Remote sensing extraction method and system for linear fracture structure under continuous visual angle

Abstract

The invention provides a remote sensing extraction method and system for a linear fracture structure under a continuous visual angle, and relates to the technical field of fracture structure interpretation, and the method comprises the steps: carrying out the preprocessing of label data and remote sensing image data in a data set, inputting the preprocessed data set into a pre-constructed fracture structure probability graph generation model, carrying out training and verification by utilizing K-fold cross validation, and obtaining K trained fracture structure probability graph generation models based on a verification set optimal result of each fold; and inputting second continuous remote sensing image data covering a second target area and the periphery of the second target area into the trained K fracture structure probability graph generation models to obtain K probability distribution result graphs, and processing the K probability distribution graphs by using a post-processing module to finally form a fracture structure graph. According to the method, the interpretation continuity and the extraction accuracy of the fracture structure are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault structure interpretation, and in particular to a method and system for extracting linear fault structures by remote sensing under a continuous viewing angle. Background Art

[0002] Fault structure is a phenomenon in which rocks lose their continuity and integrity due to mechanical fractures in a certain direction due to tectonic stress. Fault structure represents an important geological structure in the earth's crust. The interpretation of fault structure plays a vital role in earthquake research, resource exploration, and disaster prevention.

[0003] The investigation and interpretation of traditional fault structures mainly rely on field surveys, but this method has the problems of high cost and low efficiency, especially in remote areas with inconvenient transportation. Field surveys are extremely difficult. In this case, using remote sensing data for human-computer interactive interpretation becomes a more economical and efficient option. However, human-computer interactive interpretation relies on artificially set parameters, which makes the interpretation process highly dependent on manual work. This dependence may introduce subjective factors, which may lead to inaccurate interpretation results. In addition, the continuity of the fault structure may be poor in the interpretation results, which is not conducive to observation and research. Summary of the invention

[0004] The present invention aims to solve the problem that the existing technology has a strong reliance on manual labor when interpreting remote sensing data, which may lead to inaccurate interpretation results and poor continuity of fault structures in the interpretation results.

[0005] To solve the above problems, in a first aspect, the present invention provides a method for extracting linear fault structures by remote sensing under a continuous viewing angle, comprising:

[0006] Constructing a data set of the same area, wherein the data set includes first continuous remote sensing image data of a first target area and its surroundings and label data of a fracture structural surface of the first target area;

[0007] Preprocessing the label data and remote sensing image data in the data set, and based on the preprocessed data set, performing multi-fold cross-training and verification on the pre-constructed fault structure probability map generation model to obtain a plurality of trained fault structure probability map generation models;

[0008] Inputting the second continuous remote sensing image data covering the second target area and its surroundings into the trained multiple fault structure probability map generation models to obtain multiple probability distribution result maps, wherein the multiple probability distribution result maps only include the probability distribution result map of the second target area, each of the probability distribution result maps is composed of multiple pixels, and each of the pixels has a corresponding probability of the existence of a fault structure;

[0009] Based on the probability, the pixels in the multiple probability distribution result maps are screened to obtain a result map of the fracture structure surface containing high-probability fracture structure points, and the result map of the fracture structure surface is screened and refined to obtain a result map of the fracture structure line.

[0010] Optionally, the pre-constructed fracture structure probability map generation model includes a multi-scale attention semantic segmentation model.

[0011] Optionally, the preprocessing of the label data and remote sensing image data in the data set includes:

[0012] Uniformly dividing the label data and remote sensing image data in the data set;

[0013] The uniformly split data set is randomly divided into a training validation set and a test set in a preset ratio.

[0014] The performing multi-fold cross-training and verification on the pre-built fault structure probability map generation model to obtain a plurality of trained fault structure probability map generation models comprises:

[0015] The first continuous remote sensing image data in the training validation set is divided into K subsets of equal size, and each of the subsets is used as a validation set in turn, and the remaining K-1 subsets are used as training sets. The pre-constructed fault structure probability map generation model is trained and validated K times to obtain K fault structure probability map generation models, where K is a positive integer greater than 1.

[0016] Optionally, the second continuous remote sensing image data covering the second target area and its surroundings is input into a plurality of trained fault structure probability map generation models to obtain a plurality of probability distribution result maps, including: inputting a target remote sensing image block of the second target area and a plurality of surrounding remote sensing image blocks of the target remote sensing image block into a plurality of trained fault structure probability map generation models to obtain a plurality of probability distribution result maps, the second continuous remote sensing image data includes a target remote sensing image block and a plurality of surrounding remote sensing image blocks, and the target remote sensing image block and the plurality of surrounding remote sensing image blocks correspond to a plurality of continuous positions.

[0017] Optionally, the multi-scale attention semantic segmentation model includes an encoder, a decoder and a discarding structure, wherein the encoder is used to extract abstract features and spatial directionality of fracture structures in continuous remote sensing image data under receptive fields of different sizes to obtain a result feature map that is directionally aware and position sensitive and is used to characterize the fracture structure; the decoder is used to restore the result feature map to its original size to obtain the probability distribution result map; the discarding structure is used to retain and output the result feature map associated with the target remote sensing image block, while discarding the result feature map of the surrounding remote sensing image blocks; the continuous remote sensing image data includes first continuous remote sensing image data and second continuous remote sensing image data.

[0018] Optionally, the step of screening the pixels in the plurality of probability distribution result maps based on the probability to obtain a result map of the fracture structure surface containing a high-probability fracture structure point comprises:

[0019] For each identical pixel in the plurality of probability distribution result maps, determine the average value of the probability, and retain the pixels whose average value is greater than or equal to the machine learning classification probability as the key point where the fault structure occurs, and the high probability fault structure point includes the key point;

[0020] For each identical pixel in the plurality of probability distribution result maps, determining the maximum value of the probability, and retaining the pixel whose maximum value is greater than or equal to the machine learning classification probability as a possible surface where a fault structure occurs;

[0021] The pixel corresponding to the possible surface containing the key point is selected as the result map of the fracture structure surface.

[0022] Optionally, the screening and refining of the result map of the fracture structure surface to obtain the result map of the fracture structure line includes:

[0023] The result map of the fracture structure surface is converted into vector shape data, and the unconnected parts of the vector shape data are split to form a plurality of independent continuous vector data;

[0024] The continuous vector data corresponding to the possible surface containing the key point is retained, and the raster data of the fracture structural surface is formed after rasterization processing;

[0025] The grid data of the fault structure surface is refined to obtain a result map of the fault structure line.

[0026] Optionally, constructing a data set for the same area includes:

[0027] Acquire remote sensing image data of the same area that meets a preset resolution, and fault structure vector line data in a geological map that meets a preset scale;

[0028] The fracture structure vector line data is buffered to form the label data.

[0029] Optionally, the preprocessing of the label data and remote sensing image data in the data set includes:

[0030] Uniformly dividing the label data and remote sensing image data in the data set;

[0031] The uniformly segmented data set is randomly divided into a training validation set and a test set that meet a preset ratio;

[0032] The uniformly split data set is randomly divided into a training validation set and a test set in a preset ratio.

[0033] The performing multi-fold cross-training and verification on the pre-built fault structure probability map generation model to obtain a plurality of trained fault structure probability map generation models comprises:

[0034] The first continuous remote sensing image data in the training validation set is divided into K subsets of equal size, and each of the subsets is used as a validation set in turn, and the remaining K-1 subsets are used as training sets. The pre-constructed fault structure probability map generation model is trained and validated K times to obtain K fault structure probability map generation models, where K is a positive integer greater than 1.

[0035] The present invention is conducive to the model training of the fault structure probability map generation model by constructing and preprocessing a data set including remote sensing image data and label data of the fault structure surface, and improves the accuracy of subsequent extraction. The pre-constructed fault structure probability map generation model is subjected to multi-fold cross-training and verification to obtain a plurality of trained fault structure probability map generation models. The input for training the model is the first continuous remote sensing image data, which covers the first target area and multiple continuous positions around it. Therefore, the multiple first continuous remote sensing image data used for training finally appear in the form of multiple cross-covered continuous image blocks, which simulates the observation continuity of geological personnel under the perspective of fault structure interpretation, and also indirectly enlarges the observation field of interpretation, which is helpful to improve the interpretation accuracy. It should also be emphasized here that the label in the model training stage only includes the label grid data of the first target area, and does not include the label data of multiple continuous positions around it. The grid images of each target area are not overlapped in space, so the output result of the trained model also only includes the interpretation result of the target remote sensing image block. Similarly, in order to obtain the interpreted fault structure results, after the multiple fault structure probability maps trained by continuous image input are generated into models (multiple models are generated by a training verification data set and a multi-fold crossover and verification method), multiple probability distribution result maps of the second target area corresponding to the second continuous remote sensing image data are obtained (each input second continuous remote sensing image data can correspond to multiple probability distribution result maps of the second target area), the probability of the existence of fault structure under the pixel of the probability distribution result map of each target area is interpreted, and the probability distribution result map is not generated only once, based on the probability, the pixels in the generated multiple probability distribution result maps are screened and designed to obtain the result map of the fault structure surface containing the high probability fault structure point. This screening design fully considers the linear continuity characteristics of the fault structure, that is, there is a high probability of the fault structure surface near a high probability fault structure point. Therefore, the result map of the fracture structure surface containing high-probability fracture structure points more accurately interprets the distribution of the fracture structure surface. The result map of the fracture structure surface is then screened and refined to obtain a result map of the fracture structure line, thereby improving the continuity of the interpretation of the fracture structure. Compared with the processing method of setting a threshold value that is often used in fracture structure processing in the prior art, the present invention is based on a processing method of screening fracture structure surfaces based on high-probability fracture structure points, which reduces or avoids the introduction of human subjective factors and improves the accuracy of fracture structure line extraction.

[0036] In a second aspect, the present invention further provides a linear fault structure remote sensing extraction system under a continuous viewing angle, comprising:

[0037] A data set construction module, used to construct a data set of the same area, wherein the data set includes first continuous remote sensing image data of a first target area and its surroundings and label data of a fracture structural surface of the first target area;

[0038] A model optimization module is used to preprocess the label data and remote sensing image data in the data set, and based on the preprocessed data set, perform multi-fold cross-training and verification on the pre-built fault structure probability map generation model to obtain a plurality of trained fault structure probability map generation models;

[0039] A probability distribution result map generation module is used to input the second continuous remote sensing image data covering the second target area and its surroundings into the trained multiple fault structure probability map generation models to obtain multiple probability distribution result maps, wherein the multiple probability distribution result maps only include the probability distribution result map of the second target area, each of the probability distribution result maps is spliced ​​by multiple pixels, and each of the pixels has a corresponding probability of the existence of a fault structure;

[0040] The fracture structure post-processing module is used to screen the pixels in the multiple probability distribution result maps based on the probability, obtain the result map of the fracture structure surface containing high-probability fracture structure points, and screen and refine the result map of the fracture structure surface to obtain the result map of the fracture structure line.

[0041] In a third aspect, the present invention provides an electronic device, including a memory and a processor;

[0042] The memory is used to store computer programs;

[0043] The processor is used to implement the remote sensing extraction method of linear fault structure under continuous viewing angle as described in the first aspect when executing the computer program.

[0044] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for remote sensing extraction of linear fault structures under a continuous perspective as described in the first aspect is implemented.

[0045] The linear fault structure remote sensing extraction system under continuous perspective, electronic device and computer-readable storage medium provided by the present invention have the same beneficial effects as the linear fault structure remote sensing extraction method under continuous perspective relative to the prior art, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A schematic diagram of a process flow of a method for extracting linear fault structures by remote sensing under a continuous viewing angle in an embodiment of the present invention is shown;

[0047] Figure 2 A schematic diagram of extracting a fracture structure line after continuous remote sensing image data of multiple continuous positions is shown in an embodiment of the present invention;

[0048] Figure 3 A schematic diagram of model training and post-processing in an embodiment of the present invention is shown;

[0049] Figure 4 A schematic diagram of the structure of a multi-scale attention module in an embodiment of the present invention is shown;

[0050] Figure 5 A schematic diagram showing a visual comparison of key points and possible surfaces in an embodiment of the present invention is shown;

[0051] Figure 6 A schematic diagram of the structure of a linear fault structure remote sensing extraction system according to an embodiment of the present invention is shown;

[0052] Figure 7 A schematic diagram of the structure of an electronic device in an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0054] It should be noted that, in the present invention, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0055] In the description of this specification, the description with reference to the terms "embodiment", "one embodiment" and "one implementation" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or implementation are included in at least one embodiment or implementation of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or implementation. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or implementations in a suitable manner.

[0056] Reference Figure 1 As shown, the embodiment of the present invention proposes a remote sensing extraction method for linear fault structure under continuous viewing angle;

[0057] The linear fault structure remote sensing extraction method under the continuous viewing angle comprises:

[0058] S100: constructing a data set of the same area, wherein the data set includes first continuous remote sensing image data of a first target area and its surroundings and label data of a fracture structure surface of the first target area.

[0059] Specifically, after determining the research area, remote sensing image data of the area is obtained, such as very high resolution (VHR), hyperspectral (HSI), synthetic aperture radar (SAR) and other types of remote sensing image data; remote sensing image data is used as sample data in the data set, which corresponds to label data. Label data is the output result corresponding to the sample data, also called target value. Labels are used to guide model learning and help the model identify patterns and relationships in the data.

[0060] S200: Preprocessing the label data and remote sensing image data in the data set, and based on the preprocessed data set, performing multi-fold cross-training and verification on a pre-constructed fault structure probability map generation model to obtain a plurality of trained fault structure probability map generation models.

[0061] Specifically, the processing of the data set before the model is trained may include the division of the data set, the block processing of the remote sensing image data, data cleaning, etc., to ensure that the requirements of model training and verification are met. The constructed model may be a deep learning model, and the constructed model is subjected to multi-fold cross-training and verification. For each fold, a better model parameter is selected to obtain a plurality of corresponding fault structure probability map generation models, so as to output a plurality of probability distribution result maps. Different from the prior art in which one model is taken under the cross-validation framework, the present invention obtains multiple models under each verification and training result.

[0062] S300: Input the second continuous remote sensing image data covering the second target area and its surroundings into the multiple trained fault structure probability map generation models to obtain multiple probability distribution result maps, wherein the multiple probability distribution result maps only include the probability distribution result map of the second target area, each of the probability distribution result maps is composed of multiple pixels, and each of the pixels has a corresponding probability of the existence of a fault structure.

[0063] Pixel is a basic concept in Geographic Information System (GIS), which refers to the smallest resolvable unit in a map or image.

[0064] Specifically, the relationship between the second target area and the first target area is as follows: the second target area is generally an area not far from the first target area, and the two have the same or similar tectonic background. The first continuous remote sensing image data and the second continuous remote sensing image data are input in the form of a target remote sensing image block and a plurality of the surrounding remote sensing image blocks, which strengthens the observation continuity from the perspective of fault structure interpretation and indirectly enlarges the observation field of interpretation. After the second continuous remote sensing image data is input into the fault structure probability map generation model, multiple probability distribution result maps are obtained, which strengthens the observation continuity from the perspective of fault structure interpretation and generates K probability result maps instead of just one. This can improve the reliability of the results of the continuity post-processing related to the fault structure (in the next step) and the ability to assess the uncertainty of the results.

[0065] S400: Based on the probability, the pixels in the plurality of probability distribution result maps are screened to obtain a result map of the fracture structure surface containing a high-probability fracture structure point, and the result map of the fracture structure surface is screened and refined to obtain a result map of the fracture structure line.

[0066] A comprehensive analysis is performed on multiple probability distribution result maps. Through multiple probability distribution result maps, if most of the K result maps show that there is a fault structure in the area, the confidence in the result is relatively high. If the probability value of the same pixel in the K result maps is generally high, the area where the pixel is located is very likely to be a "high probability fault structure point". If the probability value of a certain pixel is higher than that of others in the K result maps, the coverage area corresponding to the pixel is very likely to be the occurrence surface of the potential fault structure, that is, the fault structure surface. Therefore, in general, the fault structure surface containing the high probability fault structure point can be used as the final fault structure surface. Since the fault structure surface may contain some noise, it can be denoised in the refinement process. Therefore, the potential fault structure can be identified from multiple probability distribution result maps, and a refined fault structure line result map can be obtained, which provides important reference information for subsequent geological research and risk assessment. High probability fault structure points can be considered as fault structure points that are judged to be very likely to have fault structures.

[0067] When this embodiment is applied in practice, by constructing and preprocessing a data set including remote sensing image data and label data of the fracture structure surface, it is beneficial to the model training of the fracture structure probability map generation model, and improve the accuracy of subsequent extraction, and the pre-constructed fracture structure probability map generation model is multi-fold cross-trained and verified to obtain a plurality of trained fracture structure probability map generation models. The input for training the model is a continuous remote sensing image, which covers the first target area and multiple continuous positions around it. Therefore, the multiple first continuous remote sensing image data used for training finally appear in the form of multiple cross-covered continuous image blocks, which simulates the observation continuity of geological personnel under the perspective of fracture structure interpretation, and indirectly amplifies the observation field of interpretation, which helps to improve the interpretation accuracy. It should also be emphasized here that the label in the model training stage only includes the label grid data of the first target area, and does not include the label data of multiple continuous positions around it. The grid data of all target areas are not overlapped in space, so the output result of the trained model also only includes the interpretation result of the target image block position (multiple probability distribution result maps only include the probability distribution result map of the second target area). Similarly, in order to obtain the interpreted fault structure results, after the multiple fault structure probability maps trained by continuous image input are generated into models (multiple models are generated by a training verification data set and a multi-fold crossover and verification method), multiple probability distribution result maps of the second target area corresponding to the second continuous remote sensing image data are obtained (each input second continuous remote sensing image data can correspond to multiple second target area probability distribution result maps), the probability of the existence of fault structure under the pixel of each probability distribution map of the second target area is interpreted, and the probability distribution result map is not generated only once, based on the probability, the pixels in the generated multiple probability distribution result maps are screened and designed to obtain the result map of the fault structure surface containing high probability fault structure points. This screening design fully considers the linear continuity characteristics of the fault structure, that is, there is a high probability of the existence of a fault structure near a high probability fault structure point. Therefore, the result map of the fracture structure surface containing high-probability fracture structure points more accurately interprets the distribution of the fracture structure surface. The result map of the fracture structure surface is then screened and refined to obtain a result map of the fracture structure line, thereby improving the continuity of the interpretation of the fracture structure. Compared with the processing method of setting a threshold value that is often used in fracture structure processing in the prior art, the present invention is based on a processing method of screening fracture structure surfaces based on high-probability fracture structure points, which reduces or avoids the introduction of human subjective factors and improves the accuracy of fracture structure line extraction.

[0068] like Figure 2 When extracting the fault structure line after inputting continuous remote sensing image data of multiple continuous positions, the ratio of the number of the first target area in the first continuous remote sensing image data and the number of the multiple surrounding remote sensing influence blocks includes 1:8.

[0069] The second continuous remote sensing image data includes the target remote sensing image block and a plurality of the surrounding remote sensing image blocks in a ratio of 1:8.

[0070] For example, the input is set to be 1 target remote sensing image block and 8 surrounding remote sensing image blocks, that is, a total of 9 image blocks are input into the model for learning and extraction.

[0071] As an optional embodiment of the present invention, the second continuous remote sensing image data covering the second target area and its surroundings is input into a plurality of the trained fault structure probability map generation models to obtain a plurality of probability distribution result maps including:

[0072] The target remote sensing image block of the second target area and multiple surrounding remote sensing image blocks of the target remote sensing image block are input into the trained multiple fault structure probability map generation models to obtain multiple probability distribution result maps, wherein the second continuous remote sensing image data includes the target remote sensing image block and multiple surrounding remote sensing image blocks, and the target remote sensing image block and multiple surrounding remote sensing image blocks correspond to multiple continuous positions. The multiple probability distribution result maps only include the probability distribution result map of the second target area (the probability distribution result map of the surrounding area is discarded), and each of the probability distribution result maps is composed of multiple pixels, and each of the pixels has a corresponding probability of the existence of a fault structure.

[0073] The prediction accuracy of the model can be improved by inputting the target area and its surrounding images into multiple models and using the information of the surrounding area to assist the prediction of the target area.

[0074] After continuous remote sensing image data is input into the trained fault structure probability map generation model, the connection between the target remote sensing image block and the fault structure of the surrounding remote sensing image blocks is strengthened, and the observation perspective is further enlarged to strengthen the observation continuity of the fault structure interpretation network perspective; the second is to use, for example, 8 image blocks instead of more image blocks to enhance the continuity of structural interpretation in order not to increase too many network parameters; the second is to design multi-scale attention modules under different levels of features to explore the features of different sizes of perspectives of the fault image blocks to better obtain their accurate spatial location information.

[0075] As an optional embodiment of the present invention, the pre-constructed fracture structure probability map generation model includes a multi-scale attention semantic segmentation model.

[0076] Specifically, the multi-scale attention semantic segmentation model performs the semantic segmentation task under the deep learning model (U-Net model-like), assigning each pixel in the image to a specific category. The multi-scale attention mechanism can help the model better capture the different scale information in the image, thereby improving the accuracy of semantic segmentation, thereby determining the probability of the existence of a fracture structure in each pixel in the target remote sensing image block in the second continuous remote sensing image data.

[0077] like Figure 3 and Figure 4 As shown, the trained model for generating the probability map of the fault structure includes an encoder, a decoder and a discarding structure, wherein, in the model training stage, the encoder is used to extract the abstract features and spatial directionality of the fault structure in the continuous remote sensing image data under receptive fields of different sizes to obtain a result feature map for characterizing the direction perception and position sensitivity of the fault structure, the decoder is used to restore the result feature map to its original size to obtain the probability distribution result map, and the discarding structure is used to retain and output the result feature map associated with the target remote sensing image block, while discarding the result feature map of the surrounding remote sensing image blocks of the target remote sensing image block. This is because when calculating the loss function, only the label corresponding to the target remote sensing image block is interpreted, and the output feature value associated with the target remote sensing image block needs to be selectively retained, while the output feature value of the surrounding remote sensing image blocks is directly discarded, so that the loss calculation can be performed. Therefore, the result feature maps of these surrounding remote sensing image blocks are discarded.

[0078] Specifically, the encoder includes four downsampling modules and two first convolution modules, the receptive fields corresponding to the four downsampling modules decrease successively, each of the downsampling modules includes two second convolution modules and a maximum pooling layer with a kernel size of 2×2, wherein the first convolution module includes a convolution layer with a kernel size of 1×1 and a step size of 1, a Bn layer and a Relu layer, and the second convolution module includes a convolution layer with a kernel size of 3×3 and a step size of 1, a Bn layer and a Relu layer, and the downsampling modules are also respectively added with multi-scale attention modules (which can be determined in combination with the pixel block size of continuous remote sensing image data. For example, in practice, the input is a 64×64 image block, and the receptive fields corresponding to the four downsampling modules are successively reduced. Therefore, the feature map obtained by the fourth sampling module is too small, and it is not suitable to add a multi-scale attention module. Therefore, in this case, a multi-scale attention module is added to at least the first three sampling modules. When the appropriate size is set, a multi-scale attention module is added to all four sampling modules).

[0079] For the encoder, corresponding to the structure of the decoder, in terms of input, a total of 9 images of the training target and its surrounding images can be input into the model, which can effectively increase the continuity of observation. In terms of model structure, as mentioned above, for the multi-attention module, that is, the "ASPP+Coordinate Attention" module, the ASPP module can learn to observe the characteristics of the fault structure from receptive fields of different sizes, and the Coordinate Attention module helps to learn the spatial directionality of the fault structure from the results of receptive fields of different sizes, and obtain the result feature map of direction perception and position sensitivity, which is expressed by the following formula:

[0080] Out=ASPPCA(Maxpooling(Conv(Conv(x))));

[0081] Among them, Out represents the output result of a downsampling module, x represents the input data of the downsampling module, ASPPCA represents the multi-scale attention module, Conv is the convolution module, and Maxpooling is the maximum pooling layer.

[0082] Reference Figure 4 , the multi-scale attention module includes a multi-view learning module (ASPP) and an attention module (Coordinate Attention);

[0083] The multi-view learning module includes an ordinary convolution module, two dilated convolution modules and a global pooling module (four modules), the ordinary convolution module and the dilated convolution module each include a convolution layer with a kernel size of 3×3 and a step size of 1, a Bn layer and a Relu layer, in the two dilated convolution modules, the dilation size of the first dilated convolution layer is 2, the patch is 2, the dilation size of the second dilated convolution layer is 4, the patch is 4, the global pooling module includes an adaptive average pooling function, a third ordinary convolution layer and a bilinear interpolation function, the third ordinary convolution layer includes a convolution layer with a kernel size of 1×1 and a step size of 1, a Bn layer and a Relu layer; then the bilinear interpolation is restored to the feature map size of the original input ASPP layer, and then the results obtained by the above four modules are spliced ​​on the channel, and then pass through a convolution layer with a kernel size of 3×3 and a step size of 1, a Bn layer, a Relu layer and a Dropout layer, wherein the Dropout layer parameter is 0.5;

[0084] The multi-view learning module is used to generate fracture structural features under receptive fields of different sizes.

[0085] The Coordinate Attention module obtains multi-scale fault structure features through the ASPP module and then inputs them into the Coordinate Attention module to learn the spatial directionality of fault structures at different scales and obtain a result feature map that is direction-aware and position-sensitive.

[0086] The attention module includes a dimensional average pooling function, a third convolution module and a normalization module. The dimensional average pooling function is used to average pool the multi-scale fault structure features in the two spatial dimensions of horizontal and vertical coordinates respectively, and after merging the two average pooling results in the third dimension, a convolution layer with a kernel size of 1×1 and a step size of 1, a Bn layer and an h_swish layer are processed to obtain new weight values ​​in the two spatial dimensions. Based on the new weight values, the weight value sizes of the two spatial dimensions are restored, and then a convolution layer with a kernel size of 1×1 and a step size of 1 and Sigmoid normalization are passed to obtain the final position weight values ​​in the two spatial dimensions. The two position weight values ​​obtained are multiplied by the input fault structure features to obtain a feature map of the position sensitive result.

[0087] More specifically, the decoder, whose main goal is to use the abstract features extracted during the encoding process to restore them to the size of the original image, consists of four upsampling modules, each of which contains a deconvolution layer with a kernel size of 2×2 and a stride of 2, a skip connection, and two convolution blocks, which consist of a convolution layer with a kernel size of 3×3 and a stride of 1, a Bn layer, and a Relu layer. The skip connection between the decoder and the encoder is used to connect features of the same resolution from different levels of the encoder, thereby alleviating the loss of spatial information caused by downsampling. Finally, a 1×1 convolution layer is applied for dimensionality reduction to reduce the number of channels to the number of categories, thereby obtaining the feature map of the target remote sensing image block plus the surrounding remote sensing image blocks.

[0088] When this embodiment is applied in practice, the composition of the neural network structure under the fault structure probability map generation model can be used to input the first continuous remote sensing image data into the model training, which can effectively increase the observation continuity and learn the spatial directionality of the fault structure from the results of receptive fields of different sizes. Thus, a result feature map representing directional perception and position sensitivity is finally obtained, and a feature map of the target remote sensing image block plus the surrounding remote sensing image blocks is output. Only the output feature map associated with the target remote sensing image block is selectively retained, and the output feature map of the surrounding remote sensing image blocks is directly discarded to obtain multiple probability score result maps.

[0089] After actual comparison, it was found that the multi-scale attention semantic segmentation model obtained after training and verification in this application is higher than the classic popular or newer semantic segmentation models. Specifically, at the pixel scale, the Precision, Recall and F1 scores of the model are improved by 28.98%, 32.83% and 31.77% respectively compared with the suboptimal model. These data fully illustrate the key role of the proposed background image input and discarding structure, the multi-scale attention module and the post-processing module based on key points and possible surfaces in accurately predicting fault structures.

[0090] Among them, ① Precision: This indicator measures the proportion of fault structure pixels that are actually accurate in the results predicted as fault structures (Precision = number of fault structure pixels predicted correctly / total number of fault structure pixels predicted). ② Recall: This indicator indicates the percentage of correct fault structure pixels successfully predicted by the model among the pixels labeled as fault structures (Recall = number of fault structure pixels predicted correctly / number of pixels labeled as fault structures). ③ F1 score: It is a comprehensive evaluation indicator of accuracy and recall, which is used to balance the performance between the two (F1 = 2×Precision×Recall / (Precision+Recall)). Among them, the lower the recall rate, the fewer fault structures in the real labels covered in the prediction results; and the lower the accuracy rate, the more fault structures with non-real labels are included in the prediction results.

[0091] As an optional embodiment of the present invention, the preprocessing of the label data and remote sensing image data in the data set includes:

[0092] The label data and remote sensing image data in the data set are uniformly divided.

[0093] Specifically, the label data of the fracture structure surface in the 1:50,000 scale geological map and the high-resolution image of the area are cut into a size of 64×64 (pixels).

[0094] The uniformly segmented data set is randomly divided into a training validation set and a test set that meet a preset ratio;

[0095] Specifically, a Gaussian distributed random number method is used to divide the training validation set and the test set according to a set ratio, or a random function is used to automatically randomly select a fixed number of sets from the data set to meet a preset ratio, such as 50% training validation set and 50% test set.

[0096] like Figure 3As shown, as an optional embodiment of the present invention, the pre-constructed fault structure probability map generation model is subjected to multi-fold cross-training and verification to obtain a plurality of trained fault structure probability map generation models, including:

[0097] The first continuous remote sensing image data in the training validation set is divided into K subsets of equal size, and each of the subsets is used as a validation set in turn, and the remaining K-1 subsets are used as training sets. The pre-constructed fault structure probability map generation model is trained and validated K times to obtain K fault structure probability map generation models, where K is a positive integer greater than 1.

[0098] During the experiment, the PyTorch1.7 deep learning framework was selected to build and train the model. All experiments were conducted in the Ubuntu20.04.3 operating system environment, and the graphics card used was GeForceRTX3080. The experiment initially used the channel average and standard deviation values ​​to normalize the image. The batch size was set to 10, the number of training iterations was set to 200, the learning rate was set to 0.001, and the Adam optimizer was used for learning rate optimization. During the training process, the model was trained using the five-fold cross-validation method, and the loss function used the cross-entropy loss. In the multi-scale attention semantic segmentation model, the selection of the optimal parameters for each fold was determined by the maximum MIOU value obtained on the validation set, and applied to the test set to generate the corresponding probability score result map, and then the final result map generated by the fracture structure post-processing module 400 was used for accuracy calculation. In comparison, the traditional semantic segmentation model does not design the fracture structure post-processing module 400, so in this application, the accuracy calculation directly uses the model parameters corresponding to the maximum MIOU value in the five-fold cross-validation method to generate the final fracture surface result map for accuracy calculation.

[0099] Specifically, by using different data subsets for training and testing, the performance of the model on different data can be more accurately evaluated to avoid overfitting the model to the training data. By applying the cross-validation method, more stable model evaluation results can be obtained based on the first continuous remote sensing image data. The choice of K value depends on the amount of data. Generally, K values ​​of 5 or 10 are more common. The larger the K value, the more stable the evaluation results.

[0100] When this embodiment is applied in practice, by dividing the first continuous remote sensing image data into K subsets and training and verifying them in turn, a more reliable model evaluation result can be obtained, which can improve the generalization ability of the model and prevent overfitting.

[0101] The fault structure remote sensing extraction model constructed by the present invention can obtain a better fault structure extraction result map with the help of background image input and discard structure, multi-scale attention module and post-processing module, providing technical support for geological surveys in difficult-to-reach areas.

[0102] like Figure 3 and Figure 5 As shown, as an optional embodiment of the present invention, the pixels in the plurality of probability distribution result maps are screened based on the probability to obtain the result map of the fracture structure surface containing the high probability fracture structure point, including:

[0103] For each identical pixel in the plurality of probability distribution result maps, determine the average value of the probability, and retain the pixels whose average value is greater than or equal to the machine learning classification probability as the key point where the fault structure occurs, and the high probability fault structure point includes the key point;

[0104] The machine learning classification probability is 0.5, indicating that the possibility of the existence of fault structure is relatively high.

[0105] Specifically, expressed in formula, K l =If[(K1+K2+……+K n ) / n>=0.5];

[0106] Among them, K l It means that a single pixel is judged as a key point, n refers to the number of n probability map results obtained by the K-fold cross-validation method, Kn refers to the probability value of the fault structure at a certain position in the nth probability map, and n is a positive integer greater than or equal to 1.

[0107] For each identical pixel in the plurality of probability distribution result maps, determining the maximum value of the probability, and retaining the pixel whose maximum value is greater than or equal to the machine learning classification probability as a possible surface where a fault structure occurs;

[0108] Specifically, expressed in formula, K m =If[max(K1,K2,……,K n )>=0.5];

[0109] Among them, K m Indicates that a single pixel is judged to be a possible surface, m is less than or equal to n.

[0110] The pixel corresponding to the possible surface containing the key point is selected as the result map of the fracture structure surface.

[0111] Specifically, first, for each 64×64 (pixel) output result map, all the output result maps are spliced ​​to generate K complete result maps; the average value of the same position of the K spliced ​​maps is taken, and the position with a probability value ≥ 0.5 is retained, representing the high probability area of ​​the occurrence of the fault structure, which is called the key point. At the same time, the maximum value of the same position of the K spliced ​​maps is taken, and the position with a probability value ≥ 0.5 is also retained, indicating the potential occurrence of the fault structure, which is called the possibility; the key points are used to screen the possible surfaces. Given the linear characteristics of the fault structure, there is a high probability that the fault structure may be near a high probability point, so the possible surface containing the key points is selected as the final fault structure surface. The determination of the machine learning classification probability avoids the influence of the final interpretation result of the fault structure line by manually setting the threshold.

[0112] First, a visual description of key points and possible faces is given, such as Figure 5 As shown in Figure A, five probability maps obtained by the K-fold (5-fold is used in the example) cross-validation method are averaged, and the areas with probabilities greater than or equal to 0.5 are selected as key points in the map. Then, after the maximum value is obtained for the five probability maps, the possible surfaces in the map are determined with probabilities greater than or equal to 0.5. It can be observed that there are many possible surfaces in the prediction results of the model, but not all possible surface locations represent a high probability of fault structures. Therefore, key points with high probabilities are used to select the range of possible surfaces. This is because fault structures are often linearly continuous, and there are high probabilities of fault structures around high-probability fault structure points. This not only takes into account the continuity of the structural line, but also avoids the selection of too many low-probability structural surfaces, and significantly reduces subjective dependence. The results of comparing the screened possible surface results with the fault structure line labels are shown in the figure. Figure 5 B. Finally, the final fracture structure line result map is obtained through the thinning algorithm ( Figure 5 C).

[0113] In summary, the above method can accurately determine the fault structural surface through the joint screening of possible surfaces and key points, so the subsequent extracted fault structural lines can also be accurately obtained.

[0114] like Figure 3 (See Figure 3 As shown in the lower half of the figure, as an optional embodiment of the present invention, the result map of the fracture structure surface is screened and refined to obtain the result map of the fracture structure line, including:

[0115] The result map of the fracture structure surface (initially in raster format) is entirely converted into vector shape data, and unconnected parts of the vector shape data are split to form multiple independent continuous vector data;

[0116] Shape data refers to a format that stores geographic information in the form of vector data in a shp file.

[0117] The continuous vector data corresponding to the possible surface containing the key point is retained, and the raster data of the fracture structural surface is formed after rasterization processing;

[0118] Specifically, after converting the raster data of the possible surface and the key point (the result map of the fracture structure surface) into vector shape data, the vector data that are not connected in space are split respectively, and finally a plurality of independent continuous vector data are formed. Then, the vector data of the key point are traversed, and the continuous vector data corresponding to the possible surface containing the key point is retained. The possible surface vector data containing the key point is retained, and the retained possible surface vector data is then rasterized to form the final possible surface raster data.

[0119] The grid data of the fault structure surface is refined to obtain a result map of the fault structure line.

[0120] Specifically, the model outputs a planar distribution. In order to be consistent with the linear representation on the geological map, the Zhang-Suen thinning algorithm is further used to refine the structural surface to obtain the fault structural line of the final interpretation result. The fault structural line is the final interpretation result of the target remote sensing image block in the second continuous remote sensing image data.

[0121] The conversion between raster data and vector data (shape data) involved in this embodiment is a common conversion format in the prior art. By vectorizing, splitting, filtering and rasterizing the result map of the fracture structure surface, an accurate result map of the fracture structure line is finally obtained, which is convenient for observation under continuous viewing angles.

[0122] As an optional embodiment of the present invention, constructing a data set of the same area includes:

[0123] Acquire remote sensing image data of the same area that meets a preset resolution, and fault structure vector line data in a geological map that meets a preset scale;

[0124] Specifically, after selecting the research area, high-resolution images of the area are collected, including multispectral remote sensing images, high-precision DEM data and SAR data. After merging the downloaded SAR, multispectral and other remote sensing images into a single image with multiple channels, they are cropped and divided according to the size of A×A (which can be set according to actual needs, such as 64×64, etc.). The range of each cropped image does not overlap, and the label data is also cropped and divided according to the same ratio.

[0125] Collect the fracture structure vector line data in the publicly published 1:50,000 scale geological map. The fracture structure vector line data is the fracture structure vector shp type line data (referring to a format that stores geographic information in the form of vector data in a shp file).

[0126] The fracture structure vector line data is buffered to form the label data.

[0127] Label data is the target value or category used to guide model training. For the fault structure probability map generation model, the label data may be information about the existence or non-existence of the fault structure, the type of fault structure, etc. Labels are usually categorical variables or numerical variables.

[0128] Specifically, ArcGIS software is used to process line buffers with xm specifications, such as 200m processing, to form fault structural surface data.

[0129] In summary, after obtaining remote sensing image data and label data of the same area that meet the preset resolution, they can serve as the basis and guiding requirements for training subsequent models.

[0130] like Figure 6 As shown, on the other hand, the present invention also provides a linear fault structure remote sensing extraction system 200 under continuous viewing angle, which uses the linear fault structure remote sensing extraction method under continuous viewing angle as described in the above embodiment, including:

[0131] A data set construction module 210 is used to construct a data set of the same area, wherein the data set includes first continuous remote sensing image data of a first target area and its surroundings and label data of a fracture structure surface of the first target area;

[0132] The model optimization module 220 is used to pre-process the label data and remote sensing image data in the data set, and based on the pre-processed data set, perform multi-fold cross-training and verification on the pre-built fault structure probability map generation model to obtain a plurality of trained fault structure probability map generation models;

[0133] The probability distribution result map generation module 230 is used to input the second continuous remote sensing image data covering the second target area and its surroundings into the trained multiple fault structure probability map generation models to obtain multiple probability distribution result maps, wherein the multiple probability distribution result maps only include the probability distribution result map of the second target area, each of the probability distribution result maps is composed of multiple pixels, and each of the pixels has a corresponding probability of the existence of a fault structure;

[0134] The fracture structure post-processing module 240 is used to screen the pixels in the multiple probability distribution result maps based on the probability, obtain the result map of the fracture structure surface containing high-probability fracture structure points, and screen and refine the result map of the fracture structure surface to obtain the result map of the fracture structure line.

[0135] The specific implementation of this embodiment can refer to the corresponding implementation method mentioned above and will not be described again here.

[0136] like Figure 7 As shown, an electronic device 300 provided by an embodiment of the present invention includes a memory 310 and a processor 320; the memory 310 is used to store a computer program; the processor 320 is used to implement the linear fault structure remote sensing extraction method under a continuous perspective as described above when executing the computer program.

[0137] In other words, an electronic device 300 includes a memory 310 and a processor 320 coupled to the memory 310; the memory 310 is configured to store a computer program; and the processor 320 is configured to perform the following operations when executing the computer program:

[0138] Constructing a data set of the same area, wherein the data set includes first continuous remote sensing image data of a first target area and its surroundings and label data of a fracture structural surface of the first target area;

[0139] Preprocessing the label data and remote sensing image data in the data set, and based on the preprocessed data set, performing multi-fold cross-training and verification on the pre-constructed fault structure probability map generation model to obtain a plurality of trained fault structure probability map generation models;

[0140] Inputting the second continuous remote sensing image data covering the second target area and its surroundings into the trained multiple fault structure probability map generation models to obtain multiple probability distribution result maps, wherein the multiple probability distribution result maps only include the probability distribution result map of the second target area, each of the probability distribution result maps is composed of multiple pixels, and each of the pixels has a corresponding probability of the existence of a fault structure;

[0141] Based on the probability, the pixels in the multiple probability distribution result maps are screened to obtain a result map of the fracture structure surface containing high-probability fracture structure points, and the result map of the fracture structure surface is screened and refined to obtain a result map of the fracture structure line.

[0142] An embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method for extracting linear fault structures by remote sensing under continuous viewing angles as described above is implemented.

[0143] In other words, a non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor performs the following operations:

[0144] Constructing a data set of the same area, wherein the data set includes first continuous remote sensing image data of a first target area and its surroundings and label data of a fracture structural surface of the first target area;

[0145] Preprocessing the label data and remote sensing image data in the data set, and based on the preprocessed data set, performing multi-fold cross-training and verification on the pre-constructed fault structure probability map generation model to obtain a plurality of trained fault structure probability map generation models;

[0146] Inputting the second continuous remote sensing image data covering the second target area and its surroundings into the trained multiple fault structure probability map generation models to obtain multiple probability distribution result maps, wherein the multiple probability distribution result maps only include the probability distribution result map of the second target area, each of the probability distribution result maps is composed of multiple pixels, and each of the pixels has a corresponding probability of the existence of a fault structure;

[0147] Based on the probability, the pixels in the multiple probability distribution result maps are screened to obtain a result map of the fracture structure surface containing high-probability fracture structure points, and the result map of the fracture structure surface is screened and refined to obtain a result map of the fracture structure line.

[0148] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0149] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

[0150] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.

Claims

1. A method for extracting linear fault structures from a continuous viewing angle using remote sensing, characterized in that: include: Constructing a data set of the same area, wherein the data set includes first continuous remote sensing image data of a first target area and its surroundings and label data of a fracture structural surface of the first target area; Preprocessing the label data and the remote sensing image data in the data set, and based on the preprocessed data set, performing multi-fold cross-training and verification on the pre-constructed fault structure probability map generation model to obtain a plurality of trained fault structure probability map generation models; Inputting the second continuous remote sensing image data covering the second target area and its surroundings into the trained multiple fault structure probability map generation models to obtain multiple probability distribution result maps, wherein the multiple probability distribution result maps only include the probability distribution result map of the second target area, each of the probability distribution result maps is composed of multiple pixels, and each of the pixels has a corresponding probability of the existence of a fault structure; Based on the probability, the pixels in the multiple probability distribution result maps are screened to obtain a result map of the fracture structure surface containing high-probability fracture structure points, and the result map of the fracture structure surface is screened and refined to obtain a result map of the fracture structure line.

2. The method for extracting linear fault structures from a continuous viewing angle according to claim 1, characterized in that: The second continuous remote sensing image data covering the second target area and its surroundings is input into the trained multiple fault structure probability map generation models to obtain multiple probability distribution result maps including: The target remote sensing image block of the second target area and multiple surrounding remote sensing image blocks of the target remote sensing image block are input into the trained multiple fault structure probability map generation models to obtain multiple probability distribution result maps. The second continuous remote sensing image data includes the target remote sensing image block and multiple surrounding remote sensing image blocks, and the target remote sensing image block and multiple surrounding remote sensing image blocks correspond to multiple continuous positions.

3. The method for extracting linear fault structures from a continuous viewing angle according to claim 2, characterized in that: The pre-built fracture structure probability map generation model includes a multi-scale attention semantic segmentation model.

4. The method for extracting linear fault structures from a continuous viewing angle by remote sensing according to claim 2, characterized in that: The preprocessing of the label data and remote sensing image data in the data set includes: Uniformly dividing the label data and remote sensing image data in the data set; The uniformly split data set is randomly divided into a training validation set and a test set in a preset ratio.

5. The method for extracting linear fault structures from a continuous viewing angle according to claim 4, characterized in that: The performing multi-fold cross-training and verification on the pre-built fault structure probability map generation model to obtain a plurality of trained fault structure probability map generation models comprises: The first continuous remote sensing image data in the training validation set is divided into K subsets of equal size, and each of the subsets is used as a validation set in turn, and the remaining K-1 subsets are used as training sets. The pre-constructed fault structure probability map generation model is trained and validated K times to obtain K fault structure probability map generation models, where K is a positive integer greater than 1.

6. The method for extracting linear fault structures from a continuous viewing angle by remote sensing according to claim 2 or 3, characterized in that: The multi-scale attention semantic segmentation model includes an encoder, a decoder and a discarding structure, wherein the encoder is used to extract abstract features and spatial directionality of fracture structures in continuous remote sensing image data under receptive fields of different sizes to obtain a result feature map for characterizing direction perception and position sensitivity of the fracture structure, the decoder is used to restore the result feature map to its original size to obtain the probability distribution result map, the discarding structure is used to retain and output the result feature map associated with the target remote sensing image block, and discard the result feature map of the surrounding remote sensing image blocks, and the continuous remote sensing image data includes first continuous remote sensing image data and second continuous remote sensing image data.

7. The method for extracting linear fault structures from a continuous viewing angle using remote sensing according to any one of claims 1 to 5, characterized in that: The step of screening the pixels in the plurality of probability distribution result maps based on the probability to obtain a result map of the fracture structure surface containing a high probability fracture structure point comprises: For each identical pixel in the plurality of probability distribution result maps, determine the average value of the probability, and retain the pixels whose average value is greater than or equal to the machine learning classification probability as the key point where the fault structure occurs, and the high probability fault structure point includes the key point; For each identical pixel in the plurality of probability distribution result maps, determining the maximum value of the probability, and retaining the pixel whose maximum value is greater than or equal to the machine learning classification probability as a possible surface where a fault structure occurs; The pixel corresponding to the possible surface containing the key point is selected as the result map of the fracture structure surface.

8. The method for extracting linear fault structures from a continuous viewing angle by remote sensing according to claim 7, characterized in that: The screening and refinement of the result map of the fracture structure surface to obtain the result map of the fracture structure line includes: The result map of the fracture structure surface is converted into vector shape data, and the unconnected parts of the vector shape data are split to form a plurality of independent continuous vector data; The continuous vector data corresponding to the possible surface containing the key point is retained, and the raster data of the fracture structural surface is formed after rasterization processing; The grid data of the fault structure surface is refined to obtain a result map of the fault structure line.

9. The method for extracting linear fault structures from a continuous viewing angle using remote sensing according to any one of claims 1 to 5, characterized in that: The data set for constructing the same area includes: Acquire remote sensing image data of the same area that meets a preset resolution, and fault structure vector line data in a geological map that meets a preset scale; The fracture structure vector line data is buffered to form the label data.

10. A linear fault structure remote sensing extraction system under continuous viewing angle, characterized in that: The method for extracting linear fault structures by remote sensing under continuous viewing angle as claimed in any one of claims 1 to 9 comprises: A data set construction module, used to construct a data set of the same area, wherein the data set includes first continuous remote sensing image data of a first target area and its surroundings and label data of a fracture structural surface of the first target area; A model optimization module is used to preprocess the label data and remote sensing image data in the data set, and based on the preprocessed data set, perform multi-fold cross-training and verification on the pre-built fault structure probability map generation model to obtain a plurality of trained fault structure probability map generation models; A probability distribution result map generation module is used to input the second continuous remote sensing image data covering the second target area and its surroundings into the trained multiple fault structure probability map generation models to obtain multiple probability distribution result maps, wherein the multiple probability distribution result maps only include the probability distribution result map of the second target area, each of the probability distribution result maps is spliced ​​by multiple pixels, and each of the pixels has a corresponding probability of the existence of a fault structure; The fracture structure post-processing module is used to screen the pixels in the multiple probability distribution result maps based on the probability, obtain the result map of the fracture structure surface containing high-probability fracture structure points, and screen and refine the result map of the fracture structure surface to obtain the result map of the fracture structure line.

Images