A Deep Learning-Based Satellite Hyperspectral Alteration Mineral Remote Sensing Quantitative Identification Method

Abstract

This invention relates to the field of hyperspectral alteration technology, and more particularly to a satellite hyperspectral alteration mineral remote sensing quantitative identification method based on deep learning. The method includes the following steps: acquiring hyperspectral data, DEM data, and atmospheric parameter data from the Gaofen-5B satellite; performing multi-source noise suppression processing on the Gaofen-5B satellite hyperspectral data to obtain a noise-corrected dataset; performing terrain shadow correction based on the noise-corrected dataset and DEM data, and atmospheric correction in conjunction with atmospheric parameter data to obtain a pre-corrected hyperspectral dataset. This invention effectively improves the quality and characterization capability of hyperspectral data through multi-source data fusion, noise and environmental interference correction, key band screening, and multi-dimensional deep feature extraction and fusion, providing a solid data foundation for subsequent refined mineral identification and analysis.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral alteration technology, and in particular to a satellite hyperspectral alteration mineral remote sensing quantitative identification method based on deep learning. Background Technology

[0002] Hyperspectral alteration minerals are secondary minerals formed when the original minerals in rocks undergo changes in chemical composition or physical properties due to geological processes (such as hydrothermal activity and weathering). They exhibit unique spectral characteristics in hyperspectral data (such as absorption peaks at specific wavelengths and abrupt changes in reflectance). Traditional methods for identifying hyperspectral alteration minerals primarily rely on the experience of remote sensing geologists to select typical spectral features of alteration minerals. These features are extracted using methods such as band ratio analysis, principal component analysis, and spectral angle classification. Experienced thresholds are then set to segment or classify the extracted features, ultimately outputting the identification results. However, these traditional methods rely on manual selection of spectral features, making it difficult to cover the complex and diverse spectral details of alteration minerals. They are prone to missing key features, have weak feature extraction capabilities for mixed pixels, and involve cumbersome processing procedures requiring manual parameter adjustments. Furthermore, they lack adaptability to different regions and types of alteration minerals, making it difficult to process large amounts of data in batches. Summary of the Invention

[0003] Therefore, it is necessary to provide a satellite hyperspectral alteration mineral remote sensing quantitative identification method based on deep learning to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a satellite hyperspectral alteration mineral remote sensing quantitative identification method based on deep learning is proposed, the method comprising the following steps:

[0005] Step S1: Acquire hyperspectral data, DEM data, and atmospheric parameter data from the Gaofen-5B satellite; perform multi-source noise suppression processing on the Gaofen-5B satellite hyperspectral data to obtain a noise correction dataset; perform terrain shadow correction based on the noise correction dataset and DEM data, and perform atmospheric correction in conjunction with atmospheric parameter data to obtain a pre-corrected hyperspectral dataset.

[0006] Step S2: Perform adaptive band selection on the pre-calibrated hyperspectral dataset to obtain a subset of key bands;

[0007] Step S3: Construct a 3D-2D hybrid convolutional network based on a subset of key bands, and input the pre-corrected hyperspectral data block into the hybrid convolutional network to obtain a 3D spectral feature tensor and a 2D spatial feature map;

[0008] Step S4: Perform attention weighting and channel splicing fusion processing on the 3D spectral feature tensor and the 2D spatial feature map to obtain the fused feature representation.

[0009] The beneficial effects of this invention are as follows: It collects hyperspectral data from the Gaofen-5B satellite, digital elevation model (DEM) data, and atmospheric parameter data, and performs multi-source noise suppression processing on the hyperspectral data, significantly reducing sensor noise and environmental interference, and improving the signal-to-noise ratio. Subsequently, terrain shadow correction is performed using DEM data to reduce spectral signal distortion caused by terrain undulations. Simultaneously, atmospheric correction is implemented based on atmospheric parameter data to correct the influence of atmospheric scattering and absorption on the spectral data, resulting in a more realistic and reliable pre-corrected hyperspectral dataset. This stage effectively eliminates systematic errors caused by environmental and observational conditions through multi-source data fusion and comprehensive correction. Step S2, based on the pre-corrected data, performs adaptive band selection to filter out a subset of key bands reflecting the physicochemical properties of the target, effectively reducing data dimensionality, redundant information, and computational complexity, providing more discriminative input features for subsequent feature extraction. Step S3 utilizes a subset of key bands to construct a 3D-2D hybrid convolutional neural network. 3D convolution operations capture the joint features of hyperspectral data in both spectral and spatial dimensions, uncovering complex correlations between spectral channels and local spatial features. Simultaneously, 2D convolution operations extract purely spatial features, enhancing the expressive power of spatial structure information. After inputting pre-corrected hyperspectral data blocks, the network outputs a 3D spectral feature tensor and a 2D spatial feature map, achieving a multi-dimensional representation of spectral and spatial features. Finally, step S4 applies attention weighting mechanisms to the two types of features, dynamically adjusting the feature response weights of each channel and spatial location to suppress irrelevant noise and highlight important information. Channel-level splicing and fusion are then used to form a fused feature representation, integrating spectral and spatial information and improving the richness and discriminative power of feature expression. The overall process achieves multi-source data fusion, noise and environmental interference correction, key band selection, and multi-dimensional deep feature extraction and fusion at the data level, effectively improving the quality and representational capabilities of hyperspectral data and providing a solid data foundation for subsequent refined mineral identification and analysis. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the steps of a satellite hyperspectral alteration mineral remote sensing quantitative identification method based on deep learning.

[0011] Figure 2 for Figure 1 A detailed flowchart illustrating the implementation steps of step S2.

[0012] Figure 3 A spatial distribution map of altered minerals in the visualization icons;

[0013] 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

[0014] 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.

[0015] 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.

[0016] 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.

[0017] To achieve the above objectives, please refer to Figures 1 to 3 A satellite hyperspectral alteration mineral remote sensing quantitative identification method based on deep learning, the method comprising the following steps:

[0018] Step S1: Acquire hyperspectral data, DEM data, and atmospheric parameter data from the Gaofen-5B satellite; perform multi-source noise suppression processing on the Gaofen-5B satellite hyperspectral data to obtain a noise correction dataset; perform terrain shadow correction based on the noise correction dataset and DEM data, and perform atmospheric correction in conjunction with atmospheric parameter data to obtain a pre-corrected hyperspectral dataset.

[0019] Step S2: Perform adaptive band selection on the pre-calibrated hyperspectral dataset to obtain a subset of key bands;

[0020] Step S3: Construct a 3D-2D hybrid convolutional network based on a subset of key bands, and input the pre-corrected hyperspectral data block into the hybrid convolutional network to obtain a 3D spectral feature tensor and a 2D spatial feature map;

[0021] Step S4: Perform attention weighting and channel splicing fusion processing on the 3D spectral feature tensor and the 2D spatial feature map to obtain the fused feature representation.

[0022] In this embodiment of the invention, reference Figure 1 The diagram shown is a flowchart illustrating the steps of the satellite hyperspectral alteration mineral remote sensing quantitative identification method based on deep learning according to the present invention. In this example, the satellite hyperspectral alteration mineral remote sensing quantitative identification method based on deep learning includes the following steps:

[0023] Step S1: Acquire hyperspectral data, DEM data, and atmospheric parameter data from the Gaofen-5B satellite; perform multi-source noise suppression processing on the Gaofen-5B satellite hyperspectral data to obtain a noise correction dataset; perform terrain shadow correction based on the noise correction dataset and DEM data, and perform atmospheric correction in conjunction with atmospheric parameter data to obtain a pre-corrected hyperspectral dataset.

[0024] In this embodiment of the invention, hyperspectral data, corresponding DEM data, and atmospheric parameter data from the Gaofen-5B satellite are acquired. These three types of data are registered and resampled in terms of spatial resolution and projected coordinate system to form a raw dataset with a consistent spatial grid. Radiometric and absorption corrections are applied to the raw radiative values ​​based on the atmospheric parameter data. An atmospheric correction parameter set (including necessary quantities such as aerosol optical thickness, gas absorption coefficient, and solar zenith angle) is calculated and output. The radiative units are converted into apparent surface reflectance for subsequent processing. Multi-source noise suppression processing is performed on the spectral sequence of each pixel of the Gaofen-5B satellite. Multi-scale decomposition is performed using the Daubechies 4th-order wavelet transform. Thresholding correction is applied to the wavelet coefficients based on noise statistical characteristics, and the spectrum is reconstructed to obtain the noise-corrected dataset. Simultaneously, known bad tracks and strong water vapor absorption bands are masked, interpolated, or spectral bands removed to eliminate outliers introduced by systematic bad tracks and atmospheric absorption. Based on the registered DEM data and image metadata, pixel slope, aspect, and incident / emission geometry are calculated to construct a shadow correction coefficient set (implemented through cosine incident angle correction or C-correction with correction factor, etc.). The shadow correction coefficient set and atmospheric correction parameter set are applied to the noise correction dataset to perform terrain shadow and atmospheric coupled correction, resulting in an integrated pre-corrected hyperspectral dataset. When needed, the above atmospheric correction parameter set, shadow correction coefficient set, and noise correction data are incorporated into an end-to-end reconstruction network for joint learning and reconstruction to output a pre-corrected hyperspectral dataset for subsequent adaptive band selection and feature extraction.

[0025] Step S2: Perform adaptive band selection on the pre-calibrated hyperspectral dataset to obtain a subset of key bands;

[0026] In embodiments of the present invention, the pre-corrected hyperspectral dataset is represented as a three-dimensional tensor or as a set of sample blocks with labeled input sample sets. A set of band discrimination and quality metrics is calculated for each band to form a band importance score set. Specifically, this includes: performing band-by-band normalization (e.g., standardization based on median and interquartile range) to offset differences in reflectance scales across different bands; then estimating the signal-to-noise ratio (SNR) of each band based on wavelet residuals in the noise-corrected dataset to obtain the SNR index; and performing continuous spectral decurving on the spectrum of each pixel. Within a predefined diagnostic absorption band window, the average and maximum absorption depths are calculated as spectral absorption response indicators. The first and second derivative energies of each band are calculated to characterize spectral inflection point sensitivity. When labeled mineral categories exist, the mutual information or F-statistic between each band and the label is calculated based on the labeled samples to obtain a supervised discrimination score. The above normalized scores are weighted or synthesized using principal components to obtain a comprehensive importance score, and a cumulative contribution curve is plotted in descending order of score. To remove redundancy, an inter-band Pearson correlation matrix is ​​constructed with 1-|r| as the distance... Hierarchical clustering is performed, and clusters are cut according to the inflection point of the cluster distance or a set relevant threshold. Within each cluster, the band with the highest comprehensive importance score is selected as the representative band. The threshold is adaptively determined by inflection point detection of the cumulative contribution curve or stability selection under cross-validation (sparse regression / L1 regularization, recursive feature elimination, and bootstrap stability selection can also be used as alternative strategies). Finally, a subset of key bands is output (represented by band index or center wavelength), and the selection result of the key band subset is consistent with the discriminative performance index of several classifiers on the cross-validation set to complete the adaptive screening closed loop.

[0027] Step S3: Construct a 3D-2D hybrid convolutional network based on a subset of key bands, and input the pre-corrected hyperspectral data block into the hybrid convolutional network to obtain a 3D spectral feature tensor and a 2D spatial feature map;

[0028] In the embodiments of the present invention, the pre-calibrated hyperspectral dataset is first block-based and batch-processed based on a subset of key bands at the data level. The pre-calibrated hyperspectral data is divided into a set of pre-calibrated hyperspectral data blocks of shape N×9×9×Bk according to the spatial neighborhood (where Bk represents the number of bands in the subset of key bands). For each data block, zero-mean-unit variance standardization and continuous spectral line correction are performed by band to eliminate scale differences and residual baseline drift. Then, the standardized data blocks are fed in parallel into the constructed 3D-2D hybrid convolutional network. The hybrid network adopts a dual-branch structure: the 3D branch aims to preserve spectral-spatial coupling information, using progressively layered 3D convolutional units to achieve joint representation of local spatial neighborhoods and continuous spectral bands; specifically, 3D convolutional layers are set sequentially (convolutional kernel sizes are (3,3,7), (3,3,5), (3,3,3), with corresponding strides of (1,1,2), (1,1,1), etc., to achieve hierarchical downsampling in the spectral dimension), and batch normalization and ReLU activation are performed in parallel after each convolution. Residual short-circuiting is introduced where necessary to preserve shallow spectral details. Finally, global average pooling or learnable spectral projection is performed on the spectral dimension at the top of the 3D branch to converge spectral energy into the channel dimension, thereby forming a 3D spectral feature tensor that preserves spectral details and encodes spatial neighborhood information. The 2D branch obtains single-channel or multi-channel spatial input by spectral convergence of key band subsets (e.g., principal component mapping or band mean / weighted sum mapping). It then uses depth-separable 2D dilated convolutional units (e.g., 3×3 kernel, dilation=2) and multi-scale pooling modules to extract spatial texture and structural features, outputting an aligned 2D spatial feature map. To ensure feature compatibility between the two branches, the network employs channel alignment, feature normalization, and learnable projection matrices at branch interactions. During training and inference, random perturbations and data augmentation strategies (spectral jitter, local noise injection, spatial flipping, etc.) are applied to intermediate features to stabilize feature distribution. The network ultimately outputs in parallel a 3D spectral feature tensor and a 2D spatial feature map for subsequent processing, which can be directly used in the subsequent attention weighting and channel concatenation fusion steps.

[0029] Step S4: Perform attention weighting and channel splicing fusion processing on the 3D spectral feature tensor and the 2D spatial feature map to obtain the fused feature representation.

[0030] In embodiments of the present invention, the 3D spectral feature tensor X s (Dimensions are denoted as N×C) s ×H×W) and 2D spatial feature map X sp (Dimensions are denoted as N×C) sp ×H×W) Implement a gradual transformation of attention weighting and channel splicing: First, for X s Global average pooling is performed along the spatial dimension to generate the channel description vector g. sThe spectral attention weight matrix A is obtained by using two fully connected layers (dimensionality reduction-ReLU-dimensionality increase) mapping and Sigmoid normalization. s (Dimensions N×C) s ×1×1); simultaneously for X sp Average pooling and max pooling are calculated along the channel dimension and concatenated along the channel dimension. After 3×3 convolution, batch normalization, and sigmoid mapping, the spatial attention weight map A is obtained. sp (Dimensions N×1×H×W); then A s Perform channel broadcast and communicate with X s Element-wise multiplication is used to complete the spectral channel reweighting for A. sp Perform pixel broadcasting and connect with X sp Element-wise multiplication is used to reweight spatial pixels, resulting in the reweighted feature X. ' s With X ' sp To ensure consistent shapes in subsequent splicing, X' was adjusted accordingly. s With X' sp Channels are projected to a uniform number of channels C using 1×1 convolution. f For any feature with inconsistent spatial resolution, bilinear interpolation or deconvolution up / downsampling is used to align the spatial dimensions. Subsequently, the projection results are batch normalized and linearly activated to produce aligned features X”. s With X” sp ; will X” s With X s p” is spliced ​​in the channel dimension to obtain X c oncat (size N×2C) f ×H×W), for X c The `oncat` function applies 1×1 convolutions to fuse channel information and then performs batch normalization and ReLU activation to generate an intermediate fused representation. Optionally, residual short-circuiting is used to add the linear mapping before projection to the fused result to preserve the original information. The final output is a fused feature representation X processed by normalization and optional dropout. fused This is used by subsequent classifiers or post-processing modules; during training, random perturbations and regularizations in the spectral and spatial domains are applied to the attention branch and fusion layer to constrain numerical stability; during inference, the above-mentioned stepwise data transformations are performed in a pipelined manner using the same operators to obtain consistent fused feature representations.

[0031] Preferably, step S1 includes the following steps:

[0032] Step S11: Collect raw hyperspectral data, corresponding DEM data and atmospheric parameter data from the Gaofen-5B satellite to construct the input sample set;

[0033] Step S12: Divide the original hyperspectral data into spatial blocks of 9×9 pixels, and label the mineral category with the center pixel of the block to obtain a labeled input sample set;

[0034] Step S13: Perform radiation correction and absorption correction on the atmospheric parameter data to obtain the atmospheric correction parameter set;

[0035] Step S14: Calculate the terrain shadow correction coefficients based on the DEM data to obtain the shadow correction coefficient set;

[0036] Step S15: Apply the atmospheric correction parameter set and the shadow correction coefficient set to the noise correction dataset to obtain the pre-corrected hyperspectral dataset.

[0037] In this embodiment of the invention, raw hyperspectral data, corresponding DEM data, and atmospheric parameter data from the Gaofen-5B satellite are acquired. Registration and resampling of these three types of raw quantities are performed on projected coordinates, spatial resolution, and timestamps to form a spatially consistent input sample set. Based on this input sample set, multi-source noise suppression is applied to the raw hyperspectral data to construct a noise correction dataset. Specifically, Daubechies 4th-order wavelet decomposition is performed on the spectrum of each pixel, and thresholding reconstruction is applied to the wavelet coefficients to remove high-frequency noise. Simultaneously, masks, spectral interpolation, or local fitting are applied to repair abnormal spectra for known bad channels and strong water vapor absorption bands. After obtaining the noise correction dataset, pixel slope, aspect, and incident / exit geometry are calculated based on the DEM data, and cosine incident angle correction is applied. The shadow correction coefficient set is obtained by fitting the data using methods such as C-correction. Simultaneously, based on atmospheric parameter data, aerosol optical thickness, water vapor content, and absorption coefficients of major gases are estimated through radiative transfer model inversion or empirical linear correction to form the atmospheric correction parameter set described in this step. Finally, the atmospheric correction parameter set and the shadow correction coefficient set are applied pixel-by-pixel and spectral-by-spectral to the noise correction dataset. Multiplication, division, and masking are performed according to the conversion relationship from radiation to apparent reflectance and the correction factor, and spectral interpolation is performed where necessary to obtain a pre-corrected hyperspectral dataset for subsequent analysis. In the sample construction stage, the pre-corrected hyperspectral data is divided into spatial blocks using a fixed 9×9 pixel window, and an labeled input sample set is generated using the lithofacies / mineral annotations of the central pixel of each block for subsequent training and validation.

[0038] Preferably, step S2 includes the following steps:

[0039] The feature importance of each band in the pre-calibrated hyperspectral dataset is evaluated to obtain a set of band importance scores;

[0040] Based on the band importance score set, an adaptive threshold screening is applied to select bands with importance higher than the preset threshold to form an initial key band subset;

[0041] Redundant bands are removed from the initial key band subset to obtain the final key band subset.

[0042] In embodiments of the present invention, the pre-calibrated hyperspectral dataset is represented as a three-dimensional data cube of size H×W×B, where H and W are the spatial dimensions and B is the number of bands. For each band b∈{1,…,B}, the corresponding two-dimensional reflectance matrix is ​​extracted, and its feature importance index is calculated based on the labeled sample set or statistical distribution characteristics. The index may include mutual information between the band and the target mineral category, F-statistic, spectral absorption depth, spectral curvature change rate, signal-to-noise ratio (SNR), etc. Different indices are mapped to the same dimension interval through normalization processing. Finally, the band importance score set S={s1,s2,…,s…} is obtained by weighted linear combination or principal component projection. B Subsequently, an adaptive threshold is calculated based on the distribution characteristics of S. The threshold can be determined by detecting the inflection point of the cumulative contribution rate curve, Otsu's method segmentation, or the optimal split point based on cross-validation accuracy, and then selecting those that satisfy S. b ≥T adaptive The set of bands forms the initial key band subset B. init To remove redundant information between bands, B init Calculate the Pearson correlation coefficient matrix R for each pair of bands in the spectrum, and then compare its absolute value with a preset correlation threshold r. th The comparison will show a high correlation (|R) ij |≥r th Bands with the same importance score are grouped into the same cluster, and within each cluster, the band with the highest importance score is retained as the representative band. Other redundant bands are removed, and the final output is a subset B of key bands without strong correlation redundancy. final This subset is stored in the form of band index or center wavelength and serves as input for subsequent feature extraction and model training.

[0043] Preferably, step S3 includes the following steps:

[0044] Step S31: Perform the first layer of 3D convolution operation on the pre-corrected hyperspectral data block to obtain the initial spectral feature tensor, where the convolution kernel size is 3,3,7 and the stride is 1,1,2;

[0045] Step S32: Perform a second-layer 3D convolution operation on the initial spectral feature tensor to obtain an enhanced spectral feature tensor, wherein the convolution kernel size is 3,3,5 and the stride is 1,1,1;

[0046] Step S33: Perform a third-layer 3D convolution operation on the enhanced spectral feature tensor, and aggregate the spectral dimensions into a two-dimensional feature map based on 3D global average pooling. Then, perform a depthwise separable dilated 2D convolution on the two-dimensional feature map to obtain a 2D spatial feature map, where the convolution kernel size is 3,3,3.

[0047] In embodiments of the present invention, the pre-calibrated hyperspectral data block is represented as having a size of N×H×W×B. k A five-dimensional tensor of size ×1, where N is the batch size, H and W are the spatial dimensions, and B... k The tensor represents the number of bands in the key band subset, and the last dimension is the single-channel identifier. In step S31, using this tensor as input, a first-layer three-dimensional convolution operator is constructed. The kernel size is set to (3,3,7), and the stride is set to (1,1,2). Local window convolution and spectral downsampling operations are implemented in the spatial and spectral dimensions, respectively. The convolution weights are updated through backpropagation during the training phase to extract the local spatial structure and spectral patterns within a continuous range of 7 bands. After batch normalization and ReLU activation, the convolution output generates a size of N×C1×H×W×B′. k The initial spectral feature tensor, where C1 is the number of output channels in the first layer, B′ k This is the result after downsampling the spectral dimension. In step S32, the initial spectral feature tensor is input into the second-layer three-dimensional convolution operator with a kernel size of (3,3,5) and a stride of (1,1,1). While maintaining spatial resolution and spectral length, the spectral context receptive field is expanded to enhance cross-band correlation. The output enhanced spectral feature tensor has a size of N×C2×H×W×B″. k The process involves batch normalization and nonlinear activation. In step S33, the enhanced spectral feature tensor is input into the third-layer three-dimensional convolution operator with a kernel size of (3,3,3). Local feature fusion is performed simultaneously in the spectral and spatial dimensions. After output, three-dimensional global average pooling is performed on the spectral dimension to aggregate the spectral features of each spatial location into a single vector, thereby compressing the five-dimensional tensor into a four-dimensional two-dimensional feature map representation N×C3×H×W. Subsequently, depth-separable dilated two-dimensional convolution is performed on the two-dimensional feature map with a kernel size of (3,3). The dilation rate is set according to the experiment to expand the spatial receptive field. The depth-separable strategy separates the spatial convolution and channel convolution to reduce parameter redundancy and retain feature resolution. The final output two-dimensional spatial feature map is used as the input data for subsequent attention weighting and fusion.

[0048] Preferably, step S4 includes the following steps:

[0049] Step S41: Calculate the spectral attention weights on the 3D spectral feature tensor to obtain the spectral attention weight matrix;

[0050] Step S42: Calculate the spatial attention weights on the 2D spatial feature map to obtain the spatial attention weight map;

[0051] Step S43: Multiply the spectral attention weight matrix element-wise with the 3D spectral feature tensor, and multiply the spatial attention weight map element-wise with the 2D spatial feature map;

[0052] Step S44: Concatenate the multiplication results according to the channel dimension to obtain the fused feature representation.

[0053] In embodiments of the present invention, the input 3D spectral feature tensor represents the composite information of data in both spatial and spectral dimensions, typically in the order of (spatial height × spatial width × spectral channels). In step S41, spectral attention weights are calculated for this 3D tensor. Essentially, this involves quantifying the importance of each spectral channel using a specific spectral dimension feature extraction method, generating a weight matrix corresponding to the spectral channel dimension. The dimension of this weight matrix is ​​consistent with the number of spectral channels, reflecting the contribution intensity of different spectral channels to the overall features. Subsequently, in step S42, spatial attention weights are calculated for the 2D spatial feature map. The spatial feature map is typically a two-dimensional projection or single-channel feature map extracted from the original 3D tensor, with a dimension of (spatial height × spatial width). The calculation process generates a two-dimensional spatial attention weight map by analyzing the information distribution in the spatial dimension, with each spatial location corresponding to a weight value to reflect the feature salience at that location. Step S43 involves performing element-wise multiplication operations between the two types of attention weights and their corresponding original feature data. Specifically, the spectral attention weight matrix and the 3D spectral feature tensor are broadcast-productted along the spectral channel dimension to enhance or suppress the feature responses of different spectral channels. Simultaneously, the spatial attention weight map is element-wise multiplied with the 2D spatial feature map to adjust the feature intensity of spatial regions, thereby achieving feature weighting along the spatial dimension. Finally, step S44 concatenates the processed spectral weighted 3D tensor and the spatial weighted 2D features along the channel dimension to form a fused feature representation. This concatenation operation expands the number of channels along the data dimension, combining attention enhancement information from both spectral and spatial perspectives, providing richer and more reasonably weighted input features for subsequent feature analysis or classification tasks. The entire process focuses on the calculation and application of multi-dimensional attention weights, finely adjusting the information distribution of different dimensions within the feature tensor through element-wise multiplication to achieve effective fusion and expression of multimodal features.

[0054] As an example of the present invention, reference is made to Figure 2 As shown, step S4 in this example includes:

[0055] Step S411: Perform global average pooling on the input 3D spectral feature tensor to obtain the spectral feature vector;

[0056] Step S412: Apply a fully connected layer to the spectral feature vector and calculate the unnormalized spectral attention score;

[0057] Step S413: Normalize the unnormalized spectral attention scores to obtain the spectral attention weight matrix.

[0058] In embodiments of this invention, the input 3D spectral feature tensor generally has a dimensional structure of (spatial height × spatial width × spectral channels), where the spatial dimension carries the spatial information of pixels, and the spectral channel dimension reflects the spectral response of different bands. Through global average pooling, the mean of this 3D tensor is calculated along the spatial dimension; specifically, the values ​​at all spatial locations within each spectral channel are averaged, ultimately yielding a one-dimensional spectral feature vector with the dimension being the number of spectral channels. This vector can be viewed as a compressed representation of the overall information along the spectral dimension, reflecting the global response intensity of each spectral channel. A fully connected layer is applied to this spectral feature vector, i.e., a set of trainable weight matrices is used to linearly transform the vector and add a bias, generating a new vector representing the attention score of this feature vector across different channels along the spectral dimension. This step essentially learns the interrelationships and importance between spectral channels through parameterized mapping, and the output unnormalized attention score retains the same dimension as the number of spectral channels. The unnormalized attention score vector is normalized, typically using the Softmax or Sigmoid function, converting the scores into weight coefficients within the interval (0,1). Simultaneously, the weight distribution of all spectral channels satisfies certain probabilistic properties (e.g., a sum of 1), forming a spectral attention weight matrix. This weight matrix's dimension corresponds to the number of spectral channels and is used for subsequent weighting operations on the 3D spectral feature tensor, achieving weighted enhancement or suppression of different spectral channels. Overall, this process achieves dimensionality reduction and compression of spectral features through global averaging aggregation of spatial dimensions, then mines inter-channel dependencies through fully connected layers, and finally normalizes the weight distribution using a normalization function to complete the calculation of spectral attention weights. This provides a data-driven attention mechanism foundation for multi-dimensional feature fusion.

[0059] Preferably, step S4 further includes the following:

[0060] The network was pre-trained and fine-tuned based on the USGS altered mineral spectral library, and the parameters of the bottom convolutional layer were frozen.

[0061] The fine-tuned fused feature representation is input into a fully connected classifier to obtain the mineral category probability set;

[0062] Applying Softmax normalization to the mineral category probability set yields a normalized mineral category probability set.

[0063] In embodiments of this invention, the fused feature representation, obtained through a preliminary feature fusion step, is typically in the form of a multidimensional feature tensor or vector, containing rich spectral and spatial information. During the fine-tuning stage, mineral spectral data labeled in the USGS alteration mineral spectral library is used as training samples, and the fused feature is input into a pre-trained neural network. This pre-trained network contains multiple convolutional layers and several fully connected layers, designed for feature extraction and classification tasks. During fine-tuning, a strategy of freezing the parameters of the bottom convolutional layers is adopted, i.e., keeping the weight parameters of the convolutional layers unchanged, only adjusting the higher-level parameters or fully connected layer parameters at the back end of the network. This preserves the pre-trained network's generalization ability to basic spectral features while adapting the network to the fine-grained classification requirements of specific mineral data. After fine-tuning, the fused feature outputs a mineral category probability set through the fully connected classifier part of the network. This probability set corresponds to the unnormalized predicted value of each mineral category, reflecting the network's confidence that the input sample belongs to each category. Subsequently, the Softmax function is used to normalize this probability set, converting the original output into a probability distribution between 0 and 1, ensuring that the sum of all category probabilities is 1, thus forming a normalized mineral category probability set. The normalization result can be used for subsequent mineral identification or decision support. The overall process emphasizes the data transfer and parameter adjustment mechanism in the pre-trained network. It achieves stability and generalization of feature extraction by freezing the bottom convolutional layers, uses fully connected layers to classify features, and finally achieves a probabilistic expression of mineral categories through probability normalization, ensuring the scientific nature of the classification results and data consistency.

[0064] Preferably, step S4 further includes the following:

[0065] A conditional random field model is constructed based on the probability set of mineral categories to extract neighborhood features of adjacent pixels;

[0066] The probability of isolated high-probability misclassified pixels is corrected using a conditional random field model, resulting in a corrected mineral category probability set.

[0067] The confidence level of the corrected mineral category probability set is evaluated to obtain a confidence-labeled dataset.

[0068] In embodiments of this invention, the mineral category probability set serves as input data, representing the probability distribution of each mineral category corresponding to each pixel. It possesses a two-dimensional spatial structure, mapped onto the spatial location of remote sensing images or spectral maps. First, a Conditional Random Field (CRF) model is constructed based on this probability set. As a discriminative probabilistic graphical model, CRF utilizes the spatial relationships and probabilistic information between adjacent pixels to define pixel nodes and their edge connections, forming a graph structure with potential dependencies. This model characterizes the mutual influence between adjacent pixel category labels through joint probability distributions, thereby integrating spatial context features and effectively capturing the spatial consistency of mineral categories within the neighborhood. Second, the constructed CRF model is used to perform probability correction on isolated high-probability misclassified pixels. Specifically, the model maximizes the conditional probability distribution, considering the joint influence of the pixel's own category probability and the category probabilities of neighboring pixels, adjusting the pixel's category prediction probability, reducing the risk of misclassification of isolated outliers, and achieving probability smoothing and spatial consistency optimization. The correction process generates a new mineral category probability set, with a spatial distribution that better conforms to geological continuity and actual mineral distribution characteristics. Finally, the confidence level of the corrected probability set is evaluated by calculating the maximum class probability or other statistical indicators for each pixel to generate a confidence-labeled dataset. This dataset provides a foundation for subsequent accuracy analysis, classification result verification, or decision support. The overall process demonstrates the spatial statistical model's ability to model and optimize the spatial structure of probabilistic data. Through the neighborhood dependency mechanism of CRF, spatial correction and confidence quantification of mineral class probabilities are achieved, thereby improving the spatial consistency of the data and the reliability of classification.

[0069] Preferably, step S4 further includes the following:

[0070] A spatial distribution map of altered minerals is generated based on the corrected mineral category probability set;

[0071] Boundary extraction is performed on the spatial distribution map of altered minerals to obtain the outline of the target mineral region;

[0072] By combining the confidence-labeled dataset, the outline of the target mineral region is labeled to obtain the final spatial distribution map of altered minerals.

[0073] In this embodiment of the invention, a modified mineral category probability set is used as input. This probability set is presented in the form of a two-dimensional matrix, where each element corresponds to a spatial pixel and its corresponding mineral category probability. Based on this probability set, a spatial distribution map of altered minerals is generated by selecting category labels with higher probability values ​​or applying the maximum probability principle. This map is expressed in the form of two-dimensional raster data, indicating that each spatial pixel belongs to a certain mineral category, thus achieving a preliminary mapping of the spatial distribution of minerals. Next, boundary extraction is performed on this spatial distribution map. Boundary extraction involves the spatial structure analysis of continuous regions of the same category in the two-dimensional raster. Contour detection algorithms from image processing, such as gradient-based edge detection or region growing algorithms, are commonly used to identify the spatial boundaries of different mineral category blocks and obtain the contour data of the target mineral region. These contours exist in the form of pixel or spatial coordinate point sequences, reflecting the spatial boundary characteristics of the mineral spatial distribution. Finally, the contours of the target mineral region are labeled using a confidence-labeled dataset. This confidence dataset contains confidence values ​​for corresponding spatial locations, used to characterize the classification accuracy and reliability of each pixel or region. During the contour annotation process, confidence level information is overlaid onto the boundaries or interior areas of the spatial distribution map. Regions with different confidence levels are classified and labeled to form the final spatial distribution map of alteration minerals with confidence level annotations. This process not only clearly defines the spatial extent of minerals but also provides a quality assessment of spatial distribution by incorporating confidence level indicators, providing a data foundation for subsequent geological interpretation and decision analysis. The overall process transforms probabilistic data into a spatial classification map, extracts spatial structure through boundary recognition, and finally integrates confidence level information to construct a complete spatial mineral distribution representation system.

[0074] like Figure 3 The image shown represents the spatial distribution of altered minerals in a visualization.

[0075] Of particular importance is the ability to generate quantitative identification reports in a visualized form by integrating feature representations, mineral category probability sets, and alteration mineral spatial distribution maps.

[0076] 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.

[0077] 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 deep learning-based satellite hyperspectral alteration mineral remote sensing quantitative identification method, characterized in that, The method comprises the following steps: Step S1: Obtain high-spectral data of GaoFen 5B satellite, DEM data and atmospheric parameter data; Perform multi-source noise suppression processing on the high-spectral data of GaoFen 5B satellite to obtain a noise-corrected data set; Perform terrain shadow correction on the noise-corrected data set based on the DEM data, and perform atmospheric correction in combination with the atmospheric parameter data to obtain a pre-corrected high-spectral data set; Step S2: Perform adaptive band selection on the pre-corrected high-spectral data set to obtain a key band subset; Step S3: Construct a 3D-2D hybrid convolution network based on the key band subset, and input the pre-corrected high-spectral data block into the hybrid convolution network to obtain a 3D spectral feature tensor and a 2D spatial feature map; wherein step S3 comprises the following steps: Step S31: Perform a first layer 3D convolution operation on the pre-corrected high-spectral data block to obtain an initial spectral feature tensor, wherein the convolution kernel size is 3, 3, 7, and the step size is 1, 1, 2; Step S32: Perform a second layer 3D convolution operation on the initial spectral feature tensor to obtain an enhanced spectral feature tensor, wherein the convolution kernel size is 3, 3, 5, and the step size is 1, 1, 1; Step S33: Perform a third layer 3D convolution operation on the enhanced spectral feature tensor, and aggregate the spectral dimension into a two-dimensional feature map based on 3D global average pooling, and then perform deep separable hollow 2D convolution on the two-dimensional feature map to obtain a 2D spatial feature map, wherein the convolution kernel size is 3, 3, 3; Step S4: Perform attention weighting and channel splicing fusion processing on the 3D spectral feature tensor and the 2D spatial feature map to obtain a fused feature representation, wherein step S4 further comprises the following: Fine-tune the fused feature representation based on a pre-training network of a USGS altered mineral spectral library, and freeze the parameters of the bottom convolution layer; Input the fine-tuned fused feature representation into a fully connected classifier to obtain a mineral class probability set; Apply Softmax normalization to the mineral class probability set to obtain a normalized mineral class probability set.

2. The deep learning based satellite hyperspectral alteration mineral remote sensing quantitative identification method according to claim 1, characterized in that, Step S1 comprises the following steps: Step S11: Collect high-spectral original data of GaoFen 5B satellite, corresponding DEM data and atmospheric parameter data to construct an input sample set; Step S12: Divide the high-spectral original data into spatial blocks according to 9*9 pixels, and label the mineral class of the block center pixel to obtain a labeled input sample set; Step S13: Perform radiation correction and absorption correction on the atmospheric parameter data to obtain an atmospheric correction parameter set; Step S14: Calculate the terrain shadow correction coefficient based on the DEM data to obtain a shadow correction coefficient set; Step S15: Apply the atmospheric correction parameter set and the shadow correction coefficient set to the noise-corrected data set to obtain a pre-corrected high-spectral data set.

3. The deep learning based satellite hyperspectral alteration mineral remote sensing quantitative identification method according to claim 1, characterized in that, Step S2 comprises: Perform feature importance evaluation on each band of the pre-corrected high-spectral data set to obtain a band importance score set; Apply adaptive threshold screening based on the band importance score set to select bands with importance higher than a preset threshold to form an initial key band subset; Remove redundant bands from the initial key band subset to obtain a final key band subset.

4. The deep learning based satellite hyperspectral alteration mineral remote sensing quantitative identification method according to claim 1, characterized in that, Step S4 comprises the following steps: Step S41: calculate the spectral attention weight on the 3D spectral feature tensor to obtain a spectral attention weight matrix; Step S42: calculate the spatial attention weight on the 2D spatial feature map to obtain a spatial attention weight map; Step S43: multiply the spectral attention weight matrix and the 3D spectral feature tensor element by element, and multiply the spatial attention weight map and the 2D spatial feature map element by element; Step S44: concatenate the multiplication results in the channel dimension to obtain the fusion feature representation.

5. The deep learning based satellite hyperspectral alteration mineral remote sensing quantitative identification method according to claim 1, characterized in that, Step S41 includes the following steps: Step S411: perform global average pooling on the input 3D spectral feature tensor to obtain a spectral feature vector; Step S412: apply a fully connected layer to the spectral feature vector to calculate an unnormalized spectral attention score; Step S413: normalize the unnormalized spectral attention score to obtain the spectral attention weight matrix.

6. The deep learning based satellite hyperspectral alteration mineral remote sensing quantitative identification method according to claim 1, characterized in that, Step S4 further includes the following: Construct a conditional random field model based on the mineral category probability set to extract adjacent pixel neighborhood features; Use the conditional random field model to correct the probability of isolated high-probability misjudgment pixels to obtain a corrected mineral category probability set; Perform confidence evaluation on the corrected mineral category probability set to obtain a confidence labeled data set.

7. The deep learning-based satellite hyperspectral alteration mineral remote sensing quantitative identification method according to claim 6, characterized in that, Step S4 further includes The following: Generate an altered mineral spatial distribution map according to the corrected mineral category probability set; Extract the boundary of the altered mineral spatial distribution map to obtain the target mineral region contour; Label the target mineral region contour in combination with the confidence labeled data set to obtain the final altered mineral spatial distribution map.

8. The deep learning based satellite hyperspectral alteration mineral remote sensing quantitative identification method according to any one of claims 1-7, characterized in that, Generate a quantitative identification report in a visual form from the fusion feature representation, the mineral category probability set and the altered mineral spatial distribution map.

Images