CN106529508A - Local and non-local multi-feature semantics-based hyperspectral image classification method - Google Patents

Abstract

The invention discloses a semantic hyperspectral image classification method based on local and non-local multi-features. It mainly solves the problems of low accuracy rate, poor robustness and weak spatial consistency in hyperspectral image classification. The steps include: inputting an image, extracting various features of the image; dividing the data set into a training set and a test set; probabilistic support vector machine mapping various features of all samples into corresponding semantic representations; constructing local and non-local neighbor sets; Construct a denoising Markov field model, perform semantic fusion and denoising processing; iteratively optimize the semantic representation; use the semantic representation to obtain the categories of all samples, and complete the accurate classification of hyperspectral images. The present invention adopts multi-feature fusion, and fully excavates and utilizes the spatial information existing in the image. In the case of small samples, it obtains very high classification accuracy, and has good robustness and spatial consistency. It is used for Military detection, map drawing, vegetation survey, mineral detection, etc.

Description

Semantic hyperspectral image classification method based on local and non-local multi-features

technical field

The invention belongs to the technical field of image processing, relates to machine learning and hyperspectral image processing, and specifically relates to a classification method for hyperspectral image based on local and non-local multi-feature semantics, which is used to classify and recognize different ground objects in the hyperspectral image.

Background technique

Hyperspectral remote sensing technology has gradually become a research hotspot in the field of earth observation in the past few decades. Hyperspectral remote sensing technology uses imaging spectrometers to image surface objects simultaneously with tens or hundreds of bands with nanoscale spectral resolution, and can obtain continuous spectral information of surface objects, and realize spatial information, radiation information, and spectral information of surface objects. Acquired synchronously, with the feature of "integration of graphs and graphs". Since different ground objects have different reflected wave information, the high resolution of the spectrum in hyperspectral images provides extremely important discriminative information for distinguishing different ground objects or targets. The ground object classification of hyperspectral images has good application prospects in geological surveys, crop disaster monitoring, air pollution, and military target strikes.

Hyperspectral remote sensing image classification is the process of classifying each pixel in hyperspectral remote sensing images into various categories. Since the images acquired by hyperspectral remote sensing technology contain rich spatial, radiometric and spectral triple information, they provide a large amount of discriminative information for classification tasks, but there are still huge challenges and difficulties. First of all, the large amount of data, at least dozens of bands, leads to high computational complexity, and also brings challenges to data storage, transmission and display; second, the dimension is too high, redundant data and some noise exist, which will cause Reduce the classification accuracy; finally, there are many bands, and the correlation between the bands is high, resulting in an increase in the number of training samples required. If the training samples are insufficient, common problems such as underfitting in machine learning will occur, resulting in a serious decline in subsequent classification accuracy. .

Among the traditional hyperspectral classification methods, classification methods such as support vector machines based on spectral information can solve the above difficulties to a certain extent, but their classification accuracy is low, and the spatial consistency of the classification result map is poor, which cannot meet the application requirements.

In recent years, at the feature level, researchers have proposed a large number of classification methods based on multi-feature information, and the classification accuracy has been improved to a certain extent. When these methods combine multi-feature information, there are mainly two ways: the first is feature The first level of fusion directly connects multiple feature vectors in series as the input of the classifier; the second is decision-level fusion, after inputting multiple feature vectors into the classifier, the classification results are fused. There is a certain amount of information loss in the process of the two fusion methods of multi-feature information, resulting in that the multi-feature information is not fully utilized effectively.

At the model level, due to the importance of spatial information in hyperspectral images, a large number of classification methods that add spatial constraints have been proposed, such as joint sparse representation model, support vector machine based on fusion kernel, Markov field model, etc. These methods use The local spatial information in the hyperspectral image greatly improves the classification accuracy. However, on the one hand, these methods are relatively rough in the processing of local spatial information, and most of them use traditional square windows as the local area of pixels, which affects the spatial consistency of the classification result map and hinders the further improvement of classification accuracy. On the other hand, a large amount of non-local spatial information is redundant in hyperspectral images, and these methods have not effectively utilized such information, making the upper limit of classification accuracy and robustness lower.

Therefore, how to extract a variety of useful features from high-dimensional redundant data, reasonably combine multiple features, and effectively use rich spatial information (local and non-local) and a small amount of precious category information to improve The accuracy and robustness of the classification results of hyperspectral images in the case of small samples and the spatial consistency of the classification results are a technical problem to be solved.

Contents of the invention

Aiming at the problems of low accuracy of classification results, poor robustness and weak spatial consistency of classification result graphs existing in the prior art, the present invention proposes a method that reasonably combines multiple feature information and can effectively utilize local Local and non-local multi-feature semantic hyperspectral image classification method based on local and non-local spatial information and related category information.

The present invention is a kind of hyperspectral image classification method based on local and non-local multi-feature semantics, it is characterized in that, comprises the following steps:

(1) Input the image and extract various features of the image: input hyperspectral image M is the total number of all pixels in the hyperspectral image, h _q is a column vector, representing the vector formed by the reflection value of pixel point q in each band; various features are extracted from the hyperspectral image, including: Original spectral features, Gabor texture features, and differential morphological features; the hyperspectral image contains c-type pixels, of which there are N marked pixels and m unmarked pixels. Each pixel of the image is a sample, and each A sample is composed of V feature vectors, which respectively represent the expression of the sample under different feature descriptors, and V is the number of feature categories;

(2) The hyperspectral image data set is divided into a training set and a test set: use N marked pixels as training samples to form a training set Its corresponding category label set is Use m unmarked pixels as test samples to form a test set Among them, x _i represents the i-th sample of the training set, y _j represents the j-th sample of the test set, l _i is the category label to which the i-th training sample belongs, D _v represents the dimension of the v-th class feature, and R represents field of real numbers;

(3) Use the probabilistic support vector machine (SVM) to map the various features of all samples into corresponding semantic representations: respectively use the training set The V feature vectors of all samples in and their corresponding class label sets Construct V probabilistic support vector machine classifiers. The kernel function of the classifier is a radial Gaussian kernel. The kernel parameters and penalty parameters are obtained by multiple cross-validation; the test set The V eigenvectors of all samples in are respectively input into the constructed V corresponding classifiers, and each test sample y _j ,j=1,2,...,m belongs to each category under different feature descriptor representations. probability, as a semantic representation for each test sample For each sample x _i in the training set, i=1,2,...,N, the probability of it belonging to its own category l _i is 1, while the probability of belonging to other categories is 0, and various semantic representations corresponding to various features are obtained in The first line of l _i is 1, and the other lines are 0; thus, various semantic representations of all samples in the hyperspectral image can be obtained

(4) Construct the local and non-local neighbor sets of all samples in the test set; for each sample y _j ,j=1,2,...,m in the test set, construct its local adaptive neighbor set B _j and non-local similarity structure Neighbor set C _j ;

(5) Construct a denoising Markov field model, perform fusion of multiple semantic representations of test samples and denoise processing of semantic representations; perform the following operations on each test sample y _j , j=1,2,...,m respectively , representing the semantics corresponding to y _j Semantic representation of all samples in locally adaptive neighbor set B _j and the semantic representation of all samples in the neighbor set _Cj of non-local similar structure Both are input into the local energy function, and the energy function is minimized to obtain the first-order noise reduction semantic representation of the test sample y _j At the same time, keep the semantic representation of the samples in the training set unchanged, and obtain the first-order denoising semantic representation of all samples in the hyperspectral image

(6) First-order noise reduction semantic representation for all samples of hyperspectral images Further iterative optimization; set the maximum number of iterations T _max , t is the current iteration algebra, and perform the following operations on each test sample: the test sample y _j , and the t-th order noise reduction of all samples in the set B _j and set C _j Semantic representation As the input of the local energy function of the denoising Markov field, the energy function is minimized to obtain the (t+1)th order semantic representation of the test sample y _j At the same time, keep the semantic representation of the samples in the training set unchanged, and then obtain the (t+1)th order denoising semantic representation of all samples in the hyperspectral image Repeat the iterative process until t=T _max -1 stops, and get the T _max order noise reduction semantic representation of all samples of the hyperspectral image, that is, the final semantic representation

(7) Utilize the final semantic representation Obtain the categories of all samples in the test set; for each sample y _j ,j=1,2,...,m in the test set, the final semantic representation is That is, the column vector composed of the probability that the test sample y _j belongs to each category, select the label at the position of the maximum element in the vector as the category of y _j So as to get the test set category prediction set Complete the classification task of this hyperspectral image.

Based on multiple features in different feature spaces of hyperspectral images, the present invention maps to the same semantic space through a weak classifier, and then uses the Markov field model to perform semantic fusion and noise reduction to obtain semantic information containing a variety of information and a small amount of noise Indicates that on this basis, the classification and recognition of different ground objects in the hyperspectral image are carried out.

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

1. Since the present invention fully excavates the interrelationships between pixels, it finds the local adaptive neighbors and non-local similar structure neighbors of each pixel. At the same time, the category information of a small number of labeled samples will also be And spread to its local and non-local neighbors, so that the hyperspectral image can still achieve high classification accuracy with a small number of labeled samples.

2. The present invention improves the Markov field model with multi-feature directions and non-local directions, and replaces the traditional discrete category similarity and difference regular term with the geodesic distance regular term between semantic vectors, so that the Markov field The model is more suitable for processing hyperspectral image classification problems, solves the problems of over-smoothing caused by traditional Markov models, and improves the spatial consistency of classification results.

3. The noise reduction Markov field model proposed in the present invention can be solved by a simple gradient descent method, and the computational complexity is lower than that of the traditional Markov field model. It can well avoid falling into the local optimal solution, improve the robustness of the model, and thus improve the robustness of the classification results.

Comparative experiments show that the invention significantly improves the classification accuracy and robustness of hyperspectral remote sensing images, and makes the classification result map have good spatial consistency.

Description of drawings

Fig. 1 is a schematic flow sheet of the present invention;

Fig. 2 is the Indian Pine data set that the simulation of the present invention adopts, and wherein, Fig. 2 a is the one-dimensional grayscale image that the Indian Pine data set obtains by principal component analysis (PCA), and Fig. 2 b is the real feature category label map of the Indian Pine data set , each color corresponds to a different feature type;

Fig. 3 is the contrast of the classification result figure on the Indian Pine data set of the present invention and existing method, and wherein, Fig. 3a-3f corresponds to based on three kinds of existing classification methods respectively: the support vector machine (SVM+CK) of fusion kernel , support vector machine combined with Markov field (SVM+MRF), joint sparse representation (SOMP); and two simplified versions of the method of the present invention, and the classification result map of the Indian Pine data set obtained by the method proposed by the present invention.

detailed description

The present invention will be described in detail below in conjunction with the accompanying drawings.

Example 1:

For the classification of ground objects in hyperspectral images, most of the current existing methods have the problems of unsatisfactory classification accuracy, poor robustness of classification results, and weak spatial consistency of classification result maps. The present invention combines various features in The fusion technology of semantic space and the local and non-local space constraint method mainly propose a semantic hyperspectral image classification method based on local and non-local multi-features to solve various problems existing in the existing methods.

The present invention is a kind of semantic hyperspectral image classification method based on local and non-local multi-features, referring to Fig. 1, comprising the following steps:

(1) Input an image and extract various features of the image.

Commonly used hyperspectral image data include the Indian Pine dataset and the Salinas dataset obtained by the airborne visible/infrared imaging spectrometer AVIRIS of NASA's Jet Propulsion Laboratory, and the University of Pavia dataset obtained by NASA's ROSIS spectrometer.

input hyperspectral image M is the total number of all pixels in the hyperspectral image, h _q is a column vector, representing the reflection value of each band of the pixel point q, which is the original spectral feature of the pixel point; Features include: original spectral features, Gabor texture features, and differential morphology features (DMP). These three features respectively reflect the spectrum, texture, and shape information of hyperspectral images. The hyperspectral image contains c-type pixels, among which there are N marked pixels and m unmarked pixels. Each pixel of the image is a sample, and each sample is composed of V feature vectors, representing the sample In the expression under different feature descriptors, V is the number of feature categories.

The number of feature categories used in this embodiment is 3, so the number V of feature categories described here and thereafter should be equal to 3 unless otherwise specified.

The N marked pixels mentioned in this embodiment are pixels selected in a moderate proportion from each type of pixels in the hyperspectral image, and all the remaining m pixels are regarded as unmarked pixels.

(2) The hyperspectral image data set is divided into a training set and a test set, and N marked pixels are used as training samples to form a training set Its corresponding category label set is Use m unmarked pixels as test samples to form a test set Among them, x _i represents the i-th labeled training sample of the labeled training set, and y _j represents the j-th unlabeled test sample of the test set. Each sample in the hyperspectral image is represented by V column vectors, each column vector represents a feature, l _i is the category label to which the i-th labeled training sample belongs, and D _v indicates the corresponding feature of the vth class of the sample dimension, and R represents the field of real numbers. Corresponding to step (1), N is the number of marked pixels, and m is the number of unmarked pixels. Here, N is the total number of marked training samples, and m is the total number of test samples.

(3) Use the probabilistic support vector machine (SVM) to map the various features of all samples to the corresponding semantic space, specifically to map to the corresponding semantic representation. Using the training set separately The V feature vectors of all samples in and their corresponding class label sets Construct V classifiers, that is, probabilistic support vector machine (SVM). The kernel function of the probabilistic support vector machine (SVM) is a radial Gaussian kernel, and the kernel parameters and penalty parameters are obtained by multiple cross-validation. the test set The V eigenvectors of all samples in are respectively input into the constructed V corresponding probabilistic support vector machine (SVM) classifiers, and each test sample y _j ,j=1,2, ..., the probability of m belonging to each class, as a semantic representation for each test sample For each sample x _i in the training set, i=1,2,...,N, the probability of belonging to its own category l _i is 1, while the probability of belonging to other categories is 0, and the various semantic expressions corresponding to its various features are expressed as in The first line of l _i is 1, and the other lines are 0. The semantic representation of the samples in the training set is a 0-1 encoded vector, and various semantic representations of all samples in the hyperspectral image are obtained Note that since the category labels of all samples in the training set are known, their corresponding semantic representations are completely correct semantic representations.

In this embodiment, the probabilistic support vector machine (SVM) is used to complete the mapping of sample features to sample semantic representation, because for the hyperspectral image object classification problem, the probabilistic support vector machine (SVM) has good robustness and comparative Strong classification ability, and it is more general as a benchmark classifier in the field of hyperspectral processing. The present invention can also use other classifiers that can obtain category probabilities to complete this step, such as multinomial logistic regression, random forest and their variant classifiers, etc., to replace the probabilistic support vector machine (SVM).

(4) In order to extract the local spatial information and non-local spatial information of all samples in the test set, local and non-local spatial constraints are introduced, and the local and non-local neighbor sets of all samples in the test set are constructed. For each sample y _j in the test set, j=1,2,...,m, construct its local adaptive neighbor set B _j and non-local similar structure neighbor set C _j , and obtain the local adaptive Neighbor Sets and Nonlocal Similarity Structure Neighbor Sets.

(5) Construct a denoising Markov field model to perform fusion of various semantic representations of test samples and semantic denoising processing, in which denoising processing is realized by introducing local and non-local spatial constraints. Combining multiple semantic representations of all samples obtained in step (3) And in step (4), the local adaptive neighbor set and the non-local similar structure neighbor set of all samples in the test set are obtained, and the multiple semantic representations of all samples in the test set and the multiple semantic representations of their local and non-local neighbors are all input In the denoising Markov field model, the joint probability of the Markov field is maximized, that is, the global energy of the Markov field is minimized. In order to minimize the global energy, the iterative conditional model method (ICM) is used to solve the problem, and the This global minimization problem is transformed into multiple local minimization problems, that is, to minimize the energy of the potential group where each sample is located. The potential group is composed of the test sample itself and all corresponding neighbor samples. For each test sample y _j , j=1,2,...,m perform the following operations respectively, and express the various semantics of the test sample y _j Multiple Semantic Representations for All Samples in Local Adaptive Neighbor Set _Bj and multiple semantic representations of all samples in the neighbor set C _j of non-locally similar structures Input it into the local energy function of the denoised Markov field, minimize the function, so as to obtain the semantic representation of the test sample y _j after the first denoising In this way, the energy of the potential group where all test samples are located is minimized, and the first-order denoising semantic representation of all test samples is obtained. In order to make the semantic representation of the samples in the training set, that is, the completely correct semantic representation, continue to positively affect its local and non-local neighbors through the denoising Markov field, so keep the semantic representation of the samples in the training set unchanged, combine all The first-order denoising semantic representation of the test sample is obtained to obtain the first-order denoising semantic representation of all samples of the hyperspectral image

(6) Since the iterative conditional model method (ICM) used in step (5) does not guarantee the minimization of the global energy after a round of local minimum energy calculations, the first-order denoising semantics for all samples of the hyperspectral image express That is, the semantic representation result obtained in step (5) is further iteratively optimized. Referring to Figure 1, set the maximum number of iterations T _max , t is the current iteration algebra, and express the t-th order noise reduction semantics of the test sample y _j The t-th order denoising semantic representation of all samples in the local adaptive neighbor set _Bj and the t-th order denoising semantic representation of all samples in the neighbor set C _j of non-local similar structure As the input of the local energy function of the denoising Markov field, the energy function is also minimized to obtain the (t+1)th order semantic representation of the test sample y _j Thus, the (t+1)th order denoising semantic representation of all test samples of the hyperspectral image is obtained. For all samples in the training set, the processing method is the same as that described in step (5), and the (t+1)th order denoising semantic representation of all samples in the hyperspectral image is obtained Repeat the iterative process, which is this step, until t=T _max -1, stop, and get the T _max order noise reduction semantic representation of all samples of the hyperspectral image, which is the final semantic representation Note that the denoising Markov field local energy function mentioned in this step is different from the denoising Markov field local energy function mentioned in step (5), because the local energy in step (5) The function needs to complete the function of multi-semantic fusion, while the local energy function in this step only needs to complete iterative optimization. For specific differences, refer to the first-order local energy function and the (t+1)th-order local energy function in Embodiment 4.

In the present invention, the parameter T _max needs to be set, which is the number of iterations mentioned in step (6). In general, if T _max is small, the result of convergence cannot be obtained, and if T _max is large, the calculation will be complicated. The higher the degree, the more unnecessary calculations are generated. The setting method of this parameter is described as follows. It is judged by whether the calculation result is stable, and a small number of forecast samples (5% to 10%) are randomly selected. According to the (t+1)th order of this part of forecast samples If the difference between the predicted category set obtained from the semantic representation and the predicted category set obtained from the t-order semantic representation is less than a certain threshold, there is no need to continue iterations, and T _max is equal to (t+1). Generally, when the number of training samples is small, the required T _max is larger, and vice versa.

(7) Use the final semantic representation obtained in step (6) Find the categories of all samples in the test set. For each sample y _j in the test set, j=1,2,...,m, its final semantic representation is (a c×1 column vector), that is, the column vector composed of the probability that the test sample y _j belongs to each category, choose The label of the position of the maximum value element in is used as the category of the test sample Thus, the categories of all samples in the test set are obtained, and the category prediction set of the test set is formed. Complete the classification task of this hyperspectral image.

For example, for the Indian Pine dataset, Figure 2b shows the real ground objects of the dataset. No matter what classification method is used, it can be compared with Figure 2b to verify the classification effect.

Since the present invention fully excavates the interrelationships between the pixels, it finds the local adaptive neighbor set and the non-local similar structure neighbor set of each pixel, and adds the local and non-local spatial constraints constructed by these two neighbor sets into In the designed denoising Markov field model, at the same time, the category information of a small number of labeled samples will also be propagated to its local and non-local neighbors along with the optimization process of the denoising Markov field. The invention enables the hyperspectral image to still achieve high classification accuracy in the case of a small number of marked samples, and the classification result map has good spatial consistency.

Example 2:

Based on local and non-local multi-feature semantic hyperspectral image classification method with embodiment 1, wherein the multiple features described in step (1), including but not limited to: original spectral features, Gabor texture features, differential morphology features (DMP ), where Gabor texture features and differential morphology features (DMP) are expressed as follows:

Gabor texture features: for hyperspectral images Perform principal component analysis (PCA) processing, take the processed first 3-dimensional principal components as 3 reference images, and perform Gabor transformation in 16 directions and 5 scales respectively, and obtain 80-dimensional texture features for each reference image, and stack Gabor texture features with a total dimension of 240 dimensions are obtained together.

Differential Morphological Features (DMP): For hyperspectral images Perform principal component analysis (PCA) processing, take the processed first 3-dimensional principal components as 3 reference images, and perform 5-scale open operations and mutual differences and 5-scale closed operations and mutual differences, each The benchmark images each get 8-dimensional differential features, and stacked together to obtain differential morphological features with a total dimension of 24 dimensions.

In the selection of various features, in addition to using the original spectral information, the present invention also focuses on texture and shape information. Since different feature descriptors have different representations for images, Gabor texture features can be well extracted The local texture information in the hyperspectral image is the correlation information between local pixels, and the differential morphology feature (DMP) can well reflect the edge and size information of the shape block in the hyperspectral image. The invention can improve the discriminability between different types of hyperspectral image pixels by combining spectrum, texture and shape features, and finally improve the classification accuracy and robustness of the hyperspectral image classification method based on local and non-local multi-feature semantics.

In addition to the three features mentioned in this embodiment, other features can also be used in the present invention, such as gray-level co-occurrence matrix, three-dimensional wavelet transform and other features commonly used in hyperspectral image analysis. At the same time, the present invention can use The number of features is not limited to three types. Although more types of features will improve the discriminative power of the method, it will also increase the computational complexity unreasonably, and a large amount of redundant information. The three features used in the present invention: original spectral features , Gabor texture features, and differential morphological features (DMP), which have basically covered most of the information of hyperspectral images.

Example 3:

Based on local and non-local multi-feature semantic hyperspectral image classification method with embodiment 1-2, wherein the local and non-local neighbor set construction method described in step (4) is as follows:

4a) For the hyperspectral image Perform principal component analysis (PCA), extract the first principal component as a reference image, that is, a grayscale image that can reflect the basic contour information of the hyperspectral image, set the number of superpixels LP, and perform entropy-based The superpixel image segmentation of , get LP superpixel blocks

In this embodiment, the entropy-based superpixel segmentation method is used to perform superpixel segmentation on the first principal component grayscale image of the hyperspectral image, and the segmented superpixel blocks can well maintain the edge information and structural information in the image, and the segmentation The size difference of the superpixel block is small, and the shape is relatively regular, which is more suitable for hyperspectral images, and is also widely used by researchers in the field of analysis and processing of hyperspectral images. In the present invention, other image segmentation methods can also be used to replace the entropy superpixel segmentation method, such as mean shift (Mean Shift) and other superpixel segmentation methods based on graph theory.

4b) Set the local window parameter W _local , for the sample y _j in the test set, it is located in the square window W _local × W _local centered on the sample y _j , and all samples belonging to the same superpixel block P _u as the sample y _j Constitute the locally adaptive neighbor set B _j of the sample y _j , so as to obtain the locally adaptive neighbor set of all samples in the test set. Compared with the traditional square window neighbor set, the local neighbor set processed in this way adds a certain degree of adaptability to the local neighbors of each test sample, so that the local neighbors corresponding to each test sample are based on their own surroundings. As determined by the situation, the local neighbor sets of some test samples will not contain a large amount of wrong spatial information due to the fixed window parameter settings. Compared with the fully adaptive local neighbor set construction method, it is possible to set appropriate local window parameters according to the different resolutions of each hyperspectral image and different object types, and introduce certain prior knowledge to improve the local space. information quality, and most of the fully adaptive local neighbor set construction methods have high complexity. The complexity of the local neighbor set construction method of the present invention is almost the same as that of the traditional method, that is, the square window neighbor set, and almost does not increase the computational complexity. But a certain amount of adaptability is introduced.

In the present invention, the local window parameter W _local needs to be set. The larger the parameter, the more local neighbors are included, the richer the local neighbor information is, and it is easier to introduce more wrong neighbors. The smaller the parameter, the fewer local neighbors are included, and it is impossible to extract enough local neighbor information. Setting this parameter requires a compromise between the two. The general values of this parameter are 3, 5, ..., 15, etc. According to the prior information such as the resolution of each hyperspectral image and the type of ground features, Adjust the parameters to varying degrees.

4c) Set the non-local structure window parameter W _nonlocal and the number of non-local neighbors K, respectively for each sample h _q in the original hyperspectral image, q=1,2,...,M in its neighborhood W _nonlocal ×W Mean pooling in _nonlocal to get the structural information of all samples

In the present invention, it is necessary to set the non-local structure window parameter W _nonlocal and the number of non-local neighbors K, wherein the selection method of the non-local structure window parameter W _nonlocal is relatively similar to the local window parameter W _local mentioned in 4b) of this embodiment, But in general, when extracting the information of the sub-block to which each sample belongs, it is necessary to include as much structural information as possible. The value of the non-local structural window parameter W _nonlocal is greater than the local window parameter W _local , such as 15, 17, .. ., 25 etc. Similarly, the parameters are adjusted to different degrees according to the prior information such as the resolution of each hyperspectral image and the type of ground objects. The value of the number of non-local neighbors K is generally 20, 30, ..., 100, etc. This parameter needs to be adjusted according to the number of samples of each type in the hyperspectral image. If the number of samples of each type is large, the non-local The value of the number of neighbors K is relatively large, and if the number of samples of some classes is small, the value of the number of non-local neighbors K is small.

The present invention uses the method of mean value pooling to extract the structural information of the sub-block to which each sample belongs, and the result after mean value pooling can generally show the basic information of the sub-block to which each sample belongs, which can reflect certain degree of distinction. Instead of mean pooling, other aggregation methods can be used to extract information about the sub-block to which each sample belongs, such as max pooling, weighted mean pooling or other more complex pooling methods.

4d) For each sample y _j ,j=1,2,...,m in the test set, use its structural information and the structural information of all remaining samples For comparison, calculate the distance between the sample structure information:

in is the geodesic distance, It represents the root operation of each row in the vector x, and the value of SGD _jq represents the distance between the test sample y _j and the sample q. Compared with the commonly used measurement methods such as Euclidean distance or cosine angle, It has been proved by research that the reflected wave information contained in the hyperspectral image can be regarded as a manifold, and the geodesic distance as a manifold distance can better express the relationship between the reflection value vectors of the hyperspectral image bands. distance. Find the K samples most similar to the test sample y _j , that is, the first K samples with the smallest SGD value, as the non-local similar structure neighbor set C _j of the test sample y _j . At the same time, because each non-local neighbor sample The degree of similarity between the test sample y _j and the test sample y j is different, so the importance of the test sample y _j is different. Each of the K non-local neighbors is given different weights according to the size of the SGD value. The higher the similarity , that is, the larger the SGD value, the greater its weight, and the weight calculation formula is as follows:

In the formula, ω _jh represents the weight corresponding to the non-local similar structure neighbor h of the test sample y _j , and γ is the Gaussian kernel parameter.

The present invention focuses on designing the local adaptive neighbor set and the non-local similar structure neighbor set of all samples in the test set, wherein the local adaptive neighbor set combines the traditional square window and the super pixel block, and fully extracts the local area near the sample location. Spatial information, reducing the introduction of wrong spatial information. The design of the neighbor set with non-local similar structure takes into account a large amount of redundant non-local spatial information in hyperspectral images, and the present invention fully extracts these information. The design of these two sets of neighbors makes the final denoising Markov field make full use of local and non-local spatial information, thereby improving the classification accuracy of hyperspectral images and the spatial consistency of classification result maps.

Example 4:

Based on local and non-local multi-feature semantic hyperspectral image classification method with embodiment 1-3, wherein in step (5) and (6), minimize the global energy of Markov field, use iterative conditional model method (ICM) to solve, The global energy minimization is transformed into minimizing the energy of each local potential group. For the test sample y _j , its corresponding first-order local energy function, which is the local energy function used in step (5), is:

The corresponding (t+1)th order local energy function, that is, the local energy function used in step (6) is:

Note, the distance measurement method in the above two local energy functions is the geodesic distance, which is the same as the geodesic distance described in Embodiment 3. Compared with traditional distance measurement methods, such as Euclidean distance, cosine angle, etc., geodesic distance has been proved to be more suitable for measuring the distance between two semantic vectors, that is, probability vectors.

Minimizing the above two local energy functions is a convex optimization problem. The present invention uses a simple and fast gradient descent method to minimize the above energy functions. Taking the first-order local energy function of the test sample y _j as an example, its gradient is:

in represent The kth element of , and:

Using the above formula, the global energy of the Markov field is minimized.

The invention improves the Markov field model with multi-feature directions and non-local directions, and replaces the traditional discrete category similarity and difference regular term with the geodesic distance regular term between semantic vectors, making the Markov field model more efficient. It is suitable for dealing with hyperspectral image classification problems, solves the over-smoothing problems caused by traditional Markov models, improves the spatial consistency of hyperspectral image classification result maps, and improves classification accuracy. After that, the global energy minimization problem is transformed into multiple local minimization problems through the iterative conditional model method (ICM), and the simple gradient descent method is used to solve the problem, and the computational complexity is lower than that of the traditional Markov field model. Algorithm, and in the optimization iteration process, it avoids falling into the local optimal solution, improves the robustness of the model, and thus improves the robustness of the classification results.

A complete embodiment is given below to further illustrate the present invention.

Example 5:

Based on local and non-local multi-feature semantic hyperspectral image classification method with embodiment 1-4,

With reference to Fig. 1, concrete implementation steps of the present invention include:

Step 1, input hyperspectral image M is the total number of all pixels in the hyperspectral image, h _q is a column vector, representing the reflection value of each band of pixel q. Use different feature descriptors to extract multiple features from hyperspectral images. In the simulation experiment of the present invention, the Gabor texture feature, the morphological difference feature (DMP) and the original spectral feature are respectively extracted.

1a) Gabor texture feature is used for the first three principal component images after dimensionality reduction of the original hyperspectral image using the principal component analysis method (PCA) to perform Gabor filter transformations in multiple directions and multiple scales. Note that 16 is used in this embodiment. direction and 5 scales, stack the results of each filter to get the Gabor texture feature of the hyperspectral image, and get the texture feature vector of each pixel correspondingly.

1b) The morphological difference feature (DMP) also performs opening and closing operations of different scales for the first three principal component images. Note, in this embodiment, 5 scales of opening operations and 5 scales of closing operations are used, and the adjacent The difference between the opening or closing operations of the two scales, and all the obtained differences are stacked to obtain the morphological difference feature of the hyperspectral image, and the corresponding morphological difference feature (DMP) vector of each pixel is obtained.

1c) The original spectral feature is to use the reflection value of each band of each pixel directly as the spectral feature of the pixel.

Step 2, select an equal proportion of pixels from each type of pixels in the hyperspectral image as marked pixels, the total number of marked pixels is N, and the remaining m pixels of the hyperspectral image are used as For unmarked pixels, V feature vectors of each pixel are used to represent the pixel, and the feature dimensions of the samples are respectively where _Dv represents the dimensionality of the vth feature.

Step 3: Select a labeled training set X and a test set Y to obtain the semantic representation set S of all samples.

3a) Use N labeled samples to form a labeled training set Its corresponding category label set is R represents the field of real numbers. using the labeled training set A single feature vector set X ^v and a category label set L in the training V probabilistic support vector machine (SVM) models. The kernel type of the support vector machine is a radial Gaussian kernel, and the kernel parameter r and the penalty parameter c are obtained by multiple cross-validation.

3b) Use m unlabeled samples to form a test set The different eigenvector sets of the test set are respectively Input them into the corresponding V probabilistic support vector machine models trained in step 3a), and obtain the probability that all samples in the test sample set belong to each category under the expression of different feature descriptors, which is the probability of the test sample Semantic representation, such as the semantic representation of the test sample y _j as For the samples in the training set, its semantic representation is a 0-1 coded vector, the position of the class label is 1, and the other positions are 0, indicating that the probability of belonging to its own category is 1, and the probability of belonging to other categories is 0. Therefore, the semantic representation set of all samples of the hyperspectral image is obtained as

Step 4. Construct a local adaptive neighbor set B and a non-local similar structure neighbor set C for each sample in the unlabeled test set Y.

4a) Use Principal Component Analysis (PCA) method to reduce dimensionality of the original hyperspectral image, select the first principal component image (a grayscale image) as the reference image, set the number of superpixels LP, and use the entropy-based superpixel The pixel segmentation method performs superpixel segmentation on the image to obtain LP superpixel blocks

4b) Set the local adaptive window parameter W _local , and construct a local adaptive neighbor set for each test sample according to the superpixel block obtained in step 4a). Taking the test sample y _j as an example, if there is a sample n that belongs to the square window of W _local × W _local size centered on the test sample y _j , and belongs to the same superpixel block P _u as the test sample y _j , then it is called This sample n is a locally adaptive neighbor of the test sample y _j , n∈B _j , and so on, to obtain the locally adaptive neighbor set of each test sample

4c) Set the non-local structural window parameter W _nonlocal and the number K of non-local neighbors. First, the original hyperspectral image M represents the number of all samples in the hyperspectral image. Mean pooling is performed in the adaptive window of W _nonlocal × W _nonlocal size, that is, the local adaptive neighbor set whose parameter is W _nonlocal , and the construction method of the adaptive window is the same as that described in 4b). Get the local structure information of each sample

4d) According to the obtained in step 4c) The local structure information of each test sample and the similarity between all remaining points are calculated separately. For each sample y _j in the test set, j=1,2,…,m, use its structural information and the structural information of all remaining samples For comparison, calculate the local structure information of the sample and the local structure information of all remaining samples The similarity between them, the similarity calculation formula is as follows:

in is the geodesic distance, Represents the root operation for each row in the column vector x. Calculate the similarity between every two samples to get a similarity matrix SD, and this matrix is a symmetrical matrix SD(j,q)=SD(q,j), which represents the relationship between the test sample y _j and the sample q the structural similarity between them.

4e) According to the similarity matrix SD obtained in step 4d), for each test sample, select its top K most similar samples as its non-local neighbors. Taking the test sample y _j as an example, select the jth column in the similarity matrix SD, each element of which represents the similarity between the test sample y _j and each sample, and select one of them except itself (the similarity between itself and itself is the largest ), the top K samples with the largest value, as the non-local neighbors of the test sample y _j , the non-local neighbor set C _j of the test sample y _j is obtained, and so on, the non-local neighbor set of each test sample is obtained

4f) At the same time, according to the size of the similarity value, give different weights to the K non-local neighbors, so that the more similar neighbors have higher weights, relatively speaking, reduce the weight of the lower similarity neighbors, In order to improve the rationality and adaptability of non-local space constraints, the weight calculation formula is as follows (where γ is a Gaussian kernel parameter):

Step 5, according to the semantic representation of each sample calculated in Step 3 and Step 4, and its corresponding local and non-local neighbor sets, construct a denoising Markov field model, where the local neighbor set is used as a local space constraint, non-local The set of local neighbors is added to the model as a non-local spatial constraint. The iterative conditional model method (ICM) is used to minimize each local energy, that is, the energy of the potential group where each sample is located, to achieve the purpose of minimizing the global energy, and calculate the denoising semantic representation of each test sample.

5a) Taking the test sample y _j as an example, the V semantic representations of the test sample y _j and the semantic representation of all samples in its local and non-local neighbor sets Input to the denoised Markov field local energy function. The energy of the potential group where the test sample y _j is located is:

Among them, the first item on the right side of formula (3) is a self-constraint item, the second item is a local space constraint item, and the third item is a non-local space constraint item. Note that all constraint items in this formula are multiple semantic cases under the constraints. Minimize this function using the gradient descent method to obtain a first-order noise-reduced semantic representation of the test sample _yj By analogy, by traversing all the test samples, the first-order noise reduction semantic representation of all test samples can be obtained, while the semantic representation of the training samples remains unchanged, so the first-order noise reduction semantic representation set of all samples can be obtained

5b) Since the iterative conditional model method (ICM) is a gradual optimization process, the global energy minimization cannot be completed in one iteration, and the final convergence result cannot be reached. The first-order noise reduction semantic representation set obtained in step 5a) Input into the denoising Markov field model to continue iterative optimization. Set the maximum number of iterations T _max , taking the test sample y _j as an example, its first-order denoising semantic representation And the first-order noise reduction semantic representation of samples in its local and non-local neighbor sets Input it into the local energy function of the denoising Markov field, and minimize the energy function to obtain the second-order denoising semantic representation of the test sample y _j By analogy, traverse all test samples. Note that when seeking the semantic representation of any order noise reduction, the semantic representation of all training samples remains unchanged, and it is still a 0-1 encoding column vector, and then the second order reduction of all samples is obtained. noisy semantic representation Since the fusion of multiple semantics has been completed in step 5a) of this example, the energy function no longer needs to fuse multiple semantics, so the local energy function is simplified as:

The meanings of the three items on the right side of formula (4) are the same as those of formula (3), but there is no need for multiple semantic fusions, so each item is a constraint under a single semantic situation. It can be seen from the formula that, according to the t-th order noise reduction semantic representation of the test sample y _j and its local and non-local neighbor samples, the (t+1)th order noise reduction semantics of the test sample y _j can be obtained by minimizing this function express. The above-mentioned first-order denoising semantic representation The second-order noise reduction semantic representation can be obtained This is a special case at t=1. This process is iterated cyclically until t=T _max -1, and the T _max order denoising semantic representation of the test sample y _j can be obtained

Repeat this process for each test sample, you can get the T _max order noise reduction semantic representation of each test sample, and then get the T _max order noise reduction semantic representation of all samples

Step 6, the semantic representation of the sample in the present invention is the column vector formed by the probability that the sample belongs to each category, so the T _max order noise reduction semantic representation obtained according to step 5 For the test sample y _j choose its semantic representation vector The position of the maximum value is the category label of the test sample y _j By analogy, the category labels of all test samples are obtained

In the present invention, a semantic hyperspectral image classification method based on local and non-local multi-features is proposed, and a new Markov field is constructed in the method, which is called denoising Markov field. The model minimizes the global energy of the Markov field through the iterative conditional model method (ICM), which can reasonably fuse the multi-feature semantics under local and non-local space constraints, and obtain a more comprehensive and comprehensive model for all test samples. Semantic representation with lower noise, and then get the predicted class labels of all test samples. This method makes full use of two aspects of information. Multi-feature information can describe hyperspectral images more comprehensively, and local and non-local spatial constraints can fully mine the relationship between hyperspectral image pixels. Finally, compared with traditional classification methods, the local and non-local multi-feature semantic hyperspectral image classification method proposed by the present invention improves the accuracy, robustness and spatial consistency of classification results.

Effect of the present invention can be further illustrated by following simulation experiments:

Embodiment 6:

The semantic hyperspectral image classification method based on local and non-local multi-features is the same as that in Embodiments 1-5.

1. Simulation conditions:

The simulation experiment uses the Indian Pine image acquired by the airborne visible/infrared imaging spectrometer AVIRIS of NASA Jet Propulsion Laboratory in June 1992 in northwestern Indiana, as shown in Figure 2a. The image size is 145×145, with a total of 220 There are 200 bands for noise removal and atmospheric and water absorption bands, see Figure 2b. After manual marking, there are a total of 16 types of ground object information.

The simulation experiment was carried out with MATLAB 2014a software on a WINDOWS7 system with Intel Core(TM) i5-4200H CPU, main frequency 2.80GHz, and 12G memory.

Table 1 gives the 16 categories of data in the Indian Pine image.

Table 1 The 16 categories of data in the Indian Pine image

category classification name Sample size Number of training samples 1 Alfalfa 46 3 2 Corn - Unploughed 1428 72 3 Corn - Irrigation 830 42 4 corn 237 12 5 grass 483 25 6 the trees 730 37 7 cut grass 28 2 8 Haystack 478 twenty four 9 buckwheat 20 1 10 Soybeans - Unploughed 972 49 11 soybeans - irrigation 2455 123 12 soy 593 30 13 wheat 205 11 14 forest 1265 64 15 buildings-grass-trees 386 20 16 stone-rebar 93 5

2. Simulation content and analysis:

Use the present invention and existing three kinds of methods to classify hyperspectral image Indian Pine, existing three kinds of methods are respectively: support vector machine (SVM+CK) based on fusion kernel, support vector machine combined Markov field (SVM+MRF) , Joint Sparse Representation (SOMP). The hyperspectral image classification method based on local and non-local multi-feature semantics proposed in the present invention uses multi-feature semantic representation and spatial constraints to classify hyperspectral images, abbreviated as NE-MFAS. In order to verify the effectiveness of the method of the present invention, the simulation Added two kinds of simplified version methods MFAS and MFS in the frame of the present invention in the experiment, wherein use MFAS method is that in the step (4) of embodiment 6, only construct local neighbor set, do not utilize the information of non-local neighbor set, To verify the impact of non-local neighbor set information on the classification results, the classification results are shown in Figure 3e. Using the MFS method in step (4) of Embodiment 6 not only does not use the non-local neighbor set information, but also does not use the superpixel information when constructing the local neighbor set, and only selects the square window as its local neighbor set to verify the superpixel The influence of constraints on the classification results, the classification results are shown in Figure 3d. Among them, the penalty factor of the support vector machine (SVM) method is Kernel parameters As determined by 5-fold cross-validation, the sparse parameter of the SOMP method was set to 30, and the spatial scale parameter was set to 7×7. The number of superpixels L used in the present invention is set to 75, the size of the local window W _local is 7×7, the size of the non-local structure window W _nonlocal is 21×21, the number of non-local neighbors K is 30, and the Gaussian kernel parameter γ is 0.05, and the maximum number of iterations T _max is set to 3. The selection method of the training samples is as follows, 5% of the pixels are randomly selected from the 16 types of data as the training samples, and the remaining 95% are used as the test samples.

As shown in Figure 3, the classification results of each method are obtained when the training samples are completely consistent, and Figures 3a-3c are the classification results of the three existing methods SVM+CK, SVM+MRF, and SOMP, respectively. It can be seen that these three classification methods perform well in most areas. Among them, the classification result map of SOMP, see Fig. 3c, there are many discrete noise points, and the spatial consistency is poor. For the classification result map of SVM+MRF, see Figure 3b. Compared with Figure 3c, there are almost no discrete noise points, but there are a lot of over-smoothing phenomena, and the edge corrosion of the classification result map is more serious. The classification result diagram of SVM+CK, see Figure 3a, compared with Figures 3b and 3c, the result is optimal, but there are still some discrete noise points and a large number of edge erosion. Figures 3d-3f correspond to the classification results of MFS, MFAS, and NE-MFAS, respectively. Comparing with Figures 3a-3c, it can be seen that the method proposed by the present invention performs better than the existing methods in terms of local area coherence, edge preservation, and small sample areas. Among them, the classification result map of MFS is shown in Figure 3d. Compared with the traditional three methods, MFS has almost no discrete noise points and the spatial consistency of most areas is good, but there is a certain edge corrosion phenomenon. Compared with MFS, MFAS adds more superpixel constraint information. See Figure 3e. Compared with Figure 3d, MFAS classification result map has a great improvement in edge preservation, and there is almost no edge corrosion phenomenon, but in some small Sample category regions, such as the yellow rectangle region on the right side of the image, are almost impossible to classify correctly. After adding non-local space constraints to MFAS, the classification result diagram of NE-MFAS of the present invention, see Figure 3f, compared with Figures 3d and 3e, while maintaining good spatial consistency, for small sample category areas, For example, the yellow rectangular area on the right side of the image is almost all correctly classified.

Through the changes in Figure 3d to 3f, it can be found that the adaptive local space constraint of the superpixel constraint proposed in the present invention can well maintain the edge information of the image and enhance the spatial consistency, while the non-local space constraint proposed in the present invention can Further improving the upper limit of classification accuracy, especially in improving the classification accuracy of categories with relatively few samples, has a very good performance.

Embodiment 7:

The semantic hyperspectral image classification method based on local and non-local multi-features is the same as in Embodiment 1-6, and the conditions and contents of the simulation are the same as in Embodiment 6. The difference between the classification result graphs described in Example 6 can only be judged by naked eyes. Here, the advantages of the method of the present invention compared with other methods are reflected from the data through comparative analysis of classification accuracy. All experiments were carried out 10 times to take the average value. Note that the training samples of each experiment are randomly selected, that is to say, the training samples of one experiment and another experiment are not exactly the same.

Table 2 provides the experimental precision contrast of existing three kinds of methods, two kinds of simplified version methods of the present invention and the inventive method:

Table 2 Three methods on the Indian Pine image and the experimental accuracy results of the present invention

method Classification accuracy (%) SVM+CK 91.59±0.94 SVM+MRF 84.91±1.43 SOMP 87.55±1.24 Simplified version 1 (MFS) of the present invention 96.87±0.79 Simplified version of the invention 2 (MFAS) 97.24±0.64 Invention (NE-MFAS) 98.20±0.58

As can be seen from Table 2, the mean value of the obtained results of the method of the present invention in 10 repeated experiments is much higher than the existing three methods, and the standard deviation is lower than the existing three methods, it can be illustrated that the method of the present invention no matter in Both classification accuracy and robustness are the best. On the one hand, the method of the present invention fully combines a large amount of discriminant information contained in multiple features. On the other hand, it is far superior to the other three methods in mining spatial information. methods for spatial information. From MFS to MFAS, the classification accuracy has increased by 0.5%, indicating that superpixel constraints have played a positive role in the extraction of spatial information. Compared with the traditional square window, the adaptive local neighbor set designed by the present invention can eliminate some negative The neighboring points of the function can extract richer local spatial information, including edge and structure information. From MFAS to NE-MFAS, the classification accuracy increased by 1%, which shows that non-local neighbors contain a lot of valuable information, and extracting this information can further improve the upper limit of classification accuracy. The classification accuracy of the invention reaches 98.20% in the case of a small number of samples of 5%, and the standard deviation is only 0.58%, which is better than most existing methods.

In summary, the present invention discloses a semantic hyperspectral image classification method based on local and non-local multi-features. It mainly solves the problems of low correct rate, poor robustness and weak spatial consistency of existing hyperspectral image classification methods. The steps include: extracting different features from the original hyperspectral image using a variety of feature extraction methods; representing each pixel of the hyperspectral image with multiple feature vectors, selecting a labeled training set and a test set; The label training set is input into the probabilistic support vector machine, and the multiple features of each pixel in the test set are mapped to the semantic space, and multiple semantic representations of each test sample are obtained; the noise reduction Markov field model is used to introduce spatial information at the same time And multi-semantic information, minimize the energy of the Markov field, get the noise-reduced semantic representation of each test sample, and then get the category information of the entire test set. The present invention utilizes the perfect fit of the semantic representation obtained by the support vector machine and the Markov field model, fully combines multi-feature information and local and non-local spatial context information, and finally can classify hyperspectral images more accurately. Compared with two simplified versions of the method of the present invention, the effectiveness of the method of the present invention is verified. Compared with the existing three methods, it shows that the method of the present invention has high accuracy, high robustness, and excellent spatial consistency in the case of small samples. It can be used in military detection, map drawing, vegetation survey, mineral detection, etc.

The parts that are not described in detail in this embodiment are commonly known and commonly used means in this industry, and will not be described here one by one. The above examples are only illustrations of the present invention, and do not constitute a limitation to the protection scope of the present invention. All designs that are the same as or similar to the present invention fall within the protection scope of the present invention.

Claims (5)

1. it is a kind of based on the semantic hyperspectral image classification method of local and non local multiple features, it is characterised in that to include as follows Step：

(1) input picture, extracts the various features of image：Input high spectrum imageM is institute in high spectrum image There are the total number of pixel, h_qFor a column vector, the reflected value vectors that constituted of the pixel q in each wave band is represented；It is right High spectrum image extracts various features respectively, includes：Original spectrum feature, Gabor textural characteristics, difference morphological feature； The high spectrum image includes c class pixels, wherein have it is N number of have a labelling pixel, m unmarked pixel, each picture of image Vegetarian refreshments is a sample, and each sample is made up of V characteristic vector, represents table of the sample in the case where different characteristic description is sub respectively State, V is the number of feature classification；

(2) hyperspectral image data collection is divided into training set and test set：There is labelling pixel to constitute instruction as training sample with N number of Practice collectionIts corresponding category label collection isWith m unmarked pixel Test set is constituted as test sampleWherein, x_iRepresent i-th sample of training set This, y_jRepresent j-th sample of test set, l_iIt is the category label belonging to i-th training sample, D_vRepresent the dimension of v category features Number, R represent real number field；

(3) various features of all samples are mapped to into corresponding semantic expressiveness using probability support vector machine (SVM)：It is sharp respectively Use training setIn all samples V characteristic vector and its corresponding category label CollectionBuild V probability supporting vector machine grader, the kernel function of the grader is radial Gaussian kernel, nuclear parameter and Punishment parameter is obtained by many times of cross validations；By test setIn all samples V Individual characteristic vector is separately input in the V correspondence grader for building, and obtains describing under sublist states in different characteristic, and each is tested Sample y_j, j=1,2 ..., m belong to the probability of each classification, used as the semantic expressiveness of each test sampleFor each sample x in training set_i, i=1,2 ..., N, it belongs to classification l itself_i's Probability is 1, and the probability for belonging to other classifications is 0, obtains the corresponding various semantic expressivenesses of various featuresWhereinL_iBehavior 1, other behaviors 0；So as to obtain the high-spectrum Various semantic expressivenesses of all samples as in

(4) local and the non local neighbour set of all samples in test set are constructed：For each sample y in test set_j,j =1,2 ..., m construct its local auto-adaptive neighbour set B_jWith non local analog structure neighbour set C_j；

(5) noise reduction markov field model is built, carries out various semantic expressiveness fusions and the drop of semantic expressiveness of test sample Make an uproar process：To each test sample y_j, j=1,2 ..., m are proceeded as follows respectively, by y_jCorresponding semantic expressivenessLocal auto-adaptive neighbour set B_jIn all samples semantic expressivenessIt is similar with non local Structure neighbour set C_jIn all samples semantic expressivenessIt is input in local energy function, it is minimum Change the energy function, obtain test sample y_jSingle order noise reduction semantic expressivenessAt the same time keep the language of sample in training set Justice represents constant, obtains the single order noise reduction semantic expressiveness of all samples of the high spectrum image

(6) the single order noise reduction semantic expressiveness to all samples of high spectrum imageFurther iteration optimization：Setting greatest iteration time Number T_max, t is current iteration algebraically, and each test sample is proceeded as follows：By test sample y_j, and set B_jAnd set C_jIn all samples t rank noise reduction semantic expressivenessesAs noise reduction Markov Random Fields local The input of energy function, minimizes the energy function, obtains test sample y_j(t+1) rank semantic expressivenessAt the same time Continue to keep the semantic expressiveness of sample in training set constant, and then obtain (t+1) rank drop of all samples of the high spectrum image Make an uproar semantic expressivenessIteration process, until t=T_max- 1 stops, and obtains the T of all samples of the high spectrum image_max Rank noise reduction semantic expressiveness, that is, final semantic expressiveness(7) using final semantic expressivenessTry to achieve test Concentrate the classification of all samples：For each sample y in test set_j, j=1,2 ..., m, its final semantic expressiveness isThat is test sample y_jBelong to the column vector constituted by the probability of each classification, maximum element is located in selecting the vector The label of position is used as y_jClassificationSo as to obtain test set class prediction setComplete the high spectrum image Classification task.

2. according to claim 1 based on the semantic hyperspectral image classification method of local and non local multiple features, its feature It is that the Gabor textural characteristics in various features and difference morphological feature described in step (1) are expressed as follows respectively：

Gabor textural characteristics：Front 3-dimensional main constituent after being processed with the high spectrum image principal component analysiss divides as benchmark image 16 directions are not carried out, the Gabor transformation of 5 yardsticks, each benchmark image respectively obtain the textural characteristics of 80 dimensions, be stacked Obtain the Gabor textural characteristics that total dimension is 240 dimensions；

Difference morphological feature：Front 3-dimensional main constituent after being processed with the high spectrum image principal component analysiss divides as benchmark image Opening operation and mutually making the difference the closed operation with 5 yardsticks and mutually make the difference for 5 yardsticks is not carried out, each benchmark image is respectively obtained The Differential Characteristics of 8 dimensions, be stacked the difference morphological feature for obtaining that total dimension is 24 dimensions.

3. according to claim 1 based on the semantic hyperspectral image classification method of local and non local multiple features, its feature It is that the local described in step (4) is as follows with non local neighbour set building method：

Principal component analysiss 4a) are carried out to the high spectrum image, first principal component is extracted as benchmark image, super-pixel number is set LP, carries out the super-pixel image segmentation based on entropy rate, obtains LP super-pixel block

4b) local window parameter W is set_local, for the sample y in test set_j, positioned at sample y_jCentered on square window W_local×W_localIn, but with sample y_jBelong to same super-pixel block P_uAll samples constitute sample y_jLocal auto-adaptive it is near Adjacent set B_j；

Non local topology window parameter W 4c) is set_nonlocal, and non local neighbour's number K, respectively to original high spectrum image In each sample h_q, q=1,2 ..., M is in its neighborhood W_nonlocal×W_nonlocalIn carry out average pond, obtain all samples Structural informationFor each sample y in test set_j, j=1,2 ..., m use its structural informationWith remaining institute There is the structural information of sampleCompare, calculate the how far between composition of sample information：

${\mathrm{SGD}}_{jq}=g(\stackrel{\‾}{h_j},\stackrel{\‾}{h_q})$

WhereinIt is geodesic distance,Each row in representing to vector x carries out out root behaviour Make, SGD_jqValue just represent test sample y_jThe how far of structural information and sample q between；Find and test sample y_jMost K similar sample, that is, the front K sample that SGD values are minimum, constitute test sample y_jNon local analog structure neighbour collection Close C_j, at the same time, according to the size of SGD values, the K different weight of non local neighbour is given respectively, weight calculation formula is such as Under：

${\mathrm{\ω}}_{jh}=\exp \left(\frac{-\left|{\mathrm{SGD}}_{jh}\right|}{\mathrm{\γ}}\right),h\∈C_j$

In formula, ω_jhRepresentative sample y_jNon local analog structure neighbour h weight size, γ is Gauss nuclear parameter.

4. according to claim 1 based on the semantic hyperspectral image classification method of local and non local multiple features, its feature It is, wherein the calculating test sample y described in step (5)_jThe single order local energy function that adopted of noise reduction semantic expressiveness for：

$E\left(\stackrel{\‾}{s_j^{\left(1\right)}};\left(B_j,C_j\right)\right)=\underset{v=\left(1,\mathrm{...},V\right)}{\Σ}g\left(\stackrel{\‾}{s_j^{\left(1\right)}},s_j^v\right)+\underset{r\∈B_j}{\Σ}\underset{v=\left(1,\mathrm{...},V\right)}{\Σ}g\left(\stackrel{\‾}{s_j^{\left(1\right)}},s_r^v\right)+\underset{h\∈C_j}{\Σ}\underset{v=\left(1,\mathrm{...},V\right)}{\Σ}g\left(\stackrel{\‾}{s_j^{\left(1\right)}},{\mathrm{\ω}}_{jh}\·s_h^v\right)$

The energy function is minimized using gradient descent method, its gradient is：

$\begin{array}{c}\frac{\∂E\left(\stackrel{\‾}{s_j^{\left(1\right)}};\left(B_j,C_j\right)\right)}{\∂\stackrel{\‾}{s_{j,k}^{\left(1\right)}}}=\frac{1}{\left|V\right|}\underset{v=\left(1,\mathrm{...},V\right)}{\Σ}f\left(\stackrel{\‾}{s_{j,k}^{\left(1\right)}},s_{j,k}^v\right)+\frac{1}{\left|V\right|}\underset{r\∈B_j}{\Σ}\underset{v=\left(1,\mathrm{...},V\right)}{\Σ}f\left(\stackrel{\‾}{s_{j,k}^{\left(1\right)}},s_{r,k}^v\right)\\ +\frac{1}{\left|V\right|}\underset{h\∈C_j}{\Σ}\underset{v=\left(1,\mathrm{...},V\right)}{\Σ}f\left(\stackrel{\‾}{s_{j,k}^{\left(1\right)}},{\mathrm{\ω}}_{jh}\·s_{h,k}^v\right)\end{array}$

WhereinRepresentK-th element, and：

$f(x,y)=\frac{\∂g(x,y)}{\∂x}=-\frac{\sqrt{y}}{2\sqrt{x}\sqrt{1-{\left(\sqrt{x}\sqrt{y}\right)}^2}}$

5. according to claim 1 based on the semantic hyperspectral image classification method of local and non local multiple features, its feature It is, wherein to test sample y in step (6)_jThe further iteration optimization of single order noise reduction semantic expressiveness in, (t+1) rank office Portion's energy function is：

$E(\stackrel{\‾}{s_j^{(t+1)}};(B_j,C_j))=g(\stackrel{\‾}{s_j^{(t+1)}},\stackrel{\‾}{s_j^{\left(t\right)}})+\underset{r\∈B_j}{\Σ}g(\stackrel{\‾}{s_j^{(t+1)}},\stackrel{\‾}{s_r^{\left(t\right)}})+\underset{h\∈C_j}{\Σ}g(\stackrel{\‾}{s_j^{(t+1)}},{\mathrm{\ω}}_{jh}\·\stackrel{\‾}{s_h^{\left(t\right)}})$