A mapping model establishing method

Abstract

The application discloses a mapping model establishing method, relates to the technical field of mineral detection, and comprises the following steps: obtaining a plurality of construction waste soil samples; synchronously collecting hyperspectral reflectivity data, close-range image data and X-ray diffraction quantitative analysis data for each construction waste soil sample; and the X-ray diffraction quantitative analysis data contains target mineral content true value, and the collection geometric parameters corresponding to the hyperspectral reflectivity data contain an incident zenith angle, an emission zenith angle and a phase angle. In the application, the hyperspectral reflectivity data, the collection geometric parameters and the Hapke model inversion are combined to generate a physical latent variable spectrum based on the inversion single scattering albedo, and a spectral index library is constructed based on the physical latent variable spectrum, sensitive spectral indices and sensitive bands are screened, so that the influence of the construction waste soil particle roughness, multiple scattering, local water content state and abnormal wave band on mineral content prediction is reduced.

Description

Technical Field

[0001] This invention relates to the field of mineral detection technology, and in particular to a method for establishing a mapping model. Background Technology

[0002] The resource utilization of construction waste soil usually requires first obtaining the mineral composition information of soil samples, and then determining its applicable direction based on the mineral content. The existing process mostly completes the mineral content analysis through on-site sampling, laboratory testing, manual comparison and experience grading, and uses the test results for raw material classification, stockpile management and recycling decisions, thus forming the need for subsequent rapid identification and accurate modeling.

[0003] Current methods for detecting mineral content in construction waste soil mainly rely on quantitative analysis in laboratories, which has a long testing cycle and is difficult to adapt to rapid on-site grading. Simple image recognition is difficult to distinguish soil particles with similar colors but different mineral compositions. Direct hyperspectral modeling is susceptible to interference from multiple scattering and particle heterogeneity. Sample-level mineral true values ​​are difficult to correspond to local areas of the image, resulting in low accuracy of mineral content prediction and poor model stability. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a mapping model establishment method.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a mapping model establishment method, characterized in that it includes:

[0006] Multiple construction waste soil samples were acquired. For each construction waste soil sample, hyperspectral reflectance data, close-up image data, and X-ray diffraction quantitative analysis data were collected simultaneously. The X-ray diffraction quantitative analysis data included the true value of the target mineral content. The acquisition geometric parameters corresponding to the hyperspectral reflectance data included the incident zenith angle, the emission zenith angle, and the phase angle.

[0007] Based on hyperspectral reflectance data and collected geometric parameters, the physical parameters of each construction waste soil sample were inverted using the Hapke model to obtain the inverted single scattering albedo corresponding to each wavelength, and the physical latent variable spectrum corresponding to each construction waste soil sample was generated based on the inverted single scattering albedo.

[0008] A spectral index library is constructed based on physical latent variable spectroscopy. The spectral indices in the library are correlated with the true values ​​of target mineral content to obtain the sensitive spectral indices and sensitive bands corresponding to the target minerals. A physical teacher model is established using the sensitive spectral indices as inputs and the true values ​​of target mineral content as outputs.

[0009] Grain boundary segmentation and texture abrupt change segmentation are performed on close-range image data to obtain multiple image segmentation units, and the color features, texture features, grain morphology features and spatial location features of each image segmentation unit are extracted.

[0010] Based on the predicted value of target mineral content, physical latent variable spectrum, and area ratio of each image segmentation unit output by the physics teacher model, a mineral content assignment label is generated for each image segmentation unit. The mineral content assignment labels of all image segmentation units in the same construction waste soil sample are aggregated by area weighting to satisfy the consistency constraint with the true value of target mineral content.

[0011] The image student model is trained by taking the color features, texture features, grain morphology features and spatial location features of each image segmentation unit as inputs, and the mineral content assignment labels and physical latent variable spectral data as joint supervision information.

[0012] The image student model performs area-weighted aggregation of the predicted mineral content of all image segmentation units within the same construction waste soil sample to obtain the sample-level aggregated predicted mineral content. The physical teacher model and the image student model are then updated based on the error between the sample-level aggregated predicted mineral content and the true value of the target mineral content, thus obtaining a mapping model that maps close-up image data to the mineral content of construction waste soil.

[0013] As a further description of the above technical solution:

[0014] The Hapke model was used to invert the physical parameters of each construction waste soil sample, obtaining the inverted single-scattering albedo for each wavelength. Based on the inverted single-scattering albedo, the physical latent variable spectrum corresponding to each construction waste soil sample was generated, including:

[0015] The hyperspectral reflectance data are arranged according to sample number and wavelength to form a sample reflectance sequence;

[0016] The incident zenith angle, emitted zenith angle, and phase angle are converted into the incident zenith angle cosine value, emitted zenith angle cosine value, and phase angle variable corresponding to the Hapke model, and the incident zenith angle cosine value, emitted zenith angle cosine value, and phase angle variable are made to establish a correspondence between the incident zenith angle cosine value, emitted zenith angle cosine value, and phase angle variable and the reflectance of each wavelength in the sample reflectance sequence.

[0017] For each wavelength position in the sample reflectance sequence, the single scattering albedo is used as the variable to be inverted, and the phase function, backscattering function, and multiple scattering function are used as calculation terms in the Hapke model. The Hapke model is used to calculate the reflectance residual between the reflectance and the hyperspectral reflectance.

[0018] According to the rule of minimizing reflectance residual, the single scattering albedo corresponding to each wavelength position is iteratively solved to obtain the inverted single scattering albedo corresponding to each wavelength.

[0019] Connect the inverted single scattering albedo corresponding to each wavelength in wavelength order to form a single scattering albedo spectral curve.

[0020] The single-scattering albedo spectral curves were band-consistentized, and wavelengths that did not participate in the inversion were removed. The single-scattering albedo spectral curves after band-consistencyization were used as the physical latent variable spectra of the construction waste soil samples.

[0021] As a further description of the above technical solution:

[0022] The single-scattering albedo corresponding to each wavelength position is iteratively solved, and the single-scattering albedo spectral curves are standardized for band consistency, including:

[0023] The reflectance in the sample reflectance sequence located in the instrument noise band, saturation band, missing band, and reflectance mutation band is marked as invalid reflectance, and the wavelength position corresponding to the invalid reflectance is marked as invalid wavelength position;

[0024] Based on the hyperspectral reflectance corresponding to the effective wavelength position, the initial range of single scattering albedo and the iterative search step size are set, and the initial range of values ​​is established in correspondence with the effective wavelength position.

[0025] For each effective wavelength position, the single-scattering albedo is adjusted in the direction of decreasing reflectivity residual in each round, and the candidate single-scattering albedo and the corresponding reflectivity residual obtained in each iteration are recorded;

[0026] When any one of the following conditions is met—that the change in reflectivity residual between two adjacent iterations is less than a preset residual change threshold or that the number of iterations reaches a preset maximum number of iterations—the iterative solution for the corresponding effective wavelength position is stopped.

[0027] The candidate single-scattering albedo with the smallest reflectance residual at the time of stopping iteration is determined as the inversion single-scattering albedo at the corresponding effective wavelength position;

[0028] The inverted single scattering albedo corresponding to all effective wavelength positions is connected in wavelength order, and the invalid wavelength positions are retained as invalid markers, generating a physical latent variable spectrum with the correspondence between sample number, effective wavelength position and invalid wavelength position.

[0029] As a further description of the above technical solution:

[0030] A spectral index library is constructed based on the physical latent variable spectrum. The spectral indices in this library are then correlated with the true values ​​of target mineral content for feature filtering. This yields the sensitive spectral indices and sensitive bands corresponding to the target minerals. A physical teacher model is then established using these sensitive spectral indices as input and the true values ​​of target mineral content as output. This model includes:

[0031] The physical latent variable spectra of each construction waste soil sample are arranged according to sample number and wavelength order to form a physical latent variable spectral matrix;

[0032] Based on the inversion of single scattering albedo corresponding to any two wavelength positions in the physical latent variable spectral matrix, a dual-band ratio index, a normalized difference index, and a difference index are constructed, and an index relationship is established between the constructed indices and the corresponding wavelength positions.

[0033] Based on the inversion of single scattering albedo corresponding to any three wavelength positions in the physical latent variable spectral matrix, a three-band combination index is constructed, and an index relationship is established between the three-band combination index and the corresponding three wavelength positions.

[0034] The dual-band ratio index, normalized difference index, difference index, and three-band combination index are merged according to the sample number to generate a spectral index library that corresponds to the true value of the target mineral content.

[0035] The correlation degree between each spectral index in the spectral index library and the true value of the target mineral content is calculated, and the spectral index correlation ranking results are generated in descending order of correlation degree.

[0036] Based on the spectral index correlation ranking results, spectral indices with correlation scores below a preset correlation score threshold are removed, and collinearity screening is performed on the remaining spectral indices to obtain the sensitive spectral indices and sensitive bands corresponding to the target minerals.

[0037] Using the sensitive spectral index as the input variable and the true value of the target mineral content as the output variable, the regression model was trained, and the trained regression model was used as the physics teacher model.

[0038] As a further description of the above technical solution:

[0039] The correlation between each spectral index in the spectral index library and the true value of the target mineral content is calculated, and collinearity is screened out based on the spectral index correlation ranking results, including:

[0040] According to the target mineral type, the spectral indices in the spectral index library that have a sample correspondence with the true value of the content of the same target mineral are divided into a candidate index set;

[0041] For each spectral index in the candidate index set, calculate its linear correlation, rank correlation, and information correlation with the true value of the target mineral content, and generate a comprehensive correlation based on the linear correlation, rank correlation, and information correlation.

[0042] The spectral indices in the candidate index set are arranged in descending order of comprehensive correlation, and the spectral index correlation ranking results are generated.

[0043] Spectral indices with a comprehensive correlation score lower than a preset correlation threshold are removed from the spectral index correlation ranking results to obtain a preliminary set of spectral indices.

[0044] Calculate the index similarity between any two spectral indices in the initial screening spectral index set. When the index similarity between two spectral indices is higher than the preset similarity threshold, retain the spectral index with a higher overall correlation to the true value of the target mineral content.

[0045] The spectral indices retained after collinearity removal are defined as sensitive spectral indices, the wavelength positions corresponding to the sensitive spectral indices are defined as sensitive bands, and the sensitive spectral indices, sensitive bands, predicted values ​​of target mineral content, and spectral fragments of physical latent variables corresponding to the sensitive bands are used as the output of the physics teacher model.

[0046] As a further description of the above technical solution:

[0047] Grain boundary segmentation and texture abrupt change segmentation are performed on close-range image data to obtain multiple image segmentation units. The color features, texture features, grain morphology features, and spatial location features of each image segmentation unit are extracted, including:

[0048] Close-up image data are collected according to sample number, and the background area, shadow area and non-soil area in the close-up image data are masked to obtain the effective image area of ​​the sample.

[0049] Based on the color gradient change and brightness gradient change in the effective image region of the sample, candidate lines of particle boundaries are extracted, and the candidate lines of particle boundaries are connected by breakpoints and closed regions are sorted to obtain the particle boundary segmentation results.

[0050] Based on the gray-level co-occurrence variation, local texture direction variation and texture intensity abrupt change location in the effective image region of the sample, texture abrupt change segmentation line is extracted, and the texture abrupt change segmentation line is fused with the particle boundary segmentation result to obtain image segmentation unit;

[0051] Calculate the percentage of pixels in the effective image region of each image segmentation unit, and use the percentage of pixels as the area percentage of the image segmentation unit;

[0052] For each image segmentation unit, extract the color mean, color variance, and color distribution histogram to form color features;

[0053] For each image segmentation unit, texture orientation, texture density, and texture contrast are extracted to form texture features;

[0054] For each image segmentation unit, the boundary length, region area, aspect ratio, roundness, and adjacency relationship are extracted to form particle morphology features, and spatial location features are formed based on the center coordinates of the image segmentation unit in the effective image region of the sample.

[0055] As a further description of the above technical solution:

[0056] Based on the predicted target mineral content output by the physics teacher model, the physical latent variable spectrum, and the area proportion of each image segmentation unit, a mineral content assignment label is generated for each image segmentation unit. The mineral content assignment labels of all image segmentation units within the same construction waste soil sample are then aggregated using area weighting to ensure consistency with the true target mineral content, including:

[0057] According to the sample number, retrieve the true value of the target mineral content, the predicted value of the target mineral content, the physical latent variable spectrum, and the area ratio of all image segmentation units corresponding to the same construction waste soil sample;

[0058] The inversion single scattering albedo response value corresponding to the sensitive band in the physical latent variable spectrum is used as the mineral response benchmark, and the color features, texture features and grain morphology features of each image segmentation unit are converted into unit image response content.

[0059] Based on the degree of matching between the unit image response content and the mineral response benchmark, the mineral response weight corresponding to each image segmentation unit is generated, and the mineral response weight is normalized according to the sum of the mineral response weights of all image segmentation units in the same sample.

[0060] Based on the normalized mineral response weights, the area proportion of each image segmentation unit, and the target mineral content prediction value output by the physics teacher model, an initial mineral content assignment label is generated for each image segmentation unit.

[0061] The initial mineral content assignment labels within the same construction waste soil sample are weighted and aggregated according to their area proportion to obtain the sample-level mineral content label aggregation value.

[0062] Based on the difference between the aggregated value of the sample-level mineral content label and the true value of the target mineral content, the initial mineral content assignment labels are proportionally corrected to obtain the mineral content assignment label corresponding to each image segmentation unit.

[0063] As a further description of the above technical solution:

[0064] Mineral response weights are generated for each image segmentation unit based on the degree of matching between the unit image response content and the mineral response benchmark. The initial mineral content assignment labels are then corrected based on the difference between the aggregated value of the sample-level mineral content labels and the true value of the target mineral content, including:

[0065] The color features, texture features, and grain morphology features of each image segmentation unit are scaled to obtain color response sub-vectors, texture response sub-vectors, and grain morphology response vectors.

[0066] The color response subvector, texture response subvector, and grain shape response subvector are connected in a preset order to generate the unit image response content corresponding to each image segmentation unit.

[0067] The inversion single scattering albedo response value corresponding to the sensitive band is extracted from the physical latent variable spectrum, and the inversion single scattering albedo response value is matched with the target mineral content prediction value output by the physical teacher model to generate the mineral response benchmark corresponding to the target mineral.

[0068] Calculate the response difference value between the unit image response content of each image segmentation unit and the mineral response benchmark, and determine the matching degree of each image segmentation unit relative to the target mineral according to the response difference value in ascending order;

[0069] The mineral response weights corresponding to each image segmentation unit are generated based on the matching degree, and the mineral response weights of all image segmentation units in the same construction waste soil sample are added together to obtain the total sample weight.

[0070] Using the total sample weight as the normalization benchmark, the mineral response weight corresponding to each image segmentation unit is normalized to obtain the normalized mineral response weight.

[0071] According to the sample number, obtain the initial mineral content allocation label and area ratio corresponding to all image segmentation units within the same construction waste soil sample;

[0072] Multiply the initial mineral content assignment label corresponding to each image segmentation unit by the area ratio to obtain the area-weighted label value corresponding to each image segmentation unit;

[0073] The sample-level mineral content label aggregate value is obtained by summing all area-weighted label values ​​within the same construction waste soil sample.

[0074] The difference between the aggregated value of the sample-level mineral content label and the true value of the target mineral content is calculated to obtain the sample-level label correction difference;

[0075] Based on the normalized mineral response weight and area ratio of each image segmentation unit, the sample-level label correction difference is allocated to each image segmentation unit in the same construction waste soil sample to obtain the label correction amount corresponding to each image segmentation unit.

[0076] The initial mineral content assignment label and label correction value corresponding to each image segmentation unit are merged to obtain the mineral content assignment label corresponding to each image segmentation unit. The mineral content assignment labels of all mineral content in the same construction waste soil sample are then aggregated by area weighting to meet the consistency constraint with the true value of the target mineral content.

[0077] As a further description of the above technical solution:

[0078] The image student model performs area-weighted aggregation of the mineral content predictions output by all image segmentation units within the same construction waste soil sample to obtain a sample-level aggregated mineral content prediction. The physics teacher model and the image student model are then updated based on the error between the sample-level aggregated mineral content prediction and the true target mineral content. This includes:

[0079] According to the sample number, retrieve all image segmentation units corresponding to the same construction waste soil sample, the area ratio of each image segmentation unit, the color features, texture features, particle morphology features and spatial location features of each image segmentation unit, the true value of the target mineral content and the physical latent variable spectrum;

[0080] The color features, texture features, grain morphology features and spatial location features of each image segmentation unit are input into the image student model to obtain the predicted value of the mineral content of each image segmentation unit, and the correspondence between the predicted value of the mineral content of the unit and the image segmentation unit is maintained.

[0081] Based on the area proportion of each image segmentation unit, the predicted mineral content of all units in the same construction waste soil sample is weighted and summarized to obtain the sample-level aggregated predicted mineral content.

[0082] Calculate the aggregation prediction error between the aggregated predicted value of mineral content at the sample level and the true value of target mineral content, and allocate the aggregation prediction error to each image segmentation unit within the same construction waste soil sample to generate the unit error write-back amount;

[0083] The mapping parameters corresponding to color features, texture features, particle morphology features, and spatial location features in the image student model are updated based on the unit error write-back amount.

[0084] The output correction parameters between the sensitive spectral index and the predicted value of the target mineral content in the physics teacher model are updated based on the aggregated prediction error. When the aggregated prediction error meets the preset error conditions, a mapping model from close-up image data to the mineral content of construction waste soil is output.

[0085] As a further description of the above technical solution:

[0086] The aggregation prediction error between the aggregated predicted value of mineral content at the sample level and the true value of the target mineral content is calculated. This aggregation prediction error is then allocated to the unit error write-back amount generated by each image segmentation unit within the same construction waste soil sample. Based on the unit error write-back amount and the aggregation prediction error, the image student model and the physics teacher model are updated, including:

[0087] According to the sample number, obtain the sample-level aggregated predicted value of mineral content, the true value of target mineral content, the predicted value of unit mineral content corresponding to each image segmentation unit, the mineral content allocation label, the area ratio, and the normalized mineral response weight.

[0088] The difference between the predicted value of the sample-level mineral content and the true value of the target mineral content is calculated to obtain the sample-level aggregation error;

[0089] The difference between the predicted mineral content value of each image segmentation unit and the mineral content assignment label is calculated to obtain the unit prediction bias.

[0090] Based on the area ratio, normalized mineral response weight and unit prediction bias of each image segmentation unit, an error allocation coefficient is generated for each image segmentation unit.

[0091] The sample-level aggregation error is allocated to each image segmentation unit within the same construction waste soil sample according to the error allocation coefficient, and the initial error write-back amount corresponding to each image segmentation unit is obtained.

[0092] The initial error write-back amount is consolidated to ensure consistency within the same sample, generating the unit error write-back amount for each image segmentation unit.

[0093] Establish a training correspondence between the unit error write-back amount and the color features, texture features, particle morphology features and spatial location features of the corresponding image segmentation unit;

[0094] The mapping parameters between image features and predicted mineral content of units in the image student model are adjusted according to the training correspondence to obtain the updated image student model.

[0095] Establish a correction correspondence between the aggregated prediction error and the true values ​​of the sensitive spectral index, sensitive band, and target mineral content in the physics teacher model;

[0096] Based on the correction correspondence, the output correction parameters between the sensitive spectral index and the predicted value of the target mineral content in the physics teacher model are adjusted to obtain the updated physics teacher model.

[0097] Repeatedly execute the image student model output, area-weighted aggregation, aggregation prediction error calculation, unit error write-back generation, image student model update and physics teacher model update until the aggregation prediction error is less than the preset error threshold in consecutive preset rounds.

[0098] The image student model, physics teacher model, sensitive band, sensitive spectral index, mineral content allocation label generation rules, and area-weighted aggregation rules that meet the preset error conditions are jointly determined as the mapping model that maps near-field image data to the mineral content of construction waste soil.

[0099] The present invention has the following beneficial effects:

[0100] 1. In this invention, firstly, by combining hyperspectral reflectance data, acquired geometric parameters, and Hapke model inversion, a physical latent variable spectrum based on the inverted single-scattering albedo is generated. Then, a spectral index library is constructed based on the physical latent variable spectrum, and sensitive spectral indices and sensitive bands are screened, thereby reducing the impact of construction waste soil particle roughness, multiple scattering, local water content, and anomalous bands on mineral content prediction. Simultaneously, the physical teacher model outputs the predicted value of target mineral content and physical latent variable spectral fragments, enabling quartz, calcite, clay minerals, and cement hydration product-related minerals to form differentiated expressions with physical constraints, improving the sensitivity, accuracy, stability, and interpretability of mineral identification.

[0101] 2. In this invention, image segmentation units are generated through particle boundary segmentation and texture mutation segmentation. Mineral content allocation labels are generated based on color features, texture features, particle morphology features, spatial location features, area ratio, and mineral response weights, enabling the sample-level target mineral content true value to be transformed into image segmentation unit-level supervisory information. Simultaneously, the mineral content allocation labels and physical latent variable spectrum are jointly learned through an image student model. The image student model and the physical teacher model are updated by using area-weighted aggregation error back-writing, which improves the localization ability, response time, and continuous correction ability of near-field image prediction of construction waste soil mineral content, providing auxiliary decision support for on-site screening, raw material grading, resource utilization matching, and stockpile management. Attached Figure Description

[0102] Figure 1 This is a diagram illustrating the method steps of the present invention. Detailed Implementation

[0103] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0104] Reference Figure 1 The present invention provides an embodiment of a mapping model establishment method, characterized in that it includes:

[0105] Multiple construction waste soil samples were acquired. For each construction waste soil sample, hyperspectral reflectance data, close-up image data, and X-ray diffraction quantitative analysis data were collected simultaneously. The X-ray diffraction quantitative analysis data included the true value of the target mineral content. The acquisition geometric parameters corresponding to the hyperspectral reflectance data included the incident zenith angle, the emission zenith angle, and the phase angle.

[0106] Based on hyperspectral reflectance data and collected geometric parameters, the physical parameters of each construction waste soil sample were inverted using the Hapke model to obtain the inverted single scattering albedo corresponding to each wavelength, and the physical latent variable spectrum corresponding to each construction waste soil sample was generated based on the inverted single scattering albedo.

[0107] A spectral index library is constructed based on physical latent variable spectroscopy. The spectral indices in the library are correlated with the true values ​​of target mineral content to obtain the sensitive spectral indices and sensitive bands corresponding to the target minerals. A physical teacher model is established using the sensitive spectral indices as inputs and the true values ​​of target mineral content as outputs.

[0108] Grain boundary segmentation and texture abrupt change segmentation are performed on close-range image data to obtain multiple image segmentation units, and the color features, texture features, grain morphology features and spatial location features of each image segmentation unit are extracted.

[0109] Based on the predicted value of target mineral content, physical latent variable spectrum, and area ratio of each image segmentation unit output by the physics teacher model, a mineral content assignment label is generated for each image segmentation unit. The mineral content assignment labels of all image segmentation units in the same construction waste soil sample are aggregated by area weighting to satisfy the consistency constraint with the true value of target mineral content.

[0110] The image student model is trained by taking the color features, texture features, grain morphology features and spatial location features of each image segmentation unit as inputs, and the mineral content assignment labels and physical latent variable spectral data as joint supervision information.

[0111] The image student model performs area-weighted aggregation of the predicted mineral content of all image segmentation units within the same construction waste soil sample to obtain the sample-level aggregated predicted mineral content. The physical teacher model and the image student model are then updated based on the error between the sample-level aggregated predicted mineral content and the true value of the target mineral content, thus obtaining a mapping model that maps close-up image data to the mineral content of construction waste soil.

[0112] The Hapke model was used to invert the physical parameters of each construction waste soil sample, obtaining the inverted single-scattering albedo for each wavelength. Based on the inverted single-scattering albedo, the physical latent variable spectrum corresponding to each construction waste soil sample was generated, including:

[0113] The hyperspectral reflectance data are arranged according to sample number and wavelength to form a sample reflectance sequence;

[0114] The incident zenith angle, emitted zenith angle, and phase angle are converted into the incident zenith angle cosine value, emitted zenith angle cosine value, and phase angle variable corresponding to the Hapke model, and the incident zenith angle cosine value, emitted zenith angle cosine value, and phase angle variable are made to establish a correspondence between the incident zenith angle cosine value, emitted zenith angle cosine value, and phase angle variable and the reflectance of each wavelength in the sample reflectance sequence.

[0115] For each wavelength position in the sample reflectance sequence, the single scattering albedo is used as the variable to be inverted, and the phase function, backscattering function, and multiple scattering function are used as calculation terms in the Hapke model. The Hapke model is used to calculate the reflectance residual between the reflectance and the hyperspectral reflectance.

[0116] According to the rule of minimizing reflectance residual, the single scattering albedo corresponding to each wavelength position is iteratively solved to obtain the inverted single scattering albedo corresponding to each wavelength.

[0117] Connect the inverted single scattering albedo corresponding to each wavelength in wavelength order to form a single scattering albedo spectral curve.

[0118] The single-scattering albedo spectral curves were band-consistentized, and wavelengths that did not participate in the inversion were removed. The single-scattering albedo spectral curves after band-consistencyization were used as the physical latent variable spectra of the construction waste soil samples.

[0119] The single-scattering albedo corresponding to each wavelength position is iteratively solved, and the single-scattering albedo spectral curves are standardized for band consistency, including:

[0120] The reflectance in the sample reflectance sequence located in the instrument noise band, saturation band, missing band, and reflectance mutation band is marked as invalid reflectance, and the wavelength position corresponding to the invalid reflectance is marked as invalid wavelength position;

[0121] Based on the hyperspectral reflectance corresponding to the effective wavelength position, the initial range of single scattering albedo and the iterative search step size are set, and the initial range of values ​​is established in correspondence with the effective wavelength position.

[0122] For each effective wavelength position, the single-scattering albedo is adjusted in the direction of decreasing reflectivity residual in each round, and the candidate single-scattering albedo and the corresponding reflectivity residual obtained in each iteration are recorded;

[0123] When any one of the following conditions is met—that the change in reflectivity residual between two adjacent iterations is less than a preset residual change threshold or that the number of iterations reaches a preset maximum number of iterations—the iterative solution for the corresponding effective wavelength position is stopped.

[0124] The candidate single-scattering albedo with the smallest reflectance residual at the time of stopping iteration is determined as the inversion single-scattering albedo at the corresponding effective wavelength position;

[0125] The inverted single scattering albedo corresponding to all effective wavelength positions is connected in wavelength order, and the invalid wavelength positions are retained as invalid markers, generating a physical latent variable spectrum with the correspondence between sample number, effective wavelength position and invalid wavelength position.

[0126] A spectral index library is constructed based on the physical latent variable spectrum. The spectral indices in this library are then correlated with the true values ​​of target mineral content for feature filtering. This yields the sensitive spectral indices and sensitive bands corresponding to the target minerals. A physical teacher model is then established using these sensitive spectral indices as input and the true values ​​of target mineral content as output. This model includes:

[0127] The physical latent variable spectra of each construction waste soil sample are arranged according to sample number and wavelength order to form a physical latent variable spectral matrix;

[0128] Based on the inversion of single scattering albedo corresponding to any two wavelength positions in the physical latent variable spectral matrix, a dual-band ratio index, a normalized difference index, and a difference index are constructed, and an index relationship is established between the constructed indices and the corresponding wavelength positions.

[0129] Based on the inversion of single scattering albedo corresponding to any three wavelength positions in the physical latent variable spectral matrix, a three-band combination index is constructed, and an index relationship is established between the three-band combination index and the corresponding three wavelength positions.

[0130] The dual-band ratio index, normalized difference index, difference index, and three-band combination index are merged according to the sample number to generate a spectral index library that corresponds to the true value of the target mineral content.

[0131] The correlation degree between each spectral index in the spectral index library and the true value of the target mineral content is calculated, and the spectral index correlation ranking results are generated in descending order of correlation degree.

[0132] Based on the spectral index correlation ranking results, spectral indices with correlation scores below a preset correlation score threshold are removed, and collinearity screening is performed on the remaining spectral indices to obtain the sensitive spectral indices and sensitive bands corresponding to the target minerals.

[0133] Using the sensitive spectral index as the input variable and the true value of the target mineral content as the output variable, the regression model was trained, and the trained regression model was used as the physics teacher model.

[0134] The correlation between each spectral index in the spectral index library and the true value of the target mineral content is calculated, and collinearity is screened out based on the spectral index correlation ranking results, including:

[0135] According to the target mineral type, the spectral indices in the spectral index library that have a sample correspondence with the true value of the content of the same target mineral are divided into a candidate index set;

[0136] For each spectral index in the candidate index set, calculate its linear correlation, rank correlation, and information correlation with the true value of the target mineral content, and generate a comprehensive correlation based on the linear correlation, rank correlation, and information correlation.

[0137] The spectral indices in the candidate index set are arranged in descending order of comprehensive correlation, and the spectral index correlation ranking results are generated.

[0138] Spectral indices with a comprehensive correlation score lower than a preset correlation threshold are removed from the spectral index correlation ranking results to obtain a preliminary set of spectral indices.

[0139] Calculate the index similarity between any two spectral indices in the initial screening spectral index set. When the index similarity between two spectral indices is higher than the preset similarity threshold, retain the spectral index with a higher overall correlation to the true value of the target mineral content.

[0140] The spectral indices retained after collinearity removal are defined as sensitive spectral indices, the wavelength positions corresponding to the sensitive spectral indices are defined as sensitive bands, and the sensitive spectral indices, sensitive bands, predicted values ​​of target mineral content, and spectral fragments of physical latent variables corresponding to the sensitive bands are used as the output of the physics teacher model.

[0141] Grain boundary segmentation and texture abrupt change segmentation are performed on close-range image data to obtain multiple image segmentation units. The color features, texture features, grain morphology features, and spatial location features of each image segmentation unit are extracted, including:

[0142] Close-up image data are collected according to sample number, and the background area, shadow area and non-soil area in the close-up image data are masked to obtain the effective image area of ​​the sample.

[0143] Based on the color gradient change and brightness gradient change in the effective image region of the sample, candidate lines of particle boundaries are extracted, and the candidate lines of particle boundaries are connected by breakpoints and closed regions are sorted to obtain the particle boundary segmentation results.

[0144] Based on the gray-level co-occurrence variation, local texture direction variation and texture intensity abrupt change location in the effective image region of the sample, texture abrupt change segmentation line is extracted, and the texture abrupt change segmentation line is fused with the particle boundary segmentation result to obtain image segmentation unit;

[0145] Calculate the percentage of pixels in the effective image region of each image segmentation unit, and use the percentage of pixels as the area percentage of the image segmentation unit;

[0146] For each image segmentation unit, extract the color mean, color variance, and color distribution histogram to form color features;

[0147] For each image segmentation unit, texture orientation, texture density, and texture contrast are extracted to form texture features;

[0148] For each image segmentation unit, the boundary length, region area, aspect ratio, roundness, and adjacency relationship are extracted to form particle morphology features, and spatial location features are formed based on the center coordinates of the image segmentation unit in the effective image region of the sample.

[0149] Based on the predicted target mineral content output by the physics teacher model, the physical latent variable spectrum, and the area proportion of each image segmentation unit, a mineral content assignment label is generated for each image segmentation unit. The mineral content assignment labels of all image segmentation units within the same construction waste soil sample are then aggregated using area weighting to ensure consistency with the true target mineral content, including:

[0150] According to the sample number, retrieve the true value of the target mineral content, the predicted value of the target mineral content, the physical latent variable spectrum, and the area ratio of all image segmentation units corresponding to the same construction waste soil sample;

[0151] The inversion single scattering albedo response value corresponding to the sensitive band in the physical latent variable spectrum is used as the mineral response benchmark, and the color features, texture features and grain morphology features of each image segmentation unit are converted into unit image response content.

[0152] Based on the degree of matching between the unit image response content and the mineral response benchmark, the mineral response weight corresponding to each image segmentation unit is generated, and the mineral response weight is normalized according to the sum of the mineral response weights of all image segmentation units in the same sample.

[0153] Based on the normalized mineral response weights, the area proportion of each image segmentation unit, and the target mineral content prediction value output by the physics teacher model, an initial mineral content assignment label is generated for each image segmentation unit.

[0154] The initial mineral content assignment labels within the same construction waste soil sample are weighted and aggregated according to their area proportion to obtain the sample-level mineral content label aggregation value.

[0155] Based on the difference between the aggregated value of the sample-level mineral content label and the true value of the target mineral content, the initial mineral content assignment labels are proportionally corrected to obtain the mineral content assignment label corresponding to each image segmentation unit.

[0156] Mineral response weights are generated for each image segmentation unit based on the degree of matching between the unit image response content and the mineral response benchmark. The initial mineral content assignment labels are then corrected based on the difference between the aggregated value of the sample-level mineral content labels and the true value of the target mineral content, including:

[0157] The color features, texture features, and grain morphology features of each image segmentation unit are scaled to obtain color response sub-vectors, texture response sub-vectors, and grain morphology response vectors.

[0158] The color response subvector, texture response subvector, and grain shape response subvector are connected in a preset order to generate the unit image response content corresponding to each image segmentation unit.

[0159] The inversion single scattering albedo response value corresponding to the sensitive band is extracted from the physical latent variable spectrum, and the inversion single scattering albedo response value is matched with the target mineral content prediction value output by the physical teacher model to generate the mineral response benchmark corresponding to the target mineral.

[0160] Calculate the response difference value between the unit image response content of each image segmentation unit and the mineral response benchmark, and determine the matching degree of each image segmentation unit relative to the target mineral according to the response difference value in ascending order;

[0161] The mineral response weights corresponding to each image segmentation unit are generated based on the matching degree, and the mineral response weights of all image segmentation units in the same construction waste soil sample are added together to obtain the total sample weight.

[0162] Using the total sample weight as the normalization benchmark, the mineral response weight corresponding to each image segmentation unit is normalized to obtain the normalized mineral response weight.

[0163] According to the sample number, obtain the initial mineral content allocation label and area ratio corresponding to all image segmentation units within the same construction waste soil sample;

[0164] Multiply the initial mineral content assignment label corresponding to each image segmentation unit by the area ratio to obtain the area-weighted label value corresponding to each image segmentation unit;

[0165] The sample-level mineral content label aggregate value is obtained by summing all area-weighted label values ​​within the same construction waste soil sample.

[0166] The difference between the aggregated value of the sample-level mineral content label and the true value of the target mineral content is calculated to obtain the sample-level label correction difference;

[0167] Based on the normalized mineral response weight and area ratio of each image segmentation unit, the sample-level label correction difference is allocated to each image segmentation unit in the same construction waste soil sample to obtain the label correction amount corresponding to each image segmentation unit.

[0168] The initial mineral content assignment label and label correction value corresponding to each image segmentation unit are merged to obtain the mineral content assignment label corresponding to each image segmentation unit. The mineral content assignment labels of all mineral content in the same construction waste soil sample are then aggregated by area weighting to meet the consistency constraint with the true value of the target mineral content.

[0169] The image student model performs area-weighted aggregation of the mineral content predictions output by all image segmentation units within the same construction waste soil sample to obtain a sample-level aggregated mineral content prediction. The physics teacher model and the image student model are then updated based on the error between the sample-level aggregated mineral content prediction and the true target mineral content. This includes:

[0170] According to the sample number, retrieve all image segmentation units corresponding to the same construction waste soil sample, the area ratio of each image segmentation unit, the color features, texture features, particle morphology features and spatial location features of each image segmentation unit, the true value of the target mineral content and the physical latent variable spectrum;

[0171] The color features, texture features, grain morphology features and spatial location features of each image segmentation unit are input into the image student model to obtain the predicted value of the mineral content of each image segmentation unit, and the correspondence between the predicted value of the mineral content of the unit and the image segmentation unit is maintained.

[0172] Based on the area proportion of each image segmentation unit, the predicted mineral content of all units in the same construction waste soil sample is weighted and summarized to obtain the sample-level aggregated predicted mineral content.

[0173] Calculate the aggregation prediction error between the aggregated predicted value of mineral content at the sample level and the true value of target mineral content, and allocate the aggregation prediction error to each image segmentation unit within the same construction waste soil sample to generate the unit error write-back amount;

[0174] The mapping parameters corresponding to color features, texture features, particle morphology features, and spatial location features in the image student model are updated based on the unit error write-back amount.

[0175] The output correction parameters between the sensitive spectral index and the predicted value of the target mineral content in the physics teacher model are updated based on the aggregated prediction error. When the aggregated prediction error meets the preset error conditions, a mapping model from close-up image data to the mineral content of construction waste soil is output.

[0176] The aggregation prediction error between the aggregated predicted value of mineral content at the sample level and the true value of the target mineral content is calculated. This aggregation prediction error is then allocated to the unit error write-back amount generated by each image segmentation unit within the same construction waste soil sample. Based on the unit error write-back amount and the aggregation prediction error, the image student model and the physics teacher model are updated, including:

[0177] According to the sample number, obtain the sample-level aggregated predicted value of mineral content, the true value of target mineral content, the predicted value of unit mineral content corresponding to each image segmentation unit, the mineral content allocation label, the area ratio, and the normalized mineral response weight.

[0178] The difference between the predicted value of the sample-level mineral content and the true value of the target mineral content is calculated to obtain the sample-level aggregation error;

[0179] The difference between the predicted mineral content value of each image segmentation unit and the mineral content assignment label is calculated to obtain the unit prediction bias.

[0180] Based on the area ratio, normalized mineral response weight and unit prediction bias of each image segmentation unit, an error allocation coefficient is generated for each image segmentation unit.

[0181] The sample-level aggregation error is allocated to each image segmentation unit within the same construction waste soil sample according to the error allocation coefficient, and the initial error write-back amount corresponding to each image segmentation unit is obtained.

[0182] The initial error write-back amount is consolidated to ensure consistency within the same sample, generating the unit error write-back amount for each image segmentation unit.

[0183] Establish a training correspondence between the unit error write-back amount and the color features, texture features, particle morphology features and spatial location features of the corresponding image segmentation unit;

[0184] The mapping parameters between image features and predicted mineral content of units in the image student model are adjusted according to the training correspondence to obtain the updated image student model.

[0185] Establish a correction correspondence between the aggregated prediction error and the true values ​​of the sensitive spectral index, sensitive band, and target mineral content in the physics teacher model;

[0186] Based on the correction correspondence, the output correction parameters between the sensitive spectral index and the predicted value of the target mineral content in the physics teacher model are adjusted to obtain the updated physics teacher model.

[0187] Repeatedly execute the image student model output, area-weighted aggregation, aggregation prediction error calculation, unit error write-back generation, image student model update and physics teacher model update until the aggregation prediction error is less than the preset error threshold in consecutive preset rounds.

[0188] The image student model, physics teacher model, sensitive band, sensitive spectral index, mineral content allocation label generation rules, and area-weighted aggregation rules that meet the preset error conditions are jointly determined as the mapping model that maps near-field image data to the mineral content of construction waste soil.

[0189] Example 1: To verify the applicability of this invention in the scenario of resource utilization of construction waste soil, it was applied to the incoming raw material screening process of a construction waste soil resource utilization treatment plant. This treatment plant receives waste soil daily from different demolition sites, foundation pit cleaning areas, and old building material sorting areas. The soil contains fine brick powder, mortar residue, cement hydration products, clay particles, weathered stone powder, and a small amount of recycled aggregate debris. The original testing method mainly relied on manual sampling followed by X-ray diffraction quantitative analysis in a laboratory. The process of sampling, sample preparation, queuing, and generating a mineral content report for a single batch of samples typically takes a long time. The treatment plant cannot complete the raw material mineral composition determination on the day of arrival, resulting in the mixing of high-calcium soil, soil with high clay mineral content, and soil with high silica content in the stockpiling area. When subsequently used as recycled filler, roadbed material, or low-strength building material admixtures, re-sorting and re-inspection are required, increasing the number of in-plant transfers and the storage area occupied.

[0190] In this scenario, construction waste soil entering the treatment site is divided into sample groups according to its source batch. Samples from each group are taken from different locations within the pile and mixed to form representative construction waste soil samples. Each sample undergoes drying, sieving, and surface leveling to ensure a relatively stable reflective surface. Subsequently, hyperspectral reflectance data, close-range image data, and quantitative X-ray diffraction analysis data are simultaneously acquired for each sample. During hyperspectral acquisition, the incident zenith angle, emission zenith angle, and phase angle are recorded, and the hyperspectral reflectance data of the same sample are arranged according to sample number and wavelength order. During close-range image acquisition, the sample is laid on a uniformly colored background, preserving particle boundaries, color differences, texture distribution, and local agglomeration morphology in the image. Quantitative X-ray diffraction analysis is used to obtain the true values ​​of target mineral content for quartz, calcite, feldspar, illite, kaolinite, dolomite, and cement hydration products.

[0191] During model building, hyperspectral reflectance data and acquired geometric parameters were input into the Hapke model to perform physical parameter inversion for each construction waste soil sample, obtaining the inverted single-scattering albedo corresponding to each wavelength position. To reduce the impact of instrument noise, saturated reflectance, and missing bands on subsequent modeling, unstable bands in the sample reflectance sequence were marked as invalid wavelength positions, and only the hyperspectral reflectance at the effective wavelength positions participated in the iterative solution. During the iteration process, the single-scattering albedo was used as the variable to be inverted, and the phase function, backscattering function, and multiple scattering function were used as calculation terms in the Hapke model. After the reflectance residual decreased in each iteration, the inverted single-scattering albedo was output, forming a physical latent variable spectrum. This physical latent variable spectrum was then used to construct a two-band ratio index, a normalized difference index, a difference index, and a three-band combination index. The spectral indices were correlated and screened with the true values ​​of target mineral content obtained from X-ray diffraction quantitative analysis to form sensitive spectral indices and sensitive bands corresponding to different target minerals.

[0192] In its horizontally expanded applications, this method is not only used for single mineral content prediction but also establishes mapping relationships for multiple resource utilization uses. For samples with a high proportion of silica, the model focuses on the content of quartz and feldspar; for samples that may affect curing performance, the model focuses on calcite, dolomite, and cement hydration product-related minerals; for samples that affect water absorption and volume stability, the model focuses on the content of kaolinite, illite, and other clay minerals. The physics teacher model is trained by the sensitive spectral index and the true value of the target mineral content, outputting the predicted value of the target mineral content and the spectral fragment of the physical latent variable corresponding to the sensitive band. After the near-field image data is masked for background, shadow, and non-soil areas, the effective image area of ​​the sample is obtained, and multiple image segmentation units are formed by particle boundary segmentation and texture abrupt segmentation. Each image segmentation unit extracts the color mean, color variance, color distribution histogram, texture direction, texture density, texture contrast, boundary length, region area, aspect ratio, roundness, adjacency relationship, and center coordinates, and calculates the pixel count ratio as the area ratio.

[0193] Since X-ray diffraction quantitative analysis data is sample-level results, while close-range image data consists of multiple image segmentation units, directly assigning sample-level ground truth values ​​to each image segmentation unit would mask local particle differences. Therefore, this embodiment generates mineral content assignment labels based on the target mineral content prediction value output by the physics teacher model, the physical latent variable spectrum, and the area proportion of each image segmentation unit. Specifically, the inverted single-scattering albedo response value corresponding to the sensitive band is used as the mineral response benchmark. The color features, texture features, and particle morphology features of the image segmentation unit are converted into unit image response content. The matching degree of the image segmentation unit relative to the target mineral is determined by the response difference value, and normalized mineral response weights are generated accordingly. After the initial mineral content assignment labels are generated, all image segmentation units within the same sample are weighted and aggregated according to their area proportions to obtain sample-level mineral content label aggregation values. Then, the difference between this aggregation value and the target mineral content ground truth value is used for proportional correction to form mineral content assignment labels for training the image student model.

[0194] During an eight-week validation period, a total of 480 sets of construction waste soil samples were collected, of which 360 sets were used for model training, 60 sets for model update validation, and 60 sets for independent testing. Six close-up images were collected from each sample set, resulting in a total of 2,880 close-up images. The hyperspectral reflectance sampling wavelength range covered the visible-near-infrared to short-wave infrared region. After removing noise and saturated bands, 1,620 effective wavelength positions were retained for each sample. Each sample set was segmented into an average of 146 image segments. Using the traditional direct hyperspectral regression method, the average absolute error for quartz content prediction was 6.8%, for calcite content prediction it was 7.3%, and for kaolinite content prediction it was 5.9%. Using the mapping model established in this invention, the average absolute error for quartz content prediction decreased to 3.1%, for calcite content prediction it decreased to 3.5%, and for kaolinite content prediction it decreased to 2.9%. When classifying test samples for resource utilization, the consistency rate between the original manual experience classification and the re-inspection results was 78.6%. After adopting the mapping model of this invention, the consistency rate between the classification of incoming samples and the subsequent X-ray diffraction verification results reached 91.8%.

[0195] In terms of processing efficiency, the average cycle from sampling upon arrival to obtaining the laboratory mineral content report for each batch of samples in the traditional process is over 36 hours, and the processing site can usually only temporarily stockpile the samples before the report is issued. In this embodiment, after the initial mapping model is established, new samples are input into the image student model through close-up image data, and the mineral content prediction results for a single group of samples are generated within an average of 120 seconds. Statistical analysis of incoming materials within the same validation cycle shows that the number of repeated transfers decreased from an average of 3.4 times per 100 tons to an average of 1.6 times per 100 tons, the proportion of samples requiring secondary re-inspection due to misclassification decreased from 21.5% to 7.2%, and the temporary mixed-up area of ​​the resource recovery stockpile decreased by approximately 28%. These results indicate that the present invention can not only establish a stable mapping relationship in laboratory sample prediction but also be horizontally extended to applications such as incoming screening, raw material grading, resource recovery use matching, and stockpile management, enabling the prediction of mineral content in construction waste soil to expand from single-point detection to rapid on-site judgment of consecutive batches.

[0196] Example 2: To further verify the feasibility and accuracy improvement of this invention in the core modeling process, it was applied to the refined modeling verification of the same batch of construction waste soil samples. This verification focused on hyperspectral physical inversion, physical teacher model establishment, image segmentation unit label de-aggregation, image student model training, and area-weighted aggregation update. The samples were fine-grained construction waste soil that had undergone preliminary screening within the same resource recovery site. The samples contained tightly mixed mortar powder, brick powder, clay particles, and carbonate particles with small particle sizes, resulting in indistinct color differences when observed with the naked eye. Some samples appeared grayish-brown or light yellowish-brown, making it difficult to directly distinguish the mineral composition using ordinary RGB images. The main problem in this scenario is that construction waste soil is a typical complex mixed medium. The color and texture of the original image are not simply linearly related to the mineral content. The original hyperspectral reflectance is also affected by particle roughness, scattering path, and local water content. Directly training the image model or the spectral regression model can easily lead to problems such as mixed responses of different minerals, sample-level labels failing to correspond to local image regions, and unstable prediction results in local areas.

[0197] In this embodiment, samples were divided into training samples, calibration samples, and test samples. Each sample underwent uniform drying, sieving, and spreading before acquisition, with the spreading thickness maintained between five and eight millimeters to ensure that close-up images could fully reveal particle boundaries and local textures. During hyperspectral acquisition, the incident zenith angle, emission zenith angle, and phase angle were recorded for each sample, and the sample number was saved simultaneously. Close-up image acquisition employed a fixed light source, fixed shooting distance, and fixed exposure settings. For each sample, images were acquired from directly above and two from oblique angles. Images with sufficient clarity and no significant obstruction were then selected for subsequent processing. X-ray diffraction quantitative analysis data served as the true values ​​for the target mineral content. The quartz content ranged from 22.4% to 59.7%, the calcite content from 5.8% to 31.2%, the feldspar content from 3.6% to 18.9%, the combined kaolinite and illite content from 4.2% to 26.5%, and the content of cement hydration product-related minerals from 2.1% to 15.4%.

[0198] In the Hapke model inversion phase, the sample reflectance sequence is first arranged in wavelength order. The incident zenith angle and emitted zenith angle are converted into corresponding cosine values, and the phase angle is used as a geometric variable in the reflectance calculation. When checking the sample reflectance sequence, wavelength positions located in the instrument noise range, reflectance saturation range, missing record range, and abrupt change range are marked as invalid wavelength positions, while the remaining valid wavelength positions are iteratively solved. The single-scattering albedo at each valid wavelength position iterates from a preset initial range. The Hapke model calculates a reflectance residual between the reflectance and hyperspectral reflectance. The iteration stops when the change in the reflectance residual between two adjacent iterations is less than a preset residual change threshold or when the maximum number of iterations is reached. Statistically, the average number of iterations for a single sample at valid wavelength positions is 28, and 94.3% of valid wavelength positions converge before the maximum number of iterations. After inversion, the inverted single-scattering albedo is connected in wavelength order to form a physical latent variable spectrum, and the invalid wavelength position markings are retained. Compared to using the raw hyperspectral reflectance directly, the average fluctuation coefficient of the physical latent variable spectrum in repeated acquisitions of adjacent samples decreased from 4.9 percent to 2.1 percent.

[0199] In the spectral index library construction phase, the physical latent variable spectral matrix is ​​arranged by sample number. Two-band ratio indices, normalized difference indices, and difference indices are constructed based on any two effective wavelength positions, and three-band combination indices are constructed based on any three effective wavelength positions. For each target mineral, spectral indices in the spectral index library that have sample correspondences with the true content value of that target mineral are divided into a candidate index set. Linear correlation, rank correlation, and information correlation are then calculated to generate a comprehensive correlation. After the initial spectral index set is formed, the similarity between indices is calculated. When two spectral indices are highly similar, the spectral index with a higher comprehensive correlation to the true content value of the target mineral is retained. Ultimately, the sensitive bands corresponding to quartz are mainly concentrated in the near-infrared reflectance stable region and the short-wave infrared local absorption response region; the sensitive bands corresponding to calcite are concentrated in the region with obvious carbonate absorption response; and the sensitive bands corresponding to kaolinite and illite are concentrated in the region with hydroxyl-containing mineral response correlation. After the physics teacher model training was completed, the prediction determination coefficient for quartz reached 0.92, the prediction determination coefficient for calcite reached 0.89, the combined prediction determination coefficient for kaolinite and illite reached 0.87, and the prediction determination coefficient for minerals related to cement hydration products reached 0.84.

[0200] In the close-up image processing stage, each image is first masked for the background, shadow, and non-soil regions to obtain the effective image regions of the sample. Then, candidate particle boundary lines are extracted from color and brightness gradient changes, breakpoints are connected, and closed regions are organized. Texture abrupt change segmentation lines are then extracted from gray-level co-occurrence changes, local texture direction changes, and texture intensity abrupt change locations. The two types of segmentation results are then fused to obtain image segmentation units. For each image segmentation unit, the following parameters are extracted: color mean, color variance, color distribution histogram, texture direction, texture density, texture contrast, boundary length, region area, aspect ratio, roundness, adjacency relationship, and center coordinates. A representative set of samples, after segmentation, forms 132 image segmentation units. Tiny units with an area percentage less than 0.1% are merged into adjacent similar units, retaining 109 effective image segmentation units. Through this processing, light-colored carbonate particles, dark-colored brick powder particles, and fine-grained clay agglomerates within the sample can form relatively independent training units.

[0201] In the mineral content allocation label generation stage, the predicted target mineral content, physical latent variable spectrum, and area proportion of each image segmentation unit output by the physics teacher model are jointly utilized. The inversion single scattering albedo response value corresponding to the sensitive band serves as the mineral response benchmark. The color features, texture features, and grain morphology features of each image segmentation unit are scaled to form the unit image response content. The system calculates the response difference between the unit image response content and the mineral response benchmark, and determines the matching degree of each image segmentation unit relative to the target mineral according to the order of response difference from smallest to largest, thereby obtaining the normalized mineral response weight. After the initial mineral content allocation label is generated, all image segmentation units are weighted and aggregated according to their area proportion to obtain the sample-level mineral content label aggregated value. If there is a difference between the label aggregated value and the true value of the target mineral content obtained from X-ray diffraction quantitative analysis, the difference is allocated to each image segmentation unit according to the normalized mineral response weight and area proportion to obtain the label correction amount, which is then merged with the initial mineral content allocation label. After correction, the average deviation between the aggregated value of the sample-level mineral content label and the true value of the target mineral content decreased from 4.6% before correction to 0.7%, ensuring that the image segmentation unit-level label can be consistent with the laboratory true value when returning to the sample level.

[0202] During the training phase of the image student model, the color features, texture features, particle morphology features, and spatial location features of each image segmentation unit are used as inputs, while mineral content assignment labels and physical latent variable spectra are used as joint supervision information. In the early stages of training, the image student model tends to misclassify some kaolinite responses as cement hydration product-related mineral responses in dark brick powder regions, with the aggregated prediction error between sample-level mineral content aggregated predictions and the true target mineral content exceeding 5% on some samples. Through unit error write-back generation, the system allocates the sample-level aggregated error to the corresponding image segmentation unit and updates the mapping parameters between image features and unit mineral content predictions in the image student model based on the unit error write-back amount. Simultaneously, it adjusts the output correction parameters between the sensitivity spectral index and the target mineral content prediction value in the physical teacher model based on the aggregated prediction error. After multiple rounds of closed-loop updates, the aggregated prediction error in consecutive rounds is consistently less than the preset error threshold, and the mapping model is determined as the final model.

[0203] In the longitudinal refinement verification, three comparative methods were used to evaluate the technical effects of the present invention. When directly using close-up images to train the model, the average absolute error for quartz content was 7.6%, for calcite content it was 8.1%, and for the combined kaolinite and illite content it was 6.4%. When directly using the original hyperspectral reflectance to train the regression model, the average absolute error for quartz content was 5.2%, for calcite content it was 5.7%, and for the combined kaolinite and illite content it was 4.8%. After using the physical teacher model, mineral content allocation labels, and image student model of the present invention for closed-loop updates, the average absolute error for quartz content decreased to 2.8%, for calcite content it decreased to 3.2%, for the combined kaolinite and illite content it decreased to 2.7%, and for cement hydration product-related minerals it decreased to 3.9%. Regarding the prediction time for a single sample, quantitative X-ray diffraction analysis typically takes several hours or more, while direct hyperspectral prediction requires over 300 seconds for data acquisition and calculation. In contrast, this invention, after model establishment, generates mineral content prediction results from close-up image data in an average of 95 seconds. For the same batch of test samples, when using the output results of this invention for mineral content classification, the accuracy rate for identifying high-silica samples reaches 93.5%, high-carbonate samples reaches 91.2%, and high-clay mineral samples reaches 89.7%. This embodiment demonstrates that this invention, through Hapke model inversion, spectral index screening, physical teacher model, image segmentation unit label de-clusting, and area-weighted closed-loop update, enables sample-level laboratory ground truth to be transformed into image segmentation unit-level training information, and further establishes a stable mapping relationship between close-up images and the mineral content of construction waste soil.

[0204] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for establishing a mapping model, characterized in that: include: Multiple construction waste soil samples were acquired. For each construction waste soil sample, hyperspectral reflectance data, close-up image data, and X-ray diffraction quantitative analysis data were collected simultaneously. The X-ray diffraction quantitative analysis data included the true value of the target mineral content. The acquisition geometric parameters corresponding to the hyperspectral reflectance data included the incident zenith angle, the emission zenith angle, and the phase angle. Based on hyperspectral reflectance data and collected geometric parameters, the physical parameters of each construction waste soil sample were inverted using the Hapke model to obtain the inverted single scattering albedo corresponding to each wavelength, and the physical latent variable spectrum corresponding to each construction waste soil sample was generated based on the inverted single scattering albedo. A spectral index library is constructed based on physical latent variable spectroscopy. The spectral indices in the library are correlated with the true values ​​of target mineral content to obtain the sensitive spectral indices and sensitive bands corresponding to the target minerals. A physical teacher model is established using the sensitive spectral indices as inputs and the true values ​​of target mineral content as outputs. Grain boundary segmentation and texture abrupt change segmentation are performed on close-range image data to obtain multiple image segmentation units, and the color features, texture features, grain morphology features and spatial location features of each image segmentation unit are extracted. Based on the predicted value of target mineral content, physical latent variable spectrum, and area ratio of each image segmentation unit output by the physics teacher model, a mineral content assignment label is generated for each image segmentation unit. The mineral content assignment labels of all image segmentation units in the same construction waste soil sample are aggregated by area weighting to satisfy the consistency constraint with the true value of target mineral content. The image student model is trained by taking the color features, texture features, grain morphology features and spatial location features of each image segmentation unit as inputs, and the mineral content assignment labels and physical latent variable spectral data as joint supervision information. The image student model performs area-weighted aggregation of the predicted mineral content of all image segmentation units within the same construction waste soil sample to obtain the sample-level aggregated predicted mineral content. The physical teacher model and the image student model are then updated based on the error between the sample-level aggregated predicted mineral content and the true value of the target mineral content, thus obtaining a mapping model that maps close-up image data to the mineral content of construction waste soil.

2. The mapping model establishment method according to claim 1, characterized in that: The Hapke model was used to invert the physical parameters of each construction waste soil sample, obtaining the inverted single-scattering albedo for each wavelength. Based on the inverted single-scattering albedo, the physical latent variable spectrum corresponding to each construction waste soil sample was generated, including: The hyperspectral reflectance data are arranged according to sample number and wavelength to form a sample reflectance sequence; The incident zenith angle, emitted zenith angle, and phase angle are converted into the incident zenith angle cosine value, emitted zenith angle cosine value, and phase angle variable corresponding to the Hapke model, and the incident zenith angle cosine value, emitted zenith angle cosine value, and phase angle variable are made to establish a correspondence between the incident zenith angle cosine value, emitted zenith angle cosine value, and phase angle variable and the reflectance of each wavelength in the sample reflectance sequence. For each wavelength position in the sample reflectance sequence, the single scattering albedo is used as the variable to be inverted, and the phase function, backscattering function, and multiple scattering function are used as calculation terms in the Hapke model. The Hapke model is used to calculate the reflectance residual between the reflectance and the hyperspectral reflectance. According to the rule of minimizing reflectance residual, the single scattering albedo corresponding to each wavelength position is iteratively solved to obtain the inverted single scattering albedo corresponding to each wavelength. Connect the inverted single scattering albedo corresponding to each wavelength in wavelength order to form a single scattering albedo spectral curve. The single-scattering albedo spectral curves were band-consistentized, and wavelengths that did not participate in the inversion were removed. The single-scattering albedo spectral curves after band-consistencyization were used as the physical latent variable spectra of the construction waste soil samples.

3. The mapping model establishment method according to claim 1, characterized in that: The single-scattering albedo corresponding to each wavelength position is iteratively solved, and the single-scattering albedo spectral curves are standardized for band consistency, including: The reflectance in the sample reflectance sequence located in the instrument noise band, saturation band, missing band, and reflectance mutation band is marked as invalid reflectance, and the wavelength position corresponding to the invalid reflectance is marked as invalid wavelength position; Based on the hyperspectral reflectance corresponding to the effective wavelength position, the initial range of single scattering albedo and the iterative search step size are set, and the initial range of values ​​is established in correspondence with the effective wavelength position. For each effective wavelength position, the single-scattering albedo is adjusted in the direction of decreasing reflectivity residual in each round, and the candidate single-scattering albedo and the corresponding reflectivity residual obtained in each iteration are recorded; When any one of the following conditions is met—that the change in reflectivity residual between two adjacent iterations is less than a preset residual change threshold or that the number of iterations reaches a preset maximum number of iterations—the iterative solution for the corresponding effective wavelength position is stopped. The candidate single-scattering albedo with the smallest reflectance residual at the time of stopping iteration is determined as the inversion single-scattering albedo at the corresponding effective wavelength position; The inverted single scattering albedo corresponding to all effective wavelength positions is connected in wavelength order, and the invalid wavelength positions are retained as invalid markers, generating a physical latent variable spectrum with the correspondence between sample number, effective wavelength position and invalid wavelength position.

4. The mapping model establishment method according to claim 1, characterized in that: A spectral index library is constructed based on the physical latent variable spectrum. The spectral indices in this library are then correlated with the true values ​​of target mineral content for feature filtering. This yields the sensitive spectral indices and sensitive bands corresponding to the target minerals. A physical teacher model is then established using these sensitive spectral indices as input and the true values ​​of target mineral content as output. This model includes: The physical latent variable spectra of each construction waste soil sample are arranged according to sample number and wavelength order to form a physical latent variable spectral matrix; Based on the inversion of single scattering albedo corresponding to any two wavelength positions in the physical latent variable spectral matrix, a dual-band ratio index, a normalized difference index, and a difference index are constructed, and an index relationship is established between the constructed indices and the corresponding wavelength positions. Based on the inversion of single scattering albedo corresponding to any three wavelength positions in the physical latent variable spectral matrix, a three-band combination index is constructed, and an index relationship is established between the three-band combination index and the corresponding three wavelength positions. The dual-band ratio index, normalized difference index, difference index, and three-band combination index are merged according to the sample number to generate a spectral index library that corresponds to the true value of the target mineral content. The correlation degree between each spectral index in the spectral index library and the true value of the target mineral content is calculated, and the spectral index correlation ranking results are generated in descending order of correlation degree. Based on the spectral index correlation ranking results, spectral indices with correlation scores below a preset correlation score threshold are removed, and collinearity screening is performed on the remaining spectral indices to obtain the sensitive spectral indices and sensitive bands corresponding to the target minerals. Using the sensitive spectral index as the input variable and the true value of the target mineral content as the output variable, the regression model was trained, and the trained regression model was used as the physics teacher model.

5. The mapping model establishment method according to claim 1, characterized in that: The correlation between each spectral index in the spectral index library and the true value of the target mineral content is calculated, and collinearity is screened out based on the spectral index correlation ranking results, including: According to the target mineral type, the spectral indices in the spectral index library that have a sample correspondence with the true value of the content of the same target mineral are divided into a candidate index set; For each spectral index in the candidate index set, calculate its linear correlation, rank correlation, and information correlation with the true value of the target mineral content, and generate a comprehensive correlation based on the linear correlation, rank correlation, and information correlation. The spectral indices in the candidate index set are arranged in descending order of comprehensive correlation, and the spectral index correlation ranking results are generated. Spectral indices with a comprehensive correlation score lower than a preset correlation threshold are removed from the spectral index correlation ranking results to obtain a preliminary set of spectral indices. Calculate the index similarity between any two spectral indices in the initial screening spectral index set. When the index similarity between two spectral indices is higher than the preset similarity threshold, retain the spectral index with a higher overall correlation to the true value of the target mineral content. The spectral indices retained after collinearity removal are defined as sensitive spectral indices, the wavelength positions corresponding to the sensitive spectral indices are defined as sensitive bands, and the sensitive spectral indices, sensitive bands, predicted values ​​of target mineral content, and spectral fragments of physical latent variables corresponding to the sensitive bands are used as the output of the physics teacher model.

6. The mapping model establishment method according to claim 1, characterized in that: Grain boundary segmentation and texture abrupt change segmentation are performed on close-range image data to obtain multiple image segmentation units. The color features, texture features, grain morphology features, and spatial location features of each image segmentation unit are extracted, including: Close-up image data are collected according to sample number, and the background area, shadow area and non-soil area in the close-up image data are masked to obtain the effective image area of ​​the sample. Based on the color gradient change and brightness gradient change in the effective image region of the sample, candidate lines of particle boundaries are extracted, and the candidate lines of particle boundaries are connected by breakpoints and closed regions are sorted to obtain the particle boundary segmentation results. Based on the gray-level co-occurrence variation, local texture direction variation and texture intensity abrupt change location in the effective image region of the sample, texture abrupt change segmentation line is extracted, and the texture abrupt change segmentation line is fused with the particle boundary segmentation result to obtain image segmentation unit; Calculate the percentage of pixels in the effective image region of each image segmentation unit, and use the percentage of pixels as the area percentage of the image segmentation unit; For each image segmentation unit, extract the color mean, color variance, and color distribution histogram to form color features; For each image segmentation unit, texture orientation, texture density, and texture contrast are extracted to form texture features; For each image segmentation unit, the boundary length, region area, aspect ratio, roundness, and adjacency relationship are extracted to form particle morphology features, and spatial location features are formed based on the center coordinates of the image segmentation unit in the effective image region of the sample.

7. The mapping model establishment method according to claim 1, characterized in that: Based on the predicted target mineral content output by the physics teacher model, the physical latent variable spectrum, and the area proportion of each image segmentation unit, a mineral content assignment label is generated for each image segmentation unit. The mineral content assignment labels of all image segmentation units within the same construction waste soil sample are then aggregated using area weighting to ensure consistency with the true target mineral content, including: According to the sample number, retrieve the true value of the target mineral content, the predicted value of the target mineral content, the physical latent variable spectrum, and the area ratio of all image segmentation units corresponding to the same construction waste soil sample; The inversion single scattering albedo response value corresponding to the sensitive band in the physical latent variable spectrum is used as the mineral response benchmark, and the color features, texture features and grain morphology features of each image segmentation unit are converted into unit image response content. Based on the degree of matching between the unit image response content and the mineral response benchmark, the mineral response weight corresponding to each image segmentation unit is generated, and the mineral response weight is normalized according to the sum of the mineral response weights of all image segmentation units in the same sample. Based on the normalized mineral response weights, the area proportion of each image segmentation unit, and the target mineral content prediction value output by the physics teacher model, an initial mineral content assignment label is generated for each image segmentation unit. The initial mineral content assignment labels within the same construction waste soil sample are weighted and aggregated according to their area proportion to obtain the sample-level mineral content label aggregation value. Based on the difference between the aggregated value of the sample-level mineral content label and the true value of the target mineral content, the initial mineral content assignment labels are proportionally corrected to obtain the mineral content assignment label corresponding to each image segmentation unit.

8. The mapping model establishment method according to claim 1, characterized in that: Mineral response weights are generated for each image segmentation unit based on the degree of matching between the unit image response content and the mineral response benchmark. The initial mineral content assignment labels are then corrected based on the difference between the aggregated value of the sample-level mineral content labels and the true value of the target mineral content, including: The color features, texture features, and grain morphology features of each image segmentation unit are scaled to obtain color response sub-vectors, texture response sub-vectors, and grain morphology response vectors. The color response subvector, texture response subvector, and grain shape response subvector are connected in a preset order to generate the unit image response content corresponding to each image segmentation unit. The inversion single scattering albedo response value corresponding to the sensitive band is extracted from the physical latent variable spectrum, and the inversion single scattering albedo response value is matched with the target mineral content prediction value output by the physical teacher model to generate the mineral response benchmark corresponding to the target mineral. Calculate the response difference value between the unit image response content of each image segmentation unit and the mineral response benchmark, and determine the matching degree of each image segmentation unit relative to the target mineral according to the response difference value in ascending order; The mineral response weights corresponding to each image segmentation unit are generated based on the matching degree, and the mineral response weights of all image segmentation units in the same construction waste soil sample are added together to obtain the total sample weight. Using the total sample weight as the normalization benchmark, the mineral response weight corresponding to each image segmentation unit is normalized to obtain the normalized mineral response weight. According to the sample number, obtain the initial mineral content allocation label and area ratio corresponding to all image segmentation units within the same construction waste soil sample; Multiply the initial mineral content assignment label corresponding to each image segmentation unit by the area ratio to obtain the area-weighted label value corresponding to each image segmentation unit; The sample-level mineral content label aggregate value is obtained by summing all area-weighted label values ​​within the same construction waste soil sample. The difference between the aggregated value of the sample-level mineral content label and the true value of the target mineral content is calculated to obtain the sample-level label correction difference; Based on the normalized mineral response weight and area ratio of each image segmentation unit, the sample-level label correction difference is allocated to each image segmentation unit in the same construction waste soil sample to obtain the label correction amount corresponding to each image segmentation unit. The initial mineral content assignment label and label correction value corresponding to each image segmentation unit are merged to obtain the mineral content assignment label corresponding to each image segmentation unit. The mineral content assignment labels of all mineral content in the same construction waste soil sample are then aggregated by area weighting to meet the consistency constraint with the true value of the target mineral content.

9. The mapping model establishment method according to claim 1, characterized in that: The image student model performs area-weighted aggregation of the mineral content predictions output by all image segmentation units within the same construction waste soil sample to obtain a sample-level aggregated mineral content prediction. The physics teacher model and the image student model are then updated based on the error between the sample-level aggregated mineral content prediction and the true target mineral content. This includes: According to the sample number, retrieve all image segmentation units corresponding to the same construction waste soil sample, the area ratio of each image segmentation unit, the color features, texture features, particle morphology features and spatial location features of each image segmentation unit, the true value of the target mineral content and the physical latent variable spectrum; The color features, texture features, grain morphology features and spatial location features of each image segmentation unit are input into the image student model to obtain the predicted value of the mineral content of each image segmentation unit, and the correspondence between the predicted value of the mineral content of the unit and the image segmentation unit is maintained. Based on the area proportion of each image segmentation unit, the predicted mineral content of all units in the same construction waste soil sample is weighted and summarized to obtain the sample-level aggregated predicted mineral content. Calculate the aggregation prediction error between the aggregated predicted value of mineral content at the sample level and the true value of target mineral content, and allocate the aggregation prediction error to each image segmentation unit within the same construction waste soil sample to generate the unit error write-back amount; The mapping parameters corresponding to color features, texture features, particle morphology features, and spatial location features in the image student model are updated based on the unit error write-back amount. The output correction parameters between the sensitive spectral index and the predicted value of the target mineral content in the physics teacher model are updated based on the aggregated prediction error. When the aggregated prediction error meets the preset error conditions, a mapping model from close-up image data to the mineral content of construction waste soil is output.

10. The mapping model establishment method according to claim 1, characterized in that: The aggregation prediction error between the aggregated predicted value of mineral content at the sample level and the true value of the target mineral content is calculated. This aggregation prediction error is then allocated to the unit error write-back amount generated by each image segmentation unit within the same construction waste soil sample. Based on the unit error write-back amount and the aggregation prediction error, the image student model and the physics teacher model are updated, including: According to the sample number, obtain the sample-level aggregated predicted value of mineral content, the true value of target mineral content, the predicted value of unit mineral content corresponding to each image segmentation unit, the mineral content allocation label, the area ratio, and the normalized mineral response weight. The difference between the predicted value of the sample-level mineral content and the true value of the target mineral content is calculated to obtain the sample-level aggregation error; The difference between the predicted mineral content value of each image segmentation unit and the mineral content assignment label is calculated to obtain the unit prediction bias. Based on the area ratio, normalized mineral response weight and unit prediction bias of each image segmentation unit, an error allocation coefficient is generated for each image segmentation unit. The sample-level aggregation error is allocated to each image segmentation unit within the same construction waste soil sample according to the error allocation coefficient, and the initial error write-back amount corresponding to each image segmentation unit is obtained. The initial error write-back amount is consolidated to ensure consistency within the same sample, generating the unit error write-back amount for each image segmentation unit. Establish a training correspondence between the unit error write-back amount and the color features, texture features, particle morphology features and spatial location features of the corresponding image segmentation unit; The mapping parameters between image features and predicted mineral content of units in the image student model are adjusted according to the training correspondence to obtain the updated image student model. Establish a correction correspondence between the aggregated prediction error and the true values ​​of the sensitive spectral index, sensitive band, and target mineral content in the physics teacher model; Based on the correction correspondence, the output correction parameters between the sensitive spectral index and the predicted value of the target mineral content in the physics teacher model are adjusted to obtain the updated physics teacher model. Repeatedly execute the image student model output, area-weighted aggregation, aggregation prediction error calculation, unit error write-back generation, image student model update and physics teacher model update until the aggregation prediction error is less than the preset error threshold in consecutive preset rounds. The image student model, physics teacher model, sensitive band, sensitive spectral index, mineral content allocation label generation rules, and area-weighted aggregation rules that meet the preset error conditions are jointly determined as the mapping model that maps near-field image data to the mineral content of construction waste soil.

Images