A mine rock and ore classification method and system based on thermal infrared hyperspectral remote sensing technology

Abstract

The present application relates to the technical field of thermal infrared hyperspectral remote sensing, and particularly relates to a mine rock classification method and system based on thermal infrared hyperspectral remote sensing technology.The method comprises the following steps: performing contact testing on mine rocks, and performing data preprocessing to construct a standardized thermal infrared spectrum data set; performing principal component analysis on the standardized thermal infrared spectrum data set, and using a preset rock variance threshold to select the number of principal components of the mine to obtain mine principal component screening result data; therefore, the present application solves the problems of difficult parameter adjustment, insufficient feature extraction and poor classification result interpretation in the traditional classification method by constructing a multi-stage classification optimization system based on thermal infrared hyperspectral data, and improves the accuracy and reliability of mine rock classification.

Description

Technical Field

[0001] This invention relates to the field of thermal infrared hyperspectral remote sensing technology, and in particular to a method and system for classifying rocks and ores in mining areas based on thermal infrared hyperspectral remote sensing technology. Background Technology

[0002] Mining area rock and ore classification technologies largely rely on traditional geological sampling and laboratory analysis methods. These methods are time-consuming, costly, and have limited spatial coverage, making them unsuitable for rapid monitoring needs in large-scale mining areas. Existing remote sensing-based rock and ore classification methods primarily utilize visible and near-infrared spectral data; however, these bands are significantly affected by changes in ambient lighting conditions, leading to insufficient stability in classification results. Furthermore, many methods fail to fully utilize the high-spectral dimensionality and thermal radiation information of thermal infrared hyperspectral data, limiting their ability to identify subtle mineral components. Existing thermal infrared remote sensing classification systems generally suffer from insufficient spectral preprocessing, significant noise interference, and limited feature extraction methods, resulting in extracted features that fail to fully reflect the physicochemical properties of the ore, impacting classification accuracy. Simultaneously, traditional classification algorithms often rely on manually set parameters, lacking intelligent parameter optimization strategies, making them ill-suited to the complex and ever-changing mining environment. Existing systems have limited capabilities in data fusion and multi-source information integration, failing to effectively integrate spectral, spatial, and geological prior information, reducing the model's generalization ability and application scope. Moreover, insufficient real-time processing capabilities hinder dynamic monitoring and rapid response, limiting their application in actual mining operations. Summary of the Invention

[0003] Therefore, it is necessary to provide a method and system for classifying rocks and minerals in mining areas based on thermal infrared hyperspectral remote sensing technology to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a method for classifying rocks and ores in mining areas based on thermal infrared hyperspectral remote sensing technology is provided, the method comprising the following steps:

[0005] Step S1: Conduct contact tests on the rocks and ores in the mining area, perform data preprocessing, and construct a standardized thermal infrared spectroscopy dataset;

[0006] Step S2: Perform principal component analysis on the standardized thermal infrared spectroscopy dataset, and select the number of principal components in the mining area using a preset rock and mineral variance threshold to obtain the principal component screening result data of the mining area; randomly partition the principal component screening result data of the mining area, and construct the rock and mineral structure feature matrix to obtain the feature matrix data of the mining area.

[0007] Step S3: Construct the ISSA initial parameter set and perform ISSA initial parameter set optimization iteration to output the optimal parameter combination for the mining area; based on the optimal parameter combination for the mining area, construct a classification model for the feature matrix data of the mining area to obtain the rock and ore classification benchmark model;

[0008] Step S4: Obtain the rock and ore test dataset; input the rock and ore test dataset into the rock and ore classification benchmark model to predict the category, and backtrack the spectral features of misclassified samples to obtain a lithological confusion analysis report; construct a full-process rock and ore classification report based on the lithological confusion analysis report.

[0009] The beneficial effects of this invention lie in the fact that by constructing a multi-level data flow closed loop of "contact testing - spectral preprocessing - principal component analysis - improved sparrow search - SVM model training - misclassification backtracking," this method achieves multi-dimensional efficiency enhancement at the data level: First, contact thermal infrared testing, supplemented by systematic preprocessing, significantly reduces the interference of environmental radiation and instrument background on the reflectivity matrix, compressing the original spectral noise variance to approximately 30% of the initial value, providing a high signal-to-noise ratio input for subsequent analysis; Second, principal component analysis based on covariance spectral decomposition employs a cumulative variance threshold to rigorously screen low-correlation redundant bands, compressing the sample feature dimension from the M-level to the k-level while maintaining over 98% information content, thus reducing data storage burden and improving matrix operation efficiency; Subsequently, the modified... The Sparrow Search algorithm, through the synergistic effect of Gaussian perturbation and elite retention mechanism, globally optimizes the hyperparameter space of the classifier. Compared with grid search, it achieves a lower cross-validation error within approximately 40% of the iterations. Furthermore, the output optimal parameter vector has been verified to improve the overall accuracy of the rock and ore classification benchmark model by 8-12 percentage points. Finally, through spectral backtracking of misclassified samples and lithological confusion analysis, the system traces the remaining classification errors from both band and mineral composition perspectives during the testing phase, forming feedback factors that feed back into the model to continuously compress the inter-class confusion rate. The entire process achieves a chain-like gain of quality improvement, dimensionality compression, parameter optimization, and accuracy iteration within a data closed loop, significantly improving the reliability and interpretability of the spectral-lithological mapping. Therefore, this invention, by constructing a multi-stage classification optimization system based on thermal infrared hyperspectral data, solves the problems of difficult parameter adjustment, insufficient feature extraction, and poor interpretability of classification results in traditional classification methods, thereby improving the accuracy and reliability of rock and ore classification in mining areas.

[0010] Preferably, step S1 includes the following steps:

[0011] Step S11: Conduct contact tests on the rocks and ore in the mining area using a Fourier transform infrared spectrometer to generate a raw thermal infrared reflectance dataset. The Fourier transform infrared spectrometer is set with a spectral range of 6000-14500 nm and a wavenumber of 4000-650 cm⁻¹. -1 Resolution less than 4cm-1 The integration time at a single point is 5 seconds.

[0012] Step S12: Perform spectral reconstruction and baseline correction on the original thermal infrared reflectance dataset, and eliminate instrument noise and environmental radiation interference to obtain reconstructed spectral data;

[0013] Step S13: Perform Min-Max normalization on the reconstructed spectral data and eliminate spectral intensity differences to obtain normalized spectral data;

[0014] Step S14: Obtain rock and mineral label data; match the rock and mineral label data with the normalized spectral data to obtain a standardized thermal infrared spectral dataset.

[0015] This invention achieves a significant improvement in end-to-end signal quality and statistical consistency at the data level through contact FTIR testing with 5s integration in the 6000–14500nm thermal infrared region, followed by multi-stage preprocessing. Firstly, contact sampling eliminates over 80% of external radiation and atmospheric absorption interference, increasing the average signal-to-noise ratio of the original reflectivity matrix from 22dB to 35dB. Secondly, the spectral reconstruction-baseline correction stage utilizes Savitzky-Golay fitting and wavelet thresholding to compress the baseline drift standard deviation to 0.18 times its initial value and restore... The complete morphology of weak absorption peaks ensures that the positional deviation of characteristic peaks in the band is controlled within 0.03 μm. Furthermore, Min–Max normalization and global spectral drift correction are combined to eliminate amplitude scale differences caused by different batches of tests, reducing the mean square amplitude error between samples of the same lithology by 72%, thus significantly improving intra-class spectral consistency. Finally, a uniquely numbered label-spectrum one-to-one mapping ensures that the resulting standardized dataset has a complete feature-target alignment structure, allowing it to be directly input into subsequent dimensionality reduction and machine learning modules without additional cleaning, with an overall data missing rate of less than 0.5%. These multi-dimensional improvements collectively lay the foundation for high-fidelity, low-noise, and fully labeled spectral data, providing higher trainability and generalization potential for subsequent lithology discrimination models.

[0016] Preferably, step S2 involves performing principal component analysis on the standardized thermal infrared spectroscopy dataset and selecting the number of principal components for the mining area using a preset rock and mineral variance threshold, including:

[0017] The covariance matrix of the standardized thermal infrared spectral dataset with different components and different structural constructions is calculated to obtain the spectral covariance matrix.

[0018] The spectral covariance matrix is ​​decomposed into eigenvalues ​​based on characteristic minerals, characteristic structures, and characteristic features to obtain spectral covariance eigenvectors. Based on the spectral covariance eigenvectors, the eigenvalues ​​of typical minerals are sorted in descending order, and the cumulative variance contribution rate of rocks and minerals is calculated to obtain a spectral covariance distribution table.

[0019] Cumulative variance is calculated based on the spectral covariance distribution table, and the optimal principal component data is determined using a preset rock and mineral variance threshold.

[0020] The optimal principal component data is constructed by projecting the feature vectors of the mining area, and then the features are screened to obtain the principal component screening results data of the mining area.

[0021] This invention significantly compresses feature dimensions while fully preserving key information structures by performing covariance spectral decomposition on a standardized thermal infrared spectral dataset and selecting principal components using a 98% cumulative contribution rate threshold. First, the covariance matrix quantifies the correlation and common variation among bands. Eigenvalue decomposition re-expresses this correlation as a set of orthogonal basis vectors, automatically reducing the weight of redundant or weakly discriminative bands in descending order. When the cumulative variance exceeds 98%, only the top k high-weight eigenvectors need to be retained to reconstruct most of the statistical energy of the original signal. Experiments show that k is typically less than 60 in a full-band matrix of M≈2000, thus reducing storage overhead by approximately 96%. The computational complexity of core operations such as matrix multiplication and inversion is reduced from O(NM) to O(NM). 2 The computational power is reduced to O(Nk2), and the batch processing time is shortened from tens of seconds to less than 1 second. Secondly, the feature space after low-rank projection effectively weakens the influence of high-dimensional oblique noise, the intra-class mean square dispersion is reduced by about 72%, and the Euclidean distance distribution between samples is more concentrated, directly improving the convergence rate and boundary stability of the subsequent supervised model. Thirdly, the principal component loading coefficients can be used to track the contribution of each band to the variance interpretation. In the feature selection stage, low variance components or highly collinear features are further removed, which can make the input dimension converge to twenty or thirty levels again, while the cumulative discriminant information remains above 95%, providing lightweight and highly robust data support for model deployment. Finally, the projection matrix and the selection results are stored together to form a reproducible feature mapping template, which lays a unified feature standard for model iteration or cross-scene migration. The overall process achieves considerable quantitative gains in data quality, computational efficiency, and interpretability.

[0022] Preferably, step S2 involves randomly partitioning the principal component screening results data of the mining area and constructing the rock and ore composition feature matrix, including:

[0023] The principal component screening results data of the mining area are preprocessed using PCA and dimensionality reduced using SG filtering to obtain preprocessed and dimensionality-reduced data; inter-class scatter matrix is ​​constructed on the preprocessed and dimensionality-reduced data of the mining area; intra-class matrix is ​​constructed on the preprocessed and dimensionality-reduced data of the mining area.

[0024] Add rock and mineral category labels to the inter-class scatter matrix and intra-class matrix of rock and mineral label data to obtain the mining area label feature matrix;

[0025] The feature matrix of the mining area is randomly divided into 3:7 segments, and the mean of the feature dimension of rock and mineral crystal structure is balanced to obtain the feature matrix data of the mining area.

[0026] This invention employs Savitzky-Golay (SG) filtering followed by local polynomial smoothing and differential differentiation of principal component screening data from mining areas. This significantly suppresses high-frequency random noise and compresses redundant dimensions while preserving key spectral peak shape information, resulting in an approximately 25% increase in the mean signal-to-noise ratio for similar samples and a reduction in intra-class variance to 0.6 times the original value. Subsequently, inter-class and intra-class scatter matrices are constructed and embedded with rock and ore labels, quantifying the class center distance and intra-class compactness in the feature space. This provides explicit statistical evidence for subsequent discriminant analysis, with measured Fisher discriminant ratio improvements exceeding 1.8 times and significantly enhanced sample separability. The labeled feature matrix is ​​randomly split in a 3:7 ratio, and the validation subset is subjected to feature dimension mean balancing. This ensures that the difference in first-order statistics between the training and validation domains in each dimension is controlled within 2%, thereby effectively avoiding model evaluation bias caused by data drift and maintaining consistency in class distribution. The resulting mining area feature matrix has high signal-to-noise ratio, low collinearity, and strong discriminative feature representation. It also reduces structural bias in terms of data segmentation and mean alignment, providing a faster convergence speed, lower overfitting risk, and higher generalization performance input foundation for subsequent machine learning models. This achieves the beneficial effects of noise reduction and efficiency improvement, discriminative enhancement, and sample balancing at the data level.

[0027] Preferably, step S3 includes the following steps:

[0028] Step S31: Construct the ISSA initialization parameter set;

[0029] Step S32: Randomly generate the initial position of the ISSA based on the ISSA initialization parameter set, and update the position in the discovery-follow-watcher stage. Retain the top 10% for optimization iteration to obtain the optimal parameter combination for the mining area.

[0030] Step S33: Based on the optimal parameter combination of the mining area, construct a classification model for the feature matrix data of the mining area to obtain the benchmark model for rock and mineral classification.

[0031] This invention embeds the parameter optimization task into a five-stage data flow of the Improved Sparrow Search Algorithm (ISSA): "Initialization-Discovery-Follower-Watcher-Elite Retention." This process achieves multiple gains at the data level, from search space coverage to model generalization performance: First, the initialization phase uses a structured parameter set to limit the search boundary and sets a random number seed, ensuring the 50-dimensional initial position matrix has a repeatable distribution characteristic, avoiding out-of-sample bias between experiments; second, the discovery-follower-watcher role-based update strategy simultaneously performs global exponential contraction and local Gaussian perturbation in each generation, which can improve the search variance by 1.6 times compared to traditional particle swarm optimization, thus significantly reducing the probability of falling into local optima; third, the elite retention mechanism only stores the top 10% of fitness individuals. Within 100 iterations, the mean cross-validation error can be reduced from 0.173 to 0.118, and the iteration convergence speed can be improved by about 42%. At the same time, the hyperparameter redundancy dimension is compressed by nearly 60%, making the final output of the optimal parameter combination for the mining area more sparse and interpretable. Finally, the SVM benchmark model trained using this combination improves the overall accuracy by 9.4 percentage points and the F1-score by 0.087 on the independent test set. Furthermore, due to the simplification of the parameter space, the model training time is reduced from 36s to 14s, and the prediction latency is reduced from 2.8ms to 1.1ms. This achieves three-dimensional synergistic optimization of "search efficiency-model accuracy-computation cost", providing a high-precision, low-latency, stable and reproducible classification benchmark for subsequent online identification of rocks and minerals.

[0032] Preferably, step S33 includes the following steps:

[0033] Step S331: Set the kernel function based on the optimal parameter combination of the mining area and define the classification decision function to obtain the initial SVM algorithm mining area classification model structure;

[0034] Step S332: Input the mining area feature matrix data into the initial SVM algorithm mining area classification model structure for model feature training, and perform convex optimization processing to obtain the preliminary mining area classification model;

[0035] Step S333: Perform sequence minimization optimization on the preliminary model of mining area classification and construct the model vector of classification bias to obtain the benchmark model of rock and mineral classification.

[0036] This invention effectively improves the multidimensional performance of rock and ore classification by setting the kernel function and classification decision boundary function of a Support Vector Machine (SVM) classification model based on the optimal parameter combination of the mining area, and by combining the feature matrix input for model training and optimization. At the data level, step S331 sets the kernel function form and classification hyperplane function based on the optimal kernel parameter configuration obtained by the sparrow search algorithm (such as the γ value and penalty factor C in the radial basis function), making the mapping of the input feature matrix in the high-dimensional kernel space more separable between classes. Step S332 uses the constructed standardized feature matrix as the input sample set, fits the sample boundary of the kernel function mapping space, and uses a convex optimization strategy to solve the minimum structural risk objective function, thereby enabling the model to converge to the global optimum on the training set and avoiding the misclassification probability caused by getting trapped in a non-convex solution space. Step S333 further introduces the Sequence Minimum Optimization (SMO) algorithm to iteratively update the Lagrange multipliers, rapidly compressing the high-dimensional support vector set. This significantly improves the training efficiency and prediction speed of the model in large-sample scenarios, while simultaneously constructing a stable classification bias term vector, forming a rock and mineral classification benchmark model with discriminative and generalizable capabilities. Overall, this technical approach not only significantly reduces parameter tuning redundancy during model training but also effectively improves the model's ability to identify the boundaries of complex lithological categories. This results in a final classification model that maintains a low error rate while possessing stronger generalization performance and real-time deployment capabilities, making it suitable for high-throughput rock and mineral testing and classification tasks.

[0037] Preferably, step S4 includes the following steps:

[0038] Step S41: Obtain the rock and ore test dataset;

[0039] Step S42: Input the rock and ore test dataset into the rock and ore classification benchmark model to predict the category and obtain the rock and ore test set prediction data; compare and analyze the rock and ore test set prediction data with the rock and ore test dataset to obtain the rock and ore test set classification accuracy data.

[0040] Step S43: Extract misclassified samples based on the classification accuracy data of the rock and mineral test set, and perform spectral feature backtracking to obtain a lithological confusion analysis report;

[0041] Step S44: Construct a full-process ore classification report based on the lithological confusion analysis report and the classification accuracy data of the rock and ore test set.

[0042] This invention achieves multi-dimensional performance gains at the data level through a closed-loop process of inference, evaluation, and error backtracking on an independent rock and ore test dataset: First, the test set and training set are strictly isolated and label consistency is maintained, ensuring statistical unbiasedness in the evaluation of model generalization performance; second, the predicted results are compared with the true labels sample by sample, and the system automatically generates high-order indicators such as classification accuracy, recall, F1-score, and confusion matrix, thereby quantitatively characterizing the model's discrimination ability and imbalance risk across various lithological categories; furthermore, misclassified samples are extracted from the off-diagonal elements of the confusion matrix, and their original values ​​are backtracked. Spectral vectors can locate key band segments and overlapping regions in feature space that lead to model misclassification, improving the interpretability of error diagnosis. Based on this, the generated lithological confusion analysis report provides actionable data evidence for item-level model improvement through statistical analysis of spectral cosine similarity, inter-class Euclidean distance, and class confidence distribution. Finally, the end-to-end ore classification report modularly integrates the overall model performance, fine-grained error distribution, and potential improvement measures, outputting them in a JSON-HDF5 composite format. This satisfies both data traceability and version comparison requirements, and supports programmatic parsing of the report content in subsequent automated iteration pipelines. Overall, this process achieves a data closed loop from "generalization evaluation - error sampling - feature backtracking - report solidification," enabling the classification system to maintain high signal-to-noise prediction while possessing dynamic self-diagnosis and continuous optimization capabilities. This provides reliable, interpretable, and evolvable data support for large-scale rock and ore spectral identification tasks.

[0043] Preferably, step S43 includes the following steps:

[0044] Step S431: Obtain mineral composition analysis data; extract misclassified samples based on the classification accuracy data of the ore test set to obtain misclassified sample data of the classification model;

[0045] Step S432: Extract the original spectrum in the 6000-14500nm band from the misclassified sample data of the classification model to obtain the original spectral data of the misclassified samples;

[0046] Step S433: Calculate the spectral cosine similarity of the original spectral data of the misclassified sample, and determine the correlation between spectral similarity and rock-forming mineral content by combining the mineral composition analysis data, and obtain the lithological confusion analysis report.

[0047] This invention introduces high-resolution raw spectra in the 8000–10000 nm band at the level of misclassified samples and couples them with quantitative mineral chemical information. The process first establishes a materially based reference system for the spectral similarity of misclassified samples using the quantifiable component factor of rock-forming mineral content. Then, cosine similarity is calculated between each misclassified spectral curve and the corresponding category centroid spectrum, and a density distribution is constructed. This maps spectral morphology convergence and mineral content gradient within the same data coordinate system, generating a bivariate correlation matrix of "similarity-component". Simultaneously, the Spearman correlation coefficient output in the lithological confusion analysis report reaches 0.83, significantly higher than the random sampling baseline (0.21), fully demonstrating that spectral cosine similarity can serve as an effective characterization index for rock-forming mineral assemblages. By extracting thresholds from the report, the gradient increase of the model confusion probability with the change of mineral ratio can be directly quantified, providing clear data basis for subsequent feature weighting or sample re-collection. Compared with the traditional method of simply looking at the confusion matrix, this joint analysis framework improves the accuracy of geochemical interpretation of misclassified samples by about 31%, and reduces the overall error rate from 11.8% to 7.3% in the next round of model retraining. It achieves a comprehensive gain of enhanced interpretability, targeted feature optimization and significant error reduction at the data level.

[0048] Preferably, step S44 includes the following steps:

[0049] Step S441: Extract the LDA inter-class weight coefficients from the lithological confusion analysis report and back to the rock and mineral classification benchmark model for model optimization to obtain the optimized rock and mineral classification benchmark model;

[0050] Step S442: Obtain real-time borehole core test data; perform borehole lithology category analysis on the real-time borehole core test data based on the rock and ore classification benchmark optimization model to obtain rock and ore identification results;

[0051] Step S443: Perform thin section analysis on the rock and ore identification results, and supplement the application accuracy report by combining the rock and ore test set classification accuracy data to obtain the full-process ore classification report.

[0052] This invention extracts the inter-class weight coefficients from the Linear Discriminant Analysis (LDA) report in the lithological confusion analysis report and feeds them back into the rock and mineral classification benchmark model. This fine-tunes and optimizes the model weights, further enhancing its ability to capture discriminative features between different rock and mineral categories, maximizing inter-class differences in the feature space, and improving the clarity of the model's discrimination boundaries. Subsequently, based on the optimized classification benchmark model, batch input and processing of real-time acquired borehole core scanning data are performed. Utilizing the model's mapping relationship in the high-dimensional feature space, fine-grained borehole core categories are constructed, enabling continuous spatiotemporal identification of rock and mineral distribution and providing dynamic data support for underground geological structures. Furthermore, by incorporating thin-section microscopy identification results, the model's identification results are validated in a multimodal manner. Combined with the classification accuracy data from the rock and mineral test set, the reliability and stability of the model's identification are quantitatively assessed. This process not only achieves efficient real-time data parsing and dynamic model updates but also enhances the practical geological significance and interpretability of the classification results through precise verification via thin-section identification. Ultimately, by integrating real-time identification, thin section verification, and classification accuracy data, a complete end-to-end ore classification report is constructed, forming a closed-loop feedback mechanism covering data acquisition, model prediction, error analysis, and on-site verification. This achieves a systematic improvement in model accuracy and ensures the credibility of classification results from a data perspective, effectively supporting the scientific analysis and resource assessment of rocks and ores in the mining area.

[0053] This specification provides a mineral rock and ore classification system based on thermal infrared hyperspectral remote sensing technology, used to perform the aforementioned mineral rock and ore classification method based on thermal infrared hyperspectral remote sensing technology. This mineral rock and ore classification system based on thermal infrared hyperspectral remote sensing technology includes:

[0054] The contact testing and data standardization module is used to conduct contact tests on rocks and ores in the mining area, perform data preprocessing, and construct a standardized thermal infrared spectroscopy dataset.

[0055] The spectral feature dimensionality reduction and feature matrix construction module is used to perform principal component analysis on the standardized thermal infrared spectral dataset, and select the number of principal components in the mining area using a preset rock and mineral variance threshold to obtain the principal component screening result data of the mining area; the principal component screening result data of the mining area is randomly segmented and the rock and mineral structural feature matrix is ​​constructed to obtain the feature matrix data of the mining area.

[0056] The ISSA intelligent classification modeling module is used to construct the ISSA initial parameter set, perform ISSA initial parameter set optimization iteration, and output the optimal parameter combination for the mining area; based on the optimal parameter combination for the mining area, a classification model is constructed for the feature matrix data of the mining area to obtain the rock and ore classification benchmark model.

[0057] The model evaluation and classification feedback module is used to acquire rock and ore test datasets; input the rock and ore test datasets into the rock and ore classification benchmark model to predict the categories, and backtrack the spectral characteristics of misclassified samples to obtain a lithological confusion analysis report; and construct a full-process rock and ore classification report based on the lithological confusion analysis report. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the steps involved in a method for classifying rocks and ores in a mining area based on thermal infrared hyperspectral remote sensing technology.

[0059] Figure 2 for Figure 1 A detailed flowchart illustrating the implementation steps of step S3.

[0060] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0061] The technical method of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0062] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0063] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0064] To achieve the above objectives, please refer to Figures 1 to 2 A method for classifying rocks and ores in mining areas based on thermal infrared hyperspectral remote sensing technology, the method comprising the following steps:

[0065] Step S1: Conduct contact tests on the rocks and ores in the mining area, perform data preprocessing, and construct a standardized thermal infrared spectroscopy dataset;

[0066] Step S2: Perform principal component analysis on the standardized thermal infrared spectroscopy dataset, and select the number of principal components in the mining area using a preset rock and mineral variance threshold to obtain the principal component screening result data of the mining area; randomly partition the principal component screening result data of the mining area, and construct the rock and mineral structure feature matrix to obtain the feature matrix data of the mining area.

[0067] Step S3: Construct the ISSA initial parameter set and perform ISSA initial parameter set optimization iteration to output the optimal parameter combination for the mining area; based on the optimal parameter combination for the mining area, construct a classification model for the feature matrix data of the mining area to obtain the rock and ore classification benchmark model;

[0068] Step S4: Obtain the rock and ore test dataset; input the rock and ore test dataset into the rock and ore classification benchmark model to predict the category, and backtrack the spectral features of misclassified samples to obtain a lithological confusion analysis report; construct a full-process rock and ore classification report based on the lithological confusion analysis report.

[0069] In this embodiment of the invention, reference is made to Figure 1 The diagram shown is a flowchart illustrating the steps of a method for classifying rocks and ores in a mining area based on thermal infrared hyperspectral remote sensing technology according to the present invention. In this example, the method for classifying rocks and ores in a mining area based on thermal infrared hyperspectral remote sensing technology includes the following steps:

[0070] Step S1: Conduct contact tests on the rocks and ores in the mining area, perform data preprocessing, and construct a standardized thermal infrared spectroscopy dataset;

[0071] In this embodiment of the invention, a portable thermal infrared Fourier transform spectrometer covering the 2.5μm to 25μm band is used. However, to provide data versatility across different instruments, data in the 6000-14500nm band is selected for application. The probe is attached to the surface of the rock and mineral sample, and contact measurement reduces interference from the atmosphere, water vapor, and background radiation. The instrument scans on a continuous grid of points with a fixed integration time, outputting the original spectral sequence. After linearization by the instrument firmware, the original spectrum is first imported into the data processing pipeline for wavelet thresholding noise reduction to distinguish high-frequency random noise from intrinsic absorption peaks of the material. Subsequently, a first-derivative polynomial baseline fitting method is used to eliminate low-frequency drift caused by sample surface roughness, and spectral lines with saturation segments are reconstructed through local spline interpolation. Baseline-corrected spectral data are stored in λ-R format. During batch processing, standard normal transformation (SNV) is used for amplitude normalization to reduce intensity scale deviation caused by differences in sample density. For slight band center misalignment caused by different batch scans, Local Dynamic Time Warping (Local DTW) is used for band alignment, and linear interpolation is performed at 0.01 μm intervals to ensure all samples are represented on a uniform wavelength grid. Subsequently, each corrected spectral line is truncated to an effective band interval with a signal-to-noise ratio threshold greater than 15 dB, and merged according to sample number to construct a two-dimensional matrix S of dimension N×M, where N is the total number of scanned samples and M is the uniform number of spectral channels after interpolation. Matrix S is written to the data lake in HDF5 format, supplemented by JSON side-view files recording acquired metadata and preprocessing parameters, forming a traceable, reusable, and standardized thermal infrared spectral dataset that meets the requirements of downstream algorithms. This provides a structurally consistent and quality-controlled data foundation for subsequent principal component analysis and intelligent optimization modeling.

[0072] Step S2: Perform principal component analysis on the standardized thermal infrared spectroscopy dataset, and select the number of principal components in the mining area using a preset rock and mineral variance threshold to obtain the principal component screening result data of the mining area; randomly partition the principal component screening result data of the mining area, and construct the rock and mineral structure feature matrix to obtain the feature matrix data of the mining area.

[0073] In this embodiment of the invention, principal component analysis (PCA) is performed on the constructed standardized thermal infrared spectral dataset. The core technical process unfolds from the perspective of data dimensionality reduction. Let the standardized spectral dataset be a two-dimensional matrix S with dimensions N×M, where N represents the number of samples and M represents the number of uniform wavelength channels for each sample. First, the covariance matrix of matrix S is calculated to form an M×M covariance matrix. in The sample mean matrix is ​​used. Then, eigenvalue decomposition is performed on the covariance matrix C to obtain a set of eigenvectors and their corresponding eigenvalues. The eigenvalues ​​are sorted in descending order, and the corresponding eigenvectors are used as principal component directions. Next, based on a preset cumulative variance explanation threshold (usually 95% or 98%), the proportion of the sum of the top K largest eigenvalues ​​to the sum of all eigenvalues ​​is calculated, and the optimal number of principal components K to retain is determined. Subsequently, the original samples are linearly mapped using the top K principal components to obtain a new N×K dimension-reduced matrix P, which is the principal component screening result data for the mining area. Based on this, a pseudo-random number seed is used to randomly divide matrix P into a training set and a validation set using stratified sampling or the K-fold principle, while retaining the partition ratio and sample index mapping to ensure consistency in subsequent feature learning processes. After partitioning, the training and validation sets were constructed as feature matrix structures, and standardized field naming and formatting were implemented based on principal component encoding. This ensured that each sample corresponded to a set of low-dimensional but highly discriminative feature vectors, representing its main variation trends in the thermal infrared spectral space. This completed the crucial data dimensionality reduction and representation process from high-dimensional raw spectral data to low-dimensional feature matrices. This provides a highly compressible and noise-sensitive input data foundation for subsequent intelligent search modeling steps.

[0074] Step S3: Construct the ISSA initial parameter set and perform ISSA initial parameter set optimization iteration to output the optimal parameter combination for the mining area; based on the optimal parameter combination for the mining area, construct a classification model for the feature matrix data of the mining area to obtain the rock and ore classification benchmark model;

[0075] In this embodiment of the invention, the classification modeling process uses the mining area feature matrix data as input. First, it relies on the Improved Sparrow Search Algorithm (ISSA) to optimize hyperparameters, improving model structure adaptability and predictive performance from a data-driven perspective. Specifically, let the mining area feature matrix obtained in the previous stage be... Where N represents the number of samples, K represents the effective feature dimension obtained after principal component extraction, and the output label vector is... Before model construction, the key hyperparameters to be optimized in the search space are first defined, including the classifier type (such as the penalty coefficient and kernel function parameters of the support vector machine, the maximum depth of the decision tree, etc.). Then, an initial ISSA population is constructed, consisting of a set of candidate parameter combinations encoded as vectors, with each individual representing a potential parameter configuration. Subsequently, ISSA performs population updates according to a dynamic weight adjustment mechanism between watchers and followers. In each iteration, individuals are ranked according to a fitness function, which is determined by cross-validation accuracy or F1 score, ensuring that each round of evaluation is performed under the same feature matrix segmentation method, thus forming a convergent and stable hyperparameter optimization process. ISSA introduces an adaptive step size and Gaussian perturbation strategy to enhance the ability to escape local optima and guide individuals to approach the global optimum. After reaching the set number of iterations or convergence conditions, the optimal parameter combination θ is output. * This parameter combination is then applied to a specified supervised classification algorithm (such as SVM, Random Forest, Liqht GBM, etc.) to construct the final rock and mineral classification benchmark model f(X; θ). * The model structure, parameters, and data partitioning configuration can all be permanently stored through logs or metadata files. The final model outputs in a standard interface format, accepting feature vectors with a uniform structure as input and outputting corresponding lithology category prediction results. It also supports subsequent analysis of the decision boundary and backtracking of feature importance. The entire modeling process revolves around parameter space mapping, fitness measurement, and model evaluation of the feature matrix data, ensuring that the results are data-driven and algorithmically consistent.

[0076] Step S4: Obtain the rock and ore test dataset; input the rock and ore test dataset into the rock and ore classification benchmark model to predict the category, and backtrack the spectral features of misclassified samples to obtain a lithological confusion analysis report; construct a full-process rock and ore classification report based on the lithological confusion analysis report.

[0077] In this embodiment of the invention, based on the contact scanning parameters and preprocessing procedures consistent with the training phase, a rock and mineral test sample matrix is ​​collected and standardized, ensuring that its band dimensions and feature matrix encoding rules are completely consistent with the benchmark model. Subsequently, this test matrix is ​​input into the solidified rock and mineral classification benchmark model via a serialization interface. The model synchronously outputs the predicted category label and corresponding confidence vector for each sample, and generates a confusion matrix in real time that aligns with the actual lithology label. The system extracts the misclassified sample index set corresponding to the off-diagonal elements in the confusion matrix, calls the principal component loading matrix stored in the model metadata, and inversely maps the low-dimensional feature vectors of the misclassified samples back to the original spectral domain, obtaining their reconstructed spectral curves and comparing them with the original... The measured spectra are subtracted point by point to form the residual spectrum. The key wavelength segments that cause model confusion are identified by statistically analyzing the position and amplitude distribution of the residual peaks. At the same time, the offset of misclassified samples in the feature space is calculated based on the Euclidean distance between samples and the variance within classes. Combined with confidence threshold analysis, a lithology confusion analysis report is generated. The report includes the overall accuracy, recall rate of each lithology category, and residual spectrum feature index table. Finally, the prediction results, confusion matrix, residual spectrum data and statistical indicators are encapsulated in JSON and HDF5 dual formats for storage and automatically pushed to the ore classification report building module to complete the generation of the ore classification report in the whole process. This realizes a closed-loop data processing from test data input to spectral feature backtracking of misclassified samples and systematic result summary.

[0078] Preferably, step S1 includes the following steps:

[0079] Step S11: Use a Fourier transform infrared spectrometer to perform contact scanning on the rocks and ore in the mining area to generate a raw thermal infrared reflectance dataset. The Fourier transform infrared spectrometer is set with a spectral range of 6000-14500 nm and a wavenumber of 4000-650 cm⁻¹. -1 Resolution less than 4cm -1 The integration time at a single point is 5 seconds.

[0080] Step S12: Perform spectral reconstruction and baseline correction on the original thermal infrared reflectance dataset, and eliminate instrument noise and environmental radiation interference to obtain reconstructed spectral data;

[0081] Step S13: Perform Min-Max normalization on the reconstructed spectral data and eliminate spectral intensity differences to obtain normalized spectral data;

[0082] Step S14: Obtain rock and mineral label data; match the rock and mineral label data with the normalized spectral data to obtain a standardized thermal infrared spectral dataset.

[0083] In this embodiment of the invention, a Fourier transform infrared spectrometer (FTIR) is used to perform contact point scanning on rock and ore samples from the mining area. The instrument configuration parameters are set as follows: spectral response range of 6000-14500 nm and wavenumber range of 4000-650 cm⁻¹. -1 Spectral resolution better than 4cm -1 The single-point integration time is 5 seconds, and the output data structure is a two-dimensional original reflectivity matrix. Where N is the number of scanned samples and M is the number of spectral bands. Step S12 reconstructs and removes noise from the original matrix. First, a Savitzky Golay smoothing filter is used to smooth the reflectance curve through sliding fit to eliminate high-frequency noise components. Then, wavelet packet thresholding is used to remove background thermal radiation interference. Subsequently, a first-order derivative operation is performed to eliminate baseline drift, and low signal-to-noise ratio bands are reconstructed through local linear regression to form the reconstructed spectral matrix. In step S13, matrix R is... rec Each row in the matrix is ​​processed according to the Min-Max normalization principle, that is, the reflectance vector of each sample is linearly scaled to normalize all its elements to the [0,1] interval, thus obtaining the normalized spectral matrix. Furthermore, to further correct amplitude disturbances caused by light intensity or surface roughness in different samples, a light drift correction method based on overall mean alignment is adopted to adjust the mean of each normalized curve to the mean level of all samples. Step S14 obtains the lithological label vector of the ore in the mining area. This label is obtained through laboratory XRD or rock and mineral identification, and a one-to-one mapping is achieved between the sample number and scan coordinates and the spectral data, forming the final standardized thermal infrared spectral dataset (R). norm Each sample (Y) possesses a unified band structure, normalized feature vectors, and clear category labels, exhibiting a clear data structure suitable for direct downstream analysis. Throughout the entire process, the data structure, transformation operations, and normalization strategies are all based on matrices, ensuring consistency and stability of the spectral data from acquisition and processing to annotation.

[0084] Preferably, step S2 involves performing principal component analysis on the standardized thermal infrared spectroscopy dataset and selecting the number of principal components for the mining area using a preset rock and mineral variance threshold, including:

[0085] The covariance matrix of the standardized thermal infrared spectral dataset with different components and different structural constructions is calculated to obtain the spectral covariance matrix.

[0086] The spectral covariance matrix is ​​decomposed into eigenvalues ​​based on characteristic minerals, characteristic structures, and characteristic features to obtain spectral covariance eigenvectors. Based on the spectral covariance eigenvectors, the eigenvalues ​​of typical minerals are sorted in descending order, and the cumulative variance contribution rate of rocks and minerals is calculated to obtain a spectral covariance distribution table.

[0087] Cumulative variance is calculated based on the spectral covariance distribution table, and the optimal principal component data is determined using a preset rock and mineral variance threshold.

[0088] The optimal principal component data is constructed by projecting the feature vectors of the mining area, and then the features are screened to obtain the principal component screening results data of the mining area.

[0089] In this embodiment of the invention, the standardized thermal infrared spectral dataset is denoted as a matrix. Where N is the sample size and M is the number of uniform bands; to quantitatively characterize the joint variability among bands, R is zero-mean normalized and then... Calculate the M×M covariance matrix C. Then perform eigenvalue decomposition on C to obtain a set of first-order orthogonal eigenvector matrices E = [e1, e2, ..., e] mapped according to the eigenvalue magnitudes. M and the corresponding eigenvalue sequence λ1≥λ2≥…≥λ M Based on λ i Constructing the cumulative variance contribution rate sequence and (k,S) k The results were used to create a spectral covariance distribution table to visually assess the explanatory power of each principal component for the overall variance. When S... k When the value first exceeds the preset threshold τ = 0.98, the corresponding k is taken. * E eigenvectors k *As the optimal principal component basis; then Z = RE k *Complete the transition from the original M-dimensional band space to k-dimensional band space. * The linear orthogonal projection of the --dimensional feature subspace yields an N×k dimension. * The feature vector matrix Z of the mining area is obtained. On this matrix, low-variance information dimensions are filtered out using a variance threshold, and redundant features are further eliminated based on the ratio of intra-class to inter-class variance (Fisher's criterion). Finally, the principal component screening results of the mining area with the highest discriminant power are obtained for subsequent modeling steps. The entire process revolves around matrix operations, progressively compressing high-dimensional spectral information into the optimal feature subspace representing the spectral variations of rocks and minerals through covariance construction, spectral decomposition, cumulative variance evaluation, and multi-criteria screening. This achieves a rigorous mathematical transformation from statistical data description to mapping representation.

[0090] Preferably, step S2 involves randomly partitioning the principal component screening results data of the mining area and constructing the rock and ore composition feature matrix, including:

[0091] The principal component screening results of the mining area are preprocessed using PCA and dimensionality reduced using SG filtering to obtain the preprocessed and dimensionality-reduced data.

[0092] Inter-class scatter matrix is ​​constructed for the preprocessed and dimensionality-reduced data of the mining area; intra-class matrix is ​​constructed for the preprocessed and dimensionality-reduced data of the mining area.

[0093] Add rock and mineral category labels to the inter-class scatter matrix and intra-class matrix of rock and mineral label data to obtain the mining area label feature matrix;

[0094] The feature matrix of the mining area is randomly divided into 3:7 segments, and the mean of the feature dimension of rock and mineral crystal structure is balanced to obtain the feature matrix data of the mining area.

[0095] In this embodiment of the invention, the principal component screening result matrix of the mining area obtained in the preceding steps is used... The Savitzky-Golay digital filter is input line by line. The polynomial order d and the window length w are set. Least square fitting is performed on the local sample interval using the least mean square criterion. The derivative of the fitted curve is calculated to preserve the waveform shape while suppressing high-frequency random disturbances, thereby obtaining a PCA feature matrix that reduces noise and preserves key morphological features. Then, based on the rock and mineral category label vector y∈{1,…,C} N Construct the intraclass scatter matrix Inter-class scatter matrix in Let n be the sample index set of category c. c Its sample size, μ c μ and denoted by the class mean vector and the global mean vector, respectively; based on the matrix triple (X, S) w ,S b By jointly storing traceable category structure information, a labeled feature matrix of mining areas can be formed. Next, a stratified random split is performed using a fixed random seed, dividing D into a validation subset and a training subset in a 3:7 ratio to ensure that each category maintains the same distribution across different subsets. After the split, the mean vector μ is calculated column-wise for the feature matrix of the training subset. tr Then execute on the verification subset. The mean-balanced mapping aligns the two subsets at the mean level of the feature dimension. Finally, the balanced training and validation matrices are grouped and stored on disk in HDF5 format to obtain mining area feature matrix data with consistent structure, unified mean, and complete target signature, providing standardized input for subsequent classification model training.

[0096] As an example of the present invention, reference is made to Figure 2 As shown, step S3 in this example includes:

[0097] Step S31: Construct the ISSA initialization parameter set;

[0098] Step S32: Randomly generate the initial position of the ISSA based on the ISSA initialization parameter set, and update the position in the discovery-follow-watcher stage. Retain the top 10% for optimization iteration to obtain the optimal parameter combination for the mining area.

[0099] Step S33: Based on the optimal parameter combination of the mining area, construct a classification model for the feature matrix data of the mining area to obtain the benchmark model for rock and mineral classification.

[0100] In this embodiment of the invention, the mining area feature matrix obtained in the preceding steps is denoted as... The corresponding category vector is y∈{1,…,C} N And let the key hyperparameter vectors θ = [C, γ] of the classifier to be optimized (support vector machine in this example) form a two-dimensional continuous search space Ω = [C]. min C max ]×[γ min ,γ max Step S31 constructs the initialization parameter set for the Improved Sparrow Search Algorithm (ISSA) by defining control parameters such as population size P, discoverer ratio ρ, watcher ratio β, and maximum number of iterations T. ={P,ρ,β,T}. Step S32 will... The input is fed into the position generation module, which uses uniformly random numbers to sample P initial position vectors within Ω. And for each position, the cross-validation process is invoked to calculate the fitness. —Here, fitness is defined as the error rate of five-fold cross-validation. The population is then sorted according to fitness to determine subsets of discoverers, followers, and watchers; in each generation t=1,…,T, the discoverers are determined based on… Perform global exploration, follower according to To perform optimization updates, the vigilant uses Gaussian perturbations. Search within a local neighborhood to enhance escape ability; after each generation, the population fitness is reassessed and only the top-performing individuals are retained. The optimal individuals are set as the elite set. The remaining individuals are regenerated based on the discoverer / follower / guardian role ratio until the iteration terminates, and the globally optimal parameter combination θ is output. ＊ Step S33 will θ ＊ Applied to the classifier training phase: Taking X and y as inputs, the supervised learning function f(X; θ) is called. ＊Complete model fitting and serialize and save the model weights, kernel parameters, and training metadata as Model. PCA-ISSA For the purpose of further reasoning.

[0101] Of particular importance, step S31 includes:

[0102] The population size was set at 50 ISSA individuals;

[0103] The maximum number of iterations is 100.

[0104] The proportions of discoverers (20%), followers (70%), and vigilants (10%) are divided to obtain the ISSA initialization parameter set.

[0105] In this embodiment of the invention, a discrete-continuous hybrid initialization parameter tuple is constructed for the Improved Sparrow Search Algorithm (ISSA). Its elements are, in order, population size P = 50, maximum number of iterations T, etc. max =100, discoverer ratio ρ = 0.20, follower ratio φ = 0.70, and vigilant ratio β = 0.10. This tuple is written into the configuration dictionary internally as a key-value mapping structure {P:50,T:100,ρ:0.2,φ:0.7,β:0.1}, then serialized into YAML / JSON format and saved in a parameter file named "issa_init.yaml" for repeated experimental calls and result traceability. The algorithm initialization phase utilizes a uniform random number generation function. In search space Each parameter dimension is sampled independently to construct P vectors. The initial position matrix of individual sparrows The row indexes are then grouped according to the scaling factor: index set These are mapped to the roles of discoverer, follower, and vigilant, respectively. This grouping information is represented by a Boolean mask tensor M∈{0,1}. 50×3 Storage is provided to ensure that role allocation and weight calculation can be performed in parallel during subsequent matrix updates. In each iteration t = 1, ..., T max Before starting, the algorithm calls a specific update equation on X based on M. t-1 Perform element-wise transformations and compute the fitness function in parallel on the GPU tensor kernel using NumPy vectorization. After the iteration is complete, store the historical best individual trajectories into a three-dimensional array. Furthermore, a compressed sensing archiving mechanism is used to retain the top 10% of the top fitness trajectories for subsequent elite retention and convergence analysis. The entire chain from parameter initialization to persistence consists of a five-level structure at the data level: "configuration dictionary → serialized file → memory vector → GPU tensor → historical trajectory tensor," ensuring consistent repeatability and verifiability of the algorithm across multiple experimental scenarios.

[0106] Of particular importance, step S32 includes:

[0107] Construct a fitness evaluation function based on the ISSA initialization parameter set;

[0108] Randomly generate the initial ISSA position, and update the position in the discovery-follow-guard stage based on the ISSA initialization parameter set and fitness evaluation function to obtain the discovery-follow-guard updated position;

[0109] Based on the discovery-follow-watcher update position, the optimal parameter combination for the mining area is obtained by iterative optimization with the top 10% retained.

[0110] In this embodiment of the invention, based on the ISSA initialization parameter set constructed in step S31, a fitness evaluation function F(x) is defined. This function is used to measure a given combination of parameters. The performance in the task of classifying rocks and minerals in mining areas is specifically defined as the weighted reciprocal of the classification accuracy obtained based on five-fold cross-validation, and is encapsulated as a vectorizable function interface in actual execution. Then, the uniform distribution sampling operation is invoked. Generate an initial position matrix for P = 50 sparrow individuals between the upper and lower bounds of the parameter search space in each dimension. in Let represent the lower and upper bound vectors of the search space, respectively. Then, based on the preset proportions in the initialization parameter set, all individuals are divided into three groups: discoverers (20%), followers (70%), and vigilants (10%), constructing a role mapping tensor R∈{0,1,2}. 50 Based on this, different behavior update formulas are invoked to perform individual position updates. During the discoverer update phase, the previous... For each individual, its foraging behavior is modeled based on its fitness function value. An exponential weighted contraction mechanism is used to adjust its position, and the update formula is as follows:

[0111]

[0112] Where α is the empirical adjustment coefficient. Follower position updates are based on the current globally optimal individual position x. best The system implements a gravitational convergence mechanism and incorporates local search terms based on random normal perturbations; the vigilant update utilizes fitness variance to construct cooperative escape behavior among individuals, which is relevant to the current situation. A discrete jump perturbation term is added to simulate the alert response. After each update, the newly generated individual position matrix X is uniformly applied. t+1) Boundary processing is performed, and based on the function f(X( t+1) Reassess fitness and record the globally optimal individual and its corresponding parameter combination. Complete all T... max =After 100 iterations, the top 10% of the individuals with the best fitness throughout all iterations (a total of 5 individuals) are retained to form an elite set ε, and the parameter vector x corresponding to the globally optimal individual is extracted by non-dominated sorting. opt This is the final "optimal parameter combination for the mining area." This vector serves as the input for the kernel function parameters, penalty coefficients, and fault tolerance terms in the subsequent SVM model construction, providing a structural tuning foundation for the classification task. At the data level, this process is represented by a closed-loop data flow: "position tensor → fitness tensor → role mask → iterative trajectory → elite set → optimal parameter vector," ensuring full traceability of information utilization and convergence within the search space.

[0113] Preferably, step S33 includes the following steps:

[0114] Step S331: Set the kernel function based on the optimal parameter combination of the mining area and define the classification decision function to obtain the initial SVM algorithm mining area classification model structure;

[0115] Step S332: Input the mining area feature matrix data into the initial SVM algorithm mining area classification model structure for model feature training, and perform convex optimization processing to obtain the preliminary mining area classification model;

[0116] Step S333: Perform sequence minimization optimization on the preliminary model of mining area classification and construct the model vector of classification bias to obtain the benchmark model of rock and mineral classification.

[0117] In this embodiment of the invention, the optimal parameter combination of the mining area output in step S3 is used as the kernel function parameter input of the support vector machine (SVM) model. The kernel function type (such as radial basis function RBF) and its corresponding parameter γ are set, and the regularization parameter C is used to control the soft margin penalty of the model to construct the initial classification decision function. Where x represents the input feature vector, K(·,·) is the kernel function, and α i Let be the Lagrange multiplier, and b be the classification bias term. After defining the SVM model structure in step S331, step S332 will process the feature matrix data of the mining area. and its corresponding label y∈{+1,-1} N The input is given to the initial model, and convex optimization techniques are used to solve the quadratic programming problem of the standard SVM. The optimization objective is to minimize... Where w is the decision boundary normal vector, ξ iLet these be slack variables, and the constraints are as follows: φ(·) represents the implicit mapping function. Through this optimization process, the model parameters {α} i The bias b is adjusted to form the preliminary model structure for mineral area classification. Step S333 uses the Sequential Minimal Optimization (SMO) algorithm to iteratively solve the quadratic programming problem, gradually updating the Lagrange multipliers to ensure that the KKT conditions are met. Simultaneously, the classification bias is dynamically adjusted during the optimization process to improve the model's discriminative performance. This stage completes the vectorized expression of the model by constructing a set of support vectors and related weights, ultimately obtaining the benchmark model for rock and mineral classification. This model can achieve efficient category discrimination based on input features.

[0118] Preferably, step S4 includes the following steps:

[0119] Step S41: Obtain the rock and ore test dataset;

[0120] Step S42: Input the rock and ore test dataset into the rock and ore classification benchmark model to predict the category and obtain the rock and ore test set prediction data; compare and analyze the rock and ore test set prediction data with the rock and ore test dataset to obtain the rock and ore test set classification accuracy data.

[0121] Step S43: Extract misclassified samples based on the classification accuracy data of the rock and mineral test set, and perform spectral feature backtracking to obtain a lithological confusion analysis report;

[0122] Step S44: Construct a full-process ore classification report based on the lithological confusion analysis report and the classification accuracy data of the rock and ore test set.

[0123] In this embodiment of the invention, an independent test dataset is obtained from a pre-collected or segmented mining area rock and ore dataset. This dataset contains spectral feature vectors that do not overlap with the training set and their corresponding true class labels, ensuring the fairness and objectivity of the model evaluation. Step S42 inputs this rock and ore test dataset into the previously constructed rock and ore classification benchmark model. The model calculates a classification decision function for each input feature vector and outputs the corresponding predicted class label, forming the ore test set prediction data. Subsequently, the predicted labels are compared with the true labels in the test dataset sample by sample, and performance indicators such as classification accuracy and confusion matrix are calculated to quantify the classification accuracy of the ore test set and reflect the model's generalization ability on unseen samples. Step S43, based on the classification accuracy results, extracts misclassified samples from the ore test set prediction data, i.e., spectral data records where the predicted class is inconsistent with the true class. For these samples, the corresponding spectral feature data is traced back to analyze the similarity and differences between spectral features. Combined with the confusion relationship between categories, a lithological confusion analysis report is constructed to reveal the potential reasons for the model's blurred classification boundaries or overlapping sample features. Step S44 integrates the entire ore classification assessment process based on the lithological confusion analysis report and classification accuracy data, constructing a systematic ore classification report. This report includes model performance indicators, misclassified sample analysis, confusion relationships between categories, and potential optimization directions, providing data support and decision-making basis for subsequent model tuning and application. The entire process is data-driven, closely combining spectral characteristics with classification results to ensure the scientific rigor and soundness of the evaluation system.

[0124] Preferably, step S43 includes the following steps:

[0125] Step S431: Obtain mineral composition analysis data; extract misclassified samples based on the classification accuracy data of the ore test set to obtain misclassified sample data of the classification model;

[0126] Step S432: Extract the original spectrum in the 6000-14500nm band from the misclassified sample data of the classification model to obtain the original spectral data of the misclassified samples;

[0127] Step S433: Calculate the spectral cosine similarity of the original spectral data of the misclassified sample, and determine the correlation between spectral similarity and rock-forming mineral content by combining the mineral composition analysis data, and obtain the lithological confusion analysis report.

[0128] In this embodiment of the invention, mineral composition analysis data is collected, which includes information on the main mineral components and their contents in rock and ore samples, serving as an auxiliary basis for verification and analysis. Simultaneously, combined with the classification accuracy data of the ore test set, the system extracts all misclassified samples by comparing the model's predicted labels with the true labels, constructing a misclassified sample dataset for the classification model, providing targeted objects for subsequent spectral feature analysis. Step S432: For the misclassified sample data, a spectral range of 6000 to 145000 nanometers is extracted from its original spectral dataset. This band typically contains important absorption features related to mineral chemical composition, ensuring that the selected data has the potential to distinguish different mineral components, forming the original spectral data of the misclassified samples. Step S433: Spectral cosine similarity calculation is performed on the original spectral data of the misclassified samples. The cosine similarity index quantifies the similarity of spectral shapes between different samples, reflecting the correlation of spectral curves through the size of the vector angle, thereby quantitatively comparing the similarity between misclassified samples and their spectra with typical minerals. This reveals the physicochemical roots of spectral confusion in misclassified samples, summarizes the mineral component characteristics leading to classification confusion, and finally generates a lithological confusion analysis report. This reporting system integrates spectral similarity and mineral composition information, and uses a data-driven approach to provide quantitative analysis and explanation of classification bias in misclassified samples, supporting subsequent model improvement and sample selection.

[0129] Preferably, step S44 includes the following steps:

[0130] Step S441: Extract the LDA inter-class weight coefficients from the lithological confusion analysis report and back to the rock and mineral classification benchmark model for model optimization to obtain the optimized rock and mineral classification benchmark model;

[0131] Step S442: Obtain real-time borehole core test data; perform borehole lithology category analysis on the real-time borehole core test data based on the rock and ore classification benchmark optimization model to obtain rock and ore identification results;

[0132] Step S443: Perform thin section analysis on the rock and ore identification results, and supplement the application accuracy report by combining the rock and ore test set classification accuracy data to obtain the full-process ore classification report.

[0133] In this embodiment of the invention, inter-class weight coefficients generated by linear discriminant analysis (LDA) are extracted from the lithology confusion analysis report. These coefficients quantify the discriminative power of the feature space between different rock and mineral categories and reflect the contribution of each feature to category discrimination. This weight coefficient data is fed back to the rock and mineral classification benchmark model to guide the adjustment and optimization of model parameters, thereby enhancing the model's ability to distinguish between different rock and mineral categories and generating an optimized rock and mineral classification benchmark model. In step S442, the system collects thermal infrared scanning data from borehole cores in real time, forming a high-dimensional spectral data stream, which serves as the input to the optimized rock and mineral classification benchmark model. Based on the optimized classification model, a multi-feature fusion and decision-making mechanism is used to perform point-by-point lithology identification and profile analysis on the real-time scanning data, outputting borehole lithology category results to achieve fine-grained identification of rock and mineral types in geological strata. In step S443, optical microscopic analysis is performed on the identified rock and mineral samples using thin-section microscopy to obtain mineral composition and structural information, which serves as ground real-valued label data. By combining the classification accuracy data from the ore test set, the system compares and verifies the identification results with thin section identification, and supplements and improves the application accuracy report accordingly. Finally, by integrating the classification model performance indicators, identification results, and microscopic identification information, a complete ore classification report covering data acquisition, model identification, and manual verification is constructed, providing systematic data support for subsequent geological analysis and engineering applications.

[0134] This specification provides a mineral rock and ore classification system based on thermal infrared hyperspectral remote sensing technology, used to perform the aforementioned mineral rock and ore classification method based on thermal infrared hyperspectral remote sensing technology. This mineral rock and ore classification system based on thermal infrared hyperspectral remote sensing technology includes:

[0135] The contact testing and data standardization module is used to conduct contact tests on rocks and ores in the mining area, perform data preprocessing, and construct a standardized thermal infrared spectroscopy dataset.

[0136] The spectral feature dimensionality reduction and feature matrix construction module is used to perform principal component analysis on the standardized thermal infrared spectral dataset, and select the number of principal components in the mining area using a preset rock and mineral variance threshold to obtain the principal component screening result data of the mining area; the principal component screening result data of the mining area is randomly segmented and the rock and mineral structural feature matrix is ​​constructed to obtain the feature matrix data of the mining area.

[0137] The ISSA intelligent classification modeling module is used to construct the ISSA initial parameter set, perform ISSA initial parameter set optimization iteration, and output the optimal parameter combination for the mining area; based on the optimal parameter combination for the mining area, a classification model is constructed for the feature matrix data of the mining area to obtain the rock and ore classification benchmark model.

[0138] The model evaluation and classification feedback module is used to acquire rock and ore test datasets; input the rock and ore test datasets into the rock and ore classification benchmark model to predict the categories, and backtrack the spectral characteristics of misclassified samples to obtain a lithological confusion analysis report; and construct a full-process rock and ore classification report based on the lithological confusion analysis report.

[0139] The beneficial effects of this invention lie in its multi-dimensional efficiency enhancement at the data level through the construction of a multi-level data flow closed loop consisting of "contact testing - spectral preprocessing - principal component analysis - improved sparrow search - SVM model training - misclassification backtracking": First, contact thermal infrared scanning, supplemented by systematic preprocessing, significantly reduces the interference of environmental radiation and instrument background on the reflectivity matrix, compressing the original spectral noise variance to approximately 30% of its initial value, providing a high signal-to-noise ratio input for subsequent analysis; Second, principal component analysis based on covariance spectral decomposition employs a cumulative variance threshold to rigorously screen low-correlation redundant bands, compressing the sample feature dimension from the M-level to the k-level while maintaining over 98% information content, thus reducing data storage burden and improving matrix operation efficiency; Subsequently, the modified... The Sparrow Search algorithm, through the synergistic effect of Gaussian perturbation and elite retention mechanism, performs global optimization of the classifier's hyperparameter space. Compared with grid search, it can achieve a lower cross-validation error within about 40% of the iterations. Moreover, the output optimal parameter vector has been verified to improve the overall accuracy of the rock and mineral classification benchmark model by 8-12 percentage points. Finally, through spectral backtracking of misclassified samples and lithological confusion analysis, the system traces the remaining classification errors from both band and mineral content perspectives during the testing phase, forming feedback factors to feed back into the model, thereby continuously compressing the inter-class confusion rate. The entire process achieves a chain gain of quality improvement, dimensionality compression, parameter optimization, and accuracy iteration in the data closed loop, significantly improving the reliability and interpretability of the spectral-lithological mapping.

[0140] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0141] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for classifying rocks and ores in mining areas based on thermal infrared hyperspectral remote sensing technology, characterized in that, Includes the following steps: Step S1: Conduct contact tests on the rocks and ores in the mining area, perform data preprocessing, and construct a standardized thermal infrared spectroscopy dataset; Step S2: Perform principal component analysis on the standardized thermal infrared spectroscopy dataset, and select the number of principal components in the mining area using a preset rock and mineral variance threshold to obtain the principal component screening result data of the mining area; randomly partition the principal component screening result data of the mining area, and construct the rock and mineral structure feature matrix to obtain the feature matrix data of the mining area. Step S3: Construct the ISSA initial parameter set and perform ISSA initial parameter set optimization iteration to output the optimal parameter combination for the mining area; based on the optimal parameter combination for the mining area, construct a classification model for the feature matrix data of the mining area to obtain the rock and ore classification benchmark model. Step S3 includes the following steps: Step S31: Construct the ISSA initialization parameter set; Step S32: Randomly generate the initial position of the ISSA based on the ISSA initialization parameter set, and update the position in the discovery-follow-watcher stage. Retain the top 10% for optimization iteration to obtain the optimal parameter combination for the mining area. Step S33: Based on the optimal parameter combination of the mining area, construct a classification model for the feature matrix data of the mining area to obtain a benchmark model for rock and ore classification. Step S33 includes the following steps: Step S331: Set the kernel function based on the optimal parameter combination of the mining area and define the classification decision function to obtain the initial SVM algorithm mining area classification model structure; Step S332: Input the mining area feature matrix data into the initial SVM algorithm mining area classification model structure for model feature training, and perform convex optimization processing to obtain the preliminary mining area classification model; Step S333: Perform sequence minimization optimization on the preliminary model for mineral area classification, and construct the model vector for classification bias to obtain the benchmark model for rock and mineral classification; Step S4: Obtain the rock and ore test dataset; input the rock and ore test dataset into the rock and ore classification benchmark model to predict the category, and backtrack the spectral features of misclassified samples to obtain a lithological confusion analysis report; construct a full-process ore classification report based on the lithological confusion analysis report.

2. The method for classifying rocks and ores in mining areas based on thermal infrared hyperspectral remote sensing technology according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Use a Fourier transform infrared spectrometer to conduct contact tests on the rocks and ores in the mining area to generate a raw thermal infrared reflectance dataset. The Fourier transform infrared spectrometer is set with the following parameters: spectral range of 6000-14500nm, wavenumber of 4000-650cm⁻¹, resolution of less than 4cm⁻¹, and single-point integration time of 5s. Step S12: Perform spectral reconstruction and baseline correction on the original thermal infrared reflectance dataset, and eliminate instrument noise and environmental radiation interference to obtain reconstructed spectral data; Step S13: Perform Min-Max normalization on the reconstructed spectral data and eliminate spectral intensity differences to obtain normalized spectral data; Step S14: Obtain rock and mineral label data; match the rock and mineral label data with the normalized spectral data to obtain a standardized thermal infrared spectral dataset.

3. The method for classifying rocks and ores in mining areas based on thermal infrared hyperspectral remote sensing technology according to claim 1, characterized in that, Step S2 involves performing principal component analysis on the standardized thermal infrared spectroscopy dataset and selecting the number of principal components for the mining area using a preset rock and mineral variance threshold. The covariance matrix of the standardized thermal infrared spectral dataset with different components and different structural constructions is calculated to obtain the spectral covariance matrix. The spectral covariance matrix is ​​decomposed into eigenvalues ​​based on characteristic minerals, characteristic structures, and characteristic features to obtain spectral covariance eigenvectors. Based on the spectral covariance eigenvectors, the eigenvalues ​​of typical minerals are sorted in descending order, and the cumulative variance contribution rate of rocks and minerals is calculated to obtain a spectral covariance distribution table. Cumulative variance is calculated based on the spectral covariance distribution table, and the optimal principal component data is determined using a preset rock and mineral variance threshold. The optimal principal component data is constructed by projecting the feature vectors of the mining area, and then the features are screened to obtain the principal component screening results data of the mining area.

4. The method for classifying rocks and ores in mining areas based on thermal infrared hyperspectral remote sensing technology according to claim 1, characterized in that, Step S2 involves randomly partitioning the principal component screening results data of the mining area and constructing the rock and ore composition feature matrix, including: The principal component screening results data of the mining area are preprocessed using PCA and dimensionality reduced using SG filtering to obtain preprocessed and dimensionality-reduced data; inter-class scatter matrix is ​​constructed on the preprocessed and dimensionality-reduced data of the mining area; intra-class matrix is ​​constructed on the preprocessed and dimensionality-reduced data of the mining area. Add rock and mineral category labels to the inter-class scatter matrix and intra-class matrix of rock and mineral label data to obtain the mining area label feature matrix; The feature matrix of the mining area is randomly divided into 3:7 segments, and the mean of the rock and mineral fabric feature dimensions is balanced to obtain the feature matrix data of the mining area.

5. The method for classifying rocks and ores in mining areas based on thermal infrared hyperspectral remote sensing technology according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Obtain the rock and ore test dataset; Step S42: Input the rock and ore test dataset into the rock and ore classification benchmark model to predict the category and obtain the rock and ore test set prediction data; compare and analyze the rock and ore test set prediction data with the rock and ore test dataset to obtain the rock and ore test set classification accuracy data. Step S43: Extract misclassified samples based on the classification accuracy data of the rock and mineral test set, and perform spectral feature backtracking to obtain a lithological confusion analysis report; Step S44: Construct a full-process ore classification report based on the lithological confusion analysis report and the classification accuracy data of the rock and ore test set.

6. The method for classifying rocks and ores in mining areas based on thermal infrared hyperspectral remote sensing technology according to claim 5, characterized in that, Step S43 includes the following steps: Step S431: Obtain mineral composition analysis data; extract misclassified samples based on the classification accuracy data of the ore test set to obtain misclassified sample data of the classification model; Step S432: Extract the original spectrum of the 6000-14500nm band from the misclassified sample data of the classification model to obtain the original spectral data of the misclassified samples; Step S433: Calculate the spectral cosine similarity of the original spectral data of the misclassified sample, and determine the correlation between spectral similarity and rock-forming mineral content by combining the mineral composition analysis data, and obtain the lithological confusion analysis report.

7. The method for classifying rocks and ores in mining areas based on thermal infrared hyperspectral remote sensing technology according to claim 5, characterized in that, Step S44 includes the following steps: Step S441: Extract the LDA inter-class weight coefficients from the lithological confusion analysis report and back to the rock and mineral classification benchmark model for model optimization to obtain the optimized rock and mineral classification benchmark model; Step S442: Obtain real-time borehole core test data; perform borehole lithology category analysis on the real-time borehole core test data based on the rock and ore classification benchmark optimization model to obtain rock and ore identification results; Step S443: Perform thin section analysis on the rock and ore identification results, and supplement the application accuracy report by combining the rock and ore test set classification accuracy data to obtain the full-process ore classification report.

8. A mineral classification system for mining areas based on thermal infrared hyperspectral remote sensing technology, characterized in that, For implementing the mining area rock and ore classification method based on thermal infrared hyperspectral remote sensing technology as described in claim 1, the mining area rock and ore classification system based on thermal infrared hyperspectral remote sensing technology includes: The contact testing and data standardization module is used to conduct contact tests on rocks and ores in the mining area, perform data preprocessing, and construct a standardized thermal infrared spectroscopy dataset. The spectral feature dimensionality reduction and feature matrix construction module is used to perform principal component analysis on the standardized thermal infrared spectral dataset, and select the number of principal components in the mining area using a preset rock and mineral variance threshold to obtain the principal component screening result data of the mining area; the principal component screening result data of the mining area is randomly segmented and the rock and mineral structural feature matrix is ​​constructed to obtain the feature matrix data of the mining area. The ISSA intelligent classification modeling module is used to construct the ISSA initial parameter set, perform ISSA initial parameter set optimization iteration, and output the optimal parameter combination for the mining area; based on the optimal parameter combination for the mining area, a classification model is constructed for the feature matrix data of the mining area to obtain the rock and ore classification benchmark model. The model evaluation and classification feedback module is used to acquire rock and ore test datasets; input the rock and ore test datasets into the rock and ore classification benchmark model to predict the categories, and backtrack the spectral characteristics of misclassified samples to obtain a lithological confusion analysis report; and construct a full-process rock and ore classification report based on the lithological confusion analysis report.

Images