Remote sensing ground object recognition mainly depends on the difference of spectral and spatial characteristics of ground objects. Due to the low spectral resolution of multi-spectra, the spectral characteristics of ground objects can not be fully expressed. Ground object recognition mainly depends on the spatial characteristics of ground objects, including gray level, color, texture, shape and spatial relationship. Information processing and information extraction mainly apply image enhancement, image transformation and image analysis methods to enhance the differences in hue, color and texture of images, so as to achieve the purpose of distinguishing ground objects to the maximum extent. With the successful development and industrialization of imaging spectrometer, remote sensing ground object information extraction has entered a new era. The recognition of ground objects by imaging spectrum mainly depends on the spectral characteristics of ground objects, and directly uses the spectral characteristics of rocks and minerals to identify ground objects and quantitatively analyze ground object information. The following two aspects are discussed: multispectral and hyperspectral remote sensing information processing.
Research progress of 1. multispectral methods
Multispectral information extraction mainly focuses on hue information extraction, texture information extraction and information fusion.
(1) Tone Information Extraction
For hue information extraction, some enhancement processes are mainly used to expand the gray difference between objects in the image, so as to highlight the target information or improve the image effect, and improve the discrimination ability of interpretation marks, such as contrast expansion, color enhancement, operation enhancement, transformation enhancement and so on. These traditional image processing methods meet the needs of application to some extent. In recent years, a series of information extraction techniques based on principal component transformation have been developed, which have played an important role in the extraction of rock and mineral information. For example, the improved direct principal component analysis proposed by Zhang Manlang (1996) can extract iron oxide information. (Kruse, 1996, Creen et al., 1988), (noise-adjusted principal component transformation) (Lee et al., 1990), block principal component transformation (Jia et al., 1999), correspondence analysis based on principal components (Carr At the same time, according to the principle of pattern recognition, supervised and unsupervised classification methods are proposed and designed, and classification recognition is carried out by decision tree (Wrbka, et al.,1999; Friedel et al.,1999; Hansen et al., 1996), these technologies and methods are based on the gray characteristics of images, and use the knowledge of mathematical statistics to classify objects and extract information.
(2) Texture information extraction
The edge and texture information of remote sensing images have a certain effect on the recognition of line ring structure, but it seems that it is not helpful to the recognition of lithology. Edge information is usually extracted by filter operator or sharpening method (Gross et al.,1998; Walber, 2000). Texture information extraction usually adopts * * * generation matrix, Fourier power spectrum and texture spectrum.
(3) Information fusion
The research of multi-source data fusion is also very popular and in-depth, and its technical methods involve different mathematical knowledge (Jimen et al.,1999; Bohr,1998; Robinson et al., 2000; Price,1999; Gross et al., 1998), such as wavelet information fusion. The application field involves non-remote sensing data (Wang Runsheng,1992; Zhu,, 1994), etc., superposition and fusion of remote sensing data, geochemical data and geophysical data. On the one hand, these methods broaden the application field of remote sensing, on the other hand, they also expand the application ability of remote sensing.
Generally speaking, the extraction of multi-spectral remote sensing information of rocks and minerals is mainly based on the gray characteristics of images, that is, based on the difference of reflectivity intensity of rocks and minerals, some mathematical transformation methods are used to enhance or highlight the target information, which is convenient for intuitive interpretation. In data processing, due to the limited band, it is impossible to effectively introduce the spectral knowledge of rock and mineral categories, and the accuracy of the results depends more on the experience of researchers.
2. Research progress of hyperspectral methods.
Imaging spectroscopy is a leap in the development of multispectral technology. By dispersing or splitting each spatial pixel, it forms dozens or even hundreds of narrow bands covered by continuous spectrum, and images the spatial characteristics of the target object at the same time. The formed remote sensing data can be described intuitively by "image cube (three-dimensional)", in which two dimensions represent space and the other dimension represents spectrum. In this way, the "continuous" spectrum and its diagnostic characteristic spectrum of the ground object can be obtained arbitrarily in the three-dimensional space where the spectrum and spatial information are fused, so that the target ground object can be directly identified based on the spectral knowledge of the ground object, and then quantitative ground object information can be obtained. In geological application, mineral identification and information processing technology can be divided into: ① characteristic parameters based on single diagnostic absorption; ② Based on complete waveform features; ③ Based on spectrum knowledge model.
The single diagnostic absorption characteristics of rock minerals can be completely characterized by absorption band position (λ), absorption depth (H), absorption width (W), absorption area (A), absorption symmetry (D), absorption number (N) and ranking parameters. According to the single diagnostic absorption waveform of end-member minerals, these parameter information can be extracted and enhanced from imaging spectral data, which can be directly used to identify rocks and mineral types. For example, the IHS coding and absorption band diagram (Kruse, 1988) is a spectral image removed by continuous method. It defines the band absorption center position image, band depth image and band semi-extreme width image, and gives their lightness (H), intensity (L) and saturation (S) in HS I space respectively, and then inversely transforms them into RGB chromaticity space. So as to directly identify the minerals according to the hue difference. When describing the single diagnostic absorption characteristic parameters of rocks and minerals, the absorption depth is a very important characteristic index, which has been paid attention to. For example, RBD image (relative absorption band depth image) (Crowley et al., 1989) uses ratio operation to enhance the absorption depth of endmembers, that is, according to the sum of shoulder reflectivity of individual diagnostic absorption peaks of endmembers to be identified, the relativity of diagnostic absorption peaks of endmembers is characterized by dividing it by the quotient image of the sum of reflectivity of corresponding wavelengths adjacent to the center of the valley. In RBD images of different endmember minerals, apart from the pixel ratio representing the possibility of endmember minerals, endmember minerals are identified by further feature enhancement and selection (such as PC transform analysis). Because of the asymmetry of absorption peak, it is difficult to describe its characteristics accurately by RBD method. The continuous interpolation band algorithm (CIBR) (De Jong, 1998) and the spectral absorption index image (SAI, Wang Jinnian, etc., 1996) are similar to the relative absorption depth map method, but a symmetry factor is introduced to make their description of absorption characteristics more reasonable. CIBR uses the radiation value at the center of the diagnostic absorption valley divided by the product of the radiation value of the left and right shoulders and the symmetry factor of absorption characteristics to generate the corresponding quotient image, which is used to enhance the diagnostic absorption depth of different minerals and identify minerals. SAI method is similar to CIBR, but it also adds a symmetry factor to the characteristics of single absorption wave shoulder. The above method is similar to the traditional ratio or color enhancement processing. The biggest difference from the conventional enhancement treatment is that it organically combines the prior knowledge of the spectral characteristics of end-member minerals, and its pertinence and purpose are more clear. Due to the influence of atmospheric radiation on the spectral characteristics of remote sensing data and the influence of spectral drift and change formed by spectral mixing on a single waveform, the recognition results contain great interference.
The biggest advantage of imaging spectrum is to reproduce the spectral curve of the corresponding phenomenon by using limited subdivision spectral bands. In this way, the use of full spectrum curve for mineral matching identification can improve the uncertain influence of single waveform (such as spectral drift and change) to a certain extent and improve the accuracy of identification. The recognition method based on full waveform is to reasonably select a metric function to measure the similarity between the standard spectrum or the measured spectrum and the image spectrum in the two-dimensional space composed of the reference spectrum and the pixel spectrum. For example, spectral matching (SM) (Baugh et al., 1998) is to use the Euclidean distance measure function of the spectral vectors of rocks and minerals, that is, to find the difference between the image pixel spectrum and the reference spectrum in the spectral space. The smaller the distance, the higher the fitting degree between the endmember spectrum of the image or the endmember spectrum to be identified and the reference spectrum measured in laboratory or field. Similarly, similarity index algorithm (Si) (Fenstermaker et al., 1994) is based on Euclidean distance laterality, and identifies ground objects according to the average sum of squares of band differences between the average spectrum of image pixels of known ground object types and the spectrum of unknown image pixels. The above two methods are more reliable than the parameter identification technology based on single absorption waveform. However, due to the influence of spectral data resolution, the spectral difference is not obvious, and it is difficult to accurately classify and identify ground objects because of the inherent defects of Euclidean distance measure. Spectral Angle Grapher (SAM) (Ben Dole et al.,1994; Crosta et al.,1998; Drake et al., 1998: Yuhas et al., 1992) is a multidimensional spectral vector space composed of rock and mineral spectra. The similarity between the endmember vector (r) of rock and mineral reference spectra and the spectral vector (t) of image pixels is solved by using an angle measure function of rock and mineral spectral vectors. Reference endmember spectra can come from laboratory and field measurements, and can also come from image pixel spectra of known categories. According to their similarity, the information of mineralization and alteration can be identified and extracted. The difficulty of this method lies in how to reasonably select the threshold of information segmentation. However, from the perspective of existing applications, this method is simple and reliable. Cross-correlation matching (Fer-rier et al.,1999; Varder Meer et al., 1997) uses correlation factor (R.) as similarity index to identify minerals through pixel-by-pixel cross-correlation matching. When the reference spectrum is completely matched with the inspection spectrum, its position m = 0；; When the reference spectrum moves to the long wave direction, its m < 0. On the contrary, m > 0. In RGB space, skewness, t-test value and correlation factor are given r, g and b respectively; If it is in the "0" matching position, its slope, t-test value and correlation factor (R.) are all close to "1", and it is displayed in white, thus identifying the end-member minerals. For intelligent identification of minerals, a complete spectral shape is usually used. For example, Tetracord mineral identification software, based on UNIX platform, automatically identifies minerals by fitting the spectrum in the spectrum database with the image spectrum. Wang Runsheng et al. (1999) used neural network to automatically identify minerals according to their complete waveforms. When there are a large number of known ground object spectra, the above method has strong adaptability. It is more useful for image feature recognition. However, the obvious deficiency is that it is difficult to accurately match the ground objects with little difference in overall spectral characteristics due to the influence of spectral changes, observation angles and granularity of actual ground objects on the obtained data, resulting in confusion and errors in rock and mineral identification and analysis.
Identification technology based on spectral model is a signal processing technology based on certain optical, spectroscopic, crystallographic and mathematical theories. It can not only overcome the shortcomings of the above methods, but also accurately quantify the physical characteristics such as the composition of surface materials while identifying the types of ground objects. For example, linear mixed spectral decomposition model (SMA/SUM)(Adams et al.,1986; Mustard, etc.,1987; Roberts et al.1997; Sabol et al.,1992; Settle et al.1993; Shipman et al; 1987: Shimaya Hei et al.,1991; Smith et al., 1985), the spectral linear decomposition model can be constructed according to the difference of spectral reflectivity response of different objects or different pixels. A pixel is not a single type of ground objects, but more composed of different types of ground objects. Therefore, in most cases, the pixel spectrum is not a linear mixture of pure ground spectrum, but more nonlinear. For single scattering, it can be decomposed into linear model, while multiple scattering is considered as nonlinear mixing. Because the average single scattering albedo abundance mainly depends on the content of different components, it can be considered as linear mixing (Mustard et al., 1987). In this way, through the single scattering albedo (SSA) conversion, the nonlinearity can be "linearized" by using the operator W=(3r+6)r/( 1 +2r)2, and then the spectrum can be decomposed. Tompkins (1996) proposed an improved spectral mixture analysis (MSMA) model. The model adopts virtual endmembers and damping least squares algorithm. According to some prior knowledge, the end members of sub-images can be effectively selected for spectral decomposition, which improves the practicability of SMA. Compared with SMA, the biggest differences of MSMA are as follows: ① endmembers and their abundance are unknown variables; ② Solve all pixels in the data set at the same time. For CEM (constrained energy minimization technology) (Fahlander et al.,1997; Fahlander et al.,1996; Resmini et al., 1997) uses the weight coefficient wk related to the pixel spectrum (ri) of the target area (or ROI area) in the imaging spectrum image sequence to describe the digital value y of the pixel vector, so as to carry out feature selection and decomposition for ground object recognition and information extraction. Like the mixed spectral decomposition model, the decomposition results not only represent the type information of the identified pixels, but also represent their abundance ratio mechanically. Different from the mixed spectral decomposition model, this method relies more on the statistical characteristics of the target area, but the results are more accurate. In a word, these methods rely more on spectral knowledge and mathematical methods, and it is difficult to determine characteristic parameters or accurately describe spectral models in practical applications, which limits the application of such technical methods. However, because this method quantifies the material composition while recognizing the ground objects, with the maturity of a series of technologies, the in-depth development of spectroscopy, crystallography and other knowledge, the improvement of recognition accuracy and quantification ability, its application will be more and more extensive.
In China, some imaging spectra have been used to directly identify minerals, but the performance of domestic sensors is not perfect enough, and the data signal-to-noise ratio is low. However, some achievements have been made in qualitative rock and mineral identification. For example, Gan Fuping (2000) used principal component analysis based on waveform feature combination to effectively divide the lithology of Hougou gold deposit in Zhangjiakou, Hebei Province. Liu Qingsheng (1999) extracted the gold-bearing alteration of a mining area in Inner Mongolia through correspondence analysis. It lags behind the developed countries such as the United States in direct quantitative mineralization identification, identification mode and identification pedigree, and there is still a certain gap.
In a word, the study of spectral mechanism of rocks and minerals, the basis of remote sensing information extraction, and the research of remote sensing information extraction methods and technologies complement each other and have a certain corresponding relationship.
With the development of remote sensing spectral imaging technology, the application foundation of remote sensing ground objects and the research of remote sensing image information extraction technology have been developed, and its research direction and trend mainly focus on the correlation between spectral feature knowledge and physical and chemical properties of ground objects and spectral physical models. The correlation between physical and chemical properties of ground objects and spectral characteristics, as well as the in-depth analysis and research on spectral physical model, can provide theoretical support for remote sensing to directly identify minerals, extract the distribution law, attributes and physical and chemical properties of ground objects, and carry out deep information mining of ground objects from different angles, and promote the development of remote sensing application technology. The practicality and industrialization of remote sensing geoscience application is the result of the mutual promotion of the application basis of remote sensing ground object spectrum and the research on information extraction technology of remote sensing ground object influence.
The development of the research on the spectral mechanism of ground objects, the basis of remote sensing information extraction and the research on the methods and technologies of remote sensing information extraction will lead to the combination of the three, which will eventually be integrated into the remote sensing application model and technology, so as to make full use of their respective advantages, improve the application ability of remote sensing, enhance the understanding of geological application, and simulate, evaluate and predict the development law of geoscience.