EMMLP-based semi-supervised classification method for hyperspectral remote sensing image, storage medium and equipment

Abstract

The application discloses an EMMLP-based hyperspectral remote sensing image semi-supervised classification method, a storage medium and equipment, and relates to a hyperspectral remote sensing image semi-supervised classification method, a storage medium and equipment. In order to solve the problem that sample points cannot be classified by using an existing hyperspectral remote sensing image classification method, the application performs dimension reduction processing on a hyperspectral image; then, an MMLP algorithm is used to preliminarily classify the dimension-reduced image; for the sample that cannot be classified by using the MMLP algorithm, an MMLP secondary classification algorithm is used; finally, the result of the MMLP secondary classification algorithm is combined with spatial neighborhood information to judge the category of all samples again. The application is used for semi-supervised classification of a hyperspectral remote sensing image.

Description

Technical Field

[0001] This invention relates to a semi-supervised classification method, storage medium, and device for hyperspectral remote sensing images. Background Technology

[0002] Hyperspectral images can be viewed as a data cube, capturing detailed spectral information while simultaneously providing the spatial distribution of the observed scene. This integrated image-spectrum characteristic allows for more accurate object identification, offering immense possibilities for image research. After nearly two decades of research and development, it has been successfully applied in various scientific research fields, including geological science, military defense, medical diagnostics, precision agriculture, and hydrology. As an important direction in image classification, its classification problem has become a research hotspot. Researchers face challenges in classifying hyperspectral images due to high correlations between adjacent spectra and a limited number of training samples. Obtaining sample class labels requires significant cost and time, and the high number of channels and computational complexity further exacerbate the problem. To address these issues, this invention employs a semi-supervised classification method to solve the hyperspectral image classification problem.

[0003] The goal of hyperspectral image classification is to assign class labels to samples based on their features. Different land features have different spectral curves, leading to the emergence of numerous semi-supervised classification methods that utilize spectral information. Wang Liguo et al. combined the difference algorithm with the "teaching and learning" algorithm, proposing a hyperspectral semi-supervised classification method that integrates multiple spatial information, improving the algorithm's classification performance to some extent. In 2020, Cao et al. addressed the problem of redundant information in hyperspectral images by proposing a semi-supervised hyperspectral image classification algorithm based on Subspace Learning with Conditional Random Field (SLCRF). This method removes redundant information from hyperspectral images by introducing a 3D autoencoder and combines neighborhood information between adjacent pixels. Finally, it uses an algorithm based on the semi-supervised conditional random field framework to classify the hyperspectral dataset. Experiments verified the effectiveness of this algorithm in removing redundant information from hyperspectral images. In 2021, Haoyu Wang et al. proposed a semi-supervised hyperspectral image classification algorithm based on Graph Convolutional Broad Network (GCBN). This method improves upon traditional spectral feature extraction methods by extracting spectral spatial features from the original hyperspectral image through graph convolution operations and then applying a combined averaging method to extract sample features again for training. Experiments verified the effectiveness of the algorithm. Hyperspectral images may contain some redundant information in each spectral channel, and there is also some correlation between spectral channels, which increases the complexity of hyperspectral image classification.

[0004] The traditional MMLP algorithm is a semi-supervised classification method based on label propagation. It propagates class labels through minimax paths. This method has low time complexity and good classification performance. However, because the algorithm uses K-nearest neighbors to construct a sparse similarity matrix, some points may become disconnected during image classification, so the algorithm cannot completely classify all samples. Summary of the Invention

[0005] This invention aims to address the problem in existing hyperspectral remote sensing image classification methods where sample points are not classified.

[0006] A semi-supervised classification method for hyperspectral remote sensing images based on EMMLP includes the following steps:

[0007] S1. Reduce the dimensionality of the high-dimensional remote sensing image to obtain the dimensionality-reduced image;

[0008] S2. Classify the dimensionality-reduced image using the MMLP algorithm; the MMLP algorithm is the Minimax Label Propagation algorithm.

[0009] S3. For samples x that are still unclassified after classification by the MMLP algorithm. ij , using x ij Update the unlabeled sample set U before MMLP classification; update the labeled sample set M before MMLP classification using the samples with class labels obtained after MMLP classification.

[0010] Define the objective function K(x) ij ) represents unlabeled sample x ij To all labeled samples x j Distance:

[0011] K(x ij )=max(c(x ij ,x j ),c(x j for j=1,...,l

[0012] Among them, unlabeled sample x ij To have a labeled sample x j The distance between them is defined as c(x) ij x j ), x j Originally unlabeled samples, labeled samples with class names were obtained through the MMLP algorithm, c(x) j ) represents unlabeled sample x j The propagation distance of the label is obtained after the MMLP algorithm is completed;

[0013] Then, for all labeled samples, max(c(x) i ,xj ),c(x j Sort the samples, select the K smallest samples based on the sorting results, and use the class of the sample with the most classes as the class label of the unlabeled sample.

[0014] S4. Re-evaluate the labels of all samples in the initial unlabeled sample set: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] j and x ij Extract target pixel x as the target pixel. j and x ij The category information of the surrounding 8 neighborhoods is used to calculate the weight sum of each category and combine it with a voting method to correct the category of the target pixel.

[0015] Furthermore, the propagation distance c(x) j )as follows:

[0016]

[0017] Among them, X ij For unlabeled sample x j From the labeled sample x i Obtain the set of labeled paths, i.e., the set of successfully labeled paths; 'a' represents a path; define the Euclidean distance between two adjacent samples as c(x). l x l+1 ), max(c(x) l x l+1 )) represents the segment in the propagation path of unlabeled samples where the distance between any two samples is the largest.

[0018] Furthermore, in S1, the PCA or MNF algorithm is used to reduce the dimensionality of the high-dimensional remote sensing image;

[0019] PCA is principal component analysis, and MNF is minimum noise separation transform.

[0020] Furthermore, the 15 smallest samples are selected from S3, i.e., K=15.

[0021] Furthermore, the specific process of S4 includes the following steps:

[0022] Extract the category information of the 8 neighboring pixels around the target pixel, and statistically analyze the categories of the 8 neighboring pixels around the target pixel using D. k Record the sum of the class weights of pixels of the same class;

[0023] In the 8 neighboring pixels of the target pixel, each unlabeled sample that was initially unlabeled but obtained a class label from other samples has its class weight set to 1; the class weight of each sample that was initially labeled is set to 2, and the class label weight of the target pixel itself is also set to 2.

[0024] The weights of all classes for the target pixel and its eight surrounding samples are statistically analyzed, and the weight of each class and D are determined. k The maximum value in the formula is used to set the class label of the target pixel as the weight and D. k The maximum value corresponds to the category label.

[0025] Furthermore, determine the weights for each category and D. k The maximum value in the formula is used to set the class label of the target pixel as the weight and D. k In the process of determining the class label of the category corresponding to the maximum value, if there are two categories with the same weight, the class label of the target pixel will not change.

[0026] A computer storage medium storing a computer program, the computer program being loaded and executed by a processor to implement the EMMLP-based semi-supervised classification method for hyperspectral remote sensing images.

[0027] A semi-supervised classification device for EMMLP-based hyperspectral remote sensing images is disclosed. The device includes a processor and a memory. The memory stores a computer program, which is loaded and executed by the processor to implement the EMMLP-based hyperspectral remote sensing image semi-supervised classification method.

[0028] Beneficial effects:

[0029] This invention first uses two algorithms, MNF and PCA, to perform dimensionality reduction on the original hyperspectral image. This reduces the data dimensionality while preserving sufficient information, effectively addressing redundancy in spectral channels and improving the spatial distance measurement between data points. This also helps the classifier to classify samples more accurately. Second, the EMMLP classification algorithm is used to classify the images after dimensionality reduction using the two different methods. This algorithm improves upon the MMLP algorithm's inability to classify all sample points, thoroughly classifying unclassified sample points left by the MMLP algorithm. Finally, by combining the category information of the spatial neighbor pixels of all samples, the category of all samples is determined again, resulting in the final classification result. Attached Figure Description

[0030] Figure 1 shows the hyperspectral image of Pavia University. Figure 1(a) is a false-color composite image of this remote sensing image; Figure 1(b) is the standard classification image.

[0031] Figure 2 shows a hyperspectral remote sensing image of Salinas; Figure 2(a) is a false-color composite image of Salinas; Figure 2(b) is the standard classification image of this image.

[0032] Figure 3 shows the hyperspectral remote sensing image of Whu-Hi-Longkou; Figure 3(a) is a false-color composite image of Whu-Hi-Longkou; Figure 3(b) is the standard classification image.

[0033] Figure 4 shows the hyperspectral remote sensing image of Whu-Hi-Hanchuan; Figure 4(a) is a false-color composite image; Figure 4(b) is a standard classification image.

[0034] Figure 5 shows the hyperspectral remote sensing image of Whu-Hi-HongHu; Figure 5(a) is a false-color composite image of Whu-Hi-HongHu; Figure 5(b) is the standard classification image, with different colors used to mark different sample categories.

[0035] Figure 6 shows the classification results of different algorithms for the Pavia University dataset; Figures 6(a) to 6(i) show the classification results of the standard classification map, AGR, MMLP, PCA-MMLP', MNF-MMLP, PCA-MMLP', MNF-MMLP', EMMLP(PCA), and EMMLP(MNF), respectively.

[0036] Figure 7 shows the classification results of different Salinas algorithms; where Figures 7(a) to 7(i) The classification results are for the standard classification map, AGR, MMLP, PCA-MMLP, MNF-MMLP, PCA-MMLP', MNF-MMLP', EMMLP(PCA), and EMMLP(MNF), respectively.

[0037] Figure 8 shows the classification results of different algorithms for Whu-Hi-Longkou; Figures 8(a) to 8(i) The classification results are for the standard classification map, AGR, MMLP, PCA-MMLP, MNF-MMLP, PCA-MMLP', MNF-MMLP', EMMLP(PCA), and EMMLP(MNF), respectively.

[0038] Figure 9 shows the classification results of different algorithms for Whu-Hi-Hanchuan; Figures 9(a) to 9(i) The classification results are for the standard classification map, AGR, MMLP, PCA-MMLP', MNF-MMLP, PCA-MMLP', MNF-MMLP', EMMLP(PCA), and EMMLP(MNF), respectively.

[0039] Figure 10 shows the classification results of Whu-Hi-Honghu using different algorithms; Figures 10(a) to 10(i)The classification results are shown for the standard classification map, AGR, MMLP, PCA-MMLP'MNF-MMLP, PCA-MMLP', MNF-MMLP', EMMLP(PCA), and EMMLP(MNF), respectively.

[0040] Figure 11 A comparison chart showing the overall classification accuracy (%) of AGR, MMLP, MNF-MMLP, PCA-MMLP, MNF-MMLP', PCA-MMLP' and EMMLP(PCA) proposed in this invention.

[0041] Figure 12 A comparison chart showing the average classification accuracy (%) of AGR, MMLP, MNF-MMLP, PCA-MMLP, MNF-MMLP', PCA-MMLP' and EMMLP (MNF) and PCA proposed in this invention.

[0042] Figure 13 A comparison chart of Kappa (%) of AGR, MMLP, MNF-MMLP, PCA-MMLP, MNF-MMLP', PCA-MMLP' and EMMLP (MNF) and EMMLP (PCA) classifications proposed in this invention.

[0043] Figure 14 shows a comparison of classification times between AGR, MMLP, MNF-MMLP, PCA-MMLP, MNF-MMLP', PCA-MMLP' and the EMMLP(MNF) and EMMLP(PCA) proposed in this invention; where Figure 14(a) shows the classification time of Pavia and Salinas, and Figure 14(b) shows the classification time of Whu-Hi-Longkou, Whu-Hi-Honghu, and Whu-Hi-Hanchuan. Detailed Implementation

[0044] To address the issue of redundant information in hyperspectral remote sensing images and the unclassified points in the MMLP algorithm, this invention proposes a semi-supervised hyperspectral image classification algorithm using Enhanced Minimax Label Propagation (EMMLP). First, to address the redundant information in hyperspectral images, dimensionality reduction is performed using Minimum Noise Fraction (MNF) and Principal Component Analysis (PCA). Second, the MMLP algorithm is used for preliminary classification of the dimensionality-reduced image. For samples that do not receive class labels after MMLP classification, a secondary MMLP classification algorithm is applied to reclassify the data, resolving the unclassified point problem. Finally, the results of the secondary MMLP classification algorithm are combined with spatial neighborhood information to re-determine the class of all samples. Experimental results based on datasets such as Pavia University, Salinas, Whu-Hi-Longkou, Whu-Hi-Hanchuan, and Whu-Hi-Honghu demonstrate that the proposed method achieves good classification performance with fewer training samples and shorter classification time. The specific implementation details are described below. Specific implementation method one:

[0046] This implementation method is a semi-supervised classification method based on EMMLP hyperspectral remote sensing images, including the following steps:

[0047] Step 1, Image Dimensionality Reduction:

[0048] Hyperspectral remote sensing images are characterized by large data volumes and high redundancy. The high dimensionality of these data negatively impacts the spatial complexity of classification algorithms. Generally, high-dimensional data implies excessively redundant data features, hindering the search for a classification hyperplane in the feature space. To address the curse of dimensionality in hyperspectral images, dimensionality reduction methods can be employed to transform the spectral information in the original space into a lower-dimensional feature space, minimizing energy loss during the transformation process. This invention utilizes both PCA and MNF dimensionality reduction algorithms for hyperspectral image dimensionality reduction. These two algorithms are independent of each other. Subsequent experiments demonstrate that different datasets show varying classification times and accuracy rates after dimensionality reduction using PCA and MNF algorithms, followed by classification using the EMMLP algorithm. This verifies that different datasets achieve different dimensionality reduction effects when using different dimensionality reduction algorithms.

[0049] Step 2, EMMLP Algorithm Classification:

[0050] After dimensionality reduction, the original remote sensing image is initially classified using an improved algorithm based on the Minimax Label Propagation (MMLP) algorithm, a semi-supervised classification algorithm based on label propagation.

[0051] The MMLP algorithm first constructs a sparse graph, in which samples are only connected to their K nearest neighbors.

[0052] Secondly, labeled samples use a minimax algorithm to find a unique propagation path to transmit their class information. The algorithm updates its own maximum distance by comparing the maximum distances of unlabeled samples with their K nearest neighbors. There may be multiple paths for the class label of a labeled sample to be transmitted to an unlabeled sample, but each path has a maximum maximum distance value. The final path that completes the class label propagation is the one with the smallest maximum distance among the multiple possible paths.

[0053] The MMLP algorithm constructs a sparse nearest neighbor graph: it finds the k nearest neighbors of all samples in the sample set as the nodes of the graph, and the sample nodes connected by edges are each other's nearest neighbors. Typically, the number of nearest neighbors k is set to a range of 5-20, thus limiting the number of edges between related nodes. The nearest neighbor graph mainly improves the computational efficiency of the algorithm by preserving the sparsity of the graph, and to some extent, prevents the propagation of erroneous information between semantically unrelated nodes. The pseudocode of this algorithm is shown in Table 1.

[0054] Table 1

[0055]

[0056]

[0057] In Table 1, B k ={k|(i,k∈ε)} is a set of samples adjacent to k, v records the maximum distance of each sample during class label propagation, and f i y is the class label of the sample. i The class label is the class label of the sample with class label. Define the maximum distance for each sample in the initial labeled sample set. f i (0) Let f be the class label of each sample in the initial labeled sample set. i (0) =y i (During subsequent label propagation, the maximum distance of initially labeled samples remains constant at 0, and their class labels also remain unchanged); the initial value of the maximum distance of unlabeled samples is defined as: The class labels for samples in the unlabeled sample set are defined initially as follows: Define Q (m) as the set of labeled samples and newly added sample points, and m as the loop variable. Each time a sample is taken from Q (m) queue, traverse the neighbors i of the sample, i ∈ B k , v i and f i are updated before iteration, and the class labels are propagated through the samples in Q (m) . When Q (m) is empty, the algorithm ends.

[0058] The MMLP algorithm cannot classify all samples. It performs well in small-scale image classification, but in large-scale images, the classification effect is not ideal. The approach of this invention is to improve its idea after the MMLP algorithm. Experimental results prove that for larger-scale images, the classification results are relatively ideal.

[0059] After the MMLP algorithm finishes classifying hyperspectral images, the class labels of some unlabeled samples and their minimax distances can be obtained. However, the MMLP algorithm has the problem of sample non-connectivity when constructing a sparse graph, resulting in the possibility that it may not be able to classify all samples when classifying images with a large number of samples. To address this problem, this invention proposes the MMLP secondary classification algorithm, namely the EMMLP algorithm. After the MMLP algorithm finishes classifying, it continues to check whether there are unclassified samples x i . If there are, then put x i into the unlabeled sample set, and add the unlabeled samples with classes obtained after the MMLP algorithm classification to the labeled sample set to update the labeled sample set. The specific process includes the following steps:

[0060] After preliminary classification by the MMLP algorithm, let the sample set of all data before classification be the total data set, which contains two types of samples: labeled samples and unlabeled samples; n represents the number of samples of all data. Define the number of labeled samples as l (l << n), and the samples other than the labeled samples are all unlabeled samples, denoted as the (l + 1)-th to the n-th samples. The purpose of this invention is to finally obtain the class labels of all unlabeled samples through semi-supervised classification. After preliminary classification by the MMLP algorithm, some unlabeled samples x j obtain class labels, and c(x j ) is the propagation cost of the sample x j obtaining the class label, that is, the minimax distance of x j :

[0061]

[0062] Among them, X ij For unlabeled sample x j From the labeled sample x i Obtain the set of labeled paths, i.e., the set of successfully labeled paths; 'a' represents a path. The Euclidean distance between two adjacent samples is defined as c(x...). l x l+1 ), max(c(x) l x l+1 )) represents the segment with the largest distance between any two samples in the propagation path of unlabeled samples.

[0063] x ij The original unlabeled samples are those that were labeled after the MMLP algorithm labeled the samples, but still have not been labeled.

[0064] Define the objective function K(x) ij ) represents unlabeled sample x ij To all labeled samples x j (Partial unlabeled samples x) j The class labels were obtained, therefore there are class-labeled samples x. j In reality, it means that a sample without a class label has obtained a labeled sample x. j The distance of x, where x j The maximum distance of a point may be the maximum distance obtained through the MMLP algorithm, or it may be 0, i.e., x. j The points belong to the initial labeled sample set after the MMLP algorithm classification is completed:

[0065] K(x ij )=max(c(x ij ,x j ),c(x j for j=1,...,l (15)

[0066] Among them, unlabeled sample x ij To have a labeled sample x j The distance between them is defined as c(x) ij x j ), x j Originally unlabeled samples, labeled samples with class names were obtained through the MMLP algorithm, c(x) j ) represents unlabeled sample x j The propagation distance of the label is obtained after the MMLP algorithm is completed.

[0067] The idea behind minimax propagation is that when label i is passed to j, it passes through countless nodes and there are many possible paths. The final path is determined by the minimum maximum distance between any two nodes across all paths. Therefore, we filter out the maximum distance between two nodes in each possible path, and the path containing the maximum distance is the shortest distance for the two labels to propagate, also known as the minimum cost.

[0068] In this invention, for node x that did not propagate during the first classification in MMLP... ij It is necessary to compare "x" ij To x j "propagation distance" and "x" j The distance obtained in the MMLP algorithm is used to filter out the maximum distance, which can improve x. ij The accuracy of obtaining the labels.

[0069] max(c(x) for all labeled samples i ,x j ),c(x j The samples are sorted, and the K smallest samples are selected based on the sorting results. In this experiment, K=15. According to the voting method, the class label corresponding to the class with the most samples of the same class label is selected as sample x. ij The class tag.

[0070] Both the MMLP secondary classification algorithm and the MMLP algorithm are suitable for image classification problems with a small number of categories and a large number of samples. As the data scale increases, the number of unclassified points also increases after the MMLP classification is completed. The MMLP secondary classification algorithm solves the problem of unclassified points based on the MMLP algorithm.

[0071] Step 3: Post-processing with neighborhood information:

[0072] After the MMLP secondary classification algorithm completes, all samples receive class labels. Combining spatial neighborhood information, and based on the characteristic that hyperspectral image pixels within the same adjacent region have a high probability of belonging to the same land cover category, a weighted voting method is used to re-determine the category of all samples after the MMLP secondary classification algorithm (re-determining the labels of all samples in the initial unlabeled sample set): Sample x... j and x ij (x j The label was obtained after the first MMLP session, x ij The labels are obtained from the initial MMLP algorithm classification and subsequent steps (regardless of which step yields the labels; this step re-evaluates the labels of all samples in the initial unlabeled sample set) and are used as the target pixels to extract the target pixel x. j and xij The category information of the surrounding 8 neighborhoods is used to calculate the weight sum of each category and combine it with a voting method to correct the category of the target pixel.

[0073] The specific steps are as follows:

[0074] (1) Sample x j and x ij As the target pixel; extract the target pixel x. j and x ij Category information of the surrounding 8 neighborhoods, for the target pixel x j and x ij The categories of the surrounding 8 neighboring pixels are statistically analyzed using D. k k = 1, ..., n', to record the sum of the class weights of pixels of the same class; n' represents the number of classes;

[0075] x j and x ij In the 8 neighboring pixels, each unlabeled sample that was initially unlabeled but obtained a class label from other samples has its class weight set to 1; each initially labeled sample has its class weight set to 2. j and x ij Its own class label weight is also set to 2;

[0076] Statistics x j and x ij The weights of all classes of the sample and its eight neighboring samples are used to determine the weight of each class and D. k The maximum value in x j and x ij The class label is set to weight and D k The maximum value corresponds to the class label of the category. A special case is also considered: when the sum of the weights of two categories is the same after statistical analysis, x... j and x ij The class tag remains unchanged.

[0077] The above method further improves the classification accuracy by combining the category information of spatially neighboring pixels to re-evaluate the classification results of the MMLP secondary classification algorithm. Specific Implementation Method Two:

[0079] This embodiment is a computer storage medium that stores a computer program. The computer program is loaded and executed by a processor to implement the semi-supervised classification method for EMMLP-based hyperspectral remote sensing images.

[0080] It should be understood that the storage media described in this embodiment include, but are not limited to, magnetic storage media and optical storage media; the magnetic storage media include, but are not limited to, RAM, ROM, and other storage media such as hard disks and USB flash drives. Specific implementation method three:

[0082] This embodiment is a semi-supervised classification device based on EMMLP hyperspectral remote sensing images. The device includes a processor and a memory. The memory stores a computer program, which is loaded and executed by the processor to implement the semi-supervised classification method based on EMMLP hyperspectral remote sensing images.

[0083] It should be understood that the device described in this embodiment includes, but is not limited to, devices including processors and memory, and may also include other devices corresponding to units or modules with information acquisition, information interaction, and control functions, such as signal acquisition devices. The device includes, but is not limited to, PCs, workstations, and mobile devices.

[0084] Example

[0085] 1. Experimental Environment and Parameter Settings

[0086] To test the performance of the EMMLP algorithm in classifying hyperspectral remote sensing datasets, this invention conducted experiments.

[0087] Experimental environment:

[0088] CPU: Intel(R) 1.9GHz dual-core; Memory: 16GB; Operating System: 64-bit Windows 10; Software Framework: Matlab2018b.

[0089] Five hyperspectral remote sensing images were selected for the experiment, including Pavia University, Salinas, Whu-Hi-Longkou, Whu-Hi-Hanchuan, and Whu-Hi-Honghu. The five selected images differed in size and the number of land cover categories they contained, in order to verify the classification performance of the proposed EMMLP algorithm on different hyperspectral remote sensing images.

[0090] To verify the performance of the proposed algorithm in classifying hyperspectral images, the following comparisons are made in several aspects: classification accuracy under different parameters, classification accuracy after dimensionality reduction, resolution of unclassified points, and joint spatial-spectral post-processing. It should be noted that background samples were not included in the calculation of the classification results during the experiments. The evaluation metrics used in the experiments included: overall classification accuracy, average classification accuracy, classification time, and Kappa coefficient.

[0091] It should be noted that, in order to distinguish between the MMLP algorithm and the MMLP quadratic classification algorithm in the experiment, the MMLP quadratic classification algorithm is denoted as MMLP', which is the EMMLP algorithm; in the experiment, the EMMLP algorithm used the PAC algorithm and the MNF algorithm for dimensionality reduction respectively. To distinguish them, the algorithms of this invention are denoted as EMMLP(PCA) and EMMLP(MNF) respectively.

[0092] 2. Introduction to the dataset

[0093] (1) Hyperspectral remote sensing image of Pavia University

[0094] The Pavia University image used in the experiment is a topographical image of the Italian university Pavia, taken in 2003 using the ROSIS sensor. The image contains 115 bands; 103 bands were used for image classification, while the remaining 12 noise bands were removed. The image has a spatial resolution of 1.3 m and a spectral range of 0.43–0.68 μm. The Pavia University image contains 9 types of land features, and the image resolution is 610 × 340 pixels.

[0095] Figure 1 shows the hyperspectral image of Pavia University. Figure 1(a) is a false-color composite image of the remote sensing image; Figure 1(b) is the standard classification image, where different sample categories are labeled with different colors.

[0096] Table 2 shows the sample category names and sample numbers for each land cover in the remote sensing image of Pavia University.

[0097] Table 2. Various land features in Pavia University remote sensing images.

[0098]

[0099]

[0100] (2) Salinas hyperspectral remote sensing images

[0101] The Salinas images used in the experiment were taken in 1998 in the Salinas Valley, California, using an AVIRIS sensor. Twenty water-absorbing bands were removed, and 204 undisturbed bands were used. The Salinas images contain 16 land cover types, with a resolution of 3.7 m and a pixel size of 512 × 217.

[0102] Figure 2 shows a hyperspectral remote sensing image of Salinas; Figure 2(a) is a false-color composite image of Salinas; Figure 2(b) is the standard classification image of this image, with different colors representing different land cover species.

[0103] Table 3 shows the sample category names and sample numbers for each land cover in the Salinas remote sensing images.

[0104] Table 3. Various land features in Salinas remote sensing images.

[0105]

[0106] (3) Hyperspectral remote sensing images of Whu-Hi-Longkou

[0107] The Whu-Hi dataset used in the experiment is a publicly shared hyperspectral and hyperspatial (dual-high) remote sensing imagery dataset for fine-grained classification of land cover, constructed by the RSIDEA team at Wuhan University. This dataset contains a wide variety of crops. The Whu-Hi-Longkou dataset is one of the three typical scene datasets of Whu-Hi, primarily a hyperspectral fine-grained classification dataset for agricultural scenes, collected and created by the RS-IDEA research group (SIRI-Whu) at Wuhan University. This dataset was acquired using an endwall nanometer hyperspectral sensor mounted on an unmanned aerial vehicle (UAV) platform in agricultural areas of different crop types in Hubei Province, China. The remote sensing images have 270 bands, a spatial resolution of 0.463 m, a spectral resolution of 6 nm, and a size of 550 × 400 pixels. There are a total of 9 different land cover categories, 6 of which are crops.

[0108] Figure 3 shows a hyperspectral remote sensing image of Whu-Hi-Longkou; Figure 3(a) is a false-color composite image of Whu-Hi-Longkou; Figure 3(b) is a standard classification image, with different colors used to mark different sample categories.

[0109] Table 4 shows the sample category names and sample numbers for each land cover in the Whu-Hi-Longkou remote sensing dataset.

[0110] Table 4. Various land features in remote sensing images of Whu-Hi-Longkou

[0111]

[0112] (4) Hyperspectral remote sensing images of Whu-Hi-Hanchuan

[0113] The Whu-Hi-Hanchuan dataset primarily depicts agricultural scenes in suburban areas. This remote sensing image contains 274 bands. The spatial resolution is 0.109 m, the spectral resolution is 6 nm, and the image size is 1217×303. It includes 16 different land cover categories, 8 of which are crops.

[0114] Figure 4 shows the hyperspectral remote sensing image of Whu-Hi-Hanchuan; Figure 4(a) is a false-color composite image; Figure 4(b) is a standard classification image.

[0115] Table 4 shows the sample category names and sample numbers for each land cover in the Whu-Hi-Hanchuan remote sensing dataset.

[0116] Table 5. Various land features in the remote sensing images of Wuhan-Hi-Hanchuan

[0117]

[0118]

[0119] (5) Hyperspectral remote sensing images of Whu-Hi-HongHu

[0120] The Whu-Hi-HongHu dataset primarily depicts complex agricultural scenes with diverse crop types. This remote sensing imagery comprises 270 spectral bands. The hyperspectral dataset includes 22 land cover categories, 18 of which are crops. The spatial resolution is 0.043 m, the spectral resolution is 6 nm, and the image pixel size is 940 × 475.

[0121] Figure 5 shows the hyperspectral remote sensing image of Whu-Hi-HongHu; Figure 5(a) is a false-color composite image of Whu-Hi-HongHu; Figure 5(b) is the standard classification image, with different colors used to mark different sample categories.

[0122] Table 6 shows the sample category names and sample numbers for each land cover in the Whu-Hi-Honghu remote sensing dataset.

[0123] Table 6. Land cover information in the remote sensing images of Whu-Hi-HongHu

[0124]

[0125]

[0126] 3. The impact of different parameters on classification performance

[0127] This section primarily evaluates the impact of the proportion of unlabeled samples and the number of nearest neighbors on classification performance. During the experiments, we found that when using PCA and MNF to reduce the dimensionality of hyperspectral images, reducing the dimension to 10 allows for the inclusion of over 95% of the principal components of the original hyperspectral image. Furthermore, the EMMLP algorithm achieves optimal classification accuracy and time when classifying the dimensionality-reduced image. Therefore, this section uniformly uses PCA and MNF to reduce the original hyperspectral dataset images to 10 dimensions for comparison, and the experimental results are the average of 10 experiments. In this subsection, 5% of the samples from each class are randomly selected as the labeled sample set, and the remaining samples are the unlabeled sample set.

[0128] Table 7 shows the classification accuracy and classification time of the algorithm of the present invention in various remote sensing datasets with different proportions of labeled samples.

[0129] Table 7 Classification accuracy and classification time under different sample proportions

[0130]

[0131] As shown in Table 7, the algorithm of this invention demonstrates good classification accuracy in most remote sensing datasets with only a small percentage (5%) of labeled samples. Furthermore, the classification accuracy increases with the increase in the sample ratio. For example, when the percentage of labeled samples increases to 30%, the classification accuracy of the Whu-Hi-Longkou dataset reaches 99.74%. Since the sample ratio affects the mapping and classification time of the MMLP algorithm, as the sample ratio increases, the number of unclassified points decreases, thus gradually reducing the overall algorithm time. However, the impact of the sample ratio on time is still relatively small compared to the overall algorithm cost. For instance, in the Whu-Hi-Hanchuan dataset, the classification time of the EMMLP (PCA) algorithm is 63.87 seconds and 61.40 seconds with a sample ratio of 5% and 30%, respectively, a difference of only 2.47 seconds, indicating that a smaller sample ratio does not significantly increase the time cost. This conclusion is particularly evident in the classification of hyperspectral remote sensing images with a large number of samples. For example, in the Whu-Hi-LongKou dataset, the EMMLP (MNF) algorithm achieved classification accuracies of 98.73% and 99.73% with 5% and 30% sample ratios, respectively, but the classification time differed by only 2.77 seconds. This also verifies that the algorithm of this invention has good classification accuracy with low time complexity and a small number of labeled samples. Therefore, the algorithm of this invention is more suitable for the classification problem of hyperspectral remote sensing images with a large number of samples.

[0132] Table 8 compares the classification accuracy and classification time of the algorithm proposed in this invention under different numbers of nearest neighbors.

[0133] Table 8 Classification accuracy and classification time for different numbers of nearest neighbors.

[0134]

[0135] As shown in Table 8, the algorithm proposed in this invention exhibits minimal impact on classification accuracy across most remote sensing datasets, regardless of the number of nearest neighbors (5-50). For example, in the Whu-Hi-Longkou dataset, the EMMLP (MNF) algorithm achieves a classification accuracy of 98.19% with 5 nearest neighbors and 98.17% with 30 nearest neighbors. However, increasing the number of nearest neighbors affects the mapping and classification time of the MMLP algorithm. The algorithm's time consumption gradually increases with the number of nearest neighbors, and the impact of the number of nearest neighbors on classification time is significant; that is, while maintaining high classification accuracy, the time consumption increases linearly with the number of nearest neighbors. For instance, in the Whu-Hi-Honghu dataset, the EMMLP (PCA) method increased classification time by 9.01 seconds when the number of nearest neighbors was adjusted from 5 to 10, but the classification accuracy only improved by 1.04%. This demonstrates that the number of nearest neighbors has a significant impact on mapping time but a relatively small impact on classification accuracy. This conclusion is particularly evident in the classification of hyperspectral remote sensing images (Whu-Hi-Honghu) with a large number of image samples. The EMMLP (PCA) algorithm achieves a classification accuracy of 82.48% when the number of nearest neighbors is 5, and 83.44% when the number of nearest neighbors is 50. Since the algorithm of this invention maintains good classification accuracy even with significant variations in the number of nearest neighbors, only increasing time overhead, the classification accuracy of this algorithm is less affected by the number of nearest neighbors during image construction when dealing with image classification problems with a large number of samples.

[0136] 4. Comparison of Experimental Results

[0137] This section verifies the classification performance of the algorithm on different datasets. The algorithms compared include: AGR, MMLP, MNF-MMLP, PCA-MMLP, MNF-MMLP', PCA-MMLP', and the proposed EMMLP (MNF) and EMMLP (PCA) algorithms. It should be noted that in the experiments, 5% of the samples in each dataset were randomly selected as labeled samples, and the rest were unlabeled samples. The experimental results of the proposed algorithm and its comparison algorithms on the Pavia University dataset for each category are shown in Table 9.

[0138] Table 9. Classification accuracy (%) of different algorithms used by Pavia University

[0139]

[0140] As shown in Table 9, in the hyperspectral remote sensing image scene of Pavia University, the PCA-MMLP and MNF-MMLP algorithms not only improved the classification accuracy compared to the MMLP algorithm, but also reduced the classification time. However, the classification time was reduced compared to the MMLP algorithm. This is because after dimensionality reduction by MNF and PCA, redundant information in the image was removed, the dimensionality of the hyperspectral image was reduced, the data size was reduced, the classification time of the algorithm was shortened, and the classification accuracy was improved.

[0141] The classification results of AGR, MMLP, MNF-MMLP, PCA-MMLP, MNF-MMLP', PCA-MMLP', and EMMLP(MNF), EMMLP(PCA) algorithms on the Pavia University dataset are compared below, as shown in Figure 6. Figure 6 shows the classification results of different algorithms on the Pavia University dataset; Figures 6(a) to 6(i) are the standard classification map, AGR, MMLP, PCA-MMLP', MNF-MMLP, PCA-MMLP', MNF-MMLP', EMMLP(PCA), and EMMLP(MNF), respectively.

[0142] The following table compares the classification results of the AGR, MMLP, MNF-MMLP, PCA-MMLP, MNF-MMLP', PCA-MMLP', EMMLP(MNF), and EMMLP(PCA) algorithms for each category in the Salinas dataset. See Table 9 for details.

[0143] Table 10. Classification accuracy (%) of different algorithms on the Salinas dataset.

[0144]

[0145] As shown in Table 10, in Salinas hyperspectral remote sensing image scenarios, the classification accuracy of the MMLP algorithm after dimensionality reduction using PCA and MNF is significantly improved. Both the PCA-MMLP' and MNF-MMLP' algorithms show improved classification accuracy, thanks to the MMLP' algorithm addressing the issue of unclassified points in the MMLP algorithm, and finally combining spatial neighborhood information for further judgment. The method of this invention further improves the classification accuracy based on the dimensionality-reduced MMLP' algorithm.

[0146] From the perspective of classification performance for different land cover types, since MMLP generally performs well in Salinas image scenes, we only selected C8 (Grapes_untrained) and C15 (Vinyard_untrained) with classification accuracy below 95% for comparison. After dimensionality reduction of the images using PCA and MNF, and then classifying the PCA and MNF-reduced data using the MMLP algorithm, the classification accuracy of C8 (Grapes_untrained) and C15 (Vinyard_untrained) showed almost no improvement.

[0147] However, after combining spatial neighborhood information to re-evaluate all sample categories of PCA-MMLP' and MNF-MMLP', the previously almost ineffective C8 (Grapes_untrained) and C15 (Vinyard_untrained) classification accuracy was significantly improved, reaching 90% and 99% respectively, showing a marked improvement in performance.

[0148] The above results demonstrate that, in Salinas image scenarios, the algorithm of this invention effectively improves the low classification accuracy of the MMLP algorithm for ground features. In terms of time overhead, the EMMLP (PCA) algorithm shows relatively low time cost increases at 5% and 30% sample ratios. The actual increase in algorithm cost is mainly due to the fact that MMLP needs to calculate the nearest neighbor distance between labeled and unlabeled samples; the significant increase in labeled samples increases the complexity of the nearest neighbor distance calculation.

[0149] Furthermore, as shown in Table 10, the EMMLP (MNF) and EMMLP (PCA) algorithms have higher Kappa coefficients compared to the MMLP algorithm. Specifically, the Kappa coefficients of both the EMMLP (MNF) and EMMLP (PCA) algorithms exceed 0.9, indicating that the classification results are almost completely consistent (Kappa coefficient of 0.81-1). This demonstrates that the algorithm of this invention has good classification performance on high-dimensional remote sensing images. Experiments also verify the classification performance of the algorithm in processing complex images.

[0150] Figure 7 shows the classification results of different Salinas algorithms; where Figures 7(a) to 7(i) The classification results are for the standard classification map, AGR, MMLP, PCA-MMLP, MNF-MMLP, PCA-MMLP', MNF-MMLP', EMMLP(PCA), and EMMLP(MNF), respectively.

[0151] As shown in Figure 7, noise in category C9 (Soil_vinyard_develop) was suppressed after dimensionality reduction. This fully demonstrates that dimensionality reduction removes redundant information in the hyperspectral image, thereby improving classification accuracy. The classification results from the EMMLP (MNF) and EMMLP (PCA) algorithms show fewer noise points and demonstrate good characterization of hyperspectral remote sensing images with a small sample size. MMLP classification results are less ideal, exhibiting more noise. For example, MMLP has more errors in category C8, resulting in more noise in the image classification results. This fully demonstrates that the method of combining spatial neighborhood information helps to further reduce noise in the classification results while improving the algorithm's classification accuracy.

[0152] The following table compares the classification results of the proposed EMMLP (MNF) and EMMLP (PCA) algorithms in the larger hyperspectral remote sensing dataset Whu-Hi-Longkou using AGR, MMLP, MNF-MMLP, PCA-MMLP, MNF-MMLP', and PCA-MMLP'. See Table 10 for details.

[0153] Table 11 Classification accuracy (%) of different algorithms in the Whu-Hi-Longkou algorithm

[0154]

[0155]

[0156] As can be seen from Table 11, PCA-MMLP' and MNF-MMLP' have improved classification accuracy compared to PCA-MMLP and MNF-MMLP algorithms. This is because the MMLP' algorithm solves the problem of unclassified points after image classification by the MMLP algorithm.

[0157] The comparison shows that the classification accuracy of the EMMLP (MNF) and EMMLP (PCA) classification methods proposed in this invention is significantly improved compared with PCA-MMLP' and MNF-MMLP', at 5.53% and 5.63% respectively. This is mainly due to the re-judgment of sample categories by combining spatial neighborhood information, especially in land cover categories C3 (Gravel) and C5, where the classification accuracy is improved by 11% and 10% respectively.

[0158] In the Whu-Hi-Longkou hyperspectral remote sensing image scenario, the MMLP algorithm, after dimensionality reduction using PCA and MNF, showed a significant improvement in classification accuracy, increasing by 29.05% and 29.37% respectively. This indicates that dimensionality reduction removes a significant amount of redundant information, resulting in a substantial improvement in classification accuracy. Both PCA-MMLP' and MNF-MMLP' algorithms show improvement compared to the PCA-MMLP and MNF-MMLP algorithms. This is mainly due to the MMLP' algorithm effectively addressing the issue of unclassified points left by MMLP. Finally, by combining spatial neighborhood information, the classification accuracy of the EMMLP(PCA) and EMMLP(MNF) algorithms was further improved.

[0159] From the classification results of different land cover types, the classification accuracy of land cover category C2 (Cotton) was significantly improved after dimensionality reduction of the image using PCA and MNF. Secondly, the classification accuracy of C2 (Cotton) was further improved by using the MMLP' algorithm to classify the data after dimensionality reduction of PCA and MNF, while the classification accuracy of C4 (Round Leaf Soybean) decreased significantly. Finally, by combining spatial neighborhood information to re-determine the categories of all samples in PCA-MMLP' and MNF-MMLP', the classification accuracy of C2 (Cotton) improved by 59% and 57% respectively, and the classification accuracy of C4 (Round Leaf Soybean) also improved significantly by 11% and 10%, reaching 99%. Ultimately, the classification accuracy of all land cover types improved to varying degrees, with land cover C7 (Water body) reaching 100%.

[0160] In terms of time consumption, the EMMLP (PCA) algorithm shows little improvement in classification time at 5% and 30% sample ratios. The improvement in classification time is mainly due to the complexity of the sample nearest neighbor relationships when MMLP constructs the sparse nearest neighbor matrix.

[0161] The results above show that the algorithm of this invention has the advantage of lower time consumption compared with other algorithms, and the overall time is even lower than that of the MMLP algorithm, with a significant improvement in efficiency.

[0162] Figure 8 shows the classification results of different algorithms for Whu-Hi-Longkou; Figures 8(a) to 8(i) The classification results are for the standard classification map, AGR, MMLP, PCA-MMLP, MNF-MMLP, PCA-MMLP', MNF-MMLP', EMMLP(PCA), and EMMLP(MNF), respectively.

[0163] As shown in Figure 8, in category C2, the algorithm of this invention effectively removes noise from the image by firstly reducing dimensionality and combining spatial neighborhood information for judgment; secondly, the MMLP' algorithm is used to classify the data after dimensionality reduction by PCA and MNF. Although the classification accuracy of C7 (Bitumen) is improved, the dimensionality reduction effect is almost unimproved in categories C1 (Asphalt) and C4 (Trees). However, after combining the spatial neighborhood information method, the classification accuracy of these two categories is improved by 19% and 10%, respectively. In particular, the classification accuracy of C3 (Gravel) is improved by 31% and 46% respectively after combining spatial neighborhood information. The above results indicate that the classification accuracy of the algorithm of this invention is significantly improved in most ground feature scenarios.

[0164] In terms of time consumption, the classification time of the EMMLP (PCA) algorithm is not significantly different with a sample ratio of 5% and 30%, indicating that a smaller sample ratio does not increase the time consumption significantly.

[0165] As shown in Figure 8, the MMLP algorithm suffers from incomplete image classification and numerous noise points due to the presence of many unclassified sample points. The algorithm of this invention maintains good overall classification accuracy, effectively suppresses skip points in the image, results in a smoother overall image, and achieves better classification performance.

[0166] The following table compares the classification results of AGR, MMLP, MNF-MMLP, PCA-MMLP, MNF-MMLP', PCA-MMLP' with the EMMLP(MNF) and EMMLP(PCA) proposed in this invention in the Whu-Hi-Hanchuan dataset. See Table 12 for details.

[0167] Table 12 Classification accuracy (%) of different algorithms for the Whu-Hi-Hanchuan dataset

[0168]

[0169]

[0170] As shown in Table 12, in the Whu-Hi-Hanchuan hyperspectral remote sensing image scene, the classification accuracy of the MMLP algorithm after dimensionality reduction using PCA and MNF is significantly improved, increasing by 26.30% and 24.51% respectively. This indicates that dimensionality reduction removes a significant amount of redundant information, resulting in a substantial improvement in classification accuracy. After using the MMLP' algorithm, PCA-MMLP and MNF-MMLP achieve improvements of 5.24% and 4.62% respectively. This is mainly due to the MMLP' algorithm effectively addressing the issue of unclassified points left by MMLP, leading to further improvements in the algorithm. Finally, by combining spatial neighborhood information, the classification accuracy of the PCA-MMLP' and MNF-MMLP' algorithms improves by 7.91% and 7.38% respectively.

[0171] From the classification results of different land cover types, after dimensionality reduction of the image using PCA and MNF, the classification accuracy of land cover category C14 (Road) improved from 40% to 80% and 79%, respectively. This indicates that dimensionality reduction using PCA and MNF helps MMLP improve its classification ability for samples with high feature complexity. Secondly, using the MMLP' algorithm to classify the data after dimensionality reduction using PCA and MNF, the classification accuracy of C14 (Road) improved from almost 0 to 80% and 78%, respectively, while the classification accuracy of C2 (Cowpea) decreased slightly. Finally, combining spatial neighborhood information to re-determine all sample categories in PCA-MMLP' and MNF-MMLP', the classification accuracy of C14 (Road) and the previously almost ineffective C2 (Cowpea) both showed significant improvements. This is mainly because spatial neighborhood information fully considers the spatial location information between samples, which helps to further remove noise interference. In terms of time cost, the algorithm of this invention maintains good stability in terms of time cost, and the increase in algorithm time cost with the increase of labeled samples is relatively small.

[0172] The above results demonstrate that the algorithm of this invention significantly improves the classification of high-dimensional remote sensing images with a wide variety of ground features and a large amount of data.

[0173] Figure 9 shows the classification results of different algorithms for Whu-Hi-Hanchuan; Figures 9(a) to 9(i) The classification results are for the standard classification map, AGR, MMLP, PCA-MMLP', MNF-MMLP, PCA-MMLP', MNF-MMLP', EMMLP(PCA), and EMMLP(MNF), respectively.

[0174] Table 12 presents a comparison of the classification results of AGR, MMLP, MNF-MMLP, PCA-MMLP, MNF-MMLP', PCA-MMLP' and EMMLP(MNF), EMMLP(PCA) algorithms for each category in the Whu-Hi-Honghu hyperspectral remote sensing dataset, which has a larger sample size. See Table 12 for details.

[0175] Table 13 Classification accuracy (%) of different algorithms for the Whu-Hi-Honghu dataset

[0176]

[0177] As shown in Table 132, in the Whu-Hi-HongHu hyperspectral remote sensing image scene, the MMLP algorithm after dimensionality reduction using PCA and MNF shows a significant improvement in classification accuracy. This indicates that dimensionality reduction removes a considerable amount of redundant information, resulting in a substantial improvement in classification accuracy. After using the MMLP' algorithm, PCA-MMLP and MNF-MMLP improved by 12.05% and 6.26%, respectively. This is mainly due to the MMLP' algorithm effectively addressing the issue of unclassified points left by MMLP. Finally, by combining spatial neighborhood information EMMLP (MNF) and EMMLP (PCA), the algorithm further removes erroneous noise points in different land cover regions, thus achieving further improvements. From the classification results of different land cover types, after dimensionality reduction of the image using PCA and MNF, the classification accuracy of C4(Cotton) improved by 9% and 4%, respectively. Secondly, further classification using the MMLP' algorithm significantly improved the classification accuracy of C4(Cotton). Finally, by combining spatial neighborhood information to re-determine the categories of all samples in PCA-MMLP' and MNF-MMLP', the classification accuracy of C4(Cotton) improved by 15%, approaching 100%. However, eight land cover types received incorrect category labels due to misclassification of spatial neighbors, resulting in a decrease in classification accuracy. In terms of time overhead, the algorithm's time overhead increased relatively steadily, and the classification time was short.

[0178] Figure 10 shows the classification results of Whu-Hi-Honghu using different algorithms; Figures 10(a) to 10(i) The classification results are shown for the standard classification map, AGR, MMLP, PCA-MMLP'MNF-MMLP, PCA-MMLP', MNF-MMLP', EMMLP(PCA), and EMMLP(MNF), respectively.

[0179] Figure 11A comparison chart showing the overall classification accuracy (%) of AGR, MMLP, MNF-MMLP, PCA-MMLP, MNF-MMLP', PCA-MMLP' and EMMLP(PCA) proposed in this invention.

[0180] according to Figure 11 It can be seen that, in the Whu-Hi-Honghu and Whu-Hi-Hanchuan datasets, the classification accuracy of the algorithm of this invention is significantly improved compared to the MMLP algorithm, and both the EMMLP (PCA) and EMMLP (MNF) algorithms show a substantial improvement in classification accuracy across all datasets. This is mainly due to:

[0181] (1) The MMLP algorithm further classifies unclassified samples;

[0182] (2) The method of combining spatial neighborhood information further classifies the data results by combining spatial information, which reduces the interference of noise in spatial features;

[0183] (3) Data dimensionality reduction also improves the distribution of data in the feature space to a certain extent, making the data distribution more consistent with the manifold hypothesis.

[0184] The classification results of the Whu-Hi-Honghu and Whu-Hi-Hanchuan datasets show that MMLP suffers from low classification accuracy due to the presence of a large amount of unclassified data. This invention significantly improves classification accuracy by combining data dimensionality reduction with MMLP for complete data classification, and further enhances the algorithm's classification accuracy through a spatial-spectral joint method. This is evident in the significant improvements achieved in the Whu-Hi-Honghu and Whu-Hi-Hanchuan datasets.

[0185] from Figure 11 It can be seen that EMMLP(PCA) has better classification accuracy than EMMLP(MNF) in classifying hyperspectral remote sensing images with large datasets. This indicates that using PCA to reduce the dimensionality of data in hyperspectral remote sensing images is more beneficial for the distribution of data in the feature space, and helps to further improve the classification accuracy of the algorithm.

[0186] Figure 12 A comparison chart showing the average classification accuracy (%) of AGR, MMLP, MNF-MMLP, PCA-MMLP, MNF-MMLP', PCA-MMLP' and EMMLP (MNF) and PCA proposed in this invention.

[0187] from Figure 12It can be seen that the average accuracy of EMMLP (MNF) and EMMLP (PCA) is higher than that of the MMLP algorithm. The average classification accuracy of the MMLP algorithm is significantly improved after dimensionality reduction by PCA and MNF. For example, after solving the problem of unclassified points in the MMLP algorithm by using the MMLP' algorithm, the average classification accuracy of the algorithm is further improved; after combining spatial neighborhood information, the average classification accuracy is significantly improved. In the Whu-Hi-Longkou dataset, the average classification accuracy of EMMLP (MNF) and EMMLP (PCA) is significantly improved compared to the MMLP algorithm; in the Whu-Hi-Hanchuan dataset, the average classification accuracy of EMMLP (MNF) and EMMLP (PCA) is improved by as much as 2.18 times and 2.02 times respectively compared to the MMLP algorithm.

[0188] Figure 13 A comparison chart of Kappa (%) of AGR, MMLP, MNF-MMLP, PCA-MMLP, MNF-MMLP', PCA-MMLP' and EMMLP (MNF) and EMMLP (PCA) classifications proposed in this invention.

[0189] from Figure 13 As can be seen, the EMMLP (MNF) and EMMLP (PCA) classification methods proposed in this invention significantly improve the Kappa coefficient compared to other classification methods in all selected hyperspectral remote sensing image scenarios. For example, in the Whu-Hi-Hanchuan hyperspectral remote sensing image scenario, the Kappa coefficients of the proposed EMMLP (MNF) and EMMLP (PCA) classification methods are significantly improved compared to the MMLP algorithm, reaching 2.63 times and 2.49 times, respectively. In the Whu-Hi-HongHu hyperspectral remote sensing image scenario, the Kappa coefficients of the EMMLP (MNF) and EMMLP (PCA) classification methods are 2.47 times and 2.16 times higher than the MMLP algorithm, respectively. The above improvements in the Kappa coefficient are mainly due to: (1) the MMLP algorithm further classifying unclassified data; (2) the method of combining spatial neighborhood information to further classify the data results by combining spatial information, reducing the interference of noise in spatial features; (3) data dimensionality reduction also improves the distribution of samples in the new feature space to a certain extent, making the samples more separated in the new hyperplane. The data after dimensionality reduction is more conducive to the classifier to classify, and the distribution of samples after dimensionality reduction is more in line with the manifold hypothesis.

[0190] Figure 14 shows a comparison of classification times between AGR, MMLP, MNF-MMLP, PCA-MMLP, MNF-MMLP', PCA-MMLP' and the EMMLP (MNF) and EMMLP (PCA) proposed in this invention; where Figure 14(a) shows the classification time of Pavia and Salinas, and Figure 14(b) shows the classification time of Whu-Hi-Longkou, Whu-Hi-Honghu, and Whu-Hi-Hanchuan.

[0191] As shown in Figure 14, although the algorithm of this invention uses data dimensionality reduction to process the data, combines the MMLP secondary classification algorithm to further solve the problem of unclassified points after MMLP classification, and uses the joint spatial neighborhood information method to improve the classification accuracy, the algorithm of this invention does not take longer than the MMLP algorithm. This is mainly due to the fact that data dimensionality reduction reduces the computational complexity of calculating the nearest neighbor sparse matrix, and the joint spatial neighborhood information method itself has low time complexity. Figure 14 also shows that the algorithm of this invention has a significant time advantage over the MMLP algorithm in classifying hyperspectral remote sensing images with large datasets, such as Whu-Hi-Honghu and Whu-Hi-Hanchuan. This also proves that the algorithm of this invention is more suitable for classifying hyperspectral remote sensing images with a large number of samples.

[0192] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A semi-supervised classification method based on EMMLP hyperspectral remote sensing image, characterized in that, The method comprises the following steps: S1, dimension reduction is performed on a high-dimensional remote sensing image to obtain a reduced image; S2, the reduced image is classified using an MMLP algorithm; the MMLP algorithm is an extreme minimum maximum label propagation algorithm; S3, the sample x that still is not classified after being classified to the MMLP algorithm ij , x ij , update the sample set U without class label before the MMLP algorithm is classified; update the sample set M with class label before the MMLP algorithm is classified by the sample with class label obtained after the MMLP is classified Define the objective function K(x ij ) as the distance from the unlabeled sample x ij to all labeled samples x j : K(x ij ) = max(c(x ij ,x j ), c(x j )) for j = 1,..., / Wherein, the distance between the classless label sample x ij and the class label sample x j is defined as c(x ij , x j ), x j is originally a classless label sample, the class label sample with a label is obtained through the MMLP algorithm, and c(x j ) is the propagation distance of the classless label sample x j after the MMLP algorithm ends; Then, the max(c(x i ,x j ),c(x j )) of all samples with class labels is sorted, and the smallest K samples are screened according to the sorting result, and the class of the sample with the most class numbers is taken as the class label of the sample without a class label. S4, rejudging the label of all samples in the initial unlabeled sample set: judging the label of sample x j and x ij As the target pixel, the class information of the 8 neighborhoods around the target pixel x j and x ij is extracted, and the class of the target pixel is corrected by calculating the weight sum of each class and combining the voting method. 2.The method of claim 1, wherein, propagation distance c(x j ) is as follows: wherein X ij is a classless sample x j From the class-labeled samples x i , a set of labeled paths, i.e., a set of successfully passed labeled paths, is obtained; a represents a path; the Euclidean distance between two adjacent samples is defined as c(x l , x l+1 ), and max(c(x l , x l+1 )) is the longest distance between any two samples in the propagation path of the classless sample. 3.The method of claim 2, wherein, In S1, PCA or MNF algorithm is used to perform dimension reduction on the high-dimensional remote sensing image. PCA is principal component analysis, and MNF is minimum noise fraction.

4. The method according to claim 3, wherein, In S3, the smallest 15 samples are selected, that is, K=15.

5. The EMMLP-based hyperspectral remote sensing image semi-supervised classification method according to claim 1, 2, 3 or 4, characterized in that, The specific process of S4 comprises the following steps: The class information of 8 neighborhoods around the target pixel is extracted, the classes of the 8 neighborhood pixels around the target pixel are counted, and D k The sum of the class weights of the pixels of the same class is recorded. In the 8 neighborhood pixels of the target pixel, each unlabeled sample which is initially unlabeled but obtains a class label from other samples is set to have a class weight of 1; each initially labeled sample is set to have a class weight of 2, and the class label weight of the target pixel itself is also set to 2. The weights of all categories of the target pixel and its 8 neighborhood samples are counted, and the maximum value of each category weight and D k is determined, and the category label of the target pixel is set as the category label of the category corresponding to the maximum value of the weight and D k . 6.The method of claim 5, wherein, determining a maximum value among the weights and D k of each class, and setting the class label of the target pixel to the class label of the class corresponding to the maximum value k In the process of determining the maximum value among the weights and D k of each class, and setting the class label of the target pixel to the class label of the class corresponding to the maximum value, if the weights and of two classes are the same, the class label of the target pixel is not changed.

7. A computer storage medium, characterized in that The storage medium stores a computer program, and the computer program is loaded and executed by the processor to implement the EMMLP-based hyperspectral remote sensing image semi-supervised classification method in any one of claims 1 to 6.

8. An equipment for semi-supervised classification of EMMLP-based hyperspectral remote sensing images, characterized in that it comprises: The device comprises a processor and a memory, and the memory stores a computer program, and the computer program is loaded and executed by the processor to implement the EMMLP-based hyperspectral remote sensing image semi-supervised classification method in any one of claims 1 to 6.

Images