A hyperspectral data processing method

Abstract

The invention provides a hyperspectral data processing method. Firstly, the number of endmembers is determined, then the MLIG algorithm is used to extract the image endmembers, and finally the MVC‑MRF algorithm is used to optimize the endmembers and spectrum unmixing. The present invention simplifies the constrained linear model, the number of endmembers is determined and the algorithm for extracting endmembers is strictly deduced through mathematical formulas, and the theory is rigorous; The algorithm is simple in calculation and high in efficiency, which can meet the application of large data sets; finally, the MVC-MRF algorithm is constructed to optimize the extracted endmembers and unmix the spectrum, and can extract all linearly independent image endmembers, which are representative and solve The mixed error is basically 0, and the classification accuracy is high.

Description

technical field [0001] The invention relates to a hyperspectral data processing method, belonging to the field of hyperspectral data information extraction. Background technique [0002] The spectral resolution of hyperspectral data is high, and many unknown signal features that cannot be determined by visual and prior information can be found. But at the same time, the spatial resolution is relatively low, and mixed pixels are common in images. The linear spectral unmixing model can give satisfactory results in hyperspectral data analysis. The accuracy of spectral unmixing mainly depends on the accuracy of endmember extraction. According to whether there are pure pixels in the assumed image data, the endmember extraction algorithms can be divided into two categories: endmember identification algorithm and endmember generation algorithm. The endmember identification algorithm based on the linear mixed model effectively utilizes the strong correspondence between the linear m...

AI Technical Summary

Problems solved by technology

This type of algorithm is mainly divided into three categories based on the geometric characteristics of the endmembers and the spectral unmixing error: one is based on the geometric characteristics of the endmembers, that is, the mixed pixel is located in the simplex with the endmember as the vertex, However, this method is a non-convex programming problem, and it is easy to fall into local optimum, and it depends a lot on the initial value.

The second is to use the unmixing error to constrain. It is necessary to pre-set the number of endmembers or the error threshold. Different parameter settings have a great impact on the results.

The third is through the common constraints of the unmixing error and the geometric characteristics of the endmembers. The non-negative matrix decomposition under the constraints is more consistent with the spectral unmixing model in the physical sense, and has obtained certain research and application in spectral unmixing. However, the dependence on the initial value is very large, and it needs to be initialized by other endmember extraction algorithms, and the estimation of the number of endmembers is insufficient, which may easily lead to deviation

The existing hyperspectral data analysis methods all assume that the number of endmembers is known or estimated by other methods, and then use the geometric characteristics of the endmembers or unmixing errors as prior information to extract the endmembers, and finally calculate the endmembers through the mixture model. Component abundance, lack of theoretical link between steps

Images