Hyperspectral mineralization alteration information extraction method and system based on semantic segmentation

Abstract

The invention discloses a hyperspectral mineralization alteration information extraction method and system based on semantic segmentation, and relates to the technical field of mineralization alteration information extraction. The method comprises the following steps: investigating a research area according to geological data, and analyzing geological information, related to mineralization, of the research area; the method comprises the following steps: preprocessing hyperspectral data of a mining area of a research area by utilizing ENVI software, and making a data set suitable for a semantic segmentation model; constructing a semantic segmentation model, and extracting alteration mineral information of the research area; remote sensing geological interpretation is carried out on the research area, and linear structure factors beneficial to mineralization are extracted in combination with principal component analysis; and in combination with remote sensing altered mineral information and linear construction factors, constructing an evidence weight model, and delineating a metallogenic prediction prospective area. The deep learning model is utilized to learn the spectral continuity and the spatial texture mode of the hyperspectral data at the same time, the defects that traditional spectral matching is sensitive to mixed pixels and poor in local noise robustness are effectively overcome, and technical guarantee is provided for efficient development and utilization of mineral resources.

Description

Technical Field

[0001] This invention relates to the field of mineralization alteration information extraction technology, and in particular to a method and system for extracting hyperspectral mineralization alteration information based on semantic segmentation. Background Technology

[0002] Mineral resources are a crucial material foundation for national economic construction and sustainable social development. A key factor restricting resource security is the shortage of reserve mineral resources. The rapid development of spatial information technology has provided a new technological foundation for mineral resource surveys and evaluations. In particular, the maturity of technologies such as satellite remote sensing, the Global Positioning System (GPS), and satellite communication has made it possible to acquire high spatiotemporal resolution geographic data. Hyperspectral remote sensing, as a new generation of Earth observation technology, can provide continuous narrow-band hyperspectral reflectance information, giving it a significant advantage in the precise identification and quantitative inversion of surface materials. In traditional methods, hyperspectral data undergoes structured reconstruction of features using linear dimensionality reduction methods such as PCA, combined with endmember extraction and spectral unmixing techniques. However, this approach suffers from recognition degradation when dealing with "different objects with the same spectrum" and mixed pixels. In recent years, the development of artificial intelligence technology has promoted the improvement of spatiotemporal joint modeling, multi-scale feature extraction, and adaptability to complex scenes. Therefore, for those skilled in the art, how to utilize deep learning technology to process hyperspectral data and extract mineralization and alteration information is an urgent problem to be solved. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for extracting hyperspectral mineralization and alteration information based on semantic segmentation, in order to solve the problems mentioned in the background art. The extraction of mineralization and alteration information based on deep learning methods can promote the effective application of hyperspectral technology in mineral exploration.

[0004] To achieve the above objectives, the present invention provides the following solution: On one hand, a method for extracting hyperspectral mineralization alteration information based on semantic segmentation, the specific steps of which include the following: The study area was explored based on geological data, and geological information related to mineralization in the study area was analyzed. Based on the geological information, hyperspectral data of the metallogenic belt in the study area were obtained; The hyperspectral data is preprocessed to obtain preprocessed hyperspectral data; The hyperspectral data is preprocessed to obtain preprocessed hyperspectral data, which is then used to create a dataset suitable for semantic segmentation. A2S2K-ResNet semantic segmentation model was constructed, and mineral alteration information of the study area was extracted using the A2S2K-ResNet semantic segmentation model. Remote sensing geological interpretation was performed on the study area, and linear tectonic factors conducive to mineralization were extracted using PCI software in conjunction with principal component analysis. Based on the extracted mineral alteration information, combined with the linear tectonic factors and mineral data, an evidence weight model was constructed to delineate the prospective mineralization area.

[0005] Preferably, the geological information related to mineralization includes favorable strata, main structural features and ore-controlling structural features, spatial distribution features of mineralization points, and regularity of mineralization and rich mineralization conditions.

[0006] Preferably, the preprocessing operations for the hyperspectral data include band removal, band combination, radiometric calibration, atmospheric correction and orthorectification, noise filtering and resampling.

[0007] Preferably, the step of radiometric calibration of the hyperspectral data is as follows: Combine visible light, near-infrared, and short-wave infrared data files into bands; Read in the wavelength range, center wavelength, and information gain value for each band according to the given satellite calibration file; The calibration parameters were automatically extracted using the radiometric calibration module of ENVI 5.6 software to complete the radiometric calibration of the hyperspectral data.

[0008] Preferably, the expression for the radiation calibration is: ; in, express Radiance values ​​for the band, express Gain value of the band, DN Represents a numerical quantization value. express The offset value of the band.

[0009] Preferably, the A2S2K-ResNet semantic segmentation model is a residual network that integrates an attention mechanism and an adaptive spectral-spatial kernel, supporting simultaneous modeling in both spatial and spectral dimensions. It extracts altered minerals by dynamically weighting the contributions of different branches through the attention mechanism.

[0010] Preferably, the A2S2K-ResNet semantic segmentation model constructs an S×S×B spectral-spatial input through neighborhood block extraction, and then uses a multi-scale residual learning structure to achieve feature representation; each residual block is configured with convolutional kernels of different scales to synchronously encode local texture, regional structure and spectral features, and achieves effective transfer between feature layers through residual connections; after pooling and fully connected mapping, the model outputs pixel-by-pixel alteration mineral grading results at the Softmax layer.

[0011] Preferably, the specific steps for constructing an evidence weighting model and delineating prospective mineralization areas are as follows: Based on the metallogenic model and metallogenic regularity of the study area, linear tectonic factors and remote sensing alteration anomalies were selected as evidence layers to establish metallogenic geological thematic layers. Establish grid cells based on the size of the study area; The prior probabilities were calculated by combining known mineral deposits and grid cells within the study area. By utilizing prior probabilities and mineralization-related evidence factors, the weight value of each factor is calculated, independence tests are performed on all factors, and a layer of evidence factors favorable to mineralization is selected. For the selected evidence factor layers, the posterior probability is calculated to obtain the posterior probability map, and the prospective mineralization area is delineated.

[0012] On the other hand, a hyperspectral mineralization and alteration information extraction system based on semantic segmentation is provided, including a geological pre-analysis module, a hyperspectral data acquisition module, a preprocessing module, an alteration information extraction module, and a mineralization prediction module; wherein... The geological pre-analysis module is used to explore the study area based on geological data and analyze the geological information related to mineralization in the study area. The hyperspectral data acquisition module is used to acquire hyperspectral data of the metallogenic belt in the study area based on the geological information. The preprocessing module is used to perform preprocessing operations on the hyperspectral data to obtain preprocessed hyperspectral data and create a dataset suitable for semantic segmentation. The alteration information extraction module is used to construct an A2S2K-ResNet semantic segmentation model and extract mineral alteration information of the study area through the A2S2K-ResNet semantic segmentation model. The mineralization prediction module is used to perform remote sensing geological interpretation of the study area, extract linear tectonic factors conducive to mineralization in PCI software by combining principal component analysis, and construct an evidence weight model based on the extracted mineral alteration information, combined with the linear tectonic factors and mineral data, to delineate the prospective mineralization area.

[0013] According to the specific embodiments provided by the present invention, the beneficial effects of the present invention are: (1) Based on a thorough study of existing geological and mineral data in the area, we will conduct research on the technical methods for extracting mineral alteration information based on semantic segmentation models. At the same time, we will combine the geological and mineralization conditions of the study area and the distribution of known mineral deposits to establish a mineralization prediction model and delineate the prospective mineralization areas of the study area, providing an important basis for mineral resource exploration.

[0014] (2) The semantic segmentation network is introduced into the hyperspectral alteration information extraction process. The deep learning model is used to learn the spectral continuity and spatial texture pattern of hyperspectral data at the same time, which effectively overcomes the shortcomings of traditional spectral matching that is sensitive to mixed pixels and lacks robustness to local noise.

[0015] (3) The high-precision alteration information extracted by the semantic segmentation model is used as the quantitative input layer of the evidence weight model. An integrated mineralization prediction framework of "hyperspectral deep learning semantic segmentation - remote sensing linear structural analysis - evidence weight prediction" is constructed, which can be applied to rapid mineral exploration prediction in medium and large scale areas. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a diagram of the A2S2K-ResNet semantic segmentation model framework of the present invention; Figure 3 This is a structural diagram of the adaptive spectral-spatial kernel attention module of the present invention; Figure 4 This is a flowchart of the mineralization prediction process using the evidence weighting model of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The purpose of this invention is to provide a method for extracting hyperspectral mineralization and alteration information based on semantic segmentation, such as... Figure 1 As shown, the specific steps include the following: S1. Conduct exploration of the study area based on geological data and analyze geological information related to mineralization in the study area; S2. Based on geological information, obtain hyperspectral data of the metallogenic belt in the study area; S3. Perform preprocessing operations on the hyperspectral data to obtain preprocessed hyperspectral data, and create a dataset suitable for semantic segmentation; S4. Construct an A2S2K-ResNet semantic segmentation model and extract mineral alteration information of the study area using the A2S2K-ResNet semantic segmentation model; S5. Remote sensing geological interpretation of the study area is carried out, and linear tectonic factors conducive to mineralization are extracted in PCI software by combining principal component analysis. Based on the extracted mineral alteration information, combined with linear tectonic factors and mineral data, an evidence weight model is constructed to delineate the prospective mineralization area.

[0020] In this embodiment, a lead-zinc mining area is taken as the study area, and ZY1-02D ​​hyperspectral remote sensing image is used as the main data source. Based on a thorough study of the existing geological and mineral data in the area, a technical method for extracting mineral alteration information based on a semantic segmentation model is developed. At the same time, combined with the geological and metallogenic conditions of the study area and the distribution of known mineral occurrences (deposits), a metallogenic prediction model is established to delineate the metallogenic prospect area of ​​the study area, providing an important basis for mineral resource exploration.

[0021] In S1, preliminary research is conducted by combining data: based on previous research results and existing geological data, the geological background, geochemical exploration, and structural information of the study area are summarized, known mineralization points are explored, and the distribution characteristics of favorable strata, main structures and ore-controlling structures related to mineralization in lead-zinc mining areas, the spatial distribution characteristics of mineralization points, and the regularity of mineralization and rich ore conditions are analyzed.

[0022] In S2, hyperspectral data is acquired. In this embodiment, ZY1-02D ​​data is used. ZY1-02D ​​satellite data can effectively acquire information on Earth's surface features and provide reliable data support for related research on mineralization and alteration information extraction technology. ZY1-02D ​​hyperspectral data contains 166 bands, with a spectral range of 400–2500 nm. The spectral resolution is 10 nm in the visible light and 20 nm in the shortwave infrared, and the spatial resolution is 30 m.

[0023] In S3, the acquired data undergoes preprocessing. The raw L1A-level hyperspectral data from the ZY-1 02D satellite contains DN values ​​with RPC correction parameters, requiring preprocessing before it can be used for alteration mineral mapping. Data processing is a prerequisite for subsequent alteration information and tectonic extraction. The identification of ground features by imaging spectra depends on the detailed spectral characteristics of the features. To obtain the true spectral characteristics of ground features from the images, relevant data preprocessing and enhancement are necessary. The preprocessing workflow mainly includes radiometric calibration, atmospheric correction, band removal, and orthorectification.

[0024] The specific procedures for each preprocessing operation will be explained below: Radiometric calibration: The DN (Digital Number) value refers to the pixel brightness value in remote sensing image data. It is the main carrier of the grayscale values ​​of ground features, is dimensionless and an integer, and its value can be used to represent the ability of ground features to reflect electromagnetic waves, i.e., the ability to distinguish ground features. Radiometric calibration refers to the specific processing of converting the DN values ​​received by the imaging spectrometer on the satellite into the apparent reflectance of the outer atmosphere. Its purpose is to eliminate errors caused by the sensor's own defects, thereby obtaining the accurate radiance value at the sensor's inlet location. The mathematical expression of radiometric calibration is shown below: ; in, express Radiance values ​​for the band, express Gain value of the band, DN Represents a numerical quantization value. express The offset value of the band.

[0025] Using ENVI 5.6 software, the visible near-infrared (VN) and shortwave infrared (SW) data files were first combined into bands. Then, the wavelength range, center wavelength, and information gain value of each band were read in according to the given satellite calibration file. The calibration parameters were automatically extracted using the software's radiometric calibration module to complete the radiometric calibration of ZY1-02D.

[0026] Atmospheric correction: Radiation signals are inevitably affected by the atmosphere during transmission, thus impacting the true spectral characteristics of remote sensing image data. Atmospheric correction essentially refers to the process of converting the apparent reflectance of the outer atmosphere into the true reflectance of the ground surface. Its basic goal is to eliminate atmospheric effects in the DN values ​​measured by spaceborne sensors, avoid distortion errors caused by the atmosphere, and improve image data quality. After atmospheric correction, the pixel spectral curve is transformed into the true spectral curve of the ground object, and the spectral characteristics are more pronounced.

[0027] The FLAASH model is currently a highly accurate atmospheric correction algorithm, widely used in multispectral and hyperspectral data preprocessing. This embodiment utilizes the FLAASH module of ENVI software to perform atmospheric correction processing on radiometrically corrected data.

[0028] Band Removal: Within the visible-near-infrared band, water vapor within a certain wavelength range can affect the quality of the acquired ZY1-02D ​​satellite hyperspectral data. These bands, containing excessive noise and insufficient ground feature information, need to be removed; these bands are located in the wavelength ranges of 1300–1400 nm and 1800–1900 nm. Simultaneously, bands containing relatively high levels of noise, specifically those before 400 nm and after 2450 nm, should also be removed. By editing the source image data, identifying the bad bands affected by water vapor absorption and noise, and then using the Resize Data module in ENVI software for removal, the process is completed.

[0029] Orthorectification: Orthorectification can simultaneously correct tilt and projection difference in an image. In this embodiment, Landsat8, covering the study area, is selected as the reference image. Combined with the DEM data of the study area, orthorectification is performed using the RPCorthorectification module in ENVI. The maximum number of GCPS to be acquired is set to 100, and the more accurate Cubic Convolution method is used for resampling.

[0030] Furthermore, while hyperspectral images provide rich spectral data, their relatively low spatial resolution often makes it difficult to identify subtle changes in minerals. Their complex spatial-spectral characteristics limit the effectiveness of traditional methods in feature extraction and multi-scale fusion, impacting recognition accuracy. To address this, this invention designs an A2S2K-ResNet semantic segmentation model for hyperspectral alteration mineral extraction, which improves the accuracy of alteration information extraction from hyperspectral data. The A2S2K-ResNet semantic segmentation model can effectively extract important feature information from geological images, providing strong support for the prediction and analysis of mining areas.

[0031] Furthermore, deep learning tasks require a suitable dataset to extract information from the data and abstract high-level information based on the extracted features, mapping it to the output. Combining the spectral feature analysis described above with existing geological data, supplemented by spectral measurements from field rock and mineral sampling, this embodiment selects representative pixels from the image to create an image label map, and then generates a sample dataset in a specific format based on the network model.

[0032] Impact on Label Data Creation: Data is an indispensable part of any semantic segmentation or deep learning algorithm, and the quality of the dataset directly affects the model's performance. This embodiment uses ZY-1 hyperspectral remote sensing imagery, selects a study area, and, based on a thorough analysis of existing geological data, identifies barite, dolomite, quartz, and anhydrite as typical alteration minerals in the area. Using alteration information extracted jointly by the Hybrid Tuned Matched Filtering (MTMF) and Spectral Angle Matching (SAM) methods as a basis, representative pixels are selected from the imagery to create label maps for the four main alteration minerals, thus conducting research on deep learning-based alteration mineral identification technology.

[0033] Hyperspectral image mineral identification mapping is similar to land cover classification. Creating image tags involves generating a single-band image with the same spatial size as the image, where different pixel values ​​represent different categories. In the tag pixel settings of this invention, pixel value 0 represents "background", 1 represents barite, 2 represents dolomite, 3 represents quartz, and 4 represents anhydrite.

[0034] This invention employs a sliding window cropping method to create a semantic segmentation dataset. Hyperspectral images and the label matrix are simultaneously cropped using a 128×128 pixel patch size and a 64-pixel sliding step. Before cropping, edge completion processing ensures the image size meets the step size divisibility requirement. Simultaneously, filtering conditions are set to retain valid patches containing sparse categories. To alleviate class imbalance, data augmentation is performed on patches containing target pixels, including geometric and spectral augmentation operations such as random rotation, vertical flipping, Gaussian noise addition, and random amplitude scaling. Finally, a hierarchical random partitioning strategy is adopted: 80% of the dataset is allocated to the training set for model training and updating model parameters; 10% is allocated to the test set, not used for training, but for evaluating model performance and preventing underfitting or overfitting parameters.

[0035] The A2S2K-ResNet semantic segmentation model is a residual network that integrates an attention mechanism and an adaptive spectral-spatial kernel, specifically designed for hyperspectral image (HSI) classification. It supports simultaneous modeling in both spatial and spectral dimensions and dynamically weights the contributions of different branches through an attention mechanism, thereby achieving superior feature extraction.

[0036] The framework of the A2S2K-ResNet network is as follows: Figure 2As shown, the model constructs an S×S×B spectral-spatial input through neighborhood block extraction, and then performs feature representation in a multi-scale residual learning structure composed of start, middle, and end residual blocks. Different scale convolutional kernels configured within each residual block simultaneously encode local texture, regional structure, and spectral features, and achieve effective transfer between feature layers through residual connections. After pooling and fully connected mapping, the model outputs pixel-wise classification results at the Softmax layer, thus achieving the final classification map. This structure effectively integrates spectral information, spatial texture, and multi-scale structural features, demonstrating significant advantages for semantic segmentation of hyperspectral images.

[0037] Adaptive Spectral-Spatial Kernel Attention Module Structure as Figure 3 As shown, a dynamic feature selection mechanism is established between the spectral and spatial domains to efficiently capture discriminative spectral-spatial fusion features in hyperspectral images. The core advantage of this module lies in its ability to dynamically allocate feature weights based on the spatial structural complexity and spectral variation characteristics of ground features.

[0038] The pure spectral branch employs a 1×1 spatial convolution kernel, operating only along the spectral direction (e.g., 7 consecutive bands) on a single pixel to learn pixel-level local spectral variation features, forming a spectral-dominated feature map. This branch focuses on reflecting subtle differences in the spectral dimension of ground materials, exhibiting stronger discriminative power for homogeneous regions. The spatial-spectral joint branch uses a 3×3 spatial window and performs joint convolution along the spectral direction, simultaneously characterizing local spatial texture structure and cross-band spectral response, thus obtaining regional-level spectral-spatial composite features. This branch is more effective in areas with mixed ground features, edge transitions, and complex textures. The feature maps obtained from both branches are then subjected to global average pooling (GAP), generating a dimension of... A vector p is used to represent the overall spectral-spatial statistics. Then, vector p is... Further processing of p generates the dimension as follows: The intermediate feature z is then processed by a fully connected layer to generate two dimensions. vector and The importance weight of each branch on the entire feature map is calculated accordingly, and finally, the two branches are normalized and weighted and fused using Softmax. The resulting fused features achieve an optimal balance between spectral consistency and spatial structure representation, thereby effectively improving the accuracy of subsequent classification and segmentation.

[0039] In the residual block stacking structure of the model, each residual block embeds an EFR (Enhanced Feature Recalibration) attention module to achieve adaptive recalibration of the convolutional feature channels. The goal of the EFR attention module is not to directly filter the original bands, but to perform weighted optimization on the "band combinations" corresponding to the convolutional features. This module compresses the spatial information of each feature channel into a single scalar through a GAP operation, forming a... The statistical vector represents the response intensity of each channel globally. The EFR module adaptively selects the convolution kernel size K based on the number of input channels B using the function ψ(B), achieving adaptive parameter tuning for different network depths and feature dimensions. Then, the adaptively determined convolution kernel size (e.g., k=3) is used to... The vectors are convolved to obtain new channel description vectors, the main purpose of which is to learn the interaction relationships between channels and model the dependency structure within the convolutional features. The convolution output is then processed by an activation function. The convolutional features undergo a nonlinear transformation to enhance the response of important feature channels and suppress redundant features. Finally, the activated vector is applied to the original input feature X through channel-wise element-wise multiplication to achieve weighted enhancement of the feature channels and complete the adaptive recalibration of key convolutional features.

[0040] To comprehensively evaluate the stability and generalization ability of the model in the alteration mineral identification task, this study introduced a multi-index performance evaluation system during the testing phase, including key indicators such as overall precision, category precision, recall, F1-score, and intersection-over-union ratio (IoU). Based on the classification results of the test samples, the semantic segmentation model showed good performance under the multi-index system, achieving both high overall classification precision and maintaining a relatively balanced and stable recognition ability across mineral categories. Specifically, the overall precision reached 0.8568, indicating that the model can correctly classify the vast majority of pixels globally. Meanwhile, the Kappa coefficient was 0.7410, significantly higher than that of random classification, indicating strong reliability and consistency of the model. The F1 values ​​for each mineral category were between 0.70 and 0.81, and the production precision and user precision for dolomite, quartz, and anhydrite were all around 0.80, indicating that the model not only identified minerals sufficiently but also reliably in these categories. The mean intersection-overlap ratio (mIoU = 0.673), as an important indicator of spatial overlap, demonstrates that the model can still achieve high spatial consistency under complex geological backgrounds and can effectively recover the spatial distribution characteristics of minerals. This invention trains an A2S2KResNet semantic segmentation model to extract four alteration minerals—barite, dolomite, quartz, and anhydrite—from two spectral bands: visible-near infrared and shortwave infrared. The specific distribution of these alteration minerals can be used to determine the location of the ore body and the intensity of mineralization.

[0041] Metallogenic prospect prediction involves the effective extraction, processing, and comprehensive analysis of mineralization anomaly information from diverse data sources such as geology, geophysical exploration, and remote sensing. Primarily relying on a GIS platform, it utilizes mathematical methods to study various factors related to mineral formation, promoting innovative and accurate mineral exploration strategies, and ultimately delineating metallogenic prospect prediction areas, thus contributing to new heights in mineral exploration. This example analyzes the regional metallogenic geological background, summarizes metallogenic regularities, and identifies seven types of factors favorable to mineralization. These factors mainly include the existence, density, and orientation of linear structures in the study area, as well as the spatial distribution of four types of alteration minerals obtained through semantic segmentation models. By appropriately weighting each ore-controlling factor, the accuracy of the metallogenic prediction model can be optimized, effectively screening potentially metallogenic areas.

[0042] The weighted evidence method is a binary image statistical analysis model based on probabilistic uncertainty and Bayes' theorem. In its application to mineralization prediction, it essentially treats each piece of geological information as an evidence weight factor for predicting potential mineralization areas. By superimposing and analyzing these evidence weight factors, the weight of each factor is determined, and the magnitude of the weight determines the contribution rate of each evidence factor to the mineralization prediction. Finally, based on the factors significant for mineralization prediction and the posterior probability of known mineral occurrences, potential mineralization areas are delineated.

[0043] The metallogenic prediction process using the weighted evidence model includes: selecting linear tectonic factors and remote sensing alteration anomalies as evidence layers based on the metallogenic model and regularity of the study area, and establishing a metallogenic geological thematic layer; establishing grid cells according to the size of the study area; calculating prior probabilities by combining known mineral occurrences within the study area and grid cells; calculating the weight value of each factor using the prior probabilities and metallogenic evidence factors, performing independence tests on all factors, and selecting the evidence factor layer favorable for metallogenesis; calculating the posterior probability of the selected evidence factor layer to obtain the posterior probability map, and delineating the prospective area for metallogenic prediction. Linear tectonic factors mainly appear as continuous or discontinuous linear volumes in remote sensing images. Combined with principal component analysis in the LINE module of PCI software, linear tectonic factors favorable for metallogenesis can be extracted.

[0044] like Figure 4 As shown, the specific process of using the evidence weight model for mineralization prediction in S5 is as follows: Step 1: Based on the metallogenic model and metallogenic regularity, select tectonic structures and remote sensing alteration anomalies as evidence layers to establish metallogenic geological thematic layers; Step 2: Establish appropriate grid cells based on the size of the study area; Step 3: Calculate the prior probability by combining the known mineral deposits and grid cells within the study area; Step 4: Using prior probabilities and mineralization-related evidence factors, calculate the weight value of each factor, perform independence tests on all factors, and screen out the layer of evidence factors favorable to mineralization. Step 5: Calculate the posterior probability of the selected evidence factor layers to obtain the posterior probability map, and delineate the prospective mineral exploration areas based on multivariate information.

[0045] This embodiment uses the ZY-1 hyperspectral remote sensing imagery to extract alteration minerals and linear structures from the lead-zinc mining area and its surrounding rocks. Based on the extracted information, a mineral exploration prediction model is established to delineate potential metallogenic areas. The prediction results are then evaluated and verified. The general steps are as follows: (1) Hyperspectral data preprocessing was carried out based on the characteristics of the ZY-1 remote sensing data. Then, based on the spectral hourglass analysis module of ENVI5.6 software, hyperspectral alteration mineral mapping was carried out in the study area, and four alteration minerals, namely barite, dolomite, quartz and anhydrite, were successfully extracted.

[0046] (2) Based on the jointly extracted alteration information using the spectral angle filling (SAM) and hybrid modulation matched filtering (MTMF) methods, representative pixels were selected from the image to create a dataset suitable for semantic segmentation models. The A2S2K-ResNet deep learning model, designed specifically for hyperspectral remote sensing images, was used to successfully extract the alteration information of the study area, thereby improving the accuracy of alteration information extraction from hyperspectral data.

[0047] (3) Linear structural factors were extracted from the study area, and alteration information extracted by the semantic segmentation model was combined to establish a mineralization prediction model based on 7 evidence layers. After independent verification of known mineralization information, the prediction results showed that the prediction results were ideal and highly credible, and had certain indicative significance for further geological prospecting in the region.

[0048] Furthermore, this embodiment also provides a hyperspectral mineralization and alteration information extraction system based on semantic segmentation, including a geological pre-analysis module, a hyperspectral data acquisition module, a preprocessing module, an alteration information extraction module, and a mineralization prediction module; wherein, The geological pre-analysis module is used to explore the study area based on geological data and analyze geological information related to mineralization in the study area; The hyperspectral data acquisition module is used to acquire hyperspectral data of the metallogenic belt in the study area based on geological information. The preprocessing module is used to preprocess the hyperspectral data to obtain preprocessed hyperspectral data, which is then used to create a dataset suitable for semantic segmentation. The alteration information extraction module is used to construct an A2S2K-ResNet semantic segmentation model and extract mineral alteration information of the study area through the A2S2K-ResNet semantic segmentation model. The mineralization prediction module is used to perform remote sensing geological interpretation of the study area. It combines principal component analysis with PCI software to extract linear tectonic factors that are conducive to mineralization. Based on the extracted mineral alteration information, combined with the linear tectonic factors and mineral data, an evidence weight model is constructed to delineate the prospective mineralization area.

[0049] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for extracting hyperspectral mineralization and alteration information based on semantic segmentation, characterized in that, The specific steps include the following: The study area was explored based on geological data, and geological information related to mineralization in the study area was analyzed. Based on the geological information, hyperspectral data of the metallogenic belt in the study area were obtained; The hyperspectral data is preprocessed to obtain preprocessed hyperspectral data, which is then used to create a dataset suitable for semantic segmentation. An A2S2K-ResNet semantic segmentation model was constructed, and mineral alteration information of the study area was extracted using the A2S2K-ResNet semantic segmentation model. Remote sensing geological interpretation was performed on the study area, and linear tectonic factors conducive to mineralization were extracted using PCI software in conjunction with principal component analysis. Based on the extracted mineral alteration information, combined with the linear tectonic factors and mineral data, an evidence weight model was constructed to delineate the prospective mineralization area.

2. The method for extracting hyperspectral mineralization and alteration information based on semantic segmentation according to claim 1, characterized in that, The geological information related to mineralization includes favorable strata, main structures and ore-controlling structures, spatial distribution characteristics of mineralization points, and regularity of mineralization and rich mineralization conditions.

3. The method for extracting hyperspectral mineralization and alteration information based on semantic segmentation according to claim 1, characterized in that, The preprocessing operations for the hyperspectral data include band removal, band combination, radiometric calibration, atmospheric correction and orthorectification, noise filtering and resampling.

4. The method for extracting hyperspectral mineralization and alteration information based on semantic segmentation according to claim 3, characterized in that, The steps for radiometric calibration of the hyperspectral data are as follows: Combine visible light, near-infrared, and short-wave infrared data files into bands; Read in the wavelength range, center wavelength, and information gain value for each band according to the given satellite calibration file; The calibration parameters were automatically extracted using the radiometric calibration module of ENVI 5.6 software to complete the radiometric calibration of the hyperspectral data.

5. The method for extracting hyperspectral mineralization and alteration information based on semantic segmentation according to claim 4, characterized in that, The expression for the radiation calibration is: ； in, express Radiance values ​​for the band, express Gain value of the band, DN Represents a numerical quantization value. express The offset value of the band.

6. The method for extracting hyperspectral mineralization and alteration information based on semantic segmentation according to claim 1, characterized in that, The A2S2K-ResNet semantic segmentation model is a residual network that integrates an attention mechanism and an adaptive spectral-spatial kernel. It supports simultaneous modeling in both spatial and spectral dimensions and extracts altered minerals by dynamically weighting the contributions of different branches through the attention mechanism.

7. The method for extracting hyperspectral mineralization and alteration information based on semantic segmentation according to claim 6, characterized in that, The A2S2K-ResNet semantic segmentation model constructs an S×S×B spectral-spatial input through neighborhood block extraction, and then uses a multi-scale residual learning structure to achieve feature representation. Each residual block is configured with convolutional kernels of different scales to synchronously encode local texture, regional structure and spectral features, and achieves effective transfer between feature layers through residual connections. After pooling and fully connected mapping, the model outputs pixel-by-pixel alteration mineral grading results at the Softmax layer.

8. The method for extracting hyperspectral mineralization and alteration information based on semantic segmentation according to claim 1, characterized in that, The specific steps for constructing an evidence weighting model to delineate prospective mineralization areas are as follows: Based on the metallogenic model and metallogenic regularity of the study area, linear tectonic factors and remote sensing alteration anomalies were selected as evidence layers to establish metallogenic geological thematic layers. Establish grid cells based on the size of the study area; The prior probabilities were calculated by combining known mineral deposits and grid cells within the study area. By utilizing prior probabilities and mineralization-related evidence factors, the weight value of each factor is calculated, independence tests are performed on all factors, and a layer of evidence factors favorable to mineralization is selected. For the selected evidence factor layers, the posterior probability is calculated to obtain the posterior probability map, and the prospective mineralization area is delineated.

9. A hyperspectral mineralization alteration information extraction system based on semantic segmentation, characterized in that, It includes a geological pre-analysis module, a hyperspectral data acquisition module, a preprocessing module, an alteration information extraction module, and a mineralization prediction module; among which, The geological pre-analysis module is used to explore the study area based on geological data and analyze the geological information related to mineralization in the study area. The hyperspectral data acquisition module is used to acquire hyperspectral data of the metallogenic belt in the study area based on the geological information. The preprocessing module is used to perform preprocessing operations on the hyperspectral data to obtain preprocessed hyperspectral data and create a dataset suitable for semantic segmentation. The alteration information extraction module is used to construct an A2S2K-ResNet semantic segmentation model and extract mineral alteration information of the study area through the A2S2K-ResNet semantic segmentation model. The mineralization prediction module is used to perform remote sensing geological interpretation of the study area, extract linear tectonic factors conducive to mineralization in PCI software by combining principal component analysis, and construct an evidence weight model based on the extracted mineral alteration information, combined with the linear tectonic factors and mineral data, to delineate the prospective mineralization area.

Images