A mineral identification and classification method based on hyperspectral unmixing technology

Abstract

The present invention belongs to the technical field of hyperspectral image processing. The present invention provides a mineral identification and classification method based on hyperspectral unmixing technology, comprising the following steps: S1, using a hyperspectral imager to obtain hyperspectral images of relevant minerals; S2, constructing a mineral end-member spectral library; S3, establishing a bilinear spectral mixing model based on the hyperspectral images of relevant minerals; S4, using a robust spectral variable bilinear spatial regularization unmixing algorithm to perform spectral unmixing, and accurately estimate the abundance of each mineral contained in the mixed pixel; S5, according to the abundance of the solved minerals, quantify the corresponding proportion of each mineral, so as to achieve accurate identification and classification. This method solves the problems of complex mixed noise interference and spectral variation in hyperspectral images.

Description

Technical Field

[0001] The present invention belongs to the technical field of hyperspectral image processing, and in particular relates to a mineral identification and classification method based on hyperspectral unmixing technology. Background Art

[0002] Imaging spectroscopy technology organically combines the spectrum of ground objects determined by material composition with the spatial distribution image reflecting the existence pattern of ground objects, so that each pixel on the remote sensing image has both spectral information representing its physical properties and spatial information reflecting its geometric characteristics. In the identification and classification of minerals, people use hyperspectral imaging technology to obtain hyperspectral images of surface cover. By interpreting these spectral sample data, the mineral composition, chemical composition and structural characteristics present in the detection data can be accurately determined, and then potential mineral resources such as iron ore, copper ore, pyrite, bauxite, phosphate ore, etc. can be identified. By comparing with the spectral database of known mineral resources, it is possible to achieve rapid identification and quantitative analysis of potential mineral resources, providing important experimental data support for the development and utilization of mineral resources. Therefore, hyperspectral technology has important application prospects in mineral analysis and is of great significance for the exploration and development of mineral resources.

[0003] The difficulties currently faced by hyperspectral mineral identification and classification technology mainly include: 1. The spatial resolution of hyperspectral remote sensing sensors is limited. At the same time, the distribution of objects in nature is complex and diverse, especially the objects themselves are small in size, resulting in pixels rarely consisting of a single uniform feature of objects. These factors lead to a large number of mixed pixels in hyperspectral remote sensing images. 2. There are complex nonlinear multiple scattering and spectral variation phenomena between various ground object components, and traditional single linear models are difficult to accurately represent the ground object mixing process. 3. Due to spectral imaging equipment failure or poor imaging conditions, hyperspectral images are susceptible to complex noise pollution such as Gaussian noise, impulse noise, dead lines, stripe noise, etc., and there are degradation phenomena such as pixel defects, resulting in a low signal-to-noise ratio of the data, which brings great challenges to mineral identification and classification.

[0004] Therefore, it is necessary to design a mineral identification and classification method based on hyperspectral unmixing technology. Summary of the invention

[0005] The purpose of the present invention is to provide a mineral identification and classification method based on hyperspectral unmixing technology to solve the difficult problems faced by the current hyperspectral mineral identification and classification technology mentioned in the background technology.

[0006] To achieve the above object, the present invention provides a mineral identification and classification method based on hyperspectral unmixing technology, comprising the following steps:

[0007] S1. Use a hyperspectral imager to obtain hyperspectral images of relevant minerals;

[0008] S2, construct mineral end-member spectral library;

[0009] S3, establishing a bilinear spectral mixing model based on the hyperspectral images of relevant minerals;

[0010] S4. Use the robust spectral variable bilinear spatial regularization demixing algorithm to perform spectral demixing and accurately estimate the abundance of each mineral contained in the mixed pixel;

[0011] S5. Based on the obtained mineral abundance, quantify the corresponding proportion of each mineral to achieve accurate identification and classification.

[0012] In a specific implementation, in S2, the hyperspectral image acquired in S1 is preprocessed by a sparse unmixing algorithm through on-site acquisition or using an existing spectral library as a priori basis, and the mineral end members that may be contained in the hyperspectral image are analyzed, so as to construct a mineral end member spectral library that can basically cover the minerals in the observed hyperspectral image. The method of extracting mineral spectral curves includes a method of on-site acquisition using hardware equipment and a method of obtaining by a sparse unmixing algorithm using an existing spectral library as a priori basis.

[0013] In a specific implementation, the sparse unmixing algorithm includes a variable splitting and augmented Lagrangian algorithm (SUnSAL).

[0014] In a specific embodiment, a bilinear spectral mixture model of minerals is constructed:

[0015] Y=EA+FB+N+S (1)

[0016] Among them, Y represents the observed hyperspectral image, E is the mineral end-member spectral library, A represents the corresponding abundance matrix, F represents the bilinear end-member matrix, B represents the interaction abundance matrix, N represents the noise matrix, and S represents the sparse noise matrix.

[0017] In a specific embodiment, when using the robust spectrally variable bilinear spatial regularization demixing algorithm to process the hyperspectral image of minerals, the following sub-steps are included:

[0018] Reconstruct coarse-scale hyperspectral images;

[0019] Construct a coarse abundance guidance matrix;

[0020] Unmixing Algorithms Spectral Unmixing.

[0021] In a specific implementation, when reconstructing a coarse-scale hyperspectral image, a median filter is used to repair the hyperspectral image that is severely damaged by noise; then, a homogeneous area of ​​the spectral image is divided by superpixel segmentation, and a quasi-mean filtering technique is used to smooth the noise interference in the superpixel block, thereby constructing a coarse-scale hyperspectral image;

[0022] When constructing a hyperspectral image at a coarse scale, the tth superpixel block, each containing n t pixels, y k Y t For any pixel in , the operation of constructing a coarse-scale spectral image can be specifically defined as:

[0023]

[0024] in, Represents superpixel block Y t The corresponding coarse superpixel block Any column in .

[0025] In a specific embodiment, when constructing the coarse abundance guidance matrix, the ADMM algorithm is used to solve the abundance estimate at the coarse scale. Construct the rough abundance guidance matrix W filter , which can be specifically expressed as:

[0026]

[0027] in, is the estimated abundance matrix at the coarse scale, W fi l ter is the coarse abundance guidance matrix, i represents the i-th row, j represents the j-th column, (i,:) represents all elements in the i-th row, ||·||2 represents the 2-norm of the calculation vector, m is the number of endmembers in the mineral endmember spectral library E, and n is the number of pixels in the hyperspectral image. ε is a minimum value to ensure that the denominator is not zero.

[0028] In a specific embodiment, when the unmixing algorithm unmixes the spectrum,

[0029] Based on the coarse abundance guidance matrix, the spatial weighting factor W is introduced spa , a sparse noise regularization term is introduced to apply a noise sparse weighting matrix W to the sparse noise S n , the final unmixing model is obtained, which is carefully solved using the ADMM method to obtain the abundance matrix A;

[0030] Spatial weighting factor W spa Based on the spatial neighborhood information, it is expressed as:

[0031]

[0032] in Represents element a ij The neighborhood set, f(·) represents the function of mining spatial correlation through the neighborhood system; the spatial correlation of the image is characterized by using neighborhood information, which is specifically expressed as:

[0033]

[0034] Among them, ∈ represents the neighborhood weight. The neighborhood weight of adjacent elements is calculated as follows:

[0035]

[0036] Among them, the function im(·) is used to measure the two elements x i and x j The importance between them; let (a,b) and (c,d) be elements of x i and x j The spatial coordinates of the two are measured by calculating the Euclidean distance between them, and the expression is:

[0037]

[0038] W n The iterative update formula can be expressed as:

[0039]

[0040] The final unmixing model is expressed as follows:

[0041]

[0042] In a specific embodiment, in S5, the abundance matrix A data of each mineral is obtained according to the abundance of the solved mineral, and it is plotted into an abundance map of the same plane size as the hyperspectral image, and the final identification and classification results are obtained by analyzing the highlighted areas presented in the abundance map of each end member.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] The present invention develops a new mineral identification and classification method. The method constructs a bilinear mixture model based on the multiple scattering effect between the features of the ground objects, and uses a robust spectrally variable bilinear spatial regularization demixing algorithm to demix the image. At the same time, the method solves the problems of complex mixed noise interference and spectral variation in hyperspectral images.

[0045] The present invention combines the filter guidance weight and the spatial weighting factor to construct a new objective function, which fully exploits the spatial information in the hyperspectral image. At the same time, the noise sparse regularization term is introduced to resist noise interference to the maximum extent, effectively improving the effect of mineral classification and identification. The experimental results on simulated hyperspectral data and real hyperspectral data sets show that the present invention has achieved good recognition and classification effects.

[0046] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0048] Figure 1 It is a schematic diagram of a flow chart of an embodiment of the present invention;

[0049] Figure 2 is a true abundance diagram of each end member of an embodiment of the present invention;

[0050] Figure 3 is a comparison of the abundance map and the difference map of end member 1 obtained by each algorithm in one embodiment of the present invention;

[0051] Figure 4 It is a mineral map prepared by the United States Geological Survey in 1995;

[0052] Figure 5 is a pseudo-color image of a preliminary abundance matrix obtained during image preprocessing in one embodiment of the present invention;

[0053] Figure 6 is an end member spectral library diagram constructed in one embodiment of the present invention;

[0054] Figure 7 It is a comparison diagram of the identification results of some minerals in an embodiment of the present invention and a reference diagram. DETAILED DESCRIPTION

[0055] The embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. The specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0056] Example 1

[0057] The specific steps of the mineral identification and classification method based on hyperspectral unmixing technology of the present invention are as follows:

[0058] Acquire mineral hyperspectral images.

[0059] By collecting data on site or using an existing large spectral library as a priori basis, the hyperspectral image is preprocessed using a sparse unmixing algorithm such as the SUnSAL algorithm to analyze the mineral end members that may be contained in the hyperspectral image. In this way, an end member spectral library E that can basically cover the minerals in the observed hyperspectral image is constructed.

[0060] Then a bilinear spectral mixture model is established, which is expressed in the following form:

[0061] Y=EA+FB+N+S (1)

[0062] Where Y represents the observed hyperspectral image, Y∈R b×n , R is a real number set, b represents the number of bands, and n represents the number of pixels. E represents a mineral spectral library with m end members, E∈R b×m , A represents the corresponding abundance matrix, A∈R m×n , F represents the bilinear end member matrix, F∈R b×m(m-1) / 2 , B represents the interaction abundance matrix, B∈R m(m-1) / 2×n , N represents the noise matrix, N∈R b×n , S is the sparse noise matrix, S∈R b×n .

[0063] Based on the spectral mixing model, sparse unmixing is introduced into hyperspectral image unmixing to establish the initial objective function:

[0064]

[0065] where min represents the minimization function, ||·|| F represents the F norm. R(A)=||A||1 is the related sparse constraint term, and λ is the penalty parameter.

[0066] In order to construct a coarse abundance guidance matrix, the present invention sets a certain threshold to detect bad points in the hyperspectral image according to the characteristics of mixed noise:

[0067] Location=Find(Y ij ≤0.05||Y ij ≥0.95) (3)

[0068] Then, a median filter with a window size of 3×3 is used to repair the pixels in the hyperspectral image that are severely interfered by sparse noise. The specific operation can be expressed as:

[0069] Y(i,j)=medianFilter 3×3 (Y(i,j)) (4)

[0070] The hyperspectral image Y is segmented into s superpixel blocks by the SLIC algorithm, and then a class mean filter operation is used to average each homogeneous area after superpixel segmentation. t is the tth superpixel block, Each block contains n t pixels, y k Y t For any pixel in k ∈R d×1 (k=1,…,n t ), the operation can be specifically defined as:

[0071]

[0072] in, Represents superpixel block Y t The corresponding coarse superpixel block The superpixel blocks in the original hyperspectral image correspond one-to-one to the coarse pixels in the reconstructed image, and each column vector in the same coarse pixel has the same value. The coarse hyperspectral image reconstructs all elements in each homogeneous area of ​​the original image according to the average value, which can effectively reduce the impact of noise.

[0073] The ADMM method is used to unmix the reconstructed coarse-scale hyperspectral image to obtain the abundance estimation at the coarse scale. The specific solution formula is as follows:

[0074]

[0075] in represents the hyperspectral image after filtering, denoising and reconstruction, Represents abundance estimates at a coarse scale.

[0076] Calculate the rough abundance guidance matrix, and its calculation process can be expressed by the following formula:

[0077]

[0078] in, is the estimated abundance matrix at a coarse scale, W filter is the coarse abundance guidance matrix, i represents the i-th row, j represents the j-th column, (i,:) represents all elements in the i-th row, ||·||2 represents the 2-norm of the calculation vector, m is the number of endmembers in the mineral endmember spectral library E, and n is the number of pixels in the hyperspectral image. ε is a minimum value to ensure that the denominator is not zero.

[0079] In order to fully explore the rich spatial information in the hyperspectral image, a spatial weighting factor W is introduced. spa .

[0080] make W spa The spatial weighting factor W based on the spatial neighborhood information is the i-th row and j-th column element (i=1,...,m,j=1,...,n) at the t+1th iteration in spa It is expressed as:

[0081]

[0082] in Represents element a ij The neighborhood set of , f(·) represents the function of mining spatial correlation through the neighborhood system.

[0083] The spatial correlation of the image is characterized by using neighborhood information, which is specifically expressed as:

[0084]

[0085] Where ∈ represents the neighborhood weight. For simplicity, the present invention uses a 3×3 8-connected neighborhood window for algorithm design and experiment. Regarding the weight metric of neighboring elements, for any i and j, it can be calculated as follows:

[0086]

[0087] Among them, the function im(·) is used to measure the two elements a i and a j Let (a,b) and (c,d) be the elements of a i and a j The weight is measured by calculating the Euclidean distance between the two, and its expression is:

[0088]

[0089] Then, the filtered denoising abundance guidance matrix is ​​combined with the spatial weighting factor to effectively reduce the interference of noise and improve the accuracy of the abundance solution. The model can be expressed as:

[0090]

[0091] In order to more effectively solve the problems of sparse noise interference and spectral variation, a noise sparse regularization term is introduced to apply a noise sparse weighting matrix W to the sparse noise S. n ∈R b×n , which improves the algorithm's ability to resist noise and its resilience to spectral outliers. n The iterative update formula can be expressed as:

[0092]

[0093] Among them, sij represents the i-th row and j-th column element of the sparse noise matrix S.

[0094] The final unmixing model can be expressed as:

[0095]

[0096] By solving the above method using the alternating direction multiplier method, the abundance A of each mineral can be obtained.

[0097] Obtain the abundance matrix A data of each row in turn and plot it into an abundance map of the same plane size as the hyperspectral image. By analyzing the corresponding material end members and their abundance maps, the type and distribution of minerals can be obtained.

[0098] The beneficial effects of the present invention are verified by experiments as follows.

[0099] First, a simulation experiment was carried out to test the method of the present invention. In the experiment of the present invention, a mineral dictionary extracted from the United States Geological Survey (USGS) library was used. The spectral library E1 contains m=240 materials (different mineral types), and its spectral characteristics are L=224 spectral bands, which are evenly distributed in the range of 0.4-2.5μm. The simulated data is generated by 100×100 pixels and 9 end members. The end members are randomly selected from the spectral library E1, and the abundance follows the Dirichlet uniform probability distribution. The abundance data has very good spatial uniformity, and the component abundances are piecewise smooth, that is, they are smooth and have sharp transitions, which can well reveal the spatial characteristics of different unmixing algorithms. Then, three kinds of mixed noise are added to the hyperspectral image. First, all bands of the hyperspectral image are contaminated with zero-mean iid Gaussian noise with a signal-to-noise ratio Then, two kinds of sparse noise, namely impulse noise and dead line, are added to the obtained hyperspectral image. The 20-30 band is polluted by 10% impulse noise, and the 80-90 band is polluted by dead line. Finally, the nonlinear coefficients are uniformly drawn in the set [0,1] to obtain the nonlinear abundance matrix, and the nonlinear terms are added to obtain the simulated hyperspectral image. Figure 2 The true abundance plots of individual endmembers for the simulated data are shown.

[0100] Comparative experiments were carried out using FCLS, SUnSAL, GDA, SEMI-NMF, BPOGM, RGBM and the algorithm of the present invention to verify the effectiveness of the model.

[0101] The signal-reconstruction error ratio (SRE) is used to quantitatively measure the quality of the measurement results, which is defined as follows:

[0102]

[0103] Among them, A and are the actual abundance and the estimated abundance, respectively. Generally speaking, the larger the SRE (dB), the better the demixing performance.

[0104] Table 1 shows the demixing results of each algorithm under simulated data with different signal-to-noise ratios:

[0105] Table 1 SRE (dB) values ​​obtained by each algorithm for unmixing the simulated data

[0106]

[0107] By comparison, it can be known that the algorithm of the present invention has the best demixing effect.

[0108] In order to display the demixing effect more intuitively, Figure 3 The abundance of end member 1 estimated by each algorithm and the difference between the actual abundance and the estimated abundance are shown. Compared with other comparative methods, the abundance image obtained in this embodiment better fits the actual situation, which proves the effectiveness of this embodiment for abundance estimation.

[0109] Real data are used to further verify the beneficial effects of the present invention.

[0110] This experiment uses the Cuprite mineral dataset, which was acquired by the Airborne Visible Infrared Imaging Spectrometer (AVIRIS) at Cuprite, Nevada, USA in June 1997. The original image has 224 spectral bands with wavelengths between 0.4 and 2.5 μm and a standard spectral resolution of 10 nm. Due to the low signal-to-noise ratio and high water absorption in some bands, the noisy channels (1-2, 221-224) and water-absorbed channels (104-113, 148-167) were removed, leaving 188 bands. Figure 4 This is a mineral map prepared by the United States Geological Survey (USGS) in 1995.

[0111] The spectral library E1 of the United States Geological Survey (USGS) is used as prior information (each spectrum has a clear mineral name label). In order to obtain a sparser initial abundance, the sparse unmixing algorithm (RMSR) is selected for initialization to obtain a preliminary abundance matrix A1. Each row of the matrix is ​​traversed, and the abundance values ​​less than 0.005 are reset to zero, and the number of abundance row data corresponding to each end member is calculated to be greater than 0. If the number of abundance row data greater than 0 is less than 1 / 200 of the total number of image pixels, the corresponding end member is deleted. Figure 5 The pseudo-color image of the abundance matrix A1 is shown, and it can be seen that there are about 20 end members with relatively obvious abundance values. Finally, 30 end members were extracted as the mineral spectrum library E of the image, as shown in Figure 6 shown.

[0112] The model of the present invention is used to solve the abundance matrix A, obtain the abundance data of each row, and draw an abundance map with the same plane size as the hyperspectral image.

[0113] The present invention also uses Tricorder 3.3 software to map the minerals in the mining area, taking the minerals Alunite (alunite), Buddingtonite (ammonium feldspar), and Muscovite (white mica) as examples. Figure 7 The distribution and estimated abundance of Alunite, Buddingtonite and Muscovite minerals drawn by Tricorder 3.3 software are shown. It can be seen that the method of the present invention can effectively identify the distribution of minerals, which can provide great help for the classification and identification of minerals.

[0114] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions and substitutions can be made without departing from the concept of the present invention, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A mineral identification and classification method based on hyperspectral unmixing technology, characterized in that: The steps include: S1. Use a hyperspectral imager to obtain hyperspectral images of relevant minerals; S2, construct mineral end-member spectral library; S3, establishing a bilinear spectral mixing model based on the hyperspectral images of relevant minerals; S4. Use the robust spectral variable bilinear spatial regularization demixing algorithm to perform spectral demixing and accurately estimate the abundance of each mineral contained in the mixed pixel; When using the robust spectrally variable bilinear spatial regularization demixing algorithm to process the hyperspectral image of minerals, the following steps are included: Reconstruct coarse-scale hyperspectral images; Construct a coarse abundance guidance matrix; Unmixing algorithm Spectral unmixing; When constructing the coarse abundance guidance matrix, the ADMM algorithm is used to solve the abundance estimate at the coarse scale. Construct the rough abundance guidance matrix W filter , which can be specifically expressed as: in, is the estimated abundance matrix at a coarse scale, W filter is the coarse abundance guidance matrix, i represents the i-th row, j represents the j-th column, (i,:) represents all elements in the i-th row, ||·||2 represents the 2-norm of the calculation vector, m is the number of endmembers in the mineral endmember spectral library E, n is the number of pixels in the hyperspectral image, and ε is a minimum value to ensure that the denominator is not 0; S5. Based on the obtained mineral abundance, quantify the corresponding proportion of each mineral to achieve accurate identification and classification.

2. The method for mineral identification and classification based on hyperspectral unmixing technology according to claim 1 is characterized in that: In S2, the hyperspectral image acquired in S1 is preprocessed by a sparse unmixing algorithm through on-site collection or by using an existing spectral library as a priori basis, and the mineral end members that may be contained in the hyperspectral image are analyzed, so as to construct a mineral end member spectral library that can basically cover the minerals in the observed hyperspectral image.

3. The method for mineral identification and classification based on hyperspectral unmixing technology according to claim 2 is characterized in that: The sparse unmixing algorithm includes the Split-and-Augmented Lagrangian Algorithm (SUnSAL).

4. The method for mineral identification and classification based on hyperspectral unmixing technology according to claim 1 is characterized in that: Construct a bilinear spectral mixture model for minerals: Y=EA+FB+N+S(1) Among them, Y represents the observed hyperspectral image, E is the mineral end-member spectral library, A represents the corresponding abundance matrix, F represents the bilinear end-member matrix, B represents the interaction abundance matrix, N represents the noise matrix, and S represents the sparse noise matrix.

5. The method for mineral identification and classification based on hyperspectral unmixing technology according to claim 1 is characterized in that: When reconstructing the coarse-scale hyperspectral image, median filtering is used to repair the hyperspectral image that is severely damaged by noise; then, the homogeneous area of ​​the spectral image is divided by superpixel segmentation, and the noise interference in the superpixel block is smoothed by the mean-like filtering technology, so as to construct the hyperspectral image at a coarse scale; When constructing a hyperspectral image at a coarse scale, the tth superpixel block, each containing n t pixels, y k Y t For any pixel in , the operation of constructing a coarse-scale spectral image can be specifically defined as: in, Represents superpixel block Y t The corresponding coarse superpixel block Any column in .

6. The method for mineral identification and classification based on hyperspectral unmixing technology according to claim 1 is characterized in that: Unmixing algorithm When unmixing spectra, Based on the coarse abundance guidance matrix, the spatial weighting factor W is introduced spa , a sparse noise regularization term is introduced to apply a noise sparse weighting matrix W to the sparse noise S n , the final unmixing model is obtained, which is carefully solved using the ADMM method to obtain the abundance matrix A; Spatial weighting factor W spa Based on the spatial neighborhood information, it is expressed as: in Represents element a ij The neighborhood set, f(·) represents the function of mining spatial correlation through the neighborhood system; the spatial correlation of the image is characterized by using neighborhood information, which is specifically expressed as: Among them, ∈ represents the neighborhood weight. The neighborhood weight of adjacent elements is calculated as follows: Among them, the function im(·) is used to measure the two elements x i and x j The importance between them; let (a,b) and (c,d) be elements of x i and x j The spatial coordinates of the two are measured by calculating the Euclidean distance between them, and the expression is: W n The iterative update formula can be expressed as: The final unmixing model is expressed as follows:

7. The method for mineral identification and classification based on hyperspectral unmixing technology according to claim 1 is characterized in that: In S5, the abundance matrix A data of each mineral is obtained according to the abundance of the solved minerals, and is plotted into an abundance map of the same plane size as the hyperspectral image, and the final recognition and classification results are obtained by analyzing the highlighted areas presented in the abundance map of each end member.

Images