Vegetation coverage area lithology segmentation method based on polarization scattering angle fusion network

Abstract

The invention discloses a vegetation coverage area lithology segmentation method based on a polarization scattering angle fusion network. According to the method, large-area vegetation-covered dual-polarization synthetic aperture radar data is taken as a data source, the strong penetrability and coherence of a dual-polarization synthetic aperture radar image are fully utilized, and the lithology segmentation model is constructed in combination with polarization decomposition knowledge. When a lithologic segmentation model is trained, a lithologic data set containing intensity information of first polarization and second polarization and polarization scattering angle parameters is input into the lithologic segmentation model for learning; the semantic information of the polarization scattering angle parameter is additionally reserved as far as possible by using the polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image, so that the lithology segmentation model learns the characteristics which are more stable and have more physical significance under the guidance of the polarization scattering angle parameter; therefore, accurate segmentation of various lithologic categories in the vegetation coverage area is realized by using the trained lithologic segmentation model.

Description

Technical Field

[0001] The present application relates to the technical field of remote sensing image processing, and in particular to a lithology segmentation method for vegetation-covered areas based on a polarization scattering angle fusion network. Background Art

[0002] As the main component of the earth's shallow layer, accurate identification and classification of rock properties are crucial in geological surveys, mineral resource exploration, and military defense. Compared with man-made structures, water bodies, vegetation and other features with obvious characteristics, rocks are very common due to the diversity of their constituent minerals, the variability of the proportions of each mineral, and the rich types. They are also easily disturbed by vegetation cover, clouds, rain and snow, which greatly increases the difficulty of rock classification.

[0003] Traditional geological mapping mainly uses field survey methods, which consumes a lot of manpower, material resources, and financial resources and is inefficient, especially in areas with complex terrain. In recent years, remote sensing (RS) technology has gradually been applied to the field of geological mapping due to its advantages of low cost, fast data acquisition speed, and large detection range. Among them, optical remote sensing satellites were first used in the lithology classification of bedrock exposed areas. However, it is difficult to use optical remote sensing images to classify lithology in vegetation covered areas. Even if the vegetation coverage rate just exceeds 10%, the lithology information on the surface becomes very vague or completely covered.

[0004] Moreover, although there are many advanced models in the field of computer vision, which perform well in natural image processing, when it comes to remote sensing image processing, these images are very different from natural images, and different bands also contain different physical meanings. It is not appropriate to directly apply relevant technologies in the field of computer vision to perform rock segmentation on remote sensing images, and the effect of rock segmentation cannot be guaranteed.

[0005] In addition, the rock segmentation areas in related technologies are mostly located in arid or semi-arid areas with a high degree of rock exposure and less research on vegetation-covered areas, making it difficult to extract rock characteristics under vegetation-covered conditions and the rock classification accuracy is low. Summary of the invention

[0006] In view of this, the present application provides a lithology segmentation method for vegetation-covered areas based on a polarization scattering angle fusion network, which takes dual-polarization synthetic aperture radar data of a large area of ​​vegetation coverage as the data source, fully utilizes the strong penetration and coherence of dual-polarization synthetic aperture radar images, and combines polarization decomposition knowledge to construct a lithology segmentation model to improve the accuracy and efficiency of lithology segmentation in vegetation-covered areas, so that the lithology segmentation model can learn more stable and physically meaningful features under the guidance of polarization scattering angle parameters, realize high-precision lithology segmentation tasks, and thus enhance the application value of synthetic aperture radar images in geological surveys and resource exploration. According to one aspect of the present application, a method for lithology segmentation of vegetation coverage areas based on a polarization scattering angle fusion network is provided, comprising: Collecting a dual-polarization synthetic aperture radar image of a vegetation-covered area, wherein the dual-polarization synthetic aperture radar image includes first polarization channel data and second polarization channel data; Preprocessing the dual-polarization synthetic aperture radar image to obtain a first polarization intensity image and a second polarization intensity image corresponding to the dual-polarization synthetic aperture radar image, and performing polarization parameter decomposition on the dual-polarization synthetic aperture radar image to obtain a polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image; Determine a synthetic image corresponding to the dual-polarization synthetic aperture radar image according to a first polarization intensity image, a second polarization intensity image, and a polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image; A lithology segmentation model is constructed based on a neural network, wherein the lithology segmentation model is used to extract lithology features in a synthetic image corresponding to the dual-polarization synthetic aperture radar image and features of a polarization scattering angle parameter in a polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image, and to fuse the lithology features with features of the polarization scattering angle parameter image; Iteratively training the lithology segmentation model according to the lithology data set and the polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image, wherein the lithology data set is determined according to the synthetic image corresponding to the dual-polarization synthetic aperture radar image; The trained lithology segmentation model is used to perform lithology segmentation on the synthetic aperture radar image to be tested in the vegetation coverage area to obtain a lithology segmentation result of the synthetic aperture radar image to be tested.

[0007] According to another aspect of the present application, a lithology segmentation device for vegetation coverage area based on a polarization scattering angle fusion network is provided, comprising: An acquisition module, used for acquiring a dual-polarization synthetic aperture radar image of a vegetation-covered area, wherein the dual-polarization synthetic aperture radar image includes first polarization channel data and second polarization channel data; a processing module, configured to preprocess the dual-polarization synthetic aperture radar image to obtain a first polarization intensity image and a second polarization intensity image corresponding to the dual-polarization synthetic aperture radar image, and perform polarization parameter decomposition on the dual-polarization synthetic aperture radar image to obtain a polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image; and Determine a synthetic image corresponding to the dual-polarization synthetic aperture radar image according to a first polarization intensity image, a second polarization intensity image, and a polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image; A construction module is used to construct a lithology segmentation model based on a neural network, wherein the lithology segmentation model is used to extract lithology features in a synthetic image corresponding to the dual-polarization synthetic aperture radar image and features of a polarization scattering angle parameter in a polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image, and fuse the lithology features with features of the polarization scattering angle parameter image; A training module, used for iteratively training the lithology segmentation model according to a lithology data set and a polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image, wherein the lithology data set is determined according to a synthetic image corresponding to the dual-polarization synthetic aperture radar image; The segmentation module is used to perform lithology segmentation on the synthetic aperture radar image to be tested in the vegetation coverage area by using the trained lithology segmentation model to obtain the lithology segmentation result of the synthetic aperture radar image to be tested.

[0008] According to another aspect of the present application, a readable storage medium is provided, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above-mentioned lithology segmentation method of vegetation coverage area based on polarization scattering angle fusion network are implemented.

[0009] According to another aspect of the present application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor. When the processor executes the program, the steps of the above-mentioned method for lithology segmentation of vegetation coverage areas based on a polarization scattering angle fusion network are implemented.

[0010] By means of the above technical scheme, the present application provides a lithology segmentation method for vegetation-covered areas based on polarization scattering angle fusion network, which makes full use of the strong penetration and coherence of dual-polarization synthetic aperture radar images, and converts the dual-polarization synthetic aperture radar images of large-area vegetation coverage into their corresponding synthetic images, so as to reflect the scattering characteristics and physical scattering mechanism of the ground objects in the dual-polarization synthetic aperture radar images through the intensity information of the first polarization and the second polarization and the polarization scattering angle parameters contained in the synthetic images. Then, a lithology segmentation model based on a neural network guided by polarization decomposition knowledge is constructed. Next, the lithology segmentation model is iteratively trained according to the lithology dataset and the polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image. When the lithology dataset containing the intensity information of the first polarization and the second polarization and the polarization scattering angle parameter is input into the lithology segmentation model to extract the lithology characteristics, the polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image is used to additionally retain the characteristics of the polarization scattering angle parameter as much as possible, and the lithology characteristics are fused with the characteristics of the polarization scattering angle parameter image, so as to ensure that the important information of the polarization scattering angle parameter is not lost during the training process, so that the lithology segmentation model can learn more stable and physically meaningful characteristics under the guidance of the polarization scattering angle parameter, and improve the reliability of the lithology segmentation model. Then, the trained lithology segmentation model can be used to realize the accurate segmentation of various lithology categories in the vegetation coverage area, and obtain high-precision lithology segmentation results, which has good practical application value in the fields of geological survey and resource exploration.

[0011] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 A schematic diagram of the process of the lithology segmentation method of the vegetation coverage area based on the polarization scattering angle fusion network provided in an embodiment of the present application is shown; Figure 2 A schematic diagram of the process of preprocessing and polarization parameter decomposition provided in an embodiment of the present application is shown; Figure 3 A schematic diagram of the structure of the lithology segmentation model provided in the embodiment of the present application is shown; Figure 4 A schematic diagram of the structure of a feature fusion module provided in an embodiment of the present application is shown; Figure 5A schematic diagram of the structure of a polarization scattering angle parameter extractor provided in an embodiment of the present application is shown; Figure 6 A schematic diagram of the experimental vegetation coverage area and a Landsat-8 visible light image provided in an embodiment of the present application are shown; Figure 7 The figure shows the vegetation coverage of the experimental vegetation coverage area provided in the embodiment of the present application; Figure 8 The lithology true value map of the experimental vegetation coverage area provided in the embodiment of the present application is shown; Fig. 9 The lithology segmentation model described in the present application and different segmentation models in the related art for the lithology segmentation mapping of the historical ALOS-1 image of the experimental vegetation coverage area are shown; Fig.10 A comparison diagram showing the lithology segmentation of the ALOS-1 image to be tested by the lithology segmentation model described in the present application provided in the embodiment of the present application is shown; Fig.11 The structural block diagram of the lithology segmentation device for vegetation coverage area based on polarization scattering angle fusion network provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0013] The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0014] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be interpreted as limiting the present application.

[0015] Those skilled in the art will appreciate that, unless expressly stated, the singular forms "a", "an", "said" and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present application refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "connected" to another element, it may be directly connected or connected to the other element, or there may be intermediate elements. In addition, the "connection" or "connection" used herein may include wireless connection or wireless fusion. The term "and / or" used herein includes all or any unit and all combinations of one or more associated listed items.

[0016] Now, exemplary embodiments according to the present application will be described in more detail with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in a variety of different forms and should not be interpreted as being limited to the embodiments set forth herein. It should be understood that these embodiments are provided to make the disclosure of the present application thorough and complete, and to fully convey the concepts of these exemplary embodiments to those of ordinary skill in the art.

[0017] In this embodiment, a lithology segmentation method for vegetation coverage area based on polarization scattering angle fusion network is provided. Figure 1 As shown, the method includes: Step 101: collect dual-polarization synthetic aperture radar images of vegetation coverage areas.

[0018] The dual-polarization synthetic aperture radar image includes first polarization channel data and second polarization channel data.

[0019] It can be understood that the vegetation coverage area described in this application is the object area of ​​this application research under vegetation coverage conditions, and can also be called the research area under vegetation coverage, for example, Liaoning Province under vegetation coverage.

[0020] In this embodiment, SAR (Synthetic Aperture Radar) active microwave remote sensing technology is used to collect high-resolution dual-polarization SAR images (Dual Polarization Synthetic Aperture Radar, i.e., dual-polarization synthetic aperture radar images) of vegetation-covered areas, thereby taking dual-polarization SAR images of large-area vegetation coverage as the data source, making full use of the strong penetration and coherence of dual-polarization SAR images, and improving the reliability of lithology segmentation.

[0021] Here, the polarization mode of SAR refers to the polarization direction of the electromagnetic wave. In this embodiment, SAR transmits horizontally polarized electromagnetic waves and simultaneously receives echo signals in horizontal and vertical polarization directions, thereby obtaining a dual-polarization SAR image. Among them, the dual-polarization SAR image is a two-channel image, with HH (horizontal transmission and horizontal reception) polarization channel data and HV (horizontal transmission and vertical reception) polarization channel data, that is, the first polarization channel data and the second polarization channel data. It should be noted that HH polarization is the first polarization and HV polarization is the second polarization.

[0022] For example, the PALSAR (Phased Arraytype L-band Synthetic Aperture Radar) sensor carried by the ALOS-1 (Advanced Land Observing Satellite) satellite is used to obtain L-band dual-polarization (HH and HV) H1.1A-level images with a spatial resolution of 10m in vegetation-covered areas to more clearly present surface details.

[0023] It is worth mentioning that compared with optical remote sensing images, SAR, as an active microwave remote sensing technology, has a much longer signal wavelength than optical signals. It can not only obtain data all day and all night, but also has penetrating power, which provides a good opportunity for lithology classification in vegetation-covered areas. Moreover, L-band signals can not only penetrate dense vegetation, but also penetrate shallow surfaces, reducing the interference of surface weathering, dust, and debris on lithology classification, and can directly obtain lithology information and improve the accuracy of lithology segmentation.

[0024] In this embodiment, the HH polarization mode can observe ground objects with low scattering rates in vegetation coverage areas, and the HV polarization mode can distinguish ground objects with different scattering characteristics in vegetation coverage areas, thereby transmitting and receiving signals through different polarization modes to obtain richer information than traditional single-polarization SAR.

[0025] Step 102: preprocess the dual-polarization synthetic aperture radar image to obtain a first polarization intensity image and a second polarization intensity image corresponding to the dual-polarization synthetic aperture radar image, and perform polarization parameter decomposition on the dual-polarization synthetic aperture radar image to obtain a polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image.

[0026] In this embodiment, the dual-polarization SAR image is preprocessed to extract the intensity (Intensity) (also referred to as power or backscatter coefficient) information of two different polarizations HH and HV of the dual-polarization SAR image (i.e., the intensity information of the first polarization and the second polarization), so as to obtain the HH polarization intensity image (i.e., the first polarization intensity image) corresponding to the dual-polarization SAR image according to the intensity information of the HH polarization of the dual-polarization SAR image. Similarly, the HV polarization intensity image (the second polarization intensity image) corresponding to the dual-polarization SAR image is obtained according to the intensity information of the HV polarization of the dual-polarization SAR image.

[0027] At the same time, the dual-polarization SAR image is subjected to polarization parameter decomposition to calculate the Alpha parameter (i.e., polarization scattering angle parameter) of the dual-polarization SAR image, thereby obtaining the Alpha parameter image (i.e., polarization scattering angle parameter image) corresponding to the dual-polarization SAR image according to the Alpha parameter of the dual-polarization SAR image. Different from the related art in which the features of the dual-polarization SAR image are obtained through iterative learning of a neural network, the Alpha parameter in this embodiment is derived strictly mathematically and has direct physical meaning.

[0028] It should be noted that the value of each pixel point in the HH polarization intensity image corresponding to the dual-polarization SAR image is the intensity information of the HH polarization extracted from the dual-polarization SAR image, and the value of each pixel point in the HV polarization intensity image corresponding to the dual-polarization SAR image is also the intensity information of the HV polarization extracted from the dual-polarization SAR image. Similarly, the value of each pixel point in the Alpha parameter image corresponding to the dual-polarization SAR image is also the calculated Alpha parameter of the dual-polarization SAR image.

[0029] For example, Figure 2 As shown in the figure, on the one hand, the ALOS-1 satellite raw data (i.e., the dual-polarization SAR image collected by the ALOS-1 satellite) is converted into single-view complex data, and then the single-view complex data is pre-processed by multi-view processing, refined Lee filtering, and terrain correction in sequence to obtain the intensity information of two different polarizations, HH and HV, of the dual-polarization SAR image, thereby obtaining the HH polarization intensity image and HV polarization intensity image corresponding to the dual-polarization SAR image. It is worth mentioning that the intensity information of HH polarization and HV polarization represents the energy density returned by the SAR signal, which is proportional to the square of the amplitude of the SAR signal. The intensity of the SAR signal not only takes into account the amplitude of the reflected signal, but also reflects the energy of the received signal, which can well display the scattering characteristics of the ground object target.

[0030] On the other hand, the polarization coherence matrix of the original data of the ALOS-1 satellite is extracted, and then refined Lee filtering is performed on the polarization coherence matrix, and the SRTM (Shuttle Radar Topography Mission) data is used to perform terrain correction on the polarization coherence matrix. Finally, the Cloude-Pottier (H / A / Alpha) polarization parameter decomposition of the polarization coherence matrix is ​​performed to obtain the polarization scattering entropy (H), anisotropy (A) and Alpha parameters of the dual-polarization SAR image, thereby obtaining the Alpha parameter image corresponding to the dual-polarization SAR image to reflect the physical scattering mechanism of the ground scatterer (that is, the ground target as a scatterer), such as surface scattering, volume scattering and secondary scattering. Different types of ground targets often correspond to different scattering mechanisms, and then the Alpha parameter image corresponding to the dual-polarization SAR image is used to obtain rich information about the scattering characteristics of the ground target.

[0031] Step 103: Determine a synthetic image corresponding to the dual-polarization synthetic aperture radar image according to the first polarization intensity image, the second polarization intensity image, and the polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image.

[0032] In this embodiment, the HH polarization intensity image, the HV polarization intensity image and the Alpha parameter image corresponding to the dual-polarization SAR image are synthesized into a three-channel false color image (synthetic image), so that in the subsequent steps, a lithology data set can be established according to the synthetic image corresponding to the dual-polarization SAR image, so as to increase the semantic information contained in the lithology data set by using the three characteristic parameters of the intensity information of HH polarization and HV polarization and the Alpha parameter, and further enable the lithology segmentation model to learn more advanced features when iteratively training the lithology segmentation model according to the lithology data set.

[0033] It can be understood that, among the multiple dual-polarization SAR images collected, each dual-polarization SAR image corresponds to a set of HH polarization intensity images, HV polarization intensity images and Alpha parameter images, and each dual-polarization SAR image corresponds to a synthetic image.

[0034] It should be noted that the HH polarization intensity image, HV polarization intensity image, Alpha parameter image and synthetic image corresponding to each dual-polarization SAR image have a one-to-one correspondence with the pixels of the dual-polarization SAR image.

[0035] For example, Figure 2As shown in FIG. 1 , the HH polarization intensity image, the HV polarization intensity image and the Alpha parameter image corresponding to the dual-polarization SAR image are resampled, cropped, linearly stretched and the like, and then the HH polarization intensity image, the HV polarization intensity image and the Alpha parameter image corresponding to the dual-polarization SAR image are synthesized into a false color image to obtain a synthetic image corresponding to the dual-polarization SAR image.

[0036] Further, as a refinement and extension of the specific implementation method of the above embodiment, in order to fully illustrate the specific implementation process of this embodiment, polarization parameter decomposition is performed on the dual-polarization SAR image to obtain a polarization scattering angle parameter image corresponding to the dual-polarization SAR image, including: extracting the polarization coherence matrix of the dual-polarization SAR image; performing eigenvalue decomposition on the polarization coherence matrix to obtain the eigenvector of the polarization coherence matrix and the eigenvalue corresponding to the eigenvector; according to the eigenvector and the eigenvalue, the polarization scattering angle parameter of the dual-polarization SAR image is obtained, and the polarization scattering angle parameter image corresponding to the dual-polarization SAR image is determined according to the polarization scattering angle parameter of the dual-polarization SAR image.

[0037] For example, electromagnetic waves are emitted by the transmitting antenna of SAR, reach the ground object target and are received. After the ground object scatterer absorbs part of the energy of the incident wave, it reflects the remaining part back in the form of a new electromagnetic wave (scattered wave) and is received by the receiving antenna of SAR. Among them, ground object targets can be divided into two categories: coherent targets (coherently scattering target) and distributed targets (distributed target). Specifically, if the electromagnetic scattering characteristics of the ground object target can be fully described by a Sinclair matrix (polarization scattering matrix), it is called a coherent target, and there is a linear relationship between its incident wave and scattered wave, which can be fully described by the Sinclair matrix. For SAR, the pixel size is larger than the wavelength, that is, the ground object target recorded in a single pixel is composed of multiple independent sub-scatterers. This ground object target is called a distributed target. The scattering of the ground object target has a certain randomness and can be described by the Sinclair matrix after spatial averaging.

[0038] Furthermore, if the ground scatterer is illuminated by a surface wave, the incident wave on the ground scatterer is: ,in, is the incident wave emitted by the transmitting antenna of the SAR radar to the ground target. Divided into two vectors in the horizontal direction and the vertical direction, namely and are the incident waves in the horizontal and vertical directions respectively; and They are the linear polarization bases in the horizontal and vertical directions respectively.

[0039] The incident wave is scattered back by the ground scatterer, and the whole scattering process can be expressed as: .in, is the scattered wave from the ground scatterer received by the SAR receiving antenna. r is the distance between the ground scatterer and the receiving antenna, k is the wave number of the SAR electromagnetic wave, S The matrix is ​​the polarization scattering matrix, which is a complex matrix. and are the scattered waves in the horizontal and vertical directions respectively, j is the imaginary unit in complex numbers, , , , are the four elements at the corresponding positions in the polarization scattering matrix S, which respectively refer to different polarization combinations and represent the scattering measurement results between specific polarization states. is the scattering coefficient of horizontal transmission and horizontal reception (Horizontal-Horizontal), is the scattering coefficient for horizontal transmission and vertical reception, is the scattering coefficient for vertical transmission and horizontal reception, is the scattering coefficient for vertical transmission and vertical reception.

[0040] Then, the polarization scattering matrix S Vectorized, expressed as an equivalent four-dimensional complex scattering vector , using the Pauli basis to expand the polarization scattering matrix S , and obtain the corresponding Pauli polarization scattering vector According to the reciprocity theorem, the polarization scattering matrix S is complex symmetric, that is , thus Transformed into a scattering vector matrix , expressed as: , Next, the scattering vector matrix Its conjugate transposed vector After performing the outer product and spatial averaging, the polarization coherence matrix is ​​obtained , used to express the characteristics of SAR signals, expressed as: .

[0041] Furthermore, the polarization coherence matrix Perform eigenvalue decomposition to decompose it into different scattering mechanisms (eigenvectors) and their corresponding relative amplitudes (eigenvalues), expressed as: , ,in, is the polarization coherence matrix after eigenvalue decomposition , ( ) is the feature vector, for The corresponding eigenvalues ​​are For Then, the characteristic vector Combined into a 3 by 3 matrix, expressed as: .in, is the scattering angle, indicating the main direction of the scattering mechanism, The value of is between 0 and 90 degrees, indicating the degree of anisotropy of the scattering process; is the auxiliary angle, used to further describe the directionality of the scattering mechanism; is the phase angle, which indicates the phase change during the scattering process; is another phase angle, related to the orientation of the polarization ellipse; is the component of the scattering mechanism in the main direction; , and is the component of the scattering mechanism in the auxiliary direction and contains phase information. and is the component of the scattering mechanism in another auxiliary direction, which also contains phase information.

[0042] Next, As the polarization parameter description operator, the probability average Alpha parameter can be obtained , expressed as: ,in, Thus, the Alpha parameter image corresponding to the dual-polarization SAR image is obtained according to the Alpha parameter of the dual-polarization SAR image.

[0043] In addition, the dispersion characteristics of the polarization state can be expressed by a scalar metric, namely the polarization scattering entropy (H), which is expressed as: .

[0044] Here, the Alpha parameter is the key parameter for identifying the main scattering mechanism and is rotationally invariant. When Alpha is close to 0, the scattering mechanism is mainly geometric optical surface scattering. When it is around 45, it is mainly dipole scattering and single scattering. When it is close to 90, it is mainly dihedral scattering on the metal surface. The polarization scattering entropy H describes the statistical disorder of different scattering mechanisms as a whole and is rotationally invariant. It reflects the statistical characteristics of the energy of the three scattering mechanisms. When H is low, it is weak depolarization, and the scattering mechanism can be equivalent to the main scattering mechanism, and other mechanisms can be ignored. When H is high, it is depolarization, and there is no single scattering mechanism. Therefore, the use of polarization scattering entropy H and Alpha parameters can explain the depolarization characteristics and polarization scattering characteristics of ground objects, such as It is high entropy multiple scattering, and the ground objects are mostly forest branches and tree crowns.

[0045] In one embodiment, the lithology segmentation method for vegetation coverage areas based on a polarization scattering angle fusion network also includes: cutting the synthetic image corresponding to the dual-polarization SAR image to obtain image samples, and performing data enhancement on the image samples, and determining the lithology data set based on the image samples after data enhancement.

[0046] Specifically, when cutting the synthetic image corresponding to the dual-polarization SAR image, the cutting window size is a square with a width of 640 pixels and a height of 640 pixels, and the windows do not overlap, so as to reflect the generalization ability of the lithology segmentation model as realistically as possible.

[0047] Then, the cut image samples are randomly divided into training set and validation set in a ratio of 9:1, and the training set and validation set are used as lithology data sets. The training set is used for the lithology segmentation model to learn lithology characteristics, and the validation set is used to test the performance of the lithology segmentation model for unknown samples, and to test the performance of the lithology segmentation model in spatial span. Random division is to ensure the randomness of the samples and to ensure that the judgment of the performance of the lithology segmentation model will not be affected by human subjective factors. Here, the image samples in the training set are used as training samples for the lithology segmentation model.

[0048] In addition, before the training samples are input into the lithology segmentation model, data enhancement can be performed on the training samples, including Random Resize, Random Crop, Random Flip, etc. Among them, when the training samples are Random Crop, the training samples are randomly cropped to a size of 512×512, and after the subsequent data enhancement method, they are input into the lithology segmentation model with this size to increase the diversity of the training set of the lithology segmentation model, so that any pixel in the training sample will have different context information each time it is input into the lithology segmentation model, so that the lithology segmentation model can learn richer semantic information.

[0049] In this embodiment, through data enhancement, the dependence of the rock segmentation model on training data can be reduced, the risk of overfitting can be reduced, and the generalization ability of unseen data can be improved. At the same time, the robustness of the rock segmentation model to lighting, angle, occlusion, scale, and noise can be enhanced, thereby improving its performance in complex environments.

[0050] In one embodiment, the lithology segmentation method for vegetation coverage areas based on a polarization scattering angle fusion network also includes: labeling the measured geological map of the vegetation coverage area according to preset lithology categories to determine the true lithology category of each pixel in the measured geological map, and determining the labeled measured geological map as a lithology true value map.

[0051] Here, the measured geological map of the vegetation coverage area also corresponds one-to-one with the pixels of the dual-polarization SAR image, so that the lithology segmentation model can be efficiently trained using the lithology truth map and the lithology dataset. The lithology truth map serves as the "truth" or ideal output data of the lithology segmentation model, and each pixel in it has a clear label of the true lithology category.

[0052] It should be noted that in the related art, measured general geological maps are used as the true value for lithology segmentation. In the measured general geological maps, the lithology composition depicted is extremely comprehensive. Specifically, the measured general geological maps divide the rock types according to the lithological characteristics of the rocks, geological ages, etc. Even after grouping, there are dozens or even nearly a hundred categories in the same area. Although such a classification method describes the geological conditions more comprehensively, it greatly increases the difficulty of lithology segmentation. In addition, the spatial distribution of geological elements is usually complex and highly fragmented. They have high similarity and severe homogeneity, and each category contains several types of rocks, which increases the complexity of the characteristics. In addition, measured general geological maps usually use field sampling to draw the approximate range. For some typical geological categories with small distribution and difficult to observe on the map, their distribution areas will be abstracted into larger areas. Therefore, if the measured general geological map is directly used as the true value for lithology segmentation, the method that is beneficial to mapping will result in erroneous results as the true value basis for learning the lithology segmentation model, and the reliability of lithology segmentation cannot be guaranteed.

[0053] In this embodiment, several lithology categories with the largest area in the vegetation coverage area are selected as preset lithology categories, and all remaining lithology categories are used as a background category, also as preset lithology categories. By using the widely distributed preset lithology categories as segmentation objects of the lithology segmentation model, the feature complexity and the difficulty of lithology segmentation are reduced, and the segmentation error is weakened.

[0054] Exemplarily, the top three lithology categories in the vegetation coverage area and the background category in the vegetation coverage area are used as four preset lithology categories. Here, all lithology categories except the top three lithology categories in the vegetation coverage area are used as a background category.

[0055] Step 104: construct a lithology segmentation model based on a neural network.

[0056] Among them, the lithology segmentation model is used to extract the lithology characteristics in the synthetic image corresponding to the dual-polarization synthetic aperture radar image, and the characteristics of the polarization scattering angle parameter in the polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image, and fuse the lithology characteristics with the characteristics of the polarization scattering angle parameter image.

[0057] In this embodiment, the lithology segmentation model is a deep learning network guided by polarization decomposition theory. Specifically, the lithology segmentation model is composed of an encoder and a decoder with a skip connection structure, which can capture the characteristic information of lithology at different scales. It is worth mentioning that the encoder has a dual-branch structure, the main branch (i.e., the first branch) extracts lithology features according to the intensity information of HH polarization and HV polarization contained in the synthetic image corresponding to the dual-polarization SAR image and the Alpha parameter, and the Alpha branch (i.e., the second branch) extracts the features of the Alpha parameter (i.e., the features of the polarization scattering angle parameter) according to the Alpha parameter image corresponding to the dual-polarization SAR image, and then dynamically fuses the lithology features with the features of the Alpha parameter in the main branch, and then outputs them to the decoder, and inputs them back to the main branch to continue to extract more advanced semantic information, thereby ensuring that the important information of the Alpha parameter will not be lost during the training process, so that the lithology segmentation model learns more stable and more physically meaningful features under the guidance of the Alpha parameter, so that the lithology segmentation model has good spatiotemporal generalization, and improves the accuracy and efficiency of lithology feature recognition in polarization SAR images of vegetation coverage areas. Therefore, the network structure of the lithology segmentation model can also be called a polarization scattering angle fusion network.

[0058] Step 105 , iteratively training the lithology segmentation model according to the lithology data set and the polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image.

[0059] Among them, the lithology dataset is determined based on the synthetic image corresponding to the dual-polarization synthetic aperture radar image.

[0060] In this embodiment, based on the structural characteristics of the lithology segmentation model, the lithology segmentation model is iteratively trained according to the lithology dataset and the Alpha parameter image corresponding to the dual-polarization SAR image. While the lithology dataset containing the intensity information of HH polarization and HV polarization and the Alpha parameter is input into the lithology segmentation model for learning, the Alpha parameter image corresponding to the dual-polarization SAR image is used to additionally retain the semantic information of the Alpha parameter as much as possible, so that the lithology segmentation model can learn more stable and more physically meaningful features under the guidance of the Alpha parameter, thereby improving the reliability of the lithology segmentation model.

[0061] Further, as a refinement and extension of the specific implementation methods of the above-mentioned embodiment, in order to fully illustrate the specific implementation process of this embodiment, the lithology segmentation model is iteratively trained according to the lithology data set and the polarization scattering angle parameter image corresponding to the dual-polarization SAR image, including: inputting the training sample and the polarization scattering angle parameter image sample corresponding to the training sample into the first branch and the second branch of the encoder respectively, to obtain the fusion feature output by the encoder, and inputting the fusion feature into the decoder to obtain the predicted lithology category of the pixel in the training sample output by the decoder; determining the target loss function of the lithology segmentation model according to the predicted lithology category of the pixel in the training sample and the actual lithology category corresponding to the pixel in the training sample in the lithology truth map; calculating the target loss function through the focal loss function and the dice loss function, and updating the model parameters of the lithology segmentation model until the target loss function converges.

[0062] Among them, the training samples are determined according to the image samples in the lithology data set, the polarization scattering angle parameter image samples corresponding to the training samples are determined according to the polarization scattering angle parameter images corresponding to the dual-polarization SAR images corresponding to the training samples, and the lithology true value map is determined according to the measured geological map of the vegetation coverage area.

[0063] It can be understood that the training samples are image samples in the lithology data set, and the image samples are obtained according to the synthetic images corresponding to the dual-polarization SAR images, so the training samples are determined according to the synthetic images corresponding to the dual-polarization SAR images. In order to extract the features of the Alpha parameters in the training samples separately, before training the lithology segmentation model, it is necessary to determine the area corresponding to the training sample position in the Alpha parameter image corresponding to the dual-polarization SAR image corresponding to the training sample, that is, the area corresponding to the pixels in the training sample one by one, so as to obtain the Alpha parameter image sample (polarization scattering angle parameter image sample) corresponding to the training sample. Then, after the training sample and the Alpha parameter image sample corresponding to the training sample are input into the main branch and the Alpha branch respectively, the features extracted by the main branch and the Alpha branch can correspond to each other, thereby ensuring the reliability of the lithology segmentation model training.

[0064] Specifically, Figure 3As shown, training samples are obtained from the lithology data set, and the training samples and the Alpha parameter image samples corresponding to the training samples are input into the encoder in the lithology segmentation model. The training samples are input into the main branch of the encoder to extract multi-scale lithology features according to the intensity information of HH polarization and HV polarization contained in the training samples and the Alpha parameters, and the Alpha parameter image samples corresponding to the training samples are input into the Alpha branch of the encoder to extract the multi-scale features of the Alpha parameters separately, and the multi-scale features of the Alpha parameters are dynamically fused with the multi-scale lithology features in the main branch, so as to obtain different levels of fusion features at four different depths of the encoder, and then each fusion feature is re-input back into the main branch to extract higher-level semantic information, successfully avoiding the loss or mixing of Alpha parameter information, and improving the accuracy of lithology feature recognition in polarimetric SAR images of vegetation coverage areas.

[0065] At the same time, based on the skip connection structure between the encoder and the decoder, the encoder can output different levels of fusion features to the decoder at its four different depths, so that the decoder can capture the characteristic information of training samples at different scales, thereby improving the accuracy of lithology segmentation.

[0066] Then, the decoder outputs the predicted lithology category of each pixel in the training sample based on the fused features output by the encoder.

[0067] Next, the AdamW (Adaptive Moment Estimation weight decay) algorithm is used as the optimization function of the lithology segmentation model, and the target loss function of the lithology segmentation model is determined according to the predicted lithology category of the pixels in the training samples and the actual lithology category corresponding to the pixels in the training samples in the lithology truth map. The target loss function is calculated by Focal loss and Dice loss until the weighted loss function converges.

[0068] Further, as a refinement and extension of the specific implementation method of the above embodiment, in order to fully illustrate the specific implementation process of this embodiment, the training samples and the polarization scattering angle parameter image samples corresponding to the training samples are respectively input into the first branch and the second branch of the encoder to obtain the fusion features output by the encoder, including: sorting the preset levels to obtain the level order, and taking the preset level at the first position in the level order as the output level; the training samples and the polarization scattering angle parameter image samples corresponding to the training samples are respectively input into the first branch and the second branch to obtain the lithological features output by the first branch at the output level and the features of the polarization scattering angle parameters output by the second branch at the output level, and the lithological features output by the first branch and the second branch at the output level are taken as the output level. The features of the lithological characteristics and the polarization scattering angle parameters are input into the feature fusion module corresponding to the output level in the first branch to obtain the fused features corresponding to the output level; if the output level is not the last preset level in the hierarchical order, the fused features corresponding to the output level are input into the next level of the feature fusion module corresponding to the output level in the first branch, and the output level is updated according to the preset level next to the output level in the hierarchical order, and the lithological characteristics output by the first branch and the second branch at the updated output level and the features of the polarization scattering angle parameters are input into the feature fusion module corresponding to the updated output level in the first branch to obtain the fused features corresponding to the updated output level, until the updated output level is the last preset level in the hierarchical order.

[0069] In this embodiment, after the training sample and the polarization scattering angle parameter image sample corresponding to the training sample are respectively input into the main branch and the Alpha branch, the main branch sequentially outputs the lithology features at different preset levels, so as to retain the detail information at the shallower level and extract the high-level semantic information at the deeper level, thereby capturing the characteristic information of the lithology at different scales. When the main branch outputs the lithology features at each preset level, the Alpha branch synchronously outputs the features of the Alpha parameters at the corresponding level to extract the multi-scale features of the Alpha parameters separately. Then, the lithology features output by the main branch and the Alpha branch at the same preset level and the features of the Alpha parameters are input into the feature fusion module connected to the preset level in the main branch, so that the two features are dynamically fused in the feature fusion module corresponding to the preset level, and the fusion features output by the feature fusion module connected to the preset level are obtained, so as to ensure that the physical meaning of the Alpha parameter (such as the scattering mechanism) is not overwhelmed by the complex calculation of the main branch, enhance the stability of the lithology segmentation model for the extraction of lithology features in the vegetation coverage area, and enable the lithology segmentation model to gradually combine the guidance information of the Alpha parameter during training, so as to maintain a high accuracy in the face of temporal and spatial differences.

[0070] It is worth mentioning that in this embodiment, after the lithology features output by the main branch at each preset level are fused with the features of the Alpha parameters of the corresponding level of the Alpha branch, the fused features are output to the decoder for lithology segmentation, and then the fused features are re-input into the main branch to continue to be used for calculation of the main branch to participate in the extraction of lithology features of subsequent preset levels in the main branch, so that the main branch can continue to extract higher-level semantic information according to the fused features until all preset levels are processed, thereby realizing the dynamic fusion of the multi-scale features of the Alpha parameters and the multi-scale lithology features, and ensuring that the physical meaning of the Alpha parameters runs through the entire network, thereby improving the accuracy and generalization ability of the lithology segmentation model for lithology segmentation in vegetation covered areas, and solving the problems of difficulty in extracting lithology features and low lithology classification accuracy under vegetation coverage conditions.

[0071] For example, Figure 3 As shown in the figure, the main branch is preset to output lithological features at the 2nd, 4th, 10th and 12th layers (preset layers) respectively. First, these preset layers are sorted according to the order of neural network processing to obtain the hierarchical order. Then, the second layer, which is at the first position in the hierarchical order, is used as the current output layer. When the main branch outputs lithological features at its second layer, the Alpha branch synchronously outputs the features of the Alpha parameters at its second layer, and then these two features are input into the feature fusion module connected to the second layer in the main branch, so that the two features are dynamically fused in the feature fusion module corresponding to the second layer in the main branch to obtain the fusion features corresponding to the second layer.

[0072] Next, it is determined whether the current output level is the last preset level (i.e., the preset level at the end of the level order). For the current output level being the 2nd level, it is not the last preset level, therefore, the fused features corresponding to the 2nd level are re-input into the next level of the feature fusion module corresponding to the 2nd level in the main branch to continue to participate in the feature extraction of subsequent levels in the main branch, and the fused features corresponding to the 2nd level are input into the decoder for lithology segmentation. Next, the 4th level (i.e., the preset level at the next level of the current output level in the level order) is updated to the current output level, and then the lithology features output by the main branch at its 4th level and the features of the Alpha parameters output by the Alpha branch at its 4th level are input into the feature fusion module corresponding to the 4th level in the main branch to obtain the fused features corresponding to the 4th level, and so on, until the main branch outputs the fused features in the feature fusion module corresponding to its 12th level (i.e., the preset level at the end of the level order). It can be understood that the fused features corresponding to each output level are the fused features output by the encoder at different depths.

[0073] Further, as a refinement and extension of the specific implementation of the above embodiment, in order to fully illustrate the specific implementation process of this embodiment, in the feature fusion module, the fusion feature is obtained according to the following formula: ; ; in, is the characteristic of the Alpha parameter, For the lithological characteristics, and All of them represent the calculations after one convolution layer and one batch normalization layer. It means that after a batch normalization layer, Gaussian error activation layer and convolution layer calculation, Represents addition between elements; represents element-wise multiplication, For fusion features.

[0074] In this embodiment, if Figure 4 As shown, the lithological features of the input feature fusion module Characteristics with Alpha parameter First, after passing through the convolution layer and BN (Batch Normalization) layer, and then through the Element-wise Multiplication (element-wise multiplication), convolution layer, BN layer and Sigmoid function, the feature is obtained. M .feature M For studying lithological characteristics Characteristics with Alpha parameter The possibility of belonging to the same object, and adjusting the lithological characteristics Characteristics with Alpha parameter If the feature M High, indicating more confidence in lithological characteristics ; The same goes for the reverse. Then add 1- M Characteristics with Alpha parameter ,feature M Lithological characteristics After each Element-wise Multiplication, Element-wise Addition is performed to fully integrate them and obtain intermediate features, so that the features can be used to MThe value of dynamically fuses the features from the main branch and the Alpha branch to effectively combine the context information while retaining the detail information. The intermediate features are passed through the BN layer, the GELU (Gaussian Error Linear Units, Gaussian error activation) layer and the convolution layer, and then combined with Y Perform an Element-wiseAddition to form a residual connection structure to ensure the efficiency of model training and effectively solve the problem of information loss during lithology segmentation. Finally, the fusion feature Y' , it is output to the encoder and then re-input into the main branch to continue training to extract more advanced semantic information.

[0075] In one embodiment, the lithology segmentation method of vegetation coverage area based on polarization scattering angle fusion network also includes: Figure 3 As shown in the figure, the main branch includes the backbone (main architecture) and feature fusion module of SwinTransformer (Shifted Window Transformer, Swin Transformer). Specifically, the backbone of SwinTransformer includes PatchPartition and Linear Embedding, Patch Merging and Swin Block.

[0076] Among them, Swin Transformer is a new type of visual Transformer, which can be used as a general backbone network for computer vision tasks. It solves the challenges of traditional Transformer applications in the field of vision by constructing hierarchical feature maps and having a computational complexity that is linearly related to the size of the input image. Among them, Patch Partition and Linear Embedding divide the input image into non-overlapping patches of fixed size (patches), and convert the pixels of each patch into a feature vector through a linear layer. This process can be regarded as converting the image from pixel space to feature space. Patch Merging In the deep stage of the model, the feature vectors of multiple patches are merged into one feature vector, which reduces the spatial size of the feature map and increases the number of channels, similar to the pooling layer in the convolutional neural network, which plays a role of downsampling. Swin Block is the core module of Swin Transformer, which includes hierarchical attention mechanism (Hierarchical Attention) and shifted window attention (Shifted Window Attention). Each Swin block includes a multi-head self-attention mechanism and a multi-layer perceptron, so that the lithology segmentation model can be efficiently modeled locally and globally.

[0077] In one embodiment, the lithology segmentation method of vegetation coverage area based on polarization scattering angle fusion network also includes: Alpha branch is composed of Alpha Extractor (Alpha extractor, i.e. polarization scattering angle parameter extractor) and HWD (Haarwavelet Downsampling). Specifically, Figure 5 As shown, Alpha Extractor extracts the features of the Alpha parameters by stacking convolutional layers, BN layers, and GELU layers, and retains low-level features through concat layers to make the extracted features richer. Figure 3 As shown in the figure, a HWD is connected after each Alpha Extractor for 1 / 2 downsampling to preserve the semantic information of the Alpha parameter as much as possible while downsampling the features of the Alpha parameter.

[0078] Among them, HWD is a simple but effective downsampling method for semantic segmentation tasks. It uses Haar wavelet transform to reduce the spatial resolution of feature maps, while minimizing the information loss in the downsampling process to retain more details (such as boundaries and textures) and context information, and improve the accuracy of pixel-level prediction. It should be noted that Figure 5 Medium Feature Map XFor the calculated Alpha parameter image sample, Alpha Extractor converts the feature map X Calculated as feature map X' , then the feature map X' The characteristics of the Alpha parameter obtained after HWD X'' .

[0079] In one embodiment, the lithology segmentation method of vegetation coverage area based on polarization scattering angle fusion network also includes: Figure 3 As shown in the figure, the decoder includes UPer Head (Unified Perceptual Parsing Head), Kernel Update Head and FCN Head (Fully - Convolutional Network Head).

[0080] It should be noted that FCNHead is an auxiliary head and only participates in the training of the lithology segmentation model. By adding the auxiliary head, the network can learn more information and provide richer gradient information to assist in training. This helps the network with the same number of parameters to learn better and improve the recall rate of the lithology segmentation model during training.

[0081] Specifically, UPerHead comes from the decoder part of UPerNet (Unified Perceptual Parsing Network) for semantic segmentation tasks. It is a multi-task network framework for UPP (Unified Perceptual Parsing) tasks. UPerNet uses the FPN (Feature Pyramid Network) structure, which can use multi-level feature representation and perform feature fusion in a pyramid hierarchy. The design of UPerNet not only improves the efficiency of the model, but also enhances the ability of the lithology segmentation model to understand the rich visual information in the image. Kernel Update Head is a key component in the K-Net (Kernel Network) framework. It segments semantic categories through a set of learnable kernels, where each kernel is responsible for a category and adaptively updates the convolution kernel so that each kernel can dynamically adjust according to the characteristics of the input image. This dynamism enhances the adaptability of the lithology segmentation model to different objects and scenes and improves segmentation accuracy. Through the iterative kernel update process, Kernel Update Head can gradually improve the expressiveness of features, making the lithology segmentation model perform better in complex scenarios.

[0082] In one embodiment, the lithology segmentation method of vegetation coverage area based on polarization scattering angle fusion network also includes: the target loss function is expressed as: ; in, is the target loss function, is the focal loss function, is the weight coefficient of the focal loss function, is the dice loss function, is the weight coefficient of the dice loss function; The focal loss function is expressed as: ; in, N is the number of pixels in the training sample, i is the number of pixels in the training sample, is the pixel in the training sample i The true lithology category in the lithology truth map The corresponding preset weights, Predict pixels in training samples for lithology segmentation model i Belongs to its true lithology category The probability of is the adjustment factor for sample difficulty; The dice loss function is expressed as: ; in, is the set of true lithology categories corresponding to the pixels in the training sample in the lithology truth map, is the set of predicted lithology categories of pixels in the training sample, express W and V The intersection between | W | and | V |Respectively W and V The number of elements in .

[0083] In this embodiment, on the one hand, using (Dice loss function) is used to deal with the problem of unbalanced lithology categories in lithology segmentation. Based on the Dice coefficient, it effectively optimizes the lithology segmentation effect by minimizing the difference between the predicted lithology category and the actual lithology category, accurately outlines the boundaries of lithology segmentation, and improves the accuracy of lithology segmentation of remote sensing images. In addition, its calculation method provides stable gradient information for lithology segmentation model training, which helps the lithology segmentation model converge faster and reduce oscillations during training, thereby improving training efficiency and stability.

[0084] On the other hand, multi-classification Focal loss is used to further solve the problem of unbalanced lithology categories. Focal loss is an improvement on the standard Cross Entropy loss function. Multi-classification Focal loss introduces a modulation factor Reduce the number of easily classified samples (i.e. For lithology categories with a large number of samples and easy to classify, the lithology segmentation model will pay less attention to them during the training process. The smaller the value of the modulation factor is, the larger the value of the modulation factor is, the more the lithology segmentation model will pay attention to these categories, thereby effectively balancing the contribution of different lithology categories to model training, which helps the lithology segmentation model to learn more representative and robust features, so that the lithology segmentation model will not overly rely on the characteristics of lithology categories with a large number of samples, but will be able to comprehensively consider the differences between each lithology category, so that the lithology segmentation model can better classify when facing new and unseen data sets, thereby improving the generalization ability of the lithology segmentation model.

[0085] For example, the weighting coefficients of the focal loss function and the dice loss function are set according to the specific conditions such as the vegetation coverage conditions, the characteristics of the training samples, and the lithology categories in the actual application scenario. For example, .

[0086] In one embodiment, the lithology segmentation method for vegetation coverage area based on polarization scattering angle fusion network further includes: updating the model parameters of the lithology segmentation model according to the following formula: ; ; ; ; ; in, For the Model parameters of the lithology segmentation model during the first iteration training; For the t The gradient of the model parameters of the lithology segmentation model during the training iteration; and Respectively t First-order and second-order momentum estimates of the gradient during training iterations; is the learning rate; and are the exponential decay rates of the first-order and second-order momentum estimates, respectively; and Respectively t Corrected bias in first-order and second-order momentum estimates of the gradient during training iterations; A small constant to prevent division by zero; is the weight attenuation coefficient; and Respectively The first-order and second-order momentum estimates of the gradient during the training iteration, For the t Model parameters of the lithology segmentation model during the training iteration.

[0087] In this embodiment, the AdamW algorithm is selected as the optimization function of the lithology segmentation model. AdamW is a variant of the Adam optimizer, which inherits the advantages of the Adam (Adaptive Moment Estimation) algorithm and can adaptively adjust the learning rate according to the gradient of the model parameters, which makes it perform well when dealing with sparse gradients and non-stationary objective functions. The AdamW algorithm contains a momentum term and an adaptive learning rate term. The momentum term improves the stability of the gradient update by estimating the first-order moment of the gradient. The adaptive learning rate term adjusts the size of the learning rate by estimating the second-order moment of the gradient, and prevents the model from overfitting through weight decay, thereby making the lithology segmentation model training more stable and efficient.

[0088] Step 106 , using the trained lithology segmentation model to perform lithology segmentation on the synthetic aperture radar image to be tested in the vegetation coverage area, to obtain a lithology segmentation result of the synthetic aperture radar image to be tested.

[0089] In this embodiment, the stable lithologic characteristics learned by the lithologic segmentation model under vegetation coverage conditions enable the lithologic segmentation model to have good spatiotemporal generalization, so that the trained lithologic segmentation model can be used to perform lithologic segmentation on the SAR image to be tested (i.e., the synthetic aperture radar image to be tested) of the vegetation coverage area to obtain lithologic mapping of the vegetation coverage area with good spatial continuity.

[0090] In one embodiment, the study area is selected as the northern part of Jinzhou City, Liaoning Province. The climate in this area is warm temperate semi-humid continental monsoon climate, with four distinct seasons, large temperature difference, annual average temperature of 7.8℃, and unstable precipitation. The area contains multiple human settlements and has a large area of ​​vegetation coverage. There are many mineral deposits in the area, including silica, iron ore, etc. The schematic diagram of the experimental vegetation coverage area and the Landsat-8 visible light image are shown in Figure 1. Figure 6 As shown in Figure 2, the vegetation coverage of the experimental vegetation coverage area is as follows: Figure 7 Here, the experimental vegetation coverage area is the northern part of Jinzhou City, Liaoning Province, which is covered by vegetation, and the Landsat-8 visible light image refers to the remote sensing image in the visible light band acquired by the land imager on the Landsat-8 satellite.

[0091] Then, three lithology categories with the widest distribution in the experimental vegetation coverage area are selected, and the remaining more than 40 lithology categories with extremely narrow distribution are classified into one category as the background category. The above four lithology categories are used as the segmentation objects of the lithology segmentation task in this embodiment, and the following is obtained: Figure 8 The true lithology map of the experimental vegetation coverage area is shown.

[0092] Among them, Figure 8 As shown, among the three main lithologies, Q4 is the Quaternary system, including sandy clay, clayey sand and gravel; The Early Jurassic includes biotite granite; the Arjnd is Archean, including the Dayingzi Formation, migmatite, mixed biotite plagioclase gneiss, felsic gneiss, biotite monzonite gneiss, amphibolite-magnetite quartzite, and marble lenses. All remaining lithology categories are grouped into one category as the background category.

[0093] Then, if Fig. 9 As shown, different segmentation models in related technologies and the lithology segmentation model described in this application were used to conduct a lithology segmentation mapping experiment on the historical ALOS-1 image of the experimental vegetation coverage area (ALOS-1 image is a dual-polarization SAR image collected by the ALOS-1 satellite) to compare the lithology segmentation results of different segmentation models for the same vegetation coverage area and verify the practical value of the lithology segmentation model described in this application. Among them, Fig. 9 (a) is the lithology truth map corresponding to the historical ALOS-1 image. Fig. 9 (b) mapping the lithology segmentation of historical ALOS-1 images using the lithology segmentation model described in this application, Fig. 9 (c) K-Net’s lithology segmentation mapping of historical ALOS-1 images. Fig. 9 (d) Lithology segmentation mapping of historical ALOS-1 images by Uper-Net Swin (Unified PerceptualParsing Net with Swin, Uper network Swin variant). Fig. 9 (e) Lithology segmentation mapping of historical ALOS-1 images by Uper-Net R50 (Unified Perceptual Parsing Net with ResNet-50Uper, a network R50 variant). Fig. 9 (f) DeepLabV3+ (DeepLab Version 3+, enhanced version of the deep laboratory v3+ network) is used to segment and map the lithology of historical ALOS-1 images. Fig. 9 (g) U-Net (U Network) lithology segmentation mapping of historical ALOS-1 images. Fig. 9 (h) PID-Net (Proportional-Integral-Derivative Network, PID controller network) lithology segmentation mapping of historical ALOS-1 images. Fig. 9 ( i) is the lithology segmentation mapping of historical ALOS-1 images by PSP-Net (Pyramid Scene Parsing Network). Here, the prediction method, mapping method, and related parameters of all segmentation models are kept consistent to ensure fair comparison.

[0094] Through Fig. 9 (b) to Fig. 9 ( i ) shown in the present application and the different segmentation models in the above-mentioned related art for the lithology segmentation mapping of historical ALOS-1 images. Fig. 9 From the comparison of the lithology truth map corresponding to the historical ALOS-1 image shown in (a), it can be seen that in the actual mapping effect, the lithology segmentation model described in this application performs best, K-Net and Uper-Net Swin are slightly worse, and other segmentation models perform poorly. In this embodiment, the area with the most obvious difference is Fig. 9 (a) to Fig. 9 ( i ), it can be observed that the results obtained by the lithology segmentation model described in this application are significantly higher than those of other segmentation models in terms of spatial continuity and prediction accuracy, and relatively continuous and high-precision boundaries are obtained at the boundaries of different lithology categories, which has a high performance in the lithology segmentation task. However, the performance of the above segmentation models with convolutional structures as the backbone is not ideal, especially for Arjnd and When predicting the two categories, the prediction accuracy is very low, and there is also a serious boundary effect. The reason may be that the area is densely covered with vegetation, and the rock structure of this type is relatively complex, resulting in complex semantic information and contextual relationships in the area, and the backbone of the convolutional structure has no advantage in this regard, resulting in the segmentation of the area is not accurate enough and the spatial continuity is also poor. However, the lithology segmentation model described in this application obtained very good segmentation results, indicating that the L-band radar signal penetrates the vegetation to a large extent and brings back enough ground lithology information, and the Swin Transformer extracts these complex features very well. At the same time, with the support of the Alpha branch and the feature fusion module, more stable and generalized features under the guidance of the Alpha parameter are extracted, and better actual mapping effects are obtained, and good performance is still maintained under the conditions of spatiotemporal differences and vegetation coverage.

[0095] In one embodiment, multiple indicators such as Acc (Accuracy), IoU (Intersection over Union), mAc (Mean Accuracy) and mIoU (Mean IoU) are used to fully evaluate the performance of the lithology segmentation model described in the present application, and to perform quality evaluation on the lithology segmentation model described in the present application.

[0096] Here, Acc is obtained by dividing the number of pixels predicted correctly for each lithology category by the number of pixels originally in this lithology category, and is used to quickly evaluate the performance of the lithology segmentation model for a certain lithology category. mAc is the average value of Acc for each lithology category to obtain an overall performance indicator, which helps to balance the performance differences between lithology categories, especially when the lithology categories are unevenly distributed, it can better reflect the overall performance of the lithology segmentation model. IoU is an indicator to measure the relationship between the predicted lithology category and the true lithology category, expressed as: , where P is the total number of pixels predicted to be a certain lithology category, and Q is the number of pixels covered by the lithology category as the true lithology category. It measures the degree of overlap between the predicted area and the true area. In multi-classification tasks, IoU can provide an accurate measure for each lithology category, helping to identify which lithology categories the lithology segmentation model performs better or needs to be improved. mIoU is the average value of IoU for all lithology categories, which provides a comprehensive measure that reflects the segmentation performance of the lithology segmentation model on all lithology categories.

[0097] Specifically, the lithology segmentation model described in this application and the different segmentation models in the above-mentioned related technologies were used to conduct lithology segmentation tests on the first historical data set. The accuracy evaluation results of each model on the validation set of the first historical data set are shown in Table 1. As can be seen from Table 1, for the validation set of the first historical data set, the lithology segmentation model described in this application achieved the best mAc and mIoU, which were 84.14% and 73.87% respectively, which were 2.75% and 3.40% higher than the slightly worse K-Net. At the same time, Table 1 also lists in detail the Acc of each model for each lithology category in the validation set of the first historical data set. Among them, BA is background Acc, QA is Acc, AA is Arjnd Acc, GA is Acc.

[0098] Table 1

[0099] Specifically, the lithology segmentation model described in this application performs better than other segmentation models in each lithology category, especially in the background and Arjnd categories. The lithology segmentation model described in this application has huge advantages. First of all, the segmentation difficulty of these two lithology categories is very high. The background category contains all lithologies except the three main lithologies, about 40 types, and non-lithology categories such as rivers. They are small and complex, and their characteristics vary greatly, which requires higher learning ability of the network; Arjnd is a mixed lithology, and its lithology characteristics are relatively more complex and changeable, requiring the network to extract deeper and more stable features to achieve higher accuracy. Secondly, the characteristic of the rock segmentation model described in the present application is that, starting from the polarization decomposition theory, the Alpha parameter with rich semantic information and physical meaning is obtained through H / A / Alpha decomposition. It is the polarization scattering angle, the most important and representative parameter in the H / A / Alpha decomposition. The difference in its value reflects different scatterer models. The rock segmentation model described in the present application extracts the features of the Alpha parameter separately through the Alpha branch, and merges it with the main branch each time it outputs features, to ensure that such important semantic information will not be lost in the iterative learning process, and the network can extract more stable features under the guidance of the Alpha parameter.

[0100] Furthermore, the lithology segmentation model described in this application trained by the first historical data set and the different segmentation models in the above-mentioned related technologies are used to directly perform lithology segmentation and evaluation on the validation set of the second historical data set to verify the cross-spatial and cross-temporal generalization performance of each model. The accuracy evaluation results of each model on the validation set of the second historical data set are shown in Table 2. Similarly, Table 2 also lists in detail the Acc of each model for each lithology category in the validation set of the second historical data set.

[0101] Table 2

[0102] Here, the first historical data set and the second historical data set are historical data collected at different sampling times in the same sampling area of ​​the experimental vegetation coverage area, and are used to test the performance of different models in both spatial span and time span. For example, the first historical data set can be constructed based on the ALOS-1 image collected in the experimental vegetation coverage area in June 2008, and the second historical data set can be constructed based on the ALOS-1 image collected in the same area as the sampling area of ​​the first historical data set in June 2010. The validation set of the first historical data set is determined based on the first historical data set, and similarly, the validation set of the second historical data set is determined based on the second historical data set.

[0103] Then, the experimental results of the first historical data set and the second historical data set are compared and analyzed. Specifically, first of all, the Acc of a single lithology category of the lithology segmentation model described in this application, as well as mAc and mIoU, are still higher than those of other models, among which mAc and mIoU are 3.48% and 4.26% higher than the slightly worse K-Net, respectively. Among the four lithology categories, Arjnd performs best. Then, compared with the experimental results of the validation set of the first historical data set, the performance of all models has declined to varying degrees, among which the lithology segmentation model described in this application has the smallest fluctuation, with mIoU dropping from 73.87% to 67.10%, and mAc dropping from 84.14% to 80.32%. Among all lithology categories, the accuracy of background and Arjnd has dropped the most, and these two categories are the main reasons for the fluctuation of model performance, while the accuracy fluctuation of the other two categories is relatively small. This is because the background and Arjnd categories have much more complex semantic information than the other two categories, especially the Arjnd rock distribution area has a large area of ​​vegetation coverage, and the dense vegetation interferes with the rock information. At the same time, the space and time between the validation set of the second historical data set and the samples previously involved in the training are different, which places higher requirements on the stability and generalization of the model. Although the performance of the lithology segmentation model described in this application has declined on the validation set of the second historical data set, compared with the significant decline in the performance of other segmentation models in related technologies, it can be considered that it still shows good performance when facing large-scale vegetation coverage and large temporal and spatial differences.

[0104] From the above, it can be seen that although the study area selected in this application has a large amount of vegetation coverage, the L band of ALOS-1 has strong penetration ability and coherence for dense vegetation coverage, so that the information carried by the radar signal after penetrating the vegetation still has the potential for lithology classification. The area covered by this type of area is densely covered with large areas of vegetation. The IoU of the validation set of the first historical data set was 88.42%, and the IoU of the validation set of the second historical data set with temporal and spatial differences was still 82.84%. It can be considered that the radar signal penetrates the vegetation and brings back the lithology information of the surface, and the lithology segmentation model described in this application learns stable lithology features from it and obtains better segmentation results.

[0105] In one embodiment, the ALOS-1 image to be tested in the experimental vegetation coverage area is segmented using the lithology segmentation model described in the present application. Fig.10 A comparison diagram of the lithology segmentation model described in the present application provided in the embodiment of the present application and the ALOS-1 image to be tested is shown. Fig.10 (a) is the lithology true value map corresponding to the ALOS-1 image to be tested. Fig.10 (b) is the lithology mapping result of the ALOS-1 image to be tested by the lithology segmentation model described in this application, Fig.10 (c) is the vegetation coverage corresponding to the ALOS-1 image to be tested. Fig.10 (a) Fig.10 (b) Fig.10 (c) It can be seen that in the case of vegetation coverage, the lithology segmentation model described in this application learns the stable characteristics of lithology and obtains good segmentation results.

[0106] It should be noted that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0107] Furthermore, if Fig.11 As shown, as a specific implementation of the above-mentioned lithology segmentation method for vegetation coverage areas based on polarization scattering angle fusion network, an embodiment of the present application provides a lithology segmentation device 1100 for vegetation coverage areas based on polarization scattering angle fusion network. The experimental lithology segmentation device 1100 for vegetation coverage areas based on polarization scattering angle fusion network includes: an acquisition module 1101, a processing module 1102, a construction module 1103, a training module 1104 and a segmentation module 1105.

[0108] The acquisition module 1101 is used to acquire dual-polarization synthetic aperture radar images of the vegetation coverage area, and the dual-polarization synthetic aperture radar images include first polarization channel data and second polarization channel data; The processing module 1102 is used to pre-process the dual-polarization synthetic aperture radar image to obtain a first polarization intensity image and a second polarization intensity image corresponding to the dual-polarization synthetic aperture radar image, and perform polarization parameter decomposition on the dual-polarization synthetic aperture radar image to obtain a polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image; and Determine a synthetic image corresponding to the dual-polarization synthetic aperture radar image according to a first polarization intensity image, a second polarization intensity image and a polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image; A construction module 1103 is used to construct a lithology segmentation model based on a neural network, wherein the lithology segmentation model is used to extract lithology features in a synthetic image corresponding to the dual-polarization synthetic aperture radar image and features of a polarization scattering angle parameter in a polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image, and fuse the lithology features with features of the polarization scattering angle parameter image; A training module 1104 is used to iteratively train the lithology segmentation model according to the lithology data set and the polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image, wherein the lithology data set is determined according to the synthetic image corresponding to the dual-polarization synthetic aperture radar image; The segmentation module 1105 is used to perform lithology segmentation on the synthetic aperture radar image to be tested in the vegetation coverage area by using the trained lithology segmentation model to obtain the lithology segmentation result of the synthetic aperture radar image to be tested.

[0109] In one embodiment, the processing module 1102 is specifically used to extract the polarization coherence matrix of the dual-polarization synthetic aperture radar image; perform eigenvalue decomposition on the polarization coherence matrix to obtain the eigenvector of the polarization coherence matrix and the eigenvalue corresponding to the eigenvector; obtain the polarization scattering angle parameter of the dual-polarization synthetic aperture radar image according to the eigenvector and the eigenvalue, and determine the polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image according to the polarization scattering angle parameter of the dual-polarization synthetic aperture radar image.

[0110] In one embodiment, the training module 1104 is specifically used to input the training samples and the polarization scattering angle parameter image samples corresponding to the training samples into the first branch and the second branch of the encoder respectively, to obtain the fusion features output by the encoder, the training samples are determined according to the image samples in the lithology data set, and the polarization scattering angle parameter image samples corresponding to the training samples are determined according to the polarization scattering angle parameter images corresponding to the dual-polarization synthetic aperture radar images corresponding to the training samples; the fusion features are input into the decoder to obtain the predicted lithology category of the pixels in the training samples output by the decoder; the target loss function of the lithology segmentation model is determined according to the predicted lithology category of the pixels in the training samples and the actual lithology category corresponding to the pixels in the training samples in the lithology truth map, and the lithology truth map is determined according to the measured geological map of the vegetation coverage area; the target loss function is calculated by the focal loss function and the dice loss function, and the model parameters of the lithology segmentation model are updated until the target loss function converges.

[0111] In one embodiment, the training module 1104 is specifically used to sort the preset levels to obtain the level order, and use the preset level at the first position in the level order as the output level; input the training sample and the polarization scattering angle parameter image sample corresponding to the training sample into the first branch and the second branch respectively, obtain the lithology characteristics output by the first branch at the output level and the characteristics of the polarization scattering angle parameters output by the second branch at the output level, and input the lithology characteristics output by the first branch and the second branch at the output level and the characteristics of the polarization scattering angle parameters into the feature fusion module corresponding to the output level in the first branch to obtain the output level corresponding to the output level. fusion features; if the output level is not the preset level at the last position in the hierarchical order, the fusion features corresponding to the output level are input into the next level of the feature fusion module corresponding to the output level in the first branch, and the output level is updated according to the preset level at the next position in the hierarchical order, and the features of the lithological features and polarization scattering angle parameters output by the first branch and the second branch at the updated output level are input into the feature fusion module corresponding to the updated output level in the first branch to obtain the fusion features corresponding to the updated output level, until the updated output level is the preset level at the last position in the hierarchical order.

[0112] In one embodiment, the training module 1104 is specifically configured to obtain the fusion feature according to the following formula: ; ; in, is the characteristic of the polarization scattering angle parameter, For the lithological characteristics, and All of them represent the calculations after one convolution layer and one batch normalization layer. It means that after a batch normalization layer, Gaussian error activation layer and convolution layer calculation, It means element-by-element addition; represents element-wise multiplication, For fusion features.

[0113] In one embodiment, the training module 1104 is specifically used for the second branch to include a polarization scattering angle parameter extractor and a Haar wavelet down-sampling connected to the polarization scattering angle parameter extractor. The polarization scattering angle parameter extractor extracts the features of the polarization scattering angle parameters in the polarization scattering angle parameter image sample by stacking convolution layers, batch normalization layers, and Gaussian error activation layers, and retains features of preset types in the features of the polarization scattering angle parameters through a splicing layer.

[0114] In one embodiment, the training module 1104 is specifically used to express the target loss function as: ; in, is the target loss function, is the focal loss function, is the weight coefficient of the focal loss function, is the dice loss function, is the weight coefficient of the dice loss function; The focal loss function is expressed as: ; in, N is the number of pixels in the training sample, i is the number of pixels in the training sample, is the pixel in the training sample i The true lithology category in the lithology truth map The corresponding preset weights, Predict pixels in training samples for lithology segmentation model i Belongs to its true lithology category The probability of is the adjustment factor for sample difficulty; The dice loss function is expressed as: ; in, is the set of true lithology categories corresponding to the pixels in the training sample in the lithology truth map, is the set of predicted lithology categories of pixels in the training sample, express W and V The intersection between | W | and | V |Respectively W and V The number of elements in .

[0115] In one embodiment, the training module 1104 is specifically used to update the model parameters of the lithology segmentation model according to the following formula: ; ; ; ; ; in, For the Model parameters of the lithology segmentation model during the first iteration training; For the t The gradient of the model parameters of the lithology segmentation model during the training iteration; and Respectively tFirst-order and second-order momentum estimates of the gradient during training iterations; is the learning rate; and are the exponential decay rates of the first-order and second-order momentum estimates, respectively; and Respectively t Corrected bias in first-order and second-order momentum estimates of the gradient during training iterations; is a constant; is the weight attenuation coefficient; and Respectively The first-order and second-order momentum estimates of the gradient during the training iteration, For the t Model parameters of the lithology segmentation model during the training iteration.

[0116] In one embodiment, the vegetation coverage area lithology segmentation device based on the polarization scattering angle fusion network also includes: A module is established to cut the synthetic image corresponding to the dual-polarization synthetic aperture radar image to obtain image samples, and to perform data enhancement on the image samples, and to determine the lithology data set according to the image samples after the data enhancement.

[0117] Regarding the specific definition of the lithology segmentation device for vegetation-covered areas based on the polarization scattering angle fusion network, please refer to the definition of the lithology segmentation method for vegetation-covered areas based on the polarization scattering angle fusion network mentioned above, which will not be repeated here. Each module in the above-mentioned lithology segmentation device for vegetation-covered areas based on the polarization scattering angle fusion network can be fully or partially implemented by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

[0118] Based on the above Figure 1 The method shown in the embodiment of the present application accordingly provides a readable storage medium on which a computer program is stored, and when the program is executed by a processor, the above-mentioned Figure 1 The rock type segmentation method of vegetation coverage area based on polarization scattering angle fusion network is shown.

[0119] Based on this understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each implementation scenario of the present application.

[0120] Based on the above Figure 1 The method shown, and Fig.11 In order to achieve the above-mentioned purpose, the embodiment of the present application further provides a computer device, which can be a personal computer, a server, a network device, etc. The computer device includes a storage medium and a processor; the storage medium is used to store a computer program; the processor is used to execute the computer program to achieve the above-mentioned Figure 1 The rock type segmentation method of vegetation coverage area based on polarization scattering angle fusion network is shown.

[0121] Optionally, the computer device may also include a user interface, a network interface, a camera, a radio frequency (RF) circuit, a sensor, an audio circuit, a WI-FI module, etc. The user interface may include a display, an input unit such as a keyboard, etc., and the optional user interface may also include a USB interface, a card reader interface, etc. The network interface may optionally include a standard wired interface, a wireless interface (such as a Bluetooth interface, a WI-FI interface), etc.

[0122] Those skilled in the art will appreciate that the computer device structure provided in this embodiment does not constitute a limitation on the computer device, and may include more or fewer components, or a combination of certain components, or different component arrangements.

[0123] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages and saves the hardware and software resources of the computer device, and supports the operation of information processing programs and other software and / or programs. The network communication module is used to realize communication between the components inside the storage medium, and communication with other hardware and software in the physical device.

[0124] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform, or the embodiments of the present application can be implemented by hardware.

[0125] Those skilled in the art will appreciate that the accompanying drawings are only schematic diagrams of a preferred implementation scenario, and the modules or processes in the accompanying drawings are not necessarily necessary for implementing the present application. Those skilled in the art will appreciate that the modules in the devices in the implementation scenario can be distributed in the devices of the implementation scenario according to the description of the implementation scenario, or can be changed accordingly and located in one or more devices different from the present implementation scenario. The modules of the above-mentioned implementation scenario can be combined into one module, or can be further split into multiple submodules.

[0126] The above serial numbers of this application are only for description and do not represent the advantages and disadvantages of the implementation scenarios. The above disclosure is only a few specific implementation scenarios of this application, but this application is not limited to them, and any changes that can be thought of by technicians in this field should fall within the scope of protection of this application.

Claims

1. A lithology segmentation method for vegetation coverage area based on polarization scattering angle fusion network, characterized in that: The method comprises: Collecting a dual-polarization synthetic aperture radar image of a vegetation-covered area, wherein the dual-polarization synthetic aperture radar image includes first polarization channel data and second polarization channel data; Preprocessing the dual-polarization synthetic aperture radar image to obtain a first polarization intensity image and a second polarization intensity image corresponding to the dual-polarization synthetic aperture radar image, and performing polarization parameter decomposition on the dual-polarization synthetic aperture radar image to obtain a polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image; Determine a synthetic image corresponding to the dual-polarization synthetic aperture radar image according to a first polarization intensity image, a second polarization intensity image, and a polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image; A lithology segmentation model is constructed based on a neural network, wherein the lithology segmentation model is used to extract lithology features in a synthetic image corresponding to the dual-polarization synthetic aperture radar image and features of a polarization scattering angle parameter in a polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image, and to fuse the lithology features with features of the polarization scattering angle parameter image; Iteratively training the lithology segmentation model according to the lithology data set and the polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image, wherein the lithology data set is determined according to the synthetic image corresponding to the dual-polarization synthetic aperture radar image; The trained lithology segmentation model is used to perform lithology segmentation on the synthetic aperture radar image to be tested in the vegetation coverage area to obtain a lithology segmentation result of the synthetic aperture radar image to be tested.

2. The lithology segmentation method for vegetation coverage area based on polarization scattering angle fusion network according to claim 1 is characterized in that: The step of performing polarization parameter decomposition on the dual-polarization synthetic aperture radar image to obtain a polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image includes: Extracting a polarization coherence matrix of the dual-polarization synthetic aperture radar image; Performing eigenvalue decomposition on the polarization coherence matrix to obtain an eigenvector of the polarization coherence matrix and an eigenvalue corresponding to the eigenvector; According to the characteristic vector and the characteristic value, the polarization scattering angle parameter of the dual-polarization synthetic aperture radar image is obtained, and according to the polarization scattering angle parameter of the dual-polarization synthetic aperture radar image, the polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image is determined.

3. The lithology segmentation method of vegetation coverage area based on polarization scattering angle fusion network according to claim 1 is characterized in that: The lithology segmentation model includes an encoder and a decoder, the encoder includes a first branch and a second branch, the encoder and the decoder are connected via a jump connection, and the lithology segmentation model is iteratively trained according to the lithology data set and the polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image, including: Inputting a training sample and a polarization scattering angle parameter image sample corresponding to the training sample into the first branch and the second branch of the encoder respectively, to obtain a fusion feature output by the encoder, wherein the training sample is determined according to the image sample in the lithology data set, and the polarization scattering angle parameter image sample corresponding to the training sample is determined according to the polarization scattering angle parameter image corresponding to the dual-polarization synthetic aperture radar image corresponding to the training sample; Inputting the fused features into the decoder to obtain the predicted lithology category of the pixels in the training sample output by the decoder; Determining a target loss function of the lithology segmentation model according to the predicted lithology category of the pixels in the training sample and the actual lithology category corresponding to the pixels in the training sample in the lithology truth map, wherein the lithology truth map is determined according to the measured geological map of the vegetation coverage area; The target loss function is calculated by using the focus loss function and the dice loss function, and the model parameters of the lithology segmentation model are updated until the target loss function converges.

4. The method for lithology segmentation of vegetation coverage area based on polarization scattering angle fusion network according to claim 3 is characterized in that: The step of inputting the training sample and the polarization scattering angle parameter image sample corresponding to the training sample into the first branch and the second branch of the encoder respectively to obtain the fusion feature output by the encoder comprises: Sorting the preset levels to obtain a level order, and taking the preset level at the first position in the level order as the output level; Input the training sample and the polarization scattering angle parameter image sample corresponding to the training sample into the first branch and the second branch respectively, obtain the lithology feature output by the first branch at the output level and the feature of the polarization scattering angle parameter output by the second branch at the output level, and input the lithology feature and the feature of the polarization scattering angle parameter output by the first branch and the second branch at the output level into the feature fusion module corresponding to the output level in the first branch, and obtain the fusion feature corresponding to the output level; If the output level is not the preset level at the last position in the level order, the fused features corresponding to the output level are input into the next level of the feature fusion module corresponding to the output level in the first branch, and the output level is updated according to the preset level at the next position in the level order, and the features of the lithological features and the polarization scattering angle parameters output by the first branch and the second branch at the updated output level are input into the feature fusion module corresponding to the updated output level in the first branch to obtain the fused features corresponding to the updated output level, until the updated output level is the preset level at the last position in the level order.

5. The method for lithology segmentation of vegetation coverage area based on polarization scattering angle fusion network according to claim 4 is characterized in that: In the feature fusion module, the fusion feature is obtained according to the following formula: ； ； in, is the characteristic of the polarization scattering angle parameter, For the lithological characteristics, and All of them represent the calculations after one convolution layer and one batch normalization layer. It means that after a batch normalization layer, Gaussian error activation layer and convolution layer calculation, It means element-by-element addition; represents element-wise multiplication, is the fusion feature.

6. The method for lithology segmentation of vegetation coverage area based on polarization scattering angle fusion network according to claim 4 is characterized in that: The second branch includes a polarization scattering angle parameter extractor and a Haar wavelet down-sampling connected to the polarization scattering angle parameter extractor. The polarization scattering angle parameter extractor extracts the features of the polarization scattering angle parameters in the polarization scattering angle parameter image sample by stacking convolution layers, batch normalization layers and Gaussian error activation layers, and retains features of preset types in the features of the polarization scattering angle parameters through a splicing layer.

7. The method for lithology segmentation of vegetation coverage area based on polarization scattering angle fusion network according to claim 3 is characterized in that: The objective loss function is expressed as: ； in, is the objective loss function, is the focal loss function, is the weighting coefficient of the focal loss function, is the dice loss function, is the weighted coefficient of the dice loss function; The focal loss function is expressed as: ； in, N is the number of pixels in the training sample, i is the number of the pixel in the training sample, is the pixel in the training sample i The true lithology category in the lithology truth map The corresponding preset weights, The lithology segmentation model predicts the pixels in the training sample i Belongs to its true lithology category The probability of is the adjustment factor for sample difficulty; The dice loss function is expressed as: ； in, is the set of real lithology categories corresponding to the pixels in the training sample in the lithology truth map, is the set of predicted lithology categories of pixels in the training sample, express W and V The intersection between | W | and | V |Respectively W and V The number of elements in .

8. The method for lithology segmentation of vegetation coverage area based on polarization scattering angle fusion network according to claim 3 is characterized in that: The model parameters of the lithology segmentation model are updated according to the following formula: ； ； ； ； ； in, For the Model parameters of the lithology segmentation model during the first iteration training; For the t The gradient of the model parameters of the lithology segmentation model during the iteration training; and Respectively t first-order and second-order momentum estimates of the gradient during the training iteration; is the learning rate; and are the exponential decay rates of the first-order and second-order momentum estimates, respectively; and Respectively t The corrected bias of the first-order and second-order momentum estimates of the gradient at the training iteration; is a constant; is the weight attenuation coefficient; and Respectively The first and second order momentum estimates of the gradient during the training iteration, For the t The model parameters of the lithology segmentation model during the iterative training.

9. The method for lithology segmentation of vegetation coverage area based on polarization scattering angle fusion network according to claim 1, characterized in that: The method further comprises: The synthetic image corresponding to the dual-polarization synthetic aperture radar image is cut to obtain an image sample, and data enhancement is performed on the image sample. The lithology data set is determined according to the image sample after data enhancement.

Images