CN111476287A - A kind of hyperspectral image small sample classification method and device - Google Patents

Abstract

The invention provides a method and a device for classifying hyperspectral image small samples, and belongs to the technical field of remote sensing image processing and application. The classification method comprises the following steps: inputting a hyperspectral image; constructing a morphological attribute section of the hyperspectral image by using a morphological attribute section algorithm; selecting a data cube in a set neighborhood range of a pixel to be classified as a morphological attribute section cube of the pixel in a morphological attribute section of the hyperspectral image, and further obtaining the morphological attribute section cube of the hyperspectral image; obtaining a feature vector of the hyperspectral image according to the morphological attribute profile cube of the hyperspectral image; and inputting the feature vectors of the hyperspectral images into a pre-trained classification model to complete hyperspectral image classification. According to the method, the empty spectrum joint information in the hyperspectral image can be fully utilized by constructing the morphological attribute profile cube of the hyperspectral image, and abundant characteristic information sources are provided for the subsequent classification process, so that higher classification accuracy can be obtained under the condition of a small sample.

Description

A kind of hyperspectral image small sample classification method and device

technical field

The invention relates to a hyperspectral image small sample classification method and device, belonging to the technical field of remote sensing image processing and application.

Background technique

The classification of hyperspectral images (also known as hyperspectral images) is a key link in hyperspectral image processing and analysis, and one of the important steps to realize earth observation of remote sensing images. Its basic task is to determine a unique category identifier for each pixel.

Traditional hyperspectral image classifiers include support vector machines, random forests, logistic regression classifiers, etc. By combining spatial features, texture features and spectral features in hyperspectral images, traditional classifiers can achieve certain classification effects. However, traditional classification methods often require complex feature design work, and rely heavily on expert experience for parameter adjustment, which has great limitations in application.

In recent years, with the rise and continuous development of deep learning, deep learning methods have been widely used in hyperspectral image classification, mainly in the following categories:

(1) Hyperspectral image classification method based on three-dimensional convolutional neural network (3DCNN). For example, in the invention patent document with the announcement number CN106022355B, a 3DCNN-based hyperspectral image space-spectrum joint classification method is disclosed. The method firstly extracts the pixels to be classified from the normalized original hyperspectral image. The data block P _n×n×L in the n×n×L neighborhood range of the center is used as the original empty spectral feature of the pixel, n represents the size of the neighborhood block, and L represents the total number of spectral segments; The 3DCNN network completes the joint classification of hyperspectral images. Although this method does not require spectral space dimensionality reduction, it can automatically extract the spatial spectral features of hyperspectral images and does not need to manually preset features, but it needs to rely on a large number of training samples, and the acquisition of hyperspectral image labeled samples in practical applications is time-consuming and laborious. , it is very inefficient to train a deep model by finding a large number of labeled samples. If there is not a sufficient number of labeled training samples for support, this method is difficult to achieve high classification accuracy.

(2) Hyperspectral image classification method based on deep forest. For example, in the invention patent application document with the publication number of CN108614992A, a classification method of hyperspectral remote sensing images is disclosed. Corresponding spectral data, spatial data and spatial spectral combined data are obtained; then the hyperspectral remote sensing images are classified by the trained deep forest classification model. Although this method uses the deep forest algorithm to classify hyperspectral remote sensing images, it has the advantages of few parameters and easy adjustment, and can also achieve good classification effect and practicability, but the data preprocessing process of this method is too complicated and time-consuming. The classification accuracy still needs to be further improved.

To sum up, among the existing hyperspectral image classification methods, traditional classification methods not only need to rely on expert experience for complex feature design work, but also have low classification accuracy; classification methods based on 3DCNN network need to rely on a large number of training samples. Under the condition of very few labeled samples, the classification accuracy is low; for the classification method based on deep forest, the data preprocessing process is too complicated and time-consuming, and the classification accuracy needs to be further improved.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a small sample classification method and device for hyperspectral images, so as to solve the problem that the classification accuracy of the existing hyperspectral image classification methods is not high under the condition of very few labeled samples.

In order to achieve the above purpose, the present invention provides a method for classifying small samples of hyperspectral images, the method comprising the following steps:

(1) Input hyperspectral images;

(2) Constructing the morphological attribute profile of the hyperspectral image by using the morphological attribute profile algorithm;

(3) In the morphological attribute profile of the hyperspectral image, select the data cube within the set neighborhood range of the pixel to be classified as the morphological attribute profile cube of the pixel, and then obtain the morphological attribute of the hyperspectral image. section cube;

(4) obtaining the feature vector of the hyperspectral image according to the morphological attribute profile cube of the hyperspectral image;

(5) Input the feature vector of the hyperspectral image into the pre-trained classification model to complete the hyperspectral image classification.

The present invention also provides a hyperspectral image small sample classification device, the device includes a processor and a memory, the processor executes a computer program stored in the memory, so as to realize the above-mentioned hyperspectral image small sample classification method.

The beneficial effects of the hyperspectral image small sample classification method and device of the present invention are as follows: the present invention first performs morphological attribute profile operation on the hyperspectral image to retain the space-spectral joint information in the image, and then selects the morphological properties of the hyperspectral image by selecting The pixels to be classified in the attribute profile are set as data cubes in the neighborhood range, and the morphological attribute profile cube of the hyperspectral image is obtained. Finally, the feature vector of the hyperspectral image is obtained according to the morphological attribute profile cube of the hyperspectral image. By constructing the morphological attribute section cube of the hyperspectral image, the space-spectral joint information in the hyperspectral image can be fully utilized, and a rich source of feature information can be provided for the subsequent classification process. to obtain higher classification accuracy.

Further, in the above-mentioned method and device for classifying small samples of hyperspectral images, sliding windows of different scales are used to perform multi-granularity scanning on the morphological attribute section cubes of the hyperspectral images, and the characteristics of the hyperspectral images are obtained according to the scanning results. vector.

The beneficial effect of this is: using sliding windows of different scales to perform multi-granularity scanning on the morphological attribute section cubes of hyperspectral images, the feature information of different scales in hyperspectral images can be better utilized, and more abstract features can be extracted. The feature is used as the discriminant information for the classification of ground objects in the hyperspectral image, which can greatly improve the classification accuracy and reduce the dimension of the feature vector of the hyperspectral image.

Further, in order to further enhance the richness and diversity of features, in the above-mentioned method and device for classifying small samples of hyperspectral images, a scan result is used as an instance, and an instance is input into a complete random forest and an ordinary random forest, respectively. The two eigenvectors of this instance are connected with the eigenvectors of all instances as the eigenvectors of the hyperspectral image.

Further, in the above method and device for classifying small samples of hyperspectral images, the pre-trained classification model is a deep forest classification model, and the deep forest classification models are connected by a cascade structure.

The beneficial effect of this is: using the deep forest classification model to use the input feature information through multi-level cascade can improve the utilization rate of the feature information, and then obtain higher classification accuracy under the condition of very few labeled samples. The classification model can avoid too many hyperparameters and improve the classification efficiency.

个特征作为候选特征，d是输入特征的个数，并选择基尼系数最好的特征进行分割并生长。Further, in order to improve the classification efficiency, in the above-mentioned method and device for classifying small samples of hyperspectral images, two complete random forests and two ordinary random forests are set in each layer of the deep forest classification model. A feature is randomly selected on each leaf node of each decision tree for segmentation and growth; each leaf node of each decision tree in the ordinary random forest is randomly selected by

A feature is used as a candidate feature, d is the number of input features, and the feature with the best Gini coefficient is selected for segmentation and growth.

Further, in the above-mentioned method and device for classifying small samples of hyperspectral images, the pre-trained classification model is a classification model trained by a support vector machine or a convolutional neural network.

Further, in the above method and device for classifying small samples of hyperspectral images, the morphological attribute profile of the hyperspectral image is obtained by the following steps: extracting the first N principal components in the hyperspectral image by using a principal component analysis method, N≥1, use the morphological attribute profile algorithm to construct a morphological attribute profile for each of the first N principal components, and then synthesize the morphological attribute profiles of the first N principal components to obtain the hyperspectral Morphological properties profile of the image.

Further, in the above method and device for classifying small samples of hyperspectral images, square structural elements with sizes of 1, 3, 5, 7, and 9 are respectively used to construct the morphological attribute profile of the hyperspectral image.

Further, in the above-mentioned method and device for classifying small samples of hyperspectral images, the operation sequence of the morphological attribute profile algorithm is: first perform the opening operation and then perform the closing operation, or perform the closing operation first and then perform the opening operation.

Description of drawings

1 is a flowchart of a method for classifying small samples of hyperspectral images in a method embodiment of the present invention;

2 is a schematic diagram of the construction process of the morphological attribute section cube of the hyperspectral image in the method embodiment of the present invention;

3 is a schematic diagram of a process for obtaining feature vectors of hyperspectral images by multi-granularity scanning in an embodiment of the method of the present invention;

4 is a schematic diagram of a process of using a deep forest classification model to complete hyperspectral image classification in an embodiment of the method of the present invention;

5 is a comparison diagram of the classification results of the hyperspectral image small sample classification method in the method embodiment of the present invention under different numbers of labeled samples;

FIG. 6 is a structural diagram of an apparatus for classifying small samples of hyperspectral images in an embodiment of the apparatus of the present invention.

Detailed ways

Method embodiment

The hyperspectral image small sample classification method of this embodiment is shown in FIG. 1 , and the method includes the following steps:

(1) Input a hyperspectral image, assuming that the size of the hyperspectral image is a×b, and there are L spectral bands;

(2) Constructing the morphological attribute profile of the hyperspectral image by using the morphological attribute profile algorithm;

In this embodiment, in order to improve the computing efficiency, the first three principal component bands (hereinafter referred to as principal components) in the hyperspectral image are extracted by principal component analysis, and then the A morphological attribute profile is constructed for each principal component, and finally the morphological attribute profiles of the first three principal components are synthesized to obtain the morphological attribute profile of the hyperspectral image.

Among them, when constructing the morphological attribute profile of a principal component, first use square structural elements with sizes 1, 3, 5, 7, and 9 to perform morphological opening operation filtering on the principal component to obtain 5 profiles, and then use the size The square structuring elements of 1, 3, 5, 7, and 9 perform morphological closing operation filtering on the principal component to obtain 5 sections. Since the section obtained when the structuring element size is 1 is the original band, in fact, the principal component has a total of 5 sections. Nine profiles are obtained, which together constitute the morphological property profile of the principal component. To sum up, since the morphological attribute profile of each principal component actually contains 9 profiles, the morphological property profile of the hyperspectral image is obtained by synthesizing the morphological attribute profiles of the first three principal components in the hyperspectral image. The attribute profile actually contains 3×9=27 profiles, that is to say, the obtained morphological property profile of the hyperspectral image is actually a data cube of a×b×27.

As another embodiment, the principal component analysis method can also be used to extract the first N principal components in the hyperspectral image, where N≥1, and the extraction of the first three principal components is not limited. Actual settings are required, for example, square structural elements of 1, 3, 5, 7, 9, and 11 are used. In this case, the morphological attribute profile of the hyperspectral image is actually a data cube of a×b×11N. The process is similar to that of extracting the first three principal components, and will not be repeated here.

As another embodiment, in order to retain the feature information in the input hyperspectral image, the step of extracting the principal components can also be omitted, and the morphological attribute profile algorithm is directly used to process the input hyperspectral image to obtain its morphological attribute profile; in addition, The shape of the structural elements in the morphological attribute profile algorithm can also be set according to actual needs, such as diamond-shaped structural elements; the operation order of the morphological attribute profile algorithm can also be set according to actual needs, for example, the closing operation can be performed first and then the opening operation can be performed.

(3) In the morphological attribute profile of the hyperspectral image, select the data cube within the set neighborhood range of the pixel to be classified as the morphological attribute profile cube of the pixel, and then obtain the morphological attribute profile cube of the hyperspectral image. ;

In this embodiment, if the set field range is 28×28, one pixel obtains a morphological attribute section cube with a size of 28×28×27, because the morphological attribute section of the hyperspectral image contains a×b morphological attribute sections in total. pixel, the obtained morphological attribute section cube of the hyperspectral image actually contains a×b morphological attribute section cubes with a size of 28×28×27.

As another implementation manner, the set neighborhood range can be set according to actual needs, and is not limited to 28×28.

(4) Multi-granularity scanning of the morphological attribute profile cube of the hyperspectral image using sliding windows of different scales; taking a scan result as an instance, and inputting an instance into a complete random forest and an ordinary random forest respectively to obtain the instance The two eigenvectors of , connect the eigenvectors of all instances as the eigenvectors of the hyperspectral image;

The morphological attribute section cube of the hyperspectral image obtained in step (3) includes a×b morphological attribute section cubes with a size of 28×28×27 in total, and the shape of 28×28×27 is used for multi-granularity scanning. The scientific property section is carried out in cubic units. In this embodiment, a 28×28×27 morphological attribute section cube is scanned with a step size of 2 using sliding windows with sizes of 7×7, 10×10 and 14×14, and 100 pieces of size can be obtained. 7×7×27 instances, 81 instances of size 10×10×27, and 49 instances of size 14×14×27, since one instance corresponds to two feature vectors (assuming the dimension of the feature vector is n) , then after connecting the eigenvectors of all instances, a 28×28×27 morphological attribute section cube can obtain a 100×2×n+81×2×n+49×2×n=460n dimension eigenvector , and then a×b morphological attribute profile cubes with a size of 28×28×27 can obtain a×b eigenvectors with a dimension of 460n, and these a×b eigenvectors with a dimension of 460n are the characteristics of the hyperspectral image. vector.

As another implementation manner, when performing multi-granularity scanning, the size, number and scanning step size of the sliding windows can be adjusted according to actual needs.

(5) The feature vector of the hyperspectral image is input into the pre-trained deep forest classification model to complete the hyperspectral image classification.

个特征作为候选特征，d是输入特征的个数，并选择基尼系数最好的特征进行分割并生长。Among them, the deep forest classification model is connected by a cascade structure, that is, each layer receives the feature vector output by the previous layer, and outputs the feature vector obtained by this layer to the next layer. Each layer in a deep forest classification model is an ensemble of decision trees. In each layer of the deep forest classification model, setting different types of forests can effectively promote the diversity of features. In this embodiment, for the sake of simplicity, each layer of the deep forest classification model is set with two complete random forests and two Ordinary random forest, a feature is randomly selected on each leaf node of each decision tree in the complete random forest to divide and grow; each leaf node of each decision tree in the ordinary random forest is randomly selected

A feature is used as a candidate feature, d is the number of input features, and the feature with the best Gini coefficient is selected for segmentation and growth.

In this embodiment, two complete random forests and two ordinary random forests are set in each layer of the deep forest classification model, so that each layer will output four feature vectors; The type and number of forests can be adjusted according to actual needs, and the feature vector output by each layer will also change accordingly.

In this embodiment, a pre-trained deep forest classification model is used to complete the hyperspectral image classification; as another implementation, a pre-trained support vector machine classification model or a convolutional neural network classification model can also be used to complete the hyperspectral image classification.

In this embodiment, after the morphological attribute section cube of the hyperspectral image is obtained, multi-granularity scanning is performed on the morphological attribute section cube, and the scanning result is further processed to obtain the feature vector of the hyperspectral image. It can make good use of the feature information of different scales in hyperspectral images, and can extract more abstract features as the discriminative information for the classification of ground objects in hyperspectral images, which can greatly improve the classification accuracy and reduce the impact of hyperspectral images. The dimension of the feature vector; as another embodiment, the morphological attribute section cube of the hyperspectral image can also be directly used as the feature vector of the hyperspectral image, or the multi-granularity scanning result can be used as the feature vector of the hyperspectral image.

The validity of the hyperspectral image small sample classification method of this embodiment is verified on the Salinas data set below, and the specific process is as follows:

Step 1. Input hyperspectral image I. The input hyperspectral image is a commonly used Salinas image with a dimension of 512×217×204. The hyperspectral image includes 16 types of ground objects and 54129 pixels to be classified.

Step 2. Construct the morphological attribute section cube of the hyperspectral image I. The specific process is shown in Figure 2:

First, the first three principal components of the hyperspectral image I (512×217×204) are extracted by principal component analysis, and the hyperspectral image I′ after the extraction of the principal components is obtained, that is, I→I′ (512×217×3) ;

Then, the square structural elements with sizes of 1, 3, 5, 7, and 9 are used to perform the opening operation and then the closing operation on each principal component, respectively, to obtain the morphological attribute profile of each principal component (see Fig. 2, respectively). The morphological attribute profiles of the first, second and third principal components), and then the morphological attribute profiles of the first three principal components are synthesized to obtain the morphological attribute profile of the hyperspectral image I, that is, I′→I _MAP (512× 217×27);

Finally, the data cube whose neighborhood range of the pixel to be classified in the morphological attribute profile of hyperspectral image I is 28 × 28 is selected as the morphological attribute profile cube of the pixel, and the input hyperspectral image I forms a total of 512 × 217 data cubes. The morphological attribute section cubes with a size of 28×28×27, and these 512×217 morphological attribute section cubes with a size of 28×28×27 constitute the morphological attribute section cubes of the hyperspectral image I.

Step 3. Use sliding windows of different scales to perform multi-granularity scanning on the morphological attribute profile cube of the hyperspectral image I, take a scan result as an instance, and input an instance into a complete random forest and an ordinary random forest respectively to obtain the instance The two eigenvectors of , connect the eigenvectors of all instances as the eigenvectors of the hyperspectral image I. The specific process is shown in Figure 3:

When performing multi-grain scanning, the unit of 28×28×27 morphological property profile cube is performed. First, a 28 × 28 × 27 morphological attribute profile cube is scanned with a step size of 2 using sliding windows of size 7 × 7, 10 × 10 and 14 × 14, and 100 pieces of size 7 × 7 × 27 instances, 81 instances of size 10 × 10 × 27, and 49 instances of size 14 × 14 × 27; each instance was then fed into a full random forest and a normal random forest (see Figure 3, respectively). In the random forest A and random forest B), each forest converts each instance into a 16-dimensional feature vector, and connects the feature vectors of all instances, so that a 28×28×27 morphological attribute profile cube is obtained. A 100×2×16+81×2×16+49×2×16=7360-dimensional feature vector; since the input hyperspectral image I forms a total of 512×217 morphological attribute profiles with a size of 28×28×27 cube, the eigenvectors of hyperspectral image I contain 512×217 7360-dimensional eigenvectors in total.

Step 4. Input the feature vector of the hyperspectral image I into the pre-trained deep forest classification model to complete the hyperspectral image classification. The specific process is shown in Figure 4:

First, a 7360-dimensional feature vector in the feature vector of hyperspectral image I is used as the input of the deep forest classification model; since each layer of the deep forest classification model includes two complete random forests and two ordinary random forests, one forest If a 16-dimensional feature vector is output, the first layer will get a 16×4=64-dimensional feature vector;

Then, connect the 64-dimensional feature vector obtained in the first layer with the input vector to obtain the input vector with the next layer dimension of 16×4+7360=7424. After this layer is processed, four 16-dimensional feature vectors will also be obtained. , connect the feature vector obtained by this layer with the input vector of this layer as the input vector of the next layer; and so on, until the last layer of the deep forest classification model generates four 16-dimensional feature vectors, and average them value to obtain the final 16-dimensional feature vector, thereby obtaining the category label of a pixel to be classified;

After inputting 512×217 7360-dimensional feature vectors into the deep forest model, the category labels of all pixels to be classified in the hyperspectral image I are obtained, and the hyperspectral image classification is completed.

The simulation conditions in this embodiment are: Intel Core i7-5700HQ, 2.7GHz CPU, GeForceGTX970M graphics processor, 32GB memory; on the Salinas data set, randomly select 5, 10, 15, 20, 25 labeled samples are used as training samples, and the rest are used as test samples, using support vector machine (SVM), Laplacian support vector machine (LapSVM), transduction support vector machine (TSVM), semi-supervised convolutional neural network (SS) -CNN), Graph Convolutional Neural Network (GCN) and the small sample classification method of hyperspectral images of the present invention were respectively carried out 20 experiments. Among them, the classification results of the hyperspectral image small sample classification method of the present invention after a certain experiment is shown in Figure 5, and the final classification results of various methods after 20 experiments are shown in Table 1. The final classification in Table 1 Results are presented as mean and variance.

Table 1 Comparison table of final classification results of various methods

It can be seen from FIG. 5 that when L=5, L=10, L=15, L=20, and L=25, that is, when the number of marked samples is 5, 10, 15, 20, and 25, respectively, the present invention The classification accuracy of the hyperspectral image small sample classification method of the present invention is 93.69%, 94.61%, 97.97%, 98.27%, 98.96%, respectively, indicating that the hyperspectral image small sample classification method of the present invention can achieve a high degree of accuracy with a very small number of labeled samples. Classification accuracy.

It can be seen from Table 1 that when L=5, L=10, L=15, L=20, and L=25, the classification accuracy of the hyperspectral image small sample classification method of the present invention is much higher than that of other methods. The hyperspectral image small sample classification method of the present invention can greatly improve the classification accuracy of the hyperspectral image under an extremely small number of labeled samples.

Device embodiment

This embodiment provides an apparatus for classifying small samples of hyperspectral images. As shown in FIG. 6 , the apparatus includes a processor and a memory. The memory stores a computer program that can be run on the processor. When the computer program is implemented, the methods in the above method embodiments are implemented.

That is to say, the methods in the above method embodiments should be understood as the process of implementing the method for classifying small samples of hyperspectral images by computer program instructions. The computer program instructions may be provided to a processor such that execution by the processor of the instructions results in the implementation of the functions specified by the above-described method flows.

The processor in this embodiment refers to a processing device such as a microprocessor MCU or a programmable logic device FPGA.

The memory referred to in this embodiment includes a physical device for storing information. Usually, the information is digitized and then stored in an electrical, magnetic, or optical medium. For example: all kinds of memories that use electrical energy to store information, RAM, ROM, etc.; all kinds of memories that use magnetic energy to store information, hard disks, floppy disks, magnetic tapes, magnetic core memories, magnetic bubble memories, U disks; use optical methods to store information of all kinds of memory, CD or DVD. Of course, there are other ways of memory, such as quantum memory, graphene memory, and so on.

The device constituted by the above-mentioned memory, processor and computer program is realized by the processor executing corresponding program instructions in the computer, and the processor can be equipped with various operating systems, such as windows operating system, linux system, android, iOS system, etc.

As another implementation manner, the apparatus may further include a display, and the display is used for displaying the classification result for reference by the staff.

Claims (10)

1. a hyperspectral image small sample classification method, is characterized in that, this method comprises the following steps: (1) Input hyperspectral images; (2) Constructing the morphological attribute profile of the hyperspectral image by using the morphological attribute profile algorithm; (3) In the morphological attribute profile of the hyperspectral image, select the data cube within the set neighborhood range of the pixel to be classified as the morphological attribute profile cube of the pixel, and then obtain the morphological attribute of the hyperspectral image. section cube; (4) obtaining the feature vector of the hyperspectral image according to the morphological attribute profile cube of the hyperspectral image; (5) Input the feature vector of the hyperspectral image into the pre-trained classification model to complete the hyperspectral image classification. 2 . The method for classifying small samples of hyperspectral images according to claim 1 , wherein the morphological attribute section cubes of the hyperspectral images are subjected to multi-granularity scanning by using sliding windows of different scales, and the said hyperspectral images are obtained according to the scanning results. 3 . Eigenvectors of hyperspectral imagery. 3. The hyperspectral image small sample classification method according to claim 2, characterized in that, taking a scan result as an instance, and inputting an instance into a complete random forest and an ordinary random forest respectively to obtain two features of the instance vector, the feature vectors of all instances are connected as the feature vector of the hyperspectral image. 4. The hyperspectral image small sample classification method according to any one of claims 1-3, wherein the pre-trained classification model is a deep forest classification model, and the deep forest classification model adopts a cascade structure to connect. 5. The method for classifying small samples of hyperspectral images according to claim 4, wherein each layer of the deep forest classification model is provided with two complete random forests and two ordinary random forests, and in the complete random forest A feature is randomly selected on each leaf node of each decision tree for segmentation and growth; each leaf node of each decision tree in the ordinary random forest is randomly selected by

A feature is used as a candidate feature, d is the number of input features, and the feature with the best Gini coefficient is selected for segmentation and growth. 6 . The method for classifying small samples of hyperspectral images according to claim 1 , wherein the pre-trained classification model is a classification model trained by using a support vector machine or a convolutional neural network. 7 . 7 . The method for classifying small samples of hyperspectral images according to claim 4 , wherein the morphological attribute profile of the hyperspectral images is obtained by the following steps: extracting the front-end images of the hyperspectral images by using a principal component analysis method. 8 . N principal components, N≥1, use the morphological attribute profile algorithm to construct a morphological attribute profile for each of the first N principal components, and then synthesize the morphological attribute profiles of the first N principal components A morphological property profile of the hyperspectral image is obtained. 8 . The method for classifying small samples of hyperspectral images according to claim 4 , wherein the morphological attribute profiles of the hyperspectral images are constructed by using square structural elements with sizes of 1, 3, 5, 7, and 9, respectively. 9 . 9 . The method for classifying small samples of hyperspectral images according to claim 4 , wherein the operation sequence of the morphological attribute profile algorithm is: first perform an opening operation and then perform a closing operation, or first perform a closing operation and then perform an opening operation. 10 . operation. 10. An apparatus for classifying small samples of hyperspectral images, characterized in that the apparatus comprises a processor and a memory, and the processor executes a computer program stored in the memory, so as to realize the method according to any one of claims 1-9. A small sample classification method for hyperspectral images.

Images