Satellite hyperspectral altered mineral remote sensing quantitative identification method based on deep learning

Abstract

The invention relates to the technical field of hyperspectral alteration, in particular to a satellite hyperspectral alteration mineral remote sensing quantitative recognition method based on deep learning. The method comprises the following steps: acquiring high-resolution 5B satellite hyperspectral data, DEM data and atmospheric parameter data; performing multi-source noise suppression processing on the high-resolution 5B satellite hyperspectral data to obtain a noise correction data set; performing terrain shadow correction based on the noise correction data set and the DEM data, and performing atmospheric correction in combination with the atmospheric parameter data to obtain a pre-corrected hyperspectral data set; through multi-source data fusion, noise and environment interference correction, key wave band screening and multi-dimensional depth feature extraction and fusion, the quality and characterization capability of hyperspectral data are effectively improved, and a solid data basis is provided for subsequent refined mineral recognition and analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hyperspectral alteration, and particularly relates to a satellite hyperspectral alteration mineral remote sensing quantitative identification method based on deep learning. BACKGROUND

[0002] Hyperspectral alteration mineral refers to a secondary mineral formed after the original mineral in the rock changes in chemical composition or physical properties under geological action (such as hydrothermal activity, weathering, etc.), which has unique spectral characteristics (such as absorption peaks of specific wavelengths, reflectivity mutations, etc.) in hyperspectral data. The traditional hyperspectral alteration mineral identification method mainly relies on the experience of remote sensing geologists to select the typical spectral characteristics of alteration minerals, extracts spectral characteristics through band ratio method, principal component analysis method, spectral angle classification method, etc., sets an empirical threshold to perform threshold segmentation or classification on the extracted characteristics, and finally outputs the identification result of the alteration mineral. The above-mentioned traditional identification method relies on manual selection of spectral characteristics, and it is difficult to cover the complex and diverse spectral details of alteration minerals, and it is easy to miss key features, the feature extraction capability for mixed pixels is weak, the processing flow is cumbersome, manual parameter adjustment is required, the adaptability to different regions and different types of alteration minerals is poor, and it is difficult to process a large amount of data in batches. SUMMARY

[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 technical problems.

[0004] To achieve the above-mentioned purpose, the satellite hyperspectral alteration mineral remote sensing quantitative identification method based on deep learning comprises the following steps:

[0005] Step S1: Obtain hyperspectral data of high-resolution 5B satellite, DEM data and atmospheric parameter data; perform multi-source noise suppression processing on the hyperspectral data of high-resolution 5B satellite to obtain a noise-corrected data set; perform terrain shadow correction based on the noise-corrected data set and the DEM data, and perform atmospheric correction in combination with the atmospheric parameter data to obtain a pre-corrected hyperspectral data set;

[0006] Step S2: Perform adaptive band selection on the pre-corrected hyperspectral data set to obtain a key band subset;

[0007] Step S3: Construct a 3D-2D hybrid convolution network based on the key band subset, and input the pre-corrected hyperspectral data block into the hybrid convolution 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 a fused feature representation.

[0009] The beneficial effects of the present application are that high-spectral data of Gao Fen 5B satellite, digital elevation model (DEM) data and atmospheric parameter data are collected, and multi-source noise suppression processing is performed on the high-spectral data, which significantly reduces sensor noise and environmental interference, and improves the signal-to-noise ratio of the data; then, the DEM data are combined to perform terrain shadow correction, which reduces the spectral signal distortion caused by the terrain undulation, and the atmospheric parameter data are used to perform atmospheric correction, which corrects the influence of atmospheric scattering and absorption on the spectral data, and a more real and reliable pre-corrected high-spectral data set is obtained, and the systematic errors caused by the environment and observation conditions are effectively eliminated through multi-source data fusion and comprehensive correction in this stage. Step S2 performs adaptive band selection based on the pre-corrected data, selects a key band subset reflecting the physical and chemical properties of the target, effectively reduces the dimension of the data, reduces the redundant information and the calculation complexity, and provides more discriminative input features for subsequent feature extraction. Step S3 uses the key band subset to construct a 3D-2D hybrid convolutional neural network, captures the joint features of the high-spectral data in the spectral and spatial dimensions through 3D convolution operation, mines the complex correlation between the spectral channels and the local space, and extracts the pure spatial dimension features through 2D convolution operation, enhances the expression ability of the spatial structure information, and outputs 3D spectral feature tensor and 2D spatial feature map after the network input pre-corrected high-spectral data block, realizing multi-dimensional representation of spectral and spatial features. Finally, step S4 applies an attention weighting mechanism to the above two types of features, dynamically adjusts the feature response weight of each channel and spatial position, suppresses irrelevant noise, highlights important information, and forms a fusion feature representation through channel dimension splicing and fusion, integrates spectral and spatial information, and improves the richness and discriminability of feature expression. The overall process realizes 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 improves the quality and representation ability of the high-spectral data, and provides a solid data foundation for subsequent refined mineral identification and analysis. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 It is a step flowchart of a satellite high-spectral alteration mineral remote sensing quantitative identification method based on deep learning.

[0011] Figure 2 It is Figure 1 It is a detailed implementation step flowchart of step S2.

[0012] Figure 3 It is an alteration mineral spatial distribution map in the visualization icon.

[0013] The implementation of the present application, functional characteristics and advantages will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0014] The technical method of the patent of the present application will be described clearly and completely in combination with the drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0015] In addition, the drawings are only schematic illustrations of the present application and are not necessarily drawn to scale. Identical reference signs in the drawings represent identical or similar parts, and thus repeated descriptions thereof will be omitted. Some of the block diagrams shown in the drawings are functional entities, which do not necessarily have to correspond to physically or logically independent entities. The functional entities can be implemented in the form of software, or 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. can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, without departing from the scope of the exemplary embodiments, a first element can be called a second element, and similarly a second element can be called a first element. 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-mentioned purpose, 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: Obtain high-resolution 5B satellite hyperspectral data, DEM data and atmospheric parameter data; perform multi-source noise suppression processing on the high-resolution 5B satellite hyperspectral data to obtain a noise-corrected data set; perform terrain shadow correction based on the noise-corrected data set and the DEM data, and perform atmospheric correction in combination with the atmospheric parameter data to obtain a pre-corrected hyperspectral data set;

[0019] Step S2: Perform adaptive band selection on the pre-corrected hyperspectral data set to obtain a key band subset;

[0020] Step S3: Construct a 3D-2D hybrid convolution network based on the key band subset, and input the pre-corrected hyperspectral data block into the hybrid convolution 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 a fused feature representation.

[0022] In the embodiment of the present application, referring to Figure 1 As shown in the figure, it is a step flow diagram of the satellite hyperspectral alteration mineral remote sensing quantitative identification method based on deep learning of the present application. In the present example, the satellite hyperspectral alteration mineral remote sensing quantitative identification method based on deep learning comprises the following steps:

[0023] Step S1: Obtain hyperspectral data of high-resolution 5B satellite, DEM data and atmospheric parameter data; perform multi-source noise suppression processing on the hyperspectral data of high-resolution 5B satellite to obtain a noise-corrected data set; perform terrain shadow correction based on the noise-corrected data set and DEM data, and perform atmospheric correction in combination with the atmospheric parameter data to obtain a pre-corrected hyperspectral data set;

[0024] In the embodiment of the present application, the hyperspectral data of high-resolution 5B satellite, corresponding DEM data and atmospheric parameter data are obtained, and the three types of data are registered and resampled on the spatial resolution and projection coordinate system to form a spatial grid consistent original data set; based on the atmospheric parameter data, the original radiation value is subjected to radiation correction and absorption correction, and an atmospheric correction parameter set (including necessary quantities such as aerosol optical thickness, gas absorption coefficient, solar zenith angle, etc.) is calculated and output, and the radiation unit is converted into ground apparent reflectance for subsequent processing; multi-source noise suppression processing is performed on the spectral sequence of each pixel of high-resolution 5B, multi-scale decomposition is performed using Daubechies 4th order wavelet transform, threshold modification is performed on the wavelet coefficients according to the noise statistical characteristics, and the spectrum is reconstructed, so that the noise-corrected data set is obtained, and the known bad track and strong water vapor absorption band are masked, interpolated or spectral segment excluded to eliminate abnormal values introduced by systematic bad track and atmospheric absorption; based on the registered DEM data and image metadata, the pixel slope, slope direction and incidence / emission geometry are calculated, the shadow correction coefficient set (realized by cosine incidence angle correction or C correction with correction factor) is constructed, and the shadow correction coefficient set and the atmospheric correction parameter set are applied to the noise-corrected data set to perform terrain shadow and atmospheric coupling correction, so that the pre-corrected hyperspectral data set is obtained. When necessary, the above atmospheric correction parameter set, shadow correction coefficient set and noise correction data are integrated into an end-to-end reconstruction network for joint learning and reconstruction to output the pre-corrected hyperspectral data set for subsequent adaptive band selection and feature extraction.

[0025] Step S2: Perform adaptive band selection on the pre-corrected hyperspectral data set to obtain a key band subset;

[0026] In the embodiments of the present application, the pre-corrected hyperspectral dataset is represented as a three-dimensional tensor or a sample block set with labeled input samples, a set of band discriminant and quality metrics is calculated for each band to form a band importance score set; specifically including: or first do wave-by-wave normalization (such as standardization according to the median and interquartile range) to offset the differences in reflectance scales of different bands, then estimate the signal-to-noise ratio (SNR) of each band based on the wavelet residual in the noise correction dataset, and obtain the SNR index; for each pixel spectrum, implement continuum removal, calculate the average absorption depth and the maximum absorption depth in the preset diagnostic absorption band window as the spectral absorption response index; calculate the first and second derivative energies of each band to represent the spectral inflection point sensitivity; when there is a labeled mineral class, calculate the mutual information or F-statistic of each band and the label based on the labeled samples to obtain the supervised discriminant score; the above several normalized scores are combined by weighting or principal component to obtain the comprehensive importance score, and the cumulative contribution curve in descending order of score is drawn; for de-redundancy, construct the Pearson correlation matrix between bands and perform hierarchical clustering with 1-|r| as the distance, cut the clusters according to the inflection point of the clustering distance or the set correlation threshold, and select the band with the highest comprehensive importance score in the cluster as the representative band; the threshold is adaptively determined by the inflection point of the cumulative contribution curve or the stability selection under cross-validation (also can use sparse regression / L1 regularization, recursive feature elimination and bootstrap stability selection as alternative strategies), finally output the key band subset (represented by band index or central wavelength), and perform consistency test on the discriminant performance index of the selected key band subset and several classifiers on the cross-validation set to complete the adaptive selection closed loop.

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

[0028] In the embodiment of the present application, the pre-corrected hyperspectral data set is first blockized and batched based on the key band subset from the data level, the pre-corrected hyperspectral data is divided into a pre-corrected hyperspectral data block set with the shape of N*9*9*Bk (where Bk represents the number of key band subset bands) according to the spatial neighborhood, the zero-mean-unit-variance standardization and continuous spectrum line correction are performed on each data block according to the band to eliminate the scale difference and residual baseline drift, and then the standardized data block is sent into the constructed 3D-2D hybrid convolution network in parallel. The hybrid network adopts a double-branch structure: the 3D branch aims to retain the spectral-spatial coupling information, and adopts a 3D convolution unit that progresses layer by layer to realize the joint representation of the local spatial neighborhood and the continuous spectral band; specifically, 3D convolution layers are arranged in sequence (the convolution kernel size is (3, 3, 7), (3, 3, 5), and (3, 3, 3) in sequence, and the corresponding step length can be set as (1, 1, 2), (1, 1, 1), etc. to realize hierarchical down-sampling in the spectral dimension), batch normalization and ReLU activation are connected in parallel after each convolution, and a residual short circuit is introduced as necessary to retain the spectral details of the shallow layer, and finally global average pooling or a learnable spectral projection is performed on the spectral dimension at the top layer of the 3D branch to converge the spectral energy into the channel dimension, thereby forming a 3D spectral feature tensor that retains spectral details and encodes spatial neighborhood information. The 2D branch obtains a single-channel or multi-channel spatial input by performing spectral domain convergence (such as principal component mapping or band mean / weighted sum mapping) on the key band subset, and adopts a depth separable 2D hollow convolution unit (such as a 3*3 kernel, dilation = 2) and a multi-scale pooling module to extract spatial texture and structure features, and outputs an aligned 2D spatial feature map; in order to ensure the compatibility of the features of the two branches, a channel alignment, feature normalization and learnable projection matrix are adopted at the branch interaction of the network. In the training and inference stage, spectral domain / spatial domain random disturbance and data enhancement strategies (spectral jitter, local noise injection, spatial flip, etc.) are applied to the intermediate features to stabilize the feature distribution. The network finally outputs the 3D spectral feature tensor and the 2D spatial feature map in parallel for subsequent processing, which are directly used for the subsequent attention weighting and channel splicing fusion steps.

[0029] Step S4: performing 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.

[0030] In the embodiment of the present application, the 3D spectral feature tensor X s (size N*C s *H*W) and the 2D spatial feature map X sp (size N*C sp *H*W) are subjected to step-by-step transformation of attention weighting and channel splicing: first, the global average pooling is performed on X s along the spatial dimension to generate a channel description vector g s, and the spectral attention weight matrix A is obtained through two layers of full connection (dimension reduction-ReLU-dimension increase) mapping and Sigmoid normalization s (size NxC s x1x1); meanwhile, average pooling and maximum pooling are calculated along the channel dimension respectively, and the spatial attention weight graph A sp (size N x 1 x H x W) is obtained through 3x3 convolution, batch normalization and Sigmoid mapping after channel dimension splicing; then, channel broadcasting is performed on A sp , and element-wise multiplication is performed with X s to complete the spectral channel reweighting; pixel broadcasting is performed on A s , and element-wise multiplication is performed with X sp to complete the spatial pixel reweighting, and the reweighted feature X sp ' s and X ' sp ; to ensure the shape consistency of subsequent splicing, 1x1 convolution is performed on X' s and X' sp to project the channel to a unified channel number C f , and bilinear interpolation or deconvolution up / down sampling is used to align the spatial size of any spatial resolution inconsistent features, and then batch normalization and linear activation are performed on the projection results to output the aligned features X" s and X" sp ; X" s and X s p" are spliced in the channel dimension to obtain X c oncat(size N x 2C f x H x W), and 1x1 convolution is applied to X c oncat to fuse channel information and generate a fusion intermediate representation again through batch normalization and ReLU activation. Optionally, a residual shortcut is used to add the linear mapping before projection to the fusion result to retain the original information, and finally the fusion feature representation X fused is output after normalization and optional dropout processing for subsequent classifier or post-processing module use; in the training process, random disturbance and regularization in the spectral domain and spatial domain are applied to the attention branch and the fusion layer to constrain the numerical stability, and in the inference process, the above step-by-step data transformation is executed according to the same operator flow to obtain consistent fusion feature representation.

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

[0032] Step S11: Collecting high-spectral original data of Gao Fen 5B satellite, corresponding DEM data and atmospheric parameter data, and constructing an input sample set;

[0033] ​Step S12: The hyperspectral raw data is divided into spatial blocks according to 9*9 pixels, and the mineral categories are labeled with the center pixels of the blocks to obtain a labeled input sample set;

[0034] Step S13: Radiance correction and absorption correction are performed on the atmospheric parameter data to obtain an atmospheric correction parameter set;

[0035] Step S14: A terrain shadow correction coefficient set is calculated based on the DEM data;

[0036] Step S15: The atmospheric correction parameter set and the shadow correction coefficient set are applied to the noise correction data set to obtain a pre-corrected hyperspectral data set.

[0037] In the embodiment of the application, hyperspectral raw data of Gao Fen 5B satellite, corresponding DEM data and atmospheric parameter data are collected, and the three types of original data are registered and resampled in projection coordinates, spatial resolution and time stamp to form a spatially consistent input sample set. Based on the input sample set, multi-source noise suppression is performed on the hyperspectral raw data to construct a noise correction data set. Specifically, Daubechies 4-order wavelet decomposition is performed on each pixel spectrum, and thresholding reconstruction is performed on the wavelet coefficients to remove high-frequency noise. At the same time, the known bad channels and strong water vapor absorption spectrum are repaired by using mask, spectrum interpolation or local fitting. After obtaining the noise correction data set, the pixel slope, slope direction and incidence / excitation geometry are calculated based on the DEM data, and the shadow correction coefficient set is fitted by using cosine incidence angle correction or C correction method. At the same time, the aerosol optical thickness, water vapor content and main gas absorption coefficient are estimated by using the radiation transfer model inversion or empirical linear correction based on the atmospheric parameter data to form the atmospheric correction parameter set of the step. Finally, the atmospheric correction parameter set and the shadow correction coefficient set are applied to the noise correction data set pixel by pixel and spectrum by spectrum. The conversion relationship from radiation to apparent reflectance is multiplied and divided by the correction factor, and the mask processing is performed, and the spectrum end interpolation is performed as necessary to obtain the pre-corrected hyperspectral data set for subsequent analysis. In the sample construction link, the pre-corrected hyperspectral data is divided into spatial blocks according to the fixed window of 9*9 pixels, and the lithofacies / mineral label of the center pixels of each block is generated to form a labeled input sample set for subsequent training and verification.

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

[0039] The feature importance of each band of the pre-corrected hyperspectral data set is evaluated to obtain a band importance score set;

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

[0041] The initial key band subset is subjected to redundant band removal to obtain a final key band subset.

[0042] In an embodiment of the present application, the pre-corrected hyperspectral dataset is represented as a three-dimensional data cube with size HxWxB, where H and W are spatial dimensions, and B is the number of bands; for each band b∈{1,…,B}, a corresponding two-dimensional reflectance matrix is extracted, and a feature importance indicator is calculated based on a labeled sample set or statistical distribution characteristics, which can include mutual information between the band and the target mineral class, F-statistics, spectral absorption depth, spectral curvature rate of change, signal-to-noise ratio (SNR), etc. Different indicators are mapped to the same dimensional interval through normalization processing, and finally a band importance score set S={s1,s2,…,sB} is obtained by weighted linear combination or principal component projection. B Subsequently, an adaptive threshold is calculated according to the distribution characteristics of S, which can be determined by inflection point detection of the cumulative contribution rate curve, Otsu method segmentation, or optimal split point based on cross-validation accuracy, and a band set that satisfies s b ≥T adaptive is selected to form an initial key band subset B init . To remove the redundancy between bands, the Pearson correlation coefficient matrix R is calculated for all bands in B init , and the absolute value of R is compared with a preset correlation threshold r th . Bands with high correlation (|R ij |≥r th ) are grouped into the same cluster, and the band with the highest importance score in each cluster is retained as the representative band, and the remaining redundant bands are removed. Finally, the key band subset B final without strong correlation redundancy is output, which is stored in the form of band index or central wavelength as input for subsequent feature extraction and model training.

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

[0044] Step S31: performing a first layer 3D convolution operation on the pre-corrected hyperspectral 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.

[0045] Step S32: performing 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.

[0046] Step S33: performing a third layer 3D convolution operation on the enhanced spectral feature tensor, and aggregating the spectral dimension into a two-dimensional feature map based on 3D global average pooling, and then performing a depth separable dilated 2D convolution on the two-dimensional feature map to obtain a 2D spatial feature map, wherein 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: concatenating the multiplication results in the channel dimension to obtain a fused feature representation.

[0053] In an embodiment of the present application, the input 3D spectral feature tensor represents the composite information of the data in the spatial dimension and the spectral dimension, and the dimension order is usually (spatial height x spatial width x spectral channel). In step S41, the spectral attention weight is calculated for the 3D tensor, which is essentially to quantify the importance of each spectral channel by a specific spectral dimension feature extraction method, and a weight matrix corresponding to the spectral channel dimension is generated. The dimension of the weight matrix is consistent with the number of spectral channels, and reflects the contribution intensity of different spectral channels in the overall feature. Then in step S42, the spatial attention weight is calculated for the 2D spatial feature map, which is usually a two-dimensional projection or a single-channel feature map extracted from the original 3D tensor, and the dimension is (spatial height x spatial width). The calculation process generates a two-dimensional spatial attention weight map by analyzing the information distribution in the spatial dimension, each spatial position corresponds to a weight value, which reflects the feature saliency of the position. Step S43 involves element-wise multiplication of the above two attention weights with the corresponding original feature data, i.e. the spectral attention weight matrix and the 3D spectral feature tensor are broadcasted in the spectral channel dimension to strengthen or suppress the feature response of different spectral channels; at the same time, the spatial attention weight map and the 2D spatial feature map are multiplied element by element to adjust the feature intensity of the spatial region, so as to realize the feature weighting in the spatial dimension. Finally, step S44 concatenates the processed spectral weighted 3D tensor and the spatial weighted 2D feature in the channel dimension to form a fused feature representation, which expands the number of channels in the data dimension and combines the attention enhancement information of the spectrum and the space, providing more rich and reasonably weighted input features for subsequent feature analysis or classification tasks. The whole process focuses on the calculation and application of multi-dimensional attention weight, which finely adjusts the information distribution in different dimensions of the feature tensor through element-wise multiplication, and realizes the effective fusion and expression of multi-modal features.

[0054] As an example of the present application, reference is made to Fig. 4, which shows a flowchart of a method for calculating a fused feature representation according to an embodiment of the present application. In this example, the step S4 comprises: Figure 2

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

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

[0057] Step S413: normalizing the unnormalized spectral attention score to obtain a spectral attention weight matrix.

[0058] ​In an embodiment of the present application, the input 3D spectral feature tensor generally has a dimensional structure of (spatial height x spatial width x spectral channel), where the spatial dimension carries the spatial information of the pixels, and the spectral channel dimension reflects the spectral response of different wavebands. Through a global average pooling operation, the mean value of the three-dimensional tensor along the spatial dimension is calculated, specifically, the average value of all spatial positions in each spectral channel is calculated, and finally a one-dimensional spectral feature vector is obtained, which has a dimension of the number of spectral channels. This vector can be regarded as a compressed representation of the overall information on the spectral dimension, reflecting the global response intensity of each spectral channel. A fully connected layer is applied to the spectral feature vector, that is, a set of trainable weight matrices is used to perform linear transformation on the vector and add a bias to generate a new vector, representing the attention score of the feature vector in different channels on the spectral dimension. This step is essentially to learn the relationship and importance between spectral channels through parameterized mapping, and the output unnormalized attention score still maintains the same dimension as the number of spectral channels. The unnormalized attention score vector is normalized, usually using the Softmax function or the Sigmoid function, to convert the score into a weight coefficient in the interval (0, 1), while ensuring that the weight distribution of all spectral channels satisfies certain probability properties (such as the sum being 1), forming a spectral attention weight matrix. The weight matrix has a dimension corresponding to the number of spectral channels, which is used for subsequent weighted operation on the 3D spectral feature tensor to achieve weighted enhancement or suppression of different spectral channels. Overall, this process realizes the dimension reduction and compression of spectral features through global average aggregation of the spatial dimension, then mines the dependence relationship between channels through fully connected layer mapping, and finally normalizes the weight distribution through the normalization function to complete the calculation of the spectral attention weight, providing a data-driven attention mechanism basis for multi-dimensional feature fusion.

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

[0060] Based on the USGS alteration mineral spectral library, the pre-training network of the fusion feature representation is fine-tuned, and the parameters of the bottom convolutional layer are frozen;

[0061] The fine-tuned fusion feature representation is input into a fully connected classifier to obtain a set of mineral class probabilities;

[0062] The set of mineral class probabilities is subjected to Softmax normalization to obtain a set of normalized mineral class probabilities.

[0063] In the embodiment of the present application, the fused feature representation is obtained through a preliminary feature fusion step, usually in the form of a multi-dimensional feature tensor or vector, containing rich spectral and spatial information. In the fine-tuning stage, the labeled mineral spectral data in the USGS alteration mineral spectral library is used as the training sample, and the fused feature is input into the pre-trained neural network. The pre-trained network contains multiple convolutional layers and several fully connected layers, designed for feature extraction and classification tasks. In the fine-tuning process, the strategy of freezing the parameters of the bottom convolutional layers is adopted, that is, the weight parameters of the convolutional layers are kept unchanged, and only the high-level parameters or fully connected layer parameters of the network backend are adjusted, so as to retain the generalization ability of the pre-trained network to the basic spectral features, while making the network adapt to the fine-grained classification needs of specific mineral data. After fine-tuning, the fused feature passes through the fully connected classifier part of the network, and outputs a set of mineral class probabilities, which correspond to the unnormalized prediction values of each mineral class, reflecting the network's confidence in the input sample belonging to each class. Then, the Softmax function is used to normalize the probability set, converting the original output into a probability distribution between 0 and 1, ensuring that the sum of all class probabilities is 1, thereby forming a normalized mineral class probability set. This normalized result can be used for subsequent mineral identification and decision support. The overall process emphasizes the data transmission and parameter adjustment mechanism in the pre-trained network, realizes the stability and generalization of feature extraction by freezing the bottom convolutional layers, uses the fully connected layers to classify the features, and finally realizes the probabilistic expression of mineral classes through probability normalization, ensuring the scientificity and data consistency of the classification results.

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

[0065] Constructing a conditional random field model based on the mineral class probability set to extract adjacent pixel neighborhood features;

[0066] Using the conditional random field model to correct the probability of isolated high-probability misjudgment pixels to obtain a corrected mineral class probability set;

[0067] Performing confidence evaluation on the corrected mineral class probability set to obtain a confidence-labeled data set.

[0068] In the embodiment of the present application, the mineral category probability set is taken as input data, representing the probability distribution of each pixel corresponding to each mineral category, having a two-dimensional spatial structure, and mapping to the spatial position of the remote sensing image or spectral map. First, a conditional random field model is constructed based on the probability set. The CRF is a discriminative probabilistic graph model, which defines the pixel nodes and their edge connections by using the spatial relationship and probability information between adjacent pixels, forming a graph structure with potential dependencies. The model describes the mutual influence between the class labels of adjacent pixels through joint probability distribution, and then integrates the spatial context features to effectively capture the spatial consistency of mineral categories in the neighborhood. Second, the CRF model is used to correct the probability of isolated high-probability misjudgment pixels. Specifically, the model adjusts the class prediction probability of the pixel by maximizing the conditional probability distribution, considering the joint influence of the class probability of the pixel itself and the class probability of the neighborhood pixels, reduces the misjudgment risk of isolated outliers, and realizes probability smoothing and spatial consistency optimization. The correction process generates a new mineral category probability set, and the spatial distribution is more consistent with the geological continuity and the actual mineral distribution characteristics. Finally, the confidence of the corrected probability set is evaluated, and the maximum class probability or other statistical indicators of each pixel are calculated to generate a confidence labeled data set. The data set provides a basis for subsequent accuracy analysis, classification result verification or decision support. The overall process embodies the spatial structure modeling and optimization capability of the spatial statistical model on the probability data. Through the neighborhood dependency mechanism of CRF, the spatial correction and confidence quantification of the mineral category probability are realized, thereby improving the spatial consistency and reliability of the classification.

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

[0070] According to the corrected mineral category probability set, an altered mineral spatial distribution map is generated.

[0071] The boundary of the altered mineral spatial distribution map is extracted to obtain a target mineral region contour.

[0072] The target mineral region contour is labeled in combination with the confidence labeled data set to obtain a final altered mineral spatial distribution map.

[0073] In the embodiments of the present application, the revised mineral class probability set is taken as input, which is presented in the form of a two-dimensional matrix, and each element in the matrix corresponds to a spatial pixel and its corresponding mineral class probability. Based on this probability set, the spatial distribution map of altered minerals is generated by selecting the class label with a higher probability value or applying the maximum probability principle. The map is expressed in the form of two-dimensional raster data, indicating that each spatial pixel belongs to a certain mineral class, and the preliminary mapping of mineral spatial distribution is realized. Then, a boundary extraction operation is performed on the spatial distribution map. Boundary extraction involves spatial structure analysis of continuous regions of the same class in a two-dimensional grid, and contour detection algorithms commonly used in image processing, such as edge detection based on gradient changes or region growing algorithms, are used to identify the spatial boundaries of different mineral class blocks to obtain the contour line data of the target mineral region. These contour lines exist in the form of pixel or spatial coordinate point sequences, reflecting the spatial boundary characteristics of mineral spatial distribution. Finally, the target mineral region contour is labeled in combination with the confidence level data set. The confidence level data set contains confidence level values corresponding to spatial positions, which are used to represent the classification accuracy and reliability of each pixel or region. In the contour labeling process, the confidence level information is superimposed on the boundaries or the interior of the spatial distribution map, and regions with different confidence levels are classified and labeled to form the final altered mineral spatial distribution map with confidence level labeling. This process not only realizes the clear demarcation of the spatial range of minerals, but also provides quality evaluation of spatial distribution by combining the confidence index, providing data basis for subsequent geological interpretation and decision analysis. The overall process converts probability data into spatial classification map, then extracts spatial structure through boundary recognition, and finally fuses confidence level information to build a complete spatial mineral distribution expression system.

[0074] As shown in Figure 3 , it represents the spatial distribution map of altered minerals in the visualization icon.

[0075] Especially important is that the fusion feature representation, mineral class probability set and altered mineral spatial distribution map are generated in a quantitative identification report in a visual form.

[0076] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the application file are intended to be included in the present application.

[0077] The foregoing is considered as illustrative only of the principles of the application. Numerous modifications and changes will readily occur to those skilled in the art, and it is intended to embrace all such modifications and changes that fall within the scope of the application. Accordingly, the application is not to be restricted in scope to the specific embodiments disclosed herein but is to be accorded the full scope that the principles and novel features request appropriately granted.

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: obtaining hyperspectral data of GaoFen 5B satellite, DEM data and atmospheric parameter data; Step S2: performing adaptive band selection on the pre-corrected hyperspectral data set to obtain a key band subset; Step S3: constructing a 3D-2D hybrid convolution network based on the key band subset, and inputting the pre-corrected hyperspectral data block into the hybrid convolution network to obtain a 3D spectral feature tensor and a 2D spatial feature map; Step S4: performing 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. Step S1 comprises the following steps: Step S11: collecting hyperspectral original data of GaoFen 5B satellite, corresponding DEM data and atmospheric parameter data, and constructing an input sample set; 2. The deep learning based satellite hyperspectral alteration mineral remote sensing quantitative identification method according to claim 1, characterized in that, Step S12: dividing the hyperspectral original data into spatial blocks according to 9*9 pixels, and labeling the mineral categories with the block center pixels to obtain a labeled input sample set; Step S13: performing radiation correction and absorption correction on the atmospheric parameter data to obtain an atmospheric correction parameter set; Step S14: calculating a terrain shadow correction coefficient based on the DEM data to obtain a shadow correction coefficient set; Step S15: applying the atmospheric correction parameter set and the shadow correction coefficient set to the noise-corrected data set to obtain a pre-corrected hyperspectral data set. Step S2 comprises: Step S21: performing feature importance evaluation on each band of the pre-corrected hyperspectral data set to obtain a band importance score set; 3. The deep learning based satellite hyperspectral alteration mineral remote sensing quantitative identification method according to claim 1, characterized in that, Step S22: applying an 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; Step S23: removing redundant bands from the initial key band subset to obtain a final key band subset. Step S3 comprises the following steps: Step S31: performing a first layer 3D convolution operation on the pre-corrected hyperspectral 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; 4. The deep learning based satellite hyperspectral alteration mineral remote sensing quantitative identification method according to claim 1, characterized in that, Step S32: performing 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: performing a third layer 3D convolution operation on the enhanced spectral feature tensor, and aggregating the spectral dimension into a two-dimensional feature map based on 3D global average pooling, and then performing a depth 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 comprises the following steps: Step S41: calculating spectral attention weights for the 3D spectral feature tensor to obtain a spectral attention weight matrix; 5. The deep learning based satellite hyperspectral alteration mineral remote sensing quantitative identification method according to claim 1, characterized in that, Step S42: calculating spatial attention weights for the 2D spatial feature map to obtain a spatial attention weight map; Step S43: multiplying the spectral attention weight matrix and the 3D spectral feature tensor element by element, and multiplying the spatial attention weight map and the 2D spatial feature map element by element. ​ ​ Step S44: concatenating the multiplication result along the channel dimension to obtain the fusion feature representation.

6. 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: performing global average pooling on the input 3D spectral feature tensor to obtain a spectral feature vector; Step S412: applying a fully connected layer to the spectral feature vector to calculate an unnormalized spectral attention score; Step S413: normalizing the unnormalized spectral attention score to obtain a spectral attention weight matrix.

7. The deep learning based satellite hyperspectral alteration mineral remote sensing quantitative identification method according to claim 1, characterized in that, Step S4 also includes the following: Fine-tuning the fusion feature representation based on the USGS alteration mineral spectral library, and freezing the parameters of the underlying convolutional layer; Inputting the fine-tuned fusion feature representation into a fully connected classifier to obtain a set of mineral class probabilities; Applying Softmax normalization to the set of mineral class probabilities to obtain a set of normalized mineral class probabilities.

8. The deep learning based satellite hyperspectral alteration mineral remote sensing quantitative identification method according to claim 7, characterized in that, Step S4 also includes the following: Based on the set of mineral class probabilities, a conditional random field model is constructed to extract the neighborhood features of adjacent pixels; Using the conditional random field model to correct the probability of isolated high-probability misjudgment pixels to obtain a set of corrected mineral class probabilities; Performing confidence evaluation on the set of corrected mineral class probabilities to obtain a confidence-labeled data set.

9. The deep learning-based satellite hyperspectral alteration mineral remote sensing quantitative identification method according to claim 8, characterized in that, Step S4 also includes The following: Generating an alteration mineral spatial distribution map according to the set of corrected mineral class probabilities; Extracting the boundary of the target mineral region from the alteration mineral spatial distribution map to obtain a target mineral region contour; Labeling the target mineral region contour in combination with the confidence-labeled data set to obtain a final alteration mineral spatial distribution map.

10. The deep learning based satellite hyperspectral alteration mineral remote sensing quantitative identification method according to any one of claims 1-9, characterized in that, Generating a quantitative identification report in a visual form from the fusion feature representation, the set of mineral class probabilities, and the alteration mineral spatial distribution map.

Images