CN104166857A - Oil spilling hyperspectral image classification method based on wavelet transformation low frequency spectrum information - Google Patents

Abstract

The invention provides an oil spilling hyperspectral image classification method based on wavelet transformation low frequency spectrum information. The method includes the steps that hyperspectral data processing and spectral characteristic preliminary analysis of three ground features are carried out; a one-dimensional signal is generated by means of spectral characteristic values of the three ground features, per-pixel five-layer wavelet decomposition is conducted in ROIs on the basis of the ROIs of the three ground features and db4 wavelet bases, and a low frequency wavelet coefficient curve of the three ground features is drawn; wave bands with high separability are selected; per-pixel five-layer wavelet decomposition based on the db4 wavelet bases is conducted on a whole image, fifth-layer low frequency wavelet coefficients of the wave bands with the high separability are extracted, and a low frequency wavelet coefficient image with high separability is generated; oil films of different thicknesses are classified according to the maximum likelihood method. By the adoption of the oil spilling hyperspectral image classification method, spectral information can be fully utilized, spectral information of hyperspectral images is sufficiently mined, and consequently both image classification efficiency and image classification accuracy are improved.

Description

Oil spilling hyperspectral image classification method based on wavelet transformation low frequency spectrum information

Technical field

The present invention relates to spectrum picture process field, relate in particular to a kind of oil spilling hyperspectral image classification method based on wavelet transformation low frequency spectrum information.

Background technology

Along with the development of earth observation technology, the spectral resolution of image, spatial resolution and temporal resolution have had very big raising, and wherein high spectrum image has obtained increasing application.High spectral resolution remote sensing refers to very narrow and continuous spectrum channel and atural object is carried out to the technology of imaging.At visible ray, near infrared range, its spectral resolution can reach nanoscale, and generally, the spectral resolution of high-spectrum remote-sensing is less than 10nm conventionally, compares with multispectral remote sensing, and its wave band is nearly dozens or even hundreds of.The data that obtained by high-spectrum remote-sensing are high-spectral data.High-spectral data can represent with data cube, and it is comprised of 3 parts, i.e. spatial image dimension, spectrum peacekeeping feature space dimension.The geographical location information of spatial image dimension essential record atural object, be applicable to general information service, the spectral signature information of spectrum dimension essential record atural object, the degree of depth that is applicable to spectral information is excavated, and feature space dimension can be excavated feature and the behavior that atural object distributes in the dimensional feature space of high-spectral data formation for the degree of depth.These features of high-spectral data are that Images Classification has brought very big advantage.First, continuous spectrum can better mate Land Surface Temperatures, finds the sensitive band of wish classification atural object interval, reduces the impact of the different spectrum of jljl and same object different images; Secondly, versatile and flexible in sorting technique, both can apply traditional mode identification method, can adopt again the Spectral matching method of mating with ground Object Spectra DataBase, can, when reducing calculated amount, improve nicety of grading.These advantages of high-spectral data can be oil spilling classification great potential are provided, spectrum characteristic for different-thickness oil film carries out band selection, application of spectral matching method can improve nicety of grading to a certain extent, if but directly use original high-spectral data signal, existence due to noise and the different spectrum of jljl and same object different images phenomenon, very easily cause different atural object classifications to obscure, degradation problem under nicety of grading, this is the difficult point of current high spectrum spilled oil monitoring just also.

Existing oil spilling classification schemes directly carries out classical way classification to image mostly, or the method for carrying out utilizing after simple denoising band selection is selected suitable wave band and is classified, nicety of grading is not high, main cause is directly to utilize original spectrum information, the spectral information between different-thickness oil film is not carried out to the deep information excavation, between class, separability is not further enhanced.

Existing method is described below:

Principal component analysis (PCA), is mainly to utilize dimensionality reduction thought, and many indexs are converted into a few overall target, consider a plurality of influence factors, as noise, correlativity, quantity of information etc., get several wave bands that integrated information is maximum as the wave band that carries out subsequent treatment, reach the object of dimensionality reduction.The method can be selected optimum wave band on the whole, but can not guarantee that the optimum wave band of selecting plays a key effect to subsequent treatment.

Minimal noise partition method (MNF), the separated transformation tool of minimal noise is for judging the dimension (being wave band number) of view data inherence, the noise in mask data, reduces with the computation requirement amount in aftertreatment.MNF is twice stacked principal component transform in essence.The advantage of the method is can try one's best noise decrease and the impact of wave band correlativity on subsequent treatment, reduces calculated amount.But the method, when removing noise, also likely remove the information that classification is worked, and the wave band number that the method obtains is fewer, is unfavorable for representing the information of whole image.

In a word, the shortcoming of existing oil spilling Hyperspectral data classification method: fail to make full use of spectral information, the spectral information of high spectrum image is excavated not, caused Images Classification efficiency and nicety of grading to decline.

Summary of the invention

The present invention is directed to the shortcoming of prior art, propose the Hyperspectral data classification method based on wavelet transformation low frequency spectrum information.This method has been given full play to the approximation coefficient that represents low-frequency component in wavelet transformation to the portraying of different wave length place atural object, and the image that utilizes low frequency spectrum signal reconstruct is when effectively removing original image noise, has improved the precision of oil identification.The advantage of wavelet transformation and the characteristic of high spectrum self have been given full play to.

An oil spilling hyperspectral image classification method based on wavelet transformation low frequency spectrum information, comprises the following steps:

Step 1: draw the curve of spectrum of oil spilling high spectrum image pixel, this curve of spectrum is configured to a continuous one-dimensional signal;

Step 2: this one-dimensional signal is carried out to wavelet decomposition, extract low frequency wavelet coefficient, obtain low frequency spectrum curve;

Step 3: select the obvious low frequency spectral coverage of classification separability to generate low-frequency image between the low frequency spectrum curve of different atural objects;

Step 4: select atural object ROI in the low-frequency image generating, to the atural object ROI the selecting classification that exercises supervision.

Further, the oil spilling hyperspectral image classification method based on wavelet transformation low frequency spectrum information as above, described step 1 comprises:

(a), read oil spilling high-spectral data, the SPECTRAL DIVERSITY reflecting according to different-thickness oil film, by the oil film atural object that is divided three classes, this three classes atural object is heavy oil film, thin oil film, seawater;

(b), select the ROI of three class atural objects, and draw spectral signature curve according to its spectrum characteristic parameter;

(c), select to distinguish the wave band interval of three class atural objects.

Further, the oil spilling hyperspectral image classification method based on wavelet transformation low frequency spectrum information as above, described step 2 comprises:

(a), respectively the pixel in the ROI of three class atural objects is carried out to 5 layers of decomposition based on db4 wavelet basis, and extract the low frequency wavelet coefficient of the 5th layer;

(b), draw the curve map of described low frequency wavelet coefficient, and choose the obvious wave band of separability according to the curve map of described wavelet coefficient.

Further, the oil spilling hyperspectral image classification method based on wavelet transformation low frequency spectrum information as above, described step 3 comprises:

Entire image is carried out to the 5 layers of decomposition based on db4 wavelet basis by pixel, extract the 5th layer of low frequency wavelet coefficient of the obvious wave band of separability, generate the obvious low frequency wavelet coefficient image of separability;

Further, the oil spilling hyperspectral image classification method based on wavelet transformation low frequency spectrum information as above, described step 4 comprises:

In the 5th layer of low frequency wavelet coefficient image, select respectively the ROI of three class atural objects, utilize the classification that exercises supervision of envi maximum likelihood method.

Beneficial effect:

1, by original spectrum data are carried out to wavelet transform process, further excavated " spectrum information " of high spectrum image, " spectrum information " of oil spilling high spectrum image is fully used.

2, by image is carried out to wavelet transformation, and generate low frequency wavelet coefficient image, can obviously reduce picture noise, make between class separability larger, effectively improved the nicety of grading of the high-spectral data of oil spilling different-thickness oil film.

Accompanying drawing explanation

Fig. 1 is the oil spilling hyperspectral image classification method process flow diagram that the present invention is based on wavelet transformation low frequency spectrum information;

Fig. 2 a is oil spilling spectrum original image;

Fig. 2 b is the classification results figure of conventional images sorting technique to oil spilling spectrum picture;

Fig. 2 c is that image classification method of the present invention is to oil spilling spectrum picture classification results figure.

Embodiment

For making the object, technical solutions and advantages of the present invention clearer, below technical scheme in the present invention be clearly and completely described, obviously, described embodiment is the present invention's part embodiment, rather than whole embodiment.Embodiment based in the present invention, those of ordinary skills, not making the every other embodiment obtaining under creative work prerequisite, belong to the scope of protection of the invention.

The present invention is directed to the SPECTRAL DIVERSITY of failing to make full use of inhomogeneity atural object in high spectrum image in current oil spilling classification hyperspectral imagery, picture depth information is failed to the deficiency that makes full use of and excavate, and we propose a kind of oil spilling hyperspectral image classification method based on wavelet transformation low frequency spectrum information.The object of this method is to give full play to the portray ability of wavelet transformation to localized variation, utilize low frequency wavelet spectrum information to regenerate image, it is little that this image has the interior feature difference of class, between class, the otherness of feature is large, thereby strengthen the differentiation of different classes of, reach the denoising of oil spilling high spectrum image and quick and precisely classification.

Method And Principle is introduced:

Below theory of wavelet transformation is introduced

(1) One Dimension Continuous Wavelet Transform

Small echo, is exactly little waveform, so-called little, is exactly that it has Decay Rate, is to be present in one compared with the ripple of zonule.If f (t) is a function that can survey, square-integrable, there is finite energy, f (t) ∈ L ²(R), L ²(R) be the vector space of f (t), R is set of real numbers.Continuous wavelet transform is defined as signal f (t) and wavelet basis function ψ _a.b(t) inner product:

$W_f(a,b)=<f,{\mathrm{\&psi;}}_{a,b}>={\left|a\right|}^{-1/2}{\&Integral;}_{R}f\left(t\right)\stackrel{\&OverBar;}{\mathrm{\&psi;}\left(\frac{t-b}{a}\right)}\mathrm{dt}$

Wherein, a > 0, f (t) ∈ L ²(R)

ψ in Wavelet transform type _a,b(t) be complex conjugate function.On mathematics, inner product characterizes the degree of two functions " similar ", therefore from above formula explanation f (t) and ψ _a.b(t) similar degree, integral kernel in (1) formula

${\mathrm{\&psi;}}_{a,b}\left(t\right)=\frac{1}{\sqrt{\left|a\right|}}\mathrm{\&psi;}\left(\frac{t-b}{a}\right)---\left(2\right)$

Be the window function of introducing, or claim small echo or basis function.Wherein variable a is contraction-expansion factor, and b is shift factor.It is actual is one group of series of functions of being stretched and being formed with translation by wavelet mother function ψ (t).Due to ψ _a,b(t) two parameter a in, b can continuously change, therefore claim that (1) formula is continuous wavelet transform.

The main process of continuous wavelet transform is:

A. the beginning of wavelet mother function ψ (t) and original signal f (t) is compared;

B. calculate wavelet coefficient W _f(a, b), this coefficient represents the degree of approximation of this part signal and small echo, coefficient W _fthe higher expression signal of value of (a, b) is more similar to small echo, so wavelet coefficient W _f(a, b) can reflect the degree of correlation of this waveform;

C. small echo is moved right, distance is k, and the wavelet function obtaining is ψ (t-k), then repeating step 1 and 2.The small echo k that moves right, obtain small echo ψ (t-2k), repeating step 1 and 2 again.By above-mentioned steps, go on, until signal f (t) finishes always

D. expand small echo ψ (t), for example, expand one times, the wavelet function obtaining is ψ (1/2)

E. repeating step a-d.

(2) one-dimensinal discrete small wave transformation

For continuous wavelet, yardstick a, time t are continuous with the side-play amount τ relevant with the time.If utilize computing machine to calculate, just must carry out discretize processing to them, obtain wavelet transform.In order to reduce the redundance of wavelet conversion coefficient, we are by wavelet basis function a, b be limited to value on some discrete points.

1. the discretize of yardstick.Current way is that yardstick is carried out to exponential level discretize at present.Even a gets a ₀> 0, j ∈ Z, and corresponding wavelet function is: j=0,1,2

2. displacement discretize.

Conventionally τ is carried out to even discrete value, to cover whole time shaft, τ meets Nyquist sampling thheorem.When a=2j, along the response sample interval of τ axle, be 2j τ ₀, at a ₀in=2 situations, j increases by 1, and yardstick a doubles, and corresponding frequency reduces half.Now sampling rate can reduce half and not cause the loss of the information that causes.So under yardstick j, due to width be ψ (t) times, so sampling interval can expand and can not cause the loss of information.ψ _{a, τ}(t) can be write as:

$a_0^{-\frac{j}{2}}\mathrm{\&psi;}\left[a_0^{-j}(t-k{a}_0^j{\mathrm{\&tau;}}_0)\right]=a_0^{-\frac{j}{2}}\mathrm{\&psi;}[a_0^{-j}t-k{\mathrm{\&tau;}}_0]---\left(3\right)$

Discrete wavelet is defined as:

${\mathrm{WT}}_f(a_0^j,k{\mathrm{\&tau;}}_0)=f\left(t\right){\mathrm{\&psi;}}_{a_0^j,k{\mathrm{\&tau;}}_0}^{*}\left(t\right)\mathrm{dt},j=\mathrm{0,1,2},...,k=\&Element;Z---\left(4\right)$

The effective ways of carrying out wavelet transform are to use wave filter, the method is developed in 1988 by Mallat, the main thought of carrying out wavelet transform with wave filter is to produce two signal A and D by two mutual incoherent wave filters, the approximate value of A representation signal wherein, the detail value of D representation signal.

(3) detailed process of wavelet transformation (to be decomposed into example based on 3 layers of db1 wavelet basis):

Suppose to have a width resolution to only have the one dimension image of 4 pixels, corresponding pixel value or the coefficient that is called picture position are respectively the wavelet conversion coefficient that [9 73 5] calculate it.Calculation procedure is as follows:

Step 1: (averaging) averages.Calculate the right mean value of neighbor, obtain the lower new images of a width resolution, its number of pixels has become 2, and the resolution of new image is original 1/2, and corresponding pixel value is: [8 4]

Step 2: ask difference (differencing).Clearly, while representing this width image by 2 pixels, the partial loss of the information of image.For the original image being comprised of 4 pixels can be gone out from the Image Reconstruction being comprised of 2 pixels, just need to store the detail coefficients (detail coefficient) of some images, to give the information of loss for change when reconstruct.Method is that first right pixel value of pixel is deducted to the mean value that this pixel is right, or uses difference that this pixel is right divided by 2.In this example, first detail coefficients is (9-8)=1, because the mean value calculating is 8, it is large 1 than 9 little 1 and than 7, stores the first two pixel value that this detail coefficients just can be recovered original image.Make to use the same method, second detail coefficients is (3-4)=-1, stores this detail coefficients and just can recover rear 2 pixel values.Therefore, original image just can represent by two mean values and two detail coefficients below, [8 4 1-1];

Step 3: repeat the 1st, 2 steps, further resolve into decomposing by the first step image obtaining image and the detail coefficients that resolution is lower.In this example, decompose finally, just by mean value 6 and three detail coefficients 2,1 of a pixel, represent entire image, [6 2 1-1] with-1.

The embodiment that this patent is concrete: the generation of low frequency spectrum and low frequency spectrum image

A pixel of take in image is example, and its curve of spectrum can form a continuous one-dimensional signal, and this signal is carried out to wavelet decomposition, extracts low frequency wavelet coefficient, can obtain low frequency spectrum curve.Between the low frequency spectrum curve of different atural objects, select the low frequency spectral coverage that classification separability is large to generate low-frequency image.Flow process please refer to Fig. 1, specific as follows:

(1) hyperspectral data processing and three kinds of spectral characteristic of ground initial analyses.First read oil spilling high-spectral data, by visual interpretation, the color distortion reflecting according to different-thickness oil film, is divided into heavy oil film, thin oil film by oil film, and another kind of is seawater.Visual interpretation is selected the ROI of three class atural objects, according to its spectrum characteristic parameter, draws spectral signature curve, between selective light spectral curve between the study area of the obvious wave band of difference interval as next step;

(2) wavelet decomposition.Utilize the spectrum characteristic parameter of three kinds of atural objects to carry out the structure of one-dimensional signal, and carry out 5 layers of decomposition (decomposition step is as follows) based on db4 wavelet basis.Respectively the pixel in seawater, heavy oil film, thin oil film ROI (region of interest) is carried out to wavelet decomposition, and extract the low frequency wavelet coefficient of the 5th layer, draw low frequency wavelet charts for finned heat.According to the difference of three class atural object low frequency wavelet coefficient curve, select the obvious spectral coverage of difference interval.

The experimental procedure of 5 layers of wavelet decomposition based on db4 wavelet function:

The first step: the spectrum matrix that obtains original image.Program is as follows:

The spectrum matrix of note: gp for obtaining

Second step: realize 5 layers of wavelet decomposition based on db4 wavelet function, obtain the 5th layer of low frequency wavelet coefficient.Program is as follows:

Note: wherein the expression formula of function db is as follows: [C, L]=wavedec (X, Y, wname).X represents the signal that need to decompose, and which kind of small echo ' wname ' represents to adopt decompose, and Y represents the number of plies of decomposing, and C represents high frequency wavelet coefficient, and L represents low frequency wavelet coefficient, is to adopt db4 small echo to carry out 5 layers of decomposition in test.

(3) generation of " spectrum "-> " figure ".Entire image is carried out to the 5 layers of decomposition (decomposition step as mentioned above) based on db4 wavelet basis by pixel, after decomposing for 5 times, obtain the low frequency wavelet coefficient of original signal, extract the 5th layer of low frequency wavelet coefficient d i of the wave band that separability is strong, each pixel is decomposed to the di producing to be rearranged according to the position of original image, be about to the low frequency wavelet coefficient d i that former pixel obtains through wavelet decomposition and be put into this pixel correspondence position, obtain the matrix D identical with original image size, D is low frequency wavelet coefficient image.The low frequency wavelet coefficient image that small echo obtains after processing can realize effective noise reduction of image, increases the degree of isolation between heavy oil film, thin oil film and seawater three class atural objects, for the high-precision classification of three class atural object provides precondition simultaneously.

(4) classification.On low frequency wavelet coefficient image D, select respectively the ROI of seawater, heavy oil film, thin oil film, utilize the maximum likelihood method classification (principle of classification of maximum likelihood method as described later) that exercises supervision, original spectrum wave band without wavelet transformation is carried out to maximum likelihood method classification, be that the nicety of grading of two kinds of classification results is relatively prepared simultaneously.

The low frequency wavelet coefficient image that small echo of the present invention obtains after processing can realize effective noise reduction of image, increases the degree of isolation between classification, for high-precision terrain classification provides precondition simultaneously.

To the image of crossing through wavelet transform process with without the original spectrum image of wavelet transformation, carry out recently verifying that small echo of the present invention processes the advantage of the image that the obtains nicety of grading when classify processing below, pass through the image of wavelet transformation, noise is low, atural object is can degree of classification high, good classification effect.

Precision evaluation

Nicety of grading evaluation is the precision of evaluating Images Classification, by nicety of grading evaluation, can grasp classifying quality, and then the validity of judgement sorting technique, the detailed process of nicety of grading evaluation is: utilize the ROI accuracy assessment method of remote sensing image processing software ENVI software to carry out the nicety of grading comparison of two kinds of classification experiments, the method for described accuracy assessment is as follows:

Being described below of maximum likelihood classification algorithm and Images Classification precision evaluation process:

Maximum likelihood classification algorithm, for classification, is mainly the distribution of satellite remote sensing multi-wavelength data to be used as to multidimensional normal distribution construct identification and classification function.Its basic thought is: the data of all kinds of known pixels form certain point group in plane or space; Every one-dimensional data of each class forms a normal distribution on the number axis of oneself, such multidimensional data just forms such multidimensional normal distribution, there is all kinds of multiple dimensional distribution models, for the data vector of any one unknown classification, all can ask conversely it to belong to all kinds of probability; The size of these probability relatively, sees that to belong to which kind of probability large, just this data vector or this pixel is classified as to such.

Images Classification precision is the ultimate criterion of classification of assessment method, and confusion matrix is the most general and obtains the extensively accuracy assessment method of approval.Conventionally evaluation classifying quality Main Basis has overall accuracy and Kappa coefficient.Overall classification accuracy refers to that the pixel summation of correctly being classified is divided by total pixel number.The pixel number of correctly being classified distributes along the diagonal line of confusion matrix, and total pixel number equals the pixel sum of all true reference sources.The higher explanation classifying quality of overall accuracy is better.

Kappa coefficient analysis is the multivariate statistical method of classification of assessment precision, the estimation of Kappa coefficient is called to KHAT statistics, it be by the pixel sum in the true classification in all earth's surfaces is multiplied by confusion matrix cornerwise and, deduct again the true pixel sum in earth's surface in a certain class with such in be classified the long-pending result to all categories summation of pixel sum, finally divided by the difference of two squares of total pixel number deduct in a certain class the true pixel sum in earth's surface with such in be classified the long-pending of pixel sum the result of all categories summation obtained.Kappa coefficient is larger, and presentation class effect is better.Computing formula is as shown in (6).

$K=\frac{N{\mathrm{\&Sigma;}}_i^r{x}_{\mathrm{ii}}-\mathrm{\&Sigma;}\left(x_{i+}{x}_{+i}\right)}{N^2-\mathrm{\&Sigma;}\left(x_{i+}{x}_{+i}\right)}---\left(6\right)$

In formula, K is Kappa coefficient, and r is the line number of error matrix, x _iithe value on the capable i row of i (principal diagonal), x _i+and x _{+ i}be respectively i capable and with i row and, N is total sample.

Table 1 spectrum picture and the contrast of low frequency spectrum frame nicety of grading

?	Original spectrum picture classification	The classification of small echo low frequency spectrum frame
			Nicety of grading	88.6％	94.7％
KAPPA coefficient	0.83	0.92

As can be seen from Table 1, the Images Classification result regenerating through extraction small echo low-frequency information is significantly better than the classification results of original spectrum image, on overall accuracy, this method precision of images has improved 6.1% compared with original spectrum image, KAPPA coefficient has improved 0.9, this explanation no matter is in integral body or on specific to each class atural object, the classification results of small echo low frequency spectrum frame will be significantly better than the classification results of original spectrum image.

By Fig. 2 a-2c, can be found out, under the identical prerequisite of sorter, Hyperspectral data classification effect based on wavelet transformation low frequency spectrum information (Fig. 2 c) is better than the classifying quality (Fig. 2 b) of the original image of prior art image classification method, wherein by Fig. 2 b, can find out because the band noise of image is obvious, the classification results of this image has directly reflected the impact that is subject to noise, and Images Classification result based on low frequency spectrum information is as Fig. 2 c, because noise is removed substantially, so classifying quality is more excellent.

Finally it should be noted that: above embodiment only, in order to technical scheme of the present invention to be described, is not intended to limit; Although the present invention is had been described in detail with reference to previous embodiment, those of ordinary skill in the art is to be understood that: its technical scheme that still can record aforementioned each embodiment is modified, or part technical characterictic is wherein equal to replacement; And these modifications or replacement do not make the essence of appropriate technical solution depart from the spirit and scope of various embodiments of the present invention technical scheme.

Claims (5)

1. the oil spilling hyperspectral image classification method based on wavelet transformation low frequency spectrum information, is characterized in that, comprises the following steps:

Step 1: draw the curve of spectrum of oil spilling high spectrum image pixel, this curve of spectrum is configured to a continuous one-dimensional signal;

Step 2: this one-dimensional signal is carried out to wavelet decomposition, extract low frequency wavelet coefficient, obtain low frequency spectrum curve;

Step 3: select the obvious low frequency spectral coverage of classification separability to generate low-frequency image between the low frequency spectrum curve of different atural objects;

Step 4: select atural object ROI in the low-frequency image generating, to the atural object ROI the selecting classification that exercises supervision.

2. the oil spilling hyperspectral image classification method based on wavelet transformation low frequency spectrum information according to claim 1, is characterized in that, described step 1 comprises:

(a), read oil spilling high-spectral data, the SPECTRAL DIVERSITY reflecting according to different-thickness oil film, by the oil film atural object that is divided three classes, this three classes atural object is heavy oil film, thin oil film, seawater;

(b), select the ROI of three class atural objects, and draw spectral signature curve according to its spectrum characteristic parameter;

(c), select to distinguish the wave band interval of three class atural objects.

3. the oil spilling hyperspectral image classification method based on wavelet transformation low frequency spectrum information according to claim 2, is characterized in that, described step 2 comprises:

(a), respectively the pixel in the ROI of three class atural objects is carried out to 5 layers of decomposition based on db4 wavelet basis, and extract the low frequency wavelet coefficient of the 5th layer;

(b), draw the curve map of described low frequency wavelet coefficient, and choose the obvious wave band of separability according to the curve map of described wavelet coefficient.

4. the oil spilling hyperspectral image classification method based on wavelet transformation low frequency spectrum information according to claim 3, is characterized in that, described step 3 comprises:

Entire image is carried out to the 5 layers of decomposition based on db4 wavelet basis by pixel, extract the 5th layer of low frequency wavelet coefficient of the obvious wave band of separability, generate the obvious low frequency wavelet coefficient image of separability.

5. the oil spilling hyperspectral image classification method based on wavelet transformation low frequency spectrum information according to claim 4, is characterized in that, described step 4 comprises:

In the 5th layer of low frequency wavelet coefficient image, select respectively the ROI of three class atural objects, utilize the classification that exercises supervision of envi maximum likelihood method.