CN104463252A - Foundation cloud classification method based on self-adaptive extreme learning machine - Google Patents

Abstract

The invention discloses a foundation cloud picture cloud classification method, and belongs to the technical field of image information processing and weather. The method comprises the following steps that (1) the textural feature, shape feature and color feature of a cloud picture are extracted to form a 21-dimensional feature vector; (2) normalization processing is carried out on each bit of the 21-dimensional feature vector; (3) a self-adaptive extreme learning machine model is built, and network training is carried out through a training sample; (4) the normalized 21-dimensional feature vector is adopted as the input of the self-adaptive extreme learning machine, and the varieties of clouds are adopted as output for cloud classification. The textural feature, shape feature and color feature of the cloud picture are comprehensively utilized, the self-adaptive extreme learning machine model based on k neighbors and the extreme learning machine is built, the foundation clouds are accurately classified, classification performance of the method is more accurate than that of an existing method, and the important application value is achieved.

Description

A kind of ground cloud classification method based on self-adaptation extreme learning machine

Technical field

The present invention relates to graphical analysis and meteorological technical field, particularly relate to a kind of cloud classification method of ground cloud atlas.

Background technology

Cloud is the most also the most general a kind of atmospheric condition as showing in earth atmosphere circle, receives the concern of people always.The earth on average there is the regional coverage of 1/3 to 1/2 cloud layer.Cloud is the important performer of synoptic process.As the saying goes " seeing that cloud knows weather ", certain weather phenomenon always links together with certain cloud.Before meteorological subject is started, the people of agricultural society are just because the close relation of weather and agricultural production starts the change paying close attention to cloud, knowwhy was not had to explain all meteors although lack meteorologic instrument at that time to the detection of atmospheric condition yet, but still summarize the relation between some cloud and Changes in weather according to people's rich experience, visible cloud is very important for the effect of research weather phenomenon and atmospheric condition.Cloud also plays a part very important to the change of weather, this effect be mainly manifested on the earth this-impact of Atmosphere System energy budget on.Determine that the element factor of ground vapour system capacity is the equilibrium relation of radiations heat energy, mainly comprise the balance between the heat effect from solar shortwave radiation and the cooling effect to the ground vapour long-wave radiation of cosmic space transmitting.Therefore judging the type of cloud, understand the distribution of cloud, is all vital for the accuracy of weather forecast, the validity of climate monitoring, the science of climate modelling and atmospheric exploration and atmospheric remote sensing.Object of the present invention is exactly set up self-adaptation limit machine learning model according to the Cloud-Picture Characteristics extracted, and distinguishes the type of the cloud mass in ground visible cloud image.

A complete cloud classification system comprises: the detection of cloud, the feature extraction of cloud and identification.At present mainly artificial visually examine is leaned on to the observation of cloud amount and cloud form, in device survey, mainly contain infrared radiation method and all band camera method can be seen by light.Infrared radiation method is carried out the information of Retrieval of Cloud by air infrared radiation in essence, is inevitably subject to the impact of atmospheric condition, needs to be undertaken demarcating by other means just can obtain good inversion result under different atmospheric conditions.In addition, infrared induction device apparatus expensive, is being difficult to large scale application at present.And utilizing visible ray all band camera method to be one " being gained as seen " method very intuitively to carry out the observation of cloud, equipment cost relative moderate is an important directions of at present ground cloud being carried out to automatic Observation.Therefore, in the present invention for be the full wave ground cloud atlas of visible ray.

The current relative maturity of detection technique of cloud, mainly utilizes empirical value or fixed threshold to realize compared with the red band ratio result of the indigo plant of cloud atlas.Feature extraction and the identification of cloud are contents most crucial in cloud classification system, also some relevant researchs have been carried out both at home and abroad at present, but the Individual features mostly only for cloud extracts, then with the domain knowledge of weather scientist for guidance is classified to extracted feature.The present invention utilizes the various features of cloud to analyze.The domestic and international Research on classifier to cloud classification mainly contains k nearest neighbor, Fuzzy strategy at present, support vector machine and neural network, and wherein the accuracy of identification of neural network classifier is generally considered higher than other sorter.Although neural-network classification method has unique advantage in numerous method, also there are some problems.Traditional neural network adopts the Gradient learning method (BP) of Error Feedback, have that pace of learning is comparatively slow, iterations too much, solve and be easy to be absorbed in the shortcomings such as local minimum.These shortcomings have had a strong impact on the application of neural network in cloud classification.In addition, in the systematic modeling of shortage, therefore easily there is the amplification of error in neural network in time inputting data and learning data difference is larger.

Summary of the invention

Goal of the invention: technical matters solved by the invention proposes a kind of ground cloud atlas cloud classification method based on self-adaptation extreme learning machine, make full use of overall textural characteristics, color characteristic, shape facility to describe cloud, adopt self-adaptation extreme learning machine as sorter, thus obtain than classic method classification performance more accurately.

Technical scheme, in order to realize above object, inventing the technical scheme taked is:

Based on a ground cloud classification method for self-adaptation extreme learning machine, it comprises following step:

Step 1, the textural characteristics extracting cloud atlas, shape facility and color characteristic, form the proper vector of a N dimension;

Step 2, each of proper vector of N dimension to be normalized;

Step 3, set up extreme learning machine sorter, utilize training sample to carry out network training;

Step 4, using normalized N dimensional feature vector as the input of self-adaptation extreme learning machine, utilize the mode of k nearest neighbor to carry out pre-service to input feature value, the N dimensional vector after process is input to extreme learning machine, obtains the final classification of cloud.

Preferably, the above-described ground cloud classification method based on self-adaptation extreme learning machine, in step 1, extract the textural characteristics of cloud atlas, textural characteristics comprises gray level co-occurrence matrixes, the Tamura textural characteristics of image; Extract the color characteristic of cloud atlas, color characteristic is that the color of cloud atlas is apart from (Stricker and Orengo proposition); Extract the shape facility of cloud atlas, shape facility is not displacement feature (moment invariants).

Preferably, the above-described ground cloud classification method based on self-adaptation extreme learning machine, proper vector normalized described in step 2, concrete grammar for: the every one dimension in the proper vector of the N that step 2 extracted dimension is normalized respectively, and all characteristic quantities are mapped as the number between 0 ~ 1.

Preferably, the above-described ground cloud classification method based on self-adaptation extreme learning machine, extreme learning machine sorter is set up described in step 3, be specially method: set up neural network structure, the link weights of input layer and hidden layer and biased Stochastic choice, obtain exporting weights by sample training, namely the sorter of cloud atlas is trained.

Preferably, the above-described ground cloud classification method based on self-adaptation extreme learning machine, the final classification of described step 4 medium cloud, concrete grammar is: utilize k near neighbor method to find the multiple neighbours of input feature value in training sample, these neighbours are integrated, obtain a new input vector, using the input of this vector as extreme learning machine, export the classification being cloud.

The method that the present invention proposes is using color characteristic, shape facility, Texture Feature Fusion together as proper vector, and adopts self-adaptation extreme learning machine as sorter, and test findings shows that the method classification accuracy is high.

The present invention adopts the e-learning mode of self-adaptation extreme learning machine to learn neural network.This method utilizes the mode of k nearest neighbor to data prediction, can reduce the distance of input data and learning data.ELM learning method is while ensureing that network has good Generalization Capability, can greatly improve the pace of learning of network, and can avoid based on problems many in Gradient Descent learning algorithm, the determination etc. of as many in Local Minimum, iterations, performance index and learning rate.Experimental result shows, the present invention propose method than traditional threshold method and dynamic thresholding method precision higher.

Beneficial effect: compared to the prior art the ground cloud classification method based on self-adaptation extreme learning machine provided by the invention has the following advantages:

(1) the present invention utilizes three kinds of features: color, shape, texture are as feature;

(2) set up adaptive model, utilize k nearest neighbor to process input amendment;

(3) based on the multidimensional characteristic of typical cloud form, set up the cloud atlas cloud form sorting algorithm of extreme learning machine classification;

(4) under the same conditions, the inventive method can obtain than traditional based on k nearest neighbor, BP neural network, independent extreme learning machine, the cloud atlas sorting technique classification performance more accurately of self-adaptive BP neural networks and SVM.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of a kind of ground cloud atlas cloud classification method based on self-adaptation extreme learning machine of the present invention.

The distribution plan of the various feature of Fig. 2 is extreme learning machine network diagram.

Fig. 3 is the inventive method and the comparing of other method conventional.

Embodiment

Below in conjunction with accompanying drawing, the present invention is illustrated.Described enforcement example is only for illustrative purposes, instead of limitation of the scope of the invention.

The present invention proposes a kind of ground cloud atlas cloud classification method based on self-adaptation extreme learning machine, object is textural characteristics, color characteristic and shape facility by making full use of cloud atlas, adopts self-adaptation extreme learning machine to divide the identification any ground cloud atlas being carried out to the classification of cloud.Setting varieties of clouds type in the invention process is 4 kinds of typical cloud forms, comprises cumuliform cloud, cirrus, stratiform clouds and clear sky.

Fig. 1 is process flow diagram of the present invention.With reference to Fig. 1, performing step of the present invention is as follows:

Step one, the textural characteristics extracting cloud atlas, shape facility and color characteristic, form the proper vector of one 21 dimension.

(1.1) extract gray level co-occurrence matrixes texture P (i, j, δ, φ) represent from gray scale i, distance be δ=(Dx, Dy) point on gray scale be the probability of j.

P(i,j,δ,θ)＝{(x,y)|z(x,y)＝i,z(x+Dx,y+Dy)＝j；x,y＝0,1,2...,N-1}

Wherein, the point of Dx to be gray level be j and gray level are the horizontal ordinate distance of the point of i, and Dy is gray level be j point and gray level is p _ijthe ordinate distance of point, different δ represents different Distance geometry directions.Extracting 6 correlated characteristics, is second moment respectively, contrast, correlativity, entropy, unfavourable balance distance and inertia distance.

(1.2) Tamura textural characteristics is extracted

Tamura texture of the present invention gets these three characteristic quantities of roughness, contrast and direction degree.

(1.3) color characteristic is extracted

Color adopts the first moment of color, second moment and third moment to describe color distribution.

${\mathrm{\μ}}_i=\frac{1}{N}{\mathrm{\Σ}}_{j=1}^N{p}_{\mathrm{ij}}$

${\mathrm{\σ}}_i={\left(\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(p_{\mathrm{ij}}-{\mathrm{\μ}}_i)}^2\right)}^{\frac{1}{2}}$

$s_i={\left(\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(p_{\mathrm{ij}}-{\mathrm{\μ}}_i)}^3\right)}^{\frac{1}{3}}$

Wherein, p _iji-th color component of a jth pixel in image.

(1.4) shape facility is extracted

Shape facility extracts morphological feature based on area invariant moment.Can be expressed as bianry image R, p+q center square:

${\mathrm{\μ}}_{p,q}=\underset{(x,y)\∈R}{\mathrm{\Σ}}{(x-x_c)}^p{(y-y_c)}^q$

(x _c, y _c) be the center of R, (x, y) belongs to R, and M.K.Hu is based on u _p,qseven features are proposed:

φ ₁＝μ _2,0+μ _0,2

${\mathrm{\φ}}_2={({\mathrm{\μ}}_{\mathrm{2,0}}-{\mathrm{\μ}}_{\mathrm{0,2}})}^2+4{\mathrm{\μ}}_{\mathrm{1,1}}^2$

${\mathrm{\φ}}_3={({\mathrm{\μ}}_{\mathrm{3,0}}-{\mathrm{\μ}}_{\mathrm{1,2}})}_8^2+{({\mathrm{\μ}}_{\mathrm{0,3}}+3{\mathrm{\μ}}_{\mathrm{2,1}})}^2$

φ ₄＝(μ _3,0+μ _1,2) ²+(μ _0,3+μ _2,1) ²

${\mathrm{\φ}}_5=[({\mathrm{\μ}}_{\mathrm{3,0}}-3{\mathrm{\μ}}_{\mathrm{1,2}})({\mathrm{\μ}}_{\mathrm{3,0}}+{\mathrm{\μ}}_{\mathrm{1,2}})+({\mathrm{\μ}}_{\mathrm{0,3}}-3{\mathrm{\μ}}_{\mathrm{2,1}})({\mathrm{\μ}}_{\mathrm{0,3}}+{\mathrm{\μ}}_{\mathrm{2,1}})]\×[{({\mathrm{\μ}}_{\mathrm{3,0}}+3{\mathrm{\μ}}_{\mathrm{1,2}})}^2-{({\mathrm{\μ}}_{\mathrm{0,3}}+{\mathrm{\μ}}_{\mathrm{2,1}})}^2]$

φ ₆＝(μ _2,0-μ _0,2)[(μ _3,0+μ _1,2) ²-(μ _0,3+μ _2,1)]+4μ _1,1(μ _3,0+μ _1,2)](μ _0,3+μ _2,1)

${\mathrm{\φ}}_7=[(3{\mathrm{\μ}}_{\mathrm{2,1}}-{\mathrm{\μ}}_{\mathrm{0,3}})({\mathrm{\μ}}_{\mathrm{3,0}}+{\mathrm{\μ}}_{\mathrm{2,1}})+({\mathrm{\μ}}_{\mathrm{3,0}}-3{\mathrm{\μ}}_{\mathrm{2,1}})({\mathrm{\μ}}_{\mathrm{0,3}}+{\mathrm{\μ}}_{\mathrm{2,1}})]\×[{({\mathrm{\μ}}_{\mathrm{3,0}}+{\mathrm{\μ}}_{\mathrm{1,2}})}^2-{3({\mathrm{\μ}}_{\mathrm{0,3}}+{\mathrm{\μ}}_{\mathrm{2,1}})}^2]$

Each of the proper vector of step 2,21 dimensions is normalized.

Utilize even normalization 21 dimensional feature vector of standard.

Step 3, set up extreme learning machine model, utilize training sample to carry out network training.

Network structure as shown in Figure 2.For N number of different sample (x _i, y _i), i=1,2 ..., N, wherein,

X _i=[x _i1, x _i2..., x _in] ∈ R ⁿ, y _i=[y _i1, y _i2..., y _im] ∈ R ^m, x _ifor input amendment, y _ifor output sample.If the hidden node number of network is activation function is that g (x) model can be expressed as:

$\underset{i=1}{\overset{\tilde{N}}{\mathrm{\Σ}}}{\mathrm{\β}}_i{g}_i(w_i{x}_j+b_i)=o_j,j=\mathrm{1,2},...,N$

Wherein, w _i=[w _i1, w _i2..., w _in] ^tfor connecting the input weights of i-th hidden layer node; B is the bias of i concealed nodes; β _i=[β _i1, β _i2..., β _im] ^tfor connecting the output weights of i-th concealed nodes; o _jfor the output valve of a jth sample.Suppose that SLFN error freely can approach N number of sample, i.e. Σ || o _j-y _j||=0, so just there is β _i, w _i, b _i, make

$\stackrel{\tilde{N}}{\mathrm{\Σ}}{\mathrm{\β}}_i{g}_i(w_i{x}_j+b_i)=y_j,j=\mathrm{1,2},...,N$

$\mathrm{\β}={\left[\begin{array}{c}{\mathrm{\β}}_1^T\\ \·\\ \·\\ \·\\ {\mathrm{\β}}_{\tilde{N}}^T\end{array}\right]}_{\tilde{N}\×m},Y={\left[\begin{array}{c}{y}_1^T\\ \·\\ \·\\ \·\\ y_N^T\end{array}\right]}_{N\×m}$

The matrix form of above-mentioned N number of equation can be written as:

Hβ＝Y

The training objective of SLFN seeks optimum network weight to make following formula minimum:

min E(w)＝min||Hβ-Y||，

Input weights and hidden layer node bias can be given at random when training and starting, now matrix H is a constant matrices, formula H β=Y is simplified as one group of linear equation, exports weights and obtains by the least square solution solving this system of linear equations minimum norm, that is: β=H ⁺y.

So far, we have learnt extreme learning machine cloud atlas sorter.

Step 4, using normalized 21 dimensional feature vectors as the input of self-adaptation extreme learning machine, utilize the mode of k nearest neighbor to carry out pre-service to input feature value, 21 dimensional vectors after process are input to extreme learning machine, obtain the final classification of cloud.

Owing to easily there is the amplification of error time ELM input data and learning data differ larger, therefore the method for k nearest neighbor and ELM is adopted to combine (self-adaptation extreme learning machine) herein, reduce the error that data variation acutely brings infinitely to expand, concrete step is:

1, for one group of test data, Q=[q ¹, q ², q ³..., q ⁿ]

By it and training data $X_i=[x_i^1,x_i^2,x_i^3,...,x_i^n]$ (i=1,2,3 ... m, m are the number of training sample) made comparisons by Euclidean distance, find out k the neighbour X of Q _q1, X _q2..., X _qk.

2, according to k neighbour X _q1, X _q2..., X _qkthe input of initialization extreme learning machine

$\mathrm{input}=[{\mathrm{\Σ}}_{i=1}^k{x}_{i1}/k,{\mathrm{\Σ}}_{i=1}^k{x}_{i2}/k,...{\mathrm{\Σ}}_{i=1}^k{x}_{\mathrm{in}}/k].$

3, initialized sample is input to extreme learning machine, obtains the output of extreme learning machine, obtain 4 classification values, choose maximal value as final classification results.Repeat step 4-5 time, get number of times of classifying maximum as net result.

Confirmatory experiment:

In order to verify the effect of the inventive method, respectively by the inventive method and conventional k nearest neighbor, BP neural network, independent extreme learning machine, self-adaptive BP neural networks and SVM method compare.As Fig. 3 experimental result shows, different sorting techniques is adopted to contrast, result shows, adopt the ground cloud atlas cloud classification classification result precision of the present invention preferably based on self-adaptation extreme learning machine the highest, therefore, sorting technique provided by the invention, performance is better than art methods on the whole, achieves extraordinary technique effect.

The above is only some embodiments of the present invention; it should be pointed out that for those skilled in the art, under the premise without departing from the principles of the invention; can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.

Claims (5)

1., based on a ground cloud classification method for self-adaptation extreme learning machine, it is characterized in that, it comprises following step:

Step 1, the textural characteristics extracting cloud atlas, shape facility and color characteristic, form the proper vector of a N dimension;

Step 2, each of proper vector of N dimension to be normalized;

Step 3, set up extreme learning machine sorter, utilize training sample to carry out network training;

Step 4, using normalized N dimensional feature vector as the input of self-adaptation extreme learning machine, utilize the mode of k nearest neighbor to carry out pre-service to input feature value, the N dimensional vector after process is input to extreme learning machine, obtains the final classification of cloud.

2. the ground cloud classification method based on self-adaptation extreme learning machine according to claim 1, is characterized in that, in step 1, extract the textural characteristics of cloud atlas, textural characteristics comprises gray level co-occurrence matrixes, the Tamura textural characteristics of image; Extract the color characteristic of cloud atlas, color characteristic is the color distance of cloud atlas; Extract the shape facility of cloud atlas, shape facility is not displacement feature.

3. the ground cloud classification method based on self-adaptation extreme learning machine according to claim 1, it is characterized in that, proper vector normalized described in step 2, concrete grammar for: the every one dimension in the proper vector of the N that step 2 extracted dimension is normalized respectively, and all characteristic quantities are mapped as the number between 0 ~ 1.

4. the ground cloud classification method based on self-adaptation extreme learning machine according to claim 1, it is characterized in that, extreme learning machine sorter is set up described in step 3, be specially method: set up neural network structure, the link weights of input layer and hidden layer and biased Stochastic choice, obtain exporting weights by sample training, namely the sorter of cloud atlas is trained.

5. the ground cloud classification method based on self-adaptation extreme learning machine according to claim 1, it is characterized in that, the final classification of described step 4 medium cloud, concrete grammar is: utilize k near neighbor method to find the multiple neighbours of input feature value in training sample, these neighbours are integrated, obtain a new input vector, using the input of this vector as extreme learning machine, export the classification being cloud.