A hyperspectral remote sensing image land cover classification method, system, device and medium

Abstract

The application discloses a hyperspectral remote sensing image land cover classification method, system, device and medium, and relates to the field of land cover classification. The method comprises the following steps: acquiring a hyperspectral remote sensing image of land to be classified, and determining a three-dimensional matrix of the hyperspectral remote sensing image of the land to be classified; converting the three-dimensional matrix into a two-dimensional matrix; dividing the two-dimensional matrix of the hyperspectral remote sensing image of the land into a first sample set with real ground object category labels and a second sample set without labels; inputting the first sample set and the second sample set into a weighted space-spectrum joint kernel collaborative representation classification algorithm to obtain a classified hyperspectral remote sensing image; wherein the weighted space-spectrum joint kernel collaborative representation classification algorithm is obtained by introducing a weighted space filtering operator and space structure information into a traditional collaborative representation classification algorithm. The method improves the classification performance of the collaborative representation model and can achieve ideal classification accuracy under small-scale labeled samples.

Description

TECHNICAL FIELD

[0001] The application relates to the field of land cover classification, and in particular to a hyperspectral remote sensing image land cover classification method, system, device and medium. BACKGROUND

[0002] With the rapid development of urbanization and human intervention, land use and land cover are constantly changing. Such changes have a significant impact on environmental factors such as climate, water balance, biodiversity, and terrestrial ecosystems. Accurate and timely monitoring and updating of land cover information is of great significance to environmental protection, sustainable development, land resource planning and management. In recent years, in response to the major needs of the country for land survey, researchers have used various remote sensing data such as satellite or airborne RGB images, multispectral images and hyperspectral remote sensing images to classify and map land cover types, and have made remarkable achievements. Among them, each pixel in the hyperspectral remote sensing image contains hundreds of narrow and continuous spectral bands, which can provide rich spectral and spatial information for ground objects, greatly reducing the phenomenon of "same object different spectrum" and "different object same spectrum", so it has attracted widespread attention from researchers.

[0003] However, most existing research results choose statistical-based classification algorithms to establish land cover classification models, such as support vector machine (SVM), random forest (RF) and sparse logistic regression (SLR) algorithms, but such algorithms usually assume that sample data follows a normal or multi-modal distribution and require a large number of labeled samples for training to fit the model. The lack of labeled samples in hyperspectral remote sensing images cannot meet the distribution assumption of the data and will affect the fitting performance of the model. In recent years, collaborative representation classification (CRC) algorithms have attracted widespread attention from researchers because they do not need to consider any prior distribution of sample data and do not require traditional training processes, thereby avoiding the influence of the number of labeled samples on the fitting performance of the model.

[0004] Given the simplicity and effectiveness of the CRC algorithm, researchers have optimized its framework and proposed a series of improved algorithms to improve its classification performance, such as collaborative representation classification algorithm based on Tikhonov regularization (CRT), CRT algorithm based on adaptive Euclidean distance (EDACRT), CRT algorithm based on kernel function (KCRT) and discriminative kernel CRT algorithm (DKCRT). These algorithms have improved the classification performance of the collaborative representation model to some extent, but only use the spectral information of the hyperspectral remote sensing image, ignoring its spatial information.

[0005] To further improve the classification performance of the collaborative representation model, researchers reconstruct the sample data by averaging or weighted averaging the spatial neighboring pixels of each sample (central pixel) to consider the spatial information of each sample in the hyperspectral remote sensing image and introduce it into the collaborative representation model, such as KCRT algorithm based on space-spectrum composite kernel (KCRT-CK), joint DKCRT algorithm (JDKCRT), weighted space-spectrum KCRT algorithm (WSSKCRT), and weighted space-spectrum DKCRT algorithm (WSSDKCRT). SUMMARY

[0006] The purpose of the present application is to provide a hyperspectral remote sensing image land cover classification method, system, device and medium to improve the accuracy of land cover classification.

[0007] To achieve the above purpose, the present application provides the following scheme:

[0008] A hyperspectral remote sensing image land cover classification method, comprising:

[0009] Obtaining a hyperspectral remote sensing image of land to be classified, and determining a three-dimensional matrix of the hyperspectral remote sensing image of the land to be classified;

[0010] Converting the three-dimensional matrix into a two-dimensional matrix;

[0011] Dividing the two-dimensional matrix of the hyperspectral remote sensing image of the land into a first sample set with real ground object class labels and a second sample set without labels;

[0012] Inputting the first sample set and the second sample set into a weighted space-spectrum joint kernel collaborative representation classification algorithm to obtain a classified hyperspectral remote sensing image;

[0013] The weighted space-spectrum joint kernel collaborative representation classification algorithm is obtained by introducing a weighted spatial filtering operator and spatial structure information into a traditional collaborative representation classification algorithm.

[0014] Optionally, the inputting the first sample set and the second sample set into a weighted space-spectrum joint kernel collaborative representation classification algorithm to obtain a classified hyperspectral remote sensing image specifically comprises:

[0015] normalizing and spatial filtering the first sample set and the second sample set to obtain a filtered first sample set and a filtered second sample set;

[0016] solving, according to the filtered first sample set and the filtered second sample set, a representation coefficient vector matrix of the filtered first sample set to each of the filtered second sample set and the corresponding spatial adjacent pixels thereof;

[0017] determining, according to the representation coefficient vector matrix, C intra-class coefficient vector matrices; wherein C represents a number of feature categories;

[0018] reconstructing, by using the filtered first sample set and the corresponding intra-class coefficient vector matrix, the filtered second sample and the corresponding spatial adjacent pixels thereof to obtain C reconstructed second samples and the corresponding spatial adjacent pixels thereof;

[0019] calculating, by using an F norm, an error between each of the reconstructed second sample and the corresponding spatial adjacent pixels thereof and the filtered second sample and the corresponding spatial adjacent pixels thereof to obtain C reconstruction errors;

[0020] taking a feature category corresponding to a minimum reconstruction error of the C reconstruction errors as a feature category of the filtered second sample;

[0021] determining a classified hyperspectral remote sensing image according to feature categories of all the filtered second samples.

[0022] Optionally, the normalizing and spatial filtering the first sample set and the second sample set to obtain a filtered first sample set and a filtered second sample set specifically comprises:

[0023] preprocessing the first sample set and the second sample set by using an amplitude normalization method to obtain a processed first sample set and a processed second sample set;

[0024] spatial filtering the processed first sample set and the processed second sample set by using a weighted spatial filtering operator to obtain a filtered first sample set and a filtered second sample set.

[0025] Optionally, the solving, according to the filtered first sample set and the filtered second sample set, a representation coefficient vector matrix of the filtered first sample set to each of the filtered second sample set and the corresponding spatial adjacent pixels thereof specifically comprises:

[0026] constructing a dictionary by using the filtered first sample set;

[0027] mapping the dictionary and the filtered second sample set into a high-dimensional feature space by using a kernel function to obtain a mapping function of the dictionary in the high-dimensional feature space and a mapping function of the filtered second sample and spatial neighboring pixels of the filtered second sample under a spatial filtering window in the high-dimensional feature space;

[0028] determining a representation coefficient vector matrix according to the mapping function of the dictionary in the high-dimensional feature space and the mapping function of the filtered second sample and spatial neighboring pixels of the filtered second sample under a spatial filtering window in the high-dimensional feature space.

[0029] Optionally, the spatial filtering of the processed first sample set and the processed second sample set by using the weighted spatial filtering operator to obtain a filtered first sample set and a filtered second sample set specifically comprises:

[0030] calculating a correlation coefficient of a processed sample and each spatial neighboring pixel of the processed sample under a spatial filtering window in the processed first sample set and the processed second sample set; the processed sample includes the processed first sample and the processed second sample;

[0031] according to the correlation coefficient, using a formula assigning a weight to the spatial neighboring pixel; wherein w 0,i represents the weight of the i th spatial neighboring pixel of the processed sample under the spatial filtering window; r 0.i represents the correlation coefficient of the processed sample and the i th spatial neighboring pixel of the processed sample under the spatial filtering window; i = 1, 2, …, n x n-1; n x n is the size of the spatial filtering window;

[0032] according to the weight of the spatial neighboring pixel, using a formula performing weighted average on the spatial neighboring pixel to obtain a filtered sample; the filtered sample includes the filtered first sample and the filtered second sample; wherein x 0,i represents the i th spatial neighboring pixel of the processed sample under the spatial filtering window; represents a reconstructed sample;

[0033] the filtered first sample set is constituted by the filtered first sample, and the filtered second sample set is constituted by the filtered second sample.

[0034] Optionally, the determining a representation coefficient vector matrix according to the mapping function of the dictionary in the high-dimensional feature space and the mapping function of the filtered second sample and spatial neighboring pixels of the filtered second sample under a spatial filtering window in the high-dimensional feature space specifically comprises:

[0035] is determined by using the formula is a representation coefficient vector matrix; wherein, ψ is a representation coefficient vector matrix; is a matrix composed of the filtered second sample and spatial neighboring pixels of the filtered second sample under a spatial filtering window, wherein is a filtered second sample, is an i-th spatial neighboring pixel of the filtered second sample under a spatial structure window; is a mapping function of the dictionary in a high-dimensional feature space; is a mapping function of the filtered second sample and spatial neighboring pixels of the filtered second sample under a spatial filtering window in a high-dimensional feature space; ψ * is an initial representation coefficient vector matrix; ||·|| F is a F norm, and λ is a global regularization parameter.

[0036] A hyperspectral remote sensing image land cover classification system, comprising:

[0037] A data acquisition module is configured to acquire a hyperspectral remote sensing image of land to be classified, and determine a three-dimensional matrix of the hyperspectral remote sensing image of the land to be classified.

[0038] A conversion module is configured to convert the three-dimensional matrix into a two-dimensional matrix.

[0039] A sample division module is configured to divide the two-dimensional matrix of the hyperspectral remote sensing image of the land into a first sample set with real ground object class labels and a second sample set without labels.

[0040] A classification module is configured to input the first sample set and the second sample set into a weighted space-spectrum joint kernel collaborative representation classification algorithm to obtain a classified hyperspectral remote sensing image.

[0041] The weighted space-spectrum joint kernel collaborative representation classification algorithm is obtained by introducing a weighted space filtering operator and spatial structure information into a traditional collaborative representation classification algorithm.

[0042] An electronic device, comprising a memory and a processor, the memory is configured to store a computer program, and the processor is configured to run the computer program to make the electronic device execute the hyperspectral remote sensing image land cover classification method.

[0043] A computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the hyperspectral remote sensing image land cover classification method.

[0044] According to the embodiments of the present application, the following technical effects are achieved:

[0045] The hyperspectral remote sensing image land cover classification method of the present application comprises the following steps: obtaining a hyperspectral remote sensing image of land to be classified, and determining a three-dimensional matrix of the hyperspectral remote sensing image of the land to be classified; converting the three-dimensional matrix into a two-dimensional matrix; dividing the two-dimensional matrix of the hyperspectral remote sensing image of the land into a first sample set with real ground object class labels and a second sample set without labels; inputting the first sample set and the second sample set into a weighted space-spectrum joint kernel collaborative representation classification algorithm to obtain a classified hyperspectral remote sensing image; wherein the weighted space-spectrum joint kernel collaborative representation classification algorithm is obtained by introducing a weighted space filtering operator and spatial structure information into a traditional collaborative representation classification algorithm. The method of the present application improves the classification performance of the collaborative representation model and can achieve ideal classification accuracy under small-scale labeled samples. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0047] Figure 1 The flowchart of the hyperspectral remote sensing image land cover classification method provided by the present application;

[0048] Figure 2 The flowchart of the hyperspectral remote sensing image land cover classification method of the present application in specific application;

[0049] Figure 3 An exemplary hyperspectral scene diagram provided for the embodiments of the present application;

[0050] Figure 4 The structure diagram of the hyperspectral remote sensing image land cover classification system provided by the present application. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0052] In order to fully tap the spatial and spectral characteristics of hyperspectral images, and further improve the performance of the collaborative representation model in land cover classification under small-scale labeled samples, the application provides a hyperspectral remote sensing image land cover classification method, system, device and medium to improve the accuracy of the collaborative representation model in land cover classification.

[0053] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the application will be further described in detail below in combination with the drawings and specific embodiments.

[0054] Embodiment one

[0055] Figure 1 The flowchart of the hyperspectral remote sensing image land cover classification method provided by the application, Figure 2 The flowchart of the hyperspectral remote sensing image land cover classification method of the application in specific application, as shown in Figure 1 and Figure 2 The hyperspectral remote sensing image land cover classification method of the application comprises:

[0056] Step 101: obtaining the hyperspectral remote sensing image of the land to be classified, and determining the three-dimensional matrix of the hyperspectral remote sensing image of the land to be classified. In this embodiment, the hyperspectral remote sensing image (land hyperspectral remote sensing image) used is the hyperspectral scene of the University of Pavia in Italy obtained by the ROSIS (Reflective Optics Spectrographic Imaging System) sensor on board. The image contains 610*340 pixels, the spatial resolution is 1.3m, and it is composed of 103 spectral bands ranging from 0.43-0.86μm. The scene includes nine ground objects, asphalt, meadows, gravel, trees, painted metal sheets, bare soil, bitumen, self-blocking bricks and shadows, a total of 42776 ground object samples, wherein each pixel represents a sample.

[0057] Step 102: converting the three-dimensional matrix into a two-dimensional matrix. The number of columns of the two-dimensional matrix represents the number of pixel points (i.e. the number of samples), and the number of rows of the two-dimensional matrix represents the dimension of the pixel (i.e. the number of bands of the hyperspectral remote sensing image).

[0058] In actual application, the obtained hyperspectral remote sensing image is as shown in Figure 3As shown, the hyperspectral remote sensing image is converted from a three-dimensional matrix to a two-dimensional matrix, and 60 samples are randomly selected from each type of ground object according to the real ground object label (ground object category label) to form a training sample set (first sample set), and a total of 540 training samples, and the remaining 42236 samples are used as a test sample set (second sample set). The training sample set and the test sample set y are input into the collaborative representation model, where d=103 represents the number of bands of the hyperspectral remote sensing image, and N=540 represents the number of labeled training samples; the labeled training samples are used to optimize the global regularization parameter λ, the spatial filtering window W f and the spatial structure window W s in the method of the application under a five-fold cross-validation strategy, and after optimization, λ is set to 10 -4 , W f is set to 13x13, and W s is set to 7x7.

[0059] Step 103: The two-dimensional matrix of the land hyperspectral remote sensing image is divided into a first sample set with real ground object category labels and a second sample set without labels.

[0060] The training sample set (first sample set) is used to construct a dictionary X=[X1, X2,..., X C ], where Xl represents the lth training sample set, N l =60 represents the number of lth training samples, l∈{1,2,...,C}, that is C=9 represents the number of ground object categories in the hyperspectral remote sensing image.

[0061] Step 104: The first sample set and the second sample set are input into the weighted space-spectrum joint kernel collaborative representation classification algorithm to obtain a classified hyperspectral remote sensing image.

[0062] The weighted space-spectrum joint kernel collaborative representation classification algorithm is obtained by introducing a weighted spatial filtering operator and spatial structure information into a conventional collaborative representation classification algorithm.

[0063] Further, the step 103 specifically comprises:

[0064] Step 1: The first sample set and the second sample set are normalized and spatially filtered to obtain a filtered first sample set and a filtered second sample set.

[0065] Further, the step 1 specifically comprises:

[0066] 1.1: Preprocessing the first sample set and the second sample set by using amplitude normalization method, to obtain the processed first sample set and the processed second sample set.

[0067] Preprocessing the original hyperspectral data (the training sample set and the test sample set) by using amplitude normalization method (AN).

[0068] 1.2: Spatially filtering the processed first sample set and the processed second sample set by using weighted spatial filter operator, to obtain the filtered first sample set and the filtered second sample set.

[0069] Spatially filtering the hyperspectral image under the spatial filtering window W f = 13x13 by using weighted spatial filter operator, and the filtered training sample set and the filtered test sample set are represented as and

[0070] Further, step 1.2 specifically includes:

[0071] 1.21: Calculating the correlation coefficient of the processed sample and each spatially adjacent pixel under the spatial filtering window of the processed sample in the processed first sample set and the processed second sample set; the processed sample includes the processed first sample and the processed second sample.

[0072] 1.22: According to the correlation coefficient, assigning weights to the spatially adjacent pixels by using the formula wherein w 0,i represents the weight of the i-th spatially adjacent pixel under the spatial filtering window of the processed sample; r 0.i represents the correlation coefficient of the processed sample and the i-th spatially adjacent pixel under the spatial filtering window of the processed sample; i = 1, 2, …, n x n-1; n x n is the size of the spatial filtering window.

[0073] 1.23: According to the weight of the spatially adjacent pixel, performing weighted average on the spatially adjacent pixel by using the formula to obtain the filtered sample; the filtered sample includes the filtered first sample and the filtered second sample. Wherein x 0,i represents the i-th spatially adjacent pixel under the spatial filtering window of the processed sample; represents the reconstructed sample.

[0074] 1.24: The filtered first sample set is composed of the filtered first sample, and the filtered second sample set is composed of the filtered second sample.

[0075] In practical applications, the weighted spatial filtering operator in step 4.2 is to calculate the correlation coefficient of each sample x (including the processed training sample and the processed test sample) and its spatial adjacent pixels under the spatial filtering window W f = n x n, where x 0,0 0,1 0,n×n-1 is the sample x itself. The result can be represented as R = {r 0,0 0,0 0,1 0,n×n-1 , where the greater the absolute value of the correlation coefficient, the higher the probability that the corresponding spatial adjacent pixel belongs to the same category as the sample (central pixel). The absolute value of the correlation coefficient is normalized to assign a weight to each spatial adjacent pixel x 0,i , i∈{1, 2, …, n x n-1}, which is expressed as follows:

[0076]

[0077] Further, the spatial adjacent pixels are weighted and averaged to reconstruct each sample, thereby alleviating the spectral shift caused by the adjacency effect in hyperspectral images, which is expressed as follows:

[0078]

[0079] Step 2: According to the filtered first sample set and the filtered second sample set, the filtered first sample set is solved for the representation coefficient vector matrix of each filtered second sample in the filtered second sample set and its corresponding spatial adjacent pixels.

[0080] In practical applications, the spatial and spectral features of the ground object samples are mined, and the constructed dictionary is used to represent the second sample and its spatial adjacent pixels to determine the representation coefficient vector matrix.

[0081] Further, the step 2 specifically includes:

[0082] 2.1: Construct a dictionary using the filtered training sample set. The dictionary constructed from the filtered training sample set is represented as where the l-th category of training sample set is represented as

[0083] 2.2: Map the dictionary and the filtered test sample set into a high-dimensional feature space using a kernel function to obtain the mapping function of the dictionary in the high-dimensional feature space and the mapping function of the filtered second sample and the spatial adjacent pixels of the filtered second sample under the spatial filtering window in the high-dimensional feature space.

[0084] ​​​​​The kernel function in the step 5.2 must satisfy the Mercer condition, then the inner product of any two samples mapped by the nonlinear mapping function Φ can be expressed by the kernel function, in this way, the sample data (the filtered first sample set and the filtered second sample set) are projected to the high-dimensional feature space induced by the kernel, so as to enhance the separability of the nonlinear sample data; the kernel function selected in the application is a radial basis function (RBF), and the specific expression is as follows:

[0085]

[0086] In the formula, γ (γ>0) is used to control the width of the RBF and is set to The median of where

[0087] 2.3: Determine the representation coefficient vector matrix according to the mapping function of the dictionary in the high-dimensional feature space and the mapping function of the spatial adjacent pixels of the filtered second sample and the filtered second sample under the spatial filtering window in the high-dimensional feature space. In actual application, in the high-dimensional space induced by the kernel, the dictionary constructed Meanwhile, the representation coefficient vector matrix is determined by estimating the representation of the filtered test sample and the spatial adjacent pixels thereof.

[0088] The representation coefficient vector matrix is determined by using the formula

[0089] Wherein, ψ is the representation coefficient vector matrix; is a matrix composed of the filtered second sample and the spatial adjacent pixels of the filtered second sample under the spatial filtering window, Wherein is the filtered second sample, is the i-th spatial adjacent pixel of the filtered second sample under the spatial structure window; is the dictionary constructed in 4.3; is the mapping function of the dictionary in the high-dimensional feature space; is the mapping function of the spatial adjacent pixels of the filtered second sample and the filtered second sample under the spatial filtering window in the high-dimensional feature space; ψ * is the initial representation coefficient vector matrix; ||·|| F The table is the F norm, and λ is a global regularization parameter.

[0090] Step 3: Determine C intra-class coefficient vector matrices according to the representation coefficient vector matrix; wherein C represents the number of ground object classes.

[0091] The coefficient vector matrix ψ in the step 3 can be solved by a closed-form solution, and the expression is as follows: ​

[0092]

[0093] where I denotes the identity matrix, denotes a Gram matrix composed of , i, j = 1, 2, …, N, denotes a matrix composed of , i = 1, 2, …, N; j = 0, 1, 2, …, n x n - 1.

[0094] Step 4: Reconstruct the filtered second samples and their corresponding spatial neighboring pixels using the filtered first sample set and the corresponding intra-class coefficient vector matrix, to obtain C reconstructed second samples and their corresponding spatial neighboring pixels. After the coefficient vector matrix ψ is solved, the intra-class training sample set and the corresponding intra-class coefficient vector matrix ψ l are used to reconstruct the filtered test samples and their spatial neighboring pixels in the kernel-induced high-dimensional feature space, so as to use the spatial structure information of the test samples to assist the classification of the test samples.

[0095] Step 5: Calculate the error between each reconstructed second sample and its corresponding spatial neighboring pixels and the filtered second sample and its corresponding spatial neighboring pixels respectively using the F-norm, to obtain C reconstruction errors.

[0096] Step 6: Take the ground object class corresponding to the minimum reconstruction error in the C reconstruction errors as the ground object class of the filtered second sample.

[0097] Finally, the class of the test sample is attributed to the class with the minimum reconstruction error, and the expression is as follows:

[0098]

[0099] where the symbol tr denotes the trace of a square matrix, denotes a Gram matrix composed of the l-th class training samples,

[0100] Step 9: Determine the classified hyperspectral remote sensing image according to the ground object classes of all the filtered second samples.

[0101] The method provided by the application not only utilizes the weighted spatial filtering operator to perform spatial filtering on the hyperspectral image to alleviate the spectral shift caused by the adjacency effect, but also utilizes the labeled training sample to represent and reconstruct each test sample and its spatial adjacent pixels at the same time, so as to consider the spatial structure information of each test sample, fully mine the spatial and spectral characteristics of the hyperspectral image, and assist in the classification of the test sample. On this basis, the sample data is projected into a high-dimensional feature space induced by a kernel by using a kernel technique (kernel function), and the separability of the nonlinear sample data is enhanced. Compared with other traditional collaborative representation models, the method provided by the application can significantly improve the performance of the collaborative representation model for land cover classification, and can achieve ideal classification accuracy under small-scale labeled samples, thereby effectively reducing the labeling cost of the hyperspectral image.

[0102] To verify the superiority of the weighted space-spectrum joint kernel collaborative representation algorithm (WSSJKCRC) provided by the application, the classification performance of land cover in the obtained hyperspectral scene is compared between the WSSJKCRC and seven traditional collaborative representation algorithms, including CRC, joint CRC (JCRC), direct average space-spectrum CRC (CRC-M), JDKCRT, KCRT-CK, WSSDKCRT and WSSKCRT, and the classification test of each algorithm is performed under the optimal parameters thereof. The classification performance of each algorithm is evaluated by using four evaluation indexes, including category accuracy (CA), overall accuracy (OA), average accuracy (AA) and Kappa statistic (Kappa), and the classification results are shown in Table 1. The test results show that, compared with other traditional collaborative representation classification algorithms, the WSSJKCRC algorithm of the application achieves the highest OA, AA and Kappa accuracy, the generated land cover classification map contains the least noise points and the most homogeneous regions, and the method provided by the application obtains an OA accuracy of more than 97% under the condition of 540 labeled training samples, which indicates that the method can achieve ideal classification effect under small-scale labeled samples, thereby effectively reducing the labeling cost of the hyperspectral image.

[0103] Table 1 Classification accuracy statistics of land cover of the hyperspectral image by each algorithm

[0104]

[0105]

[0106] Embodiment Two

[0107] To perform the method corresponding to the above-mentioned embodiment one, to realize the corresponding functions and technical effects, a hyperspectral remote sensing image land cover classification system is provided below, as shown in Figure 4 , comprising:

[0108] The data acquisition module 401 is configured to acquire hyperspectral remote sensing images of land to be classified, and determine a three-dimensional matrix of the hyperspectral remote sensing images of the land to be classified.

[0109] The conversion module 402 is configured to convert the three-dimensional matrix into a two-dimensional matrix.

[0110] The sample division module 403 is configured to divide the two-dimensional matrix of the hyperspectral remote sensing images of the land into a first sample set with real ground object category labels and a second sample set without labels.

[0111] The classification module 404 is configured to input the first sample set and the second sample set into a weighted space-spectrum joint kernel collaborative representation classification algorithm to obtain classified hyperspectral remote sensing images.

[0112] The weighted space-spectrum joint kernel collaborative representation classification algorithm is obtained by introducing a weighted space filtering operator and space structure information into a conventional collaborative representation classification algorithm.

[0113] Embodiment three

[0114] The application further provides an electronic device, including a memory and a processor, the memory is used for storing a computer program, and the processor runs the computer program to make the electronic device execute the hyperspectral remote sensing image land cover classification method in the embodiment one.

[0115] Embodiment four

[0116] The application further provides a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by a processor to realize the hyperspectral remote sensing image land cover classification method in the embodiment one.

[0117] In the specification, each embodiment is described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts of each embodiment can be referred to each other.

[0118] The principles and implementation manners of the application are described by using specific examples in the present application, and the above embodiment description is only used to help understand the method of the application and its core idea; meanwhile, for those skilled in the art, according to the idea of the application, the specific implementation manner and application range can be changed. In conclusion, the content of the specification should not be understood as a limitation of the application.

Claims

1. A hyperspectral remote sensing image land cover classification method, characterized in that, The application relates to a method for classifying hyperspectral remote sensing images of land, comprising the following steps: acquiring a hyperspectral remote sensing image of land to be classified, and determining a three-dimensional matrix of the hyperspectral remote sensing image of the land to be classified; converting the three-dimensional matrix into a two-dimensional matrix; dividing the two-dimensional matrix of the hyperspectral remote sensing image of the land into a first sample set with real ground object category labels and a second sample set without labels; inputting the first sample set and the second sample set into a weighted space-spectrum joint kernel collaborative representation classification algorithm to obtain a classified hyperspectral remote sensing image; wherein the weighted space-spectrum joint kernel collaborative representation classification algorithm is obtained by introducing a weighted space filtering operator and space structure information into a traditional collaborative representation classification algorithm; performing space filtering on the processed first sample set and the processed second sample set by using the weighted space filtering operator to obtain a filtered first sample set and a filtered second sample set, and the space filtering specifically comprises the following steps: calculating correlation coefficients of a processed sample and each spatially adjacent pixel under a spatial filtering window of the processed sample in the processed first sample set and the processed second sample set; the processed sample comprises a processed first sample and a processed second sample; According to the correlation coefficient, a formula The spatial adjacent pixels are given weights; wherein, w 0,i The weight of the i-th spatial adjacent pixel under the spatial filtering window of the processed sample is represented as w 0.i The correlation coefficient between the processed sample and the i-th spatial adjacent pixel under the spatial filtering window of the processed sample is represented as r ; i = 1, 2, …, n x n-1; n x n is the size of the spatial filtering window. According to the weight of the spatial neighboring pixel, a formula is used The spatial neighboring pixels are weighted and averaged to obtain a filtered sample; the filtered sample includes a filtered first sample and a filtered second sample; wherein x 0,i represents the i-th spatial neighboring pixel of the processed sample under a spatial filtering window; represents a reconstructed sample; the filtered first sample set is composed of the filtered first samples, and the filtered second sample set is composed of the filtered second samples.

2. The method of claim 1, wherein, the step of inputting the first sample set and the second sample set into the weighted space-spectrum joint kernel collaborative representation classification algorithm to obtain the classified hyperspectral remote sensing image specifically comprises the following steps: performing normalization processing and space filtering on the first sample set and the second sample set to obtain a filtered first sample set and a filtered second sample set; solving a representation coefficient vector matrix of the filtered first sample set for each filtered second sample and its corresponding spatially adjacent pixels in the filtered second sample set according to the filtered first sample set and the filtered second sample set; determining C intra-class coefficient vector matrices according to the representation coefficient vector matrix; wherein C represents the number of ground object categories; reconstructing the filtered second sample and its corresponding spatially adjacent pixels by using the filtered first sample set and the corresponding intra-class coefficient vector matrix to obtain C reconstructed second samples and their corresponding spatially adjacent pixels; calculating errors between each reconstructed second sample and its corresponding spatially adjacent pixels and the filtered second sample and its corresponding spatially adjacent pixels by using F norm to obtain C reconstruction errors; taking a ground object category corresponding to a minimum reconstruction error in the C reconstruction errors as the ground object category of the filtered second sample; determining the classified hyperspectral remote sensing image according to the ground object categories of all the filtered second samples.

3. The method of claim 2, wherein the method further comprises: the step of performing normalization processing and space filtering on the first sample set and the second sample set to obtain a filtered first sample set and a filtered second sample set specifically comprises the following steps: preprocessing the first sample set and the second sample set by using an amplitude normalization method to obtain a processed first sample set and a processed second sample set; Filter the processed first sample set and the processed second sample set by using a weighted spatial filtering operator to obtain a filtered first sample set and a filtered second sample set.

4. The method of claim 2, wherein the method further comprises: The expression coefficient vector matrix of the filtered first sample set to each second sample and the corresponding spatial adjacent pixels in the filtered second sample set is solved according to the filtered first sample set and the filtered second sample set, and specifically includes: A dictionary is constructed by using the filtered first sample set. The dictionary and the filtered second sample set are mapped into a high-dimensional feature space by using a kernel function to obtain a mapping function of the dictionary in the high-dimensional feature space and a mapping function of the filtered second sample and the spatial adjacent pixels of the filtered second sample under a spatial filtering window in the high-dimensional feature space. The expression coefficient vector matrix is determined according to the mapping function of the dictionary in the high-dimensional feature space and the mapping function of the filtered second sample and the spatial adjacent pixels of the filtered second sample under the spatial filtering window in the high-dimensional feature space.

5. The method of claim 4, wherein the method further comprises: The expression coefficient vector matrix is determined according to the mapping function of the dictionary in the high-dimensional feature space and the mapping function of the filtered second sample and the spatial adjacent pixels of the filtered second sample under the spatial filtering window in the high-dimensional feature space, and specifically includes: wherein the formula is a representation coefficient vector matrix; wherein ψ is a representation coefficient vector matrix; is a matrix of the filtered second sample and spatial neighboring pixels of the filtered second sample under a spatial filtering window, wherein is a filtered second sample, is an i-th spatial neighboring pixel of the filtered second sample under a spatial structure window; is a mapping function of the dictionary in a high-dimensional feature space; is a mapping function of the filtered second sample and spatial neighboring pixels of the filtered second sample under a spatial filtering window in a high-dimensional feature space; ψ * is an initial representation coefficient vector matrix; ||·|| F is a F-norm, and λ is a global regularization parameter.

6. A hyperspectral remote sensing imagery land cover classification system, comprising: It includes: The data acquisition module is used to acquire the hyperspectral remote sensing image of the land to be classified, and determine the three-dimensional matrix of the hyperspectral remote sensing image of the land to be classified. The conversion module is used to convert the three-dimensional matrix into a two-dimensional matrix. The sample division module is used to divide the two-dimensional matrix of the land hyperspectral remote sensing image into a first sample set with a real ground object class label and a second sample set without a label. The classification module is used to input the first sample set and the second sample set into a weighted space-spectrum joint kernel collaborative representation classification algorithm to obtain a classified hyperspectral remote sensing image. The weighted space-spectrum joint kernel collaborative representation classification algorithm is obtained by introducing a weighted spatial filtering operator and spatial structure information into a traditional collaborative representation classification algorithm. The processed first sample set and the processed second sample set are spatially filtered by using a weighted spatial filtering operator to obtain a filtered first sample set and a filtered second sample set, and specifically includes: The correlation coefficient of the processed sample and each spatial adjacent pixel of the processed sample under a spatial filtering window is calculated in the processed first sample set and the processed second sample set; the processed sample includes the processed first sample and the processed second sample. According to the correlation coefficient, a formula The spatial adjacent pixels are given weights; wherein, w 0,i The weight of the i-th spatial adjacent pixel under the spatial filtering window of the processed sample is represented as w 0.i The correlation coefficient between the processed sample and the i-th spatial adjacent pixel under the spatial filtering window of the processed sample is represented as r n×n is the size of the spatial filtering window; According to the weight of the spatial neighboring pixel, a formula is used The spatial neighboring pixels are weighted and averaged to obtain a filtered sample; the filtered sample includes a filtered first sample and a filtered second sample; wherein x 0,i represents the i-th spatial neighboring pixel of the processed sample under a spatial filtering window; represents the reconstructed sample; The filtered first sample set is formed by the filtered first sample, and the filtered second sample set is formed by the filtered second sample.

7. An electronic device, comprising: It includes: The memory is used to store a computer program, and the processor runs the computer program to make the electronic device execute the hyperspectral remote sensing image land cover classification method in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the land cover classification method of the hyperspectral remote sensing image according to any one of claims 1-5.

Images