Multimodal remote sensing image spectrum unmixing method based on combination of digital surface model DSM and hyperspectrum

Abstract

The invention belongs to the field of remote sensing image processing, and relates to a multimodal remote sensing image spectrum unmixing method based on combination of a digital surface model (DSM) and a hyperspectrum. The invention aims to solve the problems that the existing method does not consider that DSM and hyperspectral data are actually different-dimension information from different sensors in the same scene and does not model joint imaging of the DSM and the hyperspectral data in a model imaging process, so that the spectral unmixing precision of a remote sensing image is poor. The method comprises the following steps: acquiring multi-modal remote sensing images with the same coverage geographic area and the same resolution; the multi-modal remote sensing image comprises a hyperspectral remote sensing image and DSM remote sensing image data; based on DSM remote sensing image data, geometric component information of ground feature types is obtained; constructing a DSM and hyperspectrum combination-based multi-modal remote sensing image spectrum unmixing model; and solving the DSM and hyperspectrum combination-based multi-modal remote sensing image spectrum unmixing model by adopting an alternating direction multiplier method to obtain abundance information.

Description

Technical Field

[0001] The invention belongs to the field of remote sensing image processing and relates to a multi-modal remote sensing image spectral unmixing method. Background Art

[0002] High-precision spectral unmixing of hyperspectral remote sensing images faces a challenging problem, that is, it is difficult to distinguish different artificial target categories that present similar spectra at different altitudes without elevation information. Digital surface models (DSMs) with the same spatial resolution as hyperspectral images can provide additional geometric information, which is very useful for the spectral unmixing task of hyperspectral images. In addition, in hyperspectral images, spectral variability is also unavoidable, that is, the spectral curve of the same ground object will also change due to changes in external imaging conditions. This phenomenon is called the "same object, different spectrum" phenomenon. The hyperspectral unmixing model can obtain the shadow component and reflectance component by introducing the intrinsic decomposition model, where the shadow component is the result of the interaction between the illumination information and the ground object surface, and the reflectance component can characterize the inherent properties of the ground object target and is not affected by changes in external imaging conditions. Based on this, the reflectance component can greatly reduce the influence of spectral variability, and then can be decomposed into endmember information and abundance information using a linear mixed model. From the perspective of physical modeling, by combining DSM data, the shadow component can be modeled through the interaction of the geometric component and the illumination component of DSM. At the same time, by imposing prior constraints on the reflection component and the abundance component, more realistic reflectance information can be obtained based on the physical model terms and prior knowledge constraints, and then more accurate abundance information can be estimated.

[0003] At present, there are two main types of multimodal joint spectral unmixing methods, one is based on deep learning, and the other is based on physical models. The deep learning-based method usually inputs the geometric information provided by DSM as a spatial feature item into the network for feature enhancement, which is equivalent to introducing additional prior knowledge for assistance; the physical model-based method applies weights to improve the separability of different target types presenting similar spectra at different altitudes; however, the existing multimodal remote sensing image spectral unmixing based on the joint use of DSM and hyperspectral does not take into account that DSM data and hyperspectral data are actually different dimensional information from different sensors observing the same scene, and does not model the joint imaging of DSM data and hyperspectral data in the process of model imaging, so the model is slightly insufficient in completeness. Summary of the invention

[0004] The purpose of the present invention is to solve the problem that the existing multimodal remote sensing image spectral unmixing based on the combination of DSM and hyperspectral does not take into account that DSM data and hyperspectral data are actually different dimensional information from different sensors observing the same scene, and does not model the joint imaging of DSM data and hyperspectral data in the process of model imaging, resulting in poor spectral unmixing accuracy of remote sensing images. A multimodal remote sensing image spectral unmixing method based on the combination of digital surface model DSM and hyperspectral is proposed.

[0005] The specific process of the multimodal remote sensing image spectral unmixing method based on the digital surface model DSM and hyperspectral combination is as follows:

[0006] Step 1: Obtain multimodal remote sensing images covering the same geographical area and with the same resolution;

[0007] The multimodal remote sensing image includes hyperspectral remote sensing image data and DSM remote sensing image data;

[0008] Obtain geometric component information of ground object types based on DSM remote sensing image data;

[0009] The DSM is a digital surface model;

[0010] Step 2: Construct a multimodal remote sensing image spectral unmixing model based on the combination of DSM and hyperspectral;

[0011] Step 3: Use the alternating direction multiplier method to solve the multimodal remote sensing image spectral unmixing model based on the combination of DSM and hyperspectral constructed in step 2 to obtain abundance information.

[0012] The beneficial effects of the present invention are:

[0013] The purpose of the present invention is to improve the spectral unmixing accuracy of remote sensing images, and proposes a multimodal remote sensing image spectral unmixing method based on the combination of DSM and hyperspectral.

[0014] The present invention solves the above problems by constructing a multimodal imaging physical model (currently, there are two main types of multimodal joint spectral unmixing methods, one is a method based on deep learning, and the other is a method based on a physical model. The method based on deep learning usually inputs the geometric information provided by DSM as a spatial feature item into the network for feature enhancement, which is equivalent to introducing additional prior knowledge for assistance; the method based on the physical model applies weights to improve the separability of different target types presenting similar spectra at different altitudes; however, the existing multimodal remote sensing image spectral unmixing based on the combination of DSM and hyperspectral does not take into account that DSM data and hyperspectral data are actually different dimensional information from different sensors observing the same scene, and does not model the joint imaging of DSM data and hyperspectral data in the process of model imaging, so it is slightly insufficient in model completeness.); considering that there will inevitably be the same object with different spectra in remote sensing images, that is, the spectral curves of the same ground object target type are different due to changes in external imaging conditions. This phenomenon is not conducive to distinguishing the characteristics of different categories, and is also not conducive to unmixing. The design idea of ​​the present invention is to embed the intrinsic decomposition model in the multimodal spectral unmixing model, so as to obtain the reflectance component with greatly reduced spectral variability. This component is only related to the characteristics of the material itself and is not affected by the external imaging conditions, so as to obtain a more accurate unmixing result. Specifically, in the construction of the multimodal spectral unmixing model, the DSM and hyperspectral data are combined, and the multimodal imaging model item can be constructed using the physical model. The spectral information of the hyperspectral data and the spatial geometric information of the DSM data can be used at the same time. The connection between the DSM data and the illumination is established during the imaging process, so as to obtain the real light and dark component information, and then the intrinsic decomposition model can be used to obtain more accurate reflectance component information. Based on the global sparsity prior constraints of the local similarity of the reflectance component and the abundance information, the unmixing result with greatly reduced spectral variability can be obtained; that is, the constructed multimodal remote sensing image spectral unmixing model can theoretically significantly improve the accuracy of the hyperspectral image unmixing, and the obtained hyperspectral unmixing result is more authentic and accurate.

[0015] The method of the present invention constructs a multimodal physical imaging model, combines the intrinsic image decomposition model, and uses the spherical harmonic illumination model to obtain the light and dark components, so as to obtain reflectance information with reduced spectral variability. After being constrained by prior knowledge, the method optimizes the hyperspectral unmixing result with higher authenticity and accuracy, and finally realizes high-precision spectral unmixing of multimodal remote sensing images based on the combination of DSM and hyperspectral.

[0016] In order to verify the performance of the present invention, a set of real hyperspectral-DSM image pairs was verified. The experimental results show that compared with the current representative methods, the hyperspectral image unmixing results obtained by this method are more accurate and have the smallest reconstruction error. The experimental results verify the effectiveness of the multimodal remote sensing image spectral unmixing method based on the combination of DSM and hyperspectral proposed in the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the implementation process of the present invention;

[0018] Figure 2a This is a false color composite of a hyperspectral image;

[0019] Figure 2b This is a schematic diagram of the DSM image;

[0020] Figure 3 Abundance maps of different land feature types using different methods. DETAILED DESCRIPTION

[0021] Specific implementation method 1: The specific process of the multimodal remote sensing image spectral unmixing method based on the digital surface model DSM and hyperspectral combination in this implementation method is as follows:

[0022] Step 1: Obtain multimodal remote sensing images covering the same geographical area and with the same resolution;

[0023] The multimodal remote sensing image includes hyperspectral remote sensing image data and DSM remote sensing image data;

[0024] Obtain geometric component information of ground object types based on DSM remote sensing image data;

[0025] The DSM is a digital surface model;

[0026] Step 2: Construct a multimodal remote sensing image spectral unmixing model based on the combination of DSM and hyperspectral;

[0027] Step 3: Use the alternating direction multiplier method to solve the multimodal remote sensing image spectral unmixing model based on DSM and hyperspectral constructed in step 2 to obtain more accurate abundance information, and reconstruct a hyperspectral image consistent with the original resolution based on the abundance information.

[0028] Specific implementation method 2: This implementation method is different from specific implementation method 1 in that in step 1, multimodal remote sensing images covering the same geographical area and having the same resolution are obtained;

[0029] The multimodal remote sensing image includes hyperspectral remote sensing image data and DSM remote sensing image data;

[0030] Obtain geometric component information of ground object types based on DSM remote sensing image data;

[0031] The DSM is a digital surface model;

[0032] The specific process is:

[0033] Step 1

[0034] make and They respectively represent the acquired hyperspectral remote sensing image data and DSM remote sensing image data covering the same geographical area and having the same resolution;

[0035] in, represents the spectral characteristics of the k-th pixel in the hyperspectral remote sensing image data, k = 1, 2, ..., P, k represents the k-th pixel of the hyperspectral image, and P represents the total number of pixels in the hyperspectral image;

[0036] λ1 represents the first wavelength, λ2 represents the second wavelength, and λ3 represents the third wavelength. represents the wavelength of the first band of the k-th pixel, represents the wavelength of the second band of the k-th pixel, represents the wavelength of the dth band of the kth pixel, and the superscript T indicates the transposition;

[0037] D k Indicates the elevation corresponding to the kth pixel in the DSM remote sensing image data;

[0038] d H Represents the spectral dimension of the sample in the hyperspectral remote sensing image, d 1 Represents the spectral dimension of samples in DSM remote sensing images, Represents a domain;

[0039] Step 1 and 2: Calculate the normal of each pixel and obtain the normal vector features as follows:

[0040]

[0041] Among them, N represents the normal vector feature, n k Represents the projection of the normal of the k-th pixel on the x, y, and z spatial coordinates (spatial rectangular coordinate system), n k =[n k,x ,n k,y ,n k,z ] T ,k=1,2,...,P;n k,x Represents the projection of the normal of the k-th pixel on the x-space coordinate axis, n k,yRepresents the projection of the normal of the k-th pixel on the y-space coordinate axis, n k,z Represents the projection of the normal of the k-th pixel on the z-space coordinate axis;

[0042] Step 13: Calculate the geometric component G based on step 12:

[0043] G=[G1,G2,...,G k ,...,G P ] T

[0044] Among them, G k represents the geometric component of the kth pixel, k = 1, 2, ..., P;

[0045] G k =

[0046] [c4,2c2n k,y ,2c2n k,x ,c2n k,z ,2c1n k,x n k,y ,2c1n k,y n k,z ,c3n k,z n k,z -c5,2c1n k,x n k,z ,c1(n k, x n k,x -n k,y n k,y )] T Among them, c1, c2, c3, c4, and c5 are all constants.

[0047] The other steps and parameters are the same as those in the first embodiment.

[0048] Specific implementation method three: This implementation method is different from specific implementation methods one or two in that the values ​​of the constants c1, c2, c3, c4, and c5 are: c1=0.429, c2=0.512, c3=0.743, c4=0.886, c5=248.

[0049] The other steps and parameters are the same as those in the first or second embodiment.

[0050] Specific implementation method 4: This implementation method is different from any one of specific implementation methods 1 to 3 in that in step 2, a multimodal remote sensing image spectral unmixing model based on the combination of DSM and hyperspectral is constructed;

[0051] The specific process is:

[0052] Step 21: Introduce the intrinsic image decomposition model. The intrinsic image decomposition model transforms the hyperspectral remote sensing image Y H Decomposed into the dot product of the reflectivity component S and the brightness component R:

[0053] Among them, Y H Represents a hyperspectral remote sensing image;

[0054] Indicates the light and dark components, which can characterize changes in external imaging conditions;

[0055] Represents the reflectance component, which can characterize the reflectance characteristics of the material itself; R can more accurately represent the reflectance spectrum curve of the object and reduce the spectral uncertainty under changing external conditions;

[0056] In the traditional spectral unmixing model, the original hyperspectral image can be decomposed into the combined effect of endmember information and abundance information, where the endmember represents the type of ground object in the data, and the abundance represents the set of proportions of different ground objects in the hyperspectral image data. Therefore, the purpose of spectral unmixing is to realize the process of extracting the reflectance of the spectral information of the ground object and calculating the corresponding abundance, that is, endmember extraction and abundance estimation.

[0057] Step 22: Introduce a linear spectral unmixing model;

[0058] Step 2 and 3: Based on the intrinsic image decomposition model and the linear spectral unmixing model, construct the objective function of the multimodal remote sensing image spectral unmixing model based on the combination of DSM and hyperspectral.

[0059] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0060] Specific implementation mode 5: This implementation mode is different from the specific implementation modes 1 to 4 in that a linear spectral unmixing model is introduced in step 22; the expression is as follows:

[0061] Y H =EA+N

[0062] in, represents the endmember matrix, represents the abundance matrix, represents the measurement error; C represents the number of end members.

[0063] The abundance matrix has two basic physical constraints, namely, the abundance non-negative constraint and the abundance sum-to-one constraint. As the name implies, in natural scenes, the proportion of land object types is non-negative, and in hyperspectral image data, the sum of the proportions of all land object types is 1;

[0064] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0065] Specific implementation method 6: This implementation method is different from the specific implementation methods 1 to 5 in that, in the steps 2 and 3, based on the intrinsic image decomposition model and the linear spectral unmixing model, an objective function of a multimodal remote sensing image spectral unmixing model based on a combination of DSM and hyperspectral is constructed; the objective function is expressed as follows:

[0066]

[0067] stA≥0

[0068] in, represents the square of the Frobenius norm; α and β are penalty coefficients for balancing the importance of each constraint; M represents ambient lighting; G represents the geometric component; 1 P represents a P-dimensional all-1 matrix, T represents the transpose; Ψ(R) represents the prior constraint of the reflectivity component, Υ(A) represents the prior constraint of the abundance information, U represents the objective function; ⊙ represents the Hadamard product;

[0069] The items multiplied after the α and β coefficients are constraint items, which reflect the importance of each item by setting the penalty coefficients of these constraint items.

[0070] The first term in the objective function It is the intrinsic image decomposition model term, which is used to decompose the original hyperspectral remote sensing image into the light and dark components that characterize the changes in external imaging conditions and the reflectance components that characterize the reflectance characteristics of the material itself, and ensure that the error of the dot product of the original hyperspectral remote sensing image and the reflectance component S and the light and dark component R is as small as possible;

[0071] Item 2 is the light and dark component term, which is used to decompose the original hyperspectral remote sensing image into the interaction between the geometric component G obtained from the digital surface model DSM remote sensing image data and the ambient illumination M representing the external imaging conditions, and introduce the spherical harmonic illumination coefficient for solution;

[0072] Item 3 The spectral unmixing imaging model item is replaced by the reflectance component. Since the reflectance component is only related to the characteristics of the ground object type itself, the abundance information obtained is more realistic.

[0073] The fourth term Ψ(R) is the a priori constraint term for the reflectivity component;

[0074] The fifth term Υ(A) is the prior constraint on abundance information.

[0075] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0076] Specific implementation method 7: This implementation method is different from any one of specific implementation methods 1 to 6 in that the process of obtaining the prior constraint term Ψ(R) of the reflectivity component in the objective function is:

[0077] Constraint 1 in the prior constraint term Ψ(R) of the reflectance component: Adjacent pixels have a greater probability of having similar reflectances, so the spectrum of a pixel can be a linear combination of adjacent pixels;

[0078] R i =w ij R j

[0079] Among them, R i and R j Respectively represent the reflectivity information of pixel i and pixel j in R, w ij represents the similarity measure between pixel i and pixel j in R;

[0080] i=1,2,...,P, i represents the i-th pixel of the hyperspectral image, and P represents the total number of pixels of the hyperspectral image;

[0081] j=1,2,...,P, j represents the jth pixel of the hyperspectral image, i≠j;

[0082] Constraint 2 in the prior constraint term Ψ(R) of the reflectivity component: pixels with similar elevation information have a greater probability of having similar reflectivity;

[0083] R i =h ij R j

[0084] Among them, R i and R j Respectively represent the reflectivity information of pixel i and pixel j in R, h ij represents the elevation similarity measure between pixel i and pixel j in R;

[0085] In summary, the prior constraints of the reflectivity component are expressed as follows:

[0086]

[0087] Among them, θ w and θ h Respectively represent the a priori weights of the two reflections;

[0088] The prior constraint term Υ(A) of abundance information is based on the assumption that for a natural scene, the number of land object types contained in the scene is limited, so the distribution is sparse globally.

[0089] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0090] Specific implementation eight: This implementation differs from any one of specific implementations one to seven in that the a priori constraint term Y(A) of abundance information in the objective function is expressed as follows:

[0091] Υ(A)=γ||A|| 1,1

[0092] Among them, γ represents the coefficient, |||| 1,1 Represents a norm.

[0093] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0094] Specific implementation method 9: This implementation method is different from any one of specific implementation methods 1 to 8 in that in step 3, the alternating direction multiplier method is used to solve the multimodal remote sensing image spectral unmixing model based on DSM and hyperspectral combined constructed in step 2 to obtain more accurate abundance information, and a hyperspectral image consistent with the original resolution is reconstructed based on the abundance information; the specific process is as follows:

[0095] The objective function of the multimodal remote sensing image spectral unmixing model based on DSM and hyperspectral combination is as follows:

[0096]

[0097] stA≥0

[0098]

[0099] When using the alternating direction multiplier method, it is necessary to introduce parameters R = J and A = K, and rewrite the objective function of the multimodal remote sensing image spectral unmixing method based on the combination of DSM and hyperspectral into the following form:

[0100]

[0101] stR=J,A=K,1 T A=1

[0102] in, represents the set containing R, A, J, K;

[0103] J and K represent auxiliary variables;

[0104] The above formula is transformed into the Lagrangian augmented function of the multimodal remote sensing image spectral unmixing model based on the combination of DSM and hyperspectral:

[0105]

[0106] st1 T A=1

[0107] Among them, Λ1 and Λ2 are Lagrange multipliers, σ1 and σ2 are penalty coefficients; is the Lagrangian augmented function;

[0108] Based on the idea of ​​constraining the zero point of the first-order derivative of the model to be the extreme point, the Lagrangian augmented function is optimized to obtain the reflectivity component R, auxiliary variable J, auxiliary group variable K and abundance information A.

[0109] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0110] Specific implementation ten: This implementation is different from any one of specific implementations one to nine in that the Lagrangian augmented function is optimized based on the idea that the zero point of the first-order derivative of the constraint model is an extreme point to obtain the reflectivity component R, the auxiliary variable J, the auxiliary variable K and the abundance information A;

[0111] The specific process is:

[0112] 1) Optimize the reflectivity component R:

[0113] The objective function of the reflectivity component R is expressed as follows:

[0114]

[0115] The above formula can also be expressed as follows:

[0116]

[0117] Where T represents the auxiliary variable; S⊙T=Y, The equivalent expression is The solution results are as follows:

[0118]

[0119] T0=2W

[0120] T1=R-WR

[0121] T2=2H

[0122] T3=R-HR

[0123] Among them, T0, T1, T2, and T3 represent auxiliary variables;

[0124] 2) Optimize auxiliary variables J and K:

[0125] J=σ1R-Λ1 / σ1

[0126] K=sign(A-Λ2 / σ2)⊙max{0,A-Λ2 / σ2|-1 / σ2}

[0127] 3) Optimize abundance information A:

[0128] The objective function of abundance information A is expressed as follows:

[0129]

[0130] The solution results are as follows:

[0131] A=(2σ2K-E T R-Λ2I)×(2σ2-E T E) -1 .

[0132] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0133] The following examples are used to verify the beneficial effects of the present invention:

[0134] Embodiment 1:

[0135] A multimodal remote sensing image spectral unmixing method based on DSM and hyperspectral combination is prepared according to the following steps:

[0136] The multi-mode dataset used in the experiment includes a hyperspectral image and a DSM image. The data has been pre-processed by atmospheric and geometric correction. The hyperspectral data contains 349×1905 pixels and was acquired by the ITERS CASI-1500 sensor on the campus of the University of Houston and its surrounding rural areas in Texas, USA. The Hyperion sensor operates in the band range of 364-1046nm, with a spatial resolution of 2.5m and a total of 144 bands. The experiment uses a sub-pixel dataset of 170×170 size after cropping the image as the experimental dataset.

[0137] False color composite of hyperspectral image and DSM data image Figure 2a and Figure 2b The experiment uses the original large-scene original hyperspectral image as the baseline, and compares the reconstruction error between the abundance estimation result and the data reconstructed by the endmember information and the original image, as well as the reconstruction signal-to-noise ratio, with seven representative methods. Figure 3 The abundance maps of different types of features are given, and the quantitative indicators of different comparison methods are given in Table 1. It can be seen that the abundance estimation results of this method are more accurate in visual discrimination, and the reconstruction error and reconstructed data signal-to-noise ratio results are better than other comparison methods. The experimental results verify the effectiveness of the multimodal remote sensing image spectral unmixing method based on DSM and hyperspectral combined proposed in this paper.

[0138] Table 1 Quantitative indicators of different comparison methods

[0139]

[0140] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A multimodal remote sensing image spectral unmixing method based on a digital surface model (DSM) and a hyperspectral image, characterized in that: The specific process of the method is: Step 1: Obtain multimodal remote sensing images covering the same geographical area and with the same resolution; Obtain geometric component information of ground object types based on DSM remote sensing image data; The DSM is a digital surface model; The multimodal remote sensing image includes hyperspectral remote sensing image data and DSM remote sensing image data; Step 2: Construct a multimodal remote sensing image spectral unmixing model based on the combination of DSM and hyperspectral; Step 3: Use the alternating direction multiplier method to solve the multimodal remote sensing image spectral unmixing model based on the combination of DSM and hyperspectral constructed in step 2 to obtain abundance information.

2. The multimodal remote sensing image spectral unmixing method based on the combination of digital surface model DSM and hyperspectral according to claim 1 is characterized in that: In the step 1, multimodal remote sensing images covering the same geographical area and having the same resolution are obtained; Obtain geometric component information of ground object types based on DSM remote sensing image data; The DSM is a digital surface model; The multimodal remote sensing image includes hyperspectral remote sensing image data and DSM remote sensing image data; The specific process is: Step 1 make and They respectively represent the acquired hyperspectral remote sensing image data and DSM remote sensing image data covering the same geographical area and having the same resolution; in, represents the spectral characteristics of the k-th pixel in the hyperspectral remote sensing image data, k = 1, 2, ..., P, k represents the k-th pixel of the hyperspectral image, and P represents the total number of pixels in the hyperspectral image; λ1 represents the first wavelength, λ2 represents the second wavelength, and λ3 represents the third wavelength. represents the wavelength of the first band of the k-th pixel, represents the wavelength of the second band of the k-th pixel, represents the wavelength of the dth band of the kth pixel, and the superscript T indicates the transposition; D k Indicates the elevation corresponding to the kth pixel in the DSM remote sensing image data; d H Represents the spectral dimension of the sample in the hyperspectral remote sensing image, d 1 Represents the spectral dimension of samples in DSM remote sensing images, Represents a domain; Step 1 and 2: Calculate the normal of each pixel and obtain the normal vector features as follows: Among them, N represents the normal vector feature, n k Represents the projection of the normal of the k-th pixel on the x, y, and z spatial coordinate axes, n k =[n k,x ,n k,y ,n k,z ] T ,k=1,2,...,P;n k,x Represents the projection of the normal of the k-th pixel on the x-space coordinate axis, n k,y Represents the projection of the normal of the kth pixel on the y space coordinate axis, n k,z Represents the projection of the normal of the k-th pixel on the z-space coordinate axis; Step 13: Calculate the geometric component G based on step 12: G=[G1,G2,...,G k ,...,G P ] T Among them, G k represents the geometric component of the k-th pixel, k = 1, 2, ..., P; G k ＝ [c4,2c2n k,y ,2c2n k,x ,c2n k,z ,2c1n k,x n k,y ,2c1n k,y n k,z ,c3n k,z n k,z -c5,2c1n k,x n k,z ,c1(n k,x n k,x -n k,y n k,y )] T Among them, c1, c2, c3, c4, and c5 are all constants.

3. The multimodal remote sensing image spectral unmixing method based on the combination of digital surface model DSM and hyperspectral according to claim 2 is characterized in that: The values ​​of the constants c1, c2, c3, c4, and c5 are: c1=0.429, c2=0.512, c3=0.743, c4=0.886, c5=248.

4. The multimodal remote sensing image spectral unmixing method based on the combination of digital surface model DSM and hyperspectral spectrum according to claim 2 or 3, characterized in that: In the step 2, a multimodal remote sensing image spectral unmixing model based on the combination of DSM and hyperspectral is constructed; the specific process is: Step 21: Introduce the intrinsic image decomposition model. The intrinsic image decomposition model transforms the hyperspectral remote sensing image Y H Decomposed into the dot product of the reflectivity component S and the brightness component R: Among them, Y H Represents a hyperspectral remote sensing image; Indicates the light and dark components; represents the reflectivity component; Step 22: Introduce a linear spectral unmixing model; Step 2 and 3: Based on the intrinsic image decomposition model and the linear spectral unmixing model, construct the objective function of the multimodal remote sensing image spectral unmixing model based on the combination of DSM and hyperspectral.

5. The multimodal remote sensing image spectral unmixing method based on the combination of digital surface model DSM and hyperspectral according to claim 4 is characterized in that: The linear spectral unmixing model is introduced in step 22; the expression is as follows: Yes H =EA+N in, represents the endmember matrix, represents the abundance matrix, represents the measurement error; C represents the number of end members.

6. The multimodal remote sensing image spectral unmixing method based on digital surface model DSM and hyperspectral combination according to claim 5 is characterized by: In the steps 2 and 3, based on the intrinsic image decomposition model and the linear spectral unmixing model, an objective function of a multimodal remote sensing image spectral unmixing model based on a combination of DSM and hyperspectral is constructed; the objective function is expressed as follows: in, represents the square of the Frobenius norm; α and β are penalty coefficients for balancing the importance of each constraint; M represents ambient lighting; G represents the geometric component; 1 P represents a P-dimensional all-one matrix, T represents the transpose; Ψ(R) represents the prior constraint term of the reflectivity component, Υ(A) represents the prior constraint term of the abundance information, U represents the objective function; ⊙ represents the Hadamard product.

7. The multimodal remote sensing image spectral unmixing method based on the combination of digital surface model DSM and hyperspectral spectrum according to claim 6 is characterized in that: The acquisition process of the prior constraint term Ψ(R) of the reflectivity component in the objective function is: Constraint 1 in the prior constraint term Ψ(R) of the reflectivity component: R i =w ij R j Among them, R i and R j Respectively represent the reflectivity information of pixel i and pixel j in R, w ij represents the similarity measure between pixel i and pixel j in R; i=1,2,...,P, i represents the i-th pixel of the hyperspectral image, and P represents the total number of pixels of the hyperspectral image; j=1,2,...,P, j represents the jth pixel of the hyperspectral image, i≠j; The second constraint in the prior constraint term Ψ(R) of the reflectivity component is: R i =h ij R j Among them, R i and R j Respectively represent the reflectivity information of pixel i and pixel j in R, h ij represents the elevation similarity measure between pixel i and pixel j in R; In summary, the prior constraints of the reflectivity component are expressed as follows: Among them, θ w and θ h They represent the weights of the two reflection priors respectively.

8. The multimodal remote sensing image spectral unmixing method based on the combination of digital surface model DSM and hyperspectral spectrum according to claim 7 is characterized in that: The prior constraint term Y(A) of abundance information in the objective function is expressed as follows: Y(A)=γ||A|| 1,1 Among them, γ represents the coefficient, || || 1,1 Represents a norm.

9. The multimodal remote sensing image spectral unmixing method based on the combination of digital surface model DSM and hyperspectral spectrum according to claim 8, characterized in that: In the step 3, the alternating direction multiplier method is used to solve the multimodal remote sensing image spectral unmixing model based on DSM and hyperspectral combination constructed in step 2 to obtain abundance information; the specific process is: The objective function of the multimodal remote sensing image spectral unmixing model based on DSM and hyperspectral combination is as follows: Introducing the parameters R = J and A = K, the objective function of the multimodal remote sensing image spectral unmixing method based on the combination of DSM and hyperspectral is rewritten as follows: in, represents the set containing R, A, J, K; J and K represent auxiliary variables; The above formula is transformed into the Lagrangian augmented function of the multimodal remote sensing image spectral unmixing model based on the combination of DSM and hyperspectral: Among them, Λ1 and Λ2 are Lagrange multipliers, σ1 and σ2 are penalty coefficients; is the Lagrangian augmented function; Based on the idea of ​​constraining the zero point of the first-order derivative of the model to be the extreme point, the Lagrangian augmented function is optimized to obtain the reflectivity component R, auxiliary variable J, auxiliary variable K and abundance information A.

10. The multimodal remote sensing image spectral unmixing method based on digital surface model DSM and hyperspectral combination according to claim 9, characterized in that: The Lagrangian augmented function is optimized based on the idea that the zero point of the first-order derivative of the constraint model is an extreme point, and the reflectivity component R, the auxiliary group variable J, the auxiliary group variable K and the abundance information A are obtained; The specific process is: 1) Optimize the reflectivity component R: The objective function of the reflectivity component R is expressed as follows: The above formula can also be expressed as follows: Where T represents the auxiliary variable; S⊙T=Y, The equivalent expression is The solution results are as follows: T0=2W T1=R-WR T2=2H T3=R-HR Among them, T0, T1, T2, and T3 represent auxiliary variables; 2) Optimize auxiliary variables J and K: J=σ1R-Λ1 / σ1 K=sign(A-Λ2 / σ2)⊙max{0,|A-Λ2 / σ2|-1 / σ2} 3) Optimize abundance information A: The objective function of abundance information A is expressed as follows: The solution results are as follows: A=(2σ2K-E T R-Λ2I)×(2σ2-E T E) -1 。

Images