Multi-target multi-space collaborative hyperspectral unmixing method and system considering spectral variability
By constructing spectral feature subspace and improved multi-objective optimization algorithm, the problem of insufficient end element beam extraction accuracy in hyperspectral images is solved, and the completeness and accuracy of end element beam extraction is achieved, avoiding individual degradation and congestion Distance invalid.
Patent Information
- Application Number
- CN202410912933.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-09
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-07-09
AI Technical Summary
The existing hyperspectral image demixing methods have the problem of insufficient end element beam extraction accuracy, especially because they ignore the potential of spectral subspace and individual degradation and crowded distance failure.
By constructing a spectral feature subspace based on geospectral differentiation features, a multi-objective multi-space collaborative end-member extraction model is adopted, combined with the improved NSGA-II multi-objective optimization algorithm and gene recombination strategy, it avoids individual degeneration and crowded distance failure, and extracts More precise end element bundles.
It improves the completeness and accuracy of end element beam extraction, avoids individual degradation and crowded distance failure problems, and significantly improves the accuracy and richness of end element beam extraction.
Smart Images

Figure CN119027700A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of hyperspectral remote sensing technology, and more specifically, to a multi-target multi-space collaborative hyperspectral unmixing method and system considering spectral variability. Background Art
[0002] Mixed pixel decomposition methods can be roughly divided into two categories according to the process. The first type of method treats endmember extraction and abundance estimation as two independent steps, first performing endmember extraction and then performing abundance estimation. Common endmember extraction methods include methods based on convex geometry theory, projection-based methods, spatial-spectral joint methods, error analysis methods, and optimization-based methods. At the same time, abundance estimation usually uses constrained least squares-based methods and sparse regression methods. The second type of method integrates the two stages of endmember extraction and abundance estimation. The main methods include methods based on non-negative matrix decomposition and methods based on deep learning.
[0003] Mixed pixel decomposition methods can be further classified according to whether or not the intra-class endmember variability is considered. One class of methods considers this variability, while the other does not. Methods that do not consider endmember variability use a single endmember to represent a class of features in the mixed model, which is relatively straightforward to understand and implement. However, in real hyperspectral scenes, there are usually complex endmember variability problems, that is, there are spectral differences within a class of features. This variability is caused by factors such as intra-class material differences, lighting conditions, terrain changes, and phenological periods. Therefore, more and more research is working on dealing with the endmember variability problem.
[0004] Some methods to solve the endmember variability use existing endmember libraries and apply abundance estimation algorithms such as multi-endmember spectral mixing analysis or sparse unmixing to complete the decomposition of mixed pixels. These abundance estimation algorithms will select the most suitable set of endmembers in the endmember library for abundance estimation. However, in practical applications, the adaptability and availability of the endmember library cannot be guaranteed. Therefore, a more feasible method is to automatically extract these endmembers from the image, extract several endmembers for each land class to form an "endmember bundle", and include all intra-class changes as much as possible. This type of method is called an endmember bundle extraction algorithm.
[0005] At present, only a few endmember bundle extraction algorithms have been developed. Spatial local extraction is one of the most popular ideas in endmember bundle extraction methods. For example, the hyperspectral image is divided into several subsets in the spatial dimension, and endmembers are extracted within each subset. In this method, the number of endmembers in each subset is first estimated based on the Gram matrix method, and then the traditional endmember extraction algorithm is used to extract endmembers. Finally, the endmembers obtained in all subsets are integrated and the endmember bundle is output. There are also methods such as randomly selecting pixels in the global image to form subsets, proposing the EBE algorithm, or using multi-scale interval sampling to divide the image into multiple stem subsets, and proposing the MSREBE algorithm. However, this type of endmember bundle extraction method has certain defects, such as the inability to handle incomplete material categories in the subset, only considering spectral information, and the traditional endmember extraction algorithm used is limited by convex geometry theory, and the extracted endmembers are not accurate.
[0006] In addition, there are some endmember bundle extraction methods that use spatial information or spectral curve shape information to extract endmember bundles. For example, an endmember bundle extraction method based on spatial and spectral features (SSEBE) uses a traditional endmember extraction algorithm to extract endmembers from hyperspectral images, and develops a post-processing strategy based on the heterogeneity index to remove candidate endmembers located in heterogeneous areas. This method is also limited by the limited accuracy of the traditional endmember extraction algorithm based on convex geometry theory; there is also an endmember extraction method based on spectral curves (SCEE). This method uses the pixels with the largest or smallest wavelet transform coefficients in a certain band to form a candidate endmember set, and based on the connected component labeling method, removes the pixels in the smaller connected areas in the candidate endmember set to obtain the final endmember set. This algorithm has obvious defects, which is reflected in the fact that when it obtains a complete endmember set, a large number of redundant pixels and noise pixels will be selected, resulting in a serious decrease in the accuracy of endmember bundle extraction.
[0007] In essence, endmember extraction is a combinatorial optimization problem that requires searching for the best endmember combination in the image. In recent years, due to its powerful global search capability, multi-objective optimization algorithms have made significant progress in the field of endmember bundle extraction. Based on the assumption that the relative positions of endmembers in different feature spaces are different, endmember extraction models are constructed in different feature spaces, and endmember bundles are obtained through multi-objective optimization algorithms. The MOEBE algorithm is proposed. This method uses the difference in data distribution in different feature spaces to expose more endmembers on the convex geometry boundary formed by pixels, thereby improving the richness of endmember extraction. However, this type of technology usually roughly divides the original hyperspectral image into three spectral subsets along the spectral dimension, and regards different spectral subsets as different feature spaces. This method ignores the unique spectral characteristics of each type and does not fully utilize the potential of spectral subspaces. In addition, this type of technology directly adopts the general evolutionary algorithm framework in multi-objective optimization. In the combinatorial optimization problem of endmember extraction, the general evolutionary algorithm framework has the problem of individual degradation. At present, there is an urgent need for a more effective endmember bundle extraction framework based on multi-objective optimization, and it is necessary to improve the optimization algorithm according to the characteristics of the endmember bundle extraction task. Achieving comprehensive endmember bundling using multi-objective optimization algorithms remains an unsolved challenge.
[0008] In summary, the existing hyperspectral image unmixing methods have the problem of insufficient accuracy in extracting endmember bundles. Therefore, how to invent a high-precision endmember bundle extraction and unmixing method based on a multi-objective optimization algorithm is a technical problem that needs to be urgently solved in this technical field. Summary of the invention
[0009] In order to solve the problem of insufficient accuracy in extracting endmember bundles in the prior art, the present invention provides a multi-target multi-space collaborative hyperspectral unmixing method considering spectral variability, which has the characteristics of improving the completeness of endmember bundle extraction and avoiding individual degradation and crowding distance failure problems.
[0010] In order to achieve the above-mentioned purpose of the present invention, the technical scheme adopted is as follows:
[0011] A multi-target multi-space collaborative hyperspectral unmixing method considering spectral variability includes the following specific steps:
[0012] Based on the spectral differentiation characteristics of land types in hyperspectral images, a spectral feature subspace set is constructed;
[0013] Based on the spectral feature subspace set, a multi-objective and multi-space collaborative endmember extraction model is constructed;
[0014] Iterative optimization to solve multi-space collaborative models;
[0015] The solved endmember bundle is output as the unmixing result.
[0016] Preferably, based on the differentiated features of land spectrum, a spectral feature subspace is constructed, and the specific steps are as follows:
[0017] Using the k-medios clustering algorithm, the hyperspectral image is clustered into P pixel clusters in the spatial dimension to obtain sub-data Where P is the number of end member categories, L is the number of bands, and N i is the number of pixels contained in the i-th pixel cluster;
[0018] On the sub-data of each cluster, the fast clustering unsupervised band selection algorithm with enhanced density peak is used to obtain the band numbers selected by group P;
[0019] According to the obtained P groups of band number combinations, the sub-data is split into P parts in the spectral dimension to obtain sub-data
[0020] Use the minimum noise separation transform to transform Y j Perform dimensionality reduction to obtain the reduced dimensionality data
[0021] Will Each M sub-data in is divided into a group to obtain the subspace set under all possible combinations
[0022] Furthermore, based on the spectral feature subspace set, a multi-objective multi-space collaborative endmember extraction model is constructed, specifically:
[0023] Construct a multi-objective optimization function based on the maximum volume theory:
[0024]
[0025] in, for The end member matrix in the i-th subspace, vol(A i ) means to find A i The volume of the simplex formed;
[0026] Based on the multi-objective optimization function, a multi-objective optimization algorithm based on the improved NSGA-II is constructed;
[0027] Based on the internal and external loops, a multi-objective and multi-space collaborative endmember extraction model for endmember bundle extraction is constructed:
[0028] In the inner loop, the population is iteratively optimized through a multi-objective optimization algorithm to maximize the multi-objective function and extract the optimal endmember bundle;
[0029] The inner loop termination condition is that the number of optimization algorithm iterations reaches the set maximum number of iterations;
[0030] In the outer loop, each subspace combination is selected in order and input into the multi-objective optimization algorithm, which is used as the feature space of the inner loop end member number to extract and calculate the objective function;
[0031] The optimal endmember bundle under all spectral feature subspace combinations is extracted as the outer loop termination condition. Furthermore, based on the multi-objective optimization function, the multi-objective optimization algorithm has the following specific steps:
[0032] S1: Initialize the input endmember bundle to a population containing NP individuals; randomly initialize the individual numbers and initialize P gene pools;
[0033] S2: Adopting the gene segment allocation strategy, randomly grouping individuals in the population into pairs to form a mating pool; using simulated binary crossover and polynomial mutation in the mating pool to generate the offspring population, and merging the parent population and the offspring population in the mating pool into p', and updating the gene pool through p';
[0034] S3: Calculate the spatial crowding of each individual; select m individuals from p' using the elite-preserving environment selection method based on spatial crowding to update the population, output the updated population as an end-member bundle, and further update the gene pool based on the updated population;
[0035] S4: Return to step S2 and continue iterating the population until the set number of iterations is reached.
[0036] Furthermore, in step S1, the individual numbers are randomly initialized and P gene pools are initialized, specifically:
[0037] Encode all individuals in the input endmember bundle using random pixel numbers:
[0038] X=(g1,g2,...,g P )
[0039] g i (i=1,...,P) represents the serial number of the ith endmember in the ith pixel cluster, where the pixel cluster is the Y obtained by the clustering algorithm when constructing the spectral feature subspace. i , each X records a set of end members, encodes all Xs, and obtains NP serial numbers;
[0040] Encode the endmember combination of each individual as a chromosome, where each gene represents an endmember, and initialize P gene pools:
[0041]
[0042] where pool k (k=1,...,P) represents the kth gene library constructed by the algorithm, Represents the kth gene segment in the nth individual.
[0043] Furthermore, the spatial congestion is calculated. The specific steps are as follows:
[0044] Convert the pixel numbers of all individual records in NP populations into spatial coordinates;
[0045] Based on the spatial coordinates of each individual, its corresponding gene segment is assigned to the corresponding spatial coordinates;
[0046] Each gene library adaptively creates a corresponding adaptive grid according to the range of its spatial coordinates;
[0047] In each gene pool, the number of gene segments contained in each grid cell is taken as the crowding degree of these gene segments;
[0048] The crowding degrees of all gene segments contained in each individual are summed up, and the sum result is used as the spatial crowding degree of the individual.
[0049] Furthermore, an elite-preserving environment selection method based on spatial crowding is used to select m individuals in p'. The specific steps are:
[0050] Perform non-dominated sorting on p', and sort the individuals in p' into several different levels according to the dominance relationship;
[0051] On the basis of non-dominated sorting, if there are individuals at the same level, they are sorted again based on spatial crowding;
[0052] Keep the first m ranked individuals and eliminate the rest.
[0053] Furthermore, a replacement mechanism is provided between the inner loops to prevent duplicate extraction, and the specific steps are as follows:
[0054] In the inner loop, after each iteration of extracting the endmember bundle of the subspace, the spectral angle distance between each extracted endmember and all pixels of the original hyperspectral image is calculated;
[0055] If the spectral angle distance of a pixel is less than the set threshold, the pixel will be replaced by the spectral mean of the endmember bundle to which it belongs.
[0056] Furthermore, the spectral angle distance between each extracted endmember and all pixels of the original hyperspectral image is calculated, specifically:
[0057]
[0058] Where a represents the reference endmember spectrum, Indicates the extracted endmember spectrum.
[0059] A multi-objective multi-space collaborative hyperspectral unmixing system considering spectral variability includes a cascaded spectral feature subspace construction module, a multi-space collaborative model construction module, and an optimization solution module:
[0060] The spectral feature subspace construction module is used to construct a spectral feature subspace set based on the land-type spectral differentiation characteristics of the hyperspectral image;
[0061] The multi-space collaborative model building module is used to build a multi-objective multi-space collaborative endmember extraction model based on a spectral feature subspace set;
[0062] The optimization solution module is used for iterative optimization solution of the multi-space collaborative model, and outputs the solved end member bundle as the unmixing result.
[0063] The beneficial effects of the present invention are as follows:
[0064] The present invention discloses a multi-target multi-space collaborative hyperspectral unmixing method taking spectral variability into consideration. By constructing a spectral feature subspace based on the differentiated characteristics of land class spectra, and based on a set of spectral feature subspaces, a multi-target multi-space collaborative endmember extraction model is constructed. The completeness of endmember bundle extraction is improved, and the problem of insufficient accuracy of extracted endmember bundles is solved. The multi-space collaborative model is solved by iterative optimization, thus avoiding the problems of individual degradation and crowding distance failure. BRIEF DESCRIPTION OF THE DRAWINGS
[0065] Figure 1 It is a flow chart of a multi-target multi-space collaborative hyperspectral unmixing method considering spectral variability of the present invention.
[0066] Figure 2 This is a technical route diagram of the multi-target multi-space collaborative hyperspectral unmixing method considering spectral variability in Example 2 of the present invention.
[0067] Figure 3 It is a computational space crowding distance graph of the multi-target multi-space collaborative hyperspectral unmixing method considering spectral variability in Example 2 of the present invention.
[0068] Figure 4 It is a flow chart of a multi-objective optimization algorithm of a multi-objective multi-space collaborative hyperspectral unmixing method considering spectral variability in Example 2 of the present invention.
[0069] Figure 5 This is a schematic diagram of improving the richness of variable spectrum extraction by using a replacement mechanism of the multi-target multi-space collaborative hyperspectral unmixing method of the present invention taking into account spectral variability in Example 2. DETAILED DESCRIPTION
[0070] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0071] Example 1
[0072] like Figure 1 As shown in FIG. 1 , a multi-target multi-space collaborative hyperspectral unmixing method considering spectral variability includes the following specific steps:
[0073] Based on the spectral differentiation characteristics of land types in hyperspectral images, a spectral feature subspace set is constructed;
[0074] Based on the spectral feature subspace set, a multi-objective and multi-space collaborative endmember extraction model is constructed;
[0075] Iterative optimization to solve multi-space collaborative models;
[0076] The solved endmember bundle is output as the unmixing result.
[0077] Example 2
[0078] More specifically, Figure 2 As shown, in a specific embodiment, considering that the differentiated spectral features between land types in the hyperspectral image are distributed in different spectral subspaces, a spectral feature subspace is constructed based on the differentiated features of the land type spectra. The specific steps are:
[0079] Using the k-medios clustering algorithm, the hyperspectral image is clustered into P pixel clusters in the spatial dimension to obtain sub-data Where P is the number of end member categories, L is the number of bands, and N i is the number of pixels contained in the i-th pixel cluster;
[0080] On the sub-data of each cluster, the Enhanced Fast Density-peak-based Clustering (E-FDPC) algorithm is used to obtain the band numbers selected by group P;
[0081] According to the obtained P groups of band number combinations, the sub-data is split into P parts in the spectral dimension to obtain sub-data
[0082] Use the minimum noise separation transform to transform Y j Perform dimensionality reduction to obtain the reduced dimensionality data
[0083] Will Each M sub-data in is divided into a group to obtain the subspace set under all possible combinations
[0084] In this embodiment, M is 3.
[0085] In a specific embodiment, according to the linear mixture model, the hyperspectral image can be decomposed into the pure spectrum (endmembers) of each type of pixel and the matrix (abundance) recording the corresponding proportion. The maximum volume theory regards the distribution of the hyperspectral image in the high-dimensional spectral space as a convex geometric body, and the endmembers are often located at the vertices of the convex geometric body. In order to accurately extract the endmember set of rich and variable spectra for constructing the spectral variation model, a multi-target multi-space collaborative endmember extraction model is constructed based on the spectral feature subspace set, specifically:
[0086] Construct a multi-objective optimization function based on the maximum volume theory:
[0087]
[0088] in, for The end member matrix in the i-th subspace, vol(A i ) means to find A i The volume of the simplex formed;
[0089] Based on the multi-objective optimization function, a multi-objective optimization algorithm based on the improved NSGA-II is constructed;
[0090] Based on the internal and external loops, a multi-objective and multi-space collaborative endmember extraction model for endmember bundle extraction is constructed:
[0091] In the inner loop, the population is iteratively optimized through a multi-objective optimization algorithm to maximize the multi-objective function and extract the optimal endmember bundle;
[0092] The inner loop termination condition is that the number of optimization algorithm iterations reaches the set maximum number of iterations;
[0093] In the outer loop, each subspace combination is selected in order and input into the multi-objective optimization algorithm, which is used as the feature space of the inner loop end member number to extract and calculate the objective function;
[0094] In this embodiment, three spectral subspace combinations are selected in order and iteratively each time as an endmember bundle, and are sequentially input into the multi-objective optimization algorithm under all possible combinations.
[0095] In this embodiment, the non-dominated sorting genetic algorithm II (NSGA-II) is known for its simplicity and high efficiency, and is one of the most commonly used multi-objective optimization algorithms. The endmember bundle extraction task can be regarded as a combinatorial optimization problem, and NSGA-II is used to obtain a set of Pareto optimal solutions. However, if the existing general multi-objective optimization algorithms such as NSGA-II are directly used for the image endmember space combinatorial optimization problem, individual degradation and crowding distance failure problems will occur, affecting the accuracy of endmember bundle extraction. The present invention introduces the ideas of gene recombination and spatial crowding, and improves the NSGA-II method in a targeted manner to optimize the multi-objective function.
[0096] The optimal endmember bundle under all spectral feature subspace combinations is extracted as the termination condition of the outer loop.
[0097] In a specific embodiment, Figure 4 As shown, based on the multi-objective optimization function, the multi-objective optimization algorithm has the following specific steps:
[0098] S1: Initialize the input endmember bundle to a population containing NP individuals; randomly initialize the individual numbers and initialize P gene pools;
[0099] S2: Adopting the gene segment allocation strategy, randomly grouping individuals in the population into pairs to form a mating pool; using simulated binary crossover and polynomial mutation in the mating pool to generate the offspring population, and merging the parent population and the offspring population in the mating pool into p', and updating the gene pool through p';
[0100] S3: Calculate the spatial crowding of each individual; select m individuals from p' using the elite-preserving environment selection method based on spatial crowding to update the population, output the updated population as an end-member bundle, and further update the gene pool based on the updated population;
[0101] S4: Return to step S2 and continue iterating the population until the set number of iterations is reached.
[0102] In the multi-objective optimization algorithm, in step S1, the individual numbers are randomly initialized and P gene pools are initialized, specifically:
[0103] Encode all individuals in the input endmember bundle using random pixel numbers:
[0104] X=(g1,g2,...,g P )
[0105] g i (i=1,...,P) represents the serial number of the ith endmember in the ith pixel cluster, where the pixel cluster is the Y obtained by the clustering algorithm when constructing the spectral feature subspace.i , each X records a set of end members, encodes all Xs, and obtains NP serial numbers;
[0106] In this embodiment, population optimization is associated with the search for the optimal end member combination. And using this encoding method, each gene locus will search for end members in the corresponding cluster, greatly reducing the search space and increasing the search efficiency (the gene locus is the sequence number of the gene segment in the individual encoding);
[0107] Encode the endmember combination of each individual as a chromosome, where each gene represents an endmember, and initialize P gene pools:
[0108]
[0109] where pool k (k=1,...,P) represents the gene library constructed by the algorithm, Represents the kth gene segment in the nth individual.
[0110] In this embodiment, each gene pool is used to store all gene segments of all individuals in the population at the corresponding gene loci. At the beginning of each iteration, the individuals in the population will be completely reorganized and updated, which means that all gene segments contained in the individuals are randomly selected from the corresponding gene pool, and the gene segments in the gene pool will not be repeatedly allocated. The gene pool will be updated after each population update.
[0111] In a specific embodiment, Figure 3 As shown, the spatial crowding is calculated. The specific steps are:
[0112] Convert the pixel numbers of all individual records in NP populations into spatial coordinates;
[0113] Based on the spatial coordinates of each individual, its corresponding gene segment is assigned to the corresponding spatial coordinates;
[0114] In this embodiment, the spatial coordinates are row and column numbers in the two-dimensional space of the image.
[0115] Each gene library adaptively creates a corresponding adaptive grid according to the range of its spatial coordinates;
[0116] In each gene pool, the number of gene segments contained in each grid cell is taken as the crowding degree of these gene segments;
[0117] The crowding degrees of all gene segments contained in each individual are summed up, and the sum result is used as the spatial crowding degree of the individual.
[0118] In a specific embodiment, an elite-preserving environment selection method based on spatial crowding is used to select m individuals in p', and the specific steps are:
[0119] Perform non-dominated sorting on p', and sort the individuals in p' into several different levels according to the dominance relationship;
[0120] On the basis of non-dominated sorting, if there are individuals at the same level, they are sorted again based on spatial crowding;
[0121] Keep the first m ranked individuals and eliminate the rest.
[0122] In a specific embodiment, a replacement mechanism for preventing repeated extraction is also provided between inner loops, and the specific steps are as follows:
[0123] In the inner loop, after each iteration of extracting the endmember bundle of the subspace, the spectral angle distance between each extracted endmember and all pixels of the original hyperspectral image is calculated;
[0124] If the spectral angle distance of a pixel is less than the set threshold, the pixel will be replaced by the spectral mean of the endmember bundle to which it belongs.
[0125] In this embodiment, the proposed anti-duplicate extraction strategy is executed between each inner loop. Specifically, the spectral angular distance (SAD) between the endmember extracted in this inner loop and all pixels of the original hyperspectral image is calculated. Pixels with SAD less than a set threshold are replaced with the spectral mean of the cluster to which the pixel belongs. Figure 5 As shown in the figure, mean replacement makes the pixel necessarily updated to a mixed pixel, which can avoid the pixel from being extracted repeatedly and expose other end members that constitute the simplex. And the mean replacement pixel will contain the characteristic information of a certain type of ground feature. When all pure pixels are replaced, it can avoid the extraction of multiple types of mixed pixels. At the algorithm optimization level, the extracted end members will not interfere with the optimization process as a local optimal solution.
[0126] In summary, the beneficial effects of the present invention include:
[0127] A multi-objective multi-space collaborative endmember extraction model was constructed: the existing endmember bundle extraction algorithm based on multi-objective optimization only roughly divides the original image into three parts along the spectral dimension, does not consider the spectral differences between land types, and cannot make full use of the spectral information. In order to make full use of the image spectral information, the present invention carefully designs the feature space construction method, aiming to show the spectral variability of the endmembers of each land type as much as possible through the multi-feature space. Considering that the distribution of endmembers in different spectral subspaces has obvious differences, the spectral subspace constructed according to the characteristics of each land type can better expose the endmembers of different land types outside the data boundary, which is convenient for extraction. Specifically, the original image is first clustered into P clusters (P is the number of land types) in the spatial dimension, and then each pixel cluster is respectively band selected to obtain a band set containing the band characteristics of each land type, and then the image is divided into P subsets in the spectral dimension. In this embodiment, every three spectral subsets are taken as a combination, and endmember bundle extraction is performed sequentially under all possible combinations.
[0128] The present invention adopts a multi-strategy integrated multi-objective optimization algorithm: the existing endmember bundle extraction algorithms based on multi-objective optimization all directly use general optimization algorithms to optimize the model. The general optimization algorithm framework is designed to solve mathematical optimization problems. If it is directly used for spatial image combination optimization problems such as endmember bundle extraction, individual degradation and crowding distance failure problems will occur. Taking genetic algorithms as an example, based on genetic algorithms, endmember combinations are encoded as chromosomes, where each gene segment represents an endmember. If the gene fragments of the two parent chromosomes at the same gene locus represent endmembers of different categories, the crossover operation between the two will cause two endmembers of the same category to exist on the generated offspring chromosomes. This means that the endmember categories recorded in the offspring chromosomes are repeated, the recorded endmember categories are incomplete, and individual degradation occurs. To solve this problem, the present invention introduces a gene recombination strategy to construct a gene library for each land type, and the chromosomes will randomly select gene segments in these gene libraries to ensure that the endmember types recorded on all chromosomes at the same gene locus are consistent to avoid individual degradation problems. Moreover, since the slight differences in gene segments will seriously affect the objective function value in multi-objective optimization, the traditional crowding distance based on the objective function is invalid in the endmember bundle extraction task. Therefore, based on the two-dimensional spatial distance of the candidate endmembers on the image, the present invention proposes a crowding degree calculation strategy based on spatial distance to avoid the problem of crowding distance failure, ensure population diversity, and enhance the global search capability of the algorithm.
[0129] The present invention also adopts an anti-duplicate extraction strategy: the existing methods all extract a large number of end members at one time, and duplicate end members are inevitably present in the results. In the present invention, each outer loop extracts a small amount of accurate end members and adds them to the result set, reducing the risk of redundant extraction. In addition, in order to prevent the problem of unintentional duplicate extraction that may occur between multiple outer loops, the present invention proposes an anti-duplicate extraction strategy based on replacement. In the process of sequential extraction of different spectral subspace sets, any pixel in the original image that is similar to the subsequent end member that has been extracted is replaced with the mean of its corresponding pixel clustering cluster. This operation has two major advantages for improving the end member bundle extraction algorithm. One is that it can avoid the subsequent extraction of duplicate end members. The second is to move the extracted end members from the convex boundary of the data to the inside, so that other end members are exposed outside the boundary. This change can avoid the interference caused by the local optimal solution during the optimization process and improve the richness of the end member bundle extraction.
[0130] Example 3
[0131] A multi-objective multi-space collaborative hyperspectral unmixing system considering spectral variability includes a cascaded spectral feature subspace construction module, a multi-space collaborative model construction module, and an optimization solution module:
[0132] The spectral feature subspace construction module is used to construct a spectral feature subspace set based on the land-type spectral differentiation characteristics of the hyperspectral image;
[0133] The multi-space collaborative model building module is used to build a multi-objective multi-space collaborative endmember extraction model based on a spectral feature subspace set;
[0134] The optimization solution module is used for iterative optimization solution of the multi-space collaborative model, and outputs the solved end member bundle as the unmixing result.
[0135] Obviously, the above embodiments of the present invention are only examples for clearly explaining the present invention, and are not intended to limit the implementation methods of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention shall be included in the protection scope of the claims of the present invention.
Claims
1. A multi-target multi-space collaborative hyperspectral unmixing method considering spectral variability, characterized by: The specific steps include: Based on the spectral differentiation characteristics of land types in hyperspectral images, a spectral feature subspace set is constructed; Based on the spectral feature subspace set, a multi-objective and multi-space collaborative endmember extraction model is constructed; Iterative optimization to solve multi-space collaborative models; The solved endmember bundle is output as the unmixing result.
2. The multi-target multi-space collaborative hyperspectral unmixing method considering spectral variability according to claim 1 is characterized in that: Based on the spectral differentiation characteristics of land types, the spectral feature subspace is constructed. The specific steps are as follows: Using the k-medios clustering algorithm, the hyperspectral image is clustered into P pixel clusters in the spatial dimension to obtain sub-data Where P is the number of end member categories, L is the number of bands, and N i is the number of pixels contained in the i-th pixel cluster; On the sub-data of each cluster, the fast clustering unsupervised band selection algorithm with enhanced density peak is used to obtain the band numbers selected by group P; According to the obtained P groups of band sequence number combinations, the sub-data is split into P parts in the spectral dimension to obtain sub-data Use the minimum noise separation transform to transform Y j Perform dimensionality reduction to obtain the reduced dimensionality data Will Each M sub-data in is divided into a group to obtain the subspace set under all possible combinations 3. The multi-target multi-space collaborative hyperspectral unmixing method considering spectral variability according to claim 2 is characterized in that: Based on the spectral feature subspace set, a multi-objective multi-space collaborative endmember extraction model is constructed, specifically: Construct a multi-objective optimization function based on the maximum volume theory: in, for The end member matrix in the i-th subspace, vol(A i ) means to find A i The volume of the simplex formed; Based on the multi-objective optimization function, a multi-objective optimization algorithm based on the improved NSGA-II is constructed; Based on the internal and external loops, a multi-objective and multi-space collaborative endmember extraction model for endmember bundle extraction is constructed: In the inner loop, the population is iteratively optimized through a multi-objective optimization algorithm to maximize the multi-objective function and extract the optimal endmember bundle; The inner loop termination condition is that the number of optimization algorithm iterations reaches the set maximum number of iterations; In the outer loop, each subspace combination is selected in order and input into the multi-objective optimization algorithm, which is used as the feature space of the inner loop end member number to extract and calculate the objective function; The optimal endmember bundle under all spectral feature subspace combinations is extracted as the termination condition of the outer loop.
4. The multi-target multi-space collaborative hyperspectral unmixing method considering spectral variability according to claim 3 is characterized by: Based on the multi-objective optimization function, the multi-objective optimization algorithm has the following specific steps: S1: Initialize the input endmember bundle to a population containing NP individuals; randomly initialize the individual numbers and initialize P gene pools; S2: Adopting the gene segment allocation strategy, randomly grouping individuals in the population into pairs to form a mating pool; using simulated binary crossover and polynomial mutation in the mating pool to generate the offspring population, and merging the parent population and the offspring population in the mating pool into p', and updating the gene pool through p'; S3: Calculate the spatial crowding of each individual; select m individuals from p' using the elite-preserving environment selection method based on spatial crowding to update the population, output the updated population as an end-member bundle, and further update the gene pool based on the updated population; S4: Return to step S2 and continue iterating the population until the set number of iterations is reached.
5. The multi-target multi-space collaborative hyperspectral unmixing method considering spectral variability according to claim 4 is characterized in that: In step S1, the individual numbers are randomly initialized and P gene pools are initialized, specifically: Encode all individuals in the input endmember bundle using random pixel numbers: X=(g1,g2,...,g P ) g i (i=1,...,P) represents the serial number of the ith end member in the ith pixel cluster, where the pixel cluster is the Y obtained by the clustering algorithm when constructing the spectral feature subspace. i , each X records a set of end members, encodes all Xs, and obtains NP serial numbers; Encode the endmember combination of each individual as a chromosome, where each gene represents an endmember, and initialize P gene pools: where pool k (k=1,...,P) represents the kth gene library constructed by the algorithm, Represents the kth gene segment in the nth individual.
6. The multi-target multi-space collaborative hyperspectral unmixing method considering spectral variability according to claim 5 is characterized in that: Calculate the spatial crowding. The specific steps are: Convert the pixel numbers of all individual records in NP populations into spatial coordinates; Based on the spatial coordinates of each individual, its corresponding gene segment is assigned to the corresponding spatial coordinates; Each gene library adaptively creates a corresponding adaptive grid according to the range of its spatial coordinates; In each gene pool, the number of gene segments contained in each grid cell is taken as the crowding degree of these gene segments; The crowding degrees of all gene segments contained in each individual are summed up, and the sum result is used as the spatial crowding degree of the individual.
7. The multi-target multi-space collaborative hyperspectral unmixing method considering spectral variability according to claim 6, characterized in that: The elite-preserving environment selection method based on spatial crowding is used to select m individuals in p'. The specific steps are: Perform non-dominated sorting on p', and sort the individuals in p' into several different levels according to the dominance relationship; On the basis of non-dominated sorting, if there are individuals at the same level, they are sorted again based on spatial crowding; Keep the first m ranked individuals and eliminate the rest.
8. The multi-target multi-space collaborative hyperspectral unmixing method considering spectral variability according to claim 7 is characterized by: There is also a replacement mechanism between the inner loops to prevent duplicate extraction, and the specific steps are as follows: In the inner loop, after each iteration of extracting the endmember bundle of the subspace, the spectral angle distance between each extracted endmember and all pixels of the original hyperspectral image is calculated; If the spectral angle distance of a pixel is less than the set threshold, the pixel will be replaced by the spectral mean of the endmember bundle to which it belongs.
9. The multi-target multi-space collaborative hyperspectral unmixing method considering spectral variability according to claim 8, characterized in that: The spectral angle distance between each extracted endmember and all pixels of the original hyperspectral image is calculated as follows: Where a represents the reference endmember spectrum, Indicates the extracted endmember spectrum.
10. A multi-target multi-space collaborative hyperspectral unmixing system considering spectral variability, characterized by: Including cascaded spectral feature subspace construction module, multi-space collaborative model construction module, and optimization solution module: The spectral feature subspace construction module is used to construct a spectral feature subspace set based on the land-type spectral differentiation characteristics of the hyperspectral image; The multi-space collaborative model building module is used to build a multi-objective multi-space collaborative endmember extraction model based on a spectral feature subspace set; The optimization solution module is used for iterative optimization solution of the multi-space collaborative model, and outputs the solved end member bundle as the unmixing result.
Citation Information
Patent Citations
Non-negative matrix unmixing method based on space-spectrum combined multi-constraint optimization
CN109241843A
Hyperspectral endmember extraction method and device based on multi-objective differential evolution of sorting multiple variations
CN111797988A
End-member extraction method based on segmented vertex component analysis (VCA)
US20190392261A1
Cited By
Robust depth non-negative matrix factorization method for hyperspectral image unmixing
CN120876879A
End member beam extraction method and device, electronic equipment and storage medium
CN122200377A
