Semi-supervised weak signal unmixing method for hyperspectral images

Abstract

The present invention relates to a semi-supervised weak signal unmixing method for hyperspectral images, and belongs to the field of image processing. The method comprises the following steps: S1: obtaining a remote sensing observation image of a designated ground range from a satellite, and completing image radiation correction and atmospheric correction preprocessing; S2: using a combination of rough analysis of image spectrum peaks and valleys and field survey to estimate weak signal objects that may exist in the area observed by the satellite, including water bodies polluted by algae and ground polluted by industrial wastewater; measuring the standard spectrum represented by the weak signal in the laboratory; S3: initializing the pre-processed remote sensing image to obtain the initial spectrum of typical objects in the image and the corresponding abundance distribution range; S4: using a model including an unsupervised module, a supervised module and a sparse regularization term to estimate the spectrum of objects in the image and the corresponding distribution of objects; and obtaining the distribution range and mixed components of weak signal objects at the same time.

Description

Technical Field

[0001] The invention belongs to the field of image processing and relates to a hyperspectral image semi-supervised weak signal unmixing method. Background Art

[0002] Hyperspectral remote sensing technology, also known as imaging spectroscopy, records the radiation reflected or emitted by the surface in dozens or hundreds of continuously spaced narrow spectral bands at each pixel, and has been well applied to heavy metal pollution and precision agriculture. Due to the limited spatial resolution of the sensor, the image is prone to contain mixed pixels that are linear combinations of multiple end members. However, end members are usually considered to represent the main features in the image, and when there are weak signals of low intensity in the features due to low mixed components and limited feature coverage, the discussion and research of weak signals is usually not in the paradigm of traditional hyperspectral analysis, but this type of weak signal features often have greater detection value.

[0003] In most cases, weak signals represent some small but useful information, such as pollution or minerals. A typical example of water quality monitoring is the detection of algal blooms, which involves indirect measurement of chlorophyll concentrations. However, the spectrum of polluted water bodies with low chlorophyll concentrations, such as less than 2 mg / m3, has high reflectance only in the blue wavelength range (400–500 nm). Another example is the normalized difference vegetation index (NDVI), which can be used to estimate the green space density of an area. It is a dimensionless index that describes the difference between the visible and near-infrared reflectance of vegetation cover. As the vegetation cover increases, the NDVI tends to saturate. That is, if the vegetation cover and low chlorophyll concentration are not saturated compared to the background cover, the interpretation accuracy of the observed ground cannot meet the requirements. In these cases, due to the weak signals in practical applications, traditional decomposition strategies often encounter bottlenecks. At present, most researchers pay more attention to traditional hyperspectral interpretation methods, and there is no clear analysis paradigm for the processing of weak signals of low-intensity ground objects.

[0004] Non-negative matrix factorization (NMF)-based methods approximately decompose hyperspectral data into the product of the endmember matrix and the abundance matrix by alternately optimizing the target variable using a multiplication update rule in hyperspectral unmixing. However, when there are weak signals with low intensity in the local objects, the accuracy of image interpretation using the NMF method is insufficient. Since some targets have subtle mixed components, the difficulty in solving such problems lies in the difficulty of initialization and optimization. Secondly, when prior knowledge is introduced to detect and separate the corresponding weak signals, it may degenerate during the optimization task. In this regard, models based on sparse unmixing may be a good choice because such methods introduce existing spectral libraries and only optimize abundance. But when the spectral library is incomplete or the data is too large, it may cause mismatch, high coherence or high computational complexity problems.

[0005] Semi-supervised NMF models can be grafted into hyperspectral weak signal analysis to some extent, because such models strip off the simultaneous optimization mode of endmembers and optimize different endmembers differently. However, existing semi-supervised NMF-based methods only consider that certain land cover spectral features of a region can be retrieved from a standard library or ground survey. When some pure endmembers are known and available before performing the unmixing task, they are regarded as endmember libraries and subsequently replace the initialized endmembers. However, if the spectral correlation between the spectral library and the initialized endmembers is low, the introduced endmember library may not match the already initialized endmembers. Secondly, a large amount of previous work has demonstrated that NMF-based methods can produce the desired unmixing results at relatively low purity. Summary of the invention

[0006] In view of this, an object of the present invention is to provide a hyperspectral image semi-supervised weak signal unmixing method.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A semi-supervised weak signal unmixing method for hyperspectral images, the method comprising the following steps:

[0009] S1: Obtain satellite remote sensing observation images of the specified ground range and complete image radiation correction and atmospheric correction preprocessing;

[0010] S2: Use a combination of rough analysis of image spectrum peaks and valleys and field surveys to estimate the possible weak signal objects in the area observed by the satellite, including water bodies polluted by algae and ground polluted by industrial wastewater; measure the standard spectrum represented by the weak signal in the laboratory;

[0011] S3: Initialize the preprocessed remote sensing image to obtain the initial spectrum and corresponding abundance distribution range of typical objects in the image;

[0012] S4: A model including an unsupervised module, a supervised module and a sparse regularization term is used to estimate the spectrum of objects in the image and the corresponding distribution of objects; at the same time, the distribution range and mixed components of weak signal objects are obtained.

[0013] Optionally, the unsupervised module estimates endmembers and abundances, wherein the endmembers include endmembers initialized by Robust-OSP and prior knowledge;

[0014] set up Represents a hyperspectral image matrix containing L bands and N pixels, Indicates that it contains L bands and p Initialization terminal, Indicates that the observer has added qA priori spectrum is used to detect weak signals in hyperspectral data; the unsupervised module is described as follows:

[0015]

[0016]

[0017] The end member matrix and abundance matrix need to satisfy the non-negative constraint and the abundance matrix needs to satisfy the constraint that the abundance sum is unity. The optimization process will cause it to degrade, and a supervised module is considered to minimize the prior degradation.

[0018] Optionally, the supervised module is used to constrain the degradation of prior information and monitor weak signals in the hyperspectral image;

[0019] The supervised module is described as follows:

[0020]

[0021] Where A=[0 p×q ;I q×q ] indicates that the indicator matrix is ​​used to extract the endmember spectrum corresponding to the prior knowledge from the endmember matrix, V represents the prior knowledge; V and The difference is that the former is regarded as the observed object in the model and will not be optimized, while the latter is regarded as the prior for detecting weak signals in the model and will be updated and optimized in the model.

[0022] Optionally, the unsupervised module and the supervised module work together in the detection of weak signals in the hyperspectral spectrum, and the sparse characteristics of the abundance matrix are considered at the same time. By combining the unsupervised model and the supervised model, a semi-supervised reweighted sparse model is obtained:

[0023]

[0024] After simplification, we get the model:

[0025]

[0026] The first fidelity term is an unsupervised module, the second fidelity term is a supervised module, and the third term is a sparse regularization term; R represents a matrix that reweights the non-zero elements of the abundance matrix to force the abundance matrix to become more sparse.

[0027] Optionally, the monitoring of weak signals in the hyperspectral image is specifically:

[0028] S11: For the observed hyperspectral image data Y, the endmember initialization method Robust-OSP is used to obtain the endmembers of typical objects

[0029] S12: For the observed hyperspectral image data, obtain the weak ground object signals existing in the observation area through field surveys or ground object spectral reflection peak analysis, and obtain the imaging spectrum of the weak signal through laboratory spectral imaging and use it as prior knowledge;

[0030] S13: Obtain the distribution of ground objects corresponding to the weak signal spectrum by optimizing model (5).

[0031] Optionally, the endmember initialization method Robust-OSP is used to obtain the endmembers of typical features. Specifically:

[0032] S21: Use SVD algorithm to convert high-dimensional data Y into low-dimensional data subspace X;

[0033] S22: Calculate the data point e1 with the maximum L2 norm as the initial low-dimensional end member;

[0034] S23: Project the remaining data onto the subspace spanned by the determined end elements , obtain the second low-dimensional end member e2 and update the end member set;

[0035] S24: repeat S23 until a specified number of low-dimensional end members are obtained;

[0036] S25: Inversely transform the low-dimensional endmembers to high-dimensional endmembers.

[0037] The beneficial effects of the present invention are as follows: the HyperWeak method proposed in the present invention is based on a semi-supervised sparse non-negative matrix design, and solves weak signal scenarios by considering two fidelity terms. The first fidelity term estimates the endmembers and abundances augmented by the initialized endmembers and the priors in an unsupervised manner. The second fidelity term minimizes the prior degradation in a supervised manner. In order to improve the robustness of HyperWeak, a reweighted sparse regularization term and an endmember initialization method based on Robust-OSP are added, so that the proposed HyperWeak algorithm can effectively and accurately estimate the abundance components and distribution of weak signal features.

[0038] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0040] Figure 1 Visual comparison of six extracted ground feature spectra and standard spectra; (a) Tree; (b) Water; (c) Dirt; (d) Road; (e) Sagebrush; (f) Water+Montmor;

[0041] Figure 2 Visual comparison of the abundance of six extracted features and the standard feature abundance distribution;

[0042] Figure 3 Schematic diagram of the semi-supervised model process for weak signal analysis;

[0043] Figure 4 Schematic diagram of end member initialization of Robust-OSP algorithm. DETAILED DESCRIPTION

[0044] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0045] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0046] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0047] The present invention comprises the following steps:

[0048] S1: Obtain satellite remote sensing observation images of the specified ground range and complete image radiation correction and atmospheric correction preprocessing;

[0049] S2: Use a combination of rough analysis of image spectrum peaks and valleys and field surveys to estimate the possible weak signal objects in the area observed by the satellite, including water bodies polluted by algae and ground polluted by industrial wastewater; measure the standard spectrum represented by the weak signal in the laboratory;

[0050] S3: Initialize the preprocessed remote sensing image to obtain the initial spectrum and corresponding abundance distribution range of typical objects in the image.

[0051] S4: Model (5) is used to estimate the spectrum of the main objects in the image and the corresponding object distribution. At the same time, the distribution range and mixed components of the weak signal objects are obtained.

[0052] Unsupervised Optimization:

[0053] The unsupervised part of HyperWeak focuses on estimating endmembers and abundances, where the endmembers contain endmembers initialized by Robust-OSP and prior knowledge. Represents a hyperspectral image matrix containing L bands and N pixels, Indicates that it contains L bands and p Initialization terminal, Indicates that the observer has added q A priori spectrum is used to detect weak signals in hyperspectral data. The unsupervised optimization module is described as follows:

[0054]

[0055]

[0056] The end member matrix and the abundance matrix need to satisfy the non-negative constraint and the abundance matrix needs to satisfy the constraint that the abundance sum is one. The optimization process will cause it to degrade. Therefore, a supervised module is considered to minimize the prior degradation.

[0057] Supervised Optimization:

[0058] The supervised module is used to constrain the degradation of prior information to effectively monitor weak signals in hyperspectral images. The supervised model can be expressed as follows:

[0059]

[0060] Where A=[0 p×q ;I q×q] indicates that the indicator matrix is ​​used to extract the endmember spectrum corresponding to the prior knowledge from the endmember matrix, and V represents the prior knowledge. V and The difference is that the former is regarded as the observed object in the model and will not be optimized, while the latter is regarded as the prior for detecting weak signals in the model and will be updated and optimized in the model.

[0061] Overall model:

[0062] This scheme considers the synergy of unsupervised modules and supervised modules in the detection of hyperspectral weak signals, and at the same time considers the sparse characteristics of the abundance matrix. By combining the unsupervised model and the supervised model, the following semi-supervised reweighted sparse model is proposed:

[0063]

[0064] Simplifying the above model, we have the following model:

[0065]

[0066] The first fidelity term is an unsupervised module, the second fidelity term is a supervised module, and the third term is a sparse regularization term. R represents a matrix that reweights the non-zero elements of the abundance matrix to force the abundance matrix to become more sparse.

[0067] The weak signal detection method of the present invention and the existing traditional unmixing method are run in the standard Jasper Ridge experimental data, where Jasper Ridge contains four normal objects (Tree, Water, Dirt, Road) and two weak signal objects (Sagebrush and Water+Montmor). The experimental results are shown in Table 1.

[0068] Table 1 Experimental results of the algorithm on the Jasper Ridge dataset

[0069]

[0070] As shown in Table 1, compared with the existing traditional hyperspectral unmixing method, the present invention has a lower root mean square error result and the lowest signal degradation level in the scenario where there are weak signals in the hyperspectral image. At the same time, for normal ground objects, the present invention can also obtain better endmember extraction capabilities. Figure 1 and Figure 2 The figure shows the visual comparison of the six ground objects estimated from the image by the present invention (short solid line) and the standard spectral library (long solid line). Figure 1 and Figure 2 As shown, the endmember spectrum estimated by the present invention basically fits the real endmember spectrum, and the estimated abundance and composition of the ground objects also basically conform to the real ground object distribution information.

[0071] Figure 2 Visual comparison of the six extracted feature abundances (second row) and the standard feature abundance distribution (first row), from left to right: Tree, Water, Dirt, Road, Sagebrush, and Water+Montmor.

[0072] Figure 3 The schematic diagram of the present invention for analyzing weak signals in a hyperspectral image is shown. The schematic diagram can be briefly described as follows:

[0073] Step 1: For the observed hyperspectral image data Y, the endmember initialization method Robust-OSP is used to obtain the endmembers of typical objects

[0074] Step 2: For the observed hyperspectral image data, obtain the weak ground object signals existing in the observed ground object range through field surveys or ground object spectral reflectance peak analysis, and obtain the imaging spectrum of the weak signal through laboratory spectral imaging and use it as prior knowledge.

[0075] Step 3: Obtain the distribution of ground objects corresponding to the weak signal spectrum by optimizing model (5).

[0076] Figure 4 The important steps involved in the present invention, namely the initialization process of the Robust-OSP initialization algorithm, are shown. The process can be described as the following steps:

[0077] Step 1: Use the SVD algorithm to convert high-dimensional data Y into a low-dimensional data subspace X.

[0078] Step 2: Calculate the data point e1 with the maximum L2 norm as the initial low-dimensional end member

[0079] Step 3: Project the remaining data onto the subspace spanned by the determined end elements , obtain the second low-dimensional endmember e2 and update the endmember set.

[0080] Step 4: Repeat the third step until a specified number of low-dimensional end members are obtained.

[0081] Step 5: Inversely transform the low-dimensional endmembers to high-dimensional endmembers.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A semi-supervised weak signal unmixing method for hyperspectral images, characterized by: The method comprises the following steps: S1: Obtain satellite remote sensing observation images of the specified ground range and complete image radiation correction and atmospheric correction preprocessing; S2: Use a combination of rough analysis of image spectrum peaks and valleys and field surveys to estimate the possible weak signal objects in the area observed by the satellite, including water bodies polluted by algae and ground polluted by industrial wastewater; measure the standard spectrum represented by the weak signal in the laboratory; S3: Initialize the preprocessed remote sensing image to obtain the initial spectrum and corresponding abundance distribution range of typical objects in the image; S4: A model including an unsupervised module, a supervised module and a sparse regularization term is used to estimate the spectrum of objects in the image and the corresponding distribution of objects; and the distribution range and mixed components of weak signal objects are obtained at the same time; The unsupervised module estimates end members and abundance, wherein the end members include end members initialized by the end member initialization method Robust-OSP and prior knowledge; set up Represents a hyperspectral image matrix containing L bands and N pixels, It means that it contains L bands and p initialization end elements. It indicates that the q prior spectra added by the observer are used to detect weak signals in the hyperspectral data; the unsupervised module is expressed as follows: The end member matrix and abundance matrix need to satisfy the non-negative constraint and the abundance matrix needs to satisfy the constraint that the abundance sum is unity. The optimization process will cause it to degrade. Consider a supervised module to minimize the prior degradation. The supervised module is used to constrain the degradation of prior information and monitor weak signals in the hyperspectral image; The supervised module is described as follows: Where A=[0 p×q ;I q×q ] indicates that the indicator matrix is ​​used to extract the endmember spectrum corresponding to the prior knowledge from the endmember matrix, V represents the prior knowledge; V and The difference is that the former is regarded as the observed object in the model and will not be optimized, while the latter is regarded as the prior for detecting weak signals in the model and will be updated and optimized in the model; The unsupervised module and the supervised module work together in the detection of weak signals in the hyperspectral spectrum, and the sparse characteristics of the abundance matrix are considered at the same time. By combining the unsupervised model and the supervised model, a semi-supervised reweighted sparse model is obtained: After simplification, we get the model, which is the model of the sparse regular term: The first fidelity term is an unsupervised module, the second fidelity term is a supervised module, and the third term is a sparse regularization term; R represents a matrix that reweights the non-zero elements of the abundance matrix to force the abundance matrix to become more sparse; The endmember initialization method Robust-OSP is used to obtain the endmembers of typical features. Specifically: S21: Use SVD algorithm to convert high-dimensional data Y into low-dimensional data subspace X; S22: Calculate the data point e1 with the maximum L2 norm as the initial low-dimensional end member; S23: Project the remaining data onto the subspace spanned by the determined end elements , obtain the second low-dimensional end member e2 and update the end member set; S24: repeat S23 until a specified number of low-dimensional end members are obtained; S25: Inversely transform the low-dimensional endmembers to high-dimensional endmembers.

Images