CN101615254B - Method for extracting coal fire information by hyperspectral remote sensing based on generalized addition model - Google Patents

Abstract

The invention provides a method for extracting coal fire information by hyperspectral remote sensing based on a generalized addition model. The method comprises the following steps: analyzing geologic features of a studying region and acquiring a ground object sample; measuring and analyzing spectrum of the ground object sample by a spectrometer; acquiring satellite-borne hyperspectral remote sensing data; preprocessing the obtained hyperspectral data; carrying out image spectral features statistics, comparing and analyzing the image spectral features with actually measured ground object spectral features, and selecting a diagnostic feature factor; based on the diagnostic factor, extracting coal fire burning information by the generalized addition model; verifying the information extracting result; and delineating a coal fire region. The method based on satellite-borne high spectrum data utilizes a generalized addition model to quantitatively extract the coal fire burning information,can be used for extracting underground coal fire information from the hyperspectral data, and provides decision-making support for delineating the coal fire region.

Description

Method for extracting coal fire information by hyperspectral remote sensing based on generalized addition model

Technical field

The present invention relates to utilize spaceborne high-spectrum remote sensing data to burn the change information exchange for underground coal fire and cross the method that generalized addition model carries out quantitative extraction, can be used for extracting underground coal fire from high-spectral data, for coal fire district's delineation provides decision support.

Background technology

Underground coal fire not only consumes a large amount of coal resources, and contaminated environment also has a strong impact on people's existence and life.Utilize during remote sensing technology monitors and administer underground coal fire, the space remote sensing data of multispectral/multidate (TM, ETM etc.) are though can to reach the trend of indication coal fire development accurate relatively not; Although the airborne hyperspectral data are accurate, its flight range is limited, and expensive, obtains data and is subjected to weather effect large, and spaceborne high-spectral data can overcome top shortcoming, and is convenient, economical and obtain accurately coal fire.The correlative study of before carrying out is to adopt airborne OMIS-I high-spectral data, and the main thermal infrared wave band (3～5 μ m and 8～14 μ m) that remains that utilizes carries out the monitoring of spontaneous combustion of coal seam, and the data that the reflected sunlight spectrum is distinguished do not take full advantage of.

For underground coal fire, high-spectral data excavates with weak information extraction technology and refers to obtain research with the aspect such as weak information classification recognition technology, the spectral information of extraction and pollutant and background, radiation information, the timely phase information of spatial information etc. from the high curve of spectrum and high spectrum image by ground high-spectrum DATA REASONING and spectroscopic data Treatment Analysis, satellite Hyperspectral imaging.In the research aspect this, to burn the change information extraction model for coal fire and select, most researchs are based on analysis expert and judge to have certain subjectivity, lack a kind of quantitative extraction system of selection.

Summary of the invention

Technology of the present invention is dealt with problems and is: propose to utilize spaceborne high-spectrum remote sensing data to burn the change information exchange for underground coal fire and cross the method that generalized addition model carries out quantitative extraction, be used for extracting underground coal fire from high-spectral data, for coal fire district's delineation provides decision support.

Technical solution of the present invention is: utilize spaceborne high-spectrum remote sensing data to burn the change information exchange for underground coal fire and cross the method that generalized addition model carries out quantitative extraction, the method concrete steps are as follows:

(1) carry out geologic feature for study area and the earth's surface phenomenon is analyzed, and gather the earth's surface sample;

(2) utilize spectrometer to carry out the Laboratory Spectra test to obtaining ground object sample, and spectroscopic data is analyzed;

(3) obtain spaceborne high-spectrum remote sensing data;

(4) the spaceborne high-spectrum remote sensing data of obtain is carried out pre-service;

Can comprise: improper pixel is corrected; Atmospheric correction; Smoothing processing; Geometry correction.Wherein improper pixel is corrected to process and is comprised uncertain nominal band removal; The conversion of absolute radiation value; Loop wire is repaired; Destriping is processed; The Smile effect is removed;

(5) carry out the spectral signature statistics based on image, and compare and analyze with the actual measurement object spectrum, select the diagnostic characterization factor;

(6) with the diagnostic factor chosen as parameter, utilize generalized addition model to carry out underground coal fire and burn and become information extraction;

(7) information extraction effect and precision test;

(8) coal fire zone delineation.

Wherein, the spaceborne high-spectrum remote sensing data pre-treatment step of described step (4) comprises successively:

1) high-spectral data is carried out improper pixel and correct processing, comprising: uncertain nominal band is removed, the conversion of absolute radiation value, and the loop wire reparation, the row destriping, the Smile effect is removed;

2) high-spectral data being carried out atmospheric correction processes;

3) high-spectral data is carried out smoothing processing;

4) high-spectral data being carried out geometry correction processes.

Wherein, described step (6) with the diagnostic factor chosen as parameter, carrying out underground coal fire burns in the quantitative extraction of change information, employing utilizes non-parametric method based on generalized addition model, with some from dependent variable between exist the independent variable of complex nonlinear relation to add with different functions and the form match enter model, carry out the extraction that coal fire burns change information.

The present invention's advantage compared with prior art is:

(1) utilize high-spectral data, start with from the coal field geology environment, on the existing bases such as evolution of analyzing coalfield spontaneous combustion rule, coal measure strata and coalfield flame range, carry out measured spectra analysis and target in hyperspectral remotely sensed image spectral analysis.Carry out simultaneously target in hyperspectral remotely sensed image and carry out pre-service, improve the quality of data, be used for the statistical study of follow-up object spectrum data.

(2) based on machine learning, utilize generalized addition to carry out underground coal fire and burn the information extraction of set amount, utilize non-parametric method, with some from dependent variable between exist the independent variable of complex nonlinear relation to add with different functions and the form match enter model.

Description of drawings

Fig. 1 is embodiment of the present invention process flow diagram;

Fig. 2 is embodiment of the present invention Hyperspectral imaging pretreatment process figure;

Fig. 3 is embodiment of the present invention GAM model information extraction proof diagram;

Fig. 4 is that embodiment of the present invention result is burnt change information delineation figure.

Embodiment

Below in conjunction with accompanying drawing, technical scheme of the present invention is described further.

As shown in Figure 1, specific implementation method of the present invention comprises: study area analyzing geological features and ground object sample are obtained; Utilize spectrometer to carry out ground object sample spectral measurement and analysis; Obtain spaceborne high-spectrum remote sensing data; Obtaining high-spectral data is carried out pre-service; Image spectral signature statistics also contrasts, analyzes with the actual measurement spectral characteristic of ground, selects the diagnostic characterization factor; Utilize generalized addition model to carry out coal fire based on the diagnostic factor and burn the change information extraction; The information extraction result verification; The delineation of coal fire district.

The embodiment of the present invention is selected to utilize the Hyperion high-spectral data, reaches the area take crow and carries out example as the test site, and concrete steps are as follows:

Step 1, reach for crow that geologic feature is carried out in the area and the earth's surface phenomenon is analyzed, comprise geological tectonic environment, coal seam distribution, mining engineering and goaf situation etc., and gather ground object sample and comprise: burnt rock, gangue, sulphur etc.;

Step 2, utilize spectrometer: (the ASD of U.S. spectral analysis apparatus company, Analytical SpectralDevice) the full spectrum portable light of the FR-Por spectrometer of development carries out the Laboratory Spectra test to obtaining ground object sample, and spectroscopic data is analyzed, comprise reflectivity, centre wavelength, absorb the degree of depth, halfwidth etc.;

Step 3, obtain spaceborne high-spectrum remote sensing data---the Hyperion high-spectral data;

Step 4, the spaceborne high-spectrum remote sensing data of obtaining Hyperion is carried out pre-service;

The Hyperion high-spectral data is processed by U.S. USGS and is generated, and through a series of rough handling processes such as blotch removal, echo correction, background removal, radiation correcting, bad pixel reparation and picture quality inspections, and this product is through geometric correction.In 242 wave bands of Hyperion L1 product, 1～70 wave band is visible light-near infrared (VNIR), and 71～242 is short infrared wave band (SWIR), and wherein 198 wave bands are processed through radiation calibration, the wave band of calibration comprises VNIR8～57 wave bands, SWIR77～224 wave bands.Due to VNIR56～57 wave bands and SWIR77～78 wave bands overlapping, therefore in fact only have 196 independent wave bands.There is no the wave band of calibration is 0 value.

For the characteristics of Hyperion data self, and the status analysis of study area, thereby carry out targetedly the albedo image that pre-service obtains the Hyperion data, idiographic flow as shown in Figure 2:

(1) improper pixel is corrected

Non-calibration wave band is removed: in 242 wave bands of Hyperion raw data, 1～7 wave band, 58～76 wave bands and 225～242 wave bands are not owing to calibrating, so wave band is made as 0 value, and it is removed the new image that comprises 198 wave bands of generation; In Hyperion data process 198 wave bands of radiation calibration, because 56～57 wave bands in VNIR overlap with 77～78 wave bands in SWIR, and the noise of VNIR56～57 wave bands is less than SWIR77～78 wave bands, therefore keep the former, deletion SWIR77～78 wave bands generate an image that comprises 196 independent wave bands.

The pixel value is to the conversion of absolute radiation value: raw data is divided into respectively two files of VNIR wave band and SWIR wave band, and with the VNIR band image divided by 40, generate a new image file; The SWIR wave band generates another new image file divided by 80; Then, two image files are merged, obtain absolute radiation value image.

Bad line reparation: with the mean value reparation of its adjacent row or column.

Destriping: utilize " strip coating method is removed in the part " and threshold value is set and remove.

The principle of " strip coating method is removed in the part " is: establish m _ikBe the mean value of K-band i row, s _ikBe the standard deviation of K-band i row, and establish m _ikAnd s _ikBe respectively mean value and the standard deviation of " reference picture ".Be also the difference place of " overall situation " method and " part " method, the average of " overall situation " method " reference picture " and standard deviation replace with average and the standard deviation of entire image, that is: herein

m _ik＝m _k，s _ik＝s _k

And " part " rule choose band row near adjacent several row as " reference picture ", and with average and the standard deviation of the mean value replacement " reference picture " of the mean of mean of these several column datas of getting choosing (not comprising band is listed as) and standard deviation, that is:

m _ik＝l _mean(m _ik)，s _ik＝l _mean(s _ik)

If the gain of sensor is α _ik, side-play amount is β _ik, the radiation value that in image, the i of K-band is listed as, j is capable * _ijkShould be modified to:

x′ _ijk＝α _ikx _ijk+β _ik

Wherein gain and skew is:

${\mathrm{\α}}_{\mathrm{ik}}=\frac{\stackrel{\‾}{s_{\mathrm{ik}}}}{s_{\mathrm{ik}}},$ β _ik＝m _ik-α _ikx _ik

Certainly, before calculating, should judge at first whether the image column element is band, concrete discriminant is:

$\mathrm{test}=\frac{|m_{\mathrm{ik}}-l_{\mathrm{med}}\left(m_{\mathrm{ik}}\right)|}{l_{\mathrm{med}}\left(s_{\mathrm{ik}}\right)}$

Above-mentioned whole algorithm realizes and completes the processing of image in Matlab software, some crucial problems are arranged when coding:

1. width neighborhood threshold value:

The width neighborhood threshold value setting is namely chosen near how many row of band row most suitable as a reference.Because the sensor of VRIN and SWIR data receiver is different, the band impact that produces is different, therefore, destriping to its image will carry out respectively, the corresponding reference columns of choosing is also different, find through this research of experiment: the width neighborhood of VNIR is made as 5, SWIR and is made as the best results that 20 images are processed.

2. differentiate the band threshold value setting:

By reading and the choosing of this threshold value of study forefathers, find it is the artificial estimation selected value of carrying out for its study area data mostly, suitable its single scape map sheet used only, but might not use for other Hyperion data.And this judgment threshold plays a very important role checking the band wave band, it is also one of user's factor of being difficult to hold most, so this task, choose carefully for differentiation band threshold value and study, utilize average and standard deviation to carry out value by the statistical value to entire image test discriminant, namely choose the test standard deviation that the test average adds 0.3 times, namely thresh=mean (test)+0.3*std (test) is as judgement boundary best results.

The Smile effect is removed: remove the Smile effect and can select Global destriping (overall situation is removed band), MNF Smoothing (minimal noise separates level and smooth), Cross-tracking illuminationcorrection (cross rail gamma correction) and Interpolation (interpolation) method, this work is selected therefore to select Cross-tracking illumination correction to remove through check Smile effect and not serious.

(2) atmospheric correction: at present, atmospheric correction method roughly can be summarized as 4 kinds: based on the relative correction method of characteristics of image, based on the linear regression model method in ground, based on atmospheric radiation transmission method and composite model method.Wherein use and be based on more widely the linear regression model method in ground, the method the mathematical approach mathematics and explicit physical meaning are calculated simple, but must be take a large amount of field spectrometries as prerequisite, therefore cost is higher, and is strong to the field work dependence, and stricter to the requirement of ground scaling point.And the method that in atmospheric correction method, correction accuracy is high is the radiative transfer model method.The model that utilizes the radiation transmission principle of electromagnetic wave in atmosphere to set up carries out the method for atmospheric correction to remote sensing images.Atmospheric correction technical development based on atmospheric radiation transmission has multiple module, as ATREM, ATCOR, ACORN, HATCH, FLAASH etc.The present invention selects to utilize the FLAASH atmospheric correction module of ENVI software, eliminates atmosphere and sees through the impact that the factors such as vapor permeability, the downward diffuse reflection of skylight cause image, improves picture quality.

(3) level and smooth: through the curve of spectrum ubiquity crenellated phenomena after atmospheric correction, noise is fairly obvious, is not easy to carry out the diagnostic spectral analysis.Therefore, need to carry out curve of spectrum smoothing processing, eliminate noise effect.Remote sensing image of the present invention is through the FLAASH atmospheric correction, and FLAASH module itself is with smoothing processing, but its effect and not obvious.The pixel that the spectrum that exists in image is level and smooth because the spectrum smoothing processing in FLAASH is based on, feature obvious and be easy to differentiate, wave spectrum after such special pixel and low-pass filtering is compared the linear relationship of setting up original spectrum and level and smooth spectrum, utilize linear relationship coefficient (comprising a spectrum gain coefficient and a transfer coefficient) that view picture is affected and carry out the spectrum smoothing processing.Choosing of suitable pixel is the key of smoothing processing, not only requires spectral signature stable gently, and brightness requirement is enough high in order to therefrom extract the spectrum gain coefficient.Be directed to the view data of this test site, effect not ideal.Therefore the present invention removes/relaxes the serious crenellated phenomena on curve in image by MNF (Minimum Noise Fraction, minimal noise separates) conversion method.

(4) geometry correction: choose ground control point and proofread and correct.

Step 5, carry out spectral signature statistics based on image, and compare and analyze with the actual measurement object spectrum, select the diagnostic characterization factor;

Step 6, with the diagnostic factor chosen as parameter, utilize generalized addition model to carry out underground coal fire and burn and become information extraction.

generalized addition model (Generalized Additive Models, GAM): generalized addition model is the expansion of generalized linear model imparametrization, more flexible than generalized linear model, the predicting the outcome of its model is not to come from one and presets good model, but adopt non-parametric method to carry out match, suppose by " addition ", with some from dependent variable between exist the independent variable of complex nonlinear relation to add with different functions and the form match enter model, can explore the non-dullness between variable, nonlinear relationship, and find out rule in data, thereby better predicted the outcome.

The present invention just only exists burning and two kinds of situations of unburned for classification and the prediction of coal fire spontaneous combustion, belongs to the two-value problem, is designated respectively 1 and 0, can classify and predict with the two-valued function homing method.But, complicated nonlinear relationship often between the burning of factor of influence and coal fire in real world, the linear function during the match two-valued function returns is difficulty comparatively, can bring larger error.Generalized addition model is expanded linear two-valued function regression model, and it has introduced the linear term in smooth function f replacement Logic Regression Models, and its form is:

$g\left(\mathrm{\μ}\right)=\mathrm{\α}+{\mathrm{\Σ}}_{j=1}^n{f}_j\left(x_j\right)$

Adopt non-parametric method, with some from dependent variable between exist the independent variable of complex nonlinear relation to add with different functions and the form match enter model.

Step 7, information extraction result verification; The comparison of effect and precision will be carried out by the result of setting up model extraction, as shown in Figure 3.

Step 8, coal fire burn and become the zone delineation; Burn according to the coal fire by model extraction the burning change zone delineation that change information is carried out underground coal fire, as shown in Figure 4.

Claims (1)

1. coal fire information by hyperspectral remote sensing quantification extracting method, it is characterized in that: study area analyzing geological features and ground object sample are obtained; Utilize spectrometer to carry out ground object sample spectral measurement and analysis; Obtain spaceborne high-spectrum remote sensing data; Obtaining high-spectral data is carried out pre-service; Image spectral signature statistics also contrasts, analyzes with the actual measurement spectral characteristic of ground, selects the diagnostic characterization factor; Utilize generalized addition model to carry out coal fire based on the diagnostic factor and burn the change information extraction; The information extraction result verification; The delineation of coal fire district;

Step 1, study area analyzing geological features and ground object sample are obtained: geological tectonic environment, coal seam distribution, mining engineering and goaf situation are analyzed, and gathered the ground object sample of burnt rock, gangue and sulphur;

Step 2, utilize spectrometer to carry out ground object sample spectral measurement and analysis: the full spectrum portable light of the FR-Por spectrometer of U.S. spectral analysis apparatus company development carries out the Laboratory Spectra test to obtaining ground object sample, and to reflectivity, centre wavelength, the spectroscopic data that absorbs the degree of depth and halfwidth is analyzed;

Step 3, obtain the spaceborne high-spectrum remote sensing data of Hyperion;

Step 4, the spaceborne high-spectrum remote sensing data of obtaining Hyperion is carried out pre-service;

For the characteristics of Hyperion data self, and the status analysis of study area, thereby carrying out targetedly the albedo image that pre-service obtains the Hyperion data, idiographic flow is as follows:

(1) improper pixel is corrected

Non-calibration wave band is removed: in 242 wave bands of Hyperion raw data, 1～7 wave band, 58～76 wave bands and 225～242 wave bands are not owing to calibrating, so wave band is made as 0 value, and it is removed the new image that comprises 198 wave bands of generation; In Hyperion data process 198 wave bands of radiation calibration, because 56～57 wave bands in VNIR overlap with 77～78 wave bands in SWIR, and the noise of VNIR56～57 wave bands is less than SWIR77～78 wave bands, therefore keep the former, deletion SWIR77～78 wave bands generate an image that comprises 196 independent wave bands;

The pixel value is to the conversion of absolute radiation value: raw data is divided into respectively two files of VNIR wave band and SWIR wave band, and with the VNIR band image divided by 40, generate a new image file; The SWIR wave band generates another new image file divided by 80; Then, two image files are merged, obtain absolute radiation value image;

Bad line reparation: with the mean value reparation of its adjacent row or column;

Destriping: utilize " strip coating method is removed in the part " and threshold value is set and remove;

The principle of " strip coating method is removed in the part " is: establish m _ikBe the mean value of K-band i row, s _ikBe the standard deviation of K-band i row, and establish

With

Be respectively mean value and the standard deviation of " reference picture "; Be also the difference place of " overall situation " method and " part " method, the average of " overall situation " method " reference picture " and standard deviation replace with average and the standard deviation of entire image, that is: herein

$\stackrel{\‾}{m_{\mathrm{ik}}}={\stackrel{\‾}{m}}_k,\stackrel{\‾}{s_{\mathrm{ik}}}=\stackrel{\‾}{s_k}$

And " part " rule choose band row near adjacent several row as " reference picture ", and with average and the standard deviation of the mean value replacement " reference picture " of the mean of mean of these several column datas of getting choosing and standard deviation, that is:

$\stackrel{\‾}{m_{\mathrm{ik}}}=l_{\mathrm{mean}}\left(\stackrel{\‾}{m_{\mathrm{ik}}}\right),\stackrel{\‾}{s_{\mathrm{ik}}}=l_{\mathrm{mean}}\left(\stackrel{\‾}{s_{\mathrm{ik}}}\right)$

If the gain of sensor is α _ik, side-play amount is β _ik, the radiation value x that in image, the i of K-band is listed as, j is capable _ijkShould be modified to:

x' _ijk=α _ikx _ijk+β _ik

Wherein gain and skew is:

${\mathrm{\α}}_{\mathrm{ik}}=\frac{\stackrel{\‾}{s_{\mathrm{ik}}}}{s_{\mathrm{ik}}},{\mathrm{\β}}_{\mathrm{ik}}=\stackrel{\‾}{m_{\mathrm{ik}}}-{\mathrm{\α}}_{\mathrm{ik}}{x}_{\mathrm{ik}}$

Certainly, before calculating, should judge at first whether the image column element is band, concrete discriminant is:

$\mathrm{test}=\frac{|m_{\mathrm{ik}}-l_{\mathrm{med}}\left(m_{\mathrm{ik}}\right)|}{l_{\mathrm{med}}\left(s_{\mathrm{ik}}\right)}$

Above-mentioned whole algorithm realizes and completes the processing of image in Matlab software;

(2) atmospheric correction: select to utilize the FLAASH atmospheric correction module of ENVI software, eliminate atmosphere and see through vapor permeability, the downward irreflexive factor of skylight to the impact that image causes, improve picture quality;

(3) level and smooth: as to separate the MNF conversion method by minimal noise and remove/relax the serious crenellated phenomena on curve in image;

(4) geometry correction: choose ground control point and proofread and correct;

Step 5, carry out spectral signature statistics based on image, and compare and analyze with the actual measurement object spectrum, select the diagnostic characterization factor;

Step 6, with the diagnostic factor chosen as parameter, utilize generalized addition model to carry out underground coal fire and burn and become information extraction;

Classification and prediction for the coal fire spontaneous combustion just only exist burning and two kinds of situations of unburned, belong to the two-value problem, are designated respectively 1 and 0, classify and predict with the two-valued function homing method; But, complicated nonlinear relationship often between the burning of factor of influence and coal fire in real world, the linear function during the match two-valued function returns is difficulty comparatively, can bring larger error; Generalized addition model is expanded linear two-valued function regression model, has introduced the linear term in smooth function f replacement Logic Regression Models, and its form is:

$g\left(\mathrm{\μ}\right)=\mathrm{\α}+{\mathrm{\Σ}}_{j=1}^n{f}_j\left(x_j\right)$

Adopt non-parametric method, with some from dependent variable between exist the independent variable of complex nonlinear relation to add with different functions and the form match enter model;

Step 7, information extraction result verification; To carry out by the result of setting up model extraction the comparison of effect and precision,

Step 8, coal fire burn and become the zone delineation; Burn according to the coal fire by model extraction the burning change zone delineation that change information is carried out underground coal fire.

Images