High-precision InSAR DEM intelligent reconstruction method and system based on residual attention Transformer

Abstract

The present invention relates to a high-precision InSAR DEM intelligent reconstruction method and system based on residual attention transformer. High-resolution hole InSAR DEM generated by multi-baseline interferometry is used as training data, providing strong data support for model construction. By combining a lightweight Transformer module with a residual feature aggregation structure for high-resolution DEM reconstruction, while enhancing global perception capabilities, local residual features are aggregated to capture more comprehensive feature information, effectively improving reconstruction accuracy and training efficiency. At the same time, the spatial attention module is optimized, and an enhanced spatial attention mechanism is used to capture more complex multi-scale information. In addition, slope and aspect factors are introduced as terrain features for supervision in the loss function, which better optimizes model parameters based on surface morphological changes, enhances the model's ability to maintain terrain consistency, and provides a new solution to the problem of large-scale high-precision DEM reconstruction.

Description

Technical Field

[0001] The present invention belongs to the field of terrain surveying and mapping technology and digital surface reconstruction, and specifically relates to a high-precision InSAR DEM intelligent learning and reconstruction modeling method based on residual attention Transformer. Background Art

[0002] Digital Elevation Models (DEMs) are essential foundational data for spatial information analysis, playing a key role in digital terrain analysis, hydrological analysis, and many other geographic information modeling applications. High-precision DEMs, as the foundation of national-level terrain data, effectively support disaster management and environmental protection, such as flood simulation, earthquake assessment, and landslide and debris flow risk assessment. Furthermore, in national land planning and urban development, DEMs provide precise terrain information for realistic three-dimensional construction and land use optimization. Furthermore, their application in water resource management enables more accurate simulation of water flow paths and flood risks.

[0003] Currently, various DEM databases have been established both domestically and internationally, including the national 1:1 million, 1:250,000, 1:50,000, and 1:10,000 basic geographic information databases, the ASTER GDEM 30m resolution global DEM, and the ALOS 12.5m resolution global DEM. With the increasing number of high-resolution sensors and advancements in high-precision data acquisition methods, the scale of high-precision DEM data continues to grow. Among them, Synthetic Aperture Radar (InSAR) technology, with its all-day, all-weather capabilities and immunity to cloud, rain, and fog, can quickly and efficiently acquire high-density elevation point information, enabling the generation of high-precision DEM data. Currently, methods for generating high-precision DEMs based on InSAR technology primarily include multi-baseline interferometry and dual-satellite interferometry. Multi-baseline interferometry interferometry utilizes phase differences between SAR images acquired from different baselines to derive surface elevation information. Dual-satellite interferometry utilizes images acquired by two SAR satellites at different locations and calculates surface elevation information by analyzing the parallax between the two images. Both methods effectively mitigate the effects of spatiotemporal incoherence and atmospheric delay on DEM accuracy. Yu et al. generated high-precision InSAR DEMs using information from overlapping SAR images between different primary and vertical baselines, but the resulting results were affected by elevation ambiguity and noise. Liu et al. used multi-baseline InSAR technology to extract DEMs from earthquake-affected areas, enabling timely and accurate reflection of ground elevation changes. Zhang Changhai used dual-satellite interferometry to invert InSAR DEMs of the study area based on data from the domestically produced LuTan-1 (LT-1) satellite, providing timely data for subsequent hydrological analysis. However, due to factors such as terrain conditions, ground scattering properties, and temporal variations in the imaged area, InSAR DEMs generated using these techniques often contain holes or invalid data. For example, DEMs in these areas contain holes or missing data for certain periods due to poor signal-to-noise ratios in deserts, shadows and signal obstruction in complex terrain such as mountains and canyons, and extreme climate conditions and sea ice dynamics in the Antarctic. Overall, constructing large-scale, high-precision DEMs remains an urgent challenge.

[0004] Currently, the research methods for constructing high-precision DEM are mainly divided into three categories. The first is the interpolation-based method, including inverse distance interpolation, spline function interpolation, bicubic convolution interpolation and bilinear interpolation. This type of method is simple to calculate and easy to understand, but it has the disadvantages of obvious edge effects and insufficient detail restoration.

[0005] The second approach involves fusion and reconstruction, such as wavelet transforms and regularized frameworks. Sun Liang et al. proposed a multi-source DEM fusion method based on wavelet analysis. Using the ASTER DEM and SRTM DEM as the base data, they formed a fused data source through resampling and data registration. Low-frequency coefficients from the wavelet decomposition were fused based on the correlation between neighboring pixels, while high-frequency coefficients were fused based on the maximum absolute value of the pixel. This resulted in a fused DEM with improved accuracy. Yue et al. used a regularized framework to fuse multi-source, multi-scale DEM data, leveraging the advantages of multiple sources to improve DEM data quality and produce a seamless DEM. Yang Guangyi et al. proposed a multi-source, multi-scale DEM fusion method using valley ridges as topographic constraints. Using surface water flow simulation, they extracted valley ridges from the ASTER and WorldDEM data, then used the topographic details of the higher-resolution data to guide the interpolation of the lower-resolution data, resulting in a fused, high-precision DEM. This method combines the advantages of multiple data sources to produce higher-quality DEMs, but cannot generalize from small, fine-scale DEMs to large areas.

[0006] The third approach is based on deep learning. This approach utilizes a large amount of training data to establish a mapping relationship between low-resolution and high-resolution images. This mapping relationship is then used for predictive reconstruction to obtain high-resolution images. Compared to the previous two approaches, these approaches achieve superior reconstruction results and can expand the reconstruction range, facilitating the construction of DEM data for larger areas and higher accuracy. Convolutional Neural Networks (CNNs) have been widely used in computer vision due to their powerful information extraction and image fitting capabilities. Since Dong proposed the SRCNN, numerous CNN-based models for high-resolution image reconstruction have been proposed. These models can be categorized into two main categories: residual network-based models and generative adversarial network-based models. Residual network-based models increase network depth and provide a stable mapping from low to high resolution, typically resulting in higher accuracy. In contrast, generative adversarial networks, through the mutual parameter updates between the generator and discriminator, reconstruct high-resolution images with more high-frequency details, resulting in better visual results and avoiding smoothing artifacts, but often suffer from numerical distortion. The emergence of CNNs has also extended deep learning-based methods to high-precision DEM reconstruction research. Chen et al. used a three-layer convolutional neural network to extract, compress, and convert DEM features to reconstruct low-resolution DEMs into high-resolution DEMs, demonstrating that this method significantly outperformed traditional interpolation methods. Zhou et al. designed a dual-filter residual neural network that extracts features by combining convolution kernels with different receptive fields, thereby achieving more refined terrain modeling. Chen et al. constructed a residual network with an integrated attention mechanism for high-precision DEM reconstruction, incorporating slope and pipeline losses during training to improve model accuracy and its ability to preserve terrain features. Jiang et al. proposed a high-precision DEM reconstruction model that achieved superior reconstruction results in Asia's high mountain regions by integrating slope and curvature losses into the loss function. However, current research has encountered problems with increasing network depth, such as low training efficiency and high computational load due to excessive parameters. Furthermore, DEM data differs from general image data and exhibits variability. Some models fail to fully consider the important influence of terrain features in DEM data when optimizing network architecture. Furthermore, many models still face challenges with limited research areas and insufficient transferability. Summary of the Invention

[0007] In response to the shortcomings of existing technologies, the present invention provides a high-precision InSARDEM intelligent learning and reconstruction modeling method based on the residual attention Transformer, which reconstructs a high-precision DEM with a wide coverage. By combining the Transformer with the residual attention network and applying it to the fine-scale modeling of the DEM, while enhancing global perception capabilities, it aggregates local residual features to capture more comprehensive feature information, thereby improving the training efficiency of the model. At the same time, it introduces slope and aspect factors as terrain constraints and constructs a loss function that combines terrain characteristics, effectively improving the reconstruction accuracy of the DEM and the reconstruction effect of the terrain, providing strong technical support for the difficult problem of constructing large-scale high-precision DEMs.

[0008] To achieve the above objectives, the technical solution provided by the present invention is a high-precision InSAR DEM intelligent reconstruction method based on residual attention Transformer, comprising the following steps:

[0009] Step 1: Generate a multi-baseline InSAR DEM based on the TerraSAR-X data of the study area, and obtain the InSAR cavity DEM as a high-resolution dataset;

[0010] Step 2: Preprocess the ALOSDEM data of the study area and the cavity DEM data generated by TerraSAR-X to obtain the corresponding low-resolution DEM and high-resolution DEM datasets;

[0011] Step 3: Use the terrain feature extraction operator to calculate the slope and aspect datasets corresponding to the high-resolution DEM data, and also crop them into several terrain feature blocks of corresponding pixel sizes;

[0012] Step 4: Build a residual attention Transformer network that incorporates terrain features to perform intelligent learning and reconstruction of high-precision DEM for the study area;

[0013] Step 5: Input the data obtained in step 2 into the residual attention Transformer network constructed in step 4 to perform network training and initialize the network parameters. The slope and aspect data extracted in step 3 are then integrated into the network loss function to finally obtain a trained high-precision DEM reconstruction model.

[0014] In step 6, the ALOSDEM data is input into the network model in step 5 for prediction, and several high-resolution DEM tiles without holes are output. Finally, the tiles are spliced ​​together to obtain a predicted high-precision DEM covering the entire study area.

[0015] Furthermore, step 1 includes the following steps:

[0016] Step 1.1: Convert the format of multiple SAR image data acquired in the same study area, and perform image registration, interferometry, and flattening preprocessing to generate intensity maps, coherence maps, and differential interferograms, and then generate the unwrapped differential interferogram phase.

[0017] Step 1.2: Independent component analysis (ICA) is used to separate the phase signal of the time series of acquired images to remove atmospheric delay and noise in multi-baseline interferometry.

[0018] Step 1.3: Use the adaptive discriminant quadratic polynomial fitting method to remove the interference fringes caused by the baseline error, and then subtract the phase obtained by ICA separation from the baseline error to obtain the corrected phase signal;

[0019] In step 1.4, based on the deformation modeling error in multi-baseline interferometry, a time-series InSAR combination model is constructed between the differential interferometry phase and linear, periodic, environmental, and terrain parameters. The residual elevation is obtained by solving it and then added to the low-resolution DEM elevation value to obtain the high-resolution InSAR point measurement value.

[0020] In step 1.5, the InSAR data are converted from slant-range or ground-range projection to geographic coordinate projection by geocoding the coordinates of high-resolution InSAR point measurements, thereby generating a high-resolution InSAR hole DEM that matches the actual geographic location.

[0021] Furthermore, step 2 includes the following steps:

[0022] Step 2.1: transform the ALOS 12.5m resolution DEM of the study area and project it into the same coordinate system as the high-resolution TerraSAR-X 3m hole DEM.

[0023] Step 2.2: Select an area with rich topography and large surface undulations in the study area as the training area, and crop it into several N1×N1 pixel blocks. Discard the blocks that are less than N1×N1 pixel blocks at the region boundary, and finally obtain a high-resolution DEM dataset with an integer multiple of N1×N1.

[0024] In step 2.3, based on the regional boundary obtained in step 2.2, a low-resolution DEM of the same regional range is cropped, and based on the resolution multiple S between the high-resolution and low-resolution DEMs, a number of corresponding N2×N2 pixel-sized tiles are cropped, where N2*S=N1, to obtain a low-resolution DEM dataset of the same range size. Finally, a pair of high-resolution and low-resolution DEM datasets is obtained, which can form a one-to-one mapping relationship between the high-resolution and low-resolution tiles.

[0025] Furthermore, the residual attention Transformer network in step 4 consists of three parts: a shallow feature extraction module, a deep feature extraction module, and an upsampling module. The first part, the shallow feature extraction module, consists of a 3×3 convolutional layer, which is used to extract shallow features with 64 output channels from the low-resolution DEM, realizing the mapping of the DEM from low dimension to high dimension.

[0026] The second part of the deep feature module consists of several Transformer residual attention modules RATB and a 1×1 convolutional layer. The input shallow features and terrain features are calculated by the deep feature module to obtain the deep feature matrix;

[0027] The third part of the upsampling module consists of a sub-pixel convolution layer and a 3×3 convolution layer. The sub-pixel convolution layer rearranges the number of channels of the input deep feature matrix and reorganizes the pixels of the low-resolution feature map according to the pixel mapping relationship to achieve the mapping of low-resolution features to high-resolution features. Finally, a 3×3 convolution layer is used to remove some unnecessary artifacts or aliasing phenomena.

[0028] Furthermore, the Transformer residual attention module RATB consists of two enhanced spatial attention modules ESA at the head and tail, a residual feature aggregation module RFA, two Swin Transformer layers and a 1×1 convolution layer; ESA consists of two 1×1 convolution layers, a void convolution, a convolution group consisting of three 3×3 convolution layers and a sigmoid activation function layer.

[0029] Furthermore, RFA consists of 4 residual modules RB and a 1×1 convolution layer. The hierarchical features on the branches of the first three residual modules are concatenated with the output of the last residual module; RB consists of 3 3×3 convolutions and an ESA.

[0030] Furthermore, two STLs form the SwinTransformer module, and the STL consists of two layer normalizations, a multi-head attention mechanism, and a multi-layer perceptron layer MLP.

[0031] Furthermore, the loss function of the model training in step 4 is composed of the global loss function and the terrain feature loss function, and the calculation formula is as follows:

[0032] L=L Global +L Terrain (6)

[0033] Where, L represents the total loss; L Global represents the global loss; L Terrain Represents terrain feature loss; the global loss calculation format is as follows:

[0034]

[0035] Where y i Indicates the elevation value of the InSAR hole high-resolution DEM; represents the elevation value of the predicted high-precision DEM; n represents the number of pixels contained in an InSAR hole high-resolution DEM;

[0036] The terrain feature loss function is composed of slope loss and aspect loss. The specific calculation formula is as follows:

[0037] L Terrain =λ1L Slope +λ2L Aspect (8)

[0038]

[0039] Where, L Slope Indicates slope loss; L Aspect represents the aspect loss; λ1 and λ2 represent the weight ratios of slope and aspect loss respectively; S i Indicates the slope value of the InSAR hole high-resolution DEM; Indicates the slope value of the predicted high-precision DEM; A i Indicates the aspect value of the InSAR hole high-resolution DEM; Indicates the predicted aspect value of the high-precision DEM.

[0040] Furthermore, step 7 is included to evaluate the accuracy of the high-precision DEM predicted by the model, using the mean absolute error (MAE), root mean square error (RMSE), maximum elevation difference (E) and the maximum elevation difference (DEM) to evaluate the accuracy of the DEM. max , mean absolute error (MAE) of slope Slope As an evaluation index, the difference between the predicted high-precision DEM and the actual high-resolution DEM is compared to evaluate the reconstruction effect. The calculation formula is as follows:

[0041]

[0042] Where n represents the number of pixels contained in an InSAR hole high-resolution DEM; y i Indicates the elevation value of the InSAR hole high-resolution DEM; Indicates the predicted high-precision DEM elevation value; S i Indicates the slope value of the InSAR hole high-resolution DEM; Indicates the slope value of the predicted high-precision DEM.

[0043] The present invention also provides a high-precision InSAR DEM intelligent reconstruction system based on residual attention Transformer, comprising:

[0044] A processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the high-precision InSAR DEM intelligent reconstruction method based on residual attention Transformer as described in the above technical solution.

[0045] Compared with the prior art, the present invention has the following advantages:

[0046] 1) A high-resolution InSAR DEM was generated as high-resolution training data using the multi-baseline interferometry method, providing strong data support for the construction of a high-precision DEM intelligent learning and reconstruction model.

[0047] 2) A high-precision InSAR DEM intelligent learning reconstruction modeling method based on residual attention Transformer is proposed. The Swin Transformer module is used to capture long-range dependencies and global features, and the training efficiency of the model is improved. The integration of the residual feature aggregation structure effectively integrates local features. At the same time, an enhanced spatial attention mechanism is introduced to capture more complex multi-scale information, improve the reconstruction accuracy, and expand the reconstruction range, providing new ideas and methods for large-scale high-precision DEM reconstruction.

[0048] 3) During the model training process, slope and aspect factors are incorporated into the loss function as terrain features, taking into account the impact of terrain on DEM construction in actual situations, effectively constraining the model training process, and ensuring consistency with terrain features while improving reconstruction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Flowchart of an embodiment of the present invention.

[0050] Figure 2 This is a network structure diagram of the residual attention Transformer network of an embodiment of the present invention.

[0051] Figure 3 This is a structural diagram of the residual attention Transformer module of an embodiment of the present invention.

[0052] Figure 4 This is a structural diagram of the enhanced spatial attention module according to an embodiment of the present invention.

[0053] Figure 5 This is a structural diagram of the residual feature aggregation module according to an embodiment of the present invention.

[0054] Figure 6This is a structural diagram of the SwinTransformer module in an embodiment of the present invention.

[0055] Figure 7 2 is a structural diagram of the residual module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0056] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0057] The present invention provides a high-precision InSAR DEM intelligent learning and reconstruction modeling method based on residual attention transformer. The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments. Figure 1 As shown, the process of the embodiment of the present invention includes the following steps:

[0058] Step 1: Download the multi-view TerraSAR-X data of the study area from the EESA official website (https: / / earth.esa.int / eogateway / missions / terrasar-x-and-tandem-x), generate a multi-baseline InSAR RDEM, and obtain a 3-m resolution InSAR cavity DEM as a high-resolution dataset.

[0059] Step 1.1: Extract and convert the data of multiple SAR images of the same study area into SLC files. Then perform spatiotemporal baseline connection to select interferometric pairs. Then perform preprocessing steps including image registration, interferometry, terrain phase simulation, differential interferometry, flattening, terrain removal, and filtering to generate intensity maps, coherence maps, and differential interferograms. Finally, phase unwrapping of the differential interferograms is performed using the minimum cost flow method to obtain the unwrapped differential interferogram phase.

[0060] In step 1.2, independent component analysis (ICA) is used to separate the phase signal of the time series of the acquired image to remove the atmospheric delay and noise in the multi-baseline interferometry. The phase obtained after ICA separation can be expressed as:

[0061]

[0062] Where, represents the phase of the i-th pixel on the m-th differential interferogram after ICA separation; λ represents the radar wavelength; represents the low-pass (LP) deformation component; Bm represents the length of the vertical baseline of the mth interferometer pair; R represents the distance between the coherent target and the radar satellite; θ represents the radar incident angle; ΔH represents the elevation error; represents the residual phase;

[0063] In step 1.3, a least squares smoothing filter function is used to remove peaks in the differential interferometer phase matrix. The baseline is approximated using a quadratic polynomial fitting method with adaptive discrimination to obtain the baseline error. The interference fringes caused by the baseline error are removed. The phase signal obtained by ICA separation is then subtracted from the baseline error to obtain the corrected phase signal. The calculation formula is as follows:

[0064]

[0065] Where, represents the phase at the i-th pixel on the m-th differential interferogram after baseline correction; represents the baseline error of the polynomial fit;

[0066] In step 1.4, considering the deformation modeling error in multi-baseline interferometry, a time-series InSAR combination model between the differential interferometry phase and linear, periodic, environmental factors, and terrain parameters is constructed, which can be expressed as:

[0067]

[0068] Where, represents a linear deformation model; represents a periodic deformation model; Represents the deformation model of environmental factors; solving it can obtain the residual elevation ΔH, which can be added to the elevation value of the low-resolution DEM to obtain the high-resolution InSAR point measurement value;

[0069] In step 1.5, the InSAR data are converted from slant range or ground range projection to geographic coordinate projection by geocoding the coordinates of high-resolution InSAR point measurements, thereby generating a 3 m high-resolution InSAR hole DEM that matches the actual geographic location.

[0070] Step 2: Download the ALOS12.5mDEM of the same study area from the EarthData official website (https: / / search.asf.alaska.edu) and perform coordinate registration and cropping on it with the DEM data generated in step 1 to obtain the corresponding low-resolution DEM and high-resolution DEM datasets. The tiles in the dataset correspond one by one.

[0071] Step 2.1: Transform the coordinates of the ALOS 12.5m resolution DEM of the study area and project it into the same coordinate system as the high-resolution TerraSAR-X 3m hole DEM. To ensure that the reconstruction factor is an integer, resample it to 12m resolution.

[0072] Step 2.2: Select two areas with rich topography and large surface undulations in the study area as training areas and crop them into several 192 × 192 pixel blocks. Discard the blocks with less than 192 × 192 pixel size at the area boundary. Finally, obtain a high-resolution DEM dataset with an integer multiple of 192 × 192.

[0073] In step 2.3, based on the regional boundaries obtained in 2.2, a low-resolution DEM with the same regional range is cropped. Since the high-resolution and low-resolution DEMs are reconstructed 4 times, several corresponding 48×48 pixel blocks are cropped to obtain a low-resolution DEM dataset with the same range size. Each image block is assigned a unique serial number according to the cutting order, and finally a pair of high-resolution and low-resolution DEM datasets are obtained, which can form a one-to-one mapping relationship between the high-resolution and low-resolution blocks, and each has 2312 blocks.

[0074] Step 3: Use the terrain feature extraction operator to calculate the slope and aspect datasets corresponding to the high-resolution DEM data, and also crop them into several terrain feature blocks of 192×192 pixels. The calculation formulas for slope and aspect are as follows:

[0075]

[0076] Where Δh x Indicates the elevation difference between two adjacent pixels in the east-west direction; Δh y Indicates the elevation difference between the two adjacent pixels of the pixel in the north-south direction.

[0077] Step 4: Construct a residual attention Transformer network that combines terrain features to perform intelligent learning and reconstruction of high-precision DEM in the study area.

[0078] Residual Attention Transformer Network Figure 2As shown in the figure, the model consists of three parts: a shallow feature extraction module, a deep feature extraction module, and an upsampling module. The first part, the shallow feature extraction module, consists of a 3×3 convolutional layer, which is used to extract shallow features with 64 output channels from the low-resolution DEM, realizing the mapping of the DEM from low dimension to high dimension. The second part, the deep feature module, consists of four residual attention transformer blocks (RATB) and a 1×1 convolutional layer. The input shallow features and terrain features are calculated by the deep feature module to obtain a deep feature matrix. The third part, the upsampling module, consists of a sub-pixel convolutional layer and a 3×3 convolutional layer. The sub-pixel convolutional layer rearranges the number of channels of the input deep feature matrix and reorganizes the pixels of the low-resolution feature map according to the pixel mapping relationship to realize the mapping of low-resolution features to high-resolution features. Finally, a 3×3 convolutional layer is used to help remove some unnecessary artifacts or aliasing.

[0079] Among them, RATB is the core part of the model, which includes two Enhanced Spatial Attention (ESA) modules at the beginning and end, a Residual Features Aggregation (RFA) module, two Swin Transformer layers (STL) and a 1×1 convolutional layer. Figure 3 The specific structure of the ESA module is as shown in Figure 4As shown in the figure, it consists of two 1×1 convolutional layers, a dilated convolution, a convolution group consisting of three 3×3 convolutional layers, and a sigmoid activation function layer. It differs from the general spatial attention mechanism in the following ways: first, dilated convolution is introduced to increase the receptive field and capture multi-scale information; second, the pooling layer is deleted. The function of the pooling layer is to downsample the features, and deleting this layer helps preserve the detailed information in the features; third, tensor splicing is used instead of direct addition in feature fusion. Using direct addition for feature fusion may cause information confusion and requires an upsampling layer to ensure the consistency of tensor dimensions. The improved module uses tensor splicing function for feature fusion, increasing the tensor dimension, which can simultaneously retain the information of low-level and high-level features, helping to capture more complex patterns and features. Finally, the attention weight of each spatial position is mapped to the range of [0,1] through the sigmoid activation function, so that the weight of each position can be interpreted as "importance". Placing this module at the beginning of RATB allows the model to focus on important spatial areas from the beginning and enhance the expression of key features; placing it at the end allows the model to adjust the attention weights of different spatial positions after feature learning is completed, so that important features are emphasized in the final output, thereby improving the overall performance of the model.

[0080] The RFA module consists of four residual blocks (RB) modules and a 1×1 convolutional layer, such as Figure 5 As shown in , the hierarchical features on the first three residual module branches are spliced ​​together with the output of the last residual module, which can effectively combine the features of different levels to produce more representative features. Compared with the directly stacked residual modules, the RFA module helps the model to understand the input data more comprehensively and improve the expression ability of features. Among them, the specific structure of RB is as follows Figure 7 As shown in the figure, it consists of three 3×3 convolutions and one ESA. The three 3×3 convolutions enable the model to extract deeper and more complex feature representations. Placing ESA in the residual module enables the model to adaptively adjust the feature area of ​​interest and focus on different attention layers, thereby enhancing the flexibility of feature selection.

[0081] Two STLs form the SwinTransformer module, and STL consists of two layer normalizations (LayerNormalization, LN), a multi-head self attention mechanism (Multi-head Self Attention, MSA) and a multi-layer perceptron layer MLP (Multi-layer Perceptron), such as Figure 6As shown in the figure, the first feature addition refers to the addition of the features input to the SwinTransformer and the features obtained after LN and MSA processing; the second feature addition is the addition of the matrix obtained by the first feature addition and the matrix processed by LN and MLP on this basis. MSA is used to learn the global and local dependencies of the input features. The calculation of each layer is implemented through local window operations, and then the global information is integrated through a moving window mechanism. This method of dividing the image into moving windows also effectively reduces the sequence length of the token, which means that each self-attention calculation only focuses on the features within the window, rather than the entire feature map, which significantly reduces the amount of computation and memory requirements and improves training efficiency. MLP is used to perform nonlinear transformations and feature extraction on the output of the attention head.

[0082] In step 5, the dataset obtained in step 2 is divided into a training dataset and a test dataset in a ratio of 8:2, resulting in 1,850 training sets and 462 validation sets. The corresponding high- and low-resolution DEM data in the training sets are input into the residual attention Transformer network constructed in step 4 for model training and initialization of network parameters. The slope and aspect data extracted in step 3 are integrated into the network's loss function as a terrain loss function combined with the global loss function to form an overall loss function. Finally, a high-precision DEM reconstruction model with gradually decreasing and converging errors is obtained.

[0083] For the specific hyperparameter settings, Adam optimizer is used for optimization, β1 is set to 0.9, β2 is set to 0.999, and ε is set to 10 -8 All models were trained on the PyTorch platform with a batch size of 16 and 1200 epochs on four GeForce GTX 3090 GPUs. The initial learning rate was set to 1×10 -4 , and the learning rate strategy of cosine annealing is used for adjustment.

[0084] The loss function of model training is composed of the global loss function and the terrain feature loss function. The calculation formula is as follows:

[0085] L=L Global +L Terrain (6)

[0086] Where, L represents the total loss; L Global represents the global loss; L Terrain Represents the terrain feature loss. The global loss calculation format is as follows:

[0087]

[0088] Where y iIndicates the elevation value of the InSAR hole high-resolution DEM; It represents the elevation value of the predicted high-precision DEM; n represents the number of pixels contained in an InSAR hole high-resolution DEM.

[0089] The terrain feature loss function is composed of slope loss and aspect loss. The specific calculation formula is as follows:

[0090] L Terrain =λ1L Slope +λ2L Aspect (8)

[0091]

[0092] Where, L Slope Indicates slope loss; L Aspect represents the aspect loss; λ1 and λ2 represent the weight ratios of slope and aspect loss respectively; S i Indicates the slope value of the InSAR hole high-resolution DEM; Indicates the slope value of the predicted high-precision DEM; A i Indicates the aspect value of the InSAR hole high-resolution DEM; Indicates the predicted aspect value of the high-precision DEM.

[0093] After the loss function is calculated, the Adam optimizer is used to update the model parameters through error back propagation, and convergence is achieved through repeated iterations.

[0094] In step 6, the trained and converged model obtained in step 5 is used for prediction. The full-range 12m resolution ALOSDEM data that has completed coordinate conversion and resampling in step 2 is also cropped into 48×48 pixel blocks and input into the network model. The corresponding 3m high-resolution DEM block data without holes are output. Finally, the blocks are spliced ​​to obtain the predicted high-precision DEM covering the entire study area.

[0095] Step 7: The accuracy of the high-precision DEM predicted by the model is evaluated using the mean absolute error (MAE), root mean square error (RMSE), maximum elevation difference (E max ), mean absolute error of slope (MAE Slope ) is used as the evaluation index to compare the difference between the predicted high-precision DEM and the actual high-resolution DEM to evaluate the reconstruction effect. The calculation formula is as follows:

[0096]

[0097] Where n represents the number of pixels contained in an InSAR hole high-resolution DEM; yi Indicates the elevation value of the InSAR hole high-resolution DEM; Indicates the predicted high-precision DEM elevation value; S i Indicates the slope value of the InSAR hole high-resolution DEM; Indicates the slope value of the predicted high-precision DEM.

[0098] Based on the above indicators, the experimental results of this method are compared with some other classic image high-resolution reconstruction models. The results are shown in the following table. All models are trained on the PyTorch platform and processed for 1200 epochs on 4 GeForce GTX3090. The initial learning rate is set to 1×10 -4 , and the learning rate strategy of cosine annealing is used for adjustment, and the Adam optimizer is used for optimization, with β1 set to 0.9, β2 set to 0.999, and ε set to 10 -8 As can be seen from the table, the residual attention Transformer network is the best in all four indicators, and it can be considered to be superior to other similar models currently available.

[0099]

[0100] On the other hand, an embodiment of the present invention also provides a high-precision InSAR DEM intelligent reconstruction system based on residual attention Transformer, including:

[0101] A processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the high-precision InSAR DEM intelligent reconstruction method based on the residual attention transformer as described in the above technical solution. The specific embodiments described in this article are merely examples of the spirit of the present invention. Those skilled in the art of the present invention may make various modifications or additions to the described specific embodiments or replace them with similar methods, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A high-precision InSAR DEM intelligent reconstruction method based on residual attention transformer, characterized by: The following steps are involved: Step 1: Generate a multi-baseline InSAR DEM based on the TerraSAR-X data of the study area, and obtain the InSAR cavity DEM as a high-resolution dataset; Step 2: Preprocess the ALOSDEM data of the study area and the cavity DEM data generated by TerraSAR-X to obtain the corresponding low-resolution DEM and high-resolution DEM datasets; Step 3: Use the terrain feature extraction operator to calculate the slope and aspect datasets corresponding to the high-resolution DEM data, and also crop them into several terrain feature blocks of corresponding pixel sizes; Step 4: Build a residual attention Transformer network that incorporates terrain features to perform intelligent learning and reconstruction of high-precision DEM for the study area; The residual attention Transformer network in step 4 consists of three parts: a shallow feature extraction module, a deep feature extraction module, and an upsampling module. The first part, the shallow feature extraction module, consists of a 3×3 convolutional layer, which is used to extract shallow features with 64 output channels from the low-resolution DEM, realizing the mapping of the DEM from low dimension to high dimension. The second part of the deep feature module consists of several Transformer residual attention modules RATB and a 1×1 convolutional layer. The input shallow features and terrain features are calculated by the deep feature module to obtain the deep feature matrix; The third part of the upsampling module consists of a sub-pixel convolution layer and a 3×3 convolution layer. The sub-pixel convolution layer rearranges the number of channels of the input deep feature matrix and reorganizes the pixels of the low-resolution feature map according to the pixel mapping relationship to achieve the mapping of low-resolution features to high-resolution features. Finally, a 3×3 convolution layer is used to remove some unnecessary artifacts or aliasing. Step 5: Input the data obtained in step 2 into the residual attention Transformer network constructed in step 4 to perform network training and initialize the network parameters. The slope and aspect data extracted in step 3 are then integrated into the network loss function to finally obtain a trained high-precision DEM reconstruction model. The loss function of the model training in step 5 is composed of the global loss function and the terrain feature loss function. The calculation formula is as follows: L=L Global +L Terrain (6) Where, L represents the total loss; L Global represents the global loss; L Terrain Indicates loss of terrain features; Among them, the global loss calculation format is as follows: Where y i Indicates the elevation value of the InSAR hole high-resolution DEM; represents the elevation value of the predicted high-precision DEM; n represents the number of pixels contained in an InSAR hole high-resolution DEM; The terrain feature loss function is composed of slope loss and aspect loss. The specific calculation formula is as follows: L Terrain =λ1L Slope +λ2L Aspect (8) Where, L Slope Indicates slope loss; L Aspect represents the aspect loss; λ1 and λ2 represent the weight ratios of slope and aspect loss respectively; S i Indicates the slope value of the InSAR hole high-resolution DEM; Indicates the slope value of the predicted high-precision DEM; A i Indicates the aspect value of the InSAR hole high-resolution DEM; Indicates the predicted high-precision DEM aspect value; In step 6, the ALOSDEM data is input into the network model in step 5 for prediction, and several high-resolution DEM tiles without holes are output. Finally, the tiles are spliced ​​together to obtain a predicted high-precision DEM covering the entire study area.

2. The high-precision InSAR DEM intelligent reconstruction method based on residual attention transformer according to claim 1 is characterized by: Step 1 includes the following steps: Step 1.1: Convert the format of multiple SAR image data acquired in the same study area, and perform image registration, interferometry, and flattening preprocessing to generate intensity maps, coherence maps, and differential interferograms, and then generate the unwrapped differential interferogram phase. Step 1.2: Independent component analysis (ICA) is used to separate the phase signal of the time series of acquired images to remove atmospheric delay and noise in multi-baseline interferometry. Step 1.3: Use the adaptive discriminant quadratic polynomial fitting method to remove the interference fringes caused by the baseline error, and then subtract the phase obtained by ICA separation from the baseline error to obtain the corrected phase signal; In step 1.4, based on the deformation modeling error in multi-baseline interferometry, a time-series InSAR combination model is constructed between the differential interferometry phase and linear, periodic, environmental, and terrain parameters. The residual elevation is obtained by solving it and then added to the low-resolution DEM elevation value to obtain the high-resolution InSAR point measurement value. In step 1.5, the InSAR data are converted from slant-range or ground-range projection to geographic coordinate projection by geocoding the coordinates of high-resolution InSAR point measurements, thereby generating a high-resolution InSAR hole DEM that matches the actual geographic location.

3. The high-precision InSAR DEM intelligent reconstruction method based on residual attention transformer according to claim 1 is characterized by: Step 2 includes the following steps: Step 2.1: transform the ALOS 12.5m resolution DEM of the study area and project it into the same coordinate system as the high-resolution TerraSAR-X 3m hole DEM. Step 2.2: Select an area with rich topography and large surface undulations in the study area as the training area, and crop it into several N1×N1 pixel blocks. Discard the blocks that are less than N1×N1 pixel blocks at the region boundary, and finally obtain a high-resolution DEM dataset with an integer multiple of N1×N1. In step 2.3, based on the regional boundary obtained in step 2.2, a low-resolution DEM of the same regional range is cropped, and based on the resolution multiple S between the high-resolution and low-resolution DEMs, a number of corresponding N2×N2 pixel-sized tiles are cropped, where N2*S=N1, to obtain a low-resolution DEM dataset of the same range size. Finally, a pair of high-resolution and low-resolution DEM datasets is obtained, which can form a one-to-one mapping relationship between the high-resolution and low-resolution tiles.

4. The high-precision InSAR DEM intelligent reconstruction method based on residual attention transformer according to claim 1 is characterized by: The Transformer residual attention module RATB consists of two enhanced spatial attention modules ESA at the head and tail, a residual feature aggregation module RFA, two Swin Transformer layers and a 1×1 convolution layer; ESA consists of two 1×1 convolution layers, a void convolution, a convolution group consisting of three 3×3 convolution layers and a sigmoid activation function layer.

5. The high-precision InSAR DEM intelligent reconstruction method based on residual attention transformer according to claim 1, characterized in that: RFA consists of 4 residual modules RB and a 1×1 convolution layer. The hierarchical features on the branches of the first three residual modules are concatenated with the output of the last residual module; RB consists of 3 3×3 convolutions and an ESA.

6. The high-precision InSAR DEM intelligent reconstruction method based on residual attention transformer according to claim 1, characterized in that: Two STLs form the SwinTransformer module, and STL consists of two layer normalizations, a multi-head attention mechanism, and a multi-layer perceptron layer MLP.

7. The high-precision InSAR DEM intelligent reconstruction method based on residual attention transformer according to claim 1, characterized in that: It also includes step 7, which evaluates the accuracy of the high-precision DEM predicted by the model, using the mean absolute error (MAE), root mean square error (RMSE), and maximum elevation difference (E max , mean absolute error (MAE) of slope Slope As an evaluation index, the difference between the predicted high-precision DEM and the actual high-resolution DEM is compared to evaluate the reconstruction effect. The calculation formula is as follows: Where n represents the number of pixels contained in an InSAR hole high-resolution DEM; y i Indicates the elevation value of the InSAR hole high-resolution DEM; Indicates the predicted high-precision DEM elevation value; S i Indicates the slope value of the InSAR hole high-resolution DEM; Indicates the slope value of the predicted high-precision DEM.

8. High-precision InSAR DEM intelligent reconstruction system based on residual attention transformer, characterized by: include: A processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the high-precision InSAR DEM intelligent reconstruction method based on residual attention transformer as described in any one of claims 1 to 7.

Images