CN111695636A - Hyperspectral image classification method based on graph neural network - Google Patents

Abstract

The invention discloses a hyperspectral image classification method based on a graph neural network, comprising: S1: acquiring hyperspectral image data of a to-be-processed area, and preprocessing the hyperspectral image data; S2: processing all preprocessed hyperspectral image data Perform superpixel segmentation on the hyperspectral image data to obtain segmented superpixels; S3: construct graph data from the superpixels, and construct edges from adjacent superpixels; S4: use a graph neural network model to train the graph data to obtain Classification result; the present invention improves the precision and effect of the final classification result.

Description

A hyperspectral image classification method based on graph neural network

technical field

The invention relates to the technical field of image processing, and more particularly to a hyperspectral image classification method based on a graph neural network.

Background technique

At present, with the development of remote sensing technology and imaging spectrometer, the resolution of hyperspectral images has been continuously improved, and the application requirements have become more and more extensive. Great difficulty came.

However, the high dimensionality and highly redundant information may lead to a dramatic increase in computational complexity and may affect the classification accuracy. At the same time, the acquisition of labeled samples of hyperspectral data requires a lot of manpower and material resources, which is very difficult to obtain, which also hinders the research of hyperspectral image classification. The purpose of the study of hyperspectral image classification is to use the spectral information of a pixel to assign a feature category to each pixel in the hyperspectral image. At the beginning, the features are extracted manually, and then the data is analyzed with the help of machine learning algorithms. Modeling and classification, as researchers are required to manually extract features, experiments are slow and results are erratic.

Therefore, how to provide a hyperspectral image classification method that can solve the above problems is an urgent problem to be solved by those skilled in the art.

SUMMARY OF THE INVENTION

In view of this, the present invention provides a hyperspectral image classification method based on graph neural network, which reduces the dependence on labeled samples, and proposes a semi-supervised hyperspectral image classification algorithm based on graph neural network, which utilizes a small amount of label data to improve the final performance. The accuracy of the classification results, as well as the reduction of the appearance of bad boundary contours, results in classification results with high availability.

In order to achieve the above object, the present invention adopts the following technical solutions:

A hyperspectral image classification method based on graph neural network, including:

S1: Acquire hyperspectral image data of an area to be processed, and preprocess the hyperspectral image data;

S2: perform superpixel segmentation on the preprocessed hyperspectral image data to obtain segmented superpixels;

S3: The adjacent superpixel points are formed into edges, and graph data is constructed by the edges;

S4: Use a graph network model to train the graph data to obtain a classification result.

The beneficial effects of using the above method are as follows: the HSGACN network model, namely the graph neural network, is introduced into the hyperspectral image classification, and the effect and accuracy of the hyperspectral image classification are improved.

Preferably, the step S1 specifically includes: the preprocessing process is to remove interference bands from the hyperspectral image data, and perform normalization processing on the removed hyperspectral image data.

Preferably, in the step S3, the spatial distance and the spectral distance between the two superpixels are obtained, and the weights of the spatial distance and the spectral distance are balanced and then iterated.

Preferably, the number of iterations is 10 times.

Preferably, the graph neural network model is based on a variety of graph neural network models, such as networks designed by graph convolutional networks, graph attention networks, and the like.

As can be seen from the above technical solutions, compared with the prior art, the present invention provides a hyperspectral image classification method based on graph neural network, reduces the dependence on labeled samples, and proposes a semi-supervised hyperspectral image based on graph neural network. The classification algorithm uses a small amount of label data to improve the accuracy of the final classification result and reduce the appearance of bad boundary contours to obtain classification results with high availability.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only It is an embodiment of the present invention. For those of ordinary skill in the art, other drawings can also be obtained according to the provided drawings without creative work.

FIG. 1 is a flowchart of a method for fusing oblique photographic models provided by the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

Example 1

Referring to FIG. 1, Embodiment 1 of the present invention discloses a hyperspectral image classification method based on a graph neural network, including:

S1: Acquire hyperspectral image data of an area to be processed, and preprocess the hyperspectral image data;

S2: perform superpixel segmentation on the preprocessed hyperspectral image data to obtain segmented superpixels;

Among them, step S2 proposes a hyperspectral image superpixel segmentation algorithm based on the SLIC algorithm. By calculating the spatial distance and spectral distance between pixels, and balancing the weights, iteratively updates the superpixel cluster center and range boundary. The iteration stops when the error between the cluster center and the old cluster center is less than a certain range, and finally a hyperspectral image data composed of superpixels is obtained.

S3: The adjacent superpixel points are formed into edges, and graph data is constructed by the edges;

Among them, in the specific process of step S3: the superpixels in the segmentation graph are formed into nodes one by one, and the feature of the node is the average value of all pixel features in the superpixel, and then an edge is constructed from the nodes of the adjacent superpixels to construct the graph data.

S4: Use a graph neural network model to train the graph data to obtain a classification result.

Among them, the feature propagation characteristics of graph neural network are used to change the characteristics of unlabeled samples, so that the characteristics of superpixels of the same category are more similar. The graph attention network can change the characteristics of edge weights to gradually optimize the structure of graph data. Using the graph convolutional network can make full use of the characteristics and structural information of the graph data for the characteristics of the global operation.

Specifically, the graph neural network model in step S4 is a two-layer graph neural network structure, one is a graph attention layer, and the other is a graph convolutional layer, wherein the graph convolutional network is an existing structure

其中H^(l)是第l层的输出(即嵌入结果)，σ(·)表示激活函数，在本文中使用的是Softplus函数，W^l表示第l层包含的可训练权重矩阵。

where H ^(l) is the output of the lth layer (i.e. the embedding result), σ( ) represents the activation function, the Softplus function is used in this paper, and Wl represents the trainable weight matrix contained in the ^lth layer.

In a specific embodiment, step S1 specifically includes: the preprocessing process is to remove interference bands from the hyperspectral image data, and perform normalization processing on the removed hyperspectral image data.

Specifically, the interference band mainly excludes whether it is the absorption band of water vapor (4-7 microns). Since the 4-7 micron band is the interference absorption of atmospheric water at a long distance from the satellite, the data quality of this band is poor, and it is difficult to participate in subsequent calculations. Normalization is normalized by mean squared difference (x-mean)/variance.

In a specific embodiment, in step S3, the spatial distance and the spectral distance between the two superpixels are obtained, and the weights of the spatial distance and the spectral distance are balanced and then iterated.

Specifically, the expressions of spatial distance and spectral distance are:

In the formula, dc is the spectral square difference, d(sx, sy) is the spectral angular distance, and ds is the distance squared difference, which is balanced by adjusting m, and the value of m is generally 100.

当该误差小于0.01即可。In a specific embodiment, the number of iterations is 10, and the error expression between the cluster center and the old cluster center is:

When the error is less than 0.01.

To verify the effectiveness of the method, experiments are conducted on three public datasets, Indian pines, Pavia University, and Kennedy Space Center, and achieve 91.84%, 95.69%, and 98.42% accuracies, respectively, with only 30 labeled samples per class.

The above description of the disclosed embodiments enables any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (5)

1. A hyperspectral image classification method based on a graph neural network is characterized by comprising the following steps:

s1: the method comprises the steps of obtaining hyperspectral image data of a region to be processed, and preprocessing the hyperspectral image data;

s2: performing superpixel segmentation on the preprocessed hyperspectral image data to obtain segmented superpixels;

s3: forming edges by the adjacent super pixel points, and constructing graph data by the edges;

s4: and training the graph data by using a graph neural network model to obtain a classification result.

2. The method for classifying hyperspectral images based on a neural network of a graph according to claim 1, wherein the step S1 specifically comprises: the preprocessing process includes the steps of eliminating interference wave bands from the hyperspectral image data and performing normalization processing on the eliminated hyperspectral image data.

3. The method for classifying hyperspectral images based on a neural network of a graph according to claim 1, wherein in the step S3, a spatial distance and a spectral distance between two superpixels are obtained, and iteration is performed after the clustering center position and range are updated according to the spatial distance and the spectral distance balance weight.

4. The hyperspectral image classification method based on the graph neural network according to claim 1, wherein in the step S4, the graph neural network model is a network designed based on a plurality of graph neural network models, such as a graph convolution network, a graph attention network, and the like.

5. The method for classifying the hyperspectral images based on the graph neural network according to claim 3, wherein the number of iterations is 10.

Images