Hyperspectral remote sensing image weak end member unmixing method suitable for methane detection and quantification

Abstract

The invention discloses a hyperspectral remote sensing image weak end member unmixing method suitable for methane detection and quantification, and particularly relates to the field of computer vision technologies. Acquiring a hyperspectral remote sensing image, and determining a background end member matrix, a gas end member matrix and an abundance matrix of each pixel of the hyperspectral remote sensing image; reconstructing the hyperspectral remote sensing image; determining a reconstruction error of the hyperspectral remote sensing image and the reconstructed image; determining the degradation amount of the priori knowledge according to the priori knowledge of the gas and the gas matrix; performing sparsification on the abundance matrix to obtain a sparse abundance matrix; according to the reconstruction error, the degradation amount of the priori knowledge and the sparse abundance matrix, an objective function is established, and constraint conditions of a gas matrix, a background end member matrix and the abundance matrix are established; and obtaining an optimal gas end member matrix and an optimal abundance matrix according to the objective function. Based on the method, the detection precision of weak signals can be improved.

Description

Technical Field

[0001] The present application relates to the field of computer vision technology, and in particular to a weak endmember unmixing method for hyperspectral remote sensing images suitable for methane detection and quantification. Background Art

[0002] The methane detection process based on hyperspectral remote sensing imaging includes data acquisition, data preprocessing, gas plume detection, and point source flux velocity inversion. Gas plume detection and point source flux velocity inversion are key components, primarily due to image spatial-spectral aliasing and weak target signal radiation intensity. Current methane detection algorithms based on spectral imaging typically rely on matching the target's prior spectrum with the spectral characteristics of image pixels or using anomaly detection methods to identify gas targets as outliers. These algorithms can only accurately infer the concentration of strong plumes, while the concentration of weak plumes is often underestimated. In atmospheric scenarios, target spectra are significantly distorted during the top-down acquisition process, making current methods ineffective in identifying targets. According to the radiative transfer model, a gas pixel spectrum is a linear combination of the background radiation spectrum and the gas radiation spectrum. Hyperspectral mixed pixel decomposition technology is an effective tool for methane detection and quantitative inversion. In this process, the gas absorption spectrum, background radiation spectrum, and atmospheric absorption spectrum are demixed as endmembers, and the point source flux velocity is inverted using the abundance of the gas absorption spectrum.

[0003] Due to the limited spatial resolution of imaging sensors, images are composed of mixed pixels that are linear combinations of multiple endmembers. Endmembers are generally considered to represent the primary features in the image. Discussion and research on weak signals, caused by low-intensity mixed components and limited feature coverage, are typically not part of the traditional hyperspectral analysis paradigm. However, these weak-signal features often possess significant detection value. Non-negative matrix factorization (NMF)-based methods approximate hyperspectral remote sensing images by decomposing them into the product of an endmember matrix and an abundance matrix by alternately optimizing the target variable using a multiplicative update rule. However, when weak signals are present within features, direct NMF interpretation of spectral images suffers from insufficient accuracy. Summary of the Invention

[0004] The main purpose of this application is to provide a weak endmember unmixing method for hyperspectral remote sensing images suitable for methane detection and quantification, aiming to solve the problem of low accuracy when using the NMF method for spectral image interpretation.

[0005] To achieve the above objectives, the present application provides a method for weak endmember unmixing of hyperspectral remote sensing images suitable for methane detection and quantification, comprising: acquiring a hyperspectral remote sensing image, determining the initial endmember and corresponding abundance of each pixel of the hyperspectral remote sensing image, determining the background endmember and gas endmember based on the initial endmember, and constructing a background endmember matrix, a gas endmember matrix and an abundance matrix; reconstructing the hyperspectral remote sensing image using the gas endmember matrix, the background endmember matrix, and the abundance matrix to obtain a reconstructed image; and determining the reconstruction error between the hyperspectral remote sensing image and the reconstructed image. ; Obtain the absorption spectrum curve of the gas and determine its prior knowledge based on the absorption spectrum curve of the gas; Determine the degradation amount of the prior knowledge based on the prior knowledge of the gas and the gas matrix; Sparsify the abundance matrix to obtain a sparse abundance matrix; Establish an objective function based on the reconstruction error, the degradation amount of the prior knowledge and the sparse abundance matrix, and establish constraints for the gas matrix, background end member matrix and abundance matrix; Use the objective function to optimize the gas end member matrix, background end member matrix and abundance matrix to obtain the optimal gas end member matrix and abundance matrix.

[0006] Optionally, the method further includes: determining a gas plume coverage area according to an optimal abundance matrix; and determining a gas point source flux velocity according to the gas plume coverage area and the wind speed.

[0007] Optionally, the reconstruction error is determined according to the difference between the hyperspectral remote sensing image and the reconstructed image.

[0008] Optionally, the degradation amount of the prior knowledge is determined according to the difference between the prior knowledge of the gas and the gas matrix.

[0009] Optionally, the abundance matrix is ​​sparsely distributed to obtain a sparse abundance matrix, including: sparsely distributed to obtain a sparse abundance matrix by using the wavelength of the absorption sensitive band of the gas and a reweighted constraint matrix.

[0010] Optionally, the objective function is:

[0011] Where, , represents the reweighted constraint matrix, represents the background endmember matrix, represents the gas end member matrix, represents the abundance matrix, Represents prior knowledge of the gas.

[0012] Optionally, the prior knowledge is the absorption spectrum of the gas.

[0013] Optionally, background endmembers and gas endmembers are determined based on the initial endmembers, and a background endmember matrix, a gas endmember matrix and an abundance matrix are constructed, including: decomposing each initial endmember into a background endmember and a gas endmember; combining the background endmembers of all pixels to obtain a background endmember matrix; combining the gas endmembers of all pixels to obtain a gas endmember matrix; and combining the abundances of all pixels to obtain an abundance matrix.

[0014] Optionally, before determining the initial endmember and the corresponding abundance of each pixel of the hyperspectral remote sensing image, the method further includes performing radiation correction and atmospheric correction on the hyperspectral remote sensing image.

[0015] Optionally, the gas comprises methane, sulfur dioxide, nitrogen dioxide or hydrogen sulfide.

[0016] Compared with the prior art, the present invention has the following advantages: The invention discloses a weak endmember unmixing method for hyperspectral remote sensing images suitable for methane detection and quantification. The method comprises the following steps: determining initial endmembers based on a hyperspectral remote sensing image, decomposing the initial endmembers into a gas endmember matrix and a background endmember matrix, obtaining a reconstructed image through the gas endmember matrix, the background endmember matrix and the abundance matrix, determining a reconstruction error based on the reconstructed image, and optimizing the background endmembers by minimizing the reconstruction error; determining the degradation amount of the prior knowledge based on the prior knowledge of the gas, optimizing the gas endmembers by constraining the degradation of the prior knowledge, thereby achieving differential optimization of different endmembers, and simultaneously adding a heavily weighted sparse regularization term to achieve local enhancement of gas detection and improve the detection accuracy of weak signals; and adopting a gradient descent learning algorithm to solve the objective function and improve the speed of mixed endmember decomposition. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Schematic diagram of the process of the weak endmember unmixing method of hyperspectral remote sensing images applicable to methane detection and quantification in this application; Figure 2 This is the unmixing result diagram of Example 1.

[0018] The realization of the objectives, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this application.

[0020] The embodiments of the present invention provide a method for unmixing weak endmembers of hyperspectral remote sensing images suitable for methane detection and quantification. Figure 1 As shown, the specific steps include: Step S1, obtain a hyperspectral remote sensing image, determine the initial endmember and corresponding abundance of each pixel of the hyperspectral remote sensing image, determine the background endmember and gas endmember based on the initial endmember, and construct a background endmember matrix, a gas endmember matrix and an abundance matrix; wherein the gas endmember is the endmember of the weak signal.

[0021] Specifically, to ensure the clarity of hyperspectral remote sensing images, radiometric and atmospheric corrections are pre-processed. Next, the orthogonal subspace projection (OSP) algorithm is used to estimate the initial endmembers and corresponding abundances for each pixel in the hyperspectral remote sensing image. The initial endmembers are mixed endmembers, formed by the aliasing of background (major ground feature) endmembers and gas (weak signal) endmembers. Each initial endmember is then decomposed into background endmembers and gas endmembers using non-negative matrix factorization (NMF). The background endmembers of all pixels are combined to obtain a background endmember matrix; the gas endmembers of all pixels are combined to obtain a gas endmember matrix; and the abundances of all pixels are combined to obtain an abundance matrix.

[0022] Step S2, reconstructing the hyperspectral remote sensing image using the gas endmember matrix, the background endmember matrix, and the abundance matrix to obtain a reconstructed image; determining a reconstruction error between the hyperspectral remote sensing image and the reconstructed image; wherein the reconstruction error is determined based on the difference between the hyperspectral remote sensing image and the reconstructed image.

[0023] It is worth noting that the reconstruction error is used to establish the first objective function as an unsupervised module. The first objective function is expressed as follows:

[0024] In the formula, let Indicates inclusion bands and Hyperspectral remote sensing images with pixels, represents the background endmember matrix, represents the gas end member matrix, represents the abundance matrix, express where T represents the transposed matrix, p represents the number of endmembers in the background endmember matrix, and q represents the number of endmembers in the gas endmember matrix. Both the endmember matrix and the abundance matrix must satisfy the nonnegativity constraint, and the abundance matrix must also satisfy the sum-to-one constraint. The OSP algorithm achieves the best match for the target object while eliminating background information.

[0025] Since the background endmembers and gas endmembers obtained by decomposing the initial endmembers are not the optimal unmixing results, they need to be optimized. When prior knowledge is introduced to detect and separate the corresponding weak signals, it may degenerate during the optimization task. Therefore, the degradation of the prior information must be constrained. The specific constraint method is as follows.

[0026] Step S3: Obtain the gas absorption spectrum curve, determine the gas prior knowledge based on the gas absorption spectrum curve, and determine the degradation of the prior knowledge based on the gas prior knowledge and the gas matrix. The degradation of the prior knowledge is determined by the difference between the gas prior knowledge and the gas matrix. The prior knowledge is the gas absorption spectrum, provided by the HITRAN spectral library. Using the degradation of the prior knowledge, a second objective function is constructed as a supervised module, expressed as:

[0027] Where, The prior knowledge of the gas is considered as the true value and will not be optimized in the supervised module. The priors of the gas signals detected in the supervised module are updated and optimized in the model. The unsupervised module and the supervised module work together in methane detection based on hyperspectral remote sensing images.

[0028] Step S6, performing sparse processing on the abundance matrix to obtain a sparse abundance matrix; Specifically, the abundance matrix is ​​sparsified by the wavelength of the gas's absorption sensitive band (specific band) and the reweighted constraint matrix to obtain a sparse abundance matrix, which can locally enhance the gas abundance.

[0029] Step S7: Based on the reconstruction error, the degradation amount of prior knowledge, and the sparse abundance matrix, an objective function is established, and constraints are established for the gas matrix, background endmember matrix, and abundance matrix. The objective function is used to optimize the gas endmember matrix, background endmember matrix, and abundance matrix to obtain the optimal gas endmember matrix and abundance matrix. The objective function obtained by combining the first objective function with the second objective function and the sparse abundance matrix is ​​expressed as:

[0030] Where, , Represents the reweighted constraint matrix, which is used to reweight the non-zero elements of the abundance matrix. represents the background endmember matrix, E n represents the initial endmember matrix, represents the gas end member matrix, represents the abundance matrix, Represents the prior knowledge of the gas, parameter λ Used to balance the fidelity term and the regularization term.

[0031] For example, using a gradient descent learning algorithm to solve the objective function yields the optimal gas endmember matrix and abundance matrix. Furthermore, the initial point for the gradient descent is chosen to be closest to the global optimum of the objective function in the hyperspectral remote sensing image. This helps avoid local minima and ensures that the decomposition results more accurately approximate the true endmember signal values.

[0032] In this embodiment, an unsupervised module obtains initial endmembers and endmembers and abundances augmented with prior knowledge to optimize background endmembers. A supervised module minimizes degradation of prior knowledge and optimizes gas endmembers, thereby achieving differentiated optimization for different endmembers. Furthermore, a heavily weighted sparse regularization term is added to achieve local enhancement of gas detection. To accelerate the decomposition of mixed endmembers, a gradient descent learning algorithm is employed to solve this optimization problem, enabling accurate and efficient estimation of the distribution of weak gas signals.

[0033] Step S8: Determine the gas plume coverage area based on the optimal abundance matrix; determine the gas point source flux velocity based on the gas plume coverage area and wind speed. A pseudo-color representation of the concentration is performed based on the locally enhanced gas abundance. This can show the gradual decrease in gas concentration from the emission point source to the plume edge, with color changes indicating the decrease. The gas detection and emission point source flux velocity inversion results are then output.

[0034] Specifically, the IME method is used to invert the methane point source flux velocity. The IME method relies on the characteristic size of the methane plume and effectively reduces the influence of wind speed by linking the plume mass and the point source flux velocity. The formula for the gas point source flux velocity is:

[0035] Where, The parameter takes into account Avogadro's law and the number of pixels in the methane plume. Indicates wind speed, based on the effective simulation of the measurable 10m wind. is the square root of the area covered by the methane plume, determined from the abundance matrix.

[0036] Example To verify the effectiveness of the method of the present invention, experiments were conducted on the AVIRIS-NG dataset, and the results are as follows: Figure 2 As shown. First, the original hyperspectral remote sensing image (see Figure 2 a) Obtain the optimal methane endmember matrix and abundance matrix through steps S1-S7, determine the methane plume based on the abundance matrix, and perform pseudo-color visualization (i.e., pseudo-color representation of concentration) at a specific wavelength (2360 nm), see Figure 2b, and then the integrated quality enhancement (IME) method was used to estimate the methane emission point source flux velocity. Figure 2 c, Figure 2 The yellow box area in ac is the methane plume.

[0037] The above are only preferred embodiments of the present application and do not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for weak endmember unmixing of hyperspectral remote sensing images suitable for methane detection and quantification, characterized in that: include: Acquire a hyperspectral remote sensing image, determine an initial endmember and a corresponding abundance of each pixel of the hyperspectral remote sensing image, determine a background endmember and a gas endmember based on the initial endmember, and construct a background endmember matrix, a gas endmember matrix, and an abundance matrix; Reconstructing a hyperspectral remote sensing image using the gas endmember matrix, the background endmember matrix, and the abundance matrix to obtain a reconstructed image; Determining a reconstruction error between the hyperspectral remote sensing image and the reconstructed image; Obtaining an absorption spectrum curve of the gas, and determining prior knowledge thereof based on the absorption spectrum curve of the gas; Determine the degradation amount of the prior knowledge according to the prior knowledge of the gas and the gas matrix; performing sparse processing on the abundance matrix to obtain a sparse abundance matrix; Establishing an objective function based on the reconstruction error, the degradation amount of prior knowledge, and the sparsified abundance matrix, and establishing constraints on the gas matrix, the background endmember matrix, and the abundance matrix; The objective function is used to optimize the gas endmember matrix, the background endmember matrix and the abundance matrix to obtain the optimal gas endmember matrix and the abundance matrix.

2. The method for weak endmember unmixing of hyperspectral remote sensing images suitable for methane detection and quantification according to claim 1, characterized in that: Also includes: Determine the gas plume coverage area based on the optimal abundance matrix; The gas point source flux velocity is determined according to the gas plume coverage area and wind speed.

3. The method for weak endmember unmixing of hyperspectral remote sensing images suitable for methane detection and quantification according to claim 1, characterized in that: The reconstruction error is determined according to the difference between the hyperspectral remote sensing image and the reconstructed image.

4. The method for weak endmember unmixing of hyperspectral remote sensing images suitable for methane detection and quantification according to claim 1, characterized in that: The degradation amount of the prior knowledge is determined according to the difference between the prior knowledge of the gas and the gas matrix.

5. The method for weak endmember unmixing of hyperspectral remote sensing images suitable for methane detection and quantification according to claim 1, characterized in that: The abundance matrix is ​​sparsely processed to obtain a sparse abundance matrix, comprising: The abundance matrix is ​​sparsed by the wavelength of the gas's absorption sensitive band and the reweighted constraint matrix to obtain a sparse abundance matrix.

6. The method for weak endmember unmixing of hyperspectral remote sensing images suitable for methane detection and quantification according to claim 1, characterized in that: The objective function is: Where, , represents the reweighted constraint matrix, represents the background endmember matrix, represents the gas end member matrix, represents the abundance matrix, Represents prior knowledge of the gas.

7. The method for weak endmember unmixing of hyperspectral remote sensing images suitable for methane detection and quantification according to claim 1, characterized in that: The prior knowledge is the absorption spectrum of the gas.

8. The method for weak endmember unmixing of hyperspectral remote sensing images suitable for methane detection and quantification according to claim 1, characterized in that: Determine the background endmembers and gas endmembers according to the initial endmembers, and construct a background endmember matrix, a gas endmember matrix and an abundance matrix, including: Decomposing each of the initial end members into a background end member and a gas end member; Combining the background endmembers of all pixels to obtain the background endmember matrix; The gas endmembers of all pixels are combined to obtain the gas endmember matrix; The abundance of all pixels is combined to obtain the abundance matrix.

9. The method for weak endmember unmixing of hyperspectral remote sensing images suitable for methane detection and quantification according to claim 1, characterized in that: Before determining the initial endmember and the corresponding abundance of each pixel of the hyperspectral remote sensing image, the method further includes performing radiation correction and atmospheric correction on the hyperspectral remote sensing image.

10. The method for weak endmember unmixing of hyperspectral remote sensing images suitable for methane detection and quantification according to claim 1, characterized in that: The gas includes methane, sulfur dioxide, nitrogen dioxide or hydrogen sulfide.

Images