CN112329818B - Hyperspectral image non-supervision classification method based on graph convolution network embedded characterization - Google Patents
Hyperspectral image non-supervision classification method based on graph convolution network embedded characterization Download PDFInfo
- Publication number
- CN112329818B CN112329818B CN202011124146.9A CN202011124146A CN112329818B CN 112329818 B CN112329818 B CN 112329818B CN 202011124146 A CN202011124146 A CN 202011124146A CN 112329818 B CN112329818 B CN 112329818B
- Authority
- CN
- China
- Prior art keywords
- super
- representation
- graph
- vertex
- hyperspectral
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/28—Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A40/00—Adaptation technologies in agriculture, forestry, livestock or agroalimentary production
- Y02A40/10—Adaptation technologies in agriculture, forestry, livestock or agroalimentary production in agriculture
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Probability & Statistics with Applications (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a hyperspectral image non-supervision classification method based on graph rolling network embedded characterization, which comprises the following steps: sequentially extracting EMP and spectral features of the image to be hyperspectral to obtain spatial spectrum combined features; performing super-pixel segmentation on the spatial joint characteristics to obtain super-pixel points of the hyperspectral image to be subjected to hyperspectral; solving a solution elastic net representation of the super pixel point, taking the super pixel point associated with the non-zero component in the representation coefficient of the solution elastic net representation as a neighbor of the current point, and constructing a graph model of the super pixel point; performing graph convolutional network embedded characterization learning based on a graph model, and obtaining low-dimensional features through hierarchical vertex aggregation operation; according to the low-dimensional characteristic representation, the non-supervision classification of the hyperspectral image is realized by using a K-means algorithm, and the purpose of accurately classifying the hyperspectral image can be realized.
Description
Technical Field
The invention relates to the technical field of image processing, in particular to a hyperspectral image non-supervision classification method based on graph rolling network embedded characterization.
Background
In the 80 s of the last century, hyperspectral remote sensing technology began to rise and develop rapidly, and the ability of humans to observe and recognize things on the earth had a qualitative leap. The hyperspectral remote sensing technology can capture the corresponding spectral information while acquiring the spatial image of the observed ground object, so that the hyperspectral image is presented as a three-dimensional data cube, and the first real map-in-one imaging is realized. The multiple spectral bins of each pixel form a spectral curve that contains rich information of the earth's surface components that can be used to identify different types of terrain. Currently, hyperspectral image classification has become a popular study in the current hyperspectral remote sensing field. Classification of hyperspectral images also presents certain challenges due to their own features.
The hyperspectral image classifier learning aspect includes supervised, unsupervised, and semi-supervised classification models, differing in whether tagged data samples are used during the model training phase. Hyperspectral images require a significant cost effort to obtain a sufficient number of marked samples. The non-supervision method does not depend on label category information of the sample, only the inherent attribute and rule of uncalibrated samples are mined through a classifier, the uncalibrated samples are divided into different clusters according to the difference between different pixels, and the model is not required to be trained, so that the number of training samples is not required. Therefore, the unsupervised classification model becomes an attractive alternative to hyperspectral image classification. The classical unsupervised classification algorithm comprises a K-means clustering algorithm, an FCM algorithm, a spectral clustering algorithm and the like, wherein the K-means clustering algorithm is used for continuously and iteratively updating the class center to minimize the square sum value by calculating the average sum value of pixels near the class center to the distance of the pixels, and the K-means clustering algorithm is a simple, common and effective algorithm.
Unsupervised classification of hyperspectral images relies on efficient characterization to represent many attempts by researchers. The original method generally uses spectral information directly, and it is difficult to obtain a robust classification result. After the importance of the spatial information in hyperspectral classification is recognized, two people, namely Pesaresi and Benedicktsson, adopt a morphological transformation method to construct morphological distribution characteristics for extracting spatial structure information. In consideration of the characteristic of spatial spectrum integration of hyperspectral images, the expandable morphological spatial features EMP and spectral features are jointly extracted by Fauvel and Chanussot to a spatial spectrum joint feature representation, and then classification is carried out by adopting a support vector machine model, so that the performance improvement of algorithm classification is realized.
Aiming at the problem of over-high dimension of the air combined features, the features must be dimension-reduced, and the conventional dimension-reducing method comprises PCA and the like. However, the classical graph model cannot process a large-size hyperspectral image due to the SVD classification of the graph Laplace matrix, and meanwhile, the classical graph model is a shallow learning model, which is unfavorable for extracting inherent low-dimensional features. The graph rolling network operates on a graph, can be suitable for non-Euclidean irregular data based on the graph, fully utilizes image characteristics, and flexibly maintains category boundaries.
Disclosure of Invention
Aiming at the problems, the invention provides a hyperspectral image non-supervision classification method based on graph rolling network embedded characterization.
In order to achieve the purpose of the invention, a hyperspectral image non-supervision classification method based on graph rolling network embedded characterization is provided, which comprises the following steps:
s10, sequentially extracting EMP and spectral features of the image to be hyperspectral to obtain spatial spectrum combined features;
s20, performing super-pixel segmentation on the spatial joint characteristics to obtain super-pixel points of the hyperspectral image to be subjected to hyperspectral;
s30, solving an elastic net representation of the super-pixel points, and constructing a graph model of the super-pixel points by taking the super-pixel points associated with non-zero components in the representation coefficients of the elastic net representation as the neighbors of the current points;
s40, performing graph convolutional network embedded characterization learning based on a graph model, and obtaining low-dimensional features through hierarchical vertex aggregation operation;
s50, according to the low-dimensional characteristic representation, the non-supervision classification of the hyperspectral image is realized by using a K-means algorithm.
In one embodiment, sequentially performing EMP and spectral feature extraction on the image to be hyperspectral, and obtaining the spatial spectrum joint feature includes:
wherein V represents a spatial spectrum joint feature matrix, X represents a spectral feature matrix, EMP represents an EMP feature matrix, m represents the number of main components, N represents the number of circular structural elements with different radiuses, d represents the number of spectral bands, and N represents the number of samples.
In one embodiment, performing superpixel segmentation on the spatial joint feature to obtain a superpixel point of the image to be hyperspectral includes:
D i,c =(1-λ)×D spectral +λ×D spatial ,
D spectral =tan(SAD(x i ,x c )),
wherein D is spectral Inter-spectral distance, SAD (x) i ,x c ) Represents the angular distance of the spectrum, D spatial Representing normalized spatial Euclidean distance, r is the diagonal length of the search neighborhood, used to constrain the search range to a local neighborhood around each cluster center, D i,c Represents a weighted sum of spectral and spatial distances, lambda represents an acting parameter balancing spectral and spatial distances, x c =(x c1 ,x c2 ,…,x cn ) Representing a cluster center, which corresponds to a spatial coordinate of (cx, cy), x i =(x i1 ,x i2 ,…,x il ) The representation is located at the cluster center x c The local neighborhood of l pixels corresponds to spatial coordinates (ix, iy). The patent suggests that the parameter lambda is set to less than 0.5.
In one embodiment, solving a solution elastic net representation of a super pixel point, taking the super pixel point associated with a non-zero component in a representation coefficient of the solution elastic net representation as a neighbor of a current point, and constructing a graph model of the super pixel point includes:
s31, selecting all other super-pixel points based on each super-pixel point to construct a dictionary, and finding out elastic net representations of all the pixel points in the data set by solving the following constraint optimization problem:
s.t.sx i =SD i sc i +e i ,
wherein SD is i Is a dictionary formed by all super-pixel points, sc i Is super pixel point sx i Dictionary based SD i Obtained representation coefficient, sc= [ SC ] 1 ,sc 2 ,…,sc N ]Is a coefficient matrix, E is a characterization error matrix, lambda and gamma are regularization parameters, E i Is an error vector;
s32, constructing an elastic net graph model of the super-pixel point according to the elastic net sparse representation coefficient of each sample point, and defining according to an elastic net representation coefficient matrix SCAnd establishing edge connection between hyperspectral pixel points as an adjacency matrix of the graph model to obtain the graph model of the super pixel points.
In one embodiment, graph convolutional network embedded token learning based on a graph model, obtaining low-dimensional features through hierarchical vertex aggregation operations includes:
s41, sampling each vertex of the graph model in a layering manner, randomly sampling a fixed number of neighborhood vertices at each layer of the graph model by using a random walk mode, and approximating the vertices which are not sampled by using a historical expression of the vertices;
s42, updating self information by the GraphSAGE through aggregating neighborhood vertex information, wherein each aggregation is to aggregate the characteristics of each neighboring vertex obtained in the previous layer once, then combine the characteristics of the vertex in the previous layer to obtain the embedded characteristics of the layer, and repeatedly aggregate K times to obtain the last embedded characteristics of the vertex, wherein the vertex characteristics of the initial layer are the input sample characteristics;
s43, defining graph convolution:
wherein σ is a nonlinear activation function, W k Is a weight matrix to be learned for information propagation between different layers of a model, AG k Indicating the polymerization operation of the k-th layer,aggregation information for any neighboring vertex representing vertex v, < ->Representing the embedded features obtained from the layer above vertex v, and obtaining the final low-dimensional features obtained from the K-th layer is represented as +.>N (v) represents the set of domain points for vertex v, CONCAT () represents the join of the two matrices;
s44, designing a loss function:
wherein z is u The final embedded feature representation representing any vertex in the graph model, superscript T represents transpose, v represents vertex that has occurred with vertex u on a fixed length random walk, P n Is a negative sampling distribution, Q represents the number of negative samples, z v Representing the final low-dimensional features obtained from the acquisition of the K-th layer.
In one embodiment, implementing the unsupervised classification of hyperspectral images using the K-means algorithm based on the low-dimensional feature representation includes:
s51, clustering low-dimensional features of the super pixel points by using a K-means algorithm to obtain a label matrix of the super pixel points;
s52, restoring the super pixel point to the original pixel point, and matching the clustering result with the real class in an optimal mode through a Hungary algorithm to realize the non-supervision classification of the hyperspectral image.
According to the hyperspectral image non-supervision classification method based on graph convolution network embedded characterization, spatial features and spectral features are combined to form spatial spectrum combined features, super-pixel segmentation is carried out, elastic network decomposition is carried out on each super-pixel point, a super-pixel graph model is built, the subsequent calculation complexity is reduced, the graph convolution network is utilized to learn an embedded method in deep learning, non-supervision classification learning is carried out based on the better embedded characterization of the graph convolution network, and the purpose of precise classification of hyperspectral images is achieved.
Drawings
FIG. 1 is a flow chart of a hyperspectral image unsupervised classification method based on graph convolutional network embedded characterization, according to one embodiment;
FIG. 2 is a schematic diagram of simulation results of an embodiment;
FIG. 3 is a schematic diagram of simulation results of an embodiment.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more apparent, the present application will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the present application.
Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the present application. The appearances of such phrases in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. Those of skill in the art will explicitly and implicitly appreciate that the embodiments described herein may be combined with other embodiments.
Referring to fig. 1, fig. 1 is a flowchart of an embodiment of a hyperspectral image non-supervision classification method based on graph convolution network embedded characterization, which introduces a superpixel idea, performs elastic network decomposition on each superpixel point and builds a superpixel graph model, introduces a graph convolution network idea, processes the graph model through the graph convolution network, well learns characteristics of graph vertices and neighborhood thereof, learns a better embedded characterization, and further performs non-supervision method classification based on the graph convolution network embedded characterization, thereby obtaining a better classification result, and specifically comprises the following steps:
s10, sequentially extracting EMP and spectral features of the image to be hyperspectral to obtain spatial spectrum combined features.
S20, performing super-pixel segmentation on the spatial joint features to obtain super-pixel points of the hyperspectral image to be subjected to hyperspectral.
S30, solving the solution elastic net representation of the super pixel points, taking the super pixel points associated with the non-zero components in the representation coefficients of the solution elastic net representation as the neighbors of the current points, and constructing a graph model of the super pixel points.
S40, performing graph convolutional network embedded representation learning based on the graph model, and obtaining low-dimensional features through hierarchical vertex aggregation operation.
S50, according to the low-dimensional characteristic representation, the non-supervision classification of the hyperspectral image is realized by using a K-means algorithm.
According to the hyperspectral image non-supervision classification method based on graph convolution network embedded characterization, spatial features and spectral features are combined to form spatial spectrum combined features, super-pixel segmentation is carried out, elastic network decomposition is carried out on each super-pixel point, a super-pixel graph model is built, the subsequent calculation complexity is reduced, the graph convolution network is utilized to learn an embedded method in deep learning, non-supervision classification learning is carried out based on the better embedded characterization of the graph convolution network, and the purpose of precise classification of hyperspectral images is achieved.
In one embodiment, sequentially performing EMP and spectral feature extraction on the image to be hyperspectral, and obtaining the spatial spectrum joint feature includes:
wherein V represents a spatial spectrum joint feature matrix, X represents a spectral feature matrix, EMP represents an EMP feature matrix, m represents the number of main components, N represents the number of circular structural elements with different radiuses, d represents the number of spectral bands, and N represents the number of samples.
In one embodiment, performing superpixel segmentation on the spatial joint feature to obtain a superpixel point of the image to be hyperspectral includes:
D i,c =(1-λ)×D spectral +λ×D spatial ,
D spectral =tan(SAD(x i ,x c )),
wherein D is spectral Inter-spectral distance, SAD (x) i ,x c ) Represents the angular distance of the spectrum, D spatial Representing normalized spatial Euclidean distance, r is the diagonal length of the search neighborhood, used to constrain the search range to a local neighborhood around each cluster center, D i,c Represents a weighted sum of spectral and spatial distances, lambda represents an acting parameter balancing spectral and spatial distances, x c =(x c1 ,x c2 ,…,x cn ) Representing a cluster center, which corresponds to a spatial coordinate of (cx, cy), x i =(x i1 ,x i2 ,…,x il ) The representation is located at the cluster center x c The local neighborhood of l pixels corresponds to spatial coordinates (ix, iy). Specifically, the parameter λ is set to be less than 0.5.
In this embodiment, the similarity calculation criteria in the super-pixel segmentation are:
D i,c =(1-λ)×D spectral +λ×D spatial ,
D spectral =tan(SAD(x i ,x c )),
wherein D is spectral Inter-spectral distance, SAD (x) i ,x c ) Represents the spectrum angle (Spectral Angle Distance, SAD) distance, and the value range is [0, pi ]]。D spatial And r is the diagonal length of the search neighborhood, and the search range is constrained to be a local neighborhood around the center of each cluster. D (D) i,c The parameter lambda balances the effect of spectral and spatial distances, x, representing the weighted sum of the spectral and spatial distances c =(x c1 ,x c2 ,…,x cn ) Representing a cluster center, which corresponds to a spatial coordinate of (cx, cy), x i =(x i1 ,x i2 ,…,x il ) The representation is located at the cluster center x c The local neighborhood of l pixels corresponds to spatial coordinates (ix, iy). The present embodiment suggests that the parameter λ is set to be less than 0.5.
Further, randomly initializing a cluster center, and judging that each pixel belongs to the cluster center with the shortest distance according to the distance between the cluster center and the local neighborhood pixel. The characteristic of each super pixel point is represented by the average value of the spatial spectrum joint characteristics of all the pixel points in the local neighborhood, and the label of each super pixel point is determined by the category with the largest label number of all the pixel points in the local neighborhood. The specific process may include: firstly randomly selecting a pixel point as a dinner center, selecting a part of pixel points with the shortest distance to belong to the center according to the calculated distance between the center and the neighborhood pixel points, calculating the average value of various points to obtain a new clustering center, and repeatedly iterating to determine the final center and the pixel points belonging to the center to form a super-pixel point.
In one embodiment, solving a solution elastic net representation of a super pixel point, taking the super pixel point associated with a non-zero component in a representation coefficient of the solution elastic net representation as a neighbor of a current point, and constructing a graph model of the super pixel point includes:
s31, selecting all other super-pixel points based on each super-pixel point to construct a dictionary, and finding out elastic net representations of all the pixel points in the data set by solving the following constraint optimization problem:
s.t.sx i =SD i sc i +e i ,
wherein SD is i Is a dictionary formed by all super-pixel points, sc i Is super pixel point sx i Dictionary based SD i Obtained representation coefficient, sc= [ SC ] 1 ,sc 2 ,…,sc N ]Is a coefficient matrix, E is a characterization error matrix, lambda and gamma are regularization parameters, E i Is an error vector;
s32, constructing super pixels according to the elastic net sparse representation coefficient of each sample pointElastic net graph model of points, defined according to elastic net representation coefficient matrix SCAnd establishing edge connection between hyperspectral pixel points as an adjacency matrix of the graph model to obtain the graph model of the super pixel points.
In one embodiment, graph convolutional network embedded token learning based on a graph model, obtaining low-dimensional features through hierarchical vertex aggregation operations includes:
s41, layering (K-layer) sampling is carried out on each vertex of the graph model, and graph SAGE (SAmple and aggreGatE, sampling and aggregation of the graph) randomly samples a fixed number of neighborhood vertices at each layer of the graph model in a random walk mode, and for vertices which are not sampled, the neighborhood vertices are approximated by historical expressions of the neighborhood vertices;
s42, updating self information by the GraphSAGE through aggregating neighborhood vertex information, wherein each aggregation is to aggregate the characteristics of each neighboring vertex obtained in the previous layer once, then combine the characteristics of the vertex in the previous layer to obtain the embedded characteristics of the layer, and repeatedly aggregate K times to obtain the last embedded characteristics of the vertex, wherein the vertex characteristics of the initial layer are the input sample characteristics; wherein the vertex features of the initial layer are the input sample features. The present invention attempts two polymerization modes: average polymerization and pooling polymerization;
s43, defining graph convolution:
wherein σ is a nonlinear activation function, W k Is a weight matrix to be learned for information propagation between different layers of a model, AG k Indicating the polymerization operation of the k-th layer,aggregation information for any neighboring vertex representing vertex v, < ->Representing the embedded features obtained from the layer above vertex v, and obtaining the final low-dimensional features obtained from the K-th layer is represented as +.>N (v) represents the set of domain points for vertex v, CONCAT () represents the join of the two matrices;
s44, designing a loss function to obtain a low-dimensional characteristic according to the graph convolution and the loss function, wherein the loss function specifically comprises:
wherein z is u The final embedded feature representation representing any vertex in the graph model, superscript T represents transpose, v represents vertex that has occurred with vertex u on a fixed length random walk, P n Is a negative sampling distribution, Q represents the number of negative samples, z v Representing the final low-dimensional features obtained from the acquisition of the K-th layer.
In one example, the vertex features of the initial layer are the sample features of the input, and this example attempts two aggregation approaches: average polymerization and pooling polymerization.
Average aggregation is achieved by solving for the average value of the last layer of embedding of the neighborhood vertices, defined as follows:
wherein N (v) represents the neighbor of vertex v,aggregation information representing any neighboring vertex of vertex v.
And vector sharing weights of all adjacent vertexes in pooling aggregation are firstly processed through a nonlinear full-connection layer, and then information of a neighborhood is aggregated by maximum pooling, so that more effective embedded features are obtained. The specific formula is as follows:
where max is the maximum operator and δ is a nonlinear activation function. In principle, each neighborhood vertex can be independently passed through a multi-layer perceptron of arbitrary depth to obtain a vector, and the multi-layer perceptron can be regarded as a group of functions W pool The embedded features for each neighborhood vertex are computed.
In one embodiment, implementing the unsupervised classification of hyperspectral images using the K-means algorithm based on the low-dimensional feature representation includes:
s51, clustering low-dimensional features of the super pixel points by using a K-means algorithm to obtain a label matrix of the super pixel points;
s52, restoring the super pixel point to the original pixel point, and matching the clustering result with the real class in an optimal mode through a Hungary algorithm to realize the non-supervision classification of the hyperspectral image.
According to the hyperspectral image non-supervision classification method based on the graph rolling network embedded representation, the graph rolling network is utilized to establish a hyperspectral image non-supervision classification model based on the graph rolling network embedded representation, and the model combines spatial features and spectral features on feature representation to form spatial spectrum joint features. And super-pixel segmentation is carried out, so that elastic net decomposition is carried out on each super-pixel point, and a super-pixel graph model is constructed, and the subsequent calculation complexity is reduced. By utilizing the graph rolling network learning to learn an embedding method in deep learning, non-supervision classification learning is performed based on better embedding characterization of the graph rolling network, and the aim of accurately classifying hyperspectral images is fulfilled.
In one embodiment, to verify the effect of the above-mentioned hyperspectral image non-supervised classification method based on graph convolution network embedded characterization, simulation experiments are performed, indian pins-13 (IP-13) and S-Pavia University Scene (S-PUS) test sequence specifications are 145×145 and 306×340, respectively, a super-pixel segmentation related parameter λ=0.3 is set, a graph convolution network graph SAGE sampling layer number k=2 is set, and a neighborhood sampling layer number k=2 is setThe number is set as S 1 =25 and S 2 =10, 25 first order neighbors, 10 second order neighbors, 50 random walks with a step size of 5 are performed for each vertex, and 20 negative samples are sampled for each node with reference to word2 vec. The Batchsize is set to 512, epochs is set to 5, the weight attenuation is set to 0.0001, the implementation is carried out under a TensorFlow platform, an Adam optimizer is selected, the dimensions of the output embedded representation are all C+1, and C is the corresponding data set category number.
The evaluation of the experiment uses both qualitative and quantitative analysis methods.
As can be seen from a comparison of the classification effect of the present invention and each algorithm on the hyperspectral image, the classification effect of the present invention is significantly better than other algorithms for different hyperspectral image datasets.
For quantitative comparative analysis, OA, AA and were used for evaluation. Where OA is the Overall Accuracy (over all Accuracy) of all sample classifications, AA is the Average Accuracy (Average Accuracy) of all types of sample classifications, and Kappa coefficients are calculated as follows:
wherein c is the number of sample categories, m ii Indicating the number of samples of the ith class divided into the ith class in the classification process, N is the total number of samples, p i Representing the accuracy of each class of sample classification, N i Representing the total number of samples of class i.
After the model is used for classification, the method plays a great role in the classification of hyperspectral images, and the method obtains better low-dimensional characteristics by using the graph rolling network for learning and embedding, so that the classification accuracy is improved.
In quantitative comparison, we classify on two hyperspectral image datasets, compare the classification result of each dataset with groundtrunk, calculate the corresponding OA, AA and value. FIGS. 2 and 3 show the OA, AA and values of the algorithm of the present invention and other algorithms in the data sets Indian pins-13 (IP-13) and S-Pavia University Scene (S-PUS), respectively.
As can be seen from the summary, compared with the conventional patterning method, in this embodiment, super-pixel segmentation is introduced, and super-pixels are used to replace a plurality of pixels in a local neighborhood, so that redundant information is removed, and block patterning is avoided, so that the number of vertices is greatly reduced, and the complexity of patterning is reduced. In addition, the embodiment utilizes the graph convolution network to process the graph model, is a deep embedding method on the graph model, can well learn the characteristics of the graph vertexes and the neighborhood thereof, obtains better embedded characterization, and further improves the accuracy of subsequent clustering. The algorithm of the present embodiment presents certain advantages both from the point of view of classification accuracy and from the point of view of visual effect.
The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.
It should be noted that, the term "first\second\third" in the embodiments of the present application is merely to distinguish similar objects, and does not represent a specific order for the objects, and it is understood that "first\second\third" may interchange a specific order or sequence where allowed. It is to be understood that the "first\second\third" distinguishing objects may be interchanged where appropriate to enable embodiments of the present application described herein to be implemented in sequences other than those illustrated or described herein.
The terms "comprising" and "having" and any variations thereof, in embodiments of the present application, are intended to cover non-exclusive inclusions. For example, a process, method, apparatus, article, or device that comprises a list of steps or modules is not limited to the particular steps or modules listed and may optionally include additional steps or modules not listed or inherent to such process, method, article, or device.
The above examples merely represent a few embodiments of the present application, which are described in more detail and are not to be construed as limiting the scope of the invention. It should be noted that it would be apparent to those skilled in the art that various modifications and improvements could be made without departing from the spirit of the present application, which would be within the scope of the present application. Accordingly, the scope of protection of the present application is to be determined by the claims appended hereto.
Claims (6)
1. The hyperspectral image non-supervision classification method based on graph rolling network embedded characterization is characterized by comprising the following steps of:
s10, sequentially extracting EMP and spectral features of the image to be hyperspectral to obtain spatial spectrum combined features;
s20, performing super-pixel segmentation on the spatial joint characteristics to obtain super-pixel points of the hyperspectral image to be subjected to hyperspectral;
s30, solving an elastic net representation of the super-pixel points, and constructing a graph model of the super-pixel points by taking the super-pixel points associated with non-zero components in the representation coefficients of the elastic net representation as the neighbors of the current points;
s40, performing graph convolutional network embedded characterization learning based on a graph model, and obtaining low-dimensional features through hierarchical vertex aggregation operation;
s50, according to the low-dimensional characteristic representation, the non-supervision classification of the hyperspectral image is realized by using a K-means algorithm.
2. The hyperspectral image non-supervision classification method based on graph rolling network embedded characterization according to claim 1, wherein the steps of sequentially carrying out EMP and spectral feature extraction on the hyperspectral image to obtain spatial spectrum joint features comprise:
wherein V represents a spatial spectrum joint feature matrix, X represents a spectral feature matrix, EMP represents an EMP feature matrix, m represents the number of main components, N represents the number of circular structural elements with different radiuses, d represents the number of spectral bands, and N represents the number of samples.
3. The hyperspectral image non-supervised classification method based on graph rolling network embedded representation as claimed in claim 1, wherein the super-pixel segmentation of the spatial joint features to obtain super-pixel points of the hyperspectral image comprises:
D i,c =(1-λ)×D spectral +λ×D spatial ,
D spectral =tan(SAD(x i ,x c )),
wherein D is spectral Inter-spectral distance, SAD (x) i ,x c ) Represents the angular distance of the spectrum, D spatial Representing normalized spatial Euclidean distance, r is the diagonal length of the search neighborhood, used to constrain the search range to a local neighborhood around each cluster center, D i,c Represents a weighted sum of spectral and spatial distances, lambda represents an acting parameter balancing spectral and spatial distances, x c =(x c1 ,x c2 ,…,x cn ) Representing a cluster center, which corresponds to a spatial coordinate of (cx, cy), x i =(x i1 ,x i2 ,…,x il ) The representation is located at the cluster center x c The local neighborhood of l pixels corresponds to spatial coordinates (ix, iy).
4. The hyperspectral image non-supervised classification method based on graph rolling network embedded representation as claimed in claim 1, wherein the solving of the solution elastic net representation of the super-pixel points, taking the super-pixel points associated with non-zero components in the representation coefficients of the solution elastic net representation as the neighbors of the current points, and constructing the graph model of the super-pixel points comprises:
s31, selecting all other super-pixel points based on each super-pixel point to construct a dictionary, and finding out elastic net representations of all the pixel points in the data set by solving the following constraint optimization problem:
s.t.sx i =SD i sc i +e i ,
wherein SD is i Is a dictionary formed by all super-pixel points, sc i Is super pixel point sx i Dictionary based SD i Obtained representation coefficient, sc= [ SC ] 1 ,sc 2 ,…,sc N ]Is a coefficient matrix, E is a characterization error matrix, lambda and gamma are regularization parameters, E i Is an error vector;
s32, constructing an elastic net graph model of the super-pixel point according to the elastic net sparse representation coefficient of each sample point, and defining according to an elastic net representation coefficient matrix SCAnd establishing edge connection between hyperspectral pixel points as an adjacency matrix of the graph model to obtain the graph model of the super pixel points.
5. The method for unsupervised classification of hyperspectral images based on graph-convolution network embedded representation as claimed in claim 1, wherein performing graph-convolution network embedded representation learning based on a graph model, obtaining low-dimensional features through hierarchical vertex aggregation operation comprises:
s41, sampling each vertex of the graph model in a layering manner, randomly sampling a fixed number of neighborhood vertices at each layer of the graph model by using a random walk mode, and approximating the vertices which are not sampled by using a historical expression of the vertices;
s42, updating self information by the GraphSAGE through aggregating neighborhood vertex information, wherein each aggregation is to aggregate the characteristics of each neighboring vertex obtained in the previous layer once, then combine the characteristics of the vertex in the previous layer to obtain the embedded characteristics of the layer, and repeatedly aggregate K times to obtain the last embedded characteristics of the vertex, wherein the vertex characteristics of the initial layer are the input sample characteristics;
s43, defining graph convolution:
wherein σ is a nonlinear activation function, W k Is a weight matrix to be learned for information propagation between different layers of a model, AG k Indicating the polymerization operation of the k-th layer,aggregation information for any neighboring vertex representing vertex v, < ->Representing the embedded features obtained from the layer above vertex v, and obtaining the final low-dimensional features obtained from the K-th layer is represented as +.>N (v) represents the set of domain points for vertex v, CONCAT () represents the join of the two matrices;
s44, designing a loss function:
wherein z is u The final embedded feature representation representing any vertex in the graph model, superscript T represents transpose, v represents vertex that has occurred with vertex u on a fixed length random walk, P n Is a negative sampling distribution, Q represents the number of negative samples, z v Representation acquisition of the firstThe final low dimensional feature obtained for the K layers.
6. The method for unsupervised classification of hyperspectral images based on embedded characterization of a graph rolling network as claimed in claim 1, wherein implementing unsupervised classification of hyperspectral images using K-means algorithm based on low-dimensional feature representation comprises:
s51, clustering low-dimensional features of the super pixel points by using a K-means algorithm to obtain a label matrix of the super pixel points;
s52, restoring the super pixel point to the original pixel point, and matching the clustering result with the real class in an optimal mode through a Hungary algorithm to realize the non-supervision classification of the hyperspectral image.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011124146.9A CN112329818B (en) | 2020-10-20 | 2020-10-20 | Hyperspectral image non-supervision classification method based on graph convolution network embedded characterization |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011124146.9A CN112329818B (en) | 2020-10-20 | 2020-10-20 | Hyperspectral image non-supervision classification method based on graph convolution network embedded characterization |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112329818A CN112329818A (en) | 2021-02-05 |
CN112329818B true CN112329818B (en) | 2023-07-07 |
Family
ID=74312064
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011124146.9A Active CN112329818B (en) | 2020-10-20 | 2020-10-20 | Hyperspectral image non-supervision classification method based on graph convolution network embedded characterization |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112329818B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113298129B (en) * | 2021-05-14 | 2024-02-02 | 西安理工大学 | Polarized SAR image classification method based on superpixel and graph convolution network |
CN113628289B (en) * | 2021-07-21 | 2023-10-27 | 武汉大学 | Hyperspectral image nonlinear unmixing method and system based on graph convolution self-encoder |
CN117853769B (en) * | 2024-01-08 | 2024-10-18 | 西安交通大学 | Hyperspectral remote sensing image rapid clustering method based on multi-scale image fusion |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104392454A (en) * | 2014-12-03 | 2015-03-04 | 复旦大学 | Merging method of membership scoring based on ground object categories under spatial-spectral combined classification frame for hyper-spectral remote sensing images |
CN108009559A (en) * | 2016-11-02 | 2018-05-08 | 哈尔滨工业大学 | A kind of Hyperspectral data classification method based on empty spectrum united information |
CN110111338A (en) * | 2019-04-24 | 2019-08-09 | 广东技术师范大学 | A kind of visual tracking method based on the segmentation of super-pixel time and space significance |
CN110363236A (en) * | 2019-06-29 | 2019-10-22 | 河南大学 | The high spectrum image extreme learning machine clustering method of sky spectrum joint hypergraph insertion |
CN110399909A (en) * | 2019-07-08 | 2019-11-01 | 南京信息工程大学 | A kind of hyperspectral image classification method based on label constraint elastic network(s) graph model |
CN111126463A (en) * | 2019-12-12 | 2020-05-08 | 武汉大学 | Spectral image classification method and system based on local information constraint and sparse representation |
CN111695636A (en) * | 2020-06-15 | 2020-09-22 | 北京师范大学 | Hyperspectral image classification method based on graph neural network |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110717354B (en) * | 2018-07-11 | 2023-05-12 | 哈尔滨工业大学 | Super-pixel classification method based on semi-supervised K-SVD and multi-scale sparse representation |
CA3125790A1 (en) * | 2019-02-04 | 2020-08-13 | Farmers Edge Inc. | Shadow and cloud masking for remote sensing images in agriculture applications using multilayer perceptron |
-
2020
- 2020-10-20 CN CN202011124146.9A patent/CN112329818B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104392454A (en) * | 2014-12-03 | 2015-03-04 | 复旦大学 | Merging method of membership scoring based on ground object categories under spatial-spectral combined classification frame for hyper-spectral remote sensing images |
CN108009559A (en) * | 2016-11-02 | 2018-05-08 | 哈尔滨工业大学 | A kind of Hyperspectral data classification method based on empty spectrum united information |
CN110111338A (en) * | 2019-04-24 | 2019-08-09 | 广东技术师范大学 | A kind of visual tracking method based on the segmentation of super-pixel time and space significance |
CN110363236A (en) * | 2019-06-29 | 2019-10-22 | 河南大学 | The high spectrum image extreme learning machine clustering method of sky spectrum joint hypergraph insertion |
CN110399909A (en) * | 2019-07-08 | 2019-11-01 | 南京信息工程大学 | A kind of hyperspectral image classification method based on label constraint elastic network(s) graph model |
CN111126463A (en) * | 2019-12-12 | 2020-05-08 | 武汉大学 | Spectral image classification method and system based on local information constraint and sparse representation |
CN111695636A (en) * | 2020-06-15 | 2020-09-22 | 北京师范大学 | Hyperspectral image classification method based on graph neural network |
Non-Patent Citations (9)
Title |
---|
Spatial-Spectral Locality-Constrained Low-Rank Representation with Semi-Supervised Hypergraph Learning for Hyperspectral Image Classification;Sujuan Wang等;《INTERNATIONAL JOURNAL OF REMOTE SENSING》;第38卷(第23期);7374–7388 * |
Spatial-spectral neighbour graph for dimensionality reduction of hyperspectral image classification;Dongqing Li等;《International Journal of Remote Sensing》;1-24 * |
半监督多图嵌入的高光谱影像特征提取;黄鸿等;《光学精密工程》;第28卷(第02期);443-456 * |
基于卷积神经网络的高光谱图像谱-空联合分类;付光远等;《科学技术与工程》;第17卷(第21期);268-274 * |
基于图模型的高光谱图像分类;陈逸;《中国优秀硕士学位论文全文数据库 工程科技II辑》(第02期);C028-298 * |
基于多特征与改进子空间聚类的SAR图像分割;刘胜男;《中国优秀硕士学位论文全文数据库 信息科技辑》(第02期);I136-1435 * |
基于深度结构化学习的语义图像分割方法研究;丁福光;《中国优秀硕士学位论文全文数据库 信息科技辑》(第02期);I138-1308 * |
基于超像素分割的高光谱图像特征变换和分类算法研究;邓彬;《中国优秀硕士学位论文全文数据库 信息科技辑》(第07期);I138-882 * |
谱-空图嵌入的高光谱图像多核分类算法;郭志民等;《小型微型计算机系统》;第39卷(第11期);2545-2550 * |
Also Published As
Publication number | Publication date |
---|---|
CN112329818A (en) | 2021-02-05 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108573276B (en) | Change detection method based on high-resolution remote sensing image | |
CN110399909B (en) | Hyperspectral image classification method based on label constraint elastic network graph model | |
CN110443143B (en) | Multi-branch convolutional neural network fused remote sensing image scene classification method | |
Othman et al. | Domain adaptation network for cross-scene classification | |
CN112329818B (en) | Hyperspectral image non-supervision classification method based on graph convolution network embedded characterization | |
CN108052966B (en) | Remote sensing image scene automatic extraction and classification method based on convolutional neural network | |
Yuan et al. | Factorization-based texture segmentation | |
Awad et al. | Multicomponent image segmentation using a genetic algorithm and artificial neural network | |
CN110348399B (en) | Hyperspectral intelligent classification method based on prototype learning mechanism and multidimensional residual error network | |
CN107092870B (en) | A kind of high resolution image Semantic features extraction method | |
CN105740894B (en) | Semantic annotation method for hyperspectral remote sensing image | |
CN109325484B (en) | Flower image classification method based on background prior significance | |
CN107590515B (en) | Hyperspectral image classification method of self-encoder based on entropy rate superpixel segmentation | |
CN103927531B (en) | It is a kind of based on local binary and the face identification method of particle group optimizing BP neural network | |
CN109766858A (en) | Three-dimensional convolution neural network hyperspectral image classification method combined with bilateral filtering | |
Almogdady et al. | A flower recognition system based on image processing and neural networks | |
CN104732244B (en) | The Classifying Method in Remote Sensing Image integrated based on wavelet transformation, how tactful PSO and SVM | |
CN113627472B (en) | Intelligent garden leaf feeding pest identification method based on layered deep learning model | |
CN107832797B (en) | Multispectral image classification method based on depth fusion residual error network | |
CN102013017A (en) | Method for roughly sorting high-resolution remote sensing image scene | |
CN111695636A (en) | Hyperspectral image classification method based on graph neural network | |
Chi | Self‐organizing map‐based color image segmentation with k‐means clustering and saliency map | |
CN113239938A (en) | Hyperspectral classification method and system based on graph structure | |
CN115810149A (en) | High-resolution remote sensing image building extraction method based on superpixel and image convolution | |
CN109344777A (en) | The Optimum Classification method of target in hyperspectral remotely sensed image land use covering based on ELM |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |