CN107798286B

CN107798286B - Hyperspectral image evolution classification method based on labeled sample position

Info

Publication number: CN107798286B
Application number: CN201710568564.9A
Authority: CN
Inventors: 尚荣华; 兰雨阳; 焦李成; 刘芳; 马文萍; 王爽; 侯彪; 刘红英; 熊涛
Original assignee: Xidian University
Current assignee: Xidian University
Priority date: 2017-07-13
Filing date: 2017-07-13
Publication date: 2021-01-26
Anticipated expiration: 2037-07-13
Also published as: CN107798286A

Abstract

The invention discloses a hyperspectral image evolutionary classification method based on a marked sample position, which solves the problem of utilization of spatial information in hyperspectral image classification. The method comprises the following specific steps: inputting data, and using PCA to reduce the dimension of the base image; extracting differential morphological characteristics; coding by combining differential morphology and spectral characteristics, dividing a data set, and recording the positions of part of marked samples; initializing representative sample points; starting iteration and determining the maximum iteration times; judging a stopping condition, and if the stopping condition is met, directly classifying the unlabeled samples; if the condition is not met, designing a cross template and an elite retention strategy for evolution, and then selecting a representative point set for iteration again until the condition is met; and after the label-free samples are classified, segmenting the classification result, and further optimizing by referring to the position of the labeled sample point. The invention uses spatial information and an evolutionary algorithm to complete the classification of hyperspectral images, and the search is more based; the classification precision is improved, and the method is applied to hyperspectral image classification.

Description

Hyperspectral image evolution classification method based on labeled sample position

Technical Field

The invention belongs to the technical field of image processing, mainly relates to hyperspectral image classification, and particularly relates to a hyperspectral image evolutionary classification method based on a marked sample position. The application field is hyperspectral image classification.

Background

The hyperspectral image classification problem in remote sensing image processing, due to the fact that the number of wave bands of a hyperspectral image is large, marked samples are difficult to obtain, spectrums are mixed and the like, and many scholars conduct many researches on the problem.

In the paper "semi-collaborative subframe-Based DNA Encoding and Matching Classification for Hyperspectral removal Sensing" (IEEE Transactions on Geoscience and remove Sensing,2016,54(8): 4402-.

According to the method, each pixel point of a hyperspectral image is coded and expressed, and a pixel point set is divided into three parts, namely a marked sample set DT1, a non-marked sample set DT2 and a test sample set DT 3. The original algorithm has fewer samples in the DT1 set, DT2 is used as a candidate set for expanding the sample size, and a virtual class label is used as a real class label in the evolution step.

After entering the step of evolutionary algorithm, the representative samples of each class are initialized, and the innovation point is that the representative individuals of each class are initialized to be a plurality of ones considering the existence of subclasses in each large class. During the evolution operation, the text selects local subspaces and global subspaces according to the class labels of the marked samples and the class labels allocated to the unmarked samples, so as to remove useless information. And then one iteration is completed through the fitness calculation and the random operator. As the iteration progresses, the remaining representative set solves the classification problem better and better.

However, this method also has some disadvantages, such as it does not utilize spatial information, relies too much on spectral characteristics; the evolution algorithm lacks an elite retention strategy and the like.

Disclosure of Invention

In order to overcome the defects of the technology, the invention provides an evolutionary hyperspectral image classification method based on marked sample point position information, which improves the classification precision.

The invention relates to a hyperspectral image evolution classification method based on a marked sample position, which is characterized by comprising the following steps of:

(1) inputting data: inputting hyperspectral image data, namely original data, wherein the hyperspectral image data comprises spectral features; reducing the dimension of the image data by using Principal Component Analysis (PCA) to obtain a plurality of base images;

(2) extracting a difference morphological feature DMP (digital signature), namely a spatial feature, on each base image;

(3) encoding spectral and spatial features, segmenting the data set, recording positional information: combining the differential morphological characteristics with the complete spectral characteristics, coding together, dividing the hyperspectral image dataset into a marked sample set DT1, a non-marked sample set DT2 and a test sample set DT3, and for each type of ground object, randomly selecting part of marked sample points and recording and storing the position information of the marked sample points;

(4) initializing representative point set DNAC: randomly selecting a part of samples in the marked sample set DT1 to form a representative point set DNAC, and finishing initialization;

(5) and (4) entering iteration: setting the current iteration as the 1 st generation, and limiting the maximum iteration number to be 500; optimizing representative sample points of various ground objects;

(6) judging a stopping condition: judging whether the current iteration reaches the maximum times or whether the evolutionary algorithm converges, if not, continuing, otherwise, jumping to the step 9;

(7) constructing a genetic operator: designing replication and cross genetic operators in an evolutionary algorithm, and adding an elite retention strategy;

(8) selecting a strategy: selecting a proper fitness function, performing evolution operation on the current representative point set DNAC by using a genetic operator, selecting only a DNAC set with the maximum fitness function value from the evolution operation, entering next generation optimization, namely adding 1 to the iteration number, and returning to the step 6;

(9) intermediate results were obtained: classifying the unmarked sample set DT2 and the test sample set DT3 respectively by using the representative point set DNAC after the optimization is finished to obtain a classified image;

(10) and (3) final optimization: and (4) carrying out region-based segmentation on the classification result, and combining the classification result with the stored position information of the marked sample point to obtain a final classification result.

The method not only solves the problem of large demand on the number of marked samples in the traditional hyperspectral image classification problem, but also researches the problem of spectral mixing difficulty, applies the spatial characteristics of the image, reduces the operation cost and improves the classification precision.

Compared with the prior art, the invention has the following advantages:

firstly, the semi-supervised learning and evolutionary algorithm are adopted, so that the requirement on the number of marked sample points is not high, and the hyperspectral image classification problem can be well completed at relatively low cost for obtaining the marked samples. And the randomness of the evolutionary algorithm can avoid falling into local optimization, the fitness function strategy of the evolutionary algorithm has selectivity at the same time, and after the elite retention strategy is added, the optimization of each generation can be ensured to be smoothly carried out.

Secondly, the invention simultaneously utilizes the difference morphological characteristics and the original spectrum characteristics of the hyperspectral image base image and the position information of part of marked sample points, thereby better solving the problem of difficult classification caused by spectrum mixing in the hyperspectral image.

Drawings

FIG. 1: a flow diagram of the present invention;

FIG. 2: the invention designs a schematic diagram of evolution operations such as cross template, replication, variation and the like;

FIG. 3: a comparison of the present invention to other prior art on IndianPines datasets, where fig. 3(a) is an actual feature, fig. 3(b) is a plot of the position of labeled sample points recorded in the experiment, fig. 3(c) is the result obtained with an SVM classifier, fig. 3(d) is the result obtained with an OMP algorithm, fig. 3(e) is the result obtained with an SSDNA algorithm, and fig. 3(f) is a plot of the results of the present invention;

FIG. 4: a comparison of the present invention with other prior art on the Pavia University dataset, where fig. 4(a) is the real feature, fig. 4(b) is the plot of the position of the labeled sample points recorded in the experiment, fig. 4(c) is the result obtained with the SVM classifier, fig. 4(d) is the result obtained with the OMP algorithm, fig. 4(e) is the result obtained with the SSDNA algorithm, and fig. 4(f) is the graph of the results of the present invention.

Detailed Description

The invention is described in detail below with reference to fig. 1.

Example 1

The existing hyperspectral image evolution classification method has some defects, such as the fact that space information is not utilized, spectral features are excessively relied on, and an elite retention strategy is lacked in an evolution algorithm. The spatial information can provide a reliable basis for the classification of surrounding samples, and the spatial information is generally extracted by methods such as extracting difference morphological characteristics after a PCA dimension reduction method. The elite retention strategy can ensure the effectiveness of each iteration of the evolutionary algorithm, and an approximately optimal solution can be found after multiple iterations.

Because the hyperspectral images have more wave bands, the cost of manually obtaining the marked samples is high, and the phenomenon of under-fitting can occur when fewer marked samples are used, in order to solve the above difficulties, the invention provides a hyperspectral image evolution classification method based on the positions of the marked samples, which is shown in figure 1 and comprises the following steps:

(1) inputting data: inputting hyperspectral image data, namely original data, wherein the hyperspectral image data comprises spectral features; and using PCA data to reduce the dimension to obtain a plurality of base images, and inputting the base images into the next step.

(2) On each base image, a differential morphological feature DMP, namely a specific spatial feature, is extracted by using a basic expansion erosion operation, in the step, the base image is changed into a morphological differential feature, the spectral feature in the original data is unchanged, and the spectral feature and the differential morphological DMP feature are input into the next encoding spectrum and spatial feature, a data set is segmented, and position information is recorded.

(3) Encoding spectral and spatial features, segmenting the data set, recording positional information: combining the differential morphological features with the complete spectral featuresThe invention uses the existing method to code the spectral features into 3d-2 dimensional vectors, wherein d represents the wave band number of the hyperspectral data, and for the spatial feature of differential morphological features, the invention codes the spectral features into 7N_pcVector of dimensions, where N_pcRepresenting the number of reserved base images, completing coding after splicing with the former base images, and having the length of 7N_pc+3 d-2. The hyperspectral image data set is divided into a marked sample set DT1, an unmarked sample set DT2 and a test sample set DT3, 10 samples of each type are randomly selected to be added into DT1, the remaining 20 percent of samples are added into DT2, and 80 percent of samples are added into DT 3; for each type of ground object, randomly selecting part of marked sample points and recording and storing the position information of the marked sample points. The recorded position information may be represented as a picture, where the spatial information, i.e. the value at the coordinates (x, y) is the class mark of the marked sample, and the values at the remaining non-recorded coordinates are 0, as shown in fig. 3(b) and fig. 4 (b).

(4) Initializing representative point set DNAC: and randomly selecting a part of samples in the marked sample set DT1 to form a representative point set DNAC, and finishing initialization, wherein the samples in the DNAC set are representative sample points of various ground features.

(5) And (4) entering iteration: setting the current iteration as the 1 st generation, and limiting the maximum iteration times; the method is used for optimizing representative sample points of various ground features.

(6) Judging a stopping condition: and judging whether the current iteration reaches the maximum times or whether the evolutionary algorithm is converged, wherein the convergence means that the fitness function value of the reserved representative point set DNAC is not improved in multiple iterations. If the maximum times are not reached or the convergence is not reached, the step 7 is continuously executed, otherwise, the step 9 is skipped.

(7) Constructing a genetic operator: designing a replication and crossover genetic operator in an evolutionary algorithm, adding an elite retention strategy after 10 replications, namely, a first representative point set is not changed, and crossover operation and common mutation operation based on a small crossover template are carried out on all samples in the representative point set from 2 nd to 10 th, so that the effect of each iteration can be ensured not to be degraded, namely, the representative point set DNAC can be always better classified along with the iteration, and meanwhile, the randomness of the evolutionary algorithm for searching an optimal solution is ensured, and the situation of falling into local optimization is avoided. The main innovation of the crossover operator is that a small crossover template is designed, more information related to categories can be utilized, and the evolution operation is more purposeful.

(8) Selecting a strategy: firstly, selecting a proper fitness function, evaluating the fitness of all the 10 current representative point sets DNACs, considering that the representative point sets with higher fitness can be classified more accurately, selecting the DNAC set with the largest fitness function value from the representative point sets, entering next generation optimization, namely, adding 1 to the iteration number, returning to the step 6, and continuously judging whether the maximum iteration number is reached or whether the iteration number is converged.

(9) Intermediate results were obtained: and classifying the samples in the merged set of the unmarked sample set DT2 and the test sample set DT3 by using the optimized representative point set DNAC to obtain an intermediate result graph of the classification.

(10) And (3) final optimization: and performing region-based segmentation on the classification result, wherein the segmentation method specifically comprises the following steps: for any pixel point, if there is a pixel point with the same classification as the pixel point in the 8-neighborhood, the two pixel points are regarded as a region and the same region is allocated to the two pixel points, and all sample points without the allocated region are continuously traversed, so that each region is increased as much as possible to cover more areas until the region attributions of all the sample points are completely determined, and then the segmentation is completed. And combining the segmented classification result graph with the previously stored marked sample point position information graph to obtain a final classification result, and finishing the evolution classification of the hyperspectral image.

The invention relates to an evolutionary algorithm based on sample space information, which aims at the problem of hyperspectral image classification, and the implementation thought is as follows: firstly, carrying out PCA (principal component analysis) dimensionality reduction on hyperspectral image data, selecting a plurality of base images, extracting DMP (differential morphological characteristics), combining the base images with original spectral data, and then carrying out coding representation; and randomly dividing the sample set into three parts: a marked sample set DT1, a no-marked sample set DT2 and a test set DT 3; and recording the position information of the partial sample in DT 1; initializing various representative point sets DNACs, setting replication scale, mutation, crossover probability and a small crossover template, distributing virtual class labels for unmarked samples by using the DNAC set when designing the small crossover template, adding an elite retention strategy, namely only performing evolution operation on the second DNAC set and the subsequent DNAC set, calculating the integral accuracy of each DNAC on the DT1 set after evolution, taking the DNAC with the highest fitness as an fitness function, selecting the DNAC with the highest fitness as the input of the next iteration, adding 1 to the iteration times, performing replication again, crossover based on the small crossover template, ordinary mutation and other evolution operations; and stopping optimization when the iteration number reaches an upper limit or the algorithm converges, and classifying all samples of DT2 and DT3 by taking the representative set at the moment as a classification basis. And finally, carrying out region-based segmentation on the classification result to form a plurality of connected sets, and then optimizing the connected sets together with the position map of the marked sample to obtain a final classification result. The invention solves the problem of utilization of spatial information in hyperspectral image classification.

Example 2

The hyperspectral image evolutionary classification method based on the marked sample position is the same as that in the embodiment 1, the space and spectral characteristics are coded, the data set is segmented, and the position information is recorded in the step 3, and the method specifically comprises the following steps:

3.1. the coding for the differential morphological feature DMP is expressed as follows:

wherein the content of the first and second substances,

code for differential morphological characteristics, x, representing the ith pixel_iRepresenting the gray value of the ith pixel point of a differential image. If an image is m × n in size, it is expanded into mn-dimensional vectors, which are preferentially expanded column by column, i.e., for the columns following the second columnThe method comprises the steps that the whole column of pixel values exists, each column vector is spliced behind the first column, an image is unfolded into an mn-dimensional vector, and therefore the specific position of the ith pixel point in an original image can be determined; and the difference image is an image formed by opening or closing two adjacent structural elements with the radius lambda to obtain the absolute value of the difference at the corresponding position of the image. Because there are two kinds of operations of opening or closing, and the radius of the structural element has four kinds of values in the experiment, after taking the difference, the operation of opening or closing can produce three difference images respectively, six images are produced altogether, and then an original image without opening or closing operation, namely a base image, is added, each pixel point of one base image can be coded into a vector with the length of 7, and the multiple difference morphological characteristics of multiple base images are connected in parallel, so that the difference morphological characteristic of one pixel point can be coded into 7N_pcVector of dimensions, where N_pcIndicating the number of retained base images.

3.2. After the coding is finished, for the sample set of each type of ground feature, randomly selecting 10 marked sample sets DT1, adding the rest 20% of marked sample sets DT2 and 80% of marked sample sets DT3, and completely losing the type mark information in DT2 and DT 3;

3.3. for each type of ground object sample in DT1, a part of the samples is randomly selected, and its position information is kept recorded, that is, an all-zero matrix with the same size as the spatial resolution of the hyperspectral image data is generated first, and when a marked sample is randomly selected, the value at its corresponding coordinate position is changed to its class mark value, so that a sparser matrix can be obtained, and the non-zero position represents the position of the marked sample, see fig. 3(b) and fig. 4 (b).

The coding of spectral features is well established in the prior art and is described in detail in the following examples, so that only the coding of differential morphological features is described here. And the coding of the spatial characteristics added in the step can make up the defects of limited pure spectral characteristic representation force, spectrum mixing and the like, and provides more reference information for the classification of each sample point.

Example 3

The hyperspectral image evolution classification method based on the marked sample position is the same as the embodiment 1-2, and the genetic operator construction in the step 7 specifically includes the steps of adding an elite retention strategy and designing a cross template:

7.1. adding an elite retention strategy: in each iteration, the current representative point set DNAC is copied in a whole for several parts, in this example, after the current generation of iterative optimization is carried out, mul parts are copied from the representative point set, the value in the experiment is 10, the first part is regarded as elite and is reserved, namely, no change is made, and the evolution operation as described below is carried out on each set from the second part.

7.2. Setting a cross template: the evolutionary operation includes crossover and mutation, and since the evolutionary algorithm is an algorithm simulating nature, it is more important to perform crossover operation rather than mutation operation with a very small probability. Before the common crossing, the invention needs to determine a small crossing template, and the specific steps are as follows:

a, firstly generating a complete cross template according to a union set of a marked sample set DT1 and an unmarked sample set DT 2; the samples in the unlabeled sample set DT2 need to have their assigned virtual classmark with the currently optimal representative point set DNAC, i.e., the first DNAC, because there is no explicit classmark. The specific strategy of assigning a classmark to the DT2 set using DNAC can be achieved by measuring the similarity of two individuals as shown in the following formula:

where # represents a logical and operation between two vectors: if the values of the two vectors in a certain dimension are different, the value of the result in the dimension is 0, otherwise, the value is 1; and | | · | | represents the number of nonzero elements in the calculation vector, DNAC_jRepresents the j sample in the DNAC collection, and the DNA_iHere, the ith sample, Label (DNAC, DNA) in the DT2 collection is shown_i) Indicating the virtual classmark assigned to the ith sample in DT2 by DNAC, i.e., selecting the sum D of all samples of DNACNA_iThe class label of the most similar sample is used as the virtual class label.

7.2.b. after the complete cross-template is generated, all samples in the second and subsequent DNACs are required to be cross-template-based cross-operations and common variations as samples to be cross-template. Aiming at each sample to be crossed, a clear class mark is provided; therefore, before the crossover operation, a sample set which is the same as the sample class label to be crossed is selected from the complete crossover template to form a small crossover template corresponding to the class label, and then the common crossover operation is performed on the sample to be crossed in the small crossover template;

the normal crossover, mutation operations are the same as the prior art implementations. The invention relates to a method for crossing a sample to be crossed, which comprises the following steps that (1) common crossing is random crossing, a crossing point is randomly determined in two coding vectors of a sample to be crossed and any other sample, and all numerical values after the crossing point of the two vectors are exchanged, wherein the crossing based on a small crossing template in the invention means that an object crossed by the sample to be crossed is selected from samples with the same class marks as the sample to be crossed; in the ordinary mutation operation, a mutation point can be randomly selected from a code vector, and the value of the mutation point is randomly changed into other allowed values. After the second and later representative sample point sets DNACs are subjected to evolution operation, a plurality of representative sample point sets can be obtained and input into the next step.

The genetic operator is mainly different in that the selection range of the crossed samples is limited in one template, more information related to the categories can be utilized, the templates correspond to the category labels one by one, namely the category labels of the samples in the template are the same, the samples can be considered to contain the information related to the categories, and then the optimal solution can be found in a more guiding mode through the communication between the random crossing and the samples to be crossed. Secondly, an elite retention strategy is added, so that the phenomena of degradation and the like can be avoided in the iteration of the evolutionary algorithm, and the selected result after each iteration is not inferior to the optimal result of the previous generation.

Example 4

The hyperspectral image evolutionary classification method based on the marked sample position is the same as that in the embodiment 1-3, and the final optimization in the step 10 specifically comprises the following steps:

10.1. the intermediate result is denoted as MO and divided into a plurality of connected sets. And make assumptions: the pixels in a connected set all belong to the same class; the method of segmentation is region-based: if there is a pixel with the same classification in the 8 neighborhoods of any pixel, the two can be merged into a region. And continuously and iteratively traversing all the sample points to enable each area to cover more areas as much as possible until the area attribution of all the sample points is completely determined, wherein each area is a connected set.

10.2. Searching each connected set: in the space covered by each connected set, comparing the marked samples of the recorded position information, and judging which type the most of the marked samples belong to, so that all the pixel points in the connected set are classified into which type; if the connected set does not have a marked sample for recording the position information, the classification of the intermediate result MO is not changed;

10.3. and after the connected set is searched, ending the method to obtain a final classification result.

The hyperspectral images have more wave bands, the cost of manually obtaining the marked samples is high, and the phenomenon of under-fitting can occur when fewer marked samples are used. The hyperspectral image evolutionary classification method based on the marked sample positions can effectively solve the problems that the marked sample amount is large in demand, the simple spectral features are mixed and difficult to classify and the like, and improves the classification accuracy by using relatively few marked samples. The invention also considers that a certain type of ground objects of the hyperspectral image are almost distributed in blocks, so that the spatial information of the pixel points can provide important classification information for surrounding pixel points, and after the position information of the marked sample is added, the number of local error division points is adjusted according to the classification conditions around the positions, thereby achieving better classification effect.

Example 5

The hyperspectral image evolutionary classification method based on the marked sample position is the same as that in the embodiments 1 to 4, wherein the optimized intermediate result in the step 10 can also be expressed in the following form:

the invention adds the space characteristic code, and compared with the pure spectral characteristic, the invention can better solve the phenomenon of spectral mixing and better classify. An elite retention strategy and different crossover operators are added in the evolutionary algorithm, so that the classification precision is improved, the limit is added on the basis of random search, and the effectiveness of information can be ensured by a small crossover template; and finally, the space information of the marked samples is also used in the optimization step of the classification map, so that higher classification precision is realized.

A more complete example is given below to further illustrate the invention:

example 6

The invention relates to a Hyperspectral image evolutionary classification method based on labeled sample positions, which is an evolutionary algorithm based on sample space information and aims at the problem of Hyperspectral image (Hyperspectral image) classification, and belongs to the embodiments 1-5. Referring to fig. 1, the implementation includes the following steps:

1. inputting data:

and carrying out PCA (principal component analysis) dimensionality reduction on the input hyperspectral images, and selecting a plurality of base images. In the experiments of the present invention, two data sets were selected: IndianPines and Pavia University datasets. And selecting main components with different proportions for reservation according to the data after the dimensionality reduction. In order to obtain reasonable number of base images after dimension reduction of different data sets, the first 95% of characteristic values of an Indian pines data set are reserved in an experiment; whereas for the Pavia University dataset, the top 99% of the eigenvalues are retained.

2. Extracting a difference morphological feature DMP on each base image, wherein the basic operation of the difference morphological feature is a dilation-erosion operation:

erode(I,B)^λ＝min{f(x+x',y+y')|(x',y')∈D_b}

dilate(I,B)^λ＝max{f(x-x',y-y')|(x',y')∈D_b}

in the formulaI represents a gray image, B represents a structural element with the radius of lambda, the value of each position of the structural element can only take 0 or 1 and can be randomly generated, and min { f (x + x ', y + y') | (x ', y') ∈ D_bAnd max { f (x + x ', y + y') | (x ', y') ∈ D_bDenotes a neighborhood D of radius λ at (x, y) of the original image_bThe smallest or largest of all data at the position corresponding to 1 in (1) is taken as the value at f (x, y), and the anode (I, B)^λAnd dilate (I, B)^λRespectively, the etching and expansion operations with a structural element having a radius lambda. When the radius of the structural element changes, the difference morphological characteristics DMP can be obtained according to the following formula:

wherein^λRepresenting the dilation-erosion operation with radius λ performed on an image, | A-B | representing the difference between two images and taking the absolute value, i.e. taking the absolute value of the difference, DMP, for the corresponding location_OPAnd DMP_CLThe open-close characteristics of the DMP are respectively expressed, the two are directly connected in parallel with the original base image to form seven images, namely, any pixel point in each base image can be coded into a 7-dimensional vector. And MP is defined as:

wherein OP^λ(I) Representing the morphological opening operation, namely firstly using a structural element with the radius of lambda to carry out corrosion firstly and then expansion on the image I; and CL^λ(I) Representing a morphological closing operation, i.e. first using a structuring element of radius λ for the image IThe result of expansion followed by corrosion.

And

the difference morphological characteristics DMP can be obtained by performing a difference operation on the morphological characteristics.

3. Coding spatial and spectral features, segmenting the data set, recording positional information:

firstly, the hyperspectral image is firstly converted into a two-dimensional matrix D ═ x according to the column₁,x₂,...,x_n}. Note x_i∈

R

^d1,2, n, where d is the number of bands of the hyperspectral image and n is the number of pixel points. The encoding method of the data can be expressed as the following four parts.

a. Shape feature, its code length is d: for any pixel point x_iThe spectral values are arranged in ascending order to obtain

Then there is

According to each X_iSelecting three thresholds

Representing the j-th dimension encoding the shape feature of pixel point i, the encoding of the spectral shape feature can be determined according to the following equation

Comprises the following steps:

b. slope characteristics: all the different variations between adjacent spectral values are encoded with an encoding length d-2. If Δ is taken to be the desired spectral value tolerance, then:

in the formula, theta is a constant and is 0.5 in the experiment,

representing the j-th dimension encoding the slope characteristic of pixel point i, the encoding of the spectral slope characteristic can be determined according to

Comprises the following steps:

c. amplitude characteristics: the amplitude signature is an accumulated signature with a code length d. Specifically, for the jth band x of all the pixel points, the following description is provided^jAll the spectral values can form a gray image, all the gray values are accumulated to obtain a maximum value G, and the minimum value is x₁ ^j. The following formula represents the statistical number NO of pixels less than g after accumulating gray values:

NO＝mapping(g)

furthermore, (0.75N) ═ mapping (T)_max)，(0.5N)＝mapping(T_mid)，(0.25N)＝mapping(T_min) Wherein N represents the number of all pixel points,

representing the coding of the amplitude characteristic of the pixel point i in dimension j, the coding of the spectral amplitude characteristic

Comprises the following steps:

d. differential morphological feature coding

Comprises the following steps:

wherein the content of the first and second substances,

code for differential morphological characteristics, x, representing the ith pixel_iRepresenting the gray value of the ith pixel point of a differential image, if the size of one image is mxn, the image can be displayed as an mn-dimensional vector, the image is preferentially displayed according to columns, namely, for all the pixel values of the columns behind the second column, each column vector is integrally spliced behind the first column, the image is displayed as an mn-dimensional vector, and thus the specific position of the ith pixel point in the original image can be determined. The difference image refers to an image generated by obtaining an absolute value of a difference between corresponding positions of the image after performing an opening or closing operation using two structural elements adjacent to each other with the radius λ. Because there are two kinds of operations of opening or closing, and the radius of the structural element has four kinds of values in the experiment, after taking the difference, the operation of opening or closing can produce three images separately, produce six images altogether, and then an original image without opening and closing operation, namely the base image, each pixel point of a base image can be coded as a length of 7 vectors, connect multiple differential morphological characteristics of multiple base images in parallel, can code the differential morphological characteristic of a pixel point as a length of 7N_pcVector of dimensions, where N_pcIndicating the number of retained base images.

By combining the above 4 codes, each sample code can be represented as a 7N code_PCA +3d-2 dimensional vector, where N_pcIndicates the number of base images and d indicates the number of bands. These sample points are divided into three sets: a marked sample set DT1, an unmarked sample set DT2 and a test set DT 3. It is composed ofFor each sample type, n1 samples were included in DT1, set to 10 in the experiment, with 20% of the remaining samples added to DT2 and 80% added to DT 3. Finally the sample recording position in section DT1 was randomly selected.

4. Initializing representative point set DNAC:

by taking reference to the previous algorithm, each class of representative points is set to be nc, nc samples of all classes are randomly selected in DT1 during initialization to form a representative point set, the representative point set is marked as DNAC, initialization is completed, and nc is taken as 3 in both data sets in an experiment. Thus, for a dataset with nclass surface feature classes, the size of the DNAC collection is nc × nclass.

5. And (4) entering iteration:

setting the current iteration as the 1 st generation, and limiting the maximum iteration times, wherein the maximum iteration times in the example are 500; the method is used for optimizing representative sample points of various ground features.

In the invention, the maximum iteration number is 500, and the maximum iteration number can also be increased to another larger value, such as 1000, 1500 and the like, according to the change of the fitness function value in the optimization; and when the optimization is better, the optimization can be ended in advance, the maximum iteration number can be selected to be smaller, and the convergence of the algorithm is ensured.

6. Judging a stopping condition:

and judging whether the current iteration reaches the maximum times or whether the evolutionary algorithm reaches convergence, wherein the convergence represents that the fitness function value of the DNAC representative point set obtained by the evolutionary operation is not increased after more iterations. If not, continuing to execute, otherwise jumping to step 9;

7. constructing a genetic operator:

in this step, the representative point set is first copied as a whole, and cross mutation operators are performed from the second and subsequent representative point sets DNAC, and the first representative point set is combined as an elite reservation. The crossover operator is not completely random, but is performed according to a certain template, and a complete crossover template is generated according to the union of the marked sample set DT1 and the unmarked sample set DT 2. For the samples in the unlabeled sample set DT2, because there is no explicit classmark, the currently optimal representative point set DNAC, i.e. the first DNAC, needs to be used as a reference for the virtual classmark assigned to the samples; and the virtual classmark is obtained by assigning the classmarks to all the individuals in the DT2 set by using the current DNAC set. The similarity of the two codes can be measured to assign a classmark, as shown in the following equation:

where # represents a logical and operation between two vector corresponding locations: the two are the same, and the result is 1; the two are different, the result is 0, and | · | | represents the number of nonzero elements in the calculation vector, DNAC_jRepresents the j sample in the DNAC collection, and the DNA_iRepresents the ith sample in the DT2 collection, i.e., the sample to be assigned the virtual object class, Label (DNAC, DNA)_i) The virtual classmark assigned to the ith sample in DT2 by DNAC is shown. After a complete cross template is generated, all samples in the second and subsequent DNACs are used as samples to be crossed, and a specific class mark exists for each sample to be crossed; therefore, for each crossover operation, a sample set which is the same as the class label of the sample to be crossed needs to be selected from the complete crossover templates to form a small crossover template corresponding to the class label, and the small crossover template and each sample to be crossed are used for carrying out common crossover operation; the common crossing operation is to randomly select a crossing point from two coding vectors and to exchange the values of the two vectors before and after the crossing point; and the ordinary mutation operation means that for any vector, a mutation point is randomly selected, and the value of the mutation point is randomly changed into any one of other possible values. See figure 2.

FIG. 2 is a detailed illustration of the cross template and replication scale, mutation crossover operators, etc. of the design. Each column in fig. 2 represents a sample, and the last row is the class label for each sample, where the yellow class label represents the real class label and the green class label represents the virtual class label, which is the class label assigned to the DT2 set using DNAC. In the case of normal cross variation, it can be seen that data before and after the cross point is completely exchanged, and the cross point (blue) is randomly varied to another allowable value. In the figure nc represents the number of representatives of each class of selection, which is set to 3 in this example, nclass represents the number of surface feature classes of the data set, mul represents the replication scale, which is set to 10 in the experiment, N1 is the size of the labeled sample set, and N is the number of non-background samples.

8. Selecting a strategy:

to select a suitable fitness function first, the overall classification accuracy OA for the DT1 data set using a certain DNAC set may be used as the fitness function value, that is:

wherein DNAC_iDenotes the i-th DNAC, mul denotes the replication scale, set to 10 in the experiment, and OA_iThe overall accuracy of assigning class labels to the marked sample set DT1 using the ith DNAC is shown. After the fitness function is determined, the merits of all DNAC sets can be evaluated, with higher fitness function values giving better representative point sets. And then selecting a DNAC set with the maximum fitness function value from a plurality of DNAC sets obtained by replication and evolution, entering next generation optimization, namely adding 1 to the iteration number, and returning to the step 6.

9. Intermediate results were obtained:

after the evolutionary algorithm converges or reaches the maximum iteration number, classifying all samples in DT2 and DT3 by using the DNAC set stored at the moment, wherein the classification criterion is the same as the formula in step 7, but the DNAC at the moment_jAll samples in DT2 are not represented any more, but all samples in the set of the union of DT2 and DT3, and after the preliminary classification is finished, an intermediate classification result MO can be obtained.

10. And (3) final optimization:

based on the intermediate classification result, we can further optimize according to the position information of the marked sample recorded in step 3, which includes the following specific steps:

10.1. the intermediate result MO is partitioned, thus forming a plurality of connected sets. And make assumptions: the pixels in a connected set all belong to the same class; the method of segmentation is region-based: if there are pixels in the 8 neighborhoods of a certain pixel with the same classification as the certain pixel, the two can be regarded as being in one region. Continuously iterating and traversing all the sample points to enable each area to cover more areas as far as possible until all the area attributions of all the sample points are determined;

10.2. searching each connected set: in the area covered by each connected set, marking which type the sample is mostly, and then classifying all pixel points in the connected set into which type; if the marked sample does not exist in the connected set, the classification of the intermediate result MO is not changed;

10.3. and after the connected set is searched, finishing the method to obtain a final result.

The spatial information can provide a reliable basis for the classification of surrounding samples, and the method can extract the DMP (differential morphological feature) as the spatial information by performing morphological operation on a plurality of base images obtained after the dimensionality reduction of PCA (principal component analysis). By utilizing the selectivity and the randomness of the evolutionary algorithm, a better representative point set DNAC for classifying unmarked samples can be selected in a heuristic manner, meanwhile, the randomness of each iteration can also avoid the situation of falling into the local optimum condition, and the essence retention strategy is added into the evolutionary algorithm, so that the effectiveness of each iteration can be ensured, and the degradation phenomenon can not occur. Meanwhile, the invention adopts semi-supervised learning, so that the requirement on the number of marked sample points is not high, and the problem can be solved at relatively low cost for obtaining the marked samples. In the last step, only a small number of marked samples and their position information are used for further optimization, and higher classification precision is achieved. Due to the advantages of the points, the hyperspectral image classification method can better solve the problem of hyperspectral image classification and improve the classification precision.

Example 7

The hyperspectral image evolutionary classification method based on the marked sample positions is the same as the embodiments 1-6, and the invention is further described with reference to the attached figures 3 and 4.

Simulation experiment conditions are as follows:

the hardware platform adopted by the simulation experiment of the invention is as follows: a processor Inter Core i7, a master frequency of 2.60GHz and a memory of 8 GB;

the software platform is as follows: windows 10 family chinese version 64-bit operating system, Matlab R2016B.

The experimental contents are as follows:

the simulation experiment of the invention adopts common hyperspectral image data sets Indian pipes and Pavia University, the real ground object diagrams of which are respectively shown in the attached drawings 3(a) and 4(a), and the non-background points in the real ground object diagrams are classified. In which the Indian Pines data set was collected by AVIRIS sensors in 1992 in the northwest part of Indian pine, USA, and the scene contains two-thirds and one-third of natural vegetation such as agriculture and forest. The image size of the image after removing the water absorption wave band comprises 200 wave bands and 145 multiplied by 145, and the total number of the images comprises 16 land features. And Pavia University is the Pavia University area in italy collected by the rosss sensor in 2003. The total spectrum range is 115 bands from 0.43 μm to 0.86 μm, after 12 noisy bands are removed, 103 bands are included, the image size is 610 × 340, and 9 real ground object categories are included in total. The experimental contents are that the non-background sample points in the two data sets are classified by utilizing various coding information of the sample points and the position information of the marked sample points, and the classification effects of different methods are compared, so that the invention can achieve better classification effect.

And (3) analyzing an experimental result:

fig. 3(a) is a true terrain map of Indian Pines dataset, while fig. 3(b) is a distribution plot of recorded labeled samples, fig. 3(c) is a classification result for data using support vector machine algorithm (SVM), fig. 3(d) is a classification result for data using orthogonal matching pursuit algorithm (OMP), fig. 3(e) is a classification result for data using semi-supervised and subspace-based DNA matching algorithm (SSDNA), and fig. 3(f) is a classification result for data according to the present invention. Compared with the prior methods such as SVM, OMP, and SSDNA, the method provided by the invention can be more approximate to the real situation in principle from the visual effect of classification. Although the

parts

1,2 and 3 in fig. 3(f) all have the disorder points, which are caused by insufficient position information of the labeled sample, the disorder points at the corresponding positions in other comparison methods are more, and it is seen that the classification accuracy of the present invention at the

positions

1,2 and 3 is higher than that of other methods.

Fig. 4(a) is a real terrain map of a Pavia University dataset, while fig. 4(b) is a distribution plot of recorded labeled samples, fig. 4(c) is a classification result for data using a support vector machine algorithm (SVM), fig. 4(d) is a classification result for data using an orthogonal matching pursuit algorithm (OMP), fig. 4(e) is a classification result for data using a semi-supervised and subspace-based DNA matching algorithm (SSDNA), and fig. 4(f) is a classification result for the data according to the present invention. It can be seen from fig. 4(f) that in the large connected region, part 1 and part 2 of the figure, the labeled sample labels are used for reference, so the invention can be classified almost completely correctly, and the corresponding positions of other algorithms, fig. 4(c),4(d) and 4(e), have different degrees of disorder points, compared with the existing method, the invention has fewer wrong points, and especially the completely correct classification can be realized in the large connected region 1 and part 2.

In short, the invention discloses a hyperspectral image evolutionary classification method based on a marked sample position. The method solves the problem of utilization of spatial information in hyperspectral image classification, and comprises the following specific steps: inputting data, and using PCA to reduce the dimension of the base image; extracting differential morphological characteristics; coding by combining differential morphology and spectral characteristics, dividing a data set, and recording the positions of part of marked samples; initializing representative sample points; starting iteration and determining the maximum iteration times; judging a stopping condition, and if the stopping condition is met, directly classifying the unlabeled samples; if the condition is not met, designing a cross template and an elite retention strategy for evolution, and then selecting a representative point set for iteration again until the condition is met; and after the label-free samples are classified, segmenting the classification result, and further optimizing by referring to the position of the labeled sample point. The invention uses spatial information and an evolutionary algorithm to complete the classification of the hyperspectral images. The search is more based; the classification precision is improved. The method is applied to hyperspectral image classification.

Claims

1. A hyperspectral image evolution classification method based on a labeled sample position is characterized by comprising the following steps:

(1) inputting data: inputting hyperspectral image data, namely original data, wherein the hyperspectral image data comprises spectral features; performing Principal Component Analysis (PCA) on the image data to reduce the dimension to obtain a plurality of base images;

(6) judging a stopping condition: judging whether the current iteration reaches the maximum times or whether the evolutionary algorithm converges, if not, continuing, otherwise, jumping to the step (9);

(8) selecting a strategy: selecting a proper fitness function, taking the classification total accuracy OA of a certain DNAC set to the DT1 data set as a fitness function value, evaluating the fitness of all the representative point sets DNAC, considering that the representative point set with higher fitness can be classified more accurately, carrying out evolution operation on the representative point set DNAC by using a genetic operator, selecting only the DNAC set with the maximum fitness function value from the representative point sets, entering next generation optimization, namely adding 1 to the number of iterations, and returning to the step 6;

2. The hyperspectral image evolution classification method based on the labeled sample position as claimed in claim 1 is characterized in that the encoding spectrum and the spatial feature in the step (3) are used for segmenting a data set and recording position information, and comprises the following steps:

wherein the content of the first and second substances,

code for differential morphological characteristics, x, representing the ith pixel_iRepresenting the gray value of the ith pixel point of a differential image;

3.2. after the coding is finished, for the sample set of each type of ground feature, randomly selecting 10 marked sample sets DT1, adding 20% of the rest samples into a non-marked sample set DT2, and adding 80% into a test set DT 3;

3.3. for each type of feature in the marked sample set DT1, randomly selecting a portion of the samples keeps recording its location information.

3. The hyperspectral image evolutionary classification method based on labeled sample positions as claimed in claim 1 is characterized in that: the genetic operator construction in the step (7) is specifically to add an elite retention strategy and design a cross template for cross operation, and comprises the following steps:

7.1. adding an elite retention strategy: after the iterative optimization of the current generation is carried out, firstly, copying the representative point set by a plurality of shares, reserving the first share as elite, and carrying out evolution operation on each set from the second share;

7.2. designing a cross template to carry out cross operation: the evolution operation comprises crossing and mutation, wherein the crossing template is not completely random when crossing, but a complete crossing template is firstly generated according to a marked sample set DT1 and an unmarked sample set DT2, and for the samples of the unmarked sample set DT2, a virtual class mark is allocated to the samples by using the currently optimal representative point set DNAC; merging the DT2 set and the DT1 set to generate a complete cross template; specifically, when the crossover operation is performed on any determined sample to be crossed, namely any sample in the second and subsequent DNACs, a sample set which is the same as the sample type label to be crossed needs to be specifically selected from the complete crossover template to form a small crossover template corresponding to the sample type label, the evolution operation can be completed by randomly selecting a sample from the small crossover template and performing common crossover and mutation operations on the sample, and after a plurality of evolved DNACs are obtained, the step is finished, and the next step of the algorithm is continued.

4. The hyperspectral image evolutionary classification method based on the labeled sample position according to claim 1 is characterized in that the final optimization in the step (10) specifically comprises the following steps:

10.1. and dividing the intermediate result MO to form a plurality of connected sets: the pixels in a connected set all belong to the same class; the method of segmentation is region-based: if there is a pixel point with the same classification in the 8 neighborhoods of a certain pixel point, the two are regarded as a region, and all sample points are continuously iterated and traversed, so that each region covers more area as far as possible until the region attributions of all the sample points are completely determined;

10.2. searching each connected set: in the area covered by each connected set, the positions of the recorded marked samples are inspected, and all pixel points in the connected set are classified into the type of the samples according to the type of the samples in which most of the samples are; if the marked sample of the recording position does not exist in the connected set, the classification of the intermediate result MO is not changed;

10.3. and after the connected set is searched, finishing the algorithm to obtain a final classification result.