Method for classifying strata and lithology in semi-arid shallow coverage area by fusing multi-source and multi-temporal remote sensing

Abstract

The application discloses a kind of fusion multi-source multi-temporal remote sensing's semi-arid shallow covering area stratum lithology classification method, collects multi-period multi-spectral remote sensing image and polarization SAR data and pre-processes, calculates optical and radar vegetation index, normalizes and constructs feature dataset;Combined with field investigation label sample, construct pixel-level sample set, after inputting multi-scale time series attention network training, output GeoTIFF format classification result graph. Overall realization is based on the stratum lithology classification of semi-arid shallow covering area of multi-source multi-temporal remote sensing data, breaks through the application limit of existing remote sensing lithology classification method in semi-arid shallow covering area, gets rid of the dependence of many existing deep learning classification algorithms on regional level image sample, overcomes the training data acquisition bottleneck caused by stratum lithology unit boundary blur in semi-arid shallow covering area, improves the reliability of lithology classification under complex environmental conditions in semi-arid shallow covering area.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing information extraction and digital image processing technology, specifically relating to a method for classifying stratigraphic lithology in semi-arid shallow-covered areas by integrating multi-source and multi-temporal remote sensing. Background Technology

[0002] Semi-arid, shallowly covered areas are widely distributed across the Inner Mongolian Plateau, the Loess Plateau, and the Qinghai-Tibet Plateau in my country. These regions hold significant ecological importance and contain strategic resources. Accurately understanding their stratigraphic and lithological distribution is crucial for resource exploration, engineering construction, and environmental protection. Traditional geological survey methods face significant challenges in these areas. The surface is often covered by sparse to medium-density vegetation and aeolian deposits, resulting in poor bedrock exposure, making direct observation difficult, and field surveys are costly.

[0003] Remote sensing technology provides an effective means for large-scale geological surveys. However, existing lithology identification methods based on optical remote sensing have limited effectiveness in semi-arid shallow-covered areas, mainly due to: (1) the obscuring of surface spectral information by vegetation canopy; (2) the burial of lithological characteristics by weathered deposits; and (3) the phenomenon of "different spectra for the same material and the same spectrum for different materials" caused by changes in environmental factors.

[0004] It is worth noting that vegetation growth in semi-arid regions is closely related to the underlying lithology, and differences in soil properties controlled by different lithologies can be indirectly reflected through vegetation phenology. Current technologies often treat vegetation as interference and remove it, failing to fully utilize its value as an indirect indicator of lithology. Furthermore, using only a single temporal phase or a single type of remote sensing data (such as optical data alone) makes it difficult to comprehensively capture the response characteristics of lithology across multiple temporal phases and multiple sensors. Summary of the Invention

[0005] To address the aforementioned problems, the purpose of this invention is to provide a lithological classification method for semi-arid shallow-covered areas that integrates multi-source and multi-temporal remote sensing, thereby overcoming the limitations of traditional remote sensing lithological classification methods in semi-arid shallow-covered areas and the unique lithological identification obstacles in these areas, such as vegetation shading and cover layer interference.

[0006] To achieve the above objectives, the technical solution adopted by the present invention includes,

[0007] A method for stratigraphic and lithological classification of semi-arid shallow-cover areas integrating multi-source and multi-temporal remote sensing includes the following steps: S1. Collect time-series multispectral remote sensing images and time-series polarimetric SAR data of the target semi-arid shallow-covered area; perform the first preprocessing on the time-series multispectral remote sensing images to obtain surface reflectance images of different spectral bands in multiple periods; perform the second preprocessing on the time-series polarimetric SAR data to obtain SAR backscatter images of different polarization modes in multiple periods. S2 is used to calculate multiple optical vegetation index images from surface reflectance images of different spectral bands obtained from S1; and multiple radar vegetation index images are calculated from SAR backscatter images of different polarization modes obtained from S1. S3, normalize the surface reflectance images and SAR backscatter images with different polarization modes obtained from S1, as well as the optical vegetation index images and radar vegetation index images obtained from S2, to obtain the lithological classification feature dataset. S4. Acquire regional geological data and / or high-resolution remote sensing images for field investigation, label sample areas of different lithological types, and obtain vector sample label files containing lithological category labels for each lithological type; S5: Using the vector sample annotation file obtained in S4, extract the temporal feature data corresponding to all pixels in the sample area of ​​the lithology classification feature dataset obtained in S3, and associate it with the category label of the corresponding pixel to obtain the sample dataset; the sample dataset contains feature data and category label. S6. Construct a multi-scale temporal attention network, using the sample dataset obtained in S5 as input, and train the multi-scale temporal attention network with the cross-entropy loss function to obtain the lithology classification model. S7. Input the lithological classification feature dataset obtained in S3 into the obtained lithological classification model to perform lithological classification of the entire region and obtain the lithological classification result map.

[0008] Preferably, the multi-scale temporal attention network includes a multi-scale temporal convolution module, a temporal attention module, a feature channel attention module, a cross-dimensional interaction module, and a classifier. The training process of the multi-scale temporal attention network specifically includes: S6.1, divide the sample dataset into multiple batches, each batch contains B samples, each sample includes multiple time-series feature data and its corresponding category label; organize the time-series feature data of all samples in each batch into a time-series feature tensor, the time-series feature tensor includes time dimension and feature dimension; S6.2, Input the temporal feature tensor obtained in S6.1 into the multi-scale temporal attention network and perform the following operations: S6.2a, map the feature dimension of the temporal feature tensor to the hidden dimension to obtain the feature embedding representation. X embedded S6.2b, embed the features obtained in S6.2a into the representation. X embedded Inputting the multi-scale temporal convolution module extracts local temporal features at different time dimensions, resulting in multi-scale temporal features. X mutiscale S6.2c, embed the features obtained in S6.2a into the representation.X embedded The input temporal attention module captures the global dependencies between time steps, resulting in temporal attention features. X temporal S6.2d, embed the features obtained in S6.2a into the representation. X embedded Input the feature channel attention module, evaluate and enhance the importance of different feature channels, and obtain the feature channel attention features. X feature S6.2e, the multi-scale temporal features obtained in S6.2b. X mutiscale Temporal attention features obtained from S6.2c X temporal The feature channel attention features obtained from S6.2d X feature The input is fused and interacted with by the cross-dimensional interaction module to obtain the features after cross-dimensional interaction. X fusion S6.2f, the cross-dimensional interactive features obtained in S6.2e. X fusion Input the classifier and output the predicted probability distribution of each sample belonging to each stratigraphic lithology category; S6.3, based on the predicted probability distribution obtained in S6.2f and the corresponding class labels of the samples, calculate the cross-entropy loss value, and update the parameters in the multi-scale temporal attention network using the cross-entropy loss value; S6.4, repeat steps S6.1 to S6.3 until convergence, to obtain the lithology classification model.

[0009] Preferably, in S6.2b, the multi-scale temporal convolution module performs the following operations: S6.2b-1, embedding features into representations X embedded Transpose the dimension of the feature representation to obtain the transposed feature representation. X transposed ; S6.2b-2, construct K parallel one-dimensional convolutional layers, each with a different kernel size, where K is an integer greater than 1; S6.2b-3, uses K parallel one-dimensional convolutional layers to process the transposed feature representation. X transposed Perform convolution operations, and apply non-linear activation and global pooling to the output of each one-dimensional convolutional layer to obtain a global feature representation; S6.2b-4 concatenates the K global feature representations along the feature channel dimension to obtain multi-scale temporal features. X mutiscale.

[0010] Preferably, K=4, and the kernel sizes of each one-dimensional convolutional layer are 1, 3, 5 and 7 respectively.

[0011] Preferably, the timing attention module in S6.2c is a Transformer encoder.

[0012] Preferably, the feature channel attention module in S6.2d obtains channel attention weights through a compression-activation operation, and then integrates the channel attention weights with the feature embedding representation. X embedded By multiplying each term, a weighted channel feature representation is obtained. Then, the average of these weighted channel feature representations over time is taken to obtain the feature channel attention features. X feature .

[0013] Preferably, the cross-dimensional interaction module in S6.2e performs the following operations: S6.2e-1, multi-scale temporal features X mutiscale Temporal attention characteristics X temporal and feature channel attention features X feature The features are concatenated along the feature dimension to obtain the concatenated features. X cat ; S6.2e-2, regarding splicing features X cat Perform dimensionality reduction to obtain the reduced features; S6.2e-3 uses a multi-head self-attention mechanism to perform cross-dimensional feature interaction on the dimensionality-reduced features to obtain a self-attention output. S6.2e-4 performs a residual connection between the self-attention output and the dimensionality-reduced features to obtain the cross-dimensional interactive features. X fusion .

[0014] Preferably, the first preprocessing in S1 specifically involves performing radiometric calibration, atmospheric correction, orthorectification, projection conversion, mosaicking, cropping, and resampling on the temporal multispectral remote sensing image in sequence; the second preprocessing specifically involves performing radiometric correction, multi-view processing, filtering, terrain correction, geocoding, mosaicking, cropping, and resampling on the temporal polarimetric SAR data in sequence.

[0015] Preferably, the optical vegetation indices in the multi-phase optical vegetation index images in S2 include at least one of the following: Atmospheric Impedance Vegetation Index (ARVI), Normalized Difference Vegetation Index (NDVI), Ratio Vegetation Index (RVI), Enhanced Vegetation Index (EVI), Difference Vegetation Index (DVI), Vertical Vegetation Index (PVI), Modified Soil-Adjusted Vegetation Index (MSAVI), Green Light Normalized Difference Vegetation Index (GNDVI), and Transformed Soil-Adjusted Vegetation Index (TSAVI); the radar vegetation indices in the multi-phase radar vegetation index images include at least one of the following: Classical Radar Vegetation Index (RVI), Dual-Polarized Radar Vegetation Index (DpRVI), Vegetation Anisotropy Index (VAI), Polarimetric Difference Index (PDI), and Radar Polarimetric Scattering Index (RPSI).

[0016] Preferably, the field survey in S4 specifically includes: setting up survey points by combining regional geological data and / or high-resolution remote sensing images to ensure that there are no fewer than the preset number of survey points in each lithological unit to be classified; and creating vector surface files for annotation by combining the survey results with remote sensing image interpretation through a geographic information system.

[0017] Compared with the prior art, the advantages of the present invention are: (1) This invention provides a stratigraphic lithology classification method based on the special geographical and geological environmental conditions of semi-arid shallow-covered areas, which is based on multi-source and multi-temporal remote sensing data. It breaks through the application limitations of existing lithology classification methods that only use single-temporal or optical remote sensing in semi-arid shallow-covered areas, helps to reduce the workload of field geological surveys in semi-arid shallow-covered areas, and has strong practicality and promotion value.

[0018] (2) The present invention adopts a pixel-level sample training strategy, directly training the model based on a limited number of pixel-level samples, breaking the dependence of many existing deep learning classification methods on regional image samples, and effectively overcoming the bottleneck of training data acquisition caused by the blurred boundaries of lithological units in semi-arid shallow-covered areas and the difficulty of regional sample labeling.

[0019] (3) The deep learning framework constructed by the present invention, which combines multi-scale temporal convolution and attention mechanism, can adaptively capture the response differences of different strata lithology and their coverings in multi-source and multi-temporal remote sensing data, and realize intelligent identification and enhancement of relevant response differences, thereby improving the identification accuracy of strata lithology under complex surface cover conditions in semi-arid shallow cover areas. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof. In the drawings, Figure 1This is a flowchart of the multi-source, multi-temporal remote sensing method for stratigraphic lithology classification in semi-arid, shallowly covered areas according to the present invention. Figure 2 This is a diagram of the architecture of the Multi-Scale Temporal Attention Network (MSTA-NET). Figure 3 The image is the MSTA-NET lithological classification result based on lunar multispectral and polarimetric SAR data for the year 2024 (image-classification result). Figure 4 Image-classification results (SVM) based on multispectral and polarimetric SAR from July 2024. Figure 5 Legend for the lithological classification results. Detailed Implementation

[0021] The invention is not limited to the specific embodiments described below. All equivalent modifications made based on the technical solutions of this application fall within the protection scope of this invention. Unless otherwise specified, all components and devices in this invention utilize components and devices known in the prior art.

[0022] The multispectral images and polarimetric SAR images in this application can be multispectral or polarimetric SAR images from different platforms and different sensors, as well as combined images from different sensors and different bands.

[0023] Example This embodiment takes Duolun County in Inner Mongolia Autonomous Region as the target area. This area is a typical semi-arid shallow cover area with sparse to medium density vegetation, large areas covered by weathered layer, residual slope deposits and sand, and few rock outcrops. At the same time, it contains strategic resources and has an important ecological status, which is in line with the applicable scenario of this method.

[0024] This embodiment presents a method for stratigraphic lithology classification in semi-arid shallow-cover areas that integrates multi-source and multi-temporal remote sensing, comprising the following steps: S1, collect time-series multispectral remote sensing images and time-series polarimetric SAR data of the target semi-arid shallow-covered area; The time-series multispectral remote sensing image is subjected to a first preprocessing to obtain surface reflectance images of different spectral bands in multiple periods; the first preprocessing specifically involves performing radiometric calibration, atmospheric correction, orthorectification, projection conversion, mosaicking, cropping, and resampling on the time-series multispectral remote sensing image in sequence. The time-series polarimetric SAR data is subjected to a second preprocessing to obtain SAR backscatter images with different polarization modes in multiple periods. The second preprocessing specifically involves performing radiometric correction, multi-look processing, filtering, terrain correction, geocoding, stitching, cropping, and resampling on the time-series polarimetric SAR data in sequence.

[0025] This embodiment acquires 12 monthly Sentinel-2 multispectral remote sensing images (covering 10 spectral bands) and Sentinel-1 dual-polarization (VH, VV) SAR data of the target area in 2024 to ensure that the temporal and coordinate information of the two types of data is consistent and covers the complete target area.

[0026] First preprocessing (multispectral imagery): Radiometric calibration, atmospheric correction, and orthorectification were performed sequentially on the time-series Sentinel-2 images. The projection was then transformed to the WGS-84 coordinate system (UTM projection). After stitching and cropping were completed in each period, the images were resampled to a resolution of 10m, ultimately obtaining surface reflectance images of 12 different spectral bands.

[0027] Second preprocessing (polarimetric SAR data): The time-series Sentinel-1 data is sequentially subjected to radiometric correction, multi-look processing, enhanced Lee filtering, terrain correction, and geocoding. After stitching and cropping, it is resampled to 10m resolution to obtain SAR backscatter images in VH and VV polarization modes for 12 periods.

[0028] S2, from the surface reflectance images of different spectral bands obtained in S1, calculate the multi-period optical vegetation index images; the optical vegetation indices in the multi-period optical vegetation index images include at least one of the following: atmospheric impedance vegetation index ARVI, normalized difference vegetation index NDVI, ratio vegetation index RVI, enhanced vegetation index EVI, difference vegetation index DVI, vertical vegetation index PVI, improved soil-modified vegetation index MSAVI, green light normalized difference vegetation index GNDVI, and transformed soil-modified vegetation index TSAVI. Based on the surface reflectance image obtained in S1, this embodiment calculates 10 time-series optical vegetation indices, including atmospheric impedance vegetation index (ARVI), normalized difference vegetation index (NDVI), ratio vegetation index (RVI), enhanced vegetation index (EVI), difference vegetation index (DVI), vertical vegetation index (PVI), improved soil-regulated vegetation index (MSAVI), green light normalized difference vegetation index (GNDVI), transformed soil-regulated vegetation index (TSAVI), and normalized water index (NDWI), forming 12 periods of optical vegetation index images.

[0029] For the SAR backscatter images with different polarization modes obtained from S1, multi-phase radar vegetation index images are calculated; the radar vegetation index in the multi-phase radar vegetation index images includes at least one of the following: classic radar vegetation index RVI, dual-polarization radar vegetation index DpRVI, vegetation anisotropy index VAI, polarization difference index PDI, and radar polarization scattering index RPSI. This embodiment calculates three time-series radar vegetation indices based on the SAR backscatter image obtained from S1, including the dual-polarization radar vegetation index DpRVI, the vegetation anisotropy index VAI, and the difference / sum index DSR, generating 12 radar vegetation index images.

[0030] S3 is used to normalize the surface reflectance images and SAR backscatter images with different polarization modes obtained from S1, as well as the optical vegetation index images and radar vegetation index images obtained from S2, to obtain the lithological classification feature dataset.

[0031] In this embodiment, the surface reflectance images of 10 spectral bands and SAR backscatter images of 2 polarization modes obtained in S1, as well as the images of 10 optical vegetation indices and 3 radar vegetation indices obtained in S2, are normalized to 0-1 to ensure that all feature data are of the same magnitude. Finally, a 12-phase × 25-dimensional (10+2+10+3) lithological classification feature dataset is constructed, providing comprehensive feature support for subsequent model training.

[0032] S4. Acquire regional geological data and / or high-resolution remote sensing images for field surveys, label sample areas of different lithological types, and obtain vector sample labeling files containing lithological category labels for each lithological type; the field survey specifically includes: setting up survey points based on regional geological data and / or high-resolution remote sensing images to ensure that there are no fewer than a preset number of survey points within each lithological unit to be classified; and creating vector surface files for labeling using a geographic information system based on the survey results and remote sensing image interpretation.

[0033] The following is an example of how field surveys and sample labeling are conducted in this embodiment: Survey site layout: Based on existing regional geological data and high-resolution satellite imagery, the target area was determined to include 15 types of stratigraphic lithology and 3 types of typical land features (a total of 18 types). 276 survey sites were set up indoors to ensure that there were no fewer than 10 survey sites in each type of lithology unit to meet the sample representativeness requirements.

[0034] Sample area labeling: Created in ArcGIS by combining field survey results with visual interpretation of remote sensing imagery. The .shp format vector surface file labels sample areas of different lithologies and land features block by block. Each category can contain multiple vector surface samples. Add a Label field as a category label (numbered by Arabic numerals 1-18) to form a vector sample label file.

[0035] S5: Using the vector sample annotation file obtained in S4, extract the temporal feature data corresponding to all pixels in the sample area of ​​the lithological classification feature dataset obtained in S3, and associate it with the category label of the corresponding pixel to obtain the sample dataset; the sample dataset includes feature data part and category label part.

[0036] This embodiment is based on the vector sample annotation file of S4 and the normalized feature dataset of S3. It extracts the temporal feature data pixel values ​​corresponding to all pixels within the sample region according to category, and associates them with the corresponding category labels to construct a tabular sample dataset. To ensure category balance, 1000 pixel-level samples are randomly selected for each category, for a total of 18000 samples. Each row of the dataset corresponds to one sample pixel. The feature columns are arranged in order from 1 to 12 periods, with 25 features per period (spectral reflectance → SAR backscattering → optical vegetation index → ​​radar vegetation index). The last column is the category label. The dataset has 18000 rows × (12 × 25 + 1) columns.

[0037] S6. Construct a multi-scale temporal attention network (MSTA-NET), take the sample dataset obtained in S5 as input, and train the multi-scale temporal attention network with the cross-entropy loss function to obtain the lithology classification model. like Figure 2 As shown, the multi-scale temporal attention network includes a multi-scale temporal convolution module, a temporal attention module, a feature channel attention module, a cross-dimensional interaction module, and a classifier. The training process of the multi-scale temporal attention network specifically includes: S6.1, the sample dataset is divided into multiple batches, each batch contains B=32 samples, each sample includes multiple time-series feature data and its corresponding category label; the time-series feature data of all samples in each batch are organized into a time-series feature tensor, the time-series feature tensor includes a time dimension and a feature dimension; In this embodiment, the temporal feature data of each sample is organized as a three-dimensional tensor (B × time dimension × feature dimension) of (32, 12, 25). S6.2, Input the temporal feature tensor obtained in S6.1 into the multi-scale temporal attention network, and perform the following operations: S6.2a, map the feature dimension of the temporal feature tensor to the hidden dimension to obtain the feature embedding representation. X embedded In this embodiment, a linear transformation layer is used to map the feature dimension 25 to the hidden dimension 256, resulting in a feature embedding representation. X embedded ∈ 32×12×256 .

[0038] S6.2b, embed the features obtained in S6.2a into the representation.X embedded Inputting the multi-scale temporal convolution module extracts local temporal features at different time dimensions, resulting in multi-scale temporal features. X mutiscale In S6.2b, the multi-scale temporal convolution module specifically performs the following operations: S6.2b-1, embedding the feature representation X embedded Transpose the dimension of the feature representation to obtain the transposed feature representation. X transposed ; S6.2b-2, construct two parallel one-dimensional convolutional layers, each with a different kernel size. In this embodiment, the kernel sizes of the one-dimensional convolutional layers are 1, 3, 5, and 7, and the convolutional layers output 64 channels. S6.2b-3, the transposed feature representation is processed by four parallel one-dimensional convolutional layers respectively. X transposed Perform convolution operations, and apply non-linear activation and global pooling to the output of each one-dimensional convolutional layer to obtain a global feature representation; S6.2b-4 concatenates the four global feature representations along the feature channel dimension to obtain multi-scale temporal features. X mutiscale ∈ 32×256 .

[0039] S6.2c, embed the features obtained in S6.2a into the representation. X embedded The input temporal attention module captures the global dependencies between time steps, calculates and normalizes attention weights through linear projection, and then averages the weighted values ​​over the time dimension to obtain the temporal attention features. X temporal ∈ 32×256 ; The timing attention module in this embodiment is a 2-layer Transformer encoder.

[0040] S6.2d, embed the features obtained in S6.2a into the representation. X embedded Input the feature channel attention module, evaluate and enhance the importance of different feature channels, and obtain the feature channel attention features. X feature . The feature channel attention module uses a compression-activation operation to process the feature embedding representation. X embeddedGlobal average pooling compression is performed, followed by a bottleneck structure of "256→128→256" fully connected layers. Channel attention weights are generated via a Sigmoid function, and these channel attention weights are then embedded into the feature representation. X embedded By multiplying each term, a weighted channel feature representation is obtained. Then, the average of these weighted channel feature representations over time is taken to obtain the feature channel attention features. X feature ∈ 32×256 .

[0041] S6.2e, the multi-scale temporal features obtained in S6.2b. X mutiscale Temporal attention features obtained from S6.2c X temporal The feature channel attention features obtained from S6.2d X feature The input is fused and interacted with by the cross-dimensional interaction module to obtain the features after cross-dimensional interaction. X fusion ; The cross-dimensional interaction module performs the following operations: S6.2e-1, the multi-scale temporal features X mutiscale The temporal attention features X temporal and the feature channel attention features X feature The features are concatenated along the feature dimension to obtain the concatenated features. X cat ∈R 32×256 ; S6.2e-2, regarding the splicing feature X cat The dimensions are reduced to 512, batch normalized, and ReLU activated to obtain the reduced features. S6.2e-3, The dimensionality-reduced features are subjected to cross-dimensional feature interaction through a multi-head self-attention mechanism to obtain self-attention output; S6.2e-4, perform a residual connection between the self-attention output and the dimensionality-reduced features to obtain the cross-dimensional interactive features. X fusion ∈ 32×256 ; S6.2f, the cross-dimensional interactive features obtained in S6.2e. X fusionThe input classifier, in this embodiment, consists of two fully connected layers (512→256→128) and an output layer (128→18). Each layer is followed by batch normalization, ReLU activation and Dropout operations, and the output is the predicted probability distribution of each sample belonging to each stratigraphic lithology category. S6.3, Based on the predicted probability distribution obtained in S6.2f and the category label corresponding to the sample, calculate the cross-entropy loss value, and update the parameters in the multi-scale temporal attention network using the cross-entropy loss value; S6.4, repeat steps S6.1 to S6.3 until convergence, to obtain the lithology classification model.

[0042] S7. Input the lithological classification feature dataset obtained in S3 into the obtained lithological classification model to perform lithological classification of the entire region and obtain the lithological classification result map.

[0043] In this embodiment, the full-region feature dataset (3674×2031×12×25 dimensions) constructed by S3 is reshaped into a two-dimensional array of (3674×2031)×(12×25). A batch prediction strategy is used to input the dataset into the trained model to avoid memory overflow. All batch prediction results are collected, reorganized into a two-dimensional classification matrix according to the original spatial location, written into a GeoTIFF format file, and the complete geographic coordinate parameters (coordinate system, transformation matrix, etc.) are preserved. The final output is the lithology classification result map of the target area, as shown in the figure. Figure 3 As shown; Through experiments, Figure 3 The results of this invention show that the overall classification accuracy is high, with the overall accuracy of the validation set reaching 94.9%. For comparison, Figure 4 This image shows the lithological classification results for the same region obtained using only the July 2024 dataset and the SVM classification algorithm. The overall classification accuracy is only 89.7%. Figure 3 , Figure 4 A comparison of the stratigraphic lithology classification results across the entire region shows that... Figure 3 Compared with the stratigraphic lithology classification results Figure 4 The results shown are more accurate, and there is less fragmentation in the classification results, demonstrating that the method of this invention has improved the accuracy of the classification. Figure 3 The classification results shown are better than Figure 4 The SVM classification effect for a single period of data is shown.

[0044] The embodiments show that the content of this invention has good applicability in remote sensing lithology mapping of typical semi-arid shallow-covered areas in Duolun County, Inner Mongolia.

[0045] This embodiment also discloses a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the lithological classification method for semi-arid shallow-covered areas that integrates multi-source and multi-temporal remote sensing data of this embodiment.

[0046] This embodiment also discloses a computer program product, including a computer program / instruction, which, when executed by a processor, implements the lithological classification method for semi-arid shallow-covered areas that integrates multi-source and multi-temporal remote sensing data of this embodiment.

[0047] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0048] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0049] Furthermore, the various implementation methods disclosed in this solution can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content invented by this disclosure.

Claims

1. A method for stratigraphic and lithological classification of semi-arid shallow-cover areas integrating multi-source and multi-temporal remote sensing, characterized in that, Includes the following steps: S1, collect time-series multispectral remote sensing images and time-series polarimetric SAR data of the target semi-arid shallow-covered area; The time-series multispectral remote sensing images are subjected to a first preprocessing step to obtain surface reflectance images of different spectral bands in multiple periods. The time-series polarization SAR data is subjected to a second preprocessing to obtain SAR backscatter images with different polarization modes in multiple periods; S2, from the surface reflectance images of different spectral bands obtained in S1, calculates the optical vegetation index images of multiple periods; For the SAR backscatter images with different polarization modes obtained from S1, multi-phase radar vegetation index images are calculated. S3, normalize the surface reflectance images and SAR backscatter images with different polarization modes obtained from S1, as well as the optical vegetation index images and radar vegetation index images obtained from S2, to obtain the lithological classification feature dataset. S4. Acquire regional geological data and / or high-resolution remote sensing images for field investigation, label sample areas of different lithological types, and obtain vector sample label files containing lithological category labels for each lithological type; S5: Using the vector sample annotation file obtained in S4, extract the temporal feature data corresponding to all pixels in the sample area of ​​the lithology classification feature dataset obtained in S3, and associate it with the category label of the corresponding pixel to obtain the sample dataset. The sample dataset includes feature data and category labels. S6. Construct a multi-scale temporal attention network, take the sample dataset obtained in S5 as input, and train the multi-scale temporal attention network with the cross-entropy loss function to obtain the lithology classification model. S7. Input the lithological classification feature dataset obtained in S3 into the obtained lithological classification model to perform lithological classification of the entire region and obtain the lithological classification result map.

2. The stratigraphic lithology classification method for semi-arid shallow-cover areas integrating multi-source and multi-temporal remote sensing as described in claim 1, characterized in that, The multi-scale temporal attention network includes a multi-scale temporal convolution module, a temporal attention module, a feature channel attention module, a cross-dimensional interaction module, and a classifier. The training process of the multi-scale temporal attention network specifically includes: S6.1, the sample dataset is divided into multiple batches, each batch contains B samples, and each sample includes multiple time-series feature data and their corresponding category labels; The temporal feature data of all samples in each batch are organized into a temporal feature tensor, which includes a time dimension and a feature dimension. S6.2, Input the temporal feature tensor obtained in S6.1 into the multi-scale temporal attention network, and perform the following operations: S6.2a, map the feature dimension of the temporal feature tensor to the hidden dimension to obtain the feature embedding representation. X embedded ; S6.2b, embed the features obtained in S6.2a into the representation. X embedded Inputting the multi-scale temporal convolution module extracts local temporal features at different time dimensions, resulting in multi-scale temporal features. X mutiscale ; S6.2c, embed the features obtained in S6.2a into the representation. X embedded The input temporal attention module captures the global dependencies between time steps, resulting in temporal attention features. X temporal ; S6.2d, embed the features obtained in S6.2a into the representation. X embedded Input the feature channel attention module, evaluate and enhance the importance of different feature channels, and obtain the feature channel attention features. X feature ; S6.2e, the multi-scale temporal features obtained in S6.2b. X mutiscale Temporal attention features obtained from S6.2c X temporal The feature channel attention features obtained from S6.2d X feature The input is fused and interacted with by the cross-dimensional interaction module to obtain the features after cross-dimensional interaction. X fusion ; S6.2f, the cross-dimensional interactive features obtained in S6.2e. X fusion Input the classifier and output the predicted probability distribution of each sample belonging to each stratigraphic lithology category; S6.3, Based on the predicted probability distribution obtained in S6.2f and the category label corresponding to the sample, calculate the cross-entropy loss value, and update the parameters in the multi-scale temporal attention network using the cross-entropy loss value; S6.4, repeat steps S6.1 to S6.3 until convergence, to obtain the lithology classification model.

3. The stratigraphic and lithological classification method for semi-arid shallow-covered areas based on multi-source, multi-temporal remote sensing as described in claim 2, characterized in that... In S6.2b, the multi-scale temporal convolution module specifically performs the following operations: S6.2b-1, embedding the feature representation X embedded Transpose the dimension of the feature representation to obtain the transposed feature representation. X transposed ; S6.2b-2, construct K parallel one-dimensional convolutional layers, each of which has a different kernel size, where K is an integer greater than 1; S6.2b-3, the transposed feature representation is processed by K parallel one-dimensional convolutional layers respectively. X transposed Perform convolution operations, and apply non-linear activation and global pooling to the output of each one-dimensional convolutional layer to obtain a global feature representation; S6.2b-4 concatenates the K global feature representations along the feature channel dimension to obtain multi-scale temporal features. X mutiscale .

4. The stratigraphic lithology classification method for semi-arid shallow-covered areas based on multi-source, multi-temporal remote sensing as described in claim 3, characterized in that, With K=4, the kernel sizes of each one-dimensional convolutional layer are 1, 3, 5, and 7, respectively.

5. The stratigraphic lithology classification method for semi-arid shallow-covered areas integrating multi-source and multi-temporal remote sensing as described in claim 2, characterized in that, The timing attention module in S6.2c is a Transformer encoder.

6. The stratigraphic lithology classification method for semi-arid shallow-cover areas based on multi-source, multi-temporal remote sensing as described in claim 2, characterized in that, The feature channel attention module in S6.2d obtains channel attention weights through a compression-activation operation, and then combines the channel attention weights with the feature embedding representation. X embedded By multiplying each term, a weighted channel feature representation is obtained. Then, the average of these weighted channel feature representations over time is taken to obtain the feature channel attention features. X feature .

7. The stratigraphic lithology classification method for semi-arid shallow-covered areas based on multi-source, multi-temporal remote sensing as described in claim 2, characterized in that, The cross-dimensional interaction module in S6.2e performs the following operations: S6.2e-1, the multi-scale temporal features X mutiscale The temporal attention features X temporal and the feature channel attention features X feature The features are concatenated along the feature dimension to obtain the concatenated features. X cat ; S6.2e-2, regarding the splicing feature X cat Perform dimensionality reduction to obtain the reduced features; S6.2e-3, The dimensionality-reduced features are subjected to cross-dimensional feature interaction through a multi-head self-attention mechanism to obtain self-attention output; S6.2e-4, perform a residual connection between the self-attention output and the dimensionality-reduced features to obtain the cross-dimensional interactive features. X fusion .

8. The method for stratigraphic and lithological classification of semi-arid shallow-cover areas based on multi-source, multi-temporal remote sensing as described in any one of claims 1-7, characterized in that, The first preprocessing in S1 specifically involves sequentially performing radiometric calibration, atmospheric correction, orthorectification, projection conversion, mosaicking, cropping, and resampling on the time-series multispectral remote sensing image. The second preprocessing specifically involves sequentially performing radiometric correction, multi-view processing, filtering, terrain correction, geocoding, stitching, cropping, and resampling on the temporal polarimetric SAR data.

9. The stratigraphic and lithological classification method for semi-arid shallow-covered areas based on multi-source, multi-temporal remote sensing as described in any one of claims 1-7, characterized in that, The optical vegetation indices in the multi-phase optical vegetation index images in S2 include at least one of the following: atmospheric impedance vegetation index ARVI, normalized difference vegetation index NDVI, ratio vegetation index RVI, enhanced vegetation index EVI, difference vegetation index DVI, vertical vegetation index PVI, improved soil-modified vegetation index MSAVI, green light normalized difference vegetation index GNDVI, and transformed soil-modified vegetation index TSAVI. The radar vegetation indices in the multi-period radar vegetation index images include at least one of the following: classic radar vegetation index RVI, dual-polarization radar vegetation index DpRVI, vegetation anisotropy index VAI, polarization difference index PDI, and radar polarization scattering index RPSI.

10. The method for stratigraphic and lithological classification of semi-arid shallow-cover areas based on multi-source, multi-temporal remote sensing as described in any one of claims 1-7, characterized in that, The field survey described in S4 specifically includes: setting up survey points by combining regional geological data and / or high-resolution remote sensing images to ensure that there are no fewer than a preset number of survey points in each lithological unit to be classified; and creating vector surface files for annotation by combining the survey results with remote sensing image interpretation through a geographic information system.

Images