CN101276420A - Classification method for syncretizing optical spectrum information and multi-point simulation space information - Google Patents

Abstract

A classification method for integrating spectrum information with multipoint simulation space information is disclosed: comprising the steps of: (1) performing MLC classification to pixels of the spectrum information of a remote sense image to obtain possibility vector of each pixel corresponding to each category; (2) selecting sample data for the multipoint simulation MPS; (3) using the multipoint simulation method to establish the possibility model for each pixel in the image according to the number of condition data thereof, and storing the possibility vector; (4) adopting a data integration method to integrate two possibility vectors; (5) obtaining the result which is a possibility vector composed by corresponding each category to the home possibility, in the possibility vector of each pixel, the pixel whose home possibility corresponding to certain category is the largest, is classified to the relative category for the final classification result. The invention improves the classification precision and has extraordinarily extensive application in decoding the remote sense image so that the invention can be applied to the fields of geologic mine, weather, geography, mapping, ocean research, military spy and environmental monitor.

Description

A kind of sorting technique that merges spectral information and multiple spot simulation spatial information

Technical field

The invention belongs to the Spatial Information Technology field, specifically, relate to a kind of sorting technique that merges spectral information and multiple spot simulation spatial information.

Technical background

The computing machine remote sensing image classification is the concrete application of statistical model recognition technology in the remote sensing field.Its basic process is to extract one group of statistical characteristics of pattern to be identified, makes a policy according to certain criterion then, each pixel in the image is judged classification under it, thereby digital picture is discerned.The spectral signature that the main foundation of remote sensing image classification is, i.e. the electromagnetic multiband measured value of atural object is with its primitive character variable as remote sensing image classification.Yet owing to exist the phenomenon of a large amount of foreign matters in the Nature with spectrum and the different spectrum of jljl, in some cases, only utilize spectral information to distinguish all kinds of atural objects fully, introducing space structure information and spatial coherence information improve nicety of grading and are very important.At present, the sorting technique of consideration spatial information or image aspects mainly contains context sorting technique, textural characteristics sorting technique and utilizes traditional geostatistics sorting technique etc. on the basis of spectral information.These methods are extensive use of, and have effectively improved nicety of grading to a certain extent.But they also have limitation separately, such as, the method of context classification considers that closing on pixel is tending towards identical or close atural object classification, introduces structural information by the relation of interdependence on this statistical significance, and this method is not considered the spatial relationship on the large scale; The textural characteristics classification needs a large amount of reproductions of atural object, such as forest, and the atural object that farmland or the like occurs in flakes.Though tradition geostatistics sorting technique is not subjected to the influence of yardstick, it is not enough describing for the atural object classification with labyrinth. and the characteristic of multiple spot geostatistics is for a kind of new approach that provides is provided.

Since the sixties in 19th century, professor Matheron founded geostatistics, geostatistics was widely used in the research of numerous areas such as geography, ecology, environmental science, pedology.All these application all are based on the variation function, but the variation function can only reflect the spatial coherence of point-to-point transmission, being difficult to characterize complicated spatial structure and reproducing the geometric shape of complex target. the multiple spot geostatistics is with respect to for traditional geostatistics of variation function, propose in 1993 by Guardiano and Srivastava, thereafter, Journel, Strebelle, Zhang and Switzer etc. constantly improve on this basis.Multiple spot simulation (multiple-point simulation-MPS) is its computer implemented main algorithm, its main thought is to use training image to replace variation function to express atural object structural information and spatial autocorrelation information, thereby has overcome the deficiency that traditional geostatistics can not effectively reproduce complicated atural object geometric shape.Though it can reproduce the spatial information of atural object effectively, but this method does not propose at remote sensing image classification, do not consider the main information source of remote sensing images: spectral information, therefore the present invention combines both, spectral information and space structure information have been taken into full account simultaneously, and be not limited to simple atural object structure, can correctly classify has the complicated atural object of obvious architectural feature, has overcome the deficiency of using traditional variation function.

Summary of the invention

Technology of the present invention is dealt with problems: a kind of sorting technique that merges spectral information and multiple spot simulation spatial information is provided, this method is used spectral information and space structure information effectively simultaneously in the remote sensing image classification process, overcome and to have successfully managed the different spectrum of jljl in traditional remote sensing image classification and, improve nicety of grading with the deficiency of composing the foreign matter phenomenon.

Technical solution of the present invention: a kind of sorting technique that merges spectral information and multiple spot simulation spatial information comprises following basic step:

Step 1, the pixel of the spectral information of remote sensing images is carried out the MLC classification, obtain each pixel for probability vector of all categories;

MLC is one of supervised classification method that often uses, and in a lot of documents description is arranged all.The approximate Normal Distribution of spectral signature of its supposition training area atural object, by obtaining each pixel for belonging kinds of all categories, just pixel is subordinated to the posterior probability of each class categories, this pixel is assigned in the classification that belongs to the probability maximum gone.The present invention at first utilizes the MLC classification that the spectral information of remote sensing images is classified, and obtains the probability vector of each pixel for ownership of all categories.

Step 2, be that multiple spot simulation MPS selects sample data

For multiple spot simulation MPS selects the method for sample data to have multiplely be: such as with the sample data in the remote sensing images MLC classification as the MPS sample data, or with the sample data of pixel comparatively definite in the remote sensing images as MPS, such as a threshold value 0.8 is set, every ownership probability for a certain classification is greater than this threshold value, so just with its sample data as MPS.

Step 3, according to sample data selected in the step 2,, use the multiple spot analogy method to set up probability model and preserve probability vector all according to its condition data number for each pixel in the image;

Multiple spot analogy method wherein can be the analogy method that uniformly-spaced sorts, or the SNESIM modeling algorithm, or neural network multiple spot analogy method.The step of wherein said uniformly-spaced sort method is as follows:

(1) sets up search tree according to selected training image and training template;

(2) ordering N at interval is set;

(3) remaining photofit picture unit is sorted obtain sequence P, top n among the sequence P is waited to simulate pixel, and (wherein initial condition data is exactly a sample data according to training template to search for its condition data in the hunting zone of appointment, thereafter each pixel of simulating all can add wherein as new condition data), and then search and calculate its ownership probability in search tree for each classification, thereby obtain the probability vector of the classification results of each pixel, and when each pixel simulated after all as the remaining condition data of not simulating pixel;

(4) after sequence top n pixel simulation finishes, sequence is sorted again, and repeating step (3), all simulate until all pixels and finish.

Step 4, employing data fusion method merge probability vector that obtains in the step 2 and the probability vector in the step 3

Data fusion method can adopt the logarithm suggestion pond of Consensus theory, or linear suggestion pond, or evidence theory and based on the fusion method of statistical theory, but is not limited to the said method of mentioning.

How that one group of expert's conclusion separately is harmonious the theoretical fusion method of Consensus be to study method, and it is widely used in statistics and management science.In the Consensus theory, two main method are arranged: logarithm suggestion pond (LogarithmicOpinion Pool-Log-OP) and linear suggestion pond (Linear Opinion Pool-Linear-OP).The principle of these two methods is very simple.The decision function of Log-OP is:

$g_j\left(X\right)=\underset{i=1}{\overset{M}{\mathrm{\Π}}}{p}_i^{{\mathrm{\λ}}_i}\left({\mathrm{\ω}}_j\right|X)---\left(1\right)$

The decision function of Linear-OP is:

$g_j\left(X\right)=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\λ}}_i{p}_i\left({\mathrm{\ω}}_j\right|X)---\left(2\right)$

In above two formulas, p _i(ω _j| X) be the probability estimate of i expert, λ to classification under the X ₁..., λ _MIt is the weight of distributing to each expert.

Linear suggestion pond is more simpler comparatively speaking, and when the weight of distributing to each expert all be during less than 1 nonnegative integer, can obtain a kind of probability measure of conclusion.And logarithm suggestion pond, relative complex, and an expert's probability estimate is arranged is 0 o'clock, and the probability of whole conclusion just is 0, and this point is not suitable for MLC+MPS method needs.Therefore the method for using in the embodiment of the invention merges for the Consensus theory based on linear suggestion pond.

Linear suggestion pond data fusion method is: n class atural object is arranged, be designated as Ω={ ω ₁, ω ₂..., ω _n, pixel x probability vector in the MLC classification results is P ₁=(p ₁₁, p ₁₂..., p _1n), and the probability vector that the MPS simulation obtains is P ₂=(p ₂₁, p ₂₂..., p _2n), getting the MLC weight assignment is λ ₁, and the MPS weight is λ ₂, the following formula of having adopted of fusion process $g_j\left(X\right)=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\λ}}_i{p}_i\left({\mathrm{\ω}}_j\right|X),$ Then the ownership probability for first kind atural object is g ₁(x)=λ ₁p ₁₁+ λ ₂p ₂₁, be g for the ownership probability of the second class atural object ₂(x)=λ ₁p ₁₂+ λ ₂p ₂₂..., and be that the ownership probability of n class atural object is g for and for the ownership probability of the second class atural object _n(x)=λ ₁p _1n+ λ ₂p _2n, the probability vector after merging at last is G={g ₁(x), g ₂(x) ..., g _n(x) }.

Step 5, to obtain the result in step 4 be a probability vector of being made up of corresponding ownership probability of all categories, in the probability vector of each pixel to the ownership probability maximum of which classification, just this pixel is assigned in the corresponding classification and gone, thereby obtain final classification results.

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

(1) spectral information that had both taken into full account remote sensing images of the present invention, the space structure information and the spatial coherence information of atural object have been considered again, and be not limited to simple atural object structure, and can consider the space structure information of large scale and small scale simultaneously, thereby improved nicety of grading.

(2) the present invention is applied to the knowwhy of multiple spot simulation in the middle of the remote sensing image classification first, has improved nicety of grading.

Be analogue unit with the pixel when (3) the present invention simulates, can faithful to sampled data, make classification reliable more and true.

(4) the present invention utilizes search tree to store the conditional probability of training image, obtains local condition's probability distribution function as long as search for from search tree at every turn, has so just significantly reduced the time complexity of algorithm.

(5) owing to used MPS to extract the space structure information of atural object, the type of ground objects that the present invention was suitable for is not limited to simple atural object structure, can correctly classify to have the complicated atural object of obvious architectural feature, has overcome the deficiency of using traditional variation function.

(6) the present invention adopts data fusion method that two classification results are merged, and has so just considered spectral information and the space structure information and the spatial autocorrelation information of atural object simultaneously.Compare with common Classifying Method in Remote Sensing Image and to have the precision height, be not subjected to the advantage of space structure information dimensional constraints.Have application prospect in a lot of fields, can be applied to fields such as geological and mineral, meteorology, geography, mapping, ocean research, military surveillance and environmental monitoring.

(7) the present invention utilizes search tree to store the conditional probability of training image, greatly reduces time complexity.

Description of drawings

Fig. 1 is realization flow figure of the present invention;

Fig. 2 influences the pseudo-colours composite diagram of 4,3,2 wave bands for instantiation chosen area TM of the present invention;

Fig. 3 is a MLC sample data of the present invention, and wherein a is the road sampled point, and b is non-road sampled point, and c is sampled point not;

Fig. 4 is a MLC classification results of the present invention, and wherein a is a road, and b is non-road;

Fig. 5 a is MPS template figure of the present invention;

Fig. 5 b is a MPS training image of the present invention;

Fig. 6 is the classification results after employing the inventive method, and wherein a is a road, and b is non-road.

Embodiment

Fig. 1 is the process flow diagram of whole sorting technique of the present invention, provides specific implementation below in conjunction with example.

The experimental data of examples of algorithms is chosen Landsat Tm remote sensing image from area, China Huanghe delta on August 28th, 1999 as experimental data, imagery zone is positioned at the area, boundary of Dongying city, Shandong Province and Accessories during Binzhou, image size is the 515X515 pixel, resolution 30m, form by 7 wave bands, upper left corner latitude and longitude coordinates is 118 ° 0 ' 34.07 " 37 ° 22 ' 24.00 of E " N, lower right corner longitude and latitude is 118 ° 10 ' 52.83 " 37 ° 13 ' 58.13 of E " N.Fig. 2 is the false colored composite diagrams of 5,4,3 wave bands of Experimental Area.

Whole implementation process can be divided into three parts, is respectively MLC classified part, MPS simulation part, data fusion part.Wherein MLC classification can obtain the spectral information of atural object, and acquisition space structure information and spatial autocorrelation information that the MPS simulation can atural object, last two kinds of information combine under specific fusion method, jointly classification under the atural object are made a policy.

(1) pixel to the spectral information of remote sensing images carries out the MLC classification, obtains each pixel for probability vector of all categories;

MLC is the supervised classification method of remote sensing images relatively commonly used, present embodiment uses PCI Geomatica 9.0 to realize the MLC assorting process, also can use other software, perhaps oneself programme and realize the MLC classification, but need obtain the ownership probability of each pixel each classification.The training zone of the training sample of MLC classification is chosen as shown in Figure 3, always has 7065 sample datas, and wherein the sample data of road is 2759, and off-highroad sample data is 4306, and the MLC classification results as shown in Figure 4.Classification results is very poor as can be seen, and only have very obviously the pixel of spectral signature and correctly distinguished, road especially, a large amount of non-roads is divided for road by mistake.

(2) select sample data for multiple spot simulation MPS;

In the present embodiment, finish after the MLC classification, employed sample data in the MLC classification as the sampled data among the MPS, again in the practical application, also can be according to circumstances with the sample data of pixel comparatively definite in the remote sensing images as MPS, a threshold value promptly is set, and every ownership probability for a certain classification is greater than this threshold value, then with its sample data as MPS.

(3), according to its condition data number, use the multiple spot analogy method to set up probability model and preservation probability vector for each pixel in the image according to sample data selected in (2);

After definite MPS sampled data, need to use multiple spot simulation ground algorithm that each pixel is set up probability model and preserved probability vector.What this example used is the method that uniformly-spaced sorts and simulate,

A) use template among Fig. 5 (Fig. 5 a) and training image (Fig. 5 b) generate search tree.

Concerning the SNESIM algorithm, the template of selecting representative training image and can making full use of training image information is very important, the search tree that has only training image and data template by science just can obtain obtaining correct probability.Here training image has been represented the general features at this area road, the linear ground object that intersects anyhow just, and consider that this regional road all has certain inclination angle with respect to direct north, so training image also has certain inclination angle; On the other hand, data template shape comparison rule, and and the training image of road can be combined together the space structure information and the spatial autocorrelation information that reflect in the training image jointly to be implied preferably.In this experiment since the space structure information of the general small scale of road and spatial autocorrelation information more than abundant under the situation of large scale, and the number of sample data is bigger, so do not use many lattice point simulations, and because the abundant information under the small scale situation, though used smaller training image, but also can obtain the repetition of the data event of enough numbers, thereby also can not have influence on the effect of simulation.

After having set up search tree, not that the pixel of sample data is simulated to each.In order to make each pixel can both when simulating, obtain sufficient condition data, therefore before each pixel simulation, all need to be sorted in the path, the pixel that this condition data is few will be simulated when condition data is sufficient again, therefore can access sufficient space structure information and spatial autocorrelation information.In addition, if a pixel without any condition data, just uses the marginal probability distribution-road of estimating to account for 30%, non-road accounts for 70% as its probability vector.

B) ordering N at interval is set;

In order to obtain reasonable simulate effect, and between efficient and speed, average out, we propose to use the method for uniformly-spaced ordering to simulate, this method needs to set according to actual needs ordering N at interval, the ordering here is spaced apart 1, all needs to sort after just each point has been simulated.

C) remaining photofit picture unit is sorted obtain sequence P, top n among the sequence P is waited to simulate pixel in the hunting zone of appointment, search for its condition data according to the training template, and then search and calculate its ownership probability in search tree for each classification, thereby obtain the probability vector of the classification results of each pixel, after each pixel has been simulated, all with it as the remaining condition data of not simulating pixel;

D) after sequence top n pixel simulation finishes, sequence is sorted again, repeating step (3) is all simulated until all pixels and to be finished.

After obtaining probability vector, this probability vector is preserved, and the classification of this pixel is grouped in that class of probable value maximum in the corresponding probability vector temporarily, and the analogue value with this pixel remains the condition data of not simulating pixel as remaining then.Because the pixel that sample data is a belonging kinds to be determined, so the probability of classification was 100% under the probability vector of those sample datas was made as, other is 0%.After having simulated, all pixels just obtained the affiliated class probability vector of all pixels of entire image.Just can carry out the fusion of MLC and two kinds of results' of MPS probability vector at next step, and obtain final classification results.

(4) data fusion

After obtaining required MLC and MPS probability vector, just can use the consensus theory that two kinds of probability results are merged.The weight that present embodiment has been chosen respectively as MLC is 0.50,0.51,0.52,0.53,0.54,0.55, and corresponding M PS weight is 0.50,0.49,0.48,0.47, merges according to formula (2) respectively in 0.46,0.45 o'clock.The fusion results that finally obtains equally also is a probability vector of being made up of corresponding ownership probability of all categories.Concrete fusion method is as follows:

Present embodiment is chosen two class atural objects, and the first kind is ω ₁, second class is ω ₂, suppose that pixel x probability vector in the MLC classification results is (0.9,0.1), i.e. p ₁(ω ₁| x)=0.9, p ₁(ω ₂| x)=0.1, and the probability vector that the MPS simulation obtains is (0.6,0.4), i.e. p ₂(ω ₁| x)=0.9, p ₂(ω ₂| x)=0.1, getting the MLC weight assignment is λ ₁=0.6, and the MPS weight is λ ₂=0.4, the following formula of having adopted of fusion process $g_j\left(X\right)=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\γ}}_i{p}_i\left({\mathrm{\ω}}_j\right|X),$ Ownership probability for first kind atural object is g ₁(x)=0.60.9+0.40.6=0.78, and be g for the ownership probability of the second class atural object ₂(x)=and 0.60.1+0.40.4=0.22, the probability vector after merging at last is (0.78,0.22).

(5) in step (4), obtain with the result be a probability vector of forming by corresponding ownership probability of all categories, the ownership probability maximum of corresponding a certain classification in the probability vector of each pixel is then assigned to this pixel in corresponding this classification and is gone, thereby obtains final classification results.In all classification results, when the weight of MLC is 0.52, the MPS weight is 0.48 o'clock, and classification results is relatively good, and the result as shown in Figure 6.

By the comparison of error matrix, the overall accuracy of the inventive method is 81.161%, and the overall accuracy of MLC classification is 47.653%, therefore can prove that this algorithm has significantly improved the nicety of grading of remote sensing images.

The above is the preferred embodiments of the present invention only, is not limited to the present invention, and for a person skilled in the art, the present invention can have various changes and variation.Within the spirit and principles in the present invention all, any modification of being done, be equal to replacement, improvement etc., all should be included within the claim scope of the present invention.

Claims (6)

1, a kind of sorting technique that merges spectral information and multiple spot simulation spatial information is characterized in that may further comprise the steps:

Step 1, the pixel of the spectral information of remote sensing images is carried out the MLC classification, obtain each pixel for probability vector of all categories;

Step 2, be that multiple spot simulation MPS selects sample data;

Step 3, according to sample data selected in the step 2,, use the multiple spot analogy method to set up probability model and preserve probability vector according to its condition data number for each pixel in the image;

Step 4, employing data fusion method merge probability vector that obtains in the step 2 and the probability vector in the step 3;

Step 5, in step 4, obtain with the result be a probability vector of forming by corresponding ownership probability of all categories, the ownership probability maximum of corresponding a certain classification in the probability vector of each pixel is then assigned to this pixel in corresponding this classification and is gone, thereby obtains final classification results.

2, a kind of sorting technique that merges spectral information and multiple spot simulation spatial information according to claim 1, it is characterized in that: select the method for sample data to be in the described step 2: the sample data during remote sensing images MLC is classified is as the MPS sample data, or with the sample data of pixel comparatively definite in the remote sensing images as MPS, a threshold value promptly is set, every ownership probability for a certain classification is greater than this threshold value, then with its sample data as MPS.

3, a kind of sorting technique that merges spectral information and multiple spot simulation spatial information according to claim 1, it is characterized in that: the multiple spot analogy method in step 3 is the analogy method that uniformly-spaced sorts, or SNESIM modeling algorithm, or neural network multiple spot analogy method.

4, a kind of sorting technique that merges spectral information and multiple spot simulation spatial information according to claim 3, it is characterized in that: the step of described uniformly-spaced sort method is as follows:

(1) sets up search tree according to selected training image and training template;

(2) ordering N at interval is set;

(3) remaining photofit picture unit is sorted obtain sequence P, top n among the sequence P is waited to simulate pixel in the hunting zone of appointment, search for its condition data according to the training template, and then search and calculate its ownership probability in search tree for each classification, thereby obtain the probability vector of the classification results of each pixel, after each pixel has been simulated, all with it as the remaining condition data of not simulating pixel;

(4) after sequence top n pixel simulation finishes, sequence is sorted again, repeating step (3) is all simulated until all pixels and to be finished.

5, a kind of sorting technique that merges spectral information and multiple spot simulation spatial information according to claim 1, it is characterized in that: the data fusion method in the described step 4 comprises the logarithm suggestion pond or the linear suggestion pond of adopting the Consensus theory, or evidence theory and, but be not limited to the said method of mentioning based on the fusion method of statistical theory.

6, a kind of sorting technique that merges spectral information and multiple spot simulation spatial information according to claim 1, it is characterized in that: described linear suggestion pond data fusion method is: n class atural object is arranged, be designated as Ω={ ω ₁, ω ₂..., ω _n, pixel x probability vector in the MLC classification results is P ₁=(p ₁₁, p ₁₂..., p _1n), and the probability vector that the MPS simulation obtains is P ₂=(p ₂₁, p ₂₂..., p _2n), getting the MLC weight assignment is λ ₁, and the MPS weight is λ ₂, the following formula of having adopted of fusion process $g_j\left(X\right)=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\λ}}_i{p}_i\left({\mathrm{\ω}}_j\right|X),$ Then the ownership probability for first kind atural object is g ₁(x)=λ ₁p ₁₁+ λ ₂p ₂₁, be g for the ownership probability of the second class atural object ₂(x)=λ ₁p ₁₂+ λ ₂p ₂₂..., and be that the ownership probability of n class atural object is g for and for the ownership probability of the second class atural object _n(x)=λ ₁p _1n+ λ ₂p _2n, the probability vector after merging at last is G={g ₁(x), g ₂(x) ..., g _n(x) }.

Images