed or strong wind (>=20m/s) situation, be difficult to set up enough high-quality simultaneous observation samples for setting up geophysical model.Sea surface wind vector field measurement data comprises boat measurement data and buoy measurement data, and under high wind speed condition, sample is on the low side, and precision is not high.It is 0 ~ 20m/s that air quantity journey is surveyed in the wind sensor design that such as TAO buoy group adopts, and when wind speed is more than 30m/s, most of buoy quits work.And the wind field of Numerical weather forecasting model predictions also rarely has high wind speed data, and under high wind speed condition, forecast that wind speed is usually on the low side.The semiempirical geophysical model that this makes said method set up is more higher than actual value in the estimated value of high wind speed district to backscattering coefficient.

Summary of the invention

The invention provides a kind of modeling method for geophysical model, the accuracy of the geophysical model constructed can be improved.

The invention provides a kind of modeling method for geophysical model, comprising:

Obtain the sample of radiometer SSM/I and scatterometer QuikSCAT synchro measure;

The training sample for artificial neural network training is chosen from described sample;

Determine physical geography module function;

The value of the unknowm coefficient in described physical geography module function is determined by the described training sample chosen.

The sample of described acquisition radiometer SSM/I and scatterometer QuikSCAT synchro measure preferably includes:

Reject the sample of rainfall at more than 2mm/h;

SSM/I wind speed and the QuikSCAT wind speed deviation of rejecting same time and same place are the sample of more than 4m/s;

Carry out time match, the extraction Measuring Time of SSM/I and the measuring intervals of TIME of QuikSCAT are less than or equal to the described sample of the synchro measure of 2 hours;

Spatial match is carried out to described sample, by inverse distance weighted interpolation algorithm, SSM/I wind speed is interpolated into QuikSCAT wind vector bin.

The described training sample chosen from described sample for artificial neural network training preferably includes:

A. from described sample, extract wind speed is 16m/s ~ 26m/s, and wind direction is radiometer SSM/I and the scatterometer QuikSCAT synchro measure sample of 0 ° ~ 360 °;

B. the described synchro measure sample that a step is extracted being pressed wind speed interval 1m/s, divides into groups in 20 °, wind direction interval;

C. extract from each group, by the sample composition training sample extracted.

Described extraction from each group preferably includes:

If the number of sample of a group is more than or equal to 10, then random from this group or by setting rule extraction 10 training samples;

If the number of the sample of a group is less than 10, then from this group, repeat extraction 10 training samples at random or by setting rule.

Describedly determine that physical geography module function preferably includes:

According to wind speed V, divide different wind speed sections, described wind speed section comprises: V < 16m/s, 16m/s≤V≤20m/s, V > 20m/s tri-sections;

For different wind speed sections, determine the submodel function of this wind speed section correspondence, described submodel function comprises: QSCAT_1, (QSCAT_1+NN_GMF)/2, NN_GMF, and described QSCAT_1 is scatterometer geophysical model;

The corresponding relation of described submodel function and described wind speed section is:

${\mathrm{\&sigma;}}_0=\left\{\begin{array}{cc}\mathrm{QSCAT}\_1& (V<16m/s)\\ (\mathrm{QSCAT}\_1+\mathrm{NN}\_\mathrm{GMF})/2& (16m/s\&le;V\&le;20m/s)\\ \mathrm{NN}\_\mathrm{GMF}& (V>20m/s)\end{array}\right.;$

The defining method of described NN_GMF submodel function preferably includes:

Determine the topological structure of the described artificial neural network that NN_GMF submodel function is corresponding;

Determine the mathematic(al) representation that the topological structure of described artificial neural network is corresponding.

The topological structure of the described artificial neural network that the described NN_GMF of determination submodel function is corresponding preferably includes:

Determine the number of described topological structure input layer, middle layer, output layer respectively;

Determine the neuronic number of described input layer, middle layer, output layer respectively;

Determine the neuronic function of input layer, middle layer, output layer respectively.

The topological structure of described artificial neural network preferably includes:

An input layer, two middle layers, an output layer;

Described input layer comprises 4 input layers, the corresponding input vector of each input layer, four described input vectors are respectively: the sine value of the angle χ of wind speed V, wave beam and wind vector, the cosine value of described χ, the sine value of incidence angle θ, and wherein, χ meets for beam positional angle, φ is wind direction of ocean surface;

Each described middle layer comprises 6 middle layer neurons, and described middle layer neuron is used for transport function, and described transport function gets Logsig function, and the formula of described Logsig function is: $f\left(x\right)=\frac{1}{1+e^{-x}};$

Described output layer comprises 1 output layer neuron, and described output layer neuron is normalization backscattering coefficient σ 0, and output function is linear convergent rate function, and the formula of described output function is: f (x)=x+b;

And/or

The mathematic(al) representation that described artificial neural network topological structure is corresponding preferably includes:

${\mathrm{\&sigma;}}_0=\underset{i=\mathrm{1,6}}{\mathrm{\&Sigma;}}({\mathrm{LW}1}_i\&times;{\mathrm{VH}2}_i)+b2;$

$\mathrm{VH}2\left(i\right)=1/(1+\exp (-\underset{j=\mathrm{1,6}}{\mathrm{\&Sigma;}}({\mathrm{LW}}_{i,j}\&times;{\mathrm{VH}1}_j)+{b1}_i));$

$\mathrm{VH}1\left(i\right)=1/(1+\exp (-\underset{j=1,4}{\mathrm{\&Sigma;}}({\mathrm{IW}}_{i,j}\&times;{\mathrm{VI}}_j)+b_i));$

VI ₁=sin (χ), VI ₂=cos (χ), VI ₃=sin (θ), VI ₄=(V-15)/20; Wherein, b, b1, b2, IW, LW are described unknowm coefficient.

The value of the described unknowm coefficient determined in described physical geography module function by the described training sample chosen is preferably included:

Described V, χ, θ, σ 0 is extracted from described training sample;

Adopt Quasi-Newton algorithm to carry out matching to described mathematic(al) representation, obtain the value of unknowm coefficient.

Preferably comprise further after the described described training sample by choosing determines the value of the unknowm coefficient in described physical geography module function:

Verify by the value of other samples to described unknowm coefficient in described sample except described training sample.

By a kind of modeling method for geophysical model provided by the invention, following beneficial effect can be reached:

1. improve the accuracy of geophysical model, build high-quality geophysical model.The synchro measure sample that the present invention is obtained by radiometer SSM/I and scatterometer QuikSCAT synchro measure data, wherein there is not " saturated " effect in radiometer SSM/I under high wind speed condition, therefore, forecast that wind field or buoy data set up synchro measure sample with employing more accurate; Simultaneously, measure due to radiometer and scatterometer and all there is certain fabric width, therefore, adopt radiometer measurement result and scatterometer measurement result to carry out sampling and can obtain more synchro measure sample, more can meet a large amount of training samples set up required for geophysical model (geophysical model especially under higher wind velocity condition).

2. model calculation efficiency is high.The present invention adopts Artificial Neural Network Modeling, without any need for a priori assumption; Meanwhile, artificial neural network has stronger nonlinear fitting approximation capability, and holds good characteristic that is incomplete, denoising, for the modeling of space scatterometer Ocean Wind-field provides a kind of brand-new technological means.In addition, consider that incidence angle θ variation range is 0-90 °, when designer's artificial neural networks topological structure, only adopt V, sin (χ), cos (χ), sin (θ) four inputs parameter, remove cos of the prior art (θ) and input parameter, simplify the topological structure of artificial neural network, the operation efficiency of geophysical model is improved.

3. more accurate when describing the relation between backscattering coefficient and sea surface wind vector.The present invention adopts strong wind geophysical model (NN_GMF) pattern function when wind speed >=20m/s, backscattering coefficient obviously strengthens the sensitivity that wind speed changes, and predicts that higher defect has carried out effective correction to existing QSCAT_1 backscattering coefficient; In addition, by the wind speed of NN_GMF inverting gained, under high wind speed condition, carried out effective correction to existing QSCAT_1 model to the situation of wind speed undervalued, result is more close to the analysis result again that American National Hurricane Center is drawn by Optimal route analysis.

Accompanying drawing explanation

In order to be illustrated more clearly in the embodiment of the present invention or technical scheme of the prior art, be briefly described to the accompanying drawing used required in embodiment or description of the prior art below, apparently, accompanying drawing in below describing is only some embodiments of the present invention, for those of ordinary skills, under the prerequisite not paying creative work, other embodiment and accompanying drawing thereof can also be obtained according to these accompanying drawing illustrated embodiments.

Fig. 1 is the step schematic diagram of modeling method for geophysical model of the present invention.

Fig. 2 is the step schematic diagram of the present invention's optimal enforcement example.

Fig. 3 is NN_GMF artificial neural network topological structure schematic diagram of the present invention.

Fig. 4 is that the backscattering coefficient predicted of the QSCAT-1 of geophysical model NN-T-GMF of the present invention and traditional geophysical model Ku wave band is with wind speed situation of change comparison diagram.

Fig. 5 is the backscattering coefficient box haul variation diagram that geophysical model NN-T-GMF of the present invention predicts.

Fig. 6 is the backscattering coefficient box haul variation diagram that in prior art, geophysical model QSCAT-1 predicts.

Fig. 7 is the variation diagram utilizing the geophysical model QSCAT-1 of Ku wave band in prior art and geophysical model NN-T-GMF model inversion of the present invention to go out the typhoon Ioke intensity time respectively.

Embodiment

Carry out clear, complete description below with reference to accompanying drawing to the technical scheme of various embodiments of the present invention, obviously, described embodiment is only a part of embodiment of the present invention, instead of whole embodiments.Based on the embodiment in the present invention, other embodiments all that those of ordinary skill in the art obtain under the prerequisite not making creative work, all belong to the scope that the present invention protects.

The invention provides a kind of modeling method for geophysical model, comprising:

See Fig. 1, step 101, obtain the sample of radiometer SSM/I and scatterometer QuikSCAT synchro measure;

See Fig. 1, step 102, from described sample, choose the training sample for artificial neural network training;

See Fig. 1, step 103, determine physical geography module function;

See Fig. 1, step 104, determined the value of the unknowm coefficient in described physical geography module function by the described training sample chosen.

The whitecap that under higher wind velocity condition, Wave Breaking produces covers and foam lamellae, and this foam lamellae makes microwave reflection rate in sea increase, and the bright temperature finally causing radiometer (microwave radiometer) to receive can continue to increase with the increase of wind speed.By the synchro measure sample that radiometer SSM/I and scatterometer QuikSCAT synchro measure data obtain, wherein there is not " saturated " effect in radiometer SSM/I under high wind speed condition, therefore, forecast that wind field or buoy data set up synchro measure sample with employing more accurate.

Fig. 1, step 101, obtains the sample of radiometer SSM/I and scatterometer QuikSCAT synchro measure:

Set up geophysical model first to need to build backscattering coefficient and incidence angle θ (beams incident angle θ), beam positional angle ocean surface wind speed V, wind direction of ocean surface φ, synchro measure sample, this synchro measure sample is the sample obtaining radiometer SSM/I and scatterometer QuikSCAT synchro measure, first the data of selective radiometer SSM/I and scatterometer QuikSCAT synchro measure is needed, the present invention have employed the data information of radiometer SSM/I between 2004-2006 and scatterometer QuikSCAT synchro measure in typhoon in a specific embodiment, is chosen data information by time match, spatial match and some kick-out condition.

In a preferred embodiment of the invention, this step 101 specifically can be realized to 101.4 processes by the step 101.1 shown in Fig. 2, comprising:

See Fig. 2, step 101.1, time match: for reaching radiometer SSM/I and scatterometer QuikSCAT coupling in time, the described sample of the sample chosen only adopted the measuring intervals of TIME of SSM/I and QuikSCAT within the 2 hours synchro measure of (time interval is less than or equal to 2 hours);

See Fig. 2, step 101.2, spatial match: be reach radiometer SSM/I and scatterometer QuikSCAT coupling spatially to described sample, by inverse distance weighted interpolation algorithm, SSM/I wind speed is interpolated into QuikSCAT wind vector bin;

Then the sample of rejecting also is needed to comprise: (1) accepts a surrender the sample that rain shadow rings, and such as, see Fig. 2, step 101.3, same time, same place rainfall amount are at the sample of more than 2mm/h; (2) see Fig. 2, step 101.4, reject SSM/I wind speed and the QuikSCAT wind speed deviation sample at more than 4m/s.

The order of Fig. 2, step 101.1-101.4 can be exchanged, and such as first can carry out 101.3 and 101.4, then carry out 101.1 and 101.2.

Sample comprises interior wave beam sample and outer wave beam sample, and interior wave beam and outer wave beam are because the direction of illumination that light is different causes.Finally obtain 21234 simultaneous observation samples through choosing, wherein the sample of corresponding interior wave beam amounts to 9184,12140, outer wave beam.

Measure due to radiometer and scatterometer and all there is certain fabric width, therefore, radiometer (microwave radiometer) measurement result and scatterometer measurement result is adopted to carry out time-space registration, relative to scatterometer, prior art is adopted to the time-space registration of buoy spot measurement data, be easier to obtain a large amount of matching results, more can meet a large amount of training samples set up required for geophysical model (geophysical model especially under higher wind velocity condition).

Fig. 1, step 102, from described sample, choose the training sample for artificial neural network training:

Present invention employs artificial neural network and build geophysical model, therefore need the training sample chosen from the synchro measure sample obtained for artificial neural network training;

In a preferred embodiment of the invention, this step 102 specifically can be realized to 102.3 processes by the step 102.1 shown in Fig. 2, comprising:

Analyze the velocity distribution of the synchro measure sample obtained in table 1:

Table 1 simultaneous observation sample wind speed profile

According to the record of table 1, when wind speed is greater than 26m/s, sample size in the interval divided by 1m/s is less than 20, enough samples cannot be provided for artificial neural network training, therefore, see Fig. 2, step 102.1, from described sample, extract wind speed is 16m/s ~ 26m/s, and wind direction is radiometer SSM/I and the scatterometer QuikSCAT synchro measure sample of 0 ° ~ 360 °, forms effective synchro measure sample, on the other hand, because total number of samples amount is numerous, need to divide into groups to all samples before choosing training sample, divide into groups with certain wind speed and direction, such as: see Fig. 2, step 102.2, according to the sample that above-mentioned data is extracted, according to wind speed interval 1m/s, synchro measure sample divides into groups by 20 °, wind direction interval, be divided into some subsamples group, then Fig. 2 is seen, step 102.3, extract from each group (subsample group), by the sample composition training sample extracted, this extraction mode for randomly draw or by setting rule extraction, such as sample is sorted, extract by odd number or even number, or extract in sequence, the sample size extracted can rule of thumb be determined with the sample number in each group, the sample size extracted in a specific embodiment is wherein 10, there are two kinds of situations: a kind of situation is the sample size (such as 10) that sample number in group exceedes needs extraction in extraction process, mode that is random or that extract by setting rule extraction can be taked when extracting, another kind of situation is the sample size (such as 10) that sample number in group is less than needs and extracts, and can adopt random or repeat the mode of extraction by setting rule extraction when extracting.

Wherein, see Fig. 1, step 105, after extracting complete training sample, other samples remaining can be verified after the value of each unknowm coefficient of physical geography module function is determined by training.

It should be noted that, due to sample number quantitative limitation, the present invention is in specific implementation process, due to the incident angle fluctuating range of QuikSCAT very little (< 1 °), interior beams incident angle and outer wave beam incident angle all can be processed into definite value, such as, interior beams incident angle is 46 °, outer wave beam incident angle is 54 °, does such process, can not cause too large error.

Fig. 1, step 103, determine physical geography module function:

After obtaining the sample of complete radiometer SSM/I and scatterometer QuikSCAT synchro measure, need to determine physical geography module function, this physical geography module function preferably can describe backscattering coefficient σ ₀with incidence angle θ (beams incident angle θ), beam positional angle ocean surface wind speed V, the physical geography module function of relation between wind direction of ocean surface φ, and the training sample by choosing, utilize the method being similar to statistical fit to determine the value of each unknowm coefficient changed in physical geography module function.

In a preferred embodiment of the invention, in Fig. 1, step 103 specifically can be realized to 103.4 processes by the step 103.1 shown in Fig. 2, comprising:

Because wind speed is different, backscattering coefficient is different, and the change of backscattering coefficient and wind speed change non-linear change, therefore need to determine physical geography module function according to different wind speed and corresponding backscattering coefficient.See Fig. 2, step 103.1, according to wind speed V, divide different wind speed sections, wind speed is divided into three sections by the present invention, and three wind speed sections comprise: V < 16m/s, 16m/s≤V≤20m/s, V > 20m/s tri-sections, see Fig. 2, step 103.2, for different wind speed sections, determine the submodel function of this wind speed section correspondence, during due to wind speed V < 16m/s, for low wind conditions, prior art scatterometer geophysical model QSCAT-1 stands good, therefore the corresponding QSCAT-1 submodel function of V < 16m/s, during V > 20m/s, the present invention adopts new submodel function NN_GMF, in the interval method by linear interpolation of 16m/s≤V≤20m/s, the backscattering coefficient that the high wind speed physical geography module function finally obtained is predicted keeps continuously to the change of wind speed, smoothly.

The present invention adopts NN_GMF pattern function when wind speed V > 20m/s, and backscattering coefficient obviously strengthens the sensitivity that wind speed changes, and predicts that higher defect has carried out effective correction to existing QSCAT_1 backscattering coefficient; In addition, by the wind speed of NN_GMF inverting gained, under high wind speed condition, carried out effective correction to existing QSCAT_1 model to the situation of wind speed undervalued, result is more close to the analysis result again that American National Hurricane Center is drawn by Optimal route analysis.

Therefore the corresponding relation of described submodel function and described wind speed section is preferably:

${\mathrm{\&sigma;}}_0=\left\{\begin{array}{cc}\mathrm{QSCAT}\_1& (V<16m/s)\\ (\mathrm{QSCAT}\_1+\mathrm{NN}\_\mathrm{GMF})/2& (16m/s\&le;V\&le;20m/s)\\ \mathrm{NN}\_\mathrm{GMF}& (V>20m/s)\end{array}\right.;$

The defining method of submodel function NN_GMF comprises:

See Fig. 2, step 103.3, determine the topological structure of the described artificial neural network that NN_GMF submodel function is corresponding;

Namely, the backscattering coefficient that the wind speed adopting radiometer SSM/I to record and scatterometer synchro measure obtain, and the wind direction of scatterometer tradition inversion method gained sets up training sample, and trained the geophysical model (NN_GMF) drawn by the method for artificial neural network.

The design of artificial neural network need provide input vector and target vector according to particular problem, and the selected artificial neural network structure that will design, wherein mainly comprise the network number of plies, the neuron number of every layer, the function of every layer.The number of plies of network and every layer of neuronic number are determined by experience and experiment, target makes network topology structure simple as far as possible, specific to geophysical model modeling here, to set up exactly for artificial neural network training training sample and set up backscattering coefficient σ can be described ₀with incidence angle θ (beams incident angle θ), beam positional angle the neural network of relation between ocean surface wind speed V.Can ensure accurately to give expression to physical geography module function simultaneously.Artificial neural network of the present invention is mainly used in the foundation supporting NN_GMF submodel function.The topological structure of the described artificial neural network that the described NN_GMF of determination submodel function is corresponding preferably includes:

Determine the number of described topological structure input layer, middle layer, output layer respectively;

Determine the neuronic number of described input layer, middle layer, output layer respectively;

Determine the neuronic function of input layer, middle layer, output layer respectively.

In a specific embodiment, as shown in Figure 3, the topological structure of described artificial neural network preferably includes:

An input layer, two middle layers, an output layer;

Described input layer comprises 4 input layers, the corresponding input vector of each input layer, four described input vectors are respectively: the sine value of the angle χ of wind speed V, wave beam and wind vector, the cosine value of described χ, the sine value of incidence angle θ, and wherein, χ meets for beam positional angle, φ is wind direction of ocean surface;

Each described middle layer comprises 6 middle layer neurons, and described middle layer neuron is used for transport function, and described transport function gets Logsig function, and the formula of described Logsig function is:

Described output layer comprises 1 output layer neuron, and described output layer neuron is normalization backscattering coefficient σ 0, and output function is linear convergent rate function, and the formula of described output function is: f (x)=x+b.

Existing experience geophysical model many employings experience fits method, and according to wind speed, wind direction, frequency, polarization mode, the priori between incident angle and backscattering coefficient, utilizes a large amount of measured data to carry out matching to the parameter in empirical model, and its general type is σ ⁰=F (u, χ ..., f, p, θ).The present invention adopts Artificial Neural Network Modeling, without any need for a priori assumption; Meanwhile, artificial neural network has stronger nonlinear fitting approximation capability, and holds good characteristic that is incomplete, denoising, for the modeling of space scatterometer Ocean Wind-field provides a kind of brand-new technological means.In addition, consider that incidence angle θ variation range is 0-90 °, when designer's artificial neural networks topological structure, only adopt V, sin (χ), cos (χ), sin (θ) four inputs parameter, remove cos of the prior art (θ) and input parameter, simplify the topological structure of artificial neural network, the operation efficiency of geophysical model is improved.

See Fig. 2, step 103.4, determine the mathematic(al) representation that the topological structure of described artificial neural network is corresponding, carry out mathematical expression according to above-mentioned artificial neural network topological structure, the mathematic(al) representation that described artificial neural network topological structure is corresponding preferably includes:

${\mathrm{\&sigma;}}_0=\underset{i=\mathrm{1,6}}{\mathrm{\&Sigma;}}({\mathrm{LW}1}_i\&times;{\mathrm{VH}2}_i)+b2;$

$\mathrm{VH}2\left(i\right)=1/(1+\exp (-\underset{j=\mathrm{1,6}}{\mathrm{\&Sigma;}}({\mathrm{LW}}_{i,j}\&times;{\mathrm{VH}1}_j)+{b1}_i));$

$\mathrm{VH}1\left(i\right)=1/(1+\exp (-\underset{j=1,4}{\mathrm{\&Sigma;}}({\mathrm{IW}}_{i,j}\&times;{\mathrm{VI}}_j)+b_i));$

VI ₁=sin (χ), VI ₂=cos (χ), VI ₃=sin (θ), VI ₄=(V-15)/20; Wherein, b, b1, b2, IW, LW are described unknowm coefficient.

Fig. 1, step 104, the value of the unknowm coefficient in described physical geography module function is determined by the described training sample chosen:

In a preferred embodiment of the invention, in Fig. 1, step 104 specifically can be realized to 104.2 processes by the step 104.1 shown in Fig. 2, comprising:

The value of the described unknowm coefficient determined in described physical geography module function by the described training sample chosen is preferably included: see Fig. 2, step 104.1, extract described V, χ, θ, σ 0 from described training sample; See Fig. 2, step 104.2, adopt Quasi-Newton algorithm to carry out matching to described mathematic(al) representation, obtain the value of unknowm coefficient.Described Quasi-Newton algorithm is prior art.

Result is as follows:

$\mathrm{IW}=\left|\begin{array}{cccc}67.654& -8.035& 177.1381& 49.8384\\ -3.1209& 811.2122& -12.4052& -0.0498\\ 2.5999& -1.4096& 194.8691& 34.9457\\ 0.0208& 0.1273& 0.0005& -0.0929\\ 0.1259& 0.9328& 0.6929& 0.0452\\ -78.012& 16.4497& -17.4571& 13.1691\end{array}\right|$

$\mathrm{LW}=\left|\begin{array}{cccccc}99.7142& -2.6201& 6.6903& -13.3256& -21.5977& 9.1807\\ 100.3395& -1.7801& 7.5101& 19.1205& 105.5641& -19.4441\\ 15.3341& 4.6572& 12.6084& 13.3056& 4.5952& 0.1973\\ 0.7201& -0.5156& -0.3554& 92.5579& -42.7568& 60.2125\\ -5.343& -3.7561& 9.3147& -11.3024& -0.4588& 4.11\\ -33.7642& -2.9526& 130.6677& 32.6852& -29.0071& -3.2803\end{array}\right|$

LW1＝[90.0879 -91.4875 -7.3625 -11.8092 -4.2241 -0.6872]

b1＝[6.7153 -30.0841 7.3050 -47.2949 -16.6059 -11.6834]

b2＝[-139.8543 -33.0271 -150.0197 0.1593 -3.2267 -11.6549]

After obtaining the numerical value of unknowm coefficient by artificial neural network training, Fig. 1 can be seen, step 105, verify by the value of other samples to described unknowm coefficient in described sample except described training sample, specifically can be realized to 105.2 processes by the step 105.1 shown in Fig. 2, comprise: see Fig. 2, step 105.1, described V is extracted in other samples from described sample except described training sample, χ, θ, see Fig. 2, step 105.2, verify by the value of other samples to described unknowm coefficient in described sample except described training sample, namely Fig. 2 is utilized, the V that step 105.1 is extracted, χ, θ, the value of described unknowm coefficient is verified.

The present invention adopts NN_GMF pattern function when wind speed >=20m/s, and backscattering coefficient obviously strengthens the sensitivity that wind speed changes, and predicts that higher defect has carried out effective correction to existing QSCAT_1 backscattering coefficient; In addition, by the wind speed of NN_GMF inverting gained, under high wind speed condition, carried out effective correction to existing QSCAT_1 model to the situation of wind speed undervalued, result is more close to the analysis result again that American National Hurricane Center is drawn by Optimal route analysis.

Wherein, by the geophysical model (NN-T-GMF) of method establishment of the present invention, than existing geophysical model, more accurate when describing the relation between backscattering coefficient and sea surface wind vector.This can be verified by following two aspects.

(1) backscattering coefficient trend comparison, as shown in Figure 4

Fig. 4 illustrates the comparison diagram of the backscattering coefficient of geophysical model of the present invention (NN-T-GMF) prediction and traditional prior art QSCAT-1 model prediction.The relation that backscattering coefficient under the outer wave beam incident angle of QuikSCAT of model prediction changes with wind speed is given in figure.Wherein be with the corresponding QSCAT-1 model of triangle curve, band circle point curve NN-T-GMF.As can be seen from the figure, when wind speed is greater than 20m/s, backscattering coefficient obviously reduces the sensitivity that wind speed changes, and even trends towards saturated when wind speed is greater than 30m/s.And NN-T-GMF at the backscattering coefficient of strong wind area lower than QSCAT-1, QSCAT-1 backscattering coefficient is predicted that higher defect has obvious correction.

Fig. 5 and Fig. 6 sets forth the relation of backscattering coefficient box haul change under the condition of beams incident outside that NN-T-GMF and QSCAT-1 predicts under different wind friction velocity.As can be seen from figure equally, to backscattering coefficient, NN-T-GMF predicts that higher defect has obvious correction to QSCAT-1 model under high wind speed condition.

(2) wind vector retrieval interpretation of result

For the Wind-field Retrieval of QuikSCAT to typhoon IOKE (2006), the validity of geophysical model NN-T-GMF of the present invention is described.From on August 19th, 2006 to September 6, QuikSCAT has comparatively intactly observed typhoon IOKE totally 17 times.To each complete observed result, utilize QSCAT-1 model and NN-T-GMF model inversion wind vector respectively, and from inversion result, extract the variation diagram (as shown in Figure 7) of the intensity time of typhoon IOKE.Result shows, by NN-T-GMF inverting gained wind speed, have obvious correction, and result is more close to the analysis result again that American National Hurricane Center is drawn by Optimal route analysis to employing QSCAT-1 model under high wind speed condition to wind speed undervalued.

A rail observation data (orbit number is for 37474) for QuikSCAT on August 30th, 2006 to typhoon IOKE, adopt QSCAT-1 model inversion acquired results maximum wind power to be 36.8m/s, the maximum wind power issued far below American National Hurricane Center is the result of 65m/s; Adopt the result that NN-T-GMF inverting obtains, its typhoon wind vector maximum wind power reaches 55.1m/s, and than the result adopting QSCAT-1 model inversion, precision large increase, illustrates that NN-T-GMF is effective to QSCAT-1 in the correction of Gao Fengqu.

Simultaneously, to extract the Typhoon Wind Field that draws for standard by American National Hurricane Center by Optimal route analysis, statistical study is carried out to Wind-field Retrieval result, result is as shown in table 2, and result shows: the wind field average drawn by QACAT-1 model inversion is 20.2m/s, and root-mean-square error is 11.4m/s, relative error is 25.7%, the wind field average adopting geophysical model NN-T-GMF of the present invention inverting to draw is 23.4m/s, and root-mean-square error is 8.7m/s, and relative error is 18.1%.Statistic analysis result illustrates, adopts the Typhoon Wind Field that obtains of strong wind model NN-T-GMF inverting revised to have compared with the wind field precision that QSCAT-1 model inversion draws and significantly improves.

Table 2 is compared by the wind speed gone out of NN-T-GMF and QSCAT-1 inverting

By the invention provides a kind of modeling method for geophysical model, following beneficial effect can be reached:

1. improve the accuracy of geophysical model, build high-quality geophysical model.The synchro measure sample that the present invention is obtained by radiometer SSM/I and scatterometer QuikSCAT synchro measure data, wherein there is not " saturated " effect in radiometer SSM/I under high wind speed condition, therefore, forecast that wind field or buoy data set up synchro measure sample with employing more accurate; Simultaneously, measure due to radiometer and scatterometer and all there is certain fabric width, therefore, adopt radiometer measurement result and scatterometer measurement result to carry out sampling and can obtain more synchro measure sample, more can meet a large amount of training samples set up required for geophysical model (geophysical model especially under higher wind velocity condition).

2. model calculation efficiency is high.The present invention adopts Artificial Neural Network Modeling, without any need for a priori assumption; Meanwhile, artificial neural network has stronger nonlinear fitting approximation capability, and holds good characteristic that is incomplete, denoising, for the modeling of space scatterometer Ocean Wind-field provides a kind of brand-new technological means.In addition, consider that incidence angle θ variation range is 0-90 °, when designer's artificial neural networks topological structure, only adopt V, sin (χ), cos (χ), sin (θ) four inputs parameter, remove cos of the prior art (θ) and input parameter, simplify the topological structure of artificial neural network, the operation efficiency of geophysical model is improved.

3. more accurate when describing the relation between backscattering coefficient and sea surface wind vector.The present invention adopts NN_GMF pattern function when wind speed >=20m/s, and backscattering coefficient obviously strengthens the sensitivity that wind speed changes, and predicts that higher defect has carried out effective correction to existing QSCAT_1 backscattering coefficient; In addition, by the wind speed of NN_GMF inverting gained, under high wind speed condition, carried out effective correction to existing QSCAT_1 model to the situation of wind speed undervalued, result is more close to the analysis result again that American National Hurricane Center is drawn by Optimal route analysis.

Various embodiment provided by the invention can combine as required in any way mutually, the technical scheme obtained by this combination, also within the scope of the invention.

Obviously, those skilled in the art can carry out various change and modification to the present invention and not depart from the spirit and scope of the present invention.Like this, if belong within the scope of the claims in the present invention and equivalent technologies thereof to these amendments of the present invention and modification, then the present invention also comprises these change and modification.

Claims (9)

1. a modeling method for geophysical model, is characterized in that, comprising:

Obtain the sample of radiometer SSM/I measurement and the sample of scatterometer QuikSCAT measurement;

The sample that the sample measure described radiometer SSM/I and described scatterometer QuikSCAT measure carries out time-space registration, obtains the sample of synchro measure;

From the sample of described synchro measure, extract wind speed is 16m/s ~ 26m/s, and wind direction is radiometer SSM/I and the scatterometer QuikSCAT synchro measure sample of 0 ° ~ 360 °;

The incident angle built in described synchronized samples is 46 ° and 54 °, as new synchro measure sample;

Described new synchro measure sample is pressed wind speed interval 1m/s, divides into groups in 20 °, wind direction interval;

Extract from each group, by the sample composition training sample extracted;

Determine physical geography module function;

The value of the unknowm coefficient in described physical geography module function is determined by the described training sample chosen.

2. modeling method for geophysical model as claimed in claim 1, it is characterized in that, the sample of described acquisition radiometer SSM/I and scatterometer QuikSCAT synchro measure comprises:

Reject the sample of rainfall at more than 2mm/h;

SSM/I wind speed and the QuikSCAT wind speed deviation of rejecting same time and same place are the sample of more than 4m/s;

Carry out time match, the extraction Measuring Time of SSM/I and the measuring intervals of TIME of QuikSCAT are less than or equal to the described sample of the synchro measure of 2 hours;

Spatial match is carried out to described sample, by inverse distance weighted interpolation algorithm, SSM/I wind speed is interpolated into QuikSCAT wind vector bin.

3. modeling method for geophysical model as claimed in claim 2, it is characterized in that, described extraction from each group comprises:

If the number of sample of a group is more than or equal to 10, then random from this group or by setting rule extraction 10 training samples;

If the number of the sample of a group is less than 10, then from this group, repeat extraction 10 training samples at random or by setting rule.

4. modeling method for geophysical model as claimed in claim 1, is characterized in that, describedly determines that physical geography module function comprises:

According to wind speed V, divide different wind speed sections, described wind speed section comprises: V < 16m/s, 16m/s≤V≤20m/s, V > 20m/s tri-sections;

For different wind speed sections, determine the submodel function of this wind speed section correspondence, described submodel function comprises: QSCAT_1, (QSCAT_1+NN_GMF)/2, NN_GMF, and described QSCAT_1 is scatterometer geophysical model;

The corresponding relation of described submodel function and described wind speed section is:

${\mathrm{\&sigma;}}_0=\left\{\begin{array}{cc}\mathrm{QSCAT}\_1& (V<16m/s)\\ (\mathrm{QSCAT}\_1+\mathrm{NN}\_\mathrm{GMF})/2& (16m/s\&le;V\&le;20m/s)\\ \mathrm{NN}\_\mathrm{GMF}& (V>20m/s)\end{array}\right..$

5. modeling method for geophysical model as claimed in claim 4, it is characterized in that, the defining method of described NN_GMF submodel function comprises:

Determine the topological structure of the artificial neural network that NN_GMF submodel function is corresponding;

Determine the mathematic(al) representation that the topological structure of described artificial neural network is corresponding.

6. modeling method for geophysical model as claimed in claim 5, it is characterized in that, the topological structure of the described artificial neural network that the described NN_GMF of determination submodel function is corresponding comprises:

Determine the number of described topological structure input layer, middle layer, output layer respectively;

Determine the neuronic number of described input layer, middle layer, output layer respectively;

Determine the neuronic function of input layer, middle layer, output layer respectively.

7. modeling method for geophysical model as claimed in claim 5, it is characterized in that, the topological structure of described artificial neural network comprises:

An input layer, two middle layers, an output layer;

Described input layer comprises 4 input layers, the corresponding input vector of each input layer, four described input vectors are respectively: the sine value of the angle χ of wind speed V, wave beam and wind vector, the cosine value of described χ, the sine value of incidence angle θ, and wherein, χ meets for beam positional angle, φ is wind direction of ocean surface;

Each described middle layer comprises 6 middle layer neurons, and described middle layer neuron is used for transport function, and described transport function gets Logsig function, and the formula of described Logsig function is:

Described output layer comprises 1 output layer neuron, and described output layer neuron is normalization backscattering coefficient σ 0, and output function is linear convergent rate function, and the formula of described output function is: f (x)=x+b;

And/or

The mathematic(al) representation that described artificial neural network topological structure is corresponding comprises:

${\mathrm{\&sigma;}}_0=\underset{i=\mathrm{1,6}}{\mathrm{\&Sigma;}}({\mathrm{LW}1}_i\&times;{\mathrm{VH}2}_i)+b2;$

$\mathrm{VH}2\left(i\right)=1/(1+\exp (-\underset{j=\mathrm{1,6}}{\mathrm{\&Sigma;}}({\mathrm{LW}}_{i,j}\&times;{\mathrm{VH}1}_j)+{b1}_i));$

$\mathrm{VH}1\left(i\right)=1/(1+\exp (-\underset{j=1,4}{\mathrm{\&Sigma;}}({\mathrm{IW}}_{i,j}\&times;{\mathrm{VI}}_j)+{b1}_i));$

VI ₁=sin (χ), VI ₂=cos (χ), VI ₃=sin (θ), VI ₄=(V-15)/20; Wherein, b, b1, b2, IW, LW are described unknowm coefficient.

8. modeling method for geophysical model as claimed in claim 5, is characterized in that, the value of the described unknowm coefficient determined in described physical geography module function by the described training sample chosen is comprised:

Described V, χ, θ, σ 0 is extracted from described training sample;

Adopt Quasi-Newton algorithm to carry out matching to described mathematic(al) representation, obtain the value of unknowm coefficient.

9. the modeling method for geophysical model as described in any one of claim 1-8, is characterized in that, comprises further after the described described training sample by choosing determines the value of the unknowm coefficient in described physical geography module function:

Verify by the value of other samples to described unknowm coefficient in described sample except described training sample.