A method for hyperspectral endmember calibration and mineral analysis of martian meteorite constraints

Abstract

The present application relates to the technical field of Mars meteorite mineralogy, hyperspectral endmember calibration, Mars orbit hyperspectral mineral identification and intelligent interpretation, and specifically relates to a Mars meteorite constrained hyperspectral endmember calibration and mineral analysis method, comprising: preparing a Mars meteorite sample into a polished section sample, establishing a coordinate system and dividing to obtain a plurality of calibration regions; obtaining mineral phase confirmation results of the calibration regions in different ways from microscopic hyperspectral acquisition for each calibration region; collecting microscopic hyperspectral data in the calibration regions, converting the microscopic hyperspectral data into the coordinate system; forming a Mars meteorite hyperspectral endmember prior library, and establishing a training data set and an orbit data set to be analyzed; training to obtain a hyperspectral mineral analysis model; and using the hyperspectral mineral analysis model to obtain mineral phase categories corresponding to the orbit data; the present application can improve label credibility, model interpretability and cross-scale generalization ability.

Description

Technical Field

[0001] This invention relates to the fields of Martian meteorite mineralogy, hyperspectral endmember calibration, Martian orbit hyperspectral mineral identification and intelligent interpretation, and specifically to a Martian meteorite-constrained hyperspectral endmember calibration and mineral analysis method. Background Technology

[0002] Hyperspectral analysis of Martian minerals typically relies on orbital hyperspectral data, such as CRISM and OMEGA visible-near-infrared data, to determine mineral types through spectral band positions, absorption depths, continuum removal, endmember matching, or machine learning models. This type of method is valuable for global or regional mineral mapping on Mars, but its training labels and endmember libraries usually come primarily from publicly available spectral libraries, ground-based analog samples, expert visual interpretations, or empirical annotations in orbital images.

[0003] Martian meteorites are currently the only Martian material samples that can be directly obtained and subjected to joint petrographic, mineralogical, and spectroscopic characterization in the laboratory. Martian meteorites contain olivine, pyroxene, plagioclase glass or plagioclase, serpentine, chlorite, montmorillonite, carbonates, sulfates, and other altered minerals, reflecting Martian magmatism, water-rock reactions, and low-temperature hydrothermal processes. If a spatial correspondence can be established between the mineralogically confirmed mineral phases in Martian meteorites and the microscopic hyperspectral curves collected from the same region, hyperspectral endmembers of Martian meteorites with a clear material origin and reliable labeling can be obtained. The distribution of mineral phases and their typical hyperspectral lines in the microscopic regions of Martian meteorites are shown in the attached figure. Figure 2 As shown.

[0004] The shortcomings of existing technologies are twofold: firstly, Martian hyperspectral mineral identification largely relies on orbital spectral interpretation or publicly available endmember libraries, lacking constraints from mineral phases derived from actual Martian material samples; secondly, meteorite mineralogy research and orbital hyperspectral models are typically separate, and a coherent workflow from meteorite micro-region mineral identification and microscopic endmember calibration to orbital pixel analysis has not yet been established. Therefore, there is an urgent need to establish an endmember construction and training method based on Martian meteorite microscopic mineral phases to improve label reliability, model interpretability, and cross-scale generalization ability. Summary of the Invention

[0005] In view of the above problems, the present invention provides a hyperspectral endmember calibration and mineral analysis method constrained by Martian meteorites, which solves the technical problems of insufficient label credibility, model interpretability and cross-scale generalization ability in the prior art.

[0006] This invention provides a method for hyperspectral endmember calibration and mineral analysis constrained by Martian meteorites, comprising the following steps: Step S1: Prepare the Martian meteorite sample into a polished sheet sample, establish a coordinate system for the sample and divide the coordinate system into multiple regions to be calibrated; Step S2: For each region to be calibrated, obtain the mineral phase confirmation results of the region to be calibrated using a method different from microscopic hyperspectral acquisition; The mineral phase confirmation results include the mineral phase name, mineral phase category, spatial location, and confidence level of the mineral endmembers in the region to be calibrated; Step S3: Collect microscopic hyperspectral data in the area to be calibrated, and convert the microscopic hyperspectral data into the coordinate system; compare and analyze the mineral phase confirmation results with the corresponding converted microscopic hyperspectral data and record them together to form a Martian meteorite hyperspectral endmember prior library; The Martian meteorite hyperspectral endmember prior library includes high-confidence endmember information obtained through comparative analysis; Step S4: Based on the high-confidence end-member information, the public spectral library, and the Mars orbit hyperspectral data, perform unified preprocessing to obtain the training dataset and the orbit dataset to be analyzed; Step S5: Establish a loss function based on classification loss, endmember prior loss and region consistency loss, and train the model based on the loss function and the training dataset to obtain a hyperspectral mineral analysis model; Step S6: Use a hyperspectral mineral analysis model to process the orbital dataset to be analyzed and obtain the mineral phase categories corresponding to the orbital data.

[0007] Preferably, in step S2, the step of obtaining the mineral phase confirmation result of the region to be calibrated using a method different from microscopic hyperspectral acquisition specifically includes: Particle boundaries, cracks, and symbiotic relationships are identified based on backscattered electron imaging, and major elemental composition is confirmed based on scanning electron microscopy energy dispersive spectroscopy or electron probe microanalysis. Crystal structure and functional groups were obtained based on Raman spectroscopy, X-ray diffraction, and micro Fourier transform infrared spectroscopy. Based on one or more of the aforementioned particle boundaries, cracks, symbiotic relationships, major element composition, crystal structure, and functional groups, the mineral phase name, mineral phase category, spatial location, and confidence level are ultimately determined as the mineral phase confirmation result.

[0008] Preferably, in step S3, the comparative analysis step specifically includes: For each region to be calibrated with confirmed mineral phases, pure pixels or small regions located inside the mineral phase and far from mineral boundaries and fissures are extracted as endmembers, and the microscopic hyperspectral features of the endmembers are calculated. The microscopic hyperspectral features include: average spectrum, standard deviation, signal-to-noise ratio, and absorption characteristic parameters; the absorption characteristic parameters may include: absorption band center wavelength, absorption depth, absorption width, and absorption area after continuous spectrum removal; Based on the comparison between the microscopic hyperspectral features and the mineral phase confirmation results, it is determined whether the spectral features match the mineral phase confirmation results. If they match, the current endmember is determined as a high-confidence endmember; otherwise, the current endmember is marked as a low-confidence endmember or a verification endmember.

[0009] Preferably, step S4 specifically includes: Step S4-1: Resample the high-confidence end-member information, the public spectral library, and the Mars orbit hyperspectral data to a unified wavelength grid; Step S4-2: Normalize, remove the continuous spectrum, extract absorption feature parameters or process the first derivative of the resampled hyperspectral data to obtain standardized hyperspectral data. Step S4-3: Perform scale transfer on the standardized hyperspectral data, and divide the preprocessed data into training set, validation set, test set and orbital dataset to be analyzed.

[0010] Preferably, in step S4-3, the scale transfer step includes: (1) Hybrid pixel simulation, which involves linearly or nonlinearly mixing endmembers of multiple different mineral categories in different proportions; (2) Particle size effect simulation, simulating the influence of particle size, porosity and surface roughness on the spectrum; (3) Signal-to-noise ratio degradation: Gaussian noise, Poisson noise or measured noise that conforms to the characteristics of orbital data is added to the endmember spectrum to make the signal-to-noise ratio of the endmember comparable to that of the orbital data. (4) Set up geometric perturbations to simulate spectral changes under different incident angles, emission angles and phase angles; (5) Spectral smoothing: The endmember spectra are smoothed appropriately to match the spectral resolution of the orbital data.

[0011] Preferably, in step S5, the expression for the loss function is:

[0012] in, Represents the total loss function. These are the weighting coefficients. Represents classification loss. Indicates the prior loss of the endmember. This indicates the loss of regional consistency. The expression for classification loss is:

[0013] in, The number of training samples, This represents the total number of mineral categories. Indicates the first Each sample in category The real labels on it The model predicts the first Each sample belongs to category The probability of; Endmember prior loss constrains orbital sample features to approximate meteorite endmember prototypes:

[0014] in, For the hyperspectral mineral analysis model for the first The fusion feature vector extracted from each sample Indicates the first The true mineral category label of each sample Indicates the first The prior prototype center of each sample corresponding to the category. Represents the square of the Euclidean distance; Regional consistency loss constraint aligns the centers of similar features in different regions:

[0015] in For reference area, Indicates the region Not a reference area. Indicates the first minerals in the region Feature center in Indicates the first minerals in the region The feature center in.

[0016] Preferably, the method further includes step S7: correcting and verifying the mineral phase category corresponding to the orbital data, and updating the endmember prior library based on the verification result, specifically including: For the classification probability output by the model forward propagation, when the model classification probability is high but the endmember similarity is low or the similarity of multiple mineral endmembers is close, the pixel is marked as a low-confidence or mixed pixel; the low-confidence or mixed pixels are manually reviewed and labeled, and the reviewed data is added to the endmember prior library.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: (1) More reliable label source: This invention uses the mineral phase confirmation results as the source of microscopic hyperspectral label, avoiding the inference of label based solely on infrared spectral morphology, and improving the mineralogical credibility of training samples.

[0018] (2) Fully utilize prior Martian meteorite evidence: This invention uses Martian meteorites as the source of end-member calibration for Martian material samples, enabling a more direct connection between the training end-member and real Martian rocks and minerals. The scope of protection is not dependent on a single device: Mineral phase identification can be accomplished by multiple methods, including morphology, elemental analysis, mineral chemistry, structural spectroscopy, and comprehensive expert interpretation, adapting to different laboratory conditions and sample types.

[0019] (3) Reduce the risk of circular labeling: Infrared spectroscopy is only used for consistency checks and model input, and is not used as the sole basis for labeling, which helps to avoid self-verification using similar spectral information.

[0020] (4) Improve model interpretability: The analysis results not only provide categories, but also trace back to Martian meteorite hyperspectral endmembers, mineral phase confirmation evidence and similarity indicators, which facilitates scientific interpretation and manual verification.

[0021] (5) The model is replaceable and the process is reusable: The core of this invention lies in end-member calibration and prior construction. The analysis model can be replaced according to the task and is applicable to different orbital sensors, different mineral categories and different spatial scales.

[0022] (6) Facilitates incremental updates: New Martian meteorite samples, new mineral phase confirmation evidence, or new orbit verification samples can all be incorporated into the end-member prior library to continuously improve recognition performance. Attached Figure Description

[0023] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0024] Figure 1 The overall flowchart of hyperspectral endmember calibration and Martian mineral analysis based on Martian meteorite mineral phase constraints provided by this invention.

[0025] Figure 2 The images provided by this invention are microscopic mineral phase images of Martian meteorites and schematic diagrams of hyperspectral lines of typical minerals.

[0026] Figure 3 A flowchart illustrating the process of identifying mineral phases and assigning hyperspectral microscopic end-member tags to Martian meteorites provided by this invention.

[0027] Figure 4 The flowchart for unified processing of multi-source spectral data and construction of training sample sets provided by this invention is shown.

[0028] Figure 5 Example figures of Mars orbit hyperspectral in-situ analysis data, spectral features, and mineral identification results provided by this invention. Figure 6 Flowchart for Martian hyperspectral mineral identification and result verification based on endmember prior constraints provided by this invention Figure 7 This is a schematic diagram illustrating the multi-task joint loss optimization of the endmember prior constraint model provided by the present invention.

[0029] Figure 8 The flowchart shows the hyperspectral endmember calibration and mineral analysis method for Martian meteorite constraint provided by this invention. Detailed Implementation

[0030] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0031] To illustrate the effectiveness of the method proposed in this invention, the following detailed description of the above technical solution is provided through a specific embodiment, such as... Figure 8 As shown, a hyperspectral endmember calibration and mineral analysis method constrained by Martian meteorites is disclosed. The specific implementation steps are as follows: Step S1: Prepare the Martian meteorite sample into a polished sheet sample, establish a coordinate system for the sample and divide the coordinate system into multiple regions to be calibrated; In this step, the present invention selects Martian meteorite samples for preparation. The Martian meteorite sample may be a basaltic Martian meteorite, an orthopyroxene Martian meteorite, a Martian meteorite containing altered minerals, or other types of Martian meteorites. The selected Martian meteorite sample is prepared into a polished sheet sample suitable for microscopic analysis and microscopic hyperspectral acquisition, specifically, it may be prepared into a polished thin sheet, a polished block, or other sample form with a flat polished surface.

[0032] After sample preparation, this invention establishes a unified coordinate system for the sample. The establishment of the coordinate system can be achieved through one or more of the following methods: setting artificial markers on the sample surface or edge and determining the coordinate system based on these markers; determining the coordinate system based on mineral grain feature points in a microscopic image; establishing a spatial reference to determine the coordinate system using the sample's edge contour, crack texture, or other identifiable features; or determining the coordinate system using feature points identified by an automatic image registration algorithm.

[0033] This invention establishes a coordinate system to provide a unified spatial reference for subsequent mineral phase identification, microscopic hyperspectral acquisition, and spatial registration between different data sources.

[0034] After establishing the coordinate system, the sample surface is divided into multiple areas to be calibrated based on the mineral distribution characteristics, microstructure, and analytical requirements of the sample.

[0035] Specifically, multiple regions of a sample can be identified based on mineral grain boundaries, lithofacies regions, or scientific objectives of interest. Each region is assigned a unique region number and its spatial location and boundary extent in the coordinate system are recorded.

[0036] Step S2: For each region to be calibrated, obtain the mineral phase confirmation results of the region to be calibrated using a method different from microscopic hyperspectral acquisition; The methods different from microscopic hyperspectral acquisition include scanning electron microscopy, electron probe microscopy, polarizing microscope, and X-ray diffraction to acquire sample information; The mineral phase confirmation results include the mineral phase name, mineral phase category, spatial location, and confidence level of the mineral endmembers in the region to be calibrated; In this step, the present invention identifies the mineral phases of each region to be calibrated. This mineral phase identification employs an independent mineralogical characterization method, distinct from microscopic hyperspectral acquisition, to avoid relying solely on the infrared spectral characteristics to infer the label, thereby improving the reliability of the label's origin.

[0037] The steps for obtaining mineral phase confirmation results of the region to be calibrated using a method different from microscopic hyperspectral acquisition specifically include: Particle boundaries, cracks, and symbiotic relationships are identified based on backscattered electron imaging, and major elemental composition is confirmed based on scanning electron microscopy energy dispersive spectroscopy or electron probe microanalysis. Crystal structure and functional groups were obtained based on Raman spectroscopy, X-ray diffraction, and micro Fourier transform infrared spectroscopy. Based on one or more of the aforementioned particle boundaries, cracks, symbiotic relationships, major element composition, crystal structure, and functional groups, the mineral phase name, mineral phase category, spatial location, and confidence level are ultimately determined as the mineral phase confirmation result.

[0038] In one specific embodiment, one or more of the above-mentioned characterization methods can be used for the same area to be calibrated. For example, backscattered electron imaging can be used to identify the boundaries, morphology, and contrast of mineral grains to preliminarily determine the mineral phase type and spatial distribution. Subsequently, SEM-EDS or electron probe microanalysis can be used to confirm the elemental composition of the area, such as the content and proportion of major elements like Mg, Fe, Si, Ca, Al, and Na, thereby determining whether the mineral is olivine, pyroxene, plagioclase, feldspar glass, or other silicate minerals. For hydrous altered minerals, carbonates, or sulfates that are prone to fine-graining or mixing, Raman spectroscopy, XRD, or micro-FTIR can be used to further confirm their structural characteristics and functional group information, such as characteristic peaks of OH groups, carbonates, or sulfates. For areas with complex mineral coexistence relationships or important lithofacies backgrounds, lithofacies observation under a polarizing microscope and mineral assemblage analysis can be combined to comprehensively determine the mineral phase confirmation results.

[0039] The mineral phase confirmation results include at least the following information: the mineral phase name of the end-member in the area to be calibrated, such as olivine, low-Ca pyroxene, high-Ca pyroxene, plagioclase, maskelynite, serpentine, chlorite, smectite, carbonate, sulfate, etc.; mineral phase category, which can be classified as primary magnesian silicates, feldspar and framework silicates, hydrous altered silicates, low-temperature hydrothermal minerals, carbonates and sulfates, oxides and accessory minerals, etc.; spatial location, including the coordinate range, pixel position or vector boundary of the mineral phase in the coordinate system; confidence level, which is divided into high confidence, medium confidence, low confidence or pending review level according to the quantity, consistency and mineral boundary clarity of the mineral phase confirmation evidence.

[0040] This invention marks areas with clear boundaries, high purity, and consistent results from different verification methods as high-confidence mineral phases, which serve as the source of subsequent high-confidence end-member labels. Areas with mixed adjacent minerals, unclear alteration boundaries, or contradictions between different pieces of evidence are marked as low-confidence or areas awaiting verification. These areas are not directly included in the high-confidence end-member training samples, but are retained as mixed end-members, boundary samples, or weakly supervised samples.

[0041] Step S3: Collect microscopic hyperspectral data in the area to be calibrated, and convert the microscopic hyperspectral data into the coordinate system; compare and analyze the mineral phase confirmation results with the corresponding converted microscopic hyperspectral data and record them together to form a Martian meteorite hyperspectral endmember prior library; The Martian meteorite hyperspectral endmember prior library includes high-confidence endmember information obtained through comparative analysis; In this step, microscopic hyperspectral data is acquired in the region to be calibrated. The acquisition of the microscopic hyperspectral data is carried out in the same microscopic region as the mineral phase confirmation or in a spatially registrable region to ensure that an accurate spatial correspondence can be established between the mineral phase confirmation results and the microscopic hyperspectral curves.

[0042] The microscopic hyperspectral data can be acquired using a microscopic hyperspectral imaging system, which can operate in the visible, near-infrared, short-wave infrared, mid-infrared, or combined bands.

[0043] Because there may be differences in viewpoint, scale, rotation, or translation between the microscopic hyperspectral images and the microscopic images used for mineral facies identification, spatial registration is required. Therefore, after the microscopic hyperspectral data acquisition is completed, the microscopic hyperspectral data needs to be transformed into the coordinate system established in the aforementioned steps.

[0044] In some embodiments, the spatial registration may employ one or more of the following methods: affine or perspective transformation based on artificially marked points; automatic registration based on mineral grain boundaries, fracture textures, or feature points; non-rigid registration based on mutual information, normalized cross-correlation, or other image registration algorithms; or direct mapping using coordinates recorded by the position encoder of the microscopic platform.

[0045] After registration, the mineral phase confirmation results and the microscopic hyperspectral data achieve spatial correspondence in the same coordinate system.

[0046] The mineral phase confirmation results were compared and analyzed with the corresponding converted microscopic hyperspectral data, and the results were recorded together.

[0047] The comparative analysis steps specifically include: for each region to be calibrated with confirmed mineral phases, extracting pure pixels or small regions located within the mineral phase and far from mineral boundaries and fissures as endmembers; calculating the microscopic hyperspectral characteristics of the endmembers, including: average spectrum, standard deviation, signal-to-noise ratio, and absorption characteristic parameters; the absorption characteristic parameters may include: absorption band center wavelength, absorption depth, absorption width, absorption area after continuous spectrum removal, etc. Based on the comparative analysis of the microscopic hyperspectral characteristics and the confirmed mineral phases, determine whether the spectral characteristics match the confirmed mineral phases. If they match, the current endmember is identified as a high-confidence endmember; otherwise, the current endmember is marked as a low-confidence endmember or a verification endmember.

[0048] This invention pairs the mineral phase confirmation results of the endmembers with the corresponding converted microscopic hyperspectral data as an endmember record, and saves all endmember records as the Martian meteorite hyperspectral endmember prior library.

[0049] In some embodiments, the microscopic hyperspectral data may include spectral curves of endmembers.

[0050] In some embodiments, the Martian meteorite hyperspectral endmember prior library can be organized in a hierarchical structure: the first layer is sample-level information, recording the Martian meteorite type, petrological background, and sample processing method; the second layer is micro-area-level information, recording the microscopic region, mineral phase confirmation results, and spatial boundaries; the third layer is endmember-level information, recording the endmember spectral curves, purity, confidence level, and preprocessing version; and the fourth layer is migration-level information, recording the band matching, scale conversion, and category merging rules used when migrating microscopic endmembers to orbital hyperspectral data.

[0051] In some embodiments, the endmember prior library can be expanded and updated as new samples are added, new mineral phases are identified, or new microscopic hyperspectral data are added.

[0052] Step S4: Based on the high-confidence end-member information, the public spectral library, and the Mars orbit hyperspectral data, perform unified preprocessing to obtain the training dataset and the orbit dataset to be analyzed; Step S4-1: Resample the high-confidence end-member information, the public spectral library, and the Mars orbit hyperspectral data to a unified wavelength grid; Because Martian meteorite microscopic hyperspectral endmembers, publicly available laboratory spectral libraries such as RELAB and USGS, and Martian orbital hyperspectral data such as CRISM and OMEGA data are acquired using different sensors, their band ranges, wavelength sampling intervals, and spectral resolutions differ. Therefore, in this step, all spectral data are resampled to a unified wavelength grid. This resampling can employ linear interpolation, cubic spline interpolation, or Gaussian convolution resampling methods. The resampled wavelength grid can be selected to match the orbital hyperspectral sensor or to cover a common band range across all data sources.

[0053] Preferably, the wavelength range of the resampling to the unified wavelength grid covers 1.021μm to 2.635μm, and the number of bands after resampling is 425, which is achieved by using a linear interpolation method.

[0054] Step S4-2: Normalize, remove the continuous spectrum, extract absorption feature parameters or process the first derivative of the resampled hyperspectral data to obtain standardized hyperspectral data. In this step, the processing procedure of the present invention includes: spectral normalization, which can be achieved by maximum value normalization, L2 norm normalization, or specified band normalization, to make the spectral amplitudes of different data sources comparable; continuous spectrum removal, fitting the continuous spectral envelope of the spectrum to highlight absorption features; absorption feature parameter extraction, calculating the center wavelength, depth, width, and area of ​​key absorption bands; first-order derivative or second-order derivative processing to enhance subtle changes and inflection point features in the spectrum; and finally obtaining standardized hyperspectral data.

[0055] Step S4-3: Perform scale transfer on the standardized hyperspectral data, and divide the preprocessed data into training set, validation set, test set and orbital dataset to be analyzed.

[0056] In this step, the scale transfer process includes: pixel blending simulation, which involves linearly or nonlinearly blending endmembers of multiple mineral categories in different proportions to simulate common mineral blending in orbital pixels; grain size effect simulation, which simulates the influence of particle size, porosity, and surface roughness on the spectrum; signal-to-noise ratio (SNR) degradation, which adds Gaussian noise, Poisson noise, or measured noise that matches the characteristics of the orbital data to the endmember spectrum, making the SNR of the endmember comparable to that of the orbital data; setting geometric perturbations to simulate spectral changes under different incident angles, emission angles, and phase angles; and spectral smoothing, which moderately smooths the endmember spectrum to match the spectral resolution of the orbital data. Through these processes, the microscopic endmembers are made closer to the orbital observation samples in the feature space, improving the model's cross-scale generalization ability.

[0057] Preferably, the signal-to-noise ratio (SNR) degradation step in the scale migration step specifically includes: adding Gaussian white noise of corresponding intensity to the endmember spectrum based on the typical SNR of CRISM orbital data (in this embodiment, the SNR range is set to 10dB to 50dB).

[0058] In some embodiments, the preprocessed data can be randomly divided into a training set, a validation set, a test set, and a track dataset to be analyzed according to a preset ratio.

[0059] Step S5: Establish a loss function based on classification loss, endmember prior loss and region consistency loss, and train the model based on the loss function and the training dataset to obtain a hyperspectral mineral analysis model; In this step, the present invention first establishes a joint loss function based on classification loss, endmember prior loss, and region consistency loss; The classification loss is used to supervise the model's learning to correctly classify the input spectrum into the corresponding mineral category. In this embodiment, the classification loss uses the cross-entropy loss function, whose expression is:

[0060] in, Represents the cross-entropy loss function. The number of training samples, The total number of mineral categories (in this example, it covers 15 categories including low-calcium pyroxene, high-calcium pyroxene, olivine, plagioclase, serpentine, chlorite, montmorillonite, carbonates, sulfates, gypsum, analcime, epidote, muscovite, and prehnite). Indicates the first Each sample in category The actual label on the screen (with a value of 0 or 1). The model predicts the first Each sample belongs to category The probability of.

[0061] The endmember prior loss is used to constrain the feature representation extracted by the model to be consistent with the endmembers of the corresponding categories in the Martian meteorite hyperspectral endmember prior library. Specifically, for each mineral category, the class prototype or class center is calculated based on all high-confidence endmembers of that category in the endmember prior library.

[0062] This embodiment uses the mean of the endmember feature vectors as the category prototype. Let the mean of the endmember feature vectors be the first value. For minerals, the feature set after mapping of their external prior samples (including high-confidence endmembers of Martian meteorites and samples from the USGS and RELAB spectral libraries) by a feature encoder is: Then the first The prior prototype-centered expression of a class is:

[0063] in, Indicates the first The prior prototype center of the class, For the first The number of prior samples of the class, For prior sample index, Indicates the first Features of class-prior samples. During model training, training samples of the same class are required to be close to the endmember prototypes of that class in the feature space, while samples of different classes are separated from each other in the feature space. The expression for the endmember prior loss is:

[0064] in, Indicates the prior loss of the endmember. For the hyperspectral mineral analysis model for the first The fusion feature vector extracted from each sample Indicates the first The true mineral category label of each sample Indicates the first The prior prototype center of each sample corresponding to the category. This represents the square of the Euclidean distance. The endmember prior loss function directly constrains the feature representation of the orbital samples to converge towards meteorite endmembers with mineralogical truth.

[0065] Regional consistency loss is used to improve the model's generalization ability and prediction consistency across different Martian regions, observation conditions, or data sources. Considering that similar minerals in different Martian regions (such as Holden Crater HC, Nelly Groove NF, and Utopia Plain UP in this example) can exhibit spectral differences (i.e., domain shifts) due to variations in dust cover, grain size, or weathering degree, this step introduces regional consistency loss to mitigate this shift. The expression for regional consistency loss is:

[0066]

[0067] in, Indicates the loss of regional consistency. For reference area, Indicates the area Not a reference area. Indicates the first minerals in the region Feature center in Indicates the first minerals in the region Feature center in For the region Middle The number of samples in the set of class samples. Indicates the area Middle A set of sample classes. The region consistency loss function encourages the feature centers of similar minerals in different regions to move closer together, thereby improving the model's representation stability in cross-regional scenarios.

[0068] Ultimately, the total loss function of this invention is a weighted combination of the classification loss, endmember prior loss, and region consistency loss described above, expressed as:

[0069] in, Represents the total loss function. and These are the weighting coefficients for the prior constraint loss and the regional consistency loss, respectively (in this embodiment, they are optimized based on the performance of the validation set, preferably...). The value ranges from 0.01 to 0.1. The value range is from 0.001 to 0.01, and can be specifically set to... , This is used to balance the relationship between mineral classification, mineralogical a priori guidance, and cross-regional characterization alignment.

[0070] This invention does not limit the network structure of the hyperspectral mineral analysis model. In one specific embodiment, the hyperspectral mineral analysis model adopts a hybrid convolutional neural network (CNN) and Transformer architecture. Specifically, the CNN branch uses a one-dimensional convolutional layer (kernel size set to 3, stride set to 1) to extract local spectral features (such as absorption peak position and width); the Transformer branch divides the input spectrum into several spectral tokens and uses a multi-head self-attention mechanism (number of heads set to 8) to model global inter-spectral dependencies; the features output from the two branches are concatenated and then dimensionality-reduced by a fully connected layer to obtain the final fused features. .

[0071] The training dataset is input into the model in batches. Forward propagation calculates the prediction results and the total loss function, and backpropagation updates the model parameters. The optimizer used during training can be stochastic gradient descent (SGD), Adam, AdamW, or other optimization algorithms. This embodiment uses the Adam optimizer, with an initial learning rate of 0.001, a batch size of 64, and 100 training epochs, employing a cosine annealing decay strategy. At each training epoch, the model performance is evaluated using the validation set, recording classification accuracy, precision for each class, recall, and F1 score. An early stopping strategy is used; training stops when the validation set loss no longer decreases for 10 consecutive epochs, resulting in a trained hyperspectral mineral analysis model.

[0072] Step S6: Use a hyperspectral mineral analysis model to process the orbital dataset to be analyzed and obtain the mineral phase categories corresponding to the orbital data.

[0073] In this step, the orbital dataset to be analyzed is input into the hyperspectral mineral analysis model. For each pixel or spectral sample to be analyzed, the model forward propagates and outputs the mineral phase category and the corresponding classification probability.

[0074] Preferably, in some embodiments of the present invention, step S7 is further included: correcting and verifying the mineral phase category corresponding to the orbital data, and updating the endmember prior library based on the verification result.

[0075] In this step, a confidence threshold is set for the classification probability output by the model's forward propagation (in this embodiment, a classification probability greater than 0.7 is considered a high-confidence prediction, 0.4 to 0.7 is considered a medium-confidence prediction, and less than 0.4 is considered a low-confidence prediction). When the model's classification probability is high but the endmember similarity is low (e.g., the spectral angular distance is greater than 0.1) or the similarity of multiple mineral endmembers is close, the pixel is marked as a low-confidence or mixed pixel. The low-confidence or mixed pixels are manually reviewed and labeled, and the reviewed data is added to the endmember prior library.

[0076] The updated endmember prior library can be used for the next round of model training or incremental learning. During incremental learning, the first three feature extraction layers of the hyperspectral mineral analysis model can be frozen, and the last two fully connected classification layers can be fine-tuned using only new samples. This allows the model to absorb new samples while maintaining existing knowledge, gradually improving its recognition accuracy and cross-region generalization ability.

[0077] Figure 1 The overall technical flow of this invention is illustrated. This flow includes four stages: Martian meteorite sample preparation and microscopic region management, mineral phase identification and microscopic hyperspectral acquisition, microscopic hyperspectral endmember calibration and endmember library construction, and orbital hyperspectral identification and result verification and updating. Among these, mineral phase identification results precede hyperspectral label assignment; infrared or hyperspectral features are primarily used for consistency checks, anomaly detection, and endmember screening, rather than as the sole source of labels. This flow enables a closed loop from micro-region mineral phase identification of Martian meteorite samples to orbital hyperspectral mineral analysis.

[0078] Figure 2 This figure shows backscattered microscopic images of mineral phases and hyperspectral lines of typical minerals in a Martian meteorite sample. The backscattered images are used to illustrate the grain structure, mineral phase distribution, microscopic regions, and local mineral boundaries of the Martian meteorite sample; the typical spectral lines are used to show the differences in the spectral response of different mineral phases in the visible-near-infrared band. This figure demonstrates that Martian meteorite samples can provide microscopic hyperspectral endmembers with a clear lithofacies background and mineralogical origin, providing a data foundation for subsequent endmember labeling and endmember prior library construction.

[0079] Figure 3 This figure illustrates the process of mineral phase identification, spatial registration, and endmember labeling in microscopic regions of Martian meteorites. First, regions to be calibrated are selected and mineral phases are identified. Then, by integrating morphological and lithofacies relationships, elemental composition and mineralogy, structural spectral and functional group information, and results from automated mineral analysis or expert interpretation, a mineral phase identification result is formed. Next, the mineral phase identification result is spatially registered with the microscopic hyperspectral data, and high-confidence endmembers, mixed endmembers, and samples requiring verification are selected from the registered regions. This figure emphasizes that microscopic hyperspectral endmember labels are assigned after confirmation by independent mineralogical evidence, rather than solely inferred from infrared spectral morphology.

[0080] Figure 4This diagram illustrates the workflow for unifying the processing of Martian meteorite hyperspectral endmembers, publicly available laboratory spectral libraries, and orbital hyperspectral data to construct a training sample set. Data from different sources first enter the endmember library and spectral data processing module. After unified preprocessing, spectral resampling, scale migration simulation, and label system mapping, a dataset suitable for model training, validation, and orbital image analysis is formed. This figure illustrates the key steps in migrating microscopic Martian meteorite endmembers to orbital-scale hyperspectral analysis tasks.

[0081] Figure 5 This figure illustrates an example of in-situ or orbital hyperspectral analysis data of Mars, including hyperspectral image regions, pixel spectral curves, and mineral analysis results. It demonstrates that the training sample set not only includes Martian meteorite hyperspectral endmembers and publicly available laboratory spectral libraries, but can also incorporate manually confirmed regions or in-situ analysis samples from orbital hyperspectral data. Incorporating this type of data enhances the model's adaptability to real-world Martian observation scenarios and provides a reference for mineral identification, confidence level assessment, and anomaly screening in orbital images.

[0082] Figure 6 This paper illustrates the workflow for Martian hyperspectral mineral identification and result verification by incorporating Martian meteorite hyperspectral endmember priors. After inputting orbital hyperspectral endmembers or pixels to be analyzed, the analysis model or matching algorithm calculates the class probability, endmember similarity, and confidence level, followed by rule fusion and posterior correction. For high-confidence samples, a mineral class map and endmember matching results are output; for low-confidence or anomalous regions, manual verification is performed, and the verification results are fed back to the Martian meteorite hyperspectral endmember prior library to achieve incremental updates of class centers, similarity thresholds, and training samples.

[0083] Figure 7 This paper illustrates a joint optimization approach for training a Mars hyperspectral mineral recognition model, simultaneously incorporating mineral classification tasks, endmember category prototype constraints, and cross-regional consistency constraints. The mineral classification head outputs mineral category probabilities and calculates the classification loss. The endmember category prototype, composed of a Mars meteorite hyperspectral endmember prior library and an external prior spectral library, constrains model features to maintain consistency with the corresponding mineral category endmembers. The cross-regional consistency constraint improves the feature stability of samples from different Martian regions in identifying the same type of mineral. During model training, the classification loss, prior loss, and regional consistency loss are combined to form the total loss function, and the model parameters are updated through backpropagation.

[0084] While the specific embodiments of the present invention depict actions or steps in a particular order, this should be understood as requiring such actions or steps to be performed in the shown specific order or sequential order, or requiring all illustrated actions or steps to be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations. The above descriptions are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention.

[0085] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for hyperspectral endmember calibration and mineral analysis constrained by Martian meteorites, characterized in that, Includes the following steps: Step S1: Prepare the Martian meteorite sample into a polished sheet sample, establish a coordinate system for the sample and divide the coordinate system into multiple regions to be calibrated; Step S2: For each region to be calibrated, obtain the mineral phase confirmation results of the region to be calibrated using a method different from microscopic hyperspectral acquisition; The mineral phase confirmation results include the mineral phase name, mineral phase category, spatial location, and confidence level of the mineral endmembers in the region to be calibrated; Step S3: Collect microscopic hyperspectral data in the area to be calibrated, and transform the microscopic hyperspectral data into the coordinate system; The mineral phase confirmation results are compared and analyzed with the corresponding converted microscopic hyperspectral data and recorded together to form a Martian meteorite hyperspectral endmember prior library. The Martian meteorite hyperspectral endmember prior library includes high-confidence endmember information obtained through comparative analysis; Step S4: Based on the high-confidence end-member information, the public spectral library, and the Mars orbit hyperspectral data, perform unified preprocessing to obtain the training dataset and the orbit dataset to be analyzed; Step S5: Establish a loss function based on classification loss, endmember prior loss and region consistency loss, and train the model based on the loss function and the training dataset to obtain a hyperspectral mineral analysis model; Step S6: Use a hyperspectral mineral analysis model to process the orbital dataset to be analyzed and obtain the mineral phase categories corresponding to the orbital data.

2. The method for hyperspectral endmember calibration and mineral analysis constrained by Martian meteorites according to claim 1, characterized in that, In step S2, the step of obtaining the mineral phase confirmation results of the region to be calibrated using a method different from microscopic hyperspectral acquisition specifically includes: Particle boundaries, cracks, and symbiotic relationships are identified based on backscattered electron imaging, and major elemental composition is confirmed based on scanning electron microscopy energy dispersive spectroscopy or electron probe microanalysis. Crystal structure and functional groups were obtained based on Raman spectroscopy, X-ray diffraction, and micro Fourier transform infrared spectroscopy. Based on one or more of the aforementioned particle boundaries, cracks, symbiotic relationships, major element composition, crystal structure, and functional groups, the mineral phase name, mineral phase category, spatial location, and confidence level are ultimately determined as the mineral phase confirmation result.

3. The method for hyperspectral endmember calibration and mineral analysis constrained by Martian meteorites according to claim 2, characterized in that, Step S3, specifically the step of converting the microscopic hyperspectral data to the coordinate system, includes: One or more of the following methods: affine or perspective transformation based on manually marked points; automatic registration based on mineral grain boundaries, fracture textures, or feature points; non-rigid registration based on mutual information, normalized cross-correlation, or other image registration algorithms; or direct mapping using coordinates recorded by the position encoder of the microscopic platform.

4. The method for hyperspectral endmember calibration and mineral analysis constrained by Martian meteorites according to claim 3, characterized in that, In step S3, the comparative analysis specifically includes: For each region to be calibrated with confirmed mineral phases, pure pixels or small regions located inside the mineral phase and far from mineral boundaries and fissures are extracted as endmembers, and the microscopic hyperspectral features of the endmembers are calculated. The microscopic hyperspectral features include: average spectrum, standard deviation, signal-to-noise ratio, and absorption characteristic parameters; the absorption characteristic parameters may include: absorption band center wavelength, absorption depth, absorption width, and absorption area after continuous spectrum removal; Based on the comparison between the microscopic hyperspectral features and the mineral phase confirmation results, it is determined whether the spectral features match the mineral phase confirmation results. If they match, the current endmember is determined as a high-confidence endmember; otherwise, the current endmember is marked as a low-confidence endmember or a verification endmember.

5. The method for hyperspectral endmember calibration and mineral analysis constrained by Martian meteorites according to claim 4, characterized in that, Step S4 specifically includes: Step S4-1: Resample the high-confidence end-member information, the public spectral library, and the Mars orbit hyperspectral data to a unified wavelength grid; Step S4-2: Normalize, remove the continuous spectrum, extract absorption feature parameters or process the first derivative of the resampled hyperspectral data to obtain standardized hyperspectral data. Step S4-3: Perform scale transfer on the standardized hyperspectral data, and divide the preprocessed data into training set, validation set, test set and orbital dataset to be analyzed.

6. The method for hyperspectral endmember calibration and mineral analysis constrained by Martian meteorites according to claim 5, characterized in that, In step S4-3, the scale transfer step includes: (1) Hybrid pixel simulation, which involves linearly or nonlinearly mixing endmembers of multiple different mineral categories in different proportions; (2) Particle size effect simulation, simulating the influence of particle size, porosity and surface roughness on the spectrum; (3) Signal-to-noise ratio degradation: Gaussian noise, Poisson noise or measured noise that conforms to the characteristics of orbital data is added to the endmember spectrum to make the signal-to-noise ratio of the endmember comparable to that of the orbital data. (4) Set up geometric perturbations to simulate spectral changes under different incident angles, emission angles and phase angles; (5) Spectral smoothing: The endmember spectra are smoothed appropriately to match the spectral resolution of the orbital data.

7. The method for hyperspectral endmember calibration and mineral analysis constrained by Martian meteorites according to claim 6, characterized in that, In step S5, the expression for the loss function is: in, Represents the total loss function. These are the weighting coefficients. Represents classification loss. Indicates the prior loss of the endmember. This indicates the loss of regional consistency. The expression for classification loss is: in, The number of training samples, This represents the total number of mineral categories. Indicates the first Each sample in category The real labels on The model predicts the first Each sample belongs to category The probability of; Endmember prior loss constrains orbital sample features to approximate meteorite endmember prototypes: in, For the hyperspectral mineral analysis model for the first The fusion feature vector extracted from each sample Indicates the first The true mineral category label of each sample Indicates the first The prior prototype center of each sample corresponding to the category. Represents the square of the Euclidean distance; Regional consistency loss constraint aligns the centers of similar features in different regions: in, For reference area, Indicates the region Not a reference area. Indicates the first minerals in the region Feature center in Indicates the first minerals in the region The feature center in.

8. The method for hyperspectral endmember calibration and mineral analysis constrained by Martian meteorites according to claim 7, characterized in that, It also includes step S7, which involves correcting and verifying the mineral phase category corresponding to the orbital data, and updating the endmember prior library based on the verification results, specifically including: For the classification probability output by the model forward propagation, when the model classification probability is high but the endmember similarity is low or the similarity of multiple mineral endmembers is close, the pixel is marked as a low-confidence or mixed pixel; the low-confidence or mixed pixels are manually reviewed and labeled, and the reviewed data is added to the endmember prior library.

Images