Hyperspectral image mineral information extraction method based on pixel scale

Abstract

This invention discloses a method for extracting mineral information from hyperspectral images based on the pixel scale. The method includes the following steps: preprocessing the original hyperspectral image data to obtain a reflectance data cube; extracting spectral curves pixel by pixel based on the reflectance data cube and binding them with spatial coordinates to construct a structured pixel spectral database; processing the spectral curves in the structured pixel spectral database using envelope elimination and piecewise fitting methods to obtain diagnostic band absorption feature parameters; storing the diagnostic band absorption feature parameters and morphological feature parameters together in an extended node; identifying mineral types based on the feature parameters in the extended node; introducing virtual background endmembers in linear spectral decomposition and performing nonlinear correction on the decomposition results to obtain pixel-level mineral categories and relative contents.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing geology and image spectral big data processing technology, and particularly relates to a method for extracting mineral information from hyperspectral images based on pixel scale. Background Technology

[0002] Hyperspectral remote sensing technology, with its nanometer-level spectral resolution and continuous imaging capability in the visible to short-wave infrared range, has become a core technology for regional geological surveys, alteration information extraction, and mineral resource exploration. The formation of endogenic metallic deposits, such as hydrothermal and porphyry deposits, is often accompanied by large-scale hydrothermal alteration, resulting in distinct zoning of alteration mineral assemblages on the surface, such as sericitization, propylitization, and potassic alteration. The molecular vibrations (such as hydroxyl groups OH) within these alteration minerals (e.g., muscovite, sericite, chlorite, calcite, and iron oxides)... - carbonate CO3 2- Electron transitions and other processes produce unique spectral absorption characteristics in specific diagnostic bands. By analyzing these fine spectral "fingerprints," geologists can invert the migration paths and temperature gradient changes of hydrothermal fluids, thereby inferring the location of deep, concealed ore bodies. Therefore, hyperspectral remote sensing has become a key technological support for modern geological prospecting, transitioning from macroscopic mapping to microscopic quantitative inversion.

[0003] However, in the engineering application of massive hyperspectral data and the practice of refined mineral inversion, the existing technical system still faces the following bottlenecks and challenges that urgently need to be overcome: First, weak features are easily masked and missed. Traditional methods rely excessively on global spectral matching (such as spectral angle mapping SAM), comparing the overall similarity of the entire curve across hundreds of bands. This makes indicative mineral features such as trace amounts of sericite (weak absorption at 2200nm) at the edge of hydrothermal alteration easily submerged by the strong background of soil or the smooth spectra of featureless minerals such as quartz, resulting in the systematic loss of key geological information. Second, single-pixel quantitative inversion errors are huge. Existing technologies mostly remain at the qualitative identification stage. When processing shortwave infrared data, linear spectral unmixing models ignore the destruction of the "sum of 1" constraint condition by host rock minerals such as quartz and feldspar, which have no absorption features, after decontinuum processing. Furthermore, they fail to effectively correct the nonlinear mixing effect caused by multiple scattering between mineral particles, resulting in a serious deviation between the calculated mineral abundance and the actual geological situation. Third, data representation is unstructured. Traditional image data cubes (such as BSQ and BIP formats) focus on image rendering and lack standardized, structured storage and indexing mechanisms for massive spectral feature parameters. This leads to engineering challenges such as low I / O efficiency and severe memory overflow when performing batch training for machine learning or querying specific absorption depths, greatly hindering the practical application of AI-based mineral exploration prediction technology. Fourth, key geological laws are not embedded in the algorithm's underlying layer. For example, the slight shift in the center wavelength of the 2200nm absorption peak in sericite from greater than 2210nm to 2202-2206nm is crucial evidence for geologically identifying hydrothermal centers and mineral exploration target areas. However, existing software cannot achieve automatic Gaussian multi-peak stripping and statistical analysis of spatial displacement patterns at the pixel level, meaning this core mineral exploration indicator still relies on tedious manual interpretation.

[0004] Therefore, there is an urgent need in this field for a new paradigm for extracting hyperspectral mineral information that uses pixels as the basic deconstruction unit, has hierarchical database structured expression, performs differential recognition and Gaussian fitting for specific bands, and can overcome the bottleneck of nonlinear mixing to achieve high-precision content inversion. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a method for extracting mineral information from hyperspectral images based on the pixel scale, thereby resolving the issues present in the prior art.

[0006] To achieve the above objectives, the present invention provides a method for extracting mineral information from hyperspectral images based on pixel scale, comprising: The original hyperspectral image data is preprocessed to obtain a reflectance data cube; Based on the reflectance data cube, spectral curves are extracted pixel by pixel and bound to spatial coordinates to construct a structured pixel spectral database; The spectral curves in the structured pixel spectral database are processed using envelope elimination and piecewise fitting methods to obtain diagnostic band absorption characteristic parameters. The diagnostic band absorption characteristic parameters and morphological characteristic parameters are then stored together in the extended node. Based on the feature parameters in the extended nodes, mineral types are identified. Virtual background endmembers are introduced in the linear spectral decomposition, and the decomposition results are nonlinearly corrected to obtain pixel-level mineral categories and relative contents.

[0007] Optionally, the preprocessing of the raw hyperspectral image data includes: The original hyperspectral image data was processed using radiometric calibration and atmospheric correction to obtain surface reflectance data; The surface reflectance data is processed using minimum noise separation transform and wavelet filtering to obtain a reflectance data cube.

[0008] Optionally, the process of extracting spectral curves pixel by pixel based on the reflectance data cube and binding them to spatial coordinates to construct a structured pixel spectral database includes: The reflectance data cube is processed by scanning row by row and column by column to obtain a one-dimensional vector of single-pixel spectral reflectance; The one-dimensional vector of spectral reflectance of a single pixel is bound and encapsulated with two-dimensional spatial geographic coordinates using the HDF5 hierarchical data format to obtain a pixel spectral database with a B-tree index structure.

[0009] Optionally, the process of processing the spectral curves in the structured pixel spectral database using envelope elimination and piecewise fitting methods to obtain diagnostic band absorption characteristic parameters includes: The normalized spectral vector is obtained by processing the one-dimensional vector of the spectral reflectance of the single pixel corresponding to the hydroxyl absorption band with wavelengths of 2190–2220 nm using decontinuum processing. The set of potential absorption peak positions is obtained by processing the normalized spectral vector using second-order differentiation; Gaussian multi-peak fitting is used to process the set of potential absorption peak positions to obtain characteristic peak parameters in the converged state. These characteristic peak parameters are used as absorption characteristic parameters for the diagnostic band. The characteristic peak parameters include peak area, peak position, peak height, peak center, and half-width.

[0010] Optionally, the process of obtaining the morphological feature parameters includes: The normalized spectral curves of iron minerals were obtained by processing the spectral curves corresponding to the wavelength range of 850–1100 nm using decontinuum processing. The inflection point positions were obtained by processing the normalized spectral curve of the iron mineral using smoothing and extreme point search. The slope calculation is used to process the straight line segments before and after the inflection point to obtain the slope before and after the inflection point. The polarity change parameters of the slope before and after the inflection point are used as the absorption characteristic parameters of the diagnostic band.

[0011] Optionally, the process of processing the spectral curves in the structured pixel spectral database using envelope elimination and piecewise fitting methods to obtain diagnostic band absorption characteristic parameters further includes: The normalized spectral curves of carbonate absorption bands with wavelengths of 2300–2450 nm were obtained by processing the spectrum curves using decontinuum processing. The normalized spectral curve of carbonate was processed by bandwidth extraction and shoulder localization to obtain characteristic bandwidth parameters and shoulder position parameters, which were then used as absorption characteristic parameters for the diagnostic band.

[0012] Optionally, the process of identifying mineral types based on the feature parameters in the extended nodes, introducing virtual background endmembers in linear spectral decomposition, and performing nonlinear correction on the decomposition results to obtain pixel-level mineral categories and relative contents includes: K-Means clustering analysis was used to process the feature parameters in the extended nodes to obtain the initial mineral category discrimination results; The endmember spectral and hyperspectral data were processed by decontinuum processing to obtain the processed endmember matrix and spectral data matrix; After adding virtual background endmembers with all band values ​​of 1 to the processed endmember matrix, linear spectral decomposition using the fully constrained least squares method is performed to obtain the initial abundance results. The initial abundance results were processed using a nonlinear correction formula to obtain the corrected relative mineral content; The initial mineral category determination result and the corrected relative mineral content are used as the pixel-level mineral category and relative content.

[0013] Optionally, after obtaining the pixel-level mineral categories and relative abundances, the process further includes: using inverse mapping to map the one-dimensional array of pixel-level mineral categories and relative abundances to a geographic coordinate system to obtain a mineral spatial distribution map and a quantitative abundance layer.

[0014] The present invention also provides a pixel-scale hyperspectral image mineral information extraction system for implementing the method, the system comprising: The first processing module is used to preprocess the original hyperspectral image data to obtain a reflectance data cube; The second processing module is used to extract spectral curves pixel by pixel based on the reflectance data cube and bind them to spatial coordinates to construct a structured pixel spectral database. The third processing module is used to process the spectral curves in the structured pixel spectral database using envelope elimination and piecewise fitting methods to obtain diagnostic band absorption characteristic parameters, and to process the spectral curves of the broad and gentle absorption bands using slope calculation to obtain morphological characteristic parameters. The diagnostic band absorption characteristic parameters and morphological characteristic parameters are then stored in the extended node. The fourth processing module is used to identify mineral types based on the feature parameters in the extended nodes. It introduces virtual background endmembers in the linear spectral decomposition and performs nonlinear correction on the decomposition results to obtain pixel-level mineral categories and relative contents.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention achieves a paradigm shift of hyperspectral data from the image domain to the attribute domain by constructing a structured pixel spectral database. It employs envelope elimination and piecewise fitting methods to extract local features from diagnostic bands, effectively overcoming the problem of global spectral matching masking weak alteration features. In linear spectral decomposition, a virtual background endmember is introduced to absorb interference from featureless minerals, and nonlinear correction is combined to eliminate mixing errors caused by multiple scattering between mineral particles, significantly improving the accuracy of quantitative inversion of mineral content in single pixels and providing a high-confidence quantitative evidence chain for the prediction of hydrothermal concealed ore bodies. Attached Figure Description

[0016] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the main process control of a pixel-scale hyperspectral image mineral information extraction method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall functional module interaction architecture of a pixel-scale hyperspectral image mineral information extraction system according to an embodiment of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0019] Example 1 To address a series of technical problems in existing technologies, such as over-reliance on global band matching masking weak alteration features, excessive errors in quantitative inversion of single-pixel abundance, and severe limitations imposed by unstructured data on batch analysis by artificial intelligence and automated geological logic deduction, this invention provides a full-stack method and system for extracting mineral information from hyperspectral images based on the pixel scale. This invention is based on a paradigm shift from "image vision" to "spectral database," using the pixel as the smallest unit of attribute calculation. It constructs a highly structured spectral database using hierarchical protocols such as HDF5, implements segmented feature mining and fitting driven by geophysical mechanisms, and innovatively combines a nonlinear hybrid pixel correction model with all-1 virtual endmembers. This achieves fully automated, highly sensitive qualitative identification and high-precision quantitative inversion of alteration mineral information, providing a strong quantitative evidence chain for the prediction of concealed underground ore bodies.

[0020] This method improves the signal-to-noise ratio by preprocessing hyperspectral data, then extracts spectra pixel by pixel and binds them to spatial coordinates, innovatively constructing a structured pixel spectral database based on HDF5 format. It abandons global matching and precisely extracts parameters such as characteristic peaks of hydroxyl Gaussian fitting, carbonate bandwidth, and left and right spectral slopes and inflection points of iron mineral bands based on the physical properties of the target mineral. Finally, in the quantitative inversion stage, a full-band virtual endmember is introduced to absorb interference from featureless minerals, and a nonlinear hybrid correction formula is used to significantly reduce inversion errors. This invention achieves high-precision quantitative interpretation at the single-pixel level, significantly improving the accuracy of mineral exploration prediction and the efficiency of automated analysis of concealed ore bodies in hydrothermal deposits.

[0021] like Figure 1 As shown, this embodiment provides a method for extracting mineral information from hyperspectral images based on the pixel scale, including the following steps: Step 1: High-fidelity preprocessing of hyperspectral image data. Acquire the raw hyperspectral image data and perform image preprocessing on the raw hyperspectral image data to obtain a preprocessed hyperspectral image reflectance data cube.

[0022] Furthermore, the image preprocessing process for the original hyperspectral image data includes: converting the original digital quantization values ​​of the hyperspectral sensor into apparent radiance data using a radiometric calibration model; inverting the apparent radiance data using an atmospheric correction algorithm based on a radiative transfer model to eliminate the nonlinear effects of atmospheric water vapor, aerosols, topographic relief, and solar irradiance on the spectrum, thereby obtaining true surface reflectance data; performing destriping and noise suppression on the true surface reflectance data; projecting the image into the intrinsic feature space using minimum noise separation (MNF) transform; extracting high signal-to-noise ratio components and combining wavelet filtering to remove high-frequency spatial noise; and then reconstructing the preprocessed hyperspectral image data through inverse transform.

[0023] As a specific implementation method in this embodiment, the purpose of this step is to greatly suppress environmental and instrument noise without losing nanoscale spectral characteristics. The spectral data matrix acquired by the original airborne, spaceborne, or ground-based full-spectrum hyperspectral measuring instrument is obtained. First, based on the latest radiative transfer model, rigorous radiometric calibration and atmospheric correction are performed to convert the radiance at the sensor's entrance pupil into true surface reflectance data that has eliminated the effects of atmospheric water vapor, aerosol scattering, and differences in topographic illumination. Second, addressing the strong inter-band correlations and stripe noise commonly found in hyperspectral data, the Minimum Noise Separation Transform (MNF) is used to map the hyperspectral data to the eigenvalue space. Wavelet filtering is introduced to smooth the low eigenvalue bands rich in noise components. Finally, a high signal-to-noise ratio preprocessed reflectance data cube is recovered through inverse MNF transformation, laying an absolutely reliable data foundation for subsequent microscopic wavelength displacement analysis.

[0024] Step 2: Deconstruct the 3D image into a structured pixel spectral library. The preprocessed hyperspectral image data cube is transformed into a pixel spectral library. The complete spectral reflectance curve is extracted for each pixel, and the spatial coordinates of each pixel are strongly bound to its complete spectral vector to construct a pixel spectral database based on a hierarchical data format.

[0025] Furthermore, the process of constructing a pixel spectral database based on a hierarchical data format includes: extracting the two-dimensional spatial coordinate set and the corresponding one-dimensional spectral reflectance vector of the preprocessed image; storing the pixel spectral database in HDF5 data format, and establishing logical inclusion relationships between objects within the file by utilizing its internal hierarchical logical structure and B-tree physical storage method; and objectifying and encapsulating the spatial coordinate system metadata, sensor band metadata, and pixel-level reflectance vector to achieve structured storage that separates content from expression.

[0026] In this specific implementation, the preprocessed 3D hyperspectral image is scanned row by row and column by column, stripping it into individual pixels. For each pixel, a one-dimensional vector of its continuous "reflectance as a function of wavelength" spectral curve is extracted. The two-dimensional spatial geographic coordinates (latitude and longitude or projected coordinates) of each pixel are forcibly bound to its complete spectral vector to construct a pixel spectral database. To cope with the throughput and querying of massive amounts of data, this invention abandons traditional image formats and specifies the use of the HDF5 standard data structure for persistent storage. HDF5 has a powerful hierarchical logical structure and B-tree physical indexing method, which can separate and encapsulate the spatial coordinate system, band metadata, reflectance matrix, and future generated feature parameters in an object-oriented manner. This structure not only greatly improves I / O read efficiency, but also enables hyperspectral data to be retrieved by SQL statements or directly called for training by AI engines, just like traditional relational databases.

[0027] Step 3: Extraction of segmented feature parameters for spectral curve classification based on physical mechanisms. Based on the typical spectral characteristics of the target mineral, segmented feature parameters are extracted from the spectral curves in the pixel spectral database according to classification. For specific minerals, envelope elimination and multi-peak fitting algorithms for specific intervals are used to obtain multi-dimensional structured feature parameters, which are then stored in the extended nodes of the pixel spectral database.

[0028] Furthermore, the process of extracting segmented feature parameters for classifying the spectral curves in the pixel spectral database specifically includes: for the hydroxyl absorption band, extracting the characteristic absorption peak at the wavelength position of 2190-2220nm, and calculating its absorption depth, center position, and peak asymmetry; for the carbonate absorption band, extracting the absorption valley at the wavelength position of 2300-2450nm, and calculating its characteristic bandwidth and shoulder position; for hydration characteristics, extracting the absorption strength index of hydrous minerals in the short-wave to mid-wave infrared range; and for iron minerals, extracting the spectral slope and slope inflection point in the wavelength range of 850-1100nm.

[0029] Furthermore, the process of extracting characteristic absorption peaks at wavelengths of 2190–2220 nm from the hydroxyl absorption band specifically includes: processing the decontinuum-processed characteristic curves using second-order differentiation to find and lock each potential absorption peak; judging the absorption peaks based on the prior diagnostic wavelength positions of muscovite absorption peaks to obtain initial characteristic absorption peaks; performing multi-peak fitting on the initial characteristic absorption peaks using the Gaussian function to obtain fitting values; judging the convergence of the fitting values; if the fitting values ​​do not converge, the characteristic curve is retained; otherwise, the characteristic curve is discarded; extracting the single peak at wavelengths of 2190–2220 nm from the converged curve as characteristic peaks, and calculating the peak area, peak position, peak height, peak center, and half-peak width as structured feature parameters and storing them in the extended library.

[0030] Furthermore, the process of extracting the spectral slope and inflection point of iron minerals specifically includes: calculating the inflection point near the extreme point in the smoothed spectral curve after decontinuum, and the slope of the two straight line segments before and after the inflection point; the formula for calculating the slope of the straight line segment before the inflection point is: The formula for calculating the slope of the straight line segment after the inflection point is: ;in Reflectivity of each band The center wavelength of each band, combined with , The sum of absolute values ​​and polarity changes are used to identify iron mineral anomalies.

[0031] As a specific implementation of this embodiment, this step completely abandons the overall spectral matching method, which is prone to "blinding the view with a leaf" errors. Based on the typical crystal field and molecular vibration characteristics of the target mineral within the ore-forming system, targeted band-wise dissection and extraction are carried out, and the extracted multidimensional parameters are structured and stored in the extended node of the HDF5 database: (1) For hydroxyl (OH) - ) and the absorption band of aluminum silicates (such as muscovite and sericite): The multi-peak analysis algorithm is used to search for the minimum value of the curve after decontinuum processing by second derivative to lock the diagnostic absorption range near 2200nm. The initial characteristic absorption peak is fitted with multiple curves using Gaussian function. If the Gaussian fitting converges, it is determined to be valid and retained; otherwise, it is discarded as interference noise. The single peak parameters in the wavelength range of 2190-2220nm are accurately extracted, including: absorption valley depth (indicating the relative content of minerals), absolute center wavelength position (indicating the degree of isomorphic substitution of aluminum element), peak area, peak height and half peak width (indicating crystallinity). (2) For carbonates (CO3 2- (3) For iron oxide minerals (such as hematite): Iron minerals exhibit broad and gentle spectral characteristics in the 850-1100nm range. This invention innovatively uses second-order differential to find the poles of the smooth curve and thus determine the slope "inflection point" of the curve. The slope of the straight line segment before the inflection point and the slope of the straight line segment after the inflection point are calculated respectively. The iron staining intensity is quantitatively characterized by the slope polarity reversal and the slope difference, which effectively overcomes the problem of the difficulty in accurately locating the broad and gentle absorption valley. (4) For hydration characteristics: Determine the absorption strength index of free water and crystal water of hydrated minerals in the short-wave to mid-wave infrared range.

[0032] Step 4: Intelligent Mineral Species Identification and High-Precision Quantitative Abundance Inversion. Based on the structured feature parameters in the extended nodes, mineral species identification and pixel abundance inversion are performed, outputting the mineral category and its relative abundance for each pixel, and finally generating a mineral spatial distribution map and a quantitative abundance layer.

[0033] Furthermore, the process of mineral species identification and pixel abundance inversion specifically includes: initial mineral species identification by calling the structured feature parameters in the extended nodes through K-Means clustering analysis, spectral angle mapping, or a deep learning-based classifier; quantitative inversion of mineral content based on linear decomposition of the spectrum, before linear decomposition, the endmember spectra and hyperspectral data are decontinuum-processed; in the endmember matrix of the linear spectral decomposition, virtual endmembers with all bands equal to 1 are forcibly added, the virtual endmembers representing mineral background endmembers with no obvious absorption characteristics in all spectral decomposition bands; after completing the linear decomposition with virtual endmembers, in order to address the impact of spectral nonlinear mixing caused by multiple scattering between surface mineral particles on the mineral content inversion results, a nonlinear correction formula is introduced to numerically correct the linear inversion abundance, so that the inverted mineral content is closer to the true pixel physical abundance.

[0034] As a specific implementation of this embodiment, supported by the massive structured feature parameter matrix extracted in the preceding steps, a clustering analysis algorithm (such as the K-Means algorithm) or a benchmark comparison classification using spectral angle mapping (SAM) and spectral information divergence (SID) from a known standard spectral library is employed. Subsequently, the core quantitative abundance inversion stage begins: this invention performs initial inversion based on the linear decomposition theory of the spectrum. To eliminate the problem of unsolvable equations or negative abundance caused by minerals such as quartz, which have no obvious absorption characteristics in the short-wave infrared region, this method performs strict decontinuum envelope elimination on the hyperspectral data and endmember matrix, and then artificially and cleverly introduces a virtual background endmember with a reflectance value of 1 in all bands into the endmember matrix. After obtaining the preliminary abundance of various minerals through linear decomposition, a nonlinear empirical correction formula (covering polynomial and logarithmic term compensation) is further introduced to specifically address the internal multiple photon scattering and nonlinear mixing effects caused by close contact between mineral particles, performing numerical stretching and correction. Finally, the precise relative abundance of a specific mineral within each pixel is output, and the one-dimensional array is inversely mapped back to the geographic coordinate system to generate a high-precision spatial distribution map of minerals and a quantitative abundance layer.

[0035] As another specific implementation of this embodiment, for the application scenario of hydrothermal deposit prospecting, the process of making geological prospecting judgments based on the spatial distribution and quantitative inversion results of specific minerals specifically includes: analyzing the characteristic peak parameter data of muscovite and sericite with wavelengths around 2200nm in the inversion layer; calculating the spatial variation law of the center wavelength movement and the spatial variation law of the absorption depth of the characteristic peaks; when the wavelength of sericite on the ground shows a progressive change from greater than 2210nm in the background area to 2202-2206nm in the core area, and the absorption depth shows a spatial zonation law of gradually increasing from the center of the rock mass to the periphery, it is determined that there is a hydrothermal blind ore body or porphyry deposit in the deep part of the study area.

[0036] Example 2 To address a series of persistent problems in current remote sensing geology exploration practices, such as excessively high mineral exploration costs, severe distortion in single-pixel mineral composition inversion, global matching masking weak alteration information, and the inability to integrate massive unstructured images with advanced machine learning algorithms for large-scale analysis.

[0037] like Figure 2 As shown, based on the above information extraction method, this invention also designs an information extraction system with a multi-node topology architecture, including the following precisely coupled sub-processing modules: The first processing module (preprocessing module) is used to acquire the original hyperspectral image data, and sequentially perform radiometric calibration, atmospheric correction, and high-frequency noise suppression processing based on wavelet filtering and minimum noise separation on the original hyperspectral image data, and output a preprocessed data cube with a high signal-to-noise ratio.

[0038] The first processing module (preprocessing module) acts as the "water purifier" of the system. The configuration acquisition unit is used to load the datasets transmitted by UAVs or spacecraft; the radiation correction subunit uses an embedded large lookup table (LUT) to perform physical dimension conversion; then the noise reduction subunit schedules multiple CPU threads to perform the arduous MNF matrix covariance eigenvalue decomposition and wavelet filtering in the low-frequency feature space, and finally outputs an absolutely denoised and clean surface reflectance data stream to the bus.

[0039] The second processing module (database construction module) is used to transform the preprocessed data cube from the image domain into a spectral library in the attribute domain, extract spectral curves pixel by pixel and bind their spatial two-dimensional coordinate system to their complete spectrum, and construct a pixel spectral database based on a hierarchical protocol.

[0040] The second processing module (database construction module) includes a data extraction unit, which slices the 3D preprocessed matrix according to the pixel row spacing; the core is its embedded database read and write API engine, which strictly follows the HDF5 open source cross-platform protocol to establish a huge B-Tree index system from Metadata to Pixel_Data and then to the reserved feature extension root directory.

[0041] The third processing module (segmented micro-feature extraction module) is used to perform envelope elimination and multiple derivative peak finding calculation on the curves in the pixel spectrum database according to the diagnostic spectral absorption characteristics of the target mineral, extract the absorption depth, peak asymmetry, characteristic bandwidth and slope inflection point by category, output multi-dimensional structured feature parameters and store them as extended fields in the pixel spectrum database.

[0042] The third processing module (segmented micro-feature extraction module) includes: a second-order differential peak finding unit, used to locate potential weak absorption valleys in the spectrum; a Gaussian fitting convergence unit, used to perform rigorous mathematical fitting on the sericite characteristic band of 2190–2220 nm, preserving the high-confidence pixel spectral features of the converged state; and a derivative slope calculation unit, used to perform difference operations on the 850–1100 nm interval to extract iron staining characteristic parameters.

[0043] The third processing module (segmented micro-feature extraction module) has a continuum elimination subunit responsible for spectral normalization and shell stripping; a second-order differential Gaussian fitting unit is specifically preset and anchored near wavelengths of 2190nm–2220nm and other key infrared positions, using multiple partial derivative approximations to search for true micro-absorption depressions and extract their depth, symmetry, and center wavelength shift values; simultaneously, an iron stain slope logic controller is specifically servo-controlled in the 850nm–1100nm range, strictly applying the differential formula of the slope tangent at the left and right inflection points to amplify the gradient differences into sensitive numerical indicators. All calculated feature vectors are uniformly submitted and stored in the extended database node reserved in the second module.

[0044] The fourth processing module (intelligent judgment and inversion verification module) is used to read the structured feature parameters, drive the linear spectral decomposition algorithm based on all-1 background endmember compensation and the nonlinear hybrid model of scattering correction, calculate the pixel-level mineral classification and quantitative content inversion results, and finally generate a mineral spatial distribution map and quantitative layer that can be overlaid and analyzed by geographic information systems.

[0045] The fourth processing module (intelligent analysis and inversion verification module) incorporates a fully constrained least squares unmixed linear matrix operation array with an all-1 background endmember feed compensator, as well as a calibrator for the subsequent serial operation of the quadratic logarithmic divergence nonlinear compensation formula. This module also includes a geological mineralization experience threshold analysis machine for examining the spatial displacement distribution logic of sericite in the 2202-2206nm band. Its outermost layer includes a GIS coordinate fusion output unit, ultimately transforming the dry floating-point parameters into intuitive, geographic coordinate projection-based, layered, and high-precision mineral exploration targeted mapping.

[0046] This invention achieves the following technical advantages: 1. Structured Representation Lays the Foundation for AI Industrial Applications: It creatively breaks down the decades-old cubic barrier of hyperspectral imagery, reconstructing it into a pixel-based spectral library similar to a relational database. The powerful tree-structured management and cross-platform compatibility of the HDF5 file format enable unified encapsulation, slice reading, and rapid retrieval of complex spectral metadata and extracted feature parameters, reaching hundreds of gigabytes in size. This directly allows the system to seamlessly integrate with random forests, support vector machines, and even deep learning frameworks within the Python ecosystem, endowing hyperspectral technology with unprecedented intelligent analysis and automated batch processing capabilities.

[0047] 2. The "segmented microscopy" mechanism completely overcomes the problem of weak feature concealment: This method abandons the global comparison of hundreds of bands required by the SAM algorithm. It adopts "segmented feature parameter extraction" with clear physical meaning. For example, it mines the left / right slope parameters only for iron minerals in the corresponding diagnostic band, or it locks the Gaussian absorption depth at 2200nm only for muscovite. This high-precision local microscopy mechanism greatly filters out spectral noise interference from background-irrelevant bands, and the sensitivity of identifying trace alteration minerals increases exponentially, making it impossible for weak surface alteration halos caused by hidden deep ore bodies to escape detection.

[0048] 3. Eliminating substrate interference and achieving ultimate quantitative inversion accuracy at the single-pixel level: This invention creatively adds a "virtual background endmember with all bands being 1" to the traditional linear spectral decomposition model, solving the problem of mathematical singular matrices caused by the absence of characteristic minerals after decontinuum removal, thus making component decomposition of the entire pixel domain possible.

[0049] 4. Deeply integrated geological deposit model, directly indicating prospecting target areas: This method is specifically designed to capture the fine wavelength center shift of sericite (muscovite), accurately depicting the progressive spectral evolution from 2202nm at the hydrothermal center to the peripheral potassic zone greater than 2210nm. This technology can directly determine the geological significance (distance from the heat source and mineral enrichment) represented by the spatial variations of wavelength and peak depth, providing the most direct and conclusive evidence consistent with geological dynamics for confirming the existence of concealed porphyry copper-gold deposits underground.

[0050] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for extracting mineral information from hyperspectral images based on pixel scale, characterized in that, Includes the following steps: The original hyperspectral image data is preprocessed to obtain a reflectance data cube; Based on the reflectance data cube, spectral curves are extracted pixel by pixel and bound to spatial coordinates to construct a structured pixel spectral database; The spectral curves in the structured pixel spectral database are processed using envelope elimination and piecewise fitting methods to obtain diagnostic band absorption characteristic parameters. The diagnostic band absorption characteristic parameters and morphological characteristic parameters are then stored together in the extended node. Based on the feature parameters in the extended nodes, mineral types are identified. Virtual background endmembers are introduced in the linear spectral decomposition, and the decomposition results are nonlinearly corrected to obtain pixel-level mineral categories and relative contents.

2. The method for extracting mineral information from hyperspectral images based on pixel scale according to claim 1, characterized in that, The preprocessing of raw hyperspectral image data includes: The original hyperspectral image data was processed using radiometric calibration and atmospheric correction to obtain surface reflectance data; The surface reflectance data is processed using minimum noise separation transform and wavelet filtering to obtain a reflectance data cube.

3. The method for extracting mineral information from hyperspectral images based on pixel scale according to claim 1, characterized in that, The process of extracting spectral curves pixel by pixel based on the reflectance data cube and binding them to spatial coordinates to construct a structured pixel spectral database includes: The reflectance data cube is processed by scanning row by row and column by column to obtain a one-dimensional vector of single-pixel spectral reflectance; The one-dimensional vector of spectral reflectance of a single pixel is bound and encapsulated with two-dimensional spatial geographic coordinates using the HDF5 hierarchical data format to obtain a pixel spectral database with a B-tree index structure.

4. The method for extracting mineral information from hyperspectral images based on pixel scale according to claim 1, characterized in that, The process of processing the spectral curves in the structured pixel spectral database using envelope elimination and piecewise fitting methods to obtain diagnostic band absorption characteristic parameters includes: The normalized spectral vector is obtained by processing the one-dimensional vector of the spectral reflectance of the single pixel corresponding to the hydroxyl absorption band with wavelengths of 2190–2220 nm using decontinuum processing. The set of potential absorption peak positions is obtained by processing the normalized spectral vector using second-order differentiation; Gaussian multi-peak fitting is used to process the set of potential absorption peak positions to obtain characteristic peak parameters in the converged state. These characteristic peak parameters are used as absorption characteristic parameters for the diagnostic band. The characteristic peak parameters include peak area, peak position, peak height, peak center, and half-width.

5. The method for extracting mineral information from hyperspectral images based on pixel scale according to claim 1, characterized in that, The process of obtaining the morphological feature parameters includes: The normalized spectral curves of iron minerals were obtained by processing the spectral curves corresponding to the wavelength range of 850–1100 nm using decontinuum processing. The inflection point positions were obtained by processing the normalized spectral curve of the iron mineral using smoothing and extreme point search. The slope calculation is used to process the straight line segments before and after the inflection point to obtain the slope before and after the inflection point. The polarity change parameters of the slope before and after the inflection point are used as the absorption characteristic parameters of the diagnostic band.

6. The method for extracting mineral information from hyperspectral images based on pixel scale according to claim 1, characterized in that, The process of processing the spectral curves in the structured pixel spectral database using envelope elimination and piecewise fitting methods to obtain diagnostic band absorption characteristic parameters also includes: The normalized spectral curves of carbonate absorption bands with wavelengths of 2300–2450 nm were obtained by processing the spectrum curves using decontinuum processing. The normalized spectral curve of carbonate was processed by bandwidth extraction and shoulder localization to obtain characteristic bandwidth parameters and shoulder position parameters, which were then used as absorption characteristic parameters for the diagnostic band.

7. The method for extracting mineral information from hyperspectral images based on pixel scale according to claim 1, characterized in that, The process of identifying mineral types based on the feature parameters in the extended nodes, introducing virtual background endmembers in linear spectral decomposition and performing nonlinear correction on the decomposition results to obtain pixel-level mineral categories and relative contents includes: K-Means clustering analysis was used to process the feature parameters in the extended nodes to obtain the initial mineral category discrimination results; The endmember spectral and hyperspectral data were processed by decontinuum processing to obtain the processed endmember matrix and spectral data matrix; After adding virtual background endmembers with all band values ​​of 1 to the processed endmember matrix, linear spectral decomposition using the fully constrained least squares method is performed to obtain the initial abundance results. The initial abundance results were processed using a nonlinear correction formula to obtain the corrected relative mineral content; The initial mineral category determination result and the corrected relative mineral content are used as the pixel-level mineral category and relative content.

8. The method for extracting mineral information from hyperspectral images based on pixel scale according to claim 7, characterized in that, After obtaining the pixel-level mineral categories and relative abundances, the process also includes: using inverse mapping to map the one-dimensional array of pixel-level mineral categories and relative abundances to a geographic coordinate system, resulting in a mineral spatial distribution map and a quantitative abundance layer.

9. A system for extracting mineral information from hyperspectral images based on pixel scale, characterized in that, The system for implementing the method of claim 1, wherein the system comprises: The first processing module is used to preprocess the original hyperspectral image data to obtain a reflectance data cube; The second processing module is used to extract spectral curves pixel by pixel based on the reflectance data cube and bind them to spatial coordinates to construct a structured pixel spectral database. The third processing module is used to process the spectral curves in the structured pixel spectral database using envelope elimination and piecewise fitting methods to obtain diagnostic band absorption characteristic parameters, and to process the spectral curves of the broad and gentle absorption bands using slope calculation to obtain morphological characteristic parameters. The diagnostic band absorption characteristic parameters and morphological characteristic parameters are then stored in the extended node. The fourth processing module is used to identify mineral types based on the feature parameters in the extended nodes. It introduces virtual background endmembers in the linear spectral decomposition and performs nonlinear correction on the decomposition results to obtain pixel-level mineral categories and relative contents.

Images