A method and device for filling missing areas of high-precision DEM data

Abstract

The present invention discloses a method and device for filling a region with missing high-precision DEM data, which belongs to the field of photogrammetry and remote sensing technology, and includes: obtaining a high-precision InSAR hole DEM to be filled and a low-resolution DEM consistent with the geographical location of the high-precision InSAR hole to be filled; determining terrain feature data and SAR feature data according to the low-resolution DEM; inputting the low-resolution DEM, terrain feature data and SAR feature data into a pre-trained Transformer and CNN intelligent learning reconstruction model based on multi-source features to obtain a predicted high-resolution DEM of the hole region; combining the predicted high-resolution DEM of the hole region with the high-precision InSAR hole DEM to be filled to generate a high-precision DEM product. The implementation method of the present invention is simple and easy to implement, and the overall process structure is clear, so that the process of hole filling is fully intelligent and automated.

Description

Technical Field

[0001] The invention relates to a method and a device for filling a missing area of ​​high-precision DEM data, and belongs to the technical field of photogrammetry and remote sensing. Background Art

[0002] Digital Elevation Model (DEM) is an important support for the big data architecture of geographic information. It is widely used in many fields such as urban operation management, flood simulation and prediction, landslide and debris flow risk assessment, and real-life three-dimensional construction. With the continuous development of earth science, global digitization and informatization, the construction of high-precision DEM is of great significance. The effective methods for obtaining traditional DEM mainly include topographic map digitization, aerial photogrammetry, laser radar (LiDAR) mapping and other technologies. However, due to the fact that the monitoring mode is easily affected by weather, the monitoring range is limited, and a large amount of manpower and material resources are consumed, its application in the generation of large-scale high-precision DEM is limited. Synthetic Aperture Radar Interferometry (InSAR) technology has been proven to be one of the important remote sensing means for the construction of high-precision DEM. SAR technology has the advantages of all-day, all-weather, and not being blocked by clouds, rain and fog. In addition, the high-resolution satellite-borne SAR system can achieve sub-meter terrain mapping, which can quickly and efficiently capture the point information required for high-precision DEM.

[0003] The methods for generating high-precision DEM by InSAR mainly include two-track method, multi-baseline and dual-satellite interferometry. Since the conventional two-track method is seriously affected by the geometric structure of the SAR platform, spatiotemporal incoherence, atmospheric delay and unwrapping error, the quality of the obtained DEM is low. At present, the main methods are multi-baseline and dual-satellite interferometry. However, since the SAR DEM generation process is limited by factors such as terrain conditions, ground scattering characteristics and phase changes in the imaging area, the generated DEM will have holes due to missing data. The DEM hole problem caused by missing data is also an inherent limitation of optical remote sensing and unmanned aerial vehicle photogrammetry, and is a key factor affecting the current DEM quality.

[0004] At present, the methods for filling DEM holes caused by missing data mainly include single SAR data fusion and multi-source DEM data fusion. However, the accuracy of the reconstruction process of the existing technology needs to be further improved, and the high-frequency information of the data missing area is not fully considered, which easily leads to the problem of inconsistency between the reconstruction result and the terrain. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a method and device for filling the missing area of ​​high-precision DEM data.

[0006] In order to solve the above technical problems, the present invention is implemented by adopting the following technical solutions.

[0007] On the one hand, the present invention discloses a method for filling missing areas of high-precision DEM data, comprising:

[0008] Obtaining a high-precision InSAR hole DEM to be filled and a low-resolution DEM that is consistent with the geographical location of the high-precision InSAR hole DEM to be filled;

[0009] Determine terrain feature data and SAR feature data according to the low-resolution DEM;

[0010] Inputting the low-resolution DEM, the terrain feature data and the SAR feature data into a pre-trained Transformer and CNN intelligent learning reconstruction model based on multi-source features to obtain a predicted high-resolution DEM of the cavity area;

[0011] The processing process of the Transformer and CNN intelligent learning reconstruction model based on multi-source features includes:

[0012] Extract shallow features from low-resolution DEM to generate corresponding shallow feature matrices; perform normalization and feature fusion on terrain feature data and SAR feature data in turn to generate fused matrices; connect the shallow feature matrix and the fused matrix to generate new shallow features; perform attention Transformer and CNN feature mapping on the new shallow features to obtain a deep feature matrix; generate a predicted high-resolution DEM of the hole area based on the low-resolution DEM and the deep feature matrix;

[0013] The predicted high-resolution DEM of the hole area is combined with the high-precision InSAR hole DEM to be filled to generate a high-precision DEM product.

[0014] Furthermore, the process of obtaining the high-precision InSAR hole DEM to be filled includes:

[0015] Acquire multi-view SAR satellite data of the same target area, perform format conversion, generate single-view complex images, pre-process the single-view complex images, generate a number of differential interferograms, unwrap the differential interferograms, and generate unwrap differential interferogram phases;

[0016] Based on formula (1), the high coherence point is selected for the unwrapped differential interference phase to obtain the unwrapped phase at the high coherence point.

[0017] (1);

[0018] Where: Indicates m The first i Highly coherent points ( x , y ) at the unwrapping phase, 1≤ m ≤ M , M represents the total number of differential interferograms generated; Indicates m The first i Highly coherent points ( x , y ), , λ represents the radar wavelength, Δ d ( x , y )express( x , y ) upward line of sight deformation; Indicates m The first i Highly coherent points ( x , y ) orbit error; Indicates m The first i Highly coherent points ( x , y ) at atmospheric delay phase; Indicates m The first i Highly coherent points ( x , y ) at the noise phase; Indicates m The differential interference pattern i Highly coherent points ( x , y ) at the high-pass deformation component; Indicates m The differential interference pattern i Highly coherent points ( x , y ),

[0019] (2);

[0020] Where: R Indicates the distance between the coherent target and the satellite; θ and denote the radar incident angle and m The length of the vertical baseline of the interference pattern; Δ h (x , y ) represents the low-resolution DEM and high-resolution DEM i The difference at the high coherence points (x, y);

[0021] Use formula (3) to generate i Highly coherent points ( x , y ) at the elevation information;

[0022] (3);

[0023] Where: z ( x , y ) indicates the i Highly coherent points ( x , y ), Represents low-resolution DEM i Highly coherent points ( x , y ) at the elevation;

[0024] The elevation information is used to generate a high-precision InSAR cavity DEM that matches the target geographic location through geocoding.

[0025] Furthermore, the preprocessing includes:

[0026] The single-view complex image is subjected to baseline estimation and connection map generation, image registration, interference, flattening, terrain removal and filtering in sequence to generate intensity map, coherence map and differential interferogram. The differential interferogram is unwrapped using the minimum cost flow method to generate the unwrapped differential interferogram phase.

[0027] Furthermore, the training process of the Transformer and CNN intelligent learning reconstruction model based on multi-source features includes:

[0028] Download low-resolution DEM data that matches the actual geographical location of the high-precision InSAR hole DEM to be filled;

[0029] Constructing a DEM training sample set according to the matched low-resolution DEM data;

[0030] The terrain features corresponding to the low-resolution DEM data in the DEM training sample set are calculated using the slope and aspect extraction operators, and a terrain feature training sample set is constructed based on the corresponding terrain features;

[0031] The intensity map and coherence map generated from the single-view complex image are used as SAR features to construct a SAR feature training sample set;

[0032] Constructing a Transformer and CNN intelligent learning and reconstruction model based on multi-source features, wherein the Transformer and CNN intelligent learning and reconstruction model based on multi-source features includes: a shallow feature extraction module, a multi-source feature connection module, a Transformer, a CNN feature mapping module, an upsampling module, and a loss function calculation module;

[0033] The shallow feature extraction module is expressed as:

[0034] ;

[0035] Where: F 0 indicates the extracted shallow features; H SFE Indicates the weight is W 0 The convolutional layer; D LR Represents the input low-resolution DEM; W 0 Indicates the weight of the 3×3 convolution layer during shallow feature extraction; * indicates convolution operation;

[0036] The new shallow feature representation generated by the multi-source feature connection module is:

[0037] ;

[0038] Where: F New Represents the generated new shallow features; W New and b They represent the weight and bias of the 3×3 convolutional layer when multi-source features are fused; F TS It represents the fusion of terrain features and SAR features;

[0039] The feature representation of the output of the ATCB module in the attention transformer and CNN feature mapping module is:

[0040] ;

[0041] Where: F k Indicates k The characteristics of the ATCB module output are: K is the total number of ATCB modules, 1≤ k ≤ K ; , and Respectively represent k , No. k-1. The first ATCB module extracts nonlinear functions of features;

[0042] The feature representation after deep feature fusion is:

[0043] ;

[0044] Where: F D Represents the features after deep feature fusion; W 1 and W 2 represents the weight of the 3×3 convolution layer and the weight of the 1×1 convolution layer when deep features are fused; PReLU represents the activation function; F 1, F 2,…, F K ] represents the characteristics of each ATCB module output;

[0045] The upsampled high-resolution DEM is expressed as:

[0046] ;

[0047] Where: D HR Represents high-resolution DEM; F UP represents the pixel transformation function; W 3 represents the weight of the 3×3 convolutional layer during upsampling;

[0048] The loss function is expressed as:

[0049] ;

[0050] Where: Loss represents the loss function; L Global represents the global loss; L Terrain Indicates loss of terrain features; L SAR Indicates SAR feature loss;

[0051] The Transformer and CNN intelligent learning and reconstruction models based on multi-source features are trained according to the DEM training sample set, the terrain feature training sample set, the SAR feature training sample set and the loss function. When the loss function does not meet the convergence judgment condition, the model is retrained by error back propagation. When the loss function meets the judgment convergence condition, the trained Transformer and CNN intelligent learning and reconstruction models based on multi-source features are output.

[0052] Furthermore, the global loss is expressed as:

[0053] ;

[0054] Where: J Indicates the number of DEM training sample sets; f mask represents the hole mask function; and Respectively represent j The elevation values ​​of high-resolution DEM and InSAR high-precision DEM training samples.

[0055] Furthermore, the terrain feature loss is expressed as:

[0056] ;

[0057] Where: and Respectively represent j The terrain characteristic values ​​generated by a high-resolution DEM and an InSAR high-precision DEM.

[0058] Furthermore, the SAR feature loss is expressed as:

[0059] ;

[0060] Where: Indicates j The pixel value of the SAR feature training sample.

[0061] Furthermore, the calculation formulas for calculating the terrain features corresponding to the low-resolution DEM data in the DEM training sample set using the slope and aspect extraction operators are:

[0062] ;

[0063] ;

[0064] Where: Slope Indicates the slope, Aspect Indicates the slope direction, f x Indicates the change in horizontal distance from the pixel point elevation; f y Indicates the change in vertical distance from the pixel point elevation.

[0065] Furthermore, when constructing the DEM training sample set, the terrain feature training sample set and the SAR feature training sample set, it also includes:

[0066] Use interpolation or resampling methods to make the resolution of low-resolution DEM data meet the multiple requirements of high-resolution DEM data;

[0067] The low-resolution DEM data and terrain feature data are cropped with a pixel size of (X×X)·s, and the SAR feature data are cropped with a pixel size of X×X. For areas that do not meet the X×X pixel size at the boundary, cropping is abandoned to obtain DEM training sample sets, terrain feature training sample sets, and SAR feature training sample sets that are integer multiples of X×X, where X×X represents the pixel size and s is the ratio between the high-resolution DEM and low-resolution DEM resolutions.

[0068] In a second aspect, the present invention discloses a device for filling missing areas of high-precision DEM data, comprising:

[0069] An acquisition module is used to acquire a high-precision InSAR hole DEM to be filled and a low-resolution DEM that is consistent with the geographical location of the high-precision InSAR hole DEM to be filled;

[0070] A determination module, used for determining terrain feature data and SAR feature data according to the low-resolution DEM;

[0071] A prediction module, used for inputting the low-resolution DEM, the terrain feature data and the SAR feature data into a pre-trained Transformer and CNN intelligent learning reconstruction model based on multi-source features to obtain a predicted high-resolution DEM of the hole area;

[0072] The processing process of the Transformer and CNN intelligent learning reconstruction model based on multi-source features includes:

[0073] Extract shallow features from low-resolution DEM to generate corresponding shallow feature matrices; perform normalization and feature fusion on terrain feature data and SAR feature data in turn to generate fused matrices; connect the shallow feature matrix and the fused matrix to generate new shallow features; perform attention Transformer and CNN feature mapping on the new shallow features to obtain a deep feature matrix; generate a predicted high-resolution DEM of the hole area based on the low-resolution DEM and the deep feature matrix;

[0074] The combining module is used to combine the predicted high-resolution DEM of the hole area with the high-precision InSAR hole DEM to be filled to generate a high-precision DEM product.

[0075] The beneficial effects achieved by the present invention are:

[0076] In the present invention, the intelligent learning and reconstruction modeling of low-resolution DEM to high-resolution DEM is realized through the Transformer and CNN intelligent learning and reconstruction model based on multi-source features, thereby improving the accuracy of model training;

[0077] In the present invention, the slope and aspect in terrain features and the intensity and coherence features in SAR features are used as high-frequency information to improve the effectiveness of filling the DEM hole areas caused by data missing.

[0078] In the present invention, by organically combining high-precision InSAR DEM generation and intelligent learning reconstruction modeling, three-dimensional modeling of large-scale high-precision DEM products is achieved. Among them, InSAR technology has the advantages of wide monitoring range, high spatiotemporal resolution, and non-contact measurement, which effectively makes up for the shortcomings in traditional DEM generation. Intelligent learning reconstruction modeling based on deep learning can significantly enhance and improve the quality of DEM void areas. This application broadens the application field of InSAR technology and deep learning, provides important basic data support for geological disasters, environmental protection, soil and water conservation, etc., and provides a theoretical and methodological system that can be used as a reference for DEM modeling problems in geoscience-related fields.

[0079] The implementation scheme of the present invention is simple and easy, has strong practicality, and has a clear overall process structure, which makes the process of void filling fully intelligent and automated, and effectively solves the problems of insufficient practicality and limited application scope of previous technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 This is a flow chart of Embodiment 2 of the present invention;

[0081] Figure 2 A flow chart for constructing a training sample set in step 2 of an embodiment of the present invention;

[0082] Figure 3 This is a flow chart of modeling reconstruction based on intelligent learning of Transformer and CNN with multi-source features in step 3 of an embodiment of the present invention;

[0083] Figure 4 This is a flow chart of the attention Transformer and CNN module ATCB in step 3.3 of an embodiment of the present invention;

[0084] Figure 5 This is a flow chart of the attention structure AM in step 3.3 of an embodiment of the present invention;

[0085] Figure 6 This is the STL flow chart of the Swin Transformer structure in step 3.3 of the embodiment of the present invention. DETAILED DESCRIPTION

[0086] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0087] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and the like are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0088] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood by specific circumstances.

[0089] Embodiment 1: This embodiment introduces a method for filling missing areas of high-precision DEM data, including:

[0090] Obtaining a high-precision InSAR hole DEM to be filled and a low-resolution DEM that is consistent with the geographical location of the high-precision InSAR hole DEM to be filled;

[0091] Determine terrain feature data and SAR feature data according to the low-resolution DEM;

[0092] Inputting the low-resolution DEM, the terrain feature data and the SAR feature data into a pre-trained Transformer and CNN intelligent learning reconstruction model based on multi-source features to obtain a predicted high-resolution DEM of the cavity area;

[0093] The processing process of the Transformer and CNN intelligent learning reconstruction model based on multi-source features includes:

[0094] Extract shallow features from low-resolution DEM to generate corresponding shallow feature matrices; perform normalization and feature fusion on terrain feature data and SAR feature data in turn to generate fused matrices; connect the shallow feature matrix and the fused matrix to generate new shallow features; perform attention Transformer and CNN feature mapping on the new shallow features to obtain a deep feature matrix; generate a predicted high-resolution DEM of the hole area based on the low-resolution DEM and the deep feature matrix;

[0095] The predicted high-resolution DEM of the hole area is combined with the high-precision InSAR hole DEM to be filled to generate a high-precision DEM product.

[0096] The process of obtaining the high-precision InSAR hole DEM to be filled includes:

[0097] Acquire multi-view SAR satellite data of the same target area, perform format conversion, generate single-view complex images, pre-process the single-view complex images, generate a number of differential interferograms, unwrap the differential interferograms, and generate unwrap differential interferogram phases;

[0098] Based on formula (1), the high coherence point is selected for the unwrapped differential interference phase to obtain the unwrapped phase at the high coherence point.

[0099] (1);

[0100] Where: Indicates m The differential interference pattern i Highly coherent points ( x , y ) at the unwrapping phase, 1≤ m ≤ M , M represents the total number of differential interferograms generated; Indicates m The differential interference pattern i Highly coherent points ( x , y ), , λ represents the radar wavelength, Δ d ( x , y )express( x , y ) upward line of sight deformation; Indicates m The differential interference pattern i Highly coherent points ( x , y) orbit error; Indicates m The differential interference pattern i Highly coherent points ( x , y ) at atmospheric delay phase; Indicates m The differential interference pattern i Highly coherent points ( x , y ) at the noise phase; Indicates m The differential interference pattern i Highly coherent points ( x , y ) at the high-pass deformation component; Indicates m The differential interference pattern i Highly coherent points ( x , y ),

[0101] (2);

[0102] Where: R Indicates the distance between the coherent target and the satellite; θ and denote the radar incident angle and m The length of the vertical baseline of the interference pattern; Δ h ( x , y ) represents the low-resolution DEM and high-resolution DEM i The difference at the high coherence points (x, y);

[0103] Use formula (3) to generate i Highly coherent points ( x , y ) at the elevation information;

[0104] (3);

[0105] Where: z ( x , y ) indicates the i Highly coherent points ( x , y ), Represents low-resolution DEM i Highly coherent points ( x , y ) at the elevation;

[0106] The elevation information is used to generate a high-precision InSAR cavity DEM that matches the target geographic location through geocoding.

[0107] The preprocessing comprises:

[0108] The single-view complex image is subjected to baseline estimation and connection map generation, image registration, interference, flattening, terrain removal and filtering in sequence to generate intensity map, coherence map and differential interferogram. The differential interferogram is unwrapped using the minimum cost flow method to generate the unwrapped differential interferogram phase.

[0109] The training process of the Transformer and CNN intelligent learning reconstruction model based on multi-source features includes:

[0110] Download low-resolution DEM data that matches the actual geographical location of the high-precision InSAR hole DEM to be filled;

[0111] Constructing a DEM training sample set according to the matched low-resolution DEM data;

[0112] The terrain features corresponding to the low-resolution DEM data in the DEM training sample set are calculated using the slope and aspect extraction operators, and a terrain feature training sample set is constructed based on the corresponding terrain features;

[0113] The intensity map and coherence map generated from the single-view complex image are used as SAR features to construct a SAR feature training sample set;

[0114] Constructing a Transformer and CNN intelligent learning and reconstruction model based on multi-source features, wherein the Transformer and CNN intelligent learning and reconstruction model based on multi-source features includes: a shallow feature extraction module, a multi-source feature connection module, a Transformer, a CNN feature mapping module, an upsampling module, and a loss function calculation module;

[0115] The shallow feature extraction module is expressed as:

[0116] ;

[0117] Where: F 0 indicates the extracted shallow features; H SFE Indicates the weight is W 0 The convolutional layer; D LR Represents the input low-resolution DEM; W 0 Indicates the weight of the 3×3 convolution layer during shallow feature extraction; * indicates convolution operation;

[0118] The new shallow feature representation generated by the multi-source feature connection module is:

[0119] ;

[0120] Where: F New Represents the generated new shallow features; W New and b They represent the weight and bias of the 3×3 convolutional layer when multi-source features are fused; F TS It represents the fusion of terrain features and SAR features;

[0121] The feature representation of the output of the ATCB module in the attention transformer and CNN feature mapping module is:

[0122] ;

[0123] Where: F k Indicates k The characteristics of the ATCB module output are: K is the total number of ATCB modules, 1≤ k ≤ K ; , and Respectively represent k , No. k -1. The first ATCB module extracts nonlinear functions of features;

[0124] The feature representation after deep feature fusion is:

[0125] ;

[0126] Where: F D Represents the features after deep feature fusion; W 1 and W 2 represents the weight of the 3×3 convolution layer and the weight of the 1×1 convolution layer when deep features are fused; PReLU represents the activation function; F 1, F 2,…, F K ] represents the characteristics of each ATCB module output;

[0127] The upsampled high-resolution DEM is expressed as:

[0128] ;

[0129] Where:D HR Represents high-resolution DEM; F UP represents the pixel transformation function; W 3 represents the weight of the 3×3 convolutional layer during upsampling;

[0130] The loss function is expressed as:

[0131] ;

[0132] Where: Loss represents the loss function; L Global represents the global loss; L Terrain Indicates loss of terrain features; L SAR Indicates SAR feature loss;

[0133] The Transformer and CNN intelligent learning and reconstruction models based on multi-source features are trained according to the DEM training sample set, the terrain feature training sample set, the SAR feature training sample set and the loss function. When the loss function does not meet the convergence judgment condition, the model is retrained by error back propagation. When the loss function meets the judgment convergence condition, the trained Transformer and CNN intelligent learning and reconstruction models based on multi-source features are output.

[0134] The global loss is expressed as:

[0135] ;

[0136] Where: J Indicates the number of DEM training sample sets; f mask represents the hole mask function; and Respectively represent j The elevation values ​​of high-resolution DEM and InSAR high-precision DEM training samples.

[0137] The terrain feature loss is expressed as:

[0138] ;

[0139] Where: and Respectively represent j The terrain characteristic values ​​generated by a high-resolution DEM and an InSAR high-precision DEM.

[0140] The SAR feature loss is expressed as:

[0141] ;

[0142] Where: Indicates j The pixel value of the SAR feature training sample.

[0143] The calculation formulas for calculating the terrain features corresponding to the low-resolution DEM data in the DEM training sample set using the slope and aspect extraction operators are:

[0144] ;

[0145] ;

[0146] Where: Slope Indicates the slope, Aspect Indicates the slope direction, f x Indicates the change in horizontal distance from the pixel point elevation; f y Indicates the change in vertical distance from the pixel point elevation.

[0147] When constructing the DEM training sample set, the terrain feature training sample set and the SAR feature training sample set, it also includes:

[0148] Use interpolation or resampling methods to make the resolution of low-resolution DEM data meet the multiple requirements of high-resolution DEM data;

[0149] The low-resolution DEM data and terrain features were cropped with a pixel size of 192s×192s, and the SAR feature data was cropped with a pixel size of 192×192. For the areas that did not meet the pixel size of 192×192 at the boundary, cropping was abandoned, and the DEM training sample set, terrain feature training sample set, and SAR feature training sample set that were integer multiples of 192×192 were obtained. 192×192 represents the pixel size, and s is the ratio between the high-resolution DEM and the low-resolution DEM resolution.

[0150] Embodiment 2, based on the same inventive concept as Embodiment 1, this embodiment introduces a method for filling in areas with missing high-precision DEM data. The present invention aims to solve the problems of limited accuracy of hole filling caused by missing traditional DEM data, failure to fully consider high-frequency information in areas with missing data, and inconsistency between filling results and terrain. In response to these problems, the present invention establishes a Transformer and CNN intelligent learning and reconstruction model based on multi-source features, uses the multi-source feature connection module as feature supervision and loss function optimization, finds the rules between complex terrain features, and extracts their commonalities. Constructs a feature mapping relationship between low-precision DEM and non-hole areas of high-precision InSAR DEM, performs intelligent modeling and filling in hole areas, improves the accuracy of filling in DEM hole areas, and obtains high-precision DEM products over a wide range.

[0151] like Figure 1 As shown, the process of the embodiment of the present invention specifically includes the following steps:

[0152] Step 1: Generate high-precision InSAR DEM.

[0153] Step 1.1: Data acquisition and preprocessing. Obtain multi-view SAR (synthetic aperture radar) satellite data of the same study area (i.e. N +1 SAR image, including N +1 refers to the number of images obtained), and format conversion is performed to generate single-look complex (SLC) images, and the images are preprocessed. The preprocessing mainly includes baseline estimation and connection map generation, image registration, interference, flattening, and terrain removal (using external 12.5-meter low-resolution DEM data to remove terrain phase, and any point of the low-resolution DEM ( x , y ) can be expressed as h ref ( x , y )) and filtering are used to generate intensity maps, coherence maps and differential interference maps, and the differential interference map is phase unwrapped using the minimum cost flow method (partial interference map is 3D unwrapped) to generate the unwrapped differential interference phase.

[0154] Step 1.2: Selection of high coherence points. The intensity threshold, amplitude deviation index threshold and time threshold coherence coefficient threshold are used to select high coherence points. In this process, independent component analysis ICA (ICA) is used to remove atmospheric delay and noise, and the adaptive discriminant polynomial fitting method is used to weaken the baseline error. The time series InSAR combined model that takes into account linear deformation, periodic deformation and environmental factors is used for modeling. N +1 SAR image, can be generatedM The differential interference pattern is m Width(1≤ m ≤ M ) The differential interference pattern i Highly coherent points ( x , y ) It can be expressed as:

[0155] (1);

[0156] Where: Indicates m The differential interference pattern i Highly coherent points ( x , y ) at the unwrapping phase, 1≤ m ≤ M , M represents the total number of differential interferograms generated; Indicates m The differential interference pattern i Highly coherent points ( x , y ), , λ represents the radar wavelength, Δ d ( x , y )express( x , y ) is the upward deformation component of the line of sight (LOS) at the vertical axis, which is called the low-pass (LP) deformation component; Indicates m The differential interference pattern i Highly coherent points ( x , y ) orbit error; Indicates m The differential interference pattern i Highly coherent points ( x , y ) at atmospheric delay phase; Indicates m The differential interference pattern i Highly coherent points ( x , y ) at the noise phase; Indicates m The differential interference pattern i Highly coherent points ( x , y ) at the high-pass (HP) deformation component; Indicates m The differential interference pattern i Highly coherent points ( x , y ), and its relationship with the surface residual elevation can be expressed as:

[0157] (2);

[0158] Where: R Indicates the distance between the coherent target and the satellite; θ and denote the radar incident angle and m The length of the vertical baseline of the interference pattern; Δ h ( x , y ) represents the low-resolution DEM and high-resolution DEM i The difference between the low-resolution DEM and the high-resolution DEM at the first high-coherence point (x, y) i Highly coherent points ( x , y ) is the difference between .

[0159] Step 1.3: Generate high-precision InSARDEM. Formula (1) can be combined with formula (2) and the time series InSAR combined model that takes into account linear deformation, periodic deformation and environmental factors to solve the deformation parameters and residual elevation Δ through least squares and singular value decomposition SVD (Singular Value Decomposition, SVD). h ( x , y ), No. i Highly coherent points ( x , y ) at the height z( x , y ) can be expressed as:

[0160] (3);

[0161] Where: z ( x , y ) indicates the i Highly coherent points ( x , y ), Represents low-resolution DEM i Highly coherent points ( x , y ) at the elevation;

[0162] Formula (3) can be used to obtain the elevation information of high-precision InSAR DEM, and then a high-precision InSAR cavity DEM matching the actual geographical location is generated through steps such as geocoding.

[0163] Step 2: Construction of training sample set.

[0164] The training sample set consists of a low-resolution training sample set and a high-resolution training sample set, such as Figure 2 Specifically, it includes DEM training sample set, terrain feature training sample set and SAR feature training sample set.

[0165] Step 2.1: DEM training sample set construction. There is a wealth of low-resolution DEM data published and available for free download worldwide, such as 90m, 30m, and 12.5m DEMs, which are all available for free. Appropriate data can be selected according to the research area and actual needs. In the embodiment of the present invention, the high-precision InSAR hole DEM obtained in step 1 is used as the high-resolution DEM data, and the ALOS 12.5m DEM that matches the actual geographical location of the high-resolution DEM is used as the low-resolution DEM data.

[0166] To facilitate the calculation of the deep learning network, the resolution of the low-resolution DEM data and the high-resolution DEM data must meet the multiple relationship of 2, 3, 4, etc. Therefore, the interpolation or resampling method is used to make the DEM data meet the multiple requirements. Then the high-resolution DEM is cropped with a pixel size of 192×192, while the low-resolution DEM data is cropped with a pixel size of 192s×192s, where s is the ratio between the resolution of the high-resolution DEM and the low-resolution DEM. For areas that do not meet the 192×192 pixel size at the boundary, cropping is abandoned, and finally a DEM training sample set of integer multiples of 192×192 is obtained.

[0167] Step 2.2: Construction of terrain feature training sample set.

[0168] Use the slope and aspect extraction operators to calculate the terrain features corresponding to the low-resolution DEM data. Use the low-resolution DEM clipping method in step 2.1 to clip the slope and aspect features, and add the clipped terrain feature data to the training sample set. Slope , Slope Aspect The calculation formula can be expressed as:

[0169] (4);

[0170] (5);

[0171] Where: f xIndicates the change in pixel elevation in horizontal distance (east-west direction); f y Indicates the change in vertical distance (north-south direction) from the pixel elevation.

[0172] Step 2.3: Construction of SAR feature training sample set.

[0173] The intensity map and coherence map generated in step 1.1 are used as SAR features. The intensity feature represents the intensity of the response of the objects and landforms in the SAR image to the echo signal. The average value of the SAR image intensity information at different times is calculated and normalized. The coherence feature reflects the stability of the InSAR differential interferometry phase signal. The processing method is the same as the intensity feature. The average value of the coherence information in different differential interferograms is calculated and normalized. The SAR features are cropped according to the cropping method of the high-resolution DEM in step 2.1 and added to the training sample set.

[0174] Step 3: Build a Transformer and CNN intelligent learning reconstruction model based on multi-source features.

[0175] Combining the characteristics of DEM datasets and the voids caused by data missing areas, a Transformer and CNN intelligent learning reconstruction model based on multi-source features is constructed. Figure 3 As shown in the figure, it specifically includes shallow feature extraction module, multi-source feature connection module, Transformer and CNN feature mapping module, upsampling module and loss function calculation module.

[0176] Step 3.1: Shallow feature extraction module.

[0177] The main function of the shallow feature extraction module is to establish the mapping between DEM data from low dimension to high dimension. The design structure includes a 3×3 convolution layer and a shallow feature matrix. The shallow features are extracted from the low-resolution DEM through the 3×3 convolution layer to achieve the mapping of DEM from low dimension to high dimension and generate the corresponding shallow feature matrix. Assume that the input low-resolution DEM is D LR , then the extracted shallow features F 0 can be represented as:

[0178] (6);

[0179] Where: H SFE Indicates the weight is W 0 convolutional layer; W 0 represents the weight of the 3×3 convolutional layer during shallow feature extraction; * represents the convolution operation.

[0180] Step 3.2: Multi-source feature connection module.

[0181] Introduce a multi-source feature connection module as feature supervision in the basic structure of the model network, establish an intelligent learning reconstruction model based on multi-source features, and improve the training accuracy and consistency of terrain features. The multi-source features connected mainly include terrain features and SAR features, where the terrain features are the slope and aspect features generated by the low-resolution DEM in step 2.2, and the SAR features are the coherence features and intensity features generated in step 2.3. The feature matrix obtained after normalization of the terrain features and SAR features is generated by feature fusion to generate a fused matrix F TS ,Since the resolution of SAR features is consistent with high-resolution DEM, it cannot be fused with terrain features by simple convolution. Therefore, a feature fusion module consisting of two convolution layers and one activation layer is used to fuse terrain features and SAR features. F TS The shallow feature matrix generated in step 3.1 F 0 to generate a new shallow feature F New It can be expressed as:

[0182] (7);

[0183] Where: W New and b They represent the weight and bias of the 3×3 convolutional layer when fusing multi-source features.

[0184] Step 3.3: Attention Transformer and CNN feature mapping modules.

[0185] In order to better extract deep features and enhance feature information, an attention transformer and CNN feature mapping module is designed, the core of which is the attention transformer and CNN module ATCB (Attention Transformer andCNN Block, ATCB). F New Input attention Transformer and CNN feature mapping module, after KATCB modules, a 1×1 convolution layer, and a 3×3 convolution layer are used for calculation, and finally a deep feature matrix is ​​obtained. ATCB module, attention structure AM (Attention Module, AM) and Swin Transformer structure STL (Swin Transformer Layer, STL), such as Figure 4-Figure 6 shown.

[0186] The ATCB module consists of two attention modules AM, two Swin Transformer structures STL and a 3×3 convolutional layer, as shown in Figure 4 As shown. k ATCB modules (1≤ k ≤ K ) Output features F k It can be expressed as:

[0187] (8);

[0188] Where: K is the total number of ATCB modules; , and Respectively represent k , No. k -1. The first ATCB module extracts the nonlinear function of the feature. When a feature is input F k-1 , No. k An ATCB first extracts and selects important features from the attention structure AM, then extracts intermediate features through the Swin Transformer structure STL, passes through a 3×3 convolution layer to maintain performance, and finally introduces an attention structure AM to weight the local important area to enhance the smoothness of the terrain. The attention structure AM consists of convolution layers of different sizes and windows and a sigmoid activation function, such as Figure 5 As shown in the figure, it is used to enhance the representation information of features and improve the ability and efficiency of model training. The Swin Transformer structure STL consists of two channel normalization layers LN (Layer Norm, LN), a shifting window based multi-head self-attention mechanism SW-MSA (Shifting Window Based Multi-Head Self-Attention, SW-MSA) and a multi-layer perceptron layer MLP (Multi Layer Perceptron, MLP), as shown in the figure. Figure 6 shown.

[0189] Finally, deep feature fusion is used to combine the features output by each ATCB module and effectively concatenate them in the channel dimension. Deep feature fusion contains a 1×1 convolution layer and a 3×3 convolution layer to fuse features at all levels. In addition, the residual aggregation structure is used to ease the difficulty of learning. Features after deep feature fusion F D It can be expressed as:

[0190] (9);

[0191] Where: W 1 and W 2 represents the weight of the 3×3 convolution layer and the weight of the 1×1 convolution layer when deep features are fused; PReLU represents the activation function; F 1, F 2,…, F K ] represents the characteristics of each ATCB module output.

[0192] Step 3.4: Upsampling module.

[0193] The upsampling module consists of a 3×3 convolution layer and a sub-pixel convolution layer, which is used to map multiple low-resolution DEM feature maps into a high-resolution DEM. D HR It can be expressed as:

[0194] (10);

[0195] Where: F UP represents the pixel transformation function; W 3 represents the weight of the 3×3 convolutional layer when upsampling. The sub-pixel convolutional layer first passes C 2 convolution kernels, converting a low-resolution DEM feature map into C 2 A low-resolution DEM feature map is then mapped to C 2 The low-resolution DEM feature maps are reassembled and spliced ​​into a high-resolution DEM feature map. Finally, multiple high-resolution DEM feature maps are fused into a high-resolution DEM.

[0196] Step 3.5: Loss function calculation module.

[0197] The loss function calculation module is used to calculate the error between the actual input and expected output DEM. The error back propagation function continuously iterates and updates the parameters of the model, so that the error gradually decreases and converges. In the loss function calculation module, the hole area in the DEM is first masked out, and the mapping relationship between the low-resolution DEM in the non-hole area and the high-resolution DEM is constructed. At the same time, in order to better improve the training ability of the model and enhance the effect of multi-source features, the designed loss function calculation module mainly includes global loss, terrain feature loss and SAR feature loss. The loss function Loss It can be expressed as:

[0198] (11);

[0199] Where: Loss represents the loss function; L Global represents the global loss; L Terrain Indicates loss of terrain features; L SAR Indicates SAR feature loss.

[0200] The global loss is calculated by the error between the high-resolution DEM predicted by the model and the elevation of the non-hole area of ​​the InSAR high-precision DEM. L Global It can be expressed as:

[0201] (12);

[0202] Where: J Indicates the number of DEM training sample sets; f mask represents the hole mask function; and Respectively represent j The elevation values ​​of high-resolution DEM and InSAR high-precision DEM training samples.

[0203] The terrain feature loss is calculated by the error between the high-precision DEM predicted by the model and the terrain features generated by the InSAR high-precision DEM. L Terrain It can be expressed as:

[0204] (13);

[0205] Where: and Respectively represent jThe terrain characteristics are generated by a high-resolution DEM and an InSAR high-precision DEM. The terrain characteristics include slope characteristics and aspect characteristics, which can be calculated by equations (4) and (5) in step 2.2.

[0206] The essence of SAR features is to distinguish the characteristics of different landforms. For model training, using SAR features as prior information to weigh the quality of DEM can further supervise and guide the model. Therefore, using SAR features as the weight of DEM to measure SAR feature loss, SAR feature loss L SAR It can be expressed as:

[0207] (14);

[0208] Where: Indicates j The pixel values ​​of the SAR feature training samples. SAR features include intensity features and coherence features, which can be generated through step 1.1.

[0209] Taking into account the global loss, terrain feature loss and SAR feature loss, the loss function can be calculated by substituting equations (12), (13) and (14) into equation (11). When the loss function does not meet the convergence criteria, the model is retrained through error back propagation. When the loss function meets the convergence criteria, the predicted high-precision DEM is output.

[0210] Step 4: Model learning, training and validation.

[0211] The hole areas of the training sample set obtained in step 2 are masked, and the first 80% of the data set is used as the training set and the last 20% as the validation set. The training set is used to train the Transformer and CNN intelligent learning reconstruction models based on multi-source features constructed in step 3, initialize the network parameters, obtain the training convergence model, and verify it with the validation set. The accuracy of the validation set is evaluated by the mean absolute error (MAE), root mean square error (RMSE) of DEM elevation and the mean absolute error (MAE) of slope.

[0212] Step 5: Intelligently fill in and predict missing data areas.

[0213] The hole area in the DEM is used as the test set. The low-resolution DEM data, terrain feature data (slope and aspect generated by low-resolution DEM) and SAR feature data (intensity features and coherence features) of the test set are input into the trained model in step 4 to achieve intelligent filling and prediction of the high-precision InSAR hole DEM in step 1.

[0214] Step 6: Large-scale high-precision DEM mapping and accuracy verification.

[0215] The high-precision DEM product can be generated by combining the DEM of the data-missing area intelligently filled and predicted in step 5 with the high-precision InSAR hole DEM generated in step 1. Then, the generated high-precision DEM is subjected to coordinate transformation and reprojection to generate a DEM product that matches the actual geographical location.

[0216] Then, the maximum elevation error (E) between the generated high-precision DEM and the actual surface elevation is calculated. Max ), mean absolute error (MAE), root mean square error (RMSE) and line error at 90% probability (LE90), and compared with traditional interpolation or deep learning methods such as bicubic interpolation (Bicubic), super-resolution convolutional neural network (SRCNN), enhanced deep super-resolution network (EDSR), residual channel attention network (RCAN), etc.

[0217] In summary, the intelligent learning and reconstruction modeling method for high-precision DEM data missing areas provided in this embodiment starts from the acquired high-precision InSAR hole DEM, organically combines the multi-source feature connection module with Transformer and CNN, fully considers the high-frequency information of the DEM hole area, and realizes the intelligent modeling and filling of the hole area.

[0218] Embodiment 3 is based on the same inventive concept as Embodiment 1. This embodiment introduces a device for filling missing areas of high-precision DEM data, comprising:

[0219] An acquisition module is used to acquire a high-precision InSAR hole DEM to be filled and a low-resolution DEM that is consistent with the geographical location of the high-precision InSAR hole DEM to be filled;

[0220] A determination module, used for determining terrain feature data and SAR feature data according to the low-resolution DEM;

[0221] A prediction module, used for inputting the low-resolution DEM, the terrain feature data and the SAR feature data into a pre-trained Transformer and CNN intelligent learning reconstruction model based on multi-source features to obtain a predicted high-resolution DEM of the hole area;

[0222] The processing process of the Transformer and CNN intelligent learning reconstruction model based on multi-source features includes:

[0223] Extract shallow features from low-resolution DEM to generate corresponding shallow feature matrices; perform normalization and feature fusion on terrain feature data and SAR feature data in turn to generate fused matrices; connect the shallow feature matrix and the fused matrix to generate new shallow features; perform attention Transformer and CNN feature mapping on the new shallow features to obtain a deep feature matrix; generate a predicted high-resolution DEM of the hole area based on the low-resolution DEM and the deep feature matrix;

[0224] The combining module is used to combine the predicted high-resolution DEM of the hole area with the high-precision InSAR hole DEM to be filled to generate a high-precision DEM product.

[0225] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0226] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0227] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0228] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0229] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for filling missing areas in high-precision DEM data, characterized in that: include: Obtaining a high-precision InSAR hole DEM to be filled and a low-resolution DEM that is consistent with the geographical location of the high-precision InSAR hole DEM to be filled; Determine terrain feature data and SAR feature data according to the low-resolution DEM; Inputting the low-resolution DEM, the terrain feature data and the SAR feature data into a pre-trained Transformer and CNN intelligent learning reconstruction model based on multi-source features to obtain a predicted high-resolution DEM of the cavity area; The processing process of the Transformer and CNN intelligent learning reconstruction model based on multi-source features includes: Extract shallow features from low-resolution DEM to generate corresponding shallow feature matrices; perform normalization and feature fusion on terrain feature data and SAR feature data in turn to generate fused matrices; connect the shallow feature matrix and the fused matrix to generate new shallow features; perform attention Transformer and CNN feature mapping on the new shallow features to obtain a deep feature matrix; generate a predicted high-resolution DEM of the hole area based on the low-resolution DEM and the deep feature matrix; Combining the predicted high-resolution DEM of the hole area with the high-precision InSAR hole DEM to be filled to generate a high-precision DEM product; The training process of the Transformer and CNN intelligent learning reconstruction model based on multi-source features includes: Download low-resolution DEM data that matches the actual geographical location of the high-precision InSAR hole DEM to be filled; Constructing a DEM training sample set according to the matched low-resolution DEM data; The terrain features corresponding to the low-resolution DEM data in the DEM training sample set are calculated using the slope and aspect extraction operators, and a terrain feature training sample set is constructed based on the corresponding terrain features; The intensity map and coherence map generated from the single-view complex image are used as SAR features to construct a SAR feature training sample set; Constructing a Transformer and CNN intelligent learning and reconstruction model based on multi-source features, wherein the Transformer and CNN intelligent learning and reconstruction model based on multi-source features includes: a shallow feature extraction module, a multi-source feature connection module, a Transformer, a CNN feature mapping module, an upsampling module, and a loss function calculation module; The shallow feature extraction module is expressed as: ； Where: F 0 indicates the extracted shallow features; H SFE Represents the weight W 0 The convolutional layer; D LR Represents the input low-resolution DEM; W 0 Indicates the weight of the 3×3 convolution layer during shallow feature extraction; * indicates convolution operation; The new shallow feature representation generated by the multi-source feature connection module is: ； Where: F New Represents the generated new shallow features; W New and b They represent the weight and bias of the 3×3 convolutional layer when multi-source features are fused; F TS It represents the fusion of terrain features and SAR features; The feature representation of the output of the ATCB module in the attention transformer and CNN feature mapping module is: ； Where: F k Indicates k The characteristics of the ATCB module output are: K is the total number of ATCB modules, 1≤ k ≤ K ; , and Respectively represent k , No. k -1. The first ATCB module extracts nonlinear functions of features; The feature representation after deep feature fusion is: ； Where: F D Represents the features after deep feature fusion; W 1 and W 2 represents the weight of the 3×3 convolution layer and the weight of the 1×1 convolution layer when deep features are fused; PReLU represents the activation function; F 1, F 2,…, F K ] represents the characteristics of each ATCB module output; The upsampled high-resolution DEM is expressed as: ； Where: D HR Represents high-resolution DEM; F UP represents the pixel transformation function; W 3 represents the weight of the 3×3 convolutional layer during upsampling; The loss function is expressed as: ； Where: Loss represents the loss function; L Global represents the global loss; L Terrain Indicates loss of terrain features; L SAR Indicates SAR feature loss; The Transformer and CNN intelligent learning and reconstruction models based on multi-source features are trained according to the DEM training sample set, the terrain feature training sample set, the SAR feature training sample set and the loss function. When the loss function does not meet the convergence judgment condition, the model is retrained by error back propagation. When the loss function meets the judgment convergence condition, the trained Transformer and CNN intelligent learning and reconstruction models based on multi-source features are output.

2. The method for filling missing areas of high-precision DEM data according to claim 1 is characterized in that: The process of obtaining the high-precision InSAR hole DEM to be filled includes: Acquire multi-view SAR satellite data of the same target area, perform format conversion, generate single-view complex images, pre-process the single-view complex images, generate a number of differential interferograms, unwrap the differential interferograms, and generate unwrap differential interferogram phases; Based on formula (1), the high coherence point is selected for the unwrapped differential interference phase to obtain the unwrapped phase at the high coherence point. （1）； Where: Indicates m The differential interference pattern i Highly coherent points ( x , y ) at the unwrapping phase, 1≤ m ≤ M , M represents the total number of differential interferograms generated; Indicates m The differential interference pattern i Highly coherent points ( x , y ), , λ represents the radar wavelength, Δ d ( x , y )express( x , y ) upward line of sight deformation; Indicates m The differential interference pattern i Highly coherent points ( x , y ) orbit error; Indicates m The first i Highly coherent points ( x , y ) at atmospheric delay phase; Indicates m The first i Highly coherent points ( x , y ) at the noise phase; Indicates m The first i Highly coherent points ( x , y ) at the high-pass deformation component; Indicates m The differential interference pattern i Highly coherent points ( x , y ), （2）； Where: R Indicates the distance between the coherent target and the satellite; θ and denote the radar incident angle and m The length of the vertical baseline of the interference pattern; Δ h ( x , y ) represents the low-resolution DEM and high-resolution DEM i The difference at the high coherence points (x, y); Use formula (3) to generate i Highly coherent points ( x , y ) at the elevation information; （3）； Where: z ( x , y ) indicates the i Highly coherent points ( x , y ), Indicates low-resolution DEM i Highly coherent points ( x , y ) at the height of The elevation information is used to generate a high-precision InSAR cavity DEM that matches the target geographic location through geocoding.

3. The method for filling missing areas of high-precision DEM data according to claim 2 is characterized in that: The preprocessing comprises: The single-view complex image is subjected to baseline estimation and connection map generation, image registration, interference, flattening, terrain removal and filtering in sequence to generate intensity map, coherence map and differential interferogram. The differential interferogram is unwrapped using the minimum cost flow method to generate the unwrapped differential interferogram phase.

4. The method for filling missing areas of high-precision DEM data according to claim 3 is characterized in that: The global loss is expressed as: ； Where: J Indicates the number of DEM training sample sets; f mask represents the hole mask function; and Respectively represent j The elevation values ​​of high-resolution DEM and InSAR high-precision DEM training samples.

5. The method for filling missing areas of high-precision DEM data according to claim 4 is characterized in that: The terrain feature loss is expressed as: ； Where: and Respectively represent j The terrain characteristic values ​​generated by a high-resolution DEM and an InSAR high-precision DEM.

6. The method for filling missing areas of high-precision DEM data according to claim 5 is characterized in that: The SAR feature loss is expressed as: ； Where: Indicates j The pixel value of the SAR feature training sample.

7. The method for filling missing areas of high-precision DEM data according to claim 3 is characterized in that: The calculation formulas for calculating the terrain features corresponding to the low-resolution DEM data in the DEM training sample set using the slope and aspect extraction operators are: ； ； Where: Slope Indicates the slope, Aspect Indicates the slope direction, f x Indicates the change in horizontal distance from the pixel point elevation; f y Indicates the change in vertical distance from the pixel point elevation.

8. The method for filling missing areas of high-precision DEM data according to claim 3 is characterized in that: When constructing the DEM training sample set, the terrain feature training sample set and the SAR feature training sample set, it also includes: Use interpolation or resampling methods to make the resolution of low-resolution DEM data meet the multiple requirements of high-resolution DEM data; The low-resolution DEM data and terrain feature data are cropped with a pixel size of (X×X)·s, and the SAR feature data are cropped with a pixel size of X×X. For areas that do not meet the X×X pixel size at the boundary, cropping is abandoned to obtain DEM training sample sets, terrain feature training sample sets, and SAR feature training sample sets that are integer multiples of X×X, where X×X represents the pixel size and s is the ratio between the high-resolution DEM and low-resolution DEM resolutions.

9. A device for filling missing areas of high-precision DEM data, characterized in that: include: An acquisition module is used to acquire a high-precision InSAR hole DEM to be filled and a low-resolution DEM that is consistent with the geographical location of the high-precision InSAR hole DEM to be filled; A determination module, used for determining terrain feature data and SAR feature data according to the low-resolution DEM; A prediction module, used for inputting the low-resolution DEM, the terrain feature data and the SAR feature data into a pre-trained Transformer and CNN intelligent learning reconstruction model based on multi-source features to obtain a predicted high-resolution DEM of the hole area; The processing process of the Transformer and CNN intelligent learning reconstruction model based on multi-source features includes: Extract shallow features from low-resolution DEM to generate corresponding shallow feature matrices; perform normalization and feature fusion on terrain feature data and SAR feature data in turn to generate fused matrices; connect the shallow feature matrix and the fused matrix to generate new shallow features; perform attention Transformer and CNN feature mapping on the new shallow features to obtain a deep feature matrix; generate a predicted high-resolution DEM of the hole area based on the low-resolution DEM and the deep feature matrix; A combining module, used for combining the predicted high-resolution DEM of the hole area with the high-precision InSAR hole DEM to be filled to generate a high-precision DEM product; The training process of the Transformer and CNN intelligent learning reconstruction model based on multi-source features includes: Download low-resolution DEM data that matches the actual geographical location of the high-precision InSAR hole DEM to be filled; Constructing a DEM training sample set according to the matched low-resolution DEM data; The terrain features corresponding to the low-resolution DEM data in the DEM training sample set are calculated using the slope and aspect extraction operators, and a terrain feature training sample set is constructed based on the corresponding terrain features; The intensity map and coherence map generated from the single-view complex image are used as SAR features to construct a SAR feature training sample set; Constructing a Transformer and CNN intelligent learning and reconstruction model based on multi-source features, wherein the Transformer and CNN intelligent learning and reconstruction model based on multi-source features includes: a shallow feature extraction module, a multi-source feature connection module, a Transformer, a CNN feature mapping module, an upsampling module, and a loss function calculation module; The shallow feature extraction module is expressed as: ； Where: F 0 indicates the extracted shallow features; H SFE Represents the weight W 0 The convolutional layer; D LR Represents the input low-resolution DEM; W 0 Indicates the weight of the 3×3 convolution layer during shallow feature extraction; * indicates convolution operation; The new shallow feature representation generated by the multi-source feature connection module is: ； Where: F New Represents the generated new shallow features; W New and b They represent the weight and bias of the 3×3 convolutional layer when multi-source features are fused; F TS It represents the fusion of terrain features and SAR features; The feature representation of the output of the ATCB module in the attention transformer and CNN feature mapping module is: ； Where: F k Indicates k The characteristics of the ATCB module output are: K is the total number of ATCB modules, 1≤ k ≤ K ; , and Respectively represent k , No. k -1. The first ATCB module extracts nonlinear functions of features; The feature representation after deep feature fusion is: ； Where: F D Represents the features after deep feature fusion; W 1 and W 2 represents the weight of the 3×3 convolution layer and the weight of the 1×1 convolution layer when deep features are fused; PReLU represents the activation function; F 1, F 2,…, F K ] represents the characteristics of each ATCB module output; The upsampled high-resolution DEM is expressed as: ； Where: D HR Represents high-resolution DEM; F UP represents the pixel transformation function; W 3 represents the weight of the 3×3 convolutional layer during upsampling; The loss function is expressed as: ； Where: Loss represents the loss function; L Global represents the global loss; L Terrain Indicates loss of terrain features; L SAR Indicates SAR feature loss; The Transformer and CNN intelligent learning and reconstruction models based on multi-source features are trained according to the DEM training sample set, the terrain feature training sample set, the SAR feature training sample set and the loss function. When the loss function does not meet the convergence judgment condition, the model is retrained by error back propagation. When the loss function meets the judgment convergence condition, the trained Transformer and CNN intelligent learning and reconstruction models based on multi-source features are output.

Images