阅读说明：本技术 一种高光谱遥感图像分类方法及系统 (Hyperspectral remote sensing image classification method and system ) 是由 何芳 贾维敏 张峰干 伍宗伟 沈晓卫 赵建伟 胡豪杰 金伟 何佑明 朱玉杰 于 2020-07-15 设计创作，主要内容包括：本发明涉及一种高光谱遥感图像分类方法及系统。该方法包括获取高光谱遥感图像；根据所述高光谱遥感图像的像素点确定锚点；根据所述高光谱遥感图像的像素点和所述锚点,采用自适应邻近分配原则,构建自适应二部图；根据所述自适应二部图,采用半监督学习方法,构建二部图的半监督学习模型；获取待分类的高光谱遥感图像；采用所述二部图的半监督学习模型对所述待分类的高光谱遥感图像进行分类。本发明所提供的一种高光谱遥感图像分类方法及系统,能够降低计算复杂度,提高分类精度。(The invention relates to a hyperspectral remote sensing image classification method and a hyperspectral remote sensing image classification system. The method comprises the steps of obtaining a hyperspectral remote sensing image; determining an anchor point according to the pixel point of the hyperspectral remote sensing image; constructing a self-adaptive bipartite graph by adopting a self-adaptive adjacent distribution principle according to the pixel points and the anchor points of the hyperspectral remote sensing image; according to the self-adaptive bipartite graph, a semi-supervised learning method is adopted to construct a semi-supervised learning model of the bipartite graph; acquiring a hyperspectral remote sensing image to be classified; and classifying the hyperspectral remote sensing images to be classified by adopting a semi-supervised learning model of the bipartite graph. The hyperspectral remote sensing image classification method and the hyperspectral remote sensing image classification system can reduce the calculation complexity and improve the classification precision.)

1. A hyperspectral remote sensing image classification method is characterized by comprising the following steps:

acquiring a hyperspectral remote sensing image;

determining an anchor point according to the pixel point of the hyperspectral remote sensing image; the anchor points are pixel points of the hyperspectral remote sensing images selected randomly; the number of the anchor points is less than the number of pixel points in the hyperspectral remote sensing image;

constructing a self-adaptive bipartite graph by adopting a self-adaptive adjacent distribution principle according to the pixel points and the anchor points of the hyperspectral remote sensing image;

according to the self-adaptive bipartite graph, a semi-supervised learning method is adopted to construct a semi-supervised learning model of the bipartite graph; the semi-supervised learning model of the bipartite graph takes a self-adaptive bipartite graph as input and takes the category of the hyperspectral remote sensing image as output;

acquiring a hyperspectral remote sensing image to be classified;

and classifying the hyperspectral remote sensing images to be classified by adopting a semi-supervised learning model of the bipartite graph.

2. The hyperspectral remote sensing image classification method according to claim 1, wherein the self-adaptive bipartite graph is constructed by adopting a self-adaptive proximity distribution principle according to pixel points and anchor points of the hyperspectral remote sensing image, and specifically comprises the following steps:

constructing a similarity matrix by adopting a self-adaptive adjacent distribution principle according to the pixel points of the hyperspectral remote sensing image and the anchor points;

and constructing an adaptive bipartite graph according to the similarity matrix.

3. The hyperspectral remote sensing image classification method according to claim 1, wherein a semi-supervised learning model of the bipartite graph is constructed according to the self-adaptive bipartite graph by adopting a semi-supervised learning method, and the method further comprises the following steps:

and optimizing the self-adaptive bipartite graph by adopting a semi-supervised learning target.

4. The hyperspectral remote sensing image classification method according to claim 1, wherein a semi-supervised learning model of the bipartite graph is constructed according to the self-adaptive bipartite graph by adopting a semi-supervised learning method, and the method specifically comprises the following steps:

using formulasDetermining an objective function; z is a similarity matrix, U is a new similarity matrix,

solving the objective function by adopting an iterative optimization mode to obtain a label of a pixel point; the label is used for classifying the hyperspectral remote sensing images.

5. A hyperspectral remote sensing image classification system is characterized by comprising:

the hyperspectral remote sensing image acquisition module is used for acquiring a hyperspectral remote sensing image;

the anchor point determining module is used for determining an anchor point according to the pixel point of the hyperspectral remote sensing image; the anchor points are pixel points of the hyperspectral remote sensing images selected randomly; the number of the anchor points is less than the number of pixel points in the hyperspectral remote sensing image;

the self-adaptive bipartite graph construction module is used for constructing a self-adaptive bipartite graph by adopting a self-adaptive adjacent distribution principle according to the pixel points of the hyperspectral remote sensing image and the anchor points;

the semi-supervised learning model building module of the bipartite graph is used for building a semi-supervised learning model of the bipartite graph by adopting a semi-supervised learning method according to the self-adaptive bipartite graph; the semi-supervised learning model of the bipartite graph takes a self-adaptive bipartite graph as input and takes the category of the hyperspectral remote sensing image as output;

the hyperspectral remote sensing image classification module is used for classifying hyperspectral remote sensing images;

and the classification module is used for classifying the hyperspectral remote sensing images to be classified by adopting a semi-supervised learning model of the bipartite graph.

6. The hyperspectral remote sensing image classification system according to claim 5, wherein the adaptive bipartite graph construction module specifically comprises:

the similarity matrix construction unit is used for constructing a similarity matrix according to the pixel points of the hyperspectral remote sensing image and the anchor points by adopting a self-adaptive adjacent distribution principle;

and the self-adaptive bipartite graph constructing unit is used for constructing the self-adaptive bipartite graph according to the similarity matrix.

7. The hyperspectral remote sensing image classification system according to claim 5, further comprising:

and the optimization module is used for optimizing the self-adaptive bipartite graph by adopting a semi-supervised learning target.

8. The hyperspectral remote sensing image classification system according to claim 5, wherein the semi-supervised learning model construction module of the bipartite graph specifically comprises:

an objective function determination unit for utilizing a formulaDetermining an objective function; z is a similarity matrix, U is a new similarity matrix,

the label determining unit of the pixel point is used for solving the objective function in an iterative optimization mode to obtain a label of the pixel point; the label is used for classifying the hyperspectral remote sensing images.

Technical Field

The invention relates to the field of hyperspectral remote sensing and machine learning, in particular to a hyperspectral remote sensing image classification method and system.

Background

The hyperspectral remote sensing, namely hyperspectral resolution remote sensing, as a narrow-band imaging mode, can find substances which cannot be detected in a wide band, is popular in the 80 th century, and is an important earth observation technology. The hyperspectral remote sensing technology provides help for the actual life of people, and simultaneously brings out a series of problems of information extraction and pattern recognition, and is mainly reflected in high-dimensional data processing and analysis. With the development of imaging spectroscopy, higher spectral resolution brings more spectral band numbers, and wider coverage brings more data volume. Therefore, the hyperspectral technology provides rich spectral information and simultaneously provides a new challenge for hyperspectral data processing.

Classification is an important field of hyperspectral data processing. The hyperspectral image classification is that each pixel point in an image is endowed with a category label according to a certain classification criterion, such as 'class of things' by utilizing spectral information and spatial information of ground objects. The hyperspectral classification technology has important effects on extracting thematic information and monitoring the dynamic change of ground objects, and is widely applied to the fields of thematic map making, engineering exploration, traffic planning management, environment monitoring, land utilization, crop yield estimation and the like.

The original remote sensing image classification method is a manual visual interpretation method, which has higher requirements on the geography knowledge and the research and judgment experience of workers, and the classification result is greatly influenced by the experience and knowledge reserve of the workers. The manual visual efficiency is low, and large manpower and material resources are consumed. Due to the rapid increase in the amount of remote sensing data, manual visual methods have not been able to meet the demand. The nature of the classification problem is pattern recognition. With the continuous development of computer technology, intelligent data analysis can be realized by using a machine learning method, and a hidden relation between data is obtained. Various classification algorithms generated based on machine learning improve the classification effect of the hyperspectral images to a certain extent.

Machine Learning is a core subject in the fields of artificial intelligence and pattern recognition, and can be classified into Supervised Learning (super Learning), Unsupervised Learning (Unsupervised Learning), and Semi-Supervised Learning (SSL) according to whether class label information is used. The supervised learning carries out iterative computation by taking a probability function, an algebraic function or an artificial neural network as a basic function model by means of class mark information of input data. The unsupervised learning does not use class label information and is mainly applied to the clustering of data. Semi-supervised learning builds a learning model using a small amount of labeled information and a large amount of unlabeled information. The high-spectrum data volume is large, and the marking of the sample needs to consume large manpower and material resources, so that the semi-supervised learning method has important application in the high-spectrum image processing.

The semi-supervised learning method based on the graph has clear concept and easy realization, and has attracted wide attention in recent years. However, the conventional mapping method has high computational complexity in the composition process and the matrix inversion process, and cannot process large-scale hyperspectral images. In addition, the conventional graph method generally inputs a fixed graph, and the quality of the graph directly affects the subsequent classification effect.

Disclosure of Invention

The invention aims to provide a hyperspectral remote sensing image classification method and a hyperspectral remote sensing image classification system, which can reduce the calculation complexity and improve the classification precision.

In order to achieve the purpose, the invention provides the following scheme:

a hyperspectral remote sensing image classification method comprises the following steps:

acquiring a hyperspectral remote sensing image;

determining an anchor point according to the pixel point of the hyperspectral remote sensing image; the anchor points are pixel points of the hyperspectral remote sensing images selected randomly; the number of the anchor points is less than the number of pixel points in the hyperspectral remote sensing image;

constructing a self-adaptive bipartite graph by adopting a self-adaptive adjacent distribution principle according to the pixel points and the anchor points of the hyperspectral remote sensing image;

according to the self-adaptive bipartite graph, a semi-supervised learning method is adopted to construct a semi-supervised learning model of the bipartite graph; the semi-supervised learning model of the bipartite graph takes a self-adaptive bipartite graph as input and takes the category of the hyperspectral remote sensing image as output;

acquiring a hyperspectral remote sensing image to be classified;

and classifying the hyperspectral remote sensing images to be classified by adopting a semi-supervised learning model of the bipartite graph.

Optionally, the constructing a self-adaptive bipartite graph according to the pixel points of the hyperspectral remote sensing image and the anchor points by adopting a self-adaptive proximity allocation principle specifically includes:

constructing a similarity matrix by adopting a self-adaptive adjacent distribution principle according to the pixel points of the hyperspectral remote sensing image and the anchor points;

and constructing an adaptive bipartite graph according to the similarity matrix.

Optionally, the constructing a semi-supervised learning model of the bipartite graph by using a semi-supervised learning method according to the adaptive bipartite graph further includes:

and optimizing the self-adaptive bipartite graph by adopting a semi-supervised learning target.

Optionally, the constructing a semi-supervised learning model of the bipartite graph by using a semi-supervised learning method according to the adaptive bipartite graph specifically includes:

using formulas

Determining an objective function; z is a similarity matrix, U is a new similarity matrix,is a matrix of soft labels of the pixels,

is a soft tag matrix for an anchor point,

label matrix representing all pixel points, B is diagonal matrix, alpha is regularization parameter, L_SIs a Laplace matrix;

solving the objective function by adopting an iterative optimization mode to obtain a label of a pixel point; the label is used for classifying the hyperspectral remote sensing images.

A hyperspectral remote sensing image classification system comprising:

the hyperspectral remote sensing image acquisition module is used for acquiring a hyperspectral remote sensing image;

the anchor point determining module is used for determining an anchor point according to the pixel point of the hyperspectral remote sensing image; the anchor points are pixel points of the hyperspectral remote sensing images selected randomly; the number of the anchor points is less than the number of pixel points in the hyperspectral remote sensing image;

the self-adaptive bipartite graph construction module is used for constructing a self-adaptive bipartite graph by adopting a self-adaptive adjacent distribution principle according to the pixel points of the hyperspectral remote sensing image and the anchor points;

the semi-supervised learning model building module of the bipartite graph is used for building a semi-supervised learning model of the bipartite graph by adopting a semi-supervised learning method according to the self-adaptive bipartite graph; the semi-supervised learning model of the bipartite graph takes a self-adaptive bipartite graph as input and takes the category of the hyperspectral remote sensing image as output;

the hyperspectral remote sensing image classification module is used for classifying hyperspectral remote sensing images;

and the classification module is used for classifying the hyperspectral remote sensing images to be classified by adopting a semi-supervised learning model of the bipartite graph.

Optionally, the adaptive bipartite graph building module specifically includes:

the similarity matrix construction unit is used for constructing a similarity matrix according to the pixel points of the hyperspectral remote sensing image and the anchor points by adopting a self-adaptive adjacent distribution principle;

and the self-adaptive bipartite graph constructing unit is used for constructing the self-adaptive bipartite graph according to the similarity matrix.

Optionally, the method further includes:

and the optimization module is used for optimizing the self-adaptive bipartite graph by adopting a semi-supervised learning target.

Optionally, the module for constructing a semi-supervised learning model of the bipartite graph specifically includes:

an objective function determination unit for utilizing a formula

Determining an objective function; z is a similarity matrix, U is a new similarity matrix,

is a matrix of soft labels of the pixels,is a soft tag matrix for an anchor point,

label matrix representing all pixel points, B is diagonal matrix, alpha is regularization parameter, L_SIs a Laplace matrix;

the label determining unit of the pixel point is used for solving the objective function in an iterative optimization mode to obtain a label of the pixel point; the label is used for classifying the hyperspectral remote sensing images.

According to the specific embodiment provided by the invention, the invention discloses the following technical effects:

according to the method and the system for classifying the hyperspectral remote sensing images, the randomly selected pixel points of part of the hyperspectral remote sensing images are determined as anchor points, and the self-adaptive bipartite graph between the anchor points and the pixel points is constructed, so that the parameters of composition are reduced, and the complexity of calculation is reduced; according to the self-adaptive bipartite graph, a semi-supervised learning method is adopted to construct a semi-supervised learning model of the bipartite graph, and the constructed self-adaptive bipartite graph is continuously optimized, so that the quality of the graph is improved, and the classification precision is further improved.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.

FIG. 1 is a schematic flow chart of a hyperspectral remote sensing image classification method provided by the invention;

FIG. 2 is a schematic structural diagram of a hyperspectral remote sensing image classification system provided by the invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The invention aims to provide a hyperspectral remote sensing image classification method and a hyperspectral remote sensing image classification system, which can reduce the calculation complexity and improve the classification precision.

In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.

Fig. 1 is a schematic flow chart of a hyperspectral remote sensing image classification method provided by the invention, and as shown in fig. 1, the hyperspectral remote sensing image classification method provided by the invention comprises the following steps:

and S101, acquiring a hyperspectral remote sensing image. The pixel points of the hyperspectral remote sensing image are expressed as

n represents the number of picture elements and d represents the number of bands of each picture element.

S102, determining an anchor point according to a pixel point of the hyperspectral remote sensing image; the anchor points are pixel points of the hyperspectral remote sensing images selected randomly; the number of the anchor points is smaller than the number of the pixel points in the hyperspectral remote sensing image. That is, m anchor points are selected from original n pixel points and expressed as matrix

S103, constructing a self-adaptive bipartite graph by adopting a self-adaptive adjacent distribution principle according to the pixel points of the hyperspectral remote sensing image and the anchor points. The self-adaptive neighbor allocation principle is that the closer the distance between two points is, the higher the probability of belonging to the same class is.

S103 specifically comprises the following steps:

and constructing a similarity matrix Z by adopting a self-adaptive adjacent distribution principle according to the pixel points of the hyperspectral remote sensing image and the anchor points.

Z is:

wherein z is_ijFor the ith row and jth column element in the similarity matrix Z,

in the ith row of the similarity matrix Z, γ is a regularization parameter. Order to Is a vector whose j-th element is e_ijThus, equation (1) is written in the form of a vector as follows:

the lagrangian function for equation (2) is:

wherein, eta and beta_iA value of > 0 is the Lagrangian multiplier.

z_iShould satisfy equation (3) with respect to z_iIs equal to 0, i.e.:

for z_iThe jth element of (1) has:

according to the KKT condition, z_ijβ_ij0, obtainable from formula (5):

wherein

Thus, z_ijThe solution of (a) is:

and constructing an adaptive bipartite graph according to the similarity matrix.

The self-adaptive bipartite graph is as follows:

s104, constructing a semi-supervised learning model of the bipartite graph by adopting a semi-supervised learning method according to the self-adaptive bipartite graph; the semi-supervised learning model of the bipartite graph takes a self-adaptive bipartite graph as input and takes the category of the hyperspectral remote sensing image as output;

before S104, the method further includes:

and optimizing the self-adaptive bipartite graph by adopting a semi-supervised learning target.

The goal is to learn a new similarity matrixOrThe following were used:

the closer the optimized similarity matrix S is to the given neighbor matrix W, the better, so the following problem needs to be solved:

according to the particular structure of S and W in equations (8) and (9), equation (10) can be converted into:

s104 specifically comprises the following steps:

using formulas

Determining an objective function; z is a similarity matrix, U is a new similarity matrix,is a matrix of soft labels of the pixels,is a soft tag matrix for an anchor point,label matrix representing all pixel points, B is diagonal matrix, alpha is regularization parameter, L_SIs a laplacian matrix.

Solving the objective function by adopting an iterative optimization mode to obtain a label of a pixel point; the label is used for classifying the hyperspectral remote sensing images.

The specific determination process of the objective function is as follows:

considering that the quality of the graph is optimal in graph-based semi-supervised learning, the following model is proposed:

wherein the content of the first and second substances,soft label matrices representing the original data and anchor points, respectively.

A label matrix representing all data (raw data and anchor points). Since the training sample is a portion of the data points selected from the raw data, Y can be written asWherein

B is a diagonal matrix, the ith element on the diagonal is a regularization parameter β_i. B can also be written as

α is a regularization parameter.

In the graph theory, L_S＝D_S-S is a Laplace matrix,is a degree matrix with the element on the ith diagonal being d_i＝∑_js_ij。D_SCan also be written asWhereinIs a diagonal matrix, the elements on the diagonal are the rows and columns of U,also a diagonal matrix, the elements on the diagonal are the column sums of U,

since the constraint condition in the formula (12) isCan obtain D_r＝I_nWhereinIs an identity matrix. Therefore, the temperature of the molten metal is controlled,then, L_SNormalization can be performed in the following manner:

wherein the content of the first and second substances,is an identity matrix.

Therefore, the final objective function after normalization of the laplace matrix is:

the process of solving equation (14) by using the iterative optimization method is as follows:

when F, G are fixed, equation (14) is equivalent to:

from the basic properties of the normalized laplace matrix, the following relationship can be obtained:

wherein f is_iIs the ith row of F, d_i＝∑_js_ij，g_jIs row i of G, d_j＝∑_is_ji。

According to the particular structure of S in equation (9), equation (16) is equivalent to:

thus, equation (15) can also be written as:

since the equation (18) is independent for different i, the objective function corresponding to each i can be solved. Order tov_i、u_i、z_iRepresenting vectors in which the jth element is v_ij、u_ijAnd z_ij. Thus, for each i, equation (18) can be written in the form of a vector as follows:

equation (19) has the same form as equation (2) and can be solved in the same way.

When U is fixed, equation (14) is equivalent to:

order to

Equation (20) may be determined as:

by taking the derivative of j (q), the following equation can be obtained:

thus, the final solution is:

B_αcan be written as

Wherein

Formula (23) is equivalent to:

let L₁₁＝I_n+B_αn，L₂₂＝B_αm+I_mThe first term in equation (24) is solved using the following block matrix inversion formula:

wherein the content of the first and second substances,due to the fact thatTo findHas a computational complexity of O (n)³) For large-scale hyperspectral data, the calculation amount is too large, and therefore, the following Woodbury matrix (A-UCV) is adopted^-1＝A^-1+A^-1U(C^-1-VA^-1U)^-1VA^-1Solving the large-scale matrix C₁Solving the inverse problem, reducing the computational complexity to O (nm)²)。

Based on the above derivation, we can obtain the final soft label matrix as:

obtaining a data point x according to the soft label matrix F_iThe labels of (a) are:

the computational complexity of the whole model is O (ndmt + nm)²) While the computation complexity of the traditional graph-based semi-supervised learning model is O (n)²d+n³). Wherein n, m, d and t are the number of samples, the number of anchor points, the dimension and the number of iterations respectively. Therefore, the classification can be rapidly and accurately carried out on processing large-scale hyperspectral data.

And S105, acquiring the hyperspectral remote sensing image to be classified.

And S106, classifying the hyperspectral remote sensing images to be classified by adopting a semi-supervised learning model of the bipartite graph.

Fig. 2 is a schematic structural diagram of a hyperspectral remote sensing image classification system provided by the invention, and as shown in fig. 2, the hyperspectral remote sensing image classification system provided by the invention comprises: the system comprises a hyperspectral remote sensing image acquisition module 201, an anchor point determination module 202, a self-adaptive bipartite graph construction module 203, a bipartite graph semi-supervised learning model construction module 204, a hyperspectral remote sensing image acquisition module 205 to be classified and a classification module 206.

The hyperspectral remote sensing image acquisition module 201 is used for acquiring a hyperspectral remote sensing image.

The anchor point determining module 202 is configured to determine an anchor point according to a pixel point of the hyperspectral remote sensing image; the anchor points are pixel points of the hyperspectral remote sensing images selected randomly; the number of the anchor points is smaller than the number of the pixel points in the hyperspectral remote sensing image.

The self-adaptive bipartite graph construction module 203 is used for constructing a self-adaptive bipartite graph according to the pixel points of the hyperspectral remote sensing image and the anchor points by adopting a self-adaptive proximity distribution principle.

The semi-supervised learning model construction module 204 of the bipartite graph is used for constructing a semi-supervised learning model of the bipartite graph by adopting a semi-supervised learning method according to the self-adaptive bipartite graph; the semi-supervised learning model of the bipartite graph takes the self-adaptive bipartite graph as input and the category of the hyperspectral remote sensing image as output.

The hyperspectral remote sensing image to be classified acquisition module 205 is used for acquiring the hyperspectral remote sensing image to be classified.

The classification module 206 is configured to classify the to-be-classified hyperspectral remote sensing image by using a semi-supervised learning model of the bipartite graph.

The adaptive bipartite graph building module 203 specifically includes: a similarity matrix construction unit and an adaptive bipartite graph construction unit.

And the similarity matrix construction unit is used for constructing a similarity matrix according to the pixel points of the hyperspectral remote sensing image and the anchor points by adopting a self-adaptive adjacent distribution principle.

And the self-adaptive bipartite graph constructing unit is used for constructing a self-adaptive bipartite graph according to the similarity matrix.

The invention provides a hyperspectral remote sensing image classification system, which further comprises: and an optimization module.

The optimization module is used for optimizing the self-adaptive bipartite graph by adopting a semi-supervised learning objective.

The module 204 for constructing a semi-supervised learning model of a bipartite graph specifically comprises: an objective function determining unit and a label determining unit of the pixel point.

An objective function determination unit for utilizing the formulaDetermining an objective function; z is a similarity matrix, U is a new similarity matrix,is a matrix of soft labels of the pixels,is a soft tag matrix for an anchor point,label matrix representing all pixel points, B is diagonal matrix, alpha is regularization parameter, L_SIs a laplacian matrix.

The label determining unit of the pixel point is used for solving the objective function in an iterative optimization mode to obtain a label of the pixel point; the label is used for classifying the hyperspectral remote sensing images.

The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. For the system disclosed by the embodiment, the description is relatively simple because the system corresponds to the method disclosed by the embodiment, and the relevant points can be referred to the method part for description.

The principles and embodiments of the present invention have been described herein using specific examples, which are provided only to help understand the method and the core concept of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the above, the present disclosure should not be construed as limiting the invention.