Hyperspectral remote sensing image classification method and device based on small sample classification model

Abstract

The application relates to the technical field of remote sensing image processing, in particular to a hyperspectral remote sensing image classification method and device based on a small sample classification model, wherein the method comprises the following steps: inputting a pixel block of a hyperspectral remote sensing image into a pre-constructed adaptive mask module to obtain a corresponding mask pixel block; inputting the mask pixel block into a pre-constructed Brown distance covariance representation module to obtain a predicted label corresponding to the pixel block; training the adaptive mask module and the Brown distance covariance representation module to obtain a small sample classification model of the hyperspectral remote sensing image, so as to obtain a classification result of the hyperspectral remote sensing image. Thus, the problems in the related art that due to the complexity and diversity of ground objects, the ground object spatial coverage is inconsistent, it is difficult to accurately extract spatial context information, to mine the nonlinear correlation between different spectral dimensions, and to effectively obtain and summarize the information of limited prior samples are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image processing, and particularly relates to a hyperspectral remote sensing image classification method and device based on a small sample classification model. BACKGROUND

[0002] A hyperspectral image not only contains spatial information such as the size, shape and relative position of a ground object, but also contains the spectral characteristics of each pixel. In addition, the spectral resolution of a hyperspectral image can reach the nanometer level, and can capture subtle differences between different ground objects. Therefore, hyperspectral imaging technology is widely used in mineral exploration, medical diagnosis, aerospace and other fields.

[0003] In related technologies, small sample classification can be used for hyperspectral image classification. Small sample learning learns knowledge from limited prior samples, and then generalizes to a large number of unlabeled samples, thereby effectively solving the sample imbalance problem and improving the classification accuracy of small classes. Covariance matrix and traditional metric loss such as contrast loss and triplet loss can also be combined to realize the classification of remote sensing images.

[0004] However, in related technologies, due to the complexity and diversity of ground objects, adjacent pixels may belong to completely different classes. This uneven coverage makes it difficult for traditional fixed window or neighborhood-based processing methods to accurately extract spatial context information. Secondly, there are often complex nonlinear relationships between spectral bands, which are crucial for distinguishing different ground objects. In addition, the high spectral dimension of a hyperspectral image makes it difficult to capture the differences between samples, and establishing a relationship between samples based on Euclidean distance cannot effectively obtain and summarize the information of limited prior samples, which needs to be improved. SUMMARY

[0005] The present application provides a hyperspectral remote sensing image classification method and device based on a small sample classification model, to solve the problem in the art that due to the complexity and diversity of ground objects, the spatial coverage of ground objects is inconsistent, it is difficult to accurately extract spatial context information, to mine the nonlinear correlation between different spectral dimensions, and to effectively obtain and summarize the information of limited prior samples.

[0006] The first aspect embodiment of the present application provides a hyperspectral remote sensing image classification method based on a small sample classification model, including the following steps: inputting a pixel block of a hyperspectral remote sensing image into a pre-constructed adaptive mask module to obtain a mask pixel block corresponding to the pixel block; inputting the mask pixel block into a pre-constructed Brown distance covariance representation module to obtain a predicted label corresponding to the pixel block by using the Brown distance covariance representation module; training the adaptive mask module and the Brown distance covariance representation module based on the predicted label to obtain a small sample classification model of the hyperspectral remote sensing image, so as to obtain a classification result of the hyperspectral remote sensing image by using the small sample classification model.

[0007] Optionally, in an embodiment of the present application, before the pixel block of the hyperspectral remote sensing image is input into the pre-constructed adaptive mask module, it further includes: obtaining an initial hyperspectral remote sensing image of a sample to be classified, and pre-processing the initial hyperspectral remote sensing image to obtain a hyperspectral remote sensing image; generating a pixel block of the hyperspectral remote sensing image and a real label corresponding to the pixel block based on the hyperspectral remote sensing image.

[0008] Optionally, in an embodiment of the present application, the inputting of the pixel block of the hyperspectral remote sensing image into the pre-constructed adaptive mask module to obtain the mask pixel block corresponding to the pixel block includes: obtaining at least one pixel block of the hyperspectral remote sensing image; obtaining a center spectral vector in the pixel block based on the at least one pixel block; calculating the Euclidean distance similarity between the center spectral vector and other pixel points in the pixel block except the center spectral vector according to the pixel block and the center spectral vector; generating an adaptive sampling mask matrix based on the Euclidean distance similarity to perform masking based on the adaptive sampling mask matrix to obtain the mask pixel block.

[0009] Optionally, in an embodiment of the present application, the inputting of the mask pixel block into the pre-constructed Brown distance covariance representation module to obtain the predicted label corresponding to the pixel block by using the Brown distance covariance representation module includes: extracting spectral information of the mask pixel block by using deep convolution and deep deconvolution to obtain a first Brown distance covariance matrix and a second Brown distance covariance matrix of the mask pixel block; constructing a double-channel Brown distance covariance matrix based on the first Brown distance covariance matrix and the second Brown distance covariance matrix; obtaining the predicted label by using the double-channel Brown distance covariance matrix and a fully connected layer.

[0010] Optionally, in an embodiment of the present application, the training of the adaptive mask module and the Brown distance covariance representation module based on the predicted label to obtain the small sample classification model of the hyperspectral remote sensing image comprises: calculating a Gaussian kernel triplet loss between the predicted label and the real label by using a Gaussian kernel triplet loss function; calculating a cross-entropy loss between the predicted label and the real label by using a cross-entropy loss function; obtaining a back propagation loss between the predicted label and the real label based on the Gaussian kernel triplet loss and the cross-entropy loss, and training the adaptive mask module and the Brown distance covariance representation module by using the back propagation loss to obtain the small sample classification model.

[0011] Optionally, in an embodiment of the present application, the formula of the back propagation loss can be but is not limited to:

[0012] L = λL KT + L CE ,

[0013] wherein, L KT is the Gaussian kernel triplet loss function, L CE is the cross-entropy loss function, and λ is a hyperparameter for balancing the two losses.

[0014] The second aspect embodiment of the present application provides a hyperspectral remote sensing image classification device based on a small sample classification model, comprising: a first generation module configured to input a pixel block of a hyperspectral remote sensing image into a pre-constructed adaptive mask module to obtain a mask pixel block corresponding to the pixel block; a second generation module configured to input the mask pixel block into a pre-constructed Brown distance covariance representation module to obtain a predicted label corresponding to the pixel block by using the Brown distance covariance representation module; and a third generation module configured to train the adaptive mask module and the Brown distance covariance representation module based on the predicted label to obtain a small sample classification model of the hyperspectral remote sensing image, and to obtain a classification result of the hyperspectral remote sensing image by using the small sample classification model.

[0015] Optionally, in an embodiment of the present application, the device further comprises: a preprocessing module configured to obtain an initial hyperspectral remote sensing image of a sample to be classified before inputting a pixel block of the hyperspectral remote sensing image into the pre-constructed adaptive mask module, and to pre-process the initial hyperspectral remote sensing image to obtain a hyperspectral remote sensing image; and a fourth generation module configured to generate the pixel block of the hyperspectral remote sensing image and a real label corresponding to the pixel block based on the hyperspectral remote sensing image.

[0016] Optionally, in an embodiment of the present application, the first generating module comprises: an acquisition unit configured to acquire at least one pixel block of a hyperspectral remote sensing image; a first generating unit configured to obtain a center spectral vector in the pixel block based on the at least one pixel block; a first calculating unit configured to calculate a Euclidean distance similarity between the center spectral vector and other pixel points in the pixel block except the center spectral vector according to the pixel block and the center spectral vector; and a second generating unit configured to generate an adaptive sampling mask matrix based on the Euclidean distance similarity, and to obtain the mask pixel block by performing mask based on the adaptive sampling mask matrix.

[0017] Optionally, in an embodiment of the present application, the second generating module comprises: an extraction unit configured to extract spectral information of the mask pixel block by using deep convolution and deep deconvolution to obtain a first Brown distance covariance matrix and a second Brown distance covariance matrix of the mask pixel block; a construction unit configured to construct a double-channel Brown distance covariance matrix based on the first Brown distance covariance matrix and the second Brown distance covariance matrix; and a third generating unit configured to obtain the predicted label by using the double-channel Brown distance covariance matrix and a full connection layer.

[0018] Optionally, in an embodiment of the present application, the third generating module comprises: a second calculating unit configured to calculate a Gaussian kernel triplet loss between the predicted label and the real label by using a Gaussian kernel triplet loss function; a third calculating unit configured to calculate a cross-entropy loss between the predicted label and the real label by using a cross-entropy loss function; and a fourth generating unit configured to obtain a back propagation loss between the predicted label and the real label based on the Gaussian kernel triplet loss and the cross-entropy loss, and to train the adaptive mask module and the Brown distance covariance representation module by using the back propagation loss to obtain the small sample classification model.

[0019] Optionally, in an embodiment of the present application, a calculation formula of the back propagation loss can be but is not limited to:

[0020] L = λL KT + L CE ,

[0021] wherein, L KT is a Gaussian kernel triplet loss function, L CE is a cross-entropy loss function, and λ is a hyperparameter for balancing the two losses.

[0022] An electronic device is provided in a third aspect of the embodiments of the present application, and includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the program to implement the hyperspectral remote sensing image classification method based on a small sample classification model as described in the above embodiments.

[0023] A computer readable storage medium is provided in a fourth aspect of the embodiments of the present application, and stores a computer program. The program is executed by a processor to implement the hyperspectral remote sensing image classification method based on a small sample classification model as described above.

[0024] A computer program product is provided in a fifth aspect of the embodiments of the present application, and includes a computer program. The program is executed to implement the hyperspectral remote sensing image classification method based on a small sample classification model as described above.

[0025] The embodiments of the present application can perform adaptive masking based on pixel blocks of a hyperspectral remote sensing image to obtain a mask pixel block corresponding to the pixel blocks, generate a corresponding prediction label in a pre-constructed Brown distance covariance representation module, train an adaptive masking module and the Brown distance covariance representation module based on the prediction label, and then obtain a small sample classification model to obtain a classification result of the hyperspectral remote sensing image. The mask pixel block can alleviate the problem of inconsistent spatial coverage of ground objects, the Brown distance covariance representation module can mine the nonlinear correlation between different spectral dimensions, the adaptive masking module and the Brown distance covariance representation module can map the features of labeled samples to a high-dimensional space to represent the complex nonlinear relationship between them, and better generalize prior information to the to-be-predicted samples. Thus, the problems in the related art, such as the complexity and diversity of ground objects, inconsistent spatial coverage of ground objects, difficulty in accurately extracting spatial context information, mining nonlinear correlation between different spectral dimensions, and inability to effectively obtain and generalize information of limited prior samples, are solved.

[0026] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0027] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, including the accompanying drawings, wherein:

[0028] Figure 1 A flowchart of a hyperspectral remote sensing image classification method based on a small sample classification model according to the embodiments of the present application is provided.

[0029] Figure 2A flowchart of a working principle of a hyperspectral remote sensing image classification method based on a small sample classification model according to an embodiment of the present application is provided.

[0030] Figure 3 A block schematic diagram of a hyperspectral remote sensing image classification device based on a small sample classification model according to an embodiment of the present application is provided.

[0031] Figure 4 A structural schematic diagram of an electronic device according to an embodiment of the present application is provided. DETAILED DESCRIPTION

[0032] Embodiments of the present application are described in detail below with reference to the accompanying drawings, in which the same or similar numerals or characters represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application.

[0033] A hyperspectral remote sensing image classification method and device based on a small sample classification model according to an embodiment of the present application are described below with reference to the accompanying drawings. In view of the problem that due to the complexity and diversity of ground objects, the spatial coverage of ground objects is inconsistent, it is difficult to accurately extract spatial context information, to mine the nonlinear correlation between different spectral dimensions, and to effectively obtain and summarize the information of limited prior samples, the present application provides a hyperspectral remote sensing image classification method based on a small sample classification model. In the method, an adaptive mask can be generated based on a pixel block of a hyperspectral remote sensing image to obtain a mask pixel block corresponding to the pixel block, a corresponding prediction label is generated in a pre-constructed Brown distance covariance representation module, and the adaptive mask module and the Brown distance covariance representation module are trained based on the prediction label, and then a small sample classification model is obtained to obtain a classification result of the hyperspectral remote sensing image. The problem of inconsistent spatial coverage of ground objects is alleviated by the mask pixel block, the nonlinear correlation between different spectral dimensions is mined by the Brown distance covariance representation module, the features of labeled samples are mapped to a high-dimensional space by training the adaptive mask module and the Brown distance covariance representation module, representing the complex nonlinear relationship between them, and the prior information is better generalized to the to-be-predicted samples. Thus, the problems in the related art that due to the complexity and diversity of ground objects, the spatial coverage of ground objects is inconsistent, it is difficult to accurately extract spatial context information, to mine the nonlinear correlation between different spectral dimensions, and to effectively obtain and summarize the information of limited prior samples are solved.

[0034] Specifically, Figure 1 A flowchart of a hyperspectral remote sensing image classification method based on a small sample classification model according to an embodiment of the present application is provided.

[0035] As Figure 1As shown, the hyperspectral remote sensing image classification method based on the small sample classification model includes the following steps:

[0036] In step S101, the pixel block of the hyperspectral remote sensing image is input into the pre-constructed adaptive mask module to obtain the mask pixel block corresponding to the pixel block.

[0037] As a possible implementation manner, the pixel block of the hyperspectral remote sensing image can be input into the pre-constructed adaptive mask module for masking, and then the mask pixel block corresponding to each pixel block is obtained.

[0038] The embodiment of the present application can perform adaptive masking based on the pixel block, calculate the adaptive Euclidean distance similarity, mask the pixels that are not similar to the pixels in the center spectral vector, retain the pixels that are similar to the pixels in the center spectral vector, and thus alleviate the problem of inconsistent spatial coverage of ground objects.

[0039] Optionally, in an embodiment of the present application, before the pixel block of the hyperspectral remote sensing image is input into the pre-constructed adaptive mask module, it further includes: obtaining an initial hyperspectral remote sensing image of a sample to be classified, and pre-processing the initial hyperspectral remote sensing image to obtain the hyperspectral remote sensing image; generating the pixel block of the hyperspectral remote sensing image and the real label corresponding to the pixel block based on the hyperspectral remote sensing image.

[0040] It can be understood that the hyperspectral remote sensing image of the embodiment of the present application is a data cube, which can be represented by a tensor, and the vector of each point in the tensor is the pixel element radiation value corresponding to each band.

[0041] In some embodiments, the embodiment of the present application can pre-process the initial hyperspectral remote sensing image of the sample to be classified to obtain the hyperspectral remote sensing image, and divide the hyperspectral remote sensing image into pixel blocks to obtain at least one pixel block and the real label corresponding to each pixel block. Wherein H, W, B are the length, width and height of the hyperspectral remote sensing image respectively.

[0042] Optionally, in an embodiment of the present application, inputting the pixel block of the hyperspectral remote sensing image into the pre-constructed adaptive mask module to obtain the mask pixel block corresponding to the pixel block includes: obtaining at least one pixel block of the hyperspectral remote sensing image; obtaining a center spectral vector in the pixel block based on the at least one pixel block; calculating the Euclidean distance similarity between the center spectral vector and other pixel points in the pixel block except the center spectral vector according to the pixel block and the center spectral vector; generating an adaptive sampling mask matrix based on the Euclidean distance similarity, and performing masking based on the adaptive sampling mask matrix to obtain the mask pixel block.

[0043] As a possible implementation manner, the embodiment of the application can select a pixel block P as the input of a pre-constructed adaptive mask module, and select a center spectral vector x i,j from P i,j . Then, the Euclidean distance similarity SimED between the center spectral vector x i,j and other pixel points in P i,j is calculated to obtain a Euclidean distance similarity mapping E which describes the spatial relationship modeled by the center spectral vector:

[0044] E i，j = {e i,j (i',j') | i'=1,2,3,...,s, j'=1,2,3,...,s} i,j .

[0045] wherein e i′,j′ is the SimED value of E i,j at (i',j'), wherein i'=1,2,3,...,s, j'=1,2,3,...,s, and e i′,j′ may be but is not limited to being expressed as:

[0046]

[0047] wherein x i′,j′ is the spectral vector of a pixel point other than x i,j in P i,j , E i,j is normalized to a unified representation, and the closer the value of e i′,j′ is to 1, the higher the similarity between x i′,j′ and x i,j .

[0048] Further, in order to adaptively select high-confidence neighborhood pixels, the embodiment of the application sets a sampling threshold t i,j based on the mean m s and the standard deviation s s of the SimED mapping E s , that is, t s =m s -Cxs s , wherein λ is a hyperparameter. Based on the sampling threshold t s , an adaptive sampling mask matrix M is obtained, which may be but is not limited to being expressed as:

[0049]

[0050] wherein m i′,j′ is the value of M at (i',j').

[0051] Therefore, in this embodiment, a mask is obtained by performing a masking based on an adaptive sampling mask matrix M to obtain mask pixel blocks. It can be, but is not limited to, expressed as:

[0052]

[0053] in, This represents pixel-level multiplication along the spectral dimension.

[0054] For example, this embodiment of the application is written in Python and implemented using the classic deep learning framework PyTORCH, with Python remote sensing image read / write functions as the implementation basis. It calls the data processing libraries NUMPY, SCIPY, and SPECTRAL, inputs the filename of the remote sensing image to be read, and the remote sensing image is read into a tensor of size H×W×B. Each element in the tensor is the pixel radiance value corresponding to each band, where H is the length of the remote sensing image, W is the width of the remote sensing image, and B is the number of bands in the remote sensing image. Python remote sensing image read / write functions are well-known in this technical field and will not be described in detail here.

[0055] like Figure 2 As shown, in order to alleviate the problem of inconsistent spatial coverage of ground features, this embodiment of the application performs adaptive masking on pixel blocks. The pixel blocks of the hyperspectral remote sensing image are... As input to a pre-built adaptive mask module, and from pixel block P i,j Select the center spectral vector Then, calculate the central spectral vector x. i,j and P i,j The Euclidean distance similarity SimED between other pixels is used to obtain the Euclidean distance similarity map.

[0056] Furthermore, in order to adaptively select high-confidence neighboring pixels, embodiments of this application are based on SimED mapping E i,j The mean m s and standard deviation s s Set the sampling threshold t s And based on the sampling threshold t s Obtain the adaptive sampling mask matrix Then, a mask is created based on the adaptive sampling mask matrix M to obtain the mask pixel block.

[0057] In step S102, the mask pixel block is input into the pre-constructed Brownian distance covariance representation module to obtain the predicted label corresponding to the pixel block using the Brownian distance covariance representation module.

[0058] As a possible implementation manner, the mask pixel block is input into a pre-constructed Brown distance covariance representation module for feature extraction, and then is input into a multi-layer full connection layer to be converted into category probability information, and the maximum category probability is selected as the prediction label of the pixel block.

[0059] The embodiment of the application designs a lightweight dual-channel Brown distance covariance representation module, two channels use convolution and deconvolution respectively to extract spectral features, and then calculate the Brown distance covariance representation thereof to mine the nonlinear correlation between different spectral dimensions.

[0060] Optionally, in an embodiment of the application, the mask pixel block is input into a pre-constructed Brown distance covariance representation module to obtain the prediction label corresponding to the pixel block by using the Brown distance covariance representation module, including: extracting spectral information of the mask pixel block by using deep convolution and deep deconvolution to obtain a first Brown distance covariance matrix and a second Brown distance covariance matrix of the mask pixel block; constructing a dual-channel Brown distance covariance matrix based on the first Brown distance covariance matrix and the second Brown distance covariance matrix; and obtaining the prediction label by using the dual-channel Brown distance covariance matrix and a full connection layer.

[0061] As a possible implementation manner, the embodiment of the application inputs the mask pixel block X i,j As the pre-constructed Brown distance covariance representation module, in order to obtain more rich spectral feature representation, a parallel dual-channel structure is adopted without changing the original spatial size X i,j , wherein channel 1 extracts spectral information by deep convolution, and channel 2 extracts spectral information by deep deconvolution. The calculation process can be but is not limited to represented as:

[0062]

[0063] wherein, DepthwiseConv and DepthwiseDeConv are deep convolution and deep deconvolution respectively, are outputs of the Brown distance covariance representation modules in channel 1 and channel 2 respectively.

[0064] Further, the embodiment of the application inputs the Brown distance covariance matrix X as the input of the distance covariance matrix . Specifically, is first reshaped into a two-dimensional matrix , and then the Euclidean distance c of any two bands of X is calculated.

[0065]

[0066] wherein, and respectively represent the pth and qth spectral band of X c , is the Euclidean distance between x p and x q .

[0067] Therefore, the embodiments of the present application can generate the Euclidean distance matrix and further calculate the distance covariance matrix A c , which can be but not limited to expressed as:

[0068]

[0069] wherein, and are the row mean, column mean and mean of all elements of A , respectively.

[0070] Further, the embodiments of the present application obtain the corresponding Brownian distance covariance matrix A c1 and A c2 for channel 1 and channel 2, respectively, superimpose A c1 and A c2 in the spectral dimension to obtain the two-channel Brownian distance covariance matrix

[0071] In some embodiments, the embodiments of the present application flatten the three-dimensional matrix A0into a one-dimensional feature vector, extract spectral information through two fully connected layers, then perform nonlinear mapping using Sigmoid and Dropout, and finally obtain the predicted label A2through a fully connected layer. The calculation process can be but not limited to expressed as:

[0072] A1=FC2(FC1(Flatten(A0))),

[0073] A2=FC3(Dropout(Sigmoid(A1))),

[0074] For example, in order to mine the nonlinear correlation between different spectral dimensions, the embodiments of the present application use a pre-constructed Brownian distance covariance representation module to uniformly model the spatial spectral information. This module takes the mask pixel block X i,j as input, in order to obtain more rich spectral feature representation, a parallel two-channel structure is adopted, wherein channel 1 extracts spectral information through deep convolution, and channel 2 extracts spectral information through deep deconvolution.

[0075] In addition, the embodiments of the present application can obtain the corresponding Brownian distance covariance matrix Ac1 and A c2 , A c1 and A c2 Superimpose in the spectral dimension to get the two-channel Brown distance covariance matrix

[0076] Further, the embodiment of the present application flattens the three-dimensional matrix A0 into a one-dimensional feature vector, extracts spectral information through two fully connected layers, then uses Sigmoid and Dropout for nonlinear mapping, and finally obtains a predicted label through a fully connected layer.

[0077] In step S103, the adaptive mask module and the Brown distance covariance representation module are trained based on the predicted label to obtain a small sample classification model of the hyperspectral remote sensing image, so as to obtain a classification result of the hyperspectral remote sensing image by using the small sample classification model.

[0078] Those skilled in the art can understand that the embodiment of the present application can train the adaptive mask module and the Brown distance covariance representation module based on the predicted label, calculate the loss, and obtain the small sample classification model through multiple iterations, and then obtain the classification result of the hyperspectral remote sensing image by using the small sample classification model.

[0079] The embodiment of the present application uses Gaussian kernel measurement loss to map the features of the labeled samples to a high-dimensional space to represent the complex nonlinear relationship between them, which can better generalize the prior information to the to-be-predicted samples.

[0080] Optionally, in an embodiment of the present application, the adaptive mask module and the Brown distance covariance representation module are trained based on the predicted label to obtain a small sample classification model of the hyperspectral remote sensing image, including: calculating the Gaussian kernel triplet loss between the predicted label and the true label by using a Gaussian kernel triplet loss function; calculating the cross-entropy loss between the predicted label and the true label by using a cross-entropy loss function; obtaining the back propagation loss between the predicted label and the true label based on the Gaussian kernel triplet loss and the cross-entropy loss, and training the adaptive mask module and the Brown distance covariance representation module by using the back propagation loss to obtain the small sample classification model. The calculation formula of the back propagation loss can be, but is not limited to:

[0081] L = λL KT + L CE ,

[0082] Wherein, L KT is the Gaussian kernel triplet loss function, L CE is the cross-entropy loss function, and λ is a hyperparameter for balancing the two losses.

[0083] As a possible implementation manner, the adaptive mask module and the Brown distance covariance representation module can be repeatedly iteratively trained by using the Gaussian kernel triplet loss, the cross-entropy loss and the back propagation loss, and then a trained small sample classification model is obtained.

[0084] In the formula, the back propagation loss L can be calculated by using the Gaussian kernel triplet loss function and the cross-entropy loss function, which can be but not limited to represented as:

[0085] L = λL KT + L CE ,

[0086]

[0087] In the formula, L KT is the Gaussian kernel triplet loss function, L CE is the cross-entropy loss function, and λ is a hyperparameter for balancing the two losses. In the formula, x i represents a sample feature, and N is the sample number. (x i , x p ) represents a positive sample pair, and (x i , x n ) represents a negative sample pair. represents the loss caused by the positive sample most dissimilar to x i , represents the loss caused by the negative sample most similar to x i . h(x) = max(0, x) is a hinge loss function, and α is a hyperparameter. In the formula, U is the sample number in a small batch, C is the total number of categories in the data set, y c and are the ground truth reference map and the predicted label of the c-th category, respectively. k(x i , x j ) represents a kernel distance function, and its expression can be but not limited to:

[0088]

[0089] In the formula, f(x i ) is the feature of the sample x i , and ||f(x i )-f(x j )| 2 is the Euclidean distance between f(x i ) and f(x j ). σ k is a standard deviation, and its expression can be but not limited to:

[0090]

[0091] wherein, in the embodiment of the present application, σ k is the average distance multiplied by 2 k The embodiment of the present application uses five different values of the standard deviation, i.e. N k = 5, at this time, the range of the function value is 0 to N k The more similar the two samples are, the closer the value is to 0.

[0092] For example, the embodiment of the present application can learn the distance measurement information between prior samples, and generalize the prior information to unlabeled samples. Therefore, the whole network can be repeatedly iteratively trained by using the Gaussian kernel triplet loss function, the cross-entropy loss function and the back propagation loss, to obtain the trained small sample classification model.

[0093] Further, the embodiment of the present application is verified by a comparative experiment.

[0094] (1) Introduction of data set

[0095] The embodiment of the present application selects Indian Pine Tree, Houston 2013 and Salinas hyperspectral data sets as experimental data sets.

[0096] Indian Pine Tree is imaged by airborne visible infrared imaging spectrometer XX on a pine tree in Indiana, USA in XX, and then a size of 145x145 is intercepted for labeling as a hyperspectral image classification test purpose. The imaging wavelength range of XX imaging spectrometer is 0.4-2.5 μm, which is continuous imaging of ground objects at 220 continuous wavebands, but since the 104-108, 150-163 and 220 wavebands cannot be reflected by water, generally, the remaining 200 wavebands after removing these 20 wavebands are used as the research object. The data has a total of 21025 pixels, but only 10249 pixels are ground object pixels, and the remaining 10776 pixels are background pixels, which need to be removed in actual classification. Since the intercepted area is all crops, there are a total of 16 categories, so different ground objects have relatively similar spectral curves, and among the 16 categories, the distribution of samples is extremely uneven.

[0097] Houston 2013 data is taken by XX imaging spectrometer, and the data set covers the University of Houston campus and the surrounding urban area. Its spatial resolution is 2.5 meters, and it contains a total of 144 spectral bands, with an imaging wavelength range of 0.364-1.046 μm. The data set contains a total of 15029 ground object pixels, and contains a total of 16 categories.

[0098] The Salinas data was also acquired by the Hyperion sensor, which imaged the Salinas Valley in California, USA. Unlike the Pines data set, the spatial resolution of the Salinas data set is 3.7 m. The original image also has 224 bands, and the commonly used image is the image with 204 bands after removing the bands that cannot be reflected by water, i.e., bands 108-112, 154-167, and 224. The size of the image is 512*217, so there are a total of 111104 pixels, of which 56975 are background pixels, 54129 are pixels that can be applied to classification, and these pixels are classified into 16 categories in total.

[0099] (2) Introduction of comparison method

[0100] Among them, the embodiments of the present application respectively adopt cross-domain small sample network (method 1), class covariance measurement network (method 2), deep measurement feature embedding network (method 3), composite neighborhood perception convolution measurement network (method 4), spatial spectral twin network (method 5), double measurement multi-level relationship network (method 6) and self-pooling Transformer network (method 7) for classification comparison.

[0101] (3) Introduction of evaluation index

[0102] The classification evaluation index of the embodiments of the present application can adopt a quantitative evaluation method. In total, 80 pixels of the data set one are selected as training samples and 10169 pixels are selected as test samples, 75 pixels of the data set two are selected as training samples and 14954 pixels are selected as test samples, and 80 pixels of the data set three are selected as training samples and 54049 pixels are selected as test samples. The evaluation index can be but is not limited to kappa coefficient, OA (Overall Accuracy, overall accuracy) and AA (Average Accuracy, average accuracy).

[0103] Among them, the kappa coefficient is an authoritative evaluation index for evaluating classification problems, and the larger the kappa coefficient, the higher the accuracy.

[0104] OA is an evaluation index for evaluating classification problems, and the higher the overall accuracy, the higher the detection accuracy.

[0105] AA is an evaluation index for evaluating classification problems, which is a measure of the accuracy of the results calculated by class

[0106] (4) Introduction of verification results

[0107] Kappa coefficient:

[0108] In the comparative test, the highest Kappa coefficient obtained by the method 1-method 3 is selected to evaluate the detection capability of the method 1-method 7. The sample acquisition confusion matrix of the embodiment of the application is shown in Table 1, wherein Table 1 is a confusion matrix schematic table provided by an embodiment of the application.

[0109] Table 1

[0110]

[0111] Wherein, TTO represents the number of samples whose label is category one and prediction is also category one; TF represents the number of samples whose actual category is category two but prediction is category one. FT represents the number of samples whose actual category is category one but prediction is category two. TTT represents the number of samples whose label is category two and prediction is also category two; NCO is the sum of TTO and FT, NCT is the sum of TF and TTT, NRO is the sum of TTO and TF, NRT is the sum of FT and TTT, and N is the total number of samples.

[0112] Further, the embodiment of the application calculates the Kappa coefficient according to the confusion matrix, and the calculation formula can be but is not limited to represented as:

[0113]

[0114] OA:

[0115] The calculation method of the OA of the embodiment of the application is also based on the confusion matrix shown in Table 1, and the calculation formula can be but is not limited to represented as:

[0116]

[0117] AA:

[0118] The calculation method of the AA of the embodiment of the application is also based on the confusion matrix shown in Table 1, and the calculation formula can be but is not limited to represented as:

[0119]

[0120] Further, the embodiment of the application compares the comparative experiment of the method 1-method 7 and the embodiment of the application based on the three evaluation indexes of the Kappa coefficient, OA and AA, and the evaluation indexes are shown in Table 2, wherein Table 2 is a comparative test result schematic table provided by an embodiment of the application.

[0121] Table 2

[0122]

[0123] As can be seen from Table 2, the embodiment of the present application achieves the best results in the three indicators of OA, AA and Kappa value. Compared with other methods, the embodiment of the present application has higher classification accuracy and stronger classification ability, fully considers the correlation between prior samples and the correlation between spectral bands, and improves the classification accuracy.

[0124] The working principle of the hyperspectral remote sensing image classification method based on the small sample classification model proposed by the embodiment of the present application will be introduced below with a specific embodiment. Figure 2 The working principle of the hyperspectral remote sensing image classification method based on the small sample classification model provided by an embodiment of the present application is shown in a flowchart.

[0125] Among them, Figure 2 The working principle of the hyperspectral remote sensing image classification method based on the small sample classification model provided by an embodiment of the present application is shown in a flowchart.

[0126] Step S201: Obtain a pixel block of a hyperspectral remote sensing image and a real label corresponding to the pixel block.

[0127] Among them, the embodiment of the present application can pre-process the initial hyperspectral remote sensing image of the sample to be classified to obtain a hyperspectral remote sensing image, and divide the hyperspectral remote sensing image into pixel blocks to obtain at least one pixel block and a real label corresponding to each pixel block.

[0128] Step S202: Construct an adaptive mask module.

[0129] Among them, the embodiment of the present application can input the pixel block of the hyperspectral remote sensing image into the pre-constructed adaptive mask module to perform masking, and then obtain a mask pixel block corresponding to each pixel block.

[0130] Step S203: Construct a Brown distance covariance representation module.

[0131] Among them, the embodiment of the present application can input the mask pixel block X i,j into a parallel double-channel structure to obtain a predicted label by using the pre-constructed Brown distance covariance representation module.

[0132] Further, the embodiment of the present application obtains a small sample classification model of the hyperspectral remote sensing image based on the adaptive mask module and the Brown distance covariance representation module. Among them, the embodiment of the present application can repeatedly iteratively train the adaptive mask module and the Brown distance covariance representation module by using Gaussian kernel triplet loss, cross-entropy loss and back propagation loss, and then obtain a trained small sample classification model.

[0133] Further, the embodiment of the present application is also verified by a comparative experiment. As can be seen from Table 2, the embodiment of the present application achieves the best results in the three indicators of OA, AA and Kappa value, has higher classification accuracy and stronger classification ability.

[0134] The hyperspectral remote sensing image classification method based on a few-shot classification model proposed in this application can adaptively mask the pixel blocks of the hyperspectral remote sensing image to obtain mask pixel blocks corresponding to the pixel blocks. Corresponding predicted labels are generated through a pre-constructed Brownian distance covariance representation module. The adaptive mask module and the Brownian distance covariance representation module are then trained based on the predicted labels to obtain a few-shot classification model for the classification results of the hyperspectral remote sensing image. The mask pixel blocks alleviate the problem of inconsistent spatial coverage of ground features, and the Brownian distance covariance representation module is used to mine the nonlinear correlations between different spectral dimensions. By training the adaptive mask module and the Brownian distance covariance representation module, the features of labeled samples are mapped to a high-dimensional space, representing the complex nonlinear relationships between them, and better generalizing prior information to the samples to be predicted. Therefore, this solves the problems in related technologies, such as inconsistent spatial coverage of ground features due to their complexity and diversity, difficulty in accurately extracting spatial context information, mining nonlinear correlations between different spectral dimensions, and the inability to effectively acquire and generalize information from limited prior samples.

[0135] Next, referring to the accompanying drawings, a hyperspectral remote sensing image classification device based on a small sample classification model proposed in the embodiments of this application is described.

[0136] Figure 3 This is a block diagram of a hyperspectral remote sensing image classification device based on a small sample classification model provided in an embodiment of this application.

[0137] like Figure 3 As shown, the hyperspectral remote sensing image classification device 10 based on a small sample classification model includes: a first generation module 100, a second generation module 200, and a third generation module 300.

[0138] The first generation module 100 is used to input the pixel blocks of the hyperspectral remote sensing image into a pre-built adaptive mask module to obtain the mask pixel blocks corresponding to the pixel blocks.

[0139] The second generation module 200 is used to input the mask pixel block into the pre-constructed Brownian distance covariance representation module, so as to obtain the predicted label corresponding to the pixel block using the Brownian distance covariance representation module.

[0140] The third generation module 300 is used to train an adaptive mask module and a Brownian distance covariance representation module based on the predicted labels to obtain a few-sample classification model for hyperspectral remote sensing images, so as to obtain the classification results of hyperspectral remote sensing images using the few-sample classification model.

[0141] Optionally, in one embodiment of this application, it further includes a preprocessing module and a fourth generation module.

[0142] The preprocessing module is configured to acquire an initial hyperspectral remote sensing image of a sample to be classified, and preprocess the initial hyperspectral remote sensing image to obtain the hyperspectral remote sensing image before inputting a pixel block of the hyperspectral remote sensing image into the self-adaptive mask module.

[0143] The fourth generation module is configured to generate the pixel block of the hyperspectral remote sensing image and a real label corresponding to the pixel block based on the hyperspectral remote sensing image.

[0144] Optionally, in an embodiment of the present application, the first generation module 100 comprises an acquisition unit, a first generation unit, a first calculation unit and a second generation unit.

[0145] The acquisition unit is configured to acquire at least one pixel block of the hyperspectral remote sensing image.

[0146] The first generation unit is configured to obtain a center spectral vector in the pixel block based on the at least one pixel block.

[0147] The first calculation unit is configured to calculate a Euclidean distance similarity between the center spectral vector and other pixel points in the pixel block except the center spectral vector based on the pixel block and the center spectral vector.

[0148] The second generation unit is configured to generate a self-adaptive sampling mask matrix based on the Euclidean distance similarity, and to perform mask based on the self-adaptive sampling mask matrix to obtain a mask pixel block.

[0149] Optionally, in an embodiment of the present application, the second generation module 200 comprises an extraction unit, a construction unit and a third generation unit.

[0150] The extraction unit is configured to extract spectral information of the mask pixel block by using deep convolution and deep deconvolution to obtain a first Brown distance covariance matrix and a second Brown distance covariance matrix of the mask pixel block.

[0151] The construction unit is configured to construct a double-channel Brown distance covariance matrix based on the first Brown distance covariance matrix and the second Brown distance covariance matrix.

[0152] The third generation unit is configured to obtain a predicted label by using the double-channel Brown distance covariance matrix and a full connection layer.

[0153] Optionally, in an embodiment of the present application, the third generation module 300 comprises a second calculation unit, a third calculation unit and a fourth generation unit.

[0154] The second calculation unit is configured to calculate a Gaussian kernel triplet loss between the predicted label and a real label by using a Gaussian kernel triplet loss function.

[0155] The third calculation unit is configured to calculate a cross-entropy loss between the predicted label and the real label by using a cross-entropy loss function.

[0156] The fourth generation unit is configured to obtain a back propagation loss between the predicted label and the real label based on the Gaussian kernel triplet loss and the cross-entropy loss, and train the adaptive mask module and the Brown distance covariance representation module by using the back propagation loss, so as to obtain the small sample classification model.

[0157] Optionally, in an embodiment of the present application, the calculation formula of the back propagation loss can be but is not limited to:

[0158] L = λL KT + L CE ,

[0159] wherein, L KT is the Gaussian kernel triplet loss function, L CE is the cross-entropy loss function, and λ is a hyperparameter for balancing the two losses.

[0160] It should be noted that the foregoing explanation and description of the embodiment of the hyperspectral remote sensing image classification method based on the small sample classification model also applies to the embodiment of the hyperspectral remote sensing image classification device based on the small sample classification model, which will not be described here.

[0161] The hyperspectral remote sensing image classification device based on the small sample classification model according to the embodiment of the present application can perform adaptive mask based on the pixel block of the hyperspectral remote sensing image to obtain a mask pixel block corresponding to the pixel block, generate a corresponding predicted label through the pre-constructed Brown distance covariance representation module, train the adaptive mask module and the Brown distance covariance representation module based on the predicted label, and then obtain the small sample classification model to obtain the classification result of the hyperspectral remote sensing image. The mask pixel block can relieve the problem of inconsistent spatial coverage of ground objects, the Brown distance covariance representation module can mine the nonlinear correlation between different spectral dimensions, the adaptive mask module and the Brown distance covariance representation module can be trained to map the features of the labeled samples to a high-dimensional space to represent the complex nonlinear relationship between them, and the prior information can be better generalized to the to-be-predicted samples. Therefore, the problems in the related art, such as the complexity and diversity of ground objects, inconsistent spatial coverage of ground objects, difficulty in accurately extracting spatial context information, mining nonlinear correlation between different spectral dimensions, and inability to effectively obtain and generalize information of limited prior samples, are solved.

[0162] Figure 4 The structure schematic diagram of the electronic device provided by the embodiment of the present application is shown. The electronic device can include:

[0163] The memory 401, the processor 402 and the computer program stored on the memory 401 and executable on the processor 402.

[0164] The processor 402 implements the hyperspectral remote sensing image classification method based on the small sample classification model provided in the above embodiments when executing the program.

[0165] Further, the electronic device further comprises:

[0166] The communication interface 403 is used for communication between the memory 401 and the processor 402.

[0167] The memory 401 is used to store the computer program executable on the processor 402.

[0168] The memory 401 can include a high-speed RAM memory, and can also include a non-volatile memory, for example, at least one disk memory.

[0169] If the memory 401, the processor 402 and the communication interface 403 are independently implemented, the communication interface 403, the memory 401 and the processor 402 can be connected through a bus and complete communication between each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 4 Only one thick line is used in the figure, but it does not mean that there is only one bus or only one type of bus.

[0170] Optionally, in specific implementation, if the memory 401, the processor 402 and the communication interface 403 are integrated on a chip, the memory 401, the processor 402 and the communication interface 403 can complete communication between each other through an internal interface.

[0171] The processor 402 can be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0172] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the hyperspectral remote sensing image classification method based on the small sample classification model.

[0173] The embodiment of the present application further provides a computer program product, which comprises a computer program, and the program is executed to implement the hyperspectral remote sensing image classification method based on the small sample classification model.

[0174] In the description of the present specification, the description of the terms “one embodiment”, “some embodiments”, “an example”, “a specific example”, or “some examples” means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0175] In addition, the terms “first”, “second” are only used for description purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with “first”, “second” can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of “N” is at least two, for example, two, three, etc., unless otherwise specifically limited.

[0176] Any process or method descriptions in flow charts or otherwise described herein represent embodiments that can be understood as a set of steps, operations, or stages of implementing a custom logic function or process that can be subsequently embodied as code that can be executed by a computer, and the preferred embodiments of the present application include additional implementing functions that can not be expressly presented or discussed below. The scope of the preferred embodiments of the present application includes additional implementing functions that can not be expressly presented or discussed below, and the person skilled in the art should understand that the functions can be performed in the order shown or discussed, in a substantially simultaneous manner, or in the opposite order, according to the functions involved, which should be understood by the person skilled in the art.

[0177] The logic and / or steps represented in the flowcharts and / or described herein, for example, can be considered as a sequence of executable instructions stored in a computer readable medium, which can be executed by an instruction execution system, apparatus or device, such as a computer-based system, a processor-based system, or other system that can fetch the instructions from the instruction execution system, apparatus or device and execute the instructions, or a combination of the above. For the purposes of this specification, a "computer readable medium" can be any apparatus that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus or device. The computer readable medium can be a computer readable storage medium or a computer readable signal medium. The computer readable storage medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or a propagation medium. The computer readable signal medium can include, but is not limited to, a computer readable medium that facilitates transfer of the program from one place to another. A specific example of a computer readable medium is a non-transitory computer-readable storage medium. A specific example of a computer readable signal medium is a source or destination of the computer readable medium. Another specific example of a computer readable signal medium is a computer readable signal travelling through space. Thus, a computer readable medium can take many forms of hardware to carry out the program for use by or in connection with the instruction execution system, apparatus or device.

[0178] It should be understood that aspects of the application can be implemented in hardware, software, firmware or a combination thereof. In the above embodiments, the N steps or methods can be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. If implemented in hardware and in another embodiment, the hardware can be implemented using any or a combination of the following technologies, which are each well known in the art: a discrete logic circuit(s) having logic gates for implementing logic functions upon an application of data signals, an application specific integrated circuit having appropriate combinational logic gates, a programmable gate array(s) (PGA), a field programmable gate array (FPGA), etc.

[0179] Those of skill in the art would understand that the steps of the methods carried out above can be carried out wholly or partly by a program instructing relevant hardware, and the program can be stored in a computer readable storage medium, and when executed, includes one or a combination of the steps of the method embodiments.

[0180] In addition, each of the functional units in the various embodiments of the present application can be integrated in one processing module, or each of the units can be physically present separately, or two or more units can be integrated in one module. The integrated module can be implemented in the form of hardware or in the form of a software functional module. When the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer readable storage medium.

[0181] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A hyperspectral remote sensing image classification method based on a small sample classification model, characterized in that, The method comprises the following steps: inputting a pixel block of a hyperspectral remote sensing image into a pre-constructed adaptive mask module to obtain a mask pixel block corresponding to the pixel block; inputting the mask pixel block into a pre-constructed Brown distance covariance representation module to obtain a predicted label corresponding to the pixel block by using the Brown distance covariance representation module; training the adaptive mask module and the Brown distance covariance representation module based on the predicted label to obtain a small sample classification model of the hyperspectral remote sensing image, and obtaining a classification result of the hyperspectral remote sensing image by using the small sample classification model; wherein the inputting the pixel block of the hyperspectral remote sensing image into the pre-constructed adaptive mask module to obtain the mask pixel block corresponding to the pixel block comprises: obtaining at least one pixel block of the hyperspectral remote sensing image; obtaining a central spectral vector in the pixel block based on the at least one pixel block; calculating a Euclidean distance similarity between the central spectral vector and other pixel points in the pixel block except the central spectral vector according to the pixel block and the central spectral vector; generating an adaptive sampling mask matrix based on the Euclidean distance similarity to perform masking based on the adaptive sampling mask matrix to obtain the mask pixel block; wherein the inputting the mask pixel block into the pre-constructed Brown distance covariance representation module to obtain the predicted label corresponding to the pixel block by using the Brown distance covariance representation module comprises: extracting spectral information of the mask pixel block by using deep convolution and deep deconvolution to obtain a first Brown distance covariance matrix and a second Brown distance covariance matrix of the mask pixel block; constructing a double-channel Brown distance covariance matrix based on the first Brown distance covariance matrix and the second Brown distance covariance matrix; obtaining the predicted label by using the double-channel Brown distance covariance matrix and a fully connected layer.

2. The method of claim 1, wherein, Before the pixel block of the hyperspectral remote sensing image is inputted into the pre-constructed adaptive mask module, the method further comprises: obtaining an initial hyperspectral remote sensing image of a sample to be classified, and pre-processing the initial hyperspectral remote sensing image to obtain the hyperspectral remote sensing image; generating a pixel block of the hyperspectral remote sensing image and a real label corresponding to the pixel block based on the hyperspectral remote sensing image.

3. The method of claim 2, wherein, The training of the adaptive mask module and the Brown distance covariance representation module based on the predicted label to obtain the small sample classification model of the hyperspectral remote sensing image comprises: calculating a Gaussian kernel triplet loss between the predicted label and the real label by using a Gaussian kernel triplet loss function; calculating a cross-entropy loss between the predicted label and the real label by using a cross-entropy loss function; obtaining a back propagation loss between the predicted label and the real label based on the Gaussian kernel triplet loss and the cross-entropy loss, and training the adaptive mask module and the Brown distance covariance representation module by using the back propagation loss to obtain the small sample classification model.

4. The method of claim 3, wherein, The calculation formula of the back propagation loss is: ， where, is a Gaussian kernel triplet loss function, is a cross-entropy loss function, is a hyper-parameter for balancing the two losses.

5. A hyperspectral remote sensing image classification device based on a small sample classification model, characterized in that, The hyperspectral remote sensing image classification method based on a small sample classification model according to any one of claims 1-4, wherein the device comprises: a first generation module configured to input a pixel block of a hyperspectral remote sensing image into a pre-constructed adaptive mask module to obtain a mask pixel block corresponding to the pixel block; a second generation module configured to input the mask pixel block into a pre-constructed Brown distance covariance representation module to obtain a predicted label corresponding to the pixel block by using the Brown distance covariance representation module; a third generation module configured to train the adaptive mask module and the Brown distance covariance representation module based on the predicted label to obtain a small sample classification model of the hyperspectral remote sensing image, and to obtain a classification result of the hyperspectral remote sensing image by using the small sample classification model.

6. An electronic device, comprising: The computer program is executed by the processor to implement the hyperspectral remote sensing image classification method based on a small sample classification model according to any one of claims 1-4. The program is executed by the processor to implement the hyperspectral remote sensing image classification method based on a small sample classification model according to any one of claims 1-4.

7. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed to implement the hyperspectral remote sensing image classification method based on a small sample classification model according to any one of claims 1-4.

8. A computer program product, characterised in that, ​

Images