CN102682441A - Hyperspectral image super-resolution reconstruction method based on subpixel mapping - Google Patents

Abstract

The hyperspectral image super-resolution reconstruction method based on sub-pixel mapping first performs sub-pixel mapping, and uses the hybrid pixel decomposition method to obtain the proportion value of various ground objects in each pixel in the original hyperspectral remote sensing image, and then the original hyperspectral remote sensing image Each pixel in the hyperspectral remote sensing image, that is, the coarse pixel, is divided into n￡n sub-pixels, where n is a predefined resolution magnification; and then according to each known surface feature in each pixel The abundance value of each sub-pixel is randomly assigned to a certain kind of ground object, that is, the initialized ground object distribution image is obtained; finally, according to the iterative optimization criterion, a simulated annealing algorithm is used to obtain a resolution that is n times that of the original image Object distribution image; finally based on sub-pixel mapping for super-resolution reconstruction, the present invention considers the local and global distribution characteristics of objects at the same time, avoiding the unreality brought by only solving for the local characteristics of object distribution, making Hyperspectral remote sensing images can be obtained from surface object distribution images.

Description

Hyperspectral image super-resolution reconstruction method based on sub-pixel mapping

technical field

The invention belongs to the technical field of image processing, is applicable to hyperspectral remote sensing image reconstruction, and specifically relates to a hyperspectral image super-resolution reconstruction method based on sub-pixel mapping. the

Background technique

In recent years, satellite remote sensing technology has developed rapidly. The emerging hyperspectral remote sensing imaging technology can obtain surface observation data in many, very narrow and quasi-continuous spectral bands, and provide people with more abundant ground observation information. The ability to identify and analyze ground objects directly from remote sensing images has been enhanced. However, the spatial resolution of hyperspectral imaging technology is limited while obtaining high spectral resolution. Therefore, how to improve the spatial resolution of hyperspectral remote sensing images has become an urgent problem to be solved. the

For a space remote sensing platform that is at a distance R from the plane to be imaged, its spatial resolution ΔL can be expressed as

$\mathrm{<span\; class="notranslate">\; \&Delta;\; L</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; WR</span>}}{\mathrm{<span\; class="notranslate">\; f</span>}}$

Among them, W is the array element width of the CCD device, and f is the focal length of the optical system. Therefore, there are three ways to improve the spatial resolution of remote sensing images. One is to reduce the flight orbit height of the remote sensing platform, but this will shorten the life of the platform; The method of resolution is to increase the focal length and reduce the width of the array elements. However, increasing the focal length will increase the difficulty and cost of optical parts processing, resulting in large volume and heavy weight of the remote sensor, which brings difficulties to practical applications; CCD The size of the array element is limited by the technology. Due to the technical blockade of foreign countries, the aperture of the CCD we can get cannot be made small enough. To sum up, how to improve the spatial resolution of remote sensing images under the premise of keeping the distance R, focal length f and CCD aperture W constant is an urgent problem to be solved. Super-resolution image reconstruction technology is an effective way to solve this problem. This technology uses one or more low-resolution images obtained from different angles, different positions, and different sensors to reconstruct one or more high-resolution images. rate image. the

The hyperspectral imaging method is limited by the spatial resolution of the sensor and the complexity of the distribution of ground objects, so the spatial resolution of the obtained image is not high, which leads to the numerical information recorded by a single pixel mostly being the spectral characteristics of various types of ground objects. The resulting mixture is called a mixed pixel. The existence of the phenomenon of mixed pixels brings great difficulties to the quantitative interpretation of images, and also brings difficulties to improve the spatial resolution of hyperspectral images. The mixed pixel decomposition method is one of the effective ways to solve this difficulty. Its basic meaning is to extract the typical "pure" feature spectrum called endmember through the mixed pixel spectrum, and at the same time estimate the The proportion of , which is called the abundance value. For each type of surface object, its abundance value in each pixel forms an abundance map. However, this kind of mixed pixel decomposition method only gives the proportion of each surface feature in each pixel, and does not give how they are distributed in the pixel. The sub-pixel mapping method uses the abundance map obtained by the mixed pixel decomposition method to estimate the distribution of each surface feature in each pixel under certain restrictions, so that a higher resolution than the original image can be obtained. object classification images. The existing sub-pixel mapping methods are mainly divided into two categories, learning-based and spatial continuity-based. The former collects some known high-resolution ground object distribution images as a training set, and obtains rich images by reducing their resolution. degree map, and then use machine learning to train the relationship between the high-resolution surface object distribution image and the corresponding abundance map, and apply this relationship to the abundance map of the hyperspectral image to be processed, so as to obtain the required High-resolution ground object distribution images; the latter algorithm based on spatial continuity mostly proposes different models for the local continuity of the ground object distribution, and then obtains the optimal ground object distribution that satisfies this model through optimization. the

Contents of the invention

In order to overcome the deficiencies of the above-mentioned prior art, the object of the present invention is to provide a hyperspectral image super-resolution reconstruction method based on sub-pixel mapping, which can obtain a high-resolution hyperspectral remote sensing image with a higher resolution than the original image. the

In order to achieve the above object, the technical solution adopted in the present invention is:

A hyperspectral image super-resolution reconstruction method based on sub-pixel mapping, including the following steps:

Step 1, sub-pixel mapping;

Step 1.1, use the hybrid pixel decomposition method to obtain the proportion value of various ground objects in each pixel in the original hyperspectral remote sensing image, which is called the abundance value;

In step 1.2, each pixel in the original hyperspectral remote sensing image is called a coarse pixel, and each coarse pixel is divided into n￡n sub-pixels, where n is a predefined resolution magnification;

Step 1.3, according to the known abundance value of each surface feature in each pixel, each sub-pixel is randomly assigned as a certain type of feature, that is, the initialized feature distribution image is obtained;

Step 1.4, according to the iterative optimization criterion, the simulated annealing algorithm is used to iteratively optimize the initialized feature distribution image until the iteration termination condition is satisfied. The iterative optimization criterion is based on the local continuity of the feature distribution and the global distribution of the feature similarity, where

The local continuity of the distribution of ground features means that the same type of ground features are distributed closer than different types of ground features within a local range;

The global similarity of feature distribution refers to the similarity of feature distribution in rough pixels with similar proportions of features in the global scope, so that any pixel can be represented by a pixel similar to its feature distribution ;

The following objective function serves as the stated iterative optimization criterion:

$\mathrm{<span\; class="notranslate">\; min</span>}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; c</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&eta;</span>}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

指属于地物类别i的亚像元的横坐标和纵坐标，μ_ix，μ_iy分别指 

的平均值，η是正则化因子，q_j是与像元p具有相似性的像元，其选择依据是 

f(p)是由像元p及其邻域的丰度值组成的特征向量，th是一预设的阈值，如果某一像元q_j的特征向量属于前J(j≤J)个与f(p)相似的向量，那么该像元q_j就被认为是像元p的相似像元，因此在像元p内的亚像元分布可以由像元q_j内亚像元分布的线性组合来表示，即 α_j是er_j的归一化倒数。此目标函数第一项 

代表地物分布的局部连续性，第二项 

代表地物分布的全局相似性；  Among them, t is used to index the pixel to be processed and the sub-pixel belonging to feature i in its neighborhood, c refers to the number of feature categories, $x_t^i=\left[\begin{array}{c}{x}_t^i\\ {\mathrm{the\; y}}_t^i\end{array}\right],$ ${\mathrm{\&mu;}}_i=\left[\begin{array}{c}{\mathrm{\&mu;}}_{\mathrm{ix}}\\ {\mathrm{\&mu;}}_{\mathrm{iy}}\end{array}\right],$

refers to the abscissa and ordinate of the sub-pixel belonging to the feature category i, μ _ix and μ _iy refer to

, η is the regularization factor, q _j is the pixel with similarity to the pixel p, and its selection is based on

f(p) is a feature vector composed of the abundance values of pixel p and its neighbors, th is a preset threshold, if the feature vector of a certain pixel q _j belongs to the first J (j≤J) and f(p) similar vector, then the pixel q _j is considered as a similar pixel of pixel p, so the sub-pixel distribution in the pixel p can be determined by the linearity of the sub-pixel distribution in the pixel q _j combined to represent α _j is the normalized reciprocal of er _j . The first term of this objective function

Represents the local continuity of feature distribution, the second term

Represents the global similarity of feature distribution;

Thus, a feature distribution image whose resolution is n times that of the original image is obtained;

Step 2, super-resolution reconstruction method based on sub-pixel mapping;

Divide the object distribution image obtained in step 1 into a series of sub-image blocks whose size is m￡m; Multiply by the pure end-member surface object spectrum obtained in the hybrid pixel decomposition method, and linearly add to obtain the spectral curve of the sub-image block, so as to obtain a hyperspectral remote sensing image with a resolution n=m times that of the original image. the

The mixed pixel decomposition method in the step 1.1 refers to obtaining the proportion of each typical ground object that makes up the mixed pixel, where the mixed pixel refers to the pixel that contains different types of ground objects in the obtained hyperspectral image. the

值减小，所述终止条件是目标函数值不再下降或达到了预定的迭代次数。  The iterative optimization process in the step 1.4 is to change the distribution of ground objects through continuous iteration, wherein each iteration process will make the objective function

The value decreases, and the termination condition is that the value of the objective function no longer decreases or reaches a predetermined number of iterations.

Since the present invention aims to increase the resolution of the hyperspectral image, the image whose resolution is to be increased can be called a low-resolution image. the

Compared with prior art, the advantage of the present invention is:

1) The present invention adopts the minimum criterion of making the intra-class scatter between the same features on the local continuity model of feature distribution;

2) The present invention considers the global similarity of the feature distribution while considering the local continuity of the feature distribution, and avoids the unreality brought when only solving for the local characteristics of the feature distribution;

3) The present invention proposes a super-resolution reconstruction method based on sub-pixel mapping, which can obtain hyperspectral remote sensing images from ground object distribution images. the

Description of drawings

FIG. 1 is a schematic diagram of the division method of coarse pixels in the present invention. the

Figure 2 is a schematic diagram of the results obtained by the sub-pixel mapping method. the

Detailed ways

The present invention will be described in further detail below in conjunction with the examples. the

A hyperspectral image super-resolution reconstruction method based on sub-pixel mapping, including the following steps:

Step 1, sub-pixel mapping;

Step 1.1, use the hybrid pixel decomposition method to obtain the proportion value of various ground objects in each pixel in the original hyperspectral remote sensing image, which is called the abundance value;

The mixed pixel decomposition method is a commonly used processing technique in this field, and a more commonly used mixed pixel decomposition method is given here. This method first selects the appropriate pure ground objects that make up the hyperspectral image, called endmembers, and then extracts the spectrum of the endmembers from the image or the spectral library, and then uses the linear mixed model as the spectral mixed model, that is, r = ∑ _i a _i s _i +w 0≤a _i ≤1, ∑ _i a _i =1, where r is the spectrum of each pixel in the hyperspectral image, s _i is the spectrum of each end member, and a _i is the spectrum of each end member The proportion value of the element in the pixel, that is, the abundance value, 0≤a _i ≤1 means that the proportion value of each type of surface object should be between 0 and 1, ∑ _i a _i =1 means that in the pixel The sum of the scale values of all ground objects should be 1, and w is the error. Therefore, according to the principle of minimizing the error w, the abundance value of each object in each pixel can be obtained by solving the above equation. In this embodiment, a simple simulated image is taken as an example to illustrate the implementation of the present invention. In this example, two types of ground objects are selected, namely the target "0" and the background "1". The size of the image to be processed is 9 pixels × 9 pixels. Since it is not a real hyperspectral remote sensing image and the mixed pixel decomposition method does not belong to the content of the present invention, here we set two types of ground objects in each image The abundance values in the element are shown in Table 1 and Table 2.

Table 1, the abundance value of ground object "0" in each pixel

the 1 2 3 4 5 6 7 8 9 1 0 0 0 0 0 0 0 0 0 2 0 0 0 0 0 0 0 0 0 3 0 0 0 0.0625 0.2500 0.0625 0 0 0

[0039] 4 0 0 0.1250 0.9375 1 0.9375 0.1250 0 0 5 0 0 0.5000 1 1 1 0.5000 0 0 6 0 0 0.3750 1 1 1 0.3750 0 0 7 0 0 0 0.5000 0.7500 0.5000 0 0 0 8 0 0 0 0 0 0 0 0 0 9 0 0 0 0 0 0 0 0 0

Table 2, the abundance value of ground object "1" in each pixel

the 1 2 3 4 5 6 7 8 9 1 1 1 1 1 1 1 1 1 1 2 1 1 1 1 1 1 1 1 1 3 1 1 1 0.9375 0.7500 0.9375 1 1 1 4 1 1 0.8750 0.0625 0 0.0625 0.8750 1 1 5 1 1 0.5000 0 0 0 0.5000 1 1 6 1 1 0.6250 0 0 0 0.6250 1 1 7 1 1 1 0.5000 0.2500 0.5000 1 1 1 8 1 1 1 1 1 1 1 1 1 9 1 1 1 1 1 1 1 1 1

In step 1.2, each pixel in the original hyperspectral remote sensing image is called a coarse pixel, and each coarse pixel is divided into n￡n sub-pixels, where n is a predefined resolution magnification;

In the present invention, each pixel in the hyperspectral image to be processed is called a coarse pixel, that is, the 9×9 pixels in Table 1 are all called coarse pixels. In this embodiment, the resolution magnification factor is set to 4, that is, n=4, and then each coarse pixel is divided into 4×4 sub-pixels on average, as shown in FIG. 1 . the

Step 1.3, according to the known abundance value of each surface feature in each pixel, each sub-pixel is randomly assigned as a certain type of feature, that is, the initialized feature distribution image is obtained;

Take the pixel with an abundance value of 0.25 of the target "0" in the third row and fifth column in Table 1 as an example. After dividing each coarse pixel into 4×4 sub-pixels, there should be 4×4×0.25=4 sub-pixels belonging to the target “0”, similarly, there are 12 sub-pixels belonging to the background “1” ". Then randomly pick 4 of these 4×4 sub-pixels and assign them as the target “0”, and assign the remaining 12 sub-pixels as the background “1”. the

Step 1.4, according to the iterative optimization criterion, the simulated annealing algorithm is used to iteratively optimize the initialized feature distribution image until the iteration termination condition is satisfied. The iterative optimization criterion is based on the local continuity of the feature distribution and the global distribution of the feature similarity, where

The local continuity of the distribution of ground features means that the same type of ground features are distributed closer than different types of ground features within a local range;

The global similarity of feature distribution refers to the similarity of feature distribution in rough pixels with similar proportions of features in the global scope, so that any pixel can be represented by a pixel similar to its feature distribution ;

The following objective function serves as the stated iterative optimization criterion:

$\mathrm{<span\; class="notranslate">\; min</span>}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; c</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&eta;</span>}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

指属于地物类别i的亚像元的横坐标和纵坐标，μ_ix，μ_iy分别指 

的平均值，η是正则化因子，q_j是与像元p具有相似性的像元，其选择依据是 

f(p)是由像元p及其邻域的丰度值组成的特征向量，th是一预设的阈值，如果某一像元q_j的特征向量属于前J(j≤J)个与f(p)相似的向量，那么该像元q_j就被认为是像元p的相似像元，因此在像元p内的亚像元分布可以由像元q_j内亚像元分布的线性组合来表示，即 α_j是er_j的归一化倒数。此目标函数第一项 

代表地物分布的局部连续性，第二项 

代表地物分布的全局相似性；  Among them, t is used to index the pixel to be processed and the sub-pixel belonging to feature i in its neighborhood, c refers to the number of feature categories, $x_t^i=\left[\begin{array}{c}{x}_t^i\\ {\mathrm{the\; y}}_t^i\end{array}\right],$ ${\mathrm{\&mu;}}_i=\left[\begin{array}{c}{\mathrm{\&mu;}}_{\mathrm{ix}}\\ {\mathrm{\&mu;}}_{\mathrm{iy}}\end{array}\right],$

refers to the abscissa and ordinate of the sub-pixel belonging to the feature category i, μ _ix and μ _iy refer to

, η is the regularization factor, q _j is the pixel with similarity to the pixel p, and its selection is based on

f(p) is a feature vector composed of the abundance values of pixel p and its neighborhood, th is a preset threshold, if the feature vector of a certain pixel q _j belongs to the first J (j≤J) and f(p) similar vector, then the pixel q _j is considered as a similar pixel of pixel p, so the sub-pixel distribution in the pixel p can be determined by the linearity of the sub-pixel distribution in the pixel q _j combined to represent α _j is the normalized reciprocal of er _j . The first term of this objective function

Represents the local continuity of feature distribution, the second term

Represents the global similarity of feature distribution;

The simulated annealing algorithm is a general optimization algorithm, which is used to find the optimal solution of the objective function. In this embodiment, the optimization calculation is performed on the coarse pixels one by one, that is, firstly, the value of the objective function in the initial condition is calculated, and then the classification of the ground features of each sub-pixel in the coarse pixel is changed according to the simulated annealing algorithm, and so on. Repeat this iterative step until the value of the objective function no longer decreases or reaches the predetermined number of iterations. the

As a result, an image of the distribution of ground objects with a resolution four times that of the original image is obtained, as shown in Figure 2, in which the target "0" is represented by black, and the background 1 is represented by white. the

Step 2,

其中b_i为比例值，s_i为混合像元分解方法中得到的端元光谱。由于本实施例以模拟图像为例，所以没有给出最后得到的高分辨率的遥感图像。  In this embodiment, every 2×2 pixels in the high-resolution ground object distribution image obtained in step 1 are regarded as a whole, that is, m=2, thus forming a series of sub-image blocks, and for each sub-image The ratio of the two types of ground objects contained in the image block can be obtained again by simple calculation, and then the ratio value is multiplied by the pure ground object obtained in the mixed pixel decomposition method, that is, the spectrum of the end member, and linearly added to obtain the The spectral curve of the sub-image block, namely

where _bi is the ratio value and _si is the endmember spectrum obtained in the hybrid pixel decomposition method. Since this embodiment takes a simulated image as an example, the finally obtained high-resolution remote sensing image is not given.

Claims (3)

1. The hyperspectral image super-resolution reconstruction method based on sub-pixel mapping, is characterized in that, comprises the following steps: Step 1, sub-pixel mapping; Step 1.1, using the hybrid pixel decomposition method to obtain the proportion value of various ground objects in each pixel in the original hyperspectral remote sensing image, which is called the abundance value; Step 1.2, each pixel in the original hyperspectral remote sensing image is called a coarse pixel, and each coarse pixel is divided into n￡n sub-pixels, where n is a predefined resolution magnification; Step 1.3, according to the known abundance value of each feature in each pixel, each sub-pixel is randomly assigned as a certain kind of feature, that is, the initialized feature distribution image is obtained; Step 1.4, according to the iterative optimization criterion, the simulated annealing algorithm is used to iteratively optimize the initialized feature distribution image until the iteration termination condition is satisfied. The iterative optimization criterion is based on the local continuity of the feature distribution and the global distribution of the feature similarity, where The local continuity of the distribution of ground features means that the same type of ground features are distributed closer than different types of ground features in a local area; The global similarity of feature distribution refers to the similarity of feature distribution in rough pixels with similar proportions of features in the global scope, so that any pixel can be represented by a pixel similar to its feature distribution ; The following objective function serves as the stated iterative optimization criterion:

Among them, t is used to index the pixel to be processed and the sub-pixel belonging to feature i in its neighborhood, c refers to the number of feature categories,

refers to the abscissa and ordinate of the sub-pixel belonging to the feature category i, μ _ix and μ _iy refer to

, η is the regularization factor, q _j is the pixel with similarity to the pixel p, and its selection is based on

f(p) is a eigenvector composed of the abundance values of pixel p and its neighbors, th is a preset threshold, if the eigenvector of a certain pixel q _j belongs to the first J (j≤J) and f(p) similar vector, then the pixel q _j is considered as a similar pixel of pixel p, so the sub-pixel distribution in the pixel p can be determined by the linearity of the sub-pixel distribution in the pixel q _j combined to represent

α _j is the normalized reciprocal of er _j . The first term of this objective function Represents the local continuity of feature distribution, the second term

Represents the global similarity of feature distribution; Thus, a feature distribution image whose resolution is n times that of the original image is obtained; Step 2, super-resolution reconstruction method based on sub-pixel mapping; Divide the object distribution image obtained in step 1 into a series of sub-image blocks whose size is m￡m; Multiply by the pure end-member surface object spectrum obtained in the hybrid pixel decomposition method, and linearly add to obtain the spectral curve of the sub-image block, so as to obtain a hyperspectral remote sensing image with a resolution n=m times that of the original image. 2. The hyperspectral image super-resolution reconstruction method based on sub-pixel mapping according to claim 1, characterized in that, the mixed pixel decomposition method in the step 1.1 refers to obtaining each typical ground object that forms the mixed pixel The ratio of , where the mixed pixel refers to the pixel that contains different types of ground objects in the obtained hyperspectral image. 3. according to the hyperspectral image super-resolution reconstruction method based on sub-pixel mapping of claim 1, it is characterized in that, iterative optimization process is to change the distribution of feature by continuous iteration in the described step 1.4, wherein each iteration process will make the objective function

The value decreases, and the termination condition is that the value of the objective function no longer decreases or reaches a predetermined number of iterations.

Images