A hyperspectral-based mineral alteration feature extraction method

Abstract

The application relates to the technical field of ground object component analysis, and discloses a mineral alteration feature extraction method based on hyperspectrum, which comprises the following steps: obtaining hyperspectrum remote sensing image data of a region to be processed and performing atmospheric correction and noise reduction preprocessing; extracting initial end-member spectrum after the preprocessing; constructing a target function of nonlinear spectrum unmixing, wherein the target function comprises the following items: a data fidelity item, an abundance constraint item, a physical consistency regularization item for penalizing the deviation between the spectrum generated by a physical model and the current end-member spectrum based on a differentiable radiation transfer model; solving the target function to obtain the final end-member spectrum and an abundance matrix corresponding to the end-member spectrum; identifying alteration minerals by using the final end-member spectrum, and extracting the mineral alteration features of the region to be processed in combination with the abundance matrix corresponding to the end-member spectrum. The application can understand the complex nonlinear mixing effect existing in the surface environment by introducing a space-spectrum composite kernel function, and solving the nonlinear unmixing problem in a high-dimensional feature space.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ground object component analysis, and in particular to a mineral alteration feature extraction method based on hyperspectral. BACKGROUND

[0002] Hyperspectral remote sensing technology has the characteristics of integration of map and atlas in the field of mineral exploration and geological mapping. By analyzing the spectral curve reflected by the ground object, the surface mineral component can be identified, and the favorable area for prospecting can be delineated. However, due to the limited spatial resolution of the remote sensing sensor, a single pixel usually contains multiple ground object components, forming a mixed pixel, which seriously affects the accuracy of mineral identification.

[0003] The spectral unmixing technology is the key to solving the problem of mixed pixels, and the goal is to decompose the mixed pixel into a series of pure components (endmembers) and the proportion (abundance) of the spectrum in the pixel. The unmixing method is mostly based on the linear spectral mixing model (LSMM), which assumes that the pixel spectrum is a linear weighted sum of each endmember spectrum. However, in complex geological environments, due to factors such as terrain undulation, multiple scattering between mineral particles, and microscopic scale close mixing of minerals, the relationship between pixels and endmembers often presents a nonlinear relationship, resulting in large errors in the unmixing results of the linear model, which is difficult to meet the needs of high-precision alteration information extraction.

[0004] The methods based on kernel function and the methods based on bilinear model improve the unmixing performance to some extent, but most of the methods regard unmixing as a purely mathematical inversion problem, ignoring the specific combination rules of minerals in nature, such as the fact that some minerals often occur together, while others are mutually exclusive, and failing to effectively integrate this geological prior knowledge into the model. Moreover, the existing methods rarely consider the physical transmission process of spectral signals in the imaging chain, resulting in endmember spectra that may not have a clear physical meaning. The spectra of the same mineral will vary due to changes in lighting conditions, atmospheric conditions, and their own physical and chemical properties, i.e. endmember variability. Current models lack the ability to model this variability, often simplifying the spectrum of a mineral as a fixed and unchanging spectrum, which does not conform to the actual situation. Therefore, how to construct an unmixing framework that can simultaneously consider nonlinear effects and integrate geology, and effectively cope with endmember spectrum variability, is a technical problem that needs to be solved in the field of high-precision mineral mapping. SUMMARY

[0005] The present application provides a mineral alteration feature extraction method based on hyperspectral to solve the problem of being difficult to construct an unmixing framework that can simultaneously consider nonlinear effects and integrate geology, and effectively cope with endmember spectrum variability in the prior art.

[0006] The mineral alteration feature extraction method based on hyperspectral of the present application comprises the following steps:

[0007] Hyperspectral remote sensing image data of a region to be processed is acquired, and the hyperspectral remote sensing image data is subjected to atmospheric correction and noise reduction preprocessing; an end member extraction algorithm is used to extract initial end member spectra from the preprocessed hyperspectral remote sensing image data;

[0008] A target function of nonlinear spectral unmixing is constructed, and the target function includes: a data fidelity term in which a pixel is mapped to a high-dimensional feature space by using a space-spectrum composite kernel function; an abundance constraint term including group sparse regularization based on mineral paragenetic priori; and a physical consistency regularization term in which a deviation between a spectrum generated by a physical model and a current end member spectrum is penalized based on a differentiable radiative transfer model;

[0009] An alternating optimization algorithm is used to iteratively solve the target function, and an end member abundance matrix and an end member spectrum are alternately updated in the iteration, wherein the end member spectrum is optimized by being constrained on a preset low-dimensional manifold according to a manifold learning strategy;

[0010] When the iteration converges or a preset termination condition is reached, a final end member spectrum and an abundance matrix corresponding to the end member spectrum are obtained; the final end member spectrum is used to identify an altered mineral, and a mineral alteration feature of the region to be processed is extracted in combination with the abundance matrix corresponding to the end member spectrum.

[0011] Preferably, the space-spectrum composite kernel function is composed of a Gaussian radial basis space kernel function and a polynomial spectrum kernel function which are combined by weighting.

[0012] Preferably, the abundance constraint term of the group sparse regularization based on mineral paragenetic priori includes:

[0013] According to prior knowledge of mineral paragenetic combination, end members to be extracted having similar geological genesis or spatial paragenetic relationship are divided into different groups; an L2 norm is applied to an end member abundance vector in each group, and L2,1 mixed norms are obtained by summing L2 norms of all groups, and the L2,1 mixed norms are used as the group sparse regularization term.

[0014] Preferably, the physical consistency regularization term in which a deviation between a spectrum generated by a physical model and a current end member spectrum is penalized based on a differentiable radiative transfer model includes: a differentiable radiative transfer model is used, mineral physical parameters corresponding to the end member spectrum of the current iteration are input into the differentiable radiative transfer model, a physical model spectrum is generated, and a difference between the physical model spectrum and the current end member spectrum is calculated as the physical consistency regularization term.

[0015] Preferably, the target function is solved iteratively by using the alternating optimization algorithm, including: in each iteration, fixing the endmember spectra, solving the abundance matrix by using a multiplicative update rule to ensure the non-negativity of the abundance update, and performing a projection operation on the abundance vector of each pixel after each update to make the abundance vector satisfy the constraint that the sum is one.

[0016] Preferably, the update of the endmember spectra is based on a manifold learning strategy, and the optimization is performed by constraining the spectra on a preset low-dimensional manifold, including: using reference spectra selected from a standard mineral spectral library and the initial endmember spectra to jointly construct a training set, and learning the low-dimensional manifold structure of the training set data by using a manifold learning algorithm; and when iteratively updating the endmember spectra, projecting the gradient of the target function with respect to the endmember spectra onto the tangent space of the low-dimensional manifold, and updating along the projected gradient direction.

[0017] Preferably, the altered minerals are identified by using the last endmember spectra, including:

[0018] The similarity between the last endmember spectra and reference spectra in a standard mineral spectral library is calculated by using a spectral similarity measurement method; for each endmember spectrum, the spectrum is identified as the mineral with the highest similarity and higher than a preset similarity threshold in the reference spectra.

[0019] Preferably, the hyperspectral remote sensing image data of the region to be processed is acquired, and the hyperspectral remote sensing image data is preprocessed by atmospheric correction and noise reduction, including: acquiring the hyperspectral image data of the mining area by using an infrared imaging spectrometer AVIRIS, performing atmospheric correction by using a FLAASH model to obtain surface reflectance data, and performing noise reduction processing on the reflectance data by using a low-rank tensor approximation LRTA algorithm.

[0020] Preferably, the endmember extraction algorithm is a vertex component analysis VCA algorithm.

[0021] Preferably, the alternating optimization algorithm is an alternating direction multiplier method ADMM.

[0022] The beneficial effects of the present application are: the present application can understand the complex nonlinear mixing effect commonly existing in the actual surface environment by introducing a space-spectrum composite kernel function to solve the nonlinear unmixing problem in a high-dimensional feature space. The group sparse regularization based on the mineral paragenetic rule is used as the abundance constraint, and the geological prior knowledge is integrated into the unmixing model, so that the abundance inversion result is more in line with the geological reality, the physically unreasonable mineral combination is reduced, and the reliability of the result is improved. In addition, by constructing a physically consistent regularization term based on a differentiable radiative transfer model and updating the endmember spectrum by using a manifold learning strategy, on the one hand, the extracted endmember spectrum has a clear physical meaning, and on the other hand, the endmember variability problem caused by illumination, atmosphere and mineral properties is solved, so that the endmember extraction is closer to the real situation. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A flowchart of a mineral alteration feature extraction method based on hyperspectral provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0024] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below by referring to the drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0025] As shown in Figure 1 The mineral alteration feature extraction method based on hyperspectral provided by the embodiment of the present application specifically includes the following steps:

[0026] S1, acquiring hyperspectral remote sensing image data of a region to be processed, performing atmospheric correction and noise reduction preprocessing on the hyperspectral remote sensing image data, and extracting initial endmember spectra from the preprocessed hyperspectral remote sensing image data by using an endmember extraction algorithm.

[0027] The hyperspectral image data of a mining area is collected by an airborne visible / infrared imaging spectrometer AVIRIS. The FLAASH model is used for atmospheric correction to eliminate the influence of atmospheric scattering and absorption on the ground object spectrum, and the surface reflectance data is obtained. The low-rank tensor approximation LRTA algorithm is used for noise reduction processing on the reflectance data to remove sensor noise and bad lines, while maintaining the spatial and spectral structure of the hyperspectral image data.

[0028] The number P of endmembers existing in the hyperspectral image data is estimated by using the HFC virtual dimension method. The VCA (Vertex Component Analysis) algorithm is used to extract P initial endmember spectra from the preprocessed hyperspectral image data, and these initial endmember spectra constitute an initial endmember spectrum matrix.

[0029] S2, constructing a target function of the nonlinear spectral unmixing, the target function comprising: a data fidelity term mapping the pixels to a high-dimensional feature space using a spatial-spectral compound kernel function; an abundance constraint term including abundance non-negativity, sum-to-one, and group sparsity regularization based on mineral paragenesis prior; and a physical consistency regularization term based on a differentiable radiative transfer model to penalize the deviation between the spectrum generated by the physical model and the current endmember spectrum.

[0030] The compound kernel function adopted by the data fidelity term is the product of a spatial Gaussian kernel and a spectral radial basis function kernel, which is used to measure the error between the reconstructed pixels and the original pixels in the high-dimensional feature space. In the abundance constraint term, the group sparsity regularization term divides the geologically paragenetic minerals into groups. For example, kaolinite, alunite and montmorillonite are divided into a clay mineral group, and hematite and goethite are divided into an iron oxide group. By imposing an L2,1 norm penalty on the abundance vector within each group, it is prompted that the abundance of a group of minerals is simultaneously selected or simultaneously suppressed. The physical consistency regularization term is based on the differentiable PROSAIL model. The regularization term calculates the Euclidean distance between the spectrum generated by the PROSAIL model according to the current endmember physical parameters and the endmember spectrum estimated by the algorithm as a penalty term. The constraints on the abundance include abundance non-negativity, sum-to-one, and group sparsity regularization based on mineral paragenesis prior. Among them, the abundance non-negativity requires that the proportion of any kind of mineral in a certain pixel must be greater than or equal to zero; the sum-to-one requires that the abundance proportions of all considered minerals and other background substances that may exist in the same pixel must add up to 1; and the group sparsity regularization based on mineral paragenesis prior is a constraint term determined according to the mineral paragenesis, for example, in the alteration of propylitization, chlorite and epidote often appear at the same time. In the alteration of sericitization, quartz, sericite and pyrite are a common combination, and it is rare to find that epidote and sericite appear in large quantities in the same small rock sample.

[0031] In an optional embodiment, the spatial-spectral compound kernel function is composed of a weighted combination of a Gaussian radial basis spatial kernel function and a polynomial spectral kernel function.

[0032] The design of the spatial-spectral compound kernel function aims to simultaneously utilize the spatial distribution information and spectral characteristics of the pixels in the hyperspectral image. The Gaussian radial basis spatial kernel function measures the spatial correlation of the pixels by calculating the exponential decay of the spatial distance between the pixels, considering that the pixels geographically adjacent to each other have a higher probability of similarity. For example, in a 500 by 500 pixel image area, the spatial kernel function value of the central pixel and the pixels in the 3 by 3 neighborhood of the central pixel will be significantly higher than the spatial kernel function value of the central pixel and the pixels 100 meters away. The bandwidth parameter such as sigma of the kernel function can be set to 3.0 to control the range of spatial influence.

[0033] The polynomial spectral kernel is used to capture the complex nonlinear relationships between different spectral bands, which is crucial for distinguishing minerals with subtle but different spectral features. For example, by setting the polynomial degree to 3, the model can effectively amplify the tiny differences in the depth of specific absorption valleys between two pixels by modeling their high-order interactions on the 224-band spectral vectors. By linearly combining the results of the Gaussian spatial kernel and the polynomial spectral kernel with a weighting coefficient such as 0.4, a comprehensive similarity measure is formed, enabling the unmixing model to not only identify spectrally similar materials but also consider their spatial clustering patterns, thereby improving the accuracy and continuity of mineral mapping.

[0034] In an optional embodiment, the abundance constraint term of the group sparse regularization based on the mineral paragenesis prior includes: according to the prior knowledge of the mineral paragenesis combination, the to-be-extracted endmembers with similar geological genesis or spatial paragenetic relationship are divided into different groups; the L2 norm is applied to the endmember abundance vector in each group, and the L2 norm results of all groups are summed to obtain the L2,1 mixed norm, which is used as the group sparse regularization term.

[0035] Geological knowledge is integrated into the mathematical model to guide the solution of mineral abundance. In the exploration of a typical porphyry copper mine, according to the metallogenic theory, sericite, pyrite and kaolinite often appear in the core of the alteration zone, forming an alteration assemblage, while chlorite and epidote often appear in the periphery. According to this, among the 12 to-be-extracted endmembers, the three endmembers representing sericite, pyrite and kaolinite are divided into group 1, and the two endmembers representing chlorite and epidote are divided into group 2, and the remaining endmembers are grouped separately.

[0036] The application of the L2,1 mixed norm realizes the sparsity at the group level. For any pixel in the image, the regularization term encourages the abundance coefficients of the entire group to be zero or non-zero at the same time. For example, for a pixel located in the core of the alteration zone, the model tends to assign a non-zero abundance value to each of the three endmembers in group 1, such as 0.3, 0.2 and 0.15, while the abundance values of the two endmembers in group 2 tend to zero. This mechanism avoids the unreasonable splitting of mineral combinations that should be paragenetic in logic, making the unmixing result more consistent with geological rules and effectively suppressing the appearance of isolated abnormal points caused by noise.

[0037] In an optional embodiment, the physical consistency regularization term based on the differentiable radiative transfer model penalizes the deviation between the spectrum generated by the physical model and the current endmember spectrum, which includes: inputting the mineral physical parameters corresponding to the endmember spectrum of the current iteration into the differentiable radiative transfer model to generate a physical model spectrum; calculating the difference between the physical model spectrum and the current endmember spectrum as the physical consistency regularization term.

[0038] This regularization term ensures that the endmember spectra extracted by the algorithm are not only mathematically optimal, but also physically reasonable. At the 50th iteration of the optimization, suppose the algorithm generates an endmember spectrum of a suspected goethite. To verify the physical plausibility, a differentiable Hapke radiative transfer model is called. Typical physical parameters of goethite, such as an average grain size of 80 microns, and single-scattering albedo data of goethite measured in a laboratory setting, are input into the Hapke radiative transfer model.

[0039] The Hapke radiative transfer model simulates the spectral curve that a goethite with these physical properties should exhibit under the current imaging conditions, i.e., the physical model spectrum. The root mean square error between the physical model spectrum and the endmember spectrum generated by the algorithm is calculated. Suppose the calculated error is 0.08, which is added as a penalty term to the total objective function. This penalty term drives the next iteration to adjust the endmember spectrum so that it is closer to the physical model output, correcting spectral shapes that do not conform to the physical laws of light-matter interaction due to data noise or model limitations, such as unrealistic absorption depths or peak positions.

[0040] S3, iteratively solving the objective function using an alternating optimization algorithm, alternately updating the endmember abundance matrix and the endmember spectrum in each iteration, wherein the updating of the endmember spectrum is based on a manifold learning strategy to constrain the spectrum on a pre-set low-dimensional manifold for optimization.

[0041] The alternating direction multiplier method (ADMM) is used for iterative solution. When the endmember spectrum is fixed, the abundance matrix is updated by solving a quadratic programming problem with group sparse regularization. When the abundance matrix is fixed, the same-named mineral spectrum in the standard mineral spectrum library is learned using the Laplacian Eigenmap algorithm to construct a low-dimensional manifold embedding space. The endmember spectrum updating problem is projected onto the low-dimensional manifold, and the optimal solution is found by gradient descent method to map the endmember spectrum back to the original spectral space.

[0042] In an optional embodiment, the objective function is iteratively solved using an alternating optimization algorithm, including: in each iteration, fixing the endmember spectrum, using a multiplicative update rule to solve the abundance matrix to ensure the non-negativity of the abundance update, and after each update, projecting the abundance vector of each pixel to make the abundance vector satisfy the constraint that the sum is one.

[0043] Since the objective function is non-convex with respect to both endmember and abundance, an alternating optimization strategy is adopted to simplify the solution process. In one iteration cycle, the endmember spectra are fixed at the current iteration result, and only the abundance matrix is optimized. For example, in the kth iteration, given the 5 endmember spectra, the abundance of each mineral in each pixel is calculated by a multiplicative update formula. The design of the multiplicative update formula guarantees that the calculated abundance values, for example, 0.2, 0.5, 0.1, 0.3, 0.05, are always greater than or equal to zero, which is consistent with the physical meaning that abundance cannot be negative.

[0044] After the multiplicative update, the abundance vector of a pixel is guaranteed to be non-negative, but the sum of the vector may not be one, such as the sum of 1.15 in the above example. This does not satisfy the constraint that the sum of all component abundances must be equal to one. Therefore, a projection operation is performed on the abundance vector. The projection operation will scale the abundance vector [0.2, 0.5, 0.1, 0.3, 0.05] to the nearest point on the plane whose sum is one, to obtain a new abundance vector, such as [0.17, 0.43, 0.09, 0.26, 0.04], the sum of the components is equal to 1. After completing the process of fixing endmembers and updating abundances for all pixels in the image, the next step is to fix the abundances and update the endmembers.

[0045] In an optional embodiment, the update of the endmember spectra is based on a manifold learning strategy, which constrains the spectra to be optimized on a pre-set low-dimensional manifold, including: using reference spectra selected from a standard mineral spectrum library together with the initial endmember spectra to construct a training set, and using a manifold learning algorithm to learn the low-dimensional manifold structure of the training set data; when iteratively updating the endmember spectra, projecting the gradient of the objective function with respect to the endmember spectra onto the tangent space of the low-dimensional manifold, and updating along the projected gradient direction.

[0046] The hyperspectral data has a high dimension, such as 224 bands, but all possible spectral variants of a specific class of minerals are actually distributed on a nonlinear low-dimensional manifold much lower than 224 dimensions. Taking advantage of this characteristic, a training set is constructed including, for example, 80 spectra of epidote group minerals from the USGS spectral library and 20 initial endmember spectra. The local linear embedding algorithm is applied to analyze the training set containing 100 samples, and it is found that the intrinsic structure of the training set can be represented by a 4-dimensional manifold.

[0047] In the iterative update of a certain chlorite related endmember spectrum, the gradient of the objective function with respect to the endmember spectrum is calculated, which is a 224-dimensional vector pointing to the direction in which the function value decreases fastest. However, this direction may deviate from the 4-dimensional chlorite manifold that has been learned, resulting in an updated spectrum that loses physical meaning. Therefore, the 224-dimensional gradient vector is projected onto the tangent space of the 4-dimensional chlorite manifold at the current endmember position. This projected new gradient direction ensures that the update step is along the manifold surface, ensuring that the updated endmember spectrum is still a reasonably changed, effective chlorite family spectrum.

[0048] S4, when the iteration converges or reaches the preset termination condition, obtaining the final endmember spectrum and the abundance matrix corresponding to the endmember spectrum; identifying the altered minerals using the final endmember spectrum, and extracting the mineral alteration characteristics of the region to be processed in combination with the abundance matrix corresponding to the endmember spectrum.

[0049] The iteration termination condition is set to one of the following three: the number of iterations reaches a preset maximum value, for example, 500 times; the relative change rate of the objective function value in two consecutive iterations is less than a minimum threshold, for example, 0.00001; or the Frobenius norm change amount of the endmember spectrum matrix and the abundance matrix in two consecutive iterations is less than a preset threshold.

[0050] The P endmember spectra finally unmixed are matched with the USGS standard mineral spectrum library respectively. The spectral angle matching algorithm SAM is used to calculate the similarity of each extracted endmember with all standard mineral spectra in the library, and the standard mineral name with the highest similarity and an angle value less than a preset threshold is assigned to the extracted endmember. According to the identified kaolinite, alunite and other altered minerals, the abundance map corresponding to the minerals is taken as the mineral distribution map. By threshold segmentation and logical superposition on the abundance map of specific mineral combinations such as kaolinite and alunite, the mineral alteration characteristic region such as the advanced argillization alteration zone is circled.

[0051] In an optional embodiment, identifying the altered minerals using the final endmember spectrum comprises: using a spectral similarity measurement method to calculate the similarity of the final endmember spectrum with reference spectra in a standard mineral spectrum library; for each endmember spectrum, identifying the spectrum as the mineral with the highest similarity and higher than a preset similarity threshold in the reference spectra.

[0052] After 100 iterations of algorithm running, a set of, for example, 8 final endmember spectra are obtained. These endmember spectra are pure mathematical vectors that need to be given a geological name. For this purpose, the spectral angle matching algorithm is used to compare the 8 extracted endmember spectra one by one with a reference library containing 480 standard mineral spectra. The spectral angle matching algorithm calculates the angle between two spectral vectors, and the smaller the angle, the higher the similarity.

[0053] The implementation principle of the mineral alteration feature extraction method based on hyperspectrum in the embodiment of the present application is as follows: by introducing a space-spectrum composite kernel function, a nonlinear unmixing problem is solved in a high-dimensional feature space, and the complex nonlinear mixing effect existing in the actual surface environment can be understood. The group sparse regularization based on the mineral paragenetic law is taken as the abundance constraint, the geological prior knowledge is integrated into the unmixing model, so that the abundance inversion result is more in line with the geological reality, the physically unreasonable mineral combination is reduced, and the reliability of the result is improved. In addition, by constructing a physically consistent regularization term based on a differentiable radiative transfer model and updating the endmember spectrum by using a manifold learning strategy, the extracted endmember spectrum has a clear physical meaning, and the endmember variability problem caused by illumination, atmosphere and mineral properties is solved, so that the endmember extraction is closer to the real situation.

[0054] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A hyperspectral-based mineral alteration feature extraction method, characterized in that, The method comprises the following steps: obtaining hyperspectral remote sensing image data of a region to be processed, performing atmospheric correction and noise reduction preprocessing on the hyperspectral remote sensing image data, and extracting initial endmember spectra from the preprocessed hyperspectral remote sensing image data by using an endmember extraction algorithm; constructing a target function of nonlinear spectral unmixing, wherein the target function comprises a data fidelity term for mapping a pixel to a high-dimensional feature space by using a space-spectrum composite kernel function, an abundance constraint term comprising group sparse regularization based on a mineral paragenetic prior, and a physical consistency regularization term for penalizing deviation between a spectrum generated by a physical model and a current endmember spectrum based on a differentiable radiative transfer model; the abundance constraint term of the group sparse regularization based on the mineral paragenetic prior comprises: dividing endmembers to be extracted having similar geological genesis or spatial paragenetic relationship into different groups according to prior knowledge of mineral paragenetic combination; and applying an L2 norm to an endmember abundance vector in each group, summing L2 norm results of all groups to obtain an L2,1 mixed norm, and taking the L2,1 mixed norm as the group sparse regularization term; the physical consistency regularization term for penalizing deviation between a spectrum generated by a physical model and a current endmember spectrum based on a differentiable radiative transfer model comprises: inputting mineral physical parameters corresponding to the endmember spectrum of the current iteration into the differentiable radiative transfer model to generate a physical model spectrum; and calculating a difference between the physical model spectrum and the current endmember spectrum as the physical consistency regularization term; iterative solving of the target function is performed by using an alternating optimization algorithm, and the endmember abundance matrix and the endmember spectrum are alternately updated in the iteration, wherein the endmember spectrum is updated according to a manifold learning strategy and is optimized by being constrained on a preset low-dimensional manifold; when the iteration converges or a preset termination condition is reached, the final endmember spectrum and an abundance matrix corresponding to the endmember spectrum are obtained; and alteration minerals are identified by using the final endmember spectrum, and mineral alteration characteristics of the region to be processed are extracted in combination with the abundance matrix corresponding to the endmember spectrum.

2. The hyperspectral-based mineral alteration feature extraction method according to claim 1, characterized in that, The space-spectrum composite kernel function is composed of a Gaussian radial basis space kernel function and a polynomial spectrum kernel function.

3. The hyperspectral-based mineral alteration feature extraction method according to claim 1, characterized in that, The iterative solving of the target function by using the alternating optimization algorithm comprises: in each iteration, the endmember spectrum is fixed, the abundance matrix is solved by using a multiplicative update rule to ensure non-negativity of the abundance update, and a projection operation is performed on the abundance vector of each pixel after each update to make the abundance vector satisfy the constraint that the sum is one.

4. The hyperspectral-based mineral alteration feature extraction method according to claim 1, characterized in that, The update of the endmember spectrum according to the manifold learning strategy and the optimization by being constrained on the preset low-dimensional manifold comprises: a training set is constructed by using reference spectra selected from a standard mineral spectrum library and the initial endmember spectra, and a low-dimensional manifold structure of the training set data is learned by using a manifold learning algorithm; when the endmember spectrum is iteratively updated, a gradient of the target function with respect to the endmember spectrum is projected onto a tangent space of the low-dimensional manifold, and the update is performed along the gradient direction after the projection.

5. The hyperspectral-based mineral alteration feature extraction method according to claim 1, characterized in that, The identification of alteration minerals by using the final endmember spectrum comprises: The similarity between the last endmember spectrum and reference spectra in a standard mineral spectrum library is calculated by using a spectral similarity measurement method; For each endmember spectrum, the spectrum is identified as the mineral with the highest similarity and higher than a preset similarity threshold in the reference spectrum.

6. The hyperspectral-based mineral alteration feature extraction method according to claim 1, characterized in that, The hyperspectral remote sensing image data of the region to be processed is acquired, and the hyperspectral remote sensing image data is subjected to atmospheric correction and noise reduction preprocessing, comprising: The hyperspectral image data of the mining area is collected by an infrared imaging spectrometer AVIRIS, the FLAASH model is used for atmospheric correction to obtain surface reflectivity data, and the reflectivity data is subjected to noise reduction processing by using a low-rank tensor approximation LRTA algorithm.

7. The hyperspectral-based mineral alteration feature extraction method according to claim 1, characterized in that, The endmember extraction algorithm is a vertex component analysis algorithm VCA.

8. The hyperspectral-based mineral alteration feature extraction method of claim 1, wherein, The alternating optimization algorithm is an alternating direction multiplier method ADMM.

Images