On the one hand, the high spectral resolution imaging spectral remote sensing technology is the inheritance, development and innovation of multispectral remote sensing technology, so the vast majority of multispectral remote sensing data processing and analysis methods can still be used for hyperspectral data; on the other hand, the imaging spectral technology has the technical characteristics that are different from the multispectral technology, i.e., the high spectral resolution, the ultra-multi-bands (bands <1000, usually 100-200 or so) and massive data with very high spectral resolution (Ultra Spectral) resolution (bands >1000, mainly used for detecting atmospheric chemical constituents). Therefore, conventional multispectral data processing methods are not suitable for quantitative analysis of imaging spectral data, so imaging spectral data processing and analysis techniques have emerged. In the imaging spectral data processing and analysis methods, the key technical issues are the spectral reconstruction of features, the quantification and extraction of spectral features, the decomposition and quantitative analysis of mixed image elements, and model identification.
2.3.2.1 Spectral Reconstruction Technology
The process of inverting the spectral characteristics of a feature from imaging spectral data according to different models and algorithms is the feature spectral reconstruction technology. According to different working situations and conditions, adopting different inversion models to reconstruct the spectra of features is the first step to realize the quantitative analysis of remote sensing of imaging spectral data. If it is not inverted, there is not a unified physical quantity for comparison. At present, spectral inversion models can be roughly divided into three major types: atmospheric transport models based on atmospheric transport theory, statistical models based on statistical analysis, and empirical regression models based on simultaneous observations of ground features.
2.3.2.1.1 Models based on atmospheric transport theory
These models essentially use theoretical models to eliminate the absorption and scattering of reflectance radiation energy and atmospheric program radiation effects by molecular particles such as atmospheric gas molecules, water vapor, aerosols, and dust, and to reduce them to the reflectance radiation spectra of the ground objects. This is a complex method of generating absolute reflectance that requires both spectral and atmospheric measurements of the feature, i.e., a re-inversion of the imaging spectral data for absolute radiative calibration. Critical in this inversion process is the modeling of atmospheric transport. Since 1960, when Chndrasekhar proposed the theory of radiative transfer, many methods, such as the Ordinate method and the Variational method, have been successively developed to solve the radiative transfer problem. At present, the common atmospheric transmission models are 5 S (the Simulation of the Satellite Signal in the Solar Spectrum), 6S (the Second Simulation of the Satellite Signal in the Solar Spectrum) , LOWTRAN 7 and MODTRAN (Teillet, P.M., 1989; Vermote, E., Tanr?en, D., Deuz?e, J.L. et al. 1994, 1996; Bo?Cai Gao, K.B. Heidebrecht and A.F.H. Goetz, 1997; (Z. Qin, A. Karnieli and P. Berliner, 2001). Among them, the 6S model, which was researched and developed by Tanré et al. in France, is one of the more well-developed atmospheric radiation correction modeling algorithms in the world. The algorithm can reasonably deal with atmospheric scattering and absorption as well as produce continuous spectra to avoid large quantitative errors in spectral inversion. It also makes full use of analytical expressions and pre-selected atmospheric modes, resulting in a significant reduction in computation time. Many remote sensing experts who have used this model for spectral inversion of features believe that the model has a higher computational accuracy than other models. The disadvantage is that it is necessary to carry out the observation of spectral reflectance of typical features in the test area as well as the actual measurement of atmospheric environmental parameters, such as: atmospheric optical thickness, temperature, air pressure, water vapor content, and atmospheric distribution conditions. Comparatively speaking, the LOWT?RAN 7 and MODTRAN models do not require ground-based measurements of the reflectance of typical features, although their computational accuracy is lower. These models are generally used for the sensor selected calibration field, to carry out data absolute radiation calibration.
2.3.2.1.2 Model Based on Statistical Analysis
The model is established on the basis of analyzing the transmission characteristics of spectral remote sensing information of different features in different spectral wavelengths, and the inversion model of spectral characteristics of features is obtained by using computers to statistically analyze the spectral characteristics of typical features. Commonly used models for spectral inversion of features from imaging spectral data are the smooth field reflectance model FFR (Flat Field Reflectance) (Goetz and Srivastava, 1985; Conel, 1985; Crowley at al., 1988; Rast et al., 1991), the intrinsic average relative reflectance model IARR (Inherent Average Relative Reflectance), and the spectral inversion model IARR (Intrinsically Averaged Relative Reflectance, IARR). Reflectance model IARR (Internal Average Relative Reflectance) (Kruse et al., 1985; Kruse, 1988; Mackin and Munday, 1988; Zamoudio and Atkinson 1990), the logarithmic residual model LRC ( Logarithm Residual Correction) (Green, R.O. etal., 1985; Gower, J.F.R. etal., 1992). Among these three models, the FFR model eliminates atmospheric radiative attenuation and the zero response of the instrument by selecting the average value of features on the image that are uniformly smooth (smoothness means that the features have no spectral absorption bands and the spectral curves are flat) in both spectral and geomorphic features; the White model normalizes the image based on the average spectral curve averaged over the whole image, and then calculates the ratio of the spectral curves of each pixel to the average spectral curve ratio, which is the feature spectral characteristics; LRC model is corrected by Lyon and Lanze to eliminate solar radiation attenuation, atmospheric effects and topographic effects.
The logarithmic residual correction formula proposed by Green and Graige is as follows:
lg(Rij)=lg(DNij)-lg(aveDNi)+lg(DNi) + lg(DNg) (2-3-1)
Here Rij is the residual value of the ith band and the jth image element; DNij is the luminance value of the ith band and the jth image element; DNi is the mean value of the data of all the image elements of the ith band; DNj is the mean value of the data of the jth image element over all the bands; DNg is the mean value of all the bands and all the image elements. The method is completely based on the image itself features and does not require field feature spectral measurements. Among the first two models, the FFR model is superior to the IARR model, which overcomes the weakness of the false reflection peaks of the IARR model due to the influence of the strong absorption features of the image foo, and is less computationally intensive.
2.3.2.1.3 Empirical linear regression model
The essence of the technique of reconstructing feature spectra using this method is to carry out simultaneous reflectance observation of typical features, and then, based on the DN value of imaging spectral data and the reflectance value of the ground-measured features, find the regression equation Rij=Aj-DNij+Bj by the least-squares method (here, Aj, Bj are the linear regression coefficients for the sensor's jth band). (where Aj, Bj are the linear regression coefficients of the jth band of the sensor), and then, according to this equation, invert the reflectance spectrum of the ground. The mathematical and physical meaning of this model is clear, and the method is simple, less arithmetic, and widely used (Roberts et al. 1985; Conel et al. 1988; Elvidge 1988; Green et al. 1988; Kruse et al. 1990; Zamoudio and Atkinson 1990). For example, Abrams of JPL, USA, used the model for mineralogical infill at the Cuprite mine in Nevada State, USA; Zamudio et al. of the University of Colorado, USA, used the model for mineral identification and petrographic analysis in eastern Nevada State, USA; Mochizuki of Pasco, Japan, used the model for reflectance spectroscopy of alteration minerals in Navada (Pasco, Japan) used the model to conduct reflectance spectroscopy studies of altered minerals in Navada State, USA. The shortcomings of this model are that it is costly to carry out spectral observation of field features, and the accuracy of regression depends on the accuracy of field generalization.
In addition to these typical spectral reconstruction models, there are also models such as UA RT Code, JPL Code, Continuous Interpolation Band Ratio (CIBR) algorithm, and background method (De Jong, 1998).
2.3.2.2 Quantification and Extraction of Rock and Mineral Spectral Features, Quantitative Analysis and Recognition Models
Imaging spectral data are processed by spectral reconstruction models to obtain spectral feature spectral lines of the features. Different feature spectra have different diagnostic characteristic spectral bands, such as absorption spectral bands, calculus changes of characteristic spectral lines, waveform changes and so on. How to effectively carry out the quantitative analysis of feature characteristics and identify the features, first of all, we need to figure out how to quantify and extract the spectral features of the features. Therefore, it is very necessary to carry out the quantitative extraction based on the feature spectrum.
2.3.2.2.1 Spectral Feature Measurement, Extraction and Matching Recognition Model
(1) In terms of spectral features of a feature (in this case, the reflectance spectra of the feature), there are only two major types: the absorption bands (or reflectance valleys) and the change in the slope of spectral curves (including the change in the waveform). For the morphology and structure of these two types of spectral features, different metrics are adopted. Currently, the analytical measure of absorption bands is the shell factor method, which measures the wavelength position, depth, width and symmetry of absorption bands by normalizing the spectral curve (Lyon et al., 1985; F.A. Kruse, A.B. Lefkoff, 1993). Lefkoff, 1993). This shell coefficient method can be expressed by the ratio of the convex curve of the shell to the spectrum, or it can be derived by subtracting the shell value from the spectral reflectance value at the corresponding wavelength. Due to the asymmetry of the absorption peaks, it is difficult to accurately characterize them using the RBD method. Continuously interpolated band algorithm (CIBR, Continuun interpolated band algorithm) (De Jong, 1998) and spectral absorption index image (SAI, spectral absorption index image) (Wang Jinnian et al., 1996) are similar to the relative absorption depth map method. However, a symmetry factor was introduced to make its characterization of absorption more reasonable. In addition to these measurement parameters, there are various metrics for vegetation spectra, such as vegetation, greenness index, and so on. For the characteristics of the slope change of the spectral curve, the characterization and extraction methods are based on the overall waveform characteristics of the feature spectrum metrics, such as the Fourier transform waveform analysis method is to use a finite level of harmonic amplitude and initial phase metrics of the features of the feature spectrum; Chebyshev polynomial function-based waveform analysis is to use the polynomial function of the feature spectral curve to fit, and to extract the coefficients of a finite number of terms to represent or combine the features, or to use their ratios to represent the features, or to use their ratio to represent the features of the feature spectrum. The Chebyshev polynomial-based waveform analysis is to use the polynomial function to extract the coefficients of finite terms to represent or combine the features, or to use their ratios to represent the coefficients of the spectral waveform features of the features; the spectral angularity based on the analysis of the similarity of the waveforms (overall or segmented); and also the differential metrics, integrative metrics, and dichotomous metrics for the features of the spectral curves, etc. Of course, the absorption bands and slope features can also be characterized by statistical features, such as mean, variance, covariance matrix, eigenvalues, eigenvectors, eigenfactors, and intragroup deviation.
(2) Spectral matching recognition model is different from multispectral pattern recognition, it is based on the spectral feature metric parameters to match the recognition, is one of the characteristics of imaging spectral data processing and analysis. This characteristic model is often used in the processing of visual interactive image and spectral, spectral and standard spectral form. At present, the spectral matching recognition models are: coded matching recognition method (mean coded matching method, slope coded matching method, ratio coded matching method L absorption peak coded matching method, waveform matching method and spectral angle matching method and so on. In these matching recognition models, coding matching is basically coding, matching and recognition by binary value (0 and 1); absorption features coding matching is based on the shell coefficient method normalized to the absorption depth of each feature and wavelength position for coding; waveform matching including similarity, Fourier transform parameter, Chebyshev and other methods. In recent years the development of wavelet transform analysis in imaging spectral data analysis and processing applications are quite a lot, especially for the original signal according to different wavelet scales, decomposed into different wavelets for waveform analysis, highlighting the low-frequency weak information, is conducive to the enhancement of information, such as graphical image interpolation, fusion and hybrid pixel decomposition with wavelet transform.
2.3.2.2.2 Quantitative analysis and recognition modeling of imaging spectral data
Quantitative analysis and recognition modeling is one of the development directions of today's remote sensing technology, applied to imaging spectral data processing. Quantitative analysis and identification modeling, in addition to continuous refinement and improvement of existing quantitative and identification models based on statistical analysis (such as: improved principal component analysis, optimal band combination, change model maximum likelihood method, based on the transformation of the decision boundary feature matrix and orthogonal subspace projection), other disciplines of the new ideas, new methods are also constantly introduced to remote sensing data analysis and understanding, such as artificial intelligence expert system, fuzzy logic mapping, evidence reasoning, and fuzzy logic mapping, evidential reasoning, neural networks, fractal and dimensionalization, etc.
Artificial intelligence expert system technology is currently more popular information processing technology, especially for the solution of more complex problems have unique characteristics, Gotting and Lyon in 1986 has established the spectral information expert recognition system for the analysis of laboratory and field spectra, which is a combination of existing knowledge of the spectral characteristics of the features, by experts to determine the rules of discrimination Decision Tree (Decision Tree) to achieve the recognition of the spectral characteristics of the feature, the expert to determine the rules of discrimination (Decision Tree). Decision Tree) to achieve the purpose of recognizing features or feature categories. The Decision Tree, a knowledge-based hierarchy of discriminatory criteria, is the key to the success or failure of building an expert system. By using this system for code matching, they successfully identified 11 minerals from a large number of laboratory spectra.In 1993, F.A. Kruse and A.B. Lefkoff of the Center for Geospatial Studies (CSES) and the Institute of Environmental Sciences (IES) in the U.S. developed a knowledge-based expert system for geologic mapping of imaging spectra. Various features are selected for their roles in the recognition process and assigned corresponding weights, or recognition is carried out by establishing discrimination criteria based on the expert's knowledge and experience of the discrimination.
Currently, neural network models are favored in remote sensing feature analysis and recognition, and have a wide range of applications (Golen Giser, 1996; Giles, M.F. et al. 1995; Guo Xiaofang, 1998; Wang Runsheng et al. 2000). Since the neural network classification rules do not have specific requirements for the number of training samples and distribution characteristics, and thus can form a nonlinear discriminative boundary in the feature space, and there is also a certain degree of anti-noise, anti-jamming and adaptive ability, suitable for classification research on large data volumes, the most commonly used classification criterion is the backward propagation (BP) network model.
At present, from the point of view of the introduction of various new theories and methods of imaging spectral remote sensing data analysis and identification, most of the research and application of the model is still a kind of attempt, and the analysis is not deep enough in the research of how to combine the model with imaging spectral data.
2.3.2.3 Mixed Image Element Decomposition Model
Due to the reason of low spatial resolution, features with different components (end members), i.e., mixed image elements, appear within the image image element. Different features have different spectral characteristics, so it is necessary to improve the recognition accuracy through the hybrid spectral decomposition technique. The problem of hybrid image element is the research difficulty and hotspot of remote sensing technology. Since the spectral resolution of imaging spectroscopy technology has been improved from micrometer (μm) to nanometer (nm), the analysis and decomposition of its hybrid image element and its modeling research become more important.
At present, the methodology and technology of hyperspectral remote sensing mixed image element research firstly start from experiments to carry out the testing, analysis, numerical simulation, and decomposition model development of mixed spectra of features, and then extrapolate them to remote sensing images to carry out the analysis of mixed image elements of typical features, which mainly includes the acquisition of data of typical features (or can be manually laid out) by air-ground simultaneous observation, and after model analysis, the The decomposition of mixed-element features, or mixed-spectrum simulation synthesis. In the laboratory, by testing the mixed content of different mineral spectra, it is found that opaque minerals or dark minerals, whose spectra are proportionally mixed into other minerals, have a sharp rather than gradual decrease in the reflectance of the mix, indicating that their mixed spectra are in a nonlinear relationship with the spectra of the end-member minerals they are mixed with (magnetite and olivine). When the two minerals have similar hues, both the experimentally tested mixing spectra and the linear model synthesized mixing spectra show linear gradual changes, indicating that the mixing spectra can decompose the end-member mineral spectra according to the linear model, such as olivine and peridotite, and the wavelength positions of the absorption bands are also gradually transitioning from one wavelength position to another. Not only that, it is also found that at the wavelength of visible light and near infrared, the linear trend is observed when the low-composition end-member is mixed, and when the composition increases, the linear relationship drastically turns into a nonlinear relationship. In these three cases, the first nonlinear relationship is due to the combination of mixed spectra of the end-member components interact with each other, each other's influence after the spectrum is detected by the spectrometer; the second linear relationship is due to the end-member components do not interact with each other, and each of them independently reflect the electromagnetic wave energy contributing to the mixture of the spectra; and the third case is that the two kinds of relationship exists, and there is a critical condition (boundary condition) between the two. At present, there are very few studies on this aspect. According to these analyses, the hybrid image element decomposition models are broadly categorized into linear and nonlinear models. In remote sensing mixed image elements, the vast majority of features with similar reflectance can be decomposed into end-member components using a linear model, such as: soil and vegetation, cropland with different water content, rock outcrops and grassland, wasteland, etc. In an image, it is known in advance that there are N kinds of end-members (feature types), and the spectral reflectance of various end-members is also known, then a linear model can be used:
Technical research on imaging spectroscopic rock-mineral identification methods and analysis of influencing factors
Here, DNc is the DN value or reflectance of the mixed image element in the waveband C; Fi is the proportion of the ith kind of end-members in the mixed image element (or the weighting coefficients); DNi,c is the DN value (or reflectance) of the ith endmember on band C; Ec is the fitting error on band C. The decomposition was performed for each image element by solving the equation according to the least squares method. In the image, the DN value (or reflectivity value) of the endmember can either be taken from the training area or measured on the ground. The process of determining the endmember component is essentially an iterative process that results in minimizing the total error ε over M bands (and N ≤ M).
Imaging spectroscopy rock and mineral identification method technology research and analysis of the factors affecting
After obtaining the version of the various end-member composition, it can be quantitatively or semi-quantitatively on the end-member abundance to produce abundance and other thematic maps.
Mixed pixel decomposition with nonlinear models is not common, but there have been studies in this area, such as fuzzy segmentation model (Jin Ⅱ kim, 1996), probabilistic peng model, geometric optics model (Charles Ichoku, 1996) and based on the neural network model of the mixed pixel decomposition (Wang Xi-peng, Zhang Yangzhen et al., 1998) and so on.
The currently developed models are:
--Spectral Absorption Index Model SAI (Wang Jinnian, Tong Qingxi et al., 1996):
SAI=∑fiSAIi, ∑fi=1, fi>0 (2-3-4)
--Gaussian modeling method MGM: This model is a variety of simulation methods for modeling reflectance spectra based on the reflectance and absorption spectral properties of minerals and rocks. It is a deterministic rather than statistical method. Gaussian modified model MGM is an analytical technique developed in recent years on the basis of analyzing reflectance spectra (Cloutis, 1989, Veverka, J. et al., 2000).
m(x)=S*exp(-(xn-μn)2/2σ2), (2-3-5)
Often take n=-1.
Spectral identification and classification technology (Spectral Classification): mainly using the quantitative parameters of the hyperspectral features of the feature, combined with its distribution on the image space to extract the favorable information to achieve the purpose of classification. The main classification methods are:
--Maximum Likelihood Method MLC:
g(x)=-ln｜∑i｜+(x-mi)t∑i-1(x-mi), (2-3-6)
-- Artificial neural network technology ANN: Generally feed-forward network model, that is, the node input of the first hidden layer is equal to the weighted sum of the outputs of the nodes in the input layer. The number of iterations is based on the average error of the system being minimized.
Technical Research on Imaging Spectroscopic Rock and Mineral Identification Methods and Analysis of Influencing Factors
--Spectral Angle Mapping Method SAM (Spectral Angle Mapper): the method is based on calculating the difference between the test sample spectral vector (image element spectra) and the reference spectral vector (the trained end-member sample). spectra, or spectra from standard spectral libraries), in n-dimensional space (n bands) to determine the similarity between them. Generally the smaller the angle between the two vectors, the more similar the two spectral vectors are, which in turn identifies the two features as being of the same type, otherwise they are considered dissimilar. The mathematical model is:
Technical research on imaging spectral rock and mineral identification methods and analysis of influencing factors
Here i=1, 2, 3, ......, n, n is the number of bands.
--Spectral Dimension Feature Extraction (Spectral Dimension Feature Extraction): in hyperspectral remote sensing classification, this method is used to downscale data with multiple bands, high correlation, and high data redundancy. Related are statistical methods such as principal components, typical variables and improved PCA method.
--Optical Modeling: In addition to the aforementioned data analysis and modeling, there is a unique analysis model (optical modeling) for vegetation due to its unique reflective properties. This model mainly utilizes hyperspectral remote sensing data to predict or estimate a variety of biophysical and chemical covariates of vegetation, such as foliar index LAI, total biomass, cover, etc.; chlorophyll, water, N, P, K content, etc. The model is also an empirical statistical model. The model also belongs to the empirical statistical modeling approach. The general generic model is:
S=f(λ; θs, Φs; θv, Φv; С), (2-3-9)
Here S is the predicted biophysical and chemical parameter; λ is the wavelength; θs, Φs, θv, Φv are the parameters of the incident light and the geometric position of the sensor detections, and C describes the vegetation canopy as a characteristic parameter. The models relying on the method are PROSPECT, a spectral model of leaf optical properties, and SAIL, an arbitrary oblique scattering model of leaves, i.e., the LIBERTY model of biochemical parameter inversion.
Hyperspectral in vegetation applications in addition to biological and chemical covariates inversion analysis, but also focus on the use of vegetation spectral properties spectral lines of the blue edge, reflectance peaks, yellow edges, red absorption valley, red edge, near infrared reflectance plateau region and other changes and data normalization, logarithmic, differential and other transformations, to monitor the growth of the vegetation and pests and diseases, to identify, classify, and map the forest (Clark, R.N.. Roush.T.L., 1984).
2.3.2.4 Spectral Data Application Processing and Analysis Software
Through carrying out the research on testing and analysis of hyperspectral properties of rock and mineral and the technology and application analysis of imaging spectral methods, the following data processing and analysis software has been developed and developed:
2.3.2.4.1 Spectral Database and Analysis Software (400-2500φ)
Foreign countries: Standard mineral spectral libraries (including airborne spectra) and spectral analysis management software SPAM, IRIS of USGS and JPL of the U.S. Geological Survey, rock mineral spectral libraries of the Japan Institute of Geological Survey, etc. (http://speclib.jpl.nasa.gov; http://speclab.cr.usgs.gov; Kruse F A et al.1993).
Domestic: Anhui Institute of Optics of the Chinese Academy of Sciences, Institute of Remote Sensing Application of the Chinese Academy of Sciences, and Remote Sensing Center for Aerospace Physical Exploration of the former Ministry of Geology and Minerals and other scientific research institutes have built their own spectral libraries (Wang Runsheng et al., 2000).
2.3.2.4.2 Image Processing and Analyzing Software
At present, the commonly used spectral image processing and analyzing software at home and abroad are Erdas, PCI, ENVI and so on. Among them, both PCI and ENVI have hyperspectral analysis processing function (ENVI User's Guide., 2000). In addition, there are like Tetricorder (Clark, R.N., G.A. Swayze, K.E. Livo, 2003). Domestically, through the hyperspectral remote sensing method technology and demonstration application research, the Institute of Remote Sensing Application of the Chinese Academy of Sciences and the Aerospace Physical Exploration Remote Sensing Center of the Ministry of Land and Resources have successively established imaging spectral data analysis and processing systems, such as: HIPAS, ISDPS and so on.