CN104268838B - Wavelet denoising algorithm oriented to hyperspectral databases - Google Patents

Abstract

The invention discloses a wavelet denoising algorithm oriented to hyperspectral databases and belongs to processing methods for hyperspectral data. The wavelet denoising algorithm oriented to the hyperspectral databases is used for denoising hyperspectral curves with noise. The wavelet denoising algorithm oriented to the hyperspectral databases sequentially comprises the steps of wavelet denoising parameter selection and wavelet denoising. In the wavelet denoising parameter learning process, characteristics of the hyperspectral curves of various substances in the hyperspectral databases are fully considered, wavelet denoising parameters suitable for most of the hyperspectral curves of the hyperspectral databases are rapidly selected, and then the selected parameters are used for denoising the hyperspectral curves obtained through measurement. According to the wavelet denoising algorithm, the wavelet denoising parameters suitable for the hyperspectral databases can be rapidly selected according to the hyperspectral databases of various different application scenarios, the denoising effect is improved, and the accuracy rate of spectrum matching is increased.

Description

A kind of Wavelet Algorithm towards hyperspectral data storehouse

Technical field

The invention belongs to hyperspectral data processing technology field, it is related to a kind of processing method of hyperspectral data, specifically relates to And a kind of Wavelet Algorithm towards hyperspectral data storehouse, for carrying noisy ultraphotic spectral curve to carry out noise reduction.

Background technology

High-spectrum remote-sensing is an emerging earth observation technology, and it is in the ultraviolet of electromagnetic spectrum, visible ray, closely red Outward, in the range of mid-infrared and Thermal infrared bands, many fine, continuously narrow spectral coverage glazing modal data technology are obtained.Fortune With the airborne or spaceborne instrument with high spectral resolution, remote sensing is carried out to earth surface and can obtain a lot of ground observations The information hardly resulting in.In order to realize the identification of material, it usually needs the hyperspectral data that measurement is obtained and known ultraphotic The hyperspectral data of each material in modal data storehouse is mated.But the hyperspectral data that measurement obtains can be subject to various unavoidably Noise before atmospheric reflectance, sunlight etc. receive can be introduced in the middle of the interference of noise, such as process from signal source to receptor for the signal, After signal is received, internal system circuit also can produce noise.In order to improve the accuracy rate of coupling, the ultraphotic that measurement obtains is composed Data carries out noise reduction and is very important.

Wavelet transformation has good Time-Frequency Localization property, is widely applied in signal denoising.Wavelet transformation Noise under multi-scale wavelet transformation can be utilized, and the wavelet coefficient under each yardstick is different with signal, thus preferably remove making an uproar Sound, reaches the purpose recovering primary signal, and the fast algorithm that small echo signal has, the requirement of calculating time can be met. Wavelet transformation have very strong remove data dependence energy, make the energy of spectral signal concentrate on larger little of some amplitudes of wavelet field In wave system number, and the wavelet coefficient of noise is then uniformly distributed in whole wavelet field.Therefore, signals and associated noises are through wavelet decomposition Afterwards it is believed that the larger wavelet coefficient of amplitude is based on signal, the less coefficient of amplitude is largely noise.To belong to It is possible to reduction obtains eliminating the signal of noise after the wavelet coefficient of noise is suppressed.

However, wavelet decomposition algorithm needs by setting the parameters such as different basic functions, Decomposition order and threshold value to noisy Signal is decomposed, and is separated with noise signal with realizing primary signal.The selection of these parameters will directly affect denoising effect Quality.These parameters have up to thousands of kinds of combination.For different applications, that is, different hyperspectral data storehouse, it is required for Find a kind of parameter combination being adapted therewith so that can realize relatively to the hyperspectral data in this data base by Wavelet Denoising Method Good noise reduction.Common practice is to travel through all combinations, selects optimal parameter combination, this will expend several hours.With When, when choosing optimal parameter combination, the judgment criteria of employing is not quite reasonable, only considered the absolute effect of noise reduction.Due to The spectrum that there are some materials from meeting in data base has the noise-like feature of more classes in itself, no matter is adopted which kind of noise reduction parameters The noise reduction of these spectrum is all more poor than other spectrum, is therefore gone in judge meeting ability selection course using absolute noise reduction More consider the noise reduction of this kind of small part peculiar spectrum, do not choose the noise reduction parameters being suitable for major part spectrum. It is, thus, sought for one kind is according to different hyperspectral data Sink Characteristics, can be suitable for this data base's to one by fast searching The method of Wavelet Denoising Method parameter.

Content of the invention

In order to solve above-mentioned technical problem, the invention provides a kind of Wavelet Denoising Method towards hyperspectral data storehouse is calculated Method is it is therefore an objective to quickly select suitable wavelet de-noising parameter to carry out noise reduction to the noisy hyperspectral data of band.

The technical solution adopted in the present invention is: a kind of Wavelet Algorithm towards hyperspectral data storehouse, and its feature exists In comprising the following steps:

Step 1: select wavelet de-noising parameter, described wavelet de-noising parameter includes threshold selection function, threshold function table, threshold Value Adjusted Option, Decomposition order, wavelet basis function；Fast search is carried out to various wavelet de-noising parameter combinations, selects to be applied to The wavelet de-noising parameter combination in the corresponding hyperspectral data storehouse of current application background；

Step 2: wavelet de-noising is carried out to the hyperspectral data measuring with selected wavelet de-noising parameter combination.

Preferably, fast search, its fast search side are carried out to various wavelet de-noising parameter combinations described in step 1 The process of realizing of method includes following sub-step:

Step 1.1: set the spectrum of total k kind material in hyperspectral data storehouse, be designated as set s={ s₁,s₂,…,s_k, Add white Gaussian noise, the spectra collection with noise for the note is combined into t={ t in this k bar spectrum₁,t₂,…,t_k}；If treating the small echo of selection Each total m kind of parameter group, remembers that the collection that this m kind parameter group is formed is combined into g={ g₁,g₂,…,g_m}；

Step 1.2: randomly select a curve of spectrum t from t_x, using parameter combinations all in g to t_xCarry out small echo fall Make an uproar, and calculate spectrum and s after noise reduction respectively_xSpectrum intervals, according to the ascending sequence of spectrum intervals, before selection, c% is Little spectrum intervals, obtains the parameter combination of corresponding front c% simultaneously, is designated as set b, wherein c is predetermined parameter；

Step 1.3: judge, in set b, whether parameter combination number is equal to 1；

If parameter combination number is 1 in set b, choose this parameter combination be preferably after wavelet de-noising parameter, this stream Journey terminates；

If the parameter combination in set b is more than 1, execute following step 1.4；

Step 1.4: update set g, using set b as new set g, by t_xReject from set t；

Step 1.5: judge, reject t_xAfterwards, if whether still having element in set t；

If still there being element in set t, the described step 1.2 of revolution execution；

If set t is null set, execute following step 1.6；

Step 1.6: calculate using each parameter combination in set b to after k Banded improvement spectrum noise reduction with s_xSpectrum intervals Averaged spectrum apart from ansa, choose the corresponding parameter combination of minimum average B configuration spectrum intervals ansk be preferably after wavelet de-noising ginseng Number, this flow process terminates.

Preferably, the averaged spectrum described in step 1.6 is as follows apart from ansa computational methods:

$\mathrm{ansa}=\frac{{\mathrm{\σ}}_{n=1}^k{\mathrm{nsa}}_n}{k}$

Wherein, nsa_nBe using parameter current combination to nth bar band din-light spectrum noise reduction after with s_nSpectrum intervals, its calculating Formula is as follows:

${\mathrm{nsa}}_n=\frac{{\mathrm{sa}}_n-{\mathrm{sa}}_{\min }}{{\mathrm{sa}}_{\max }-{\mathrm{sa}}_{\min }}$

Wherein,

${\mathrm{sa}}_n={\cos }^{-1}\frac{<d_s,s_n>}{\left|d_n\right|\left|s_n\right|}$

sa_min=min (sa₁,sa₂,...,sa_k)

sa_max=max (sa₁,sa₂,...,sa_k)

Wherein, d_nIt is, using parameter current combination, with din-light, the noise reduction spectrum obtaining after noise reduction, < d are composed to nth bar_n,s_n> be d_nWith s_nInner product, | | represent take two norms.

The present invention includes the selection of wavelet de-noising parameter, wavelet de-noising step, due to during wavelet de-noising parameter learning Take into full account the characteristic of various material ultraphotic spectral curves in hyperspectral data storehouse, introduce the thought of Fast Learning, thus can Select to be suitable for the wavelet de-noising parameter of the hyperspectral data storehouse overwhelming majority curve of spectrum with quick, then obtained using selection Parameter carries out noise reduction to the ultraphotic spectral curve measuring.This method can quickly be applied in various hyperspectral data storehouses, fall Effect of making an uproar is good, improves the accuracy rate of Spectral matching.

Brief description

Fig. 1: for the flow chart of the embodiment of the present invention.

Fig. 2: for the method for fast searching flow chart of the embodiment of the present invention.

Specific embodiment

Understand for the ease of those of ordinary skill in the art and implement the present invention, below in conjunction with the accompanying drawings and embodiment is to this Bright be described in further detail it will be appreciated that described herein enforcement example be merely to illustrate and explain the present invention, not For limiting the present invention.

Ask for an interview Fig. 1, the technical solution adopted in the present invention is: a kind of Wavelet Algorithm towards hyperspectral data storehouse, Comprise the following steps:

Step 1: select wavelet de-noising parameter, described wavelet de-noising parameter includes threshold selection function, threshold function table, threshold Value Adjusted Option, Decomposition order, wavelet basis function；Fast search is carried out to various wavelet de-noising parameter combinations, selects to be applied to The wavelet de-noising parameter combination in the corresponding hyperspectral data storehouse of current application background；Ask for an interview Fig. 2, method for fast searching realize process Including following sub-step:

Step 1.1: set the spectrum of total k kind material in hyperspectral data storehouse, be designated as set s={ s₁,s₂,…,s_k, Add white Gaussian noise, the spectra collection with noise for the note is combined into t={ t in this k bar spectrum₁,t₂,…,t_k}；If treating the small echo of selection Each total m kind of parameter group, remembers that the collection that this m kind parameter group is formed is combined into g={ g₁,g₂,…,g_m}；

Step 1.2: randomly select a curve of spectrum t from t_x, using parameter combinations all in g to t_xCarry out small echo fall Make an uproar, and calculate spectrum and s after noise reduction respectively_xSpectrum intervals, according to the ascending sequence of spectrum intervals, choose first 10% Little spectrum intervals, obtains the parameter combination of corresponding first 10% simultaneously, is designated as set b；

Step 1.3: judge, in set b, whether parameter combination number is equal to 1；

If parameter combination number is 1 in set b, choose this parameter combination be preferably after wavelet de-noising parameter, this stream Journey terminates；

If the parameter combination in set b is more than 1, execute following step 1.4；

Step 1.4: update set g, using set b as new set g, by t_xReject from set t；

Step 1.5: judge, reject t_xAfterwards, if whether still having element in set t；

If still there being element in set t, the described step 1.2 of revolution execution；

If set t is null set, execute following step 1.6；

Step 1.6: calculate using each parameter combination in set b to after k Banded improvement spectrum noise reduction with s_xSpectrum intervals Averaged spectrum apart from ansa, choose the corresponding parameter combination of minimum average B configuration spectrum intervals ansk be preferably after wavelet de-noising ginseng Number, this flow process terminates.

Wherein averaged spectrum is as follows apart from ansa computational methods:

$\mathrm{ansa}=\frac{{\mathrm{\&sigma;}}_{n=1}^k{\mathrm{nsa}}_n}{k}$

Wherein, nsa_nBe using parameter current combination to nth bar band din-light spectrum noise reduction after with s_nSpectrum intervals, its calculating Formula is as follows:

${\mathrm{nsa}}_n=\frac{{\mathrm{sa}}_n-{\mathrm{sa}}_{\min }}{{\mathrm{sa}}_{\max }-{\mathrm{sa}}_{\min }}$

Wherein,

${\mathrm{sa}}_n={\cos }^{-1}\frac{<d_s,s_n>}{\left|d_n\right|\left|s_n\right|}$

sa_min=min (sa₁,sa₂,...,sa_k)

sa_max=max (sa₁,sa₂,...,sa_k)

Wherein, d_nIt is, using parameter current combination, with din-light, the noise reduction spectrum obtaining after noise reduction, < d are composed to nth bar_n,s_n> be d_nWith s_nInner product, | | represent take two norms.

Step 2: wavelet de-noising is carried out to the hyperspectral data measuring with selected wavelet de-noising parameter combination.

The k of the present invention is different according to different spectra database number, and m is the number of parameter combination to be selected it is also possible to root According to need adjust.

It should be appreciated that the part that this specification does not elaborate belongs to prior art.

It should be appreciated that the above-mentioned description for preferred embodiment is more detailed, can not therefore be considered to this The restriction of invention patent protection scope, those of ordinary skill in the art, under the enlightenment of the present invention, is weighing without departing from the present invention Profit requires under protected ambit, can also make replacement or deform, each fall within protection scope of the present invention, this Bright scope is claimed should be defined by claims.

Claims (2)

1. a kind of Wavelet Algorithm towards hyperspectral data storehouse is it is characterised in that comprise the following steps:

Step 1: select wavelet de-noising parameter, described wavelet de-noising parameter includes wavelet basis function, threshold selection function, decomposition The number of plies, adjusting thresholds scheme, threshold function table；Various wavelet de-noising parameter combinations are carried out with fast search, selects to be applied to currently The wavelet de-noising parameter combination in the corresponding hyperspectral data storehouse of application background；

Described various wavelet de-noising parameter combinations are carried out with fast search, its method for fast searching realize process include following Sub-step:

Step 1.1: set the spectrum of total k kind material in hyperspectral data storehouse, be designated as set s={ s₁,s₂,…,s_k, in this k bar Add white Gaussian noise, the spectra collection with noise for the note is combined into t={ t in spectrum₁,t₂,…,t_k}；If treating the wavelet de-noising ginseng of selection Array has amounted to m kind, remembers that the collection that this m kind parameter group is formed is combined into g={ g₁,g₂,…,g_m}；

Step 1.2: randomly select a curve of spectrum t from t_x, using parameter combinations all in g to t_xCarry out wavelet de-noising, and Calculate spectrum and s after noise reduction respectively_xSpectrum intervals, according to the ascending sequence of spectrum intervals, c% minimum light before selection Spectrum distance from, obtain the parameter combination of corresponding front c% simultaneously, be designated as set b, wherein c is predetermined parameter, x takes 1～ k；

Step 1.3: judge, in set b, whether parameter combination number is equal to 1；

If parameter combination number is 1 in set b, choose this parameter combination be preferably after wavelet de-noising parameter, this flow process ties Bundle；

If the parameter combination in set b is more than 1, execute following step 1.4；

Step 1.4: update set g, using set b as new set g, by t_xReject from set t；

Step 1.5: judge, reject t_xAfterwards, whether still there is element in set t；

If still there being element in set t, the described step 1.2 of revolution execution；

If set t is null set, execute following step 1.6；

Step 1.6: calculate using each parameter combination in set b to after k Banded improvement spectrum noise reduction with s_xAveraged spectrum distance Ansa, choose the corresponding parameter combination of minimum average B configuration spectrum intervals ansk be preferably after wavelet de-noising parameter, this flow process terminates；

Step 2: wavelet de-noising is carried out to the hyperspectral data measuring with selected wavelet de-noising parameter combination.

2. the Wavelet Algorithm towards hyperspectral data storehouse according to claim 1 it is characterised in that: in step 1.6 Described averaged spectrum is as follows apart from ansa computational methods:

$ansa=\frac{{\mathrm{\&sigma;}}_{n=1}^k{\mathrm{nsa}}_n}{k}$

Wherein, nsa_nBe using parameter current combination to nth bar band din-light spectrum noise reduction after with s_nSpectrum intervals, n takes 1～k；Its Computing formula is as follows:

${\mathrm{nsa}}_n=\frac{{\mathrm{sa}}_n-{\mathrm{sa}}_{min}}{{\mathrm{sa}}_{max}-{\mathrm{sa}}_{min}}$

Wherein,

${\mathrm{sa}}_n={\cos }^{-1}\frac{<d_n,s_n>}{\left|d_n\right|\left|s_n\right|}$

sa_min=min (sa₁,sa₂,...,sa_k)

sa_max=max (sa₁,sa₂,...,sa_k)

Wherein, d_nIt is, using parameter current combination, with din-light, the noise reduction spectrum obtaining after noise reduction, < d are composed to nth bar_n,s_n> it is d_nWith s_nInner product, | | represent take two norms, | d_n||s_n| it is that two norms are multiplied.