Digital elevation model reconstruction method and system

Abstract

The application provides a digital elevation model reconstruction method and system, the method comprising: obtaining random undersampled low-resolution elevation data corresponding to a target terrain area; inputting the random undersampled low-resolution elevation data corresponding to the target terrain area into a pre-trained reconstruction network model, and obtaining super-resolution elevation data through task cooperation operation of a spatial interpolation reconstruction branch and a super-resolution reconstruction branch in the reconstruction network model, so as to reconstruct a digital elevation model of the target terrain area. The technical scheme provided by the application can directly reconstruct super-resolution elevation data with high precision, high resolution and high integrity from random discrete undersampled elevation data through task cooperation operation of the spatial interpolation reconstruction branch and the super-resolution reconstruction branch in the reconstruction network model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geographic information science and artificial intelligence, and particularly relates to a digital elevation model reconstruction method and system. BACKGROUND

[0002] A digital elevation model (DEM) is a kind of solid ground model represented by a set of two-dimensional ordered numerical arrays, which accurately reflects the terrain undulations and plays an important role in engineering construction, geological research and analysis, and terrain three-dimensional reconstruction. For example, plate movement research, underwater sediment migration pattern research, water surface and underwater vehicle safe navigation, oil and gas exploration, and other work all need accurate ground elevation data support.

[0003] DEM data can be generated from a variety of sensor data, including radar, light detection and ranging (LiDAR), optical remote sensing images, etc. LiDAR combined with unmanned aerial photography technology can obtain DEM data with large range, high precision and high resolution, so it has been widely used. However, extreme weather or complex terrain (such as deep valleys, steep terrain, etc.) will cause uneven scattering and reflection of laser beams, which will make the laser radar system unable to effectively receive signals, thereby affecting the integrity of the obtained DEM data. Therefore, it is of great significance to select a suitable spatial interpolation method to reconstruct the random discrete undersampling elevation data into complete DEM data, and then reconstruct it into high-precision and high-resolution DEM data through a super-resolution reconstruction algorithm.

[0004] Currently, research on digital elevation model (DEM) reconstruction can be mainly divided into two categories: interpolation-based methods and learning-based methods. Interpolation-based methods include traditional spatial interpolation methods such as inverse distance interpolation, polynomial interpolation, spline interpolation, and Kriging interpolation. These methods are simple and easy to use, but due to their lack of consideration for the specific characteristics of terrain features and low model complexity, they perform poorly in interpolation tasks involving complex terrain. Because of the continuity and heterogeneity of geographic space, its model complexity is high, making it impossible for traditional interpolation methods to meet the requirements of higher accuracy and lower error. In recent years, deep learning technology has made significant progress in the field of image processing, especially demonstrating powerful performance in tasks such as image enhancement, image restoration, and image generation. Learning-based methods learn the inherent knowledge of terrain from large elevation datasets and utilize powerful deep learning models to accurately recover the details and high-frequency information of DEMs. Some researchers have used conditional generative adversarial networks (GANs) to transform the interpolation task into an image generation task, achieving DEM spatial interpolation completion based on random measurement points. Other researchers, inspired by the SRGAN network, proposed the D-SRGAN model, which reconstructs the spatial resolution of DEM data from 50 feet to 3 feet. Still others, by modeling terrain features, proposed a terrain-feature-aware DEM super-resolution reconstruction network model—TfaSR. This model integrates a deep residual module and a deformable convolutional module, achieving the extraction of both deep and adaptive features, and yielding better results than SRGAN.

[0005] However, current research typically treats spatial interpolation completion and super-resolution reconstruction of digital elevation models as highly dependent linear sequence tasks, performing spatial interpolation and super-resolution reconstruction operations sequentially, while ignoring their correlation and complementarity, resulting in low accuracy of digital elevation model reconstruction. Summary of the Invention

[0006] This application provides a digital elevation model reconstruction method and system to solve the problem of low accuracy in digital elevation model reconstruction caused by sequentially performing spatial interpolation and super-resolution reconstruction operations in the prior art.

[0007] Firstly, this application provides a digital elevation model reconstruction method based on attention and multi-task mechanisms, including:

[0008] Acquire low-resolution elevation data of randomly undersampled terrain regions corresponding to the target terrain area;

[0009] The random undersampled low-resolution elevation data corresponding to the target terrain region is input into a pre-trained reconstruction network model. Through the collaborative operation of the spatial interpolation reconstruction branch and the super-resolution reconstruction branch in the reconstruction network model, super-resolution elevation data is obtained to reconstruct the digital elevation model of the target terrain region.

[0010] Optionally, the training process of the reconstruction network model comprises:

[0011] obtaining a training sample set, each sample in the training sample set being random undersampling low-resolution elevation data and original high-resolution elevation data of a known local terrain area;

[0012] constructing a digital elevation model reconstruction network based on attention and multi-task mechanism comprising a spatial interpolation reconstruction branch and a super-resolution reconstruction branch;

[0013] based on the training sample set, iteratively training the digital elevation model reconstruction network to obtain a reconstruction network model.

[0014] Optionally, the digital elevation model reconstruction network based on attention and multi-task mechanism comprising a spatial interpolation reconstruction branch and a super-resolution reconstruction branch is constructed by:

[0015] constructing a spatial interpolation reconstruction branch, the spatial interpolation reconstruction branch comprising a plurality of spatial interpolation reconstruction modules, each spatial interpolation module comprising at least two sequentially arranged dense feature extraction units, each dense feature extraction unit comprising a batch normalization layer, a ReLU activation function layer and a convolution layer connected in sequence, for local feature extraction and terrain feature reconstruction, and each dense feature extraction unit being connected by a skip connection;

[0016] constructing a super-resolution reconstruction branch, the super-resolution reconstruction branch comprising a plurality of super-resolution reconstruction modules and an up-sampling module, each super-resolution reconstruction module comprising at least two sequentially arranged dense feature extraction units, each dense feature extraction unit comprising a batch normalization layer, a ReLU activation function layer and a convolution layer connected in sequence, for local feature extraction and terrain feature reconstruction, and each dense feature extraction unit being connected by a skip connection;

[0017] constructing a task transfer module based on an attention mechanism, the number of the task transfer modules being the same as the number of the super-resolution reconstruction modules, the input of an i-th task transfer module being the output of an i-th super-resolution reconstruction module and an i-th spatial interpolation reconstruction module, and the output of the i-th task transfer module being the input of an i+1-th super-resolution reconstruction module;

[0018] a convolution layer with a convolution kernel size of 1x1 is connected in series at the head of the spatial interpolation reconstruction branch and the super-resolution reconstruction branch, respectively, and a long skip connection is connected at the end of the spatial interpolation reconstruction branch and the super-resolution reconstruction branch, respectively, to obtain a digital elevation model reconstruction network based on attention and multi-task mechanism.

[0019] Optionally, the iterative training of the digital elevation model reconstruction network based on the training sample set obtains a reconstruction network model, including:

[0020] inputting one training sample randomly selected from the training sample set into the digital elevation model reconstruction network, and performing spatial interpolation reconstruction and super-resolution reconstruction through the digital elevation model reconstruction network to obtain reconstructed low-resolution elevation data and reconstructed super-resolution elevation data;

[0021] based on the original high-resolution elevation data and the original low-resolution elevation data of the training sample, calculating a first content loss corresponding to the reconstructed low-resolution elevation data and a second content loss corresponding to the reconstructed super-resolution elevation data by using a first sub-loss calculation equation set;

[0022] calculating a first perceptual loss corresponding to the reconstructed low-resolution elevation data and a second perceptual loss corresponding to the reconstructed super-resolution elevation data by using a second sub-loss calculation equation set;

[0023] performing weighted summation processing on the first content loss, the second content loss, the first perceptual loss and the second perceptual loss by using a joint loss function to obtain a joint loss;

[0024] in the training process, when the change rate of the joint loss is greater than a preset change rate threshold, updating the parameters of the digital elevation model reconstruction network by using a gradient descent algorithm until the change rate of the joint loss is less than or equal to the preset change rate threshold, and obtaining optimal network model parameters;

[0025] determining the digital elevation model reconstruction network with the optimal network model parameters as the reconstruction network model.

[0026] Optionally, the expression of the first sub-loss calculation equation set is:

[0027] ;

[0028] ;

[0029] wherein, is the first content loss, W and H are the width and height of an input image respectively, the input image is an image corresponding to the randomly undersampled low-resolution elevation data in the training sample, i and j are the horizontal and vertical coordinate positions of a pixel point in the image, is the original low-resolution elevation data of the pixel point at the (i, j) position, is the reconstructed low-resolution elevation data of the pixel point at the (i, j) position, is the second content loss, original high-resolution elevation data of a pixel point located at (i, j), reconstructed high-resolution elevation data of a pixel point located at (i, j).

[0030] Optionally, the expression of the second sub-loss calculation equation group is:

[0031] ;

[0032] ;

[0033] wherein, the first perception loss, a feature extraction network, the feature extraction network comprising a pre-trained neural network, N being a size of a number of layers of convolutional layers in the pre-trained neural network, and in the expression, i represents an i-th layer in the pre-trained neural network, original low-resolution elevation data, a feature representation of the original low-resolution elevation data at an i-th layer in the pre-trained neural network, reconstructed low-resolution elevation data, a feature representation of the reconstructed low-resolution elevation data at an i-th layer in the pre-trained neural network, the second perception loss, original high-resolution elevation data, a feature representation of the original high-resolution elevation data at an i-th layer in the pre-trained neural network, reconstructed high-resolution elevation data, a feature representation of the reconstructed high-resolution elevation data at an i-th layer in the pre-trained neural network.

[0034] Optionally, the training sample set is obtained, comprising:

[0035] obtaining high-resolution elevation data of a known overall terrain area;

[0036] segmenting the high-resolution elevation data of the known overall terrain area to obtain original high-resolution elevation data of a known local terrain area;

[0037] performing down-sampling processing on the original high-resolution elevation data of the known local terrain area to obtain original low-resolution elevation data of the known local terrain area;

[0038] performing random sampling processing on the original low-resolution elevation data of the known local terrain area to obtain random under-sampling low-resolution elevation data of the known local terrain area.

[0039] Secondly, this application provides a digital elevation model reconstruction system, comprising:

[0040] The acquisition module is used to acquire randomly undersampled low-resolution elevation data corresponding to the target terrain area;

[0041] The reconstruction module is used to input the randomly undersampled low-resolution elevation data corresponding to the target terrain region into a pre-trained reconstruction network model. Through the collaborative operation of the spatial interpolation reconstruction branch and the super-resolution reconstruction branch in the reconstruction network model, super-resolution elevation data is obtained to reconstruct the digital elevation model of the target terrain region.

[0042] Thirdly, this application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are to be invoked and executed by the processing component to implement a digital elevation model reconstruction method as described in any of the first aspects.

[0043] Fourthly, this application provides a computer storage medium storing a computer program, which, when executed by a computer, implements a digital elevation model reconstruction method as described in any of the first aspects.

[0044] This application provides a method for reconstructing a digital elevation model, which includes: acquiring randomly undersampled low-resolution elevation data corresponding to a target terrain region; inputting the randomly undersampled low-resolution elevation data corresponding to the target terrain region into a pre-trained reconstruction network model; and obtaining complete high-resolution elevation data through the collaborative operation of the spatial interpolation reconstruction branch and the super-resolution reconstruction branch in the reconstruction network model, so as to reconstruct the digital elevation model of the target terrain region.

[0045] The above technical solution adopted in this application has the following advantages:

[0046] In this application, the reconstruction network model includes two branches: a spatial interpolation reconstruction branch and a super-resolution reconstruction branch. The pre-trained reconstruction network model performs a parallel processing flow based on multi-task learning, jointly performing spatial interpolation reconstruction and super-resolution reconstruction. Therefore, it can simultaneously complete DEM spatial interpolation and super-resolution reconstruction tasks. Furthermore, this application can directly reconstruct high-precision, high-resolution, and high-completeness DEM data from randomly discrete undersampled elevation data.

[0047] Furthermore, the reconstruction network model allows for the sharing of feature representations between the two branches through a task transfer module. This network uses a specially designed task transfer module to embed and synthesize the correlation between the two tasks, leveraging knowledge from the spatial interpolation task to accelerate the learning process of the super-resolution reconstruction task and improve reconstruction accuracy. Moreover, compared to various sequential combination methods of other DEM spatial interpolation and super-resolution reconstruction network models, the multi-task model proposed in this application achieves better results. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of a digital elevation model reconstruction method provided by an embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of a digital elevation model reconstruction method based on attention and multi-task mechanism provided by an embodiment of the present invention;

[0051] Figure 3 This is a schematic diagram of the data acquisition for the training digital elevation model provided in an embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram of the structure of the super-resolution reconstruction and spatial interpolation module provided in an embodiment of the present invention;

[0053] Figure 5 This is a schematic diagram of the upsampling module in the super-resolution reconstruction branch provided in an embodiment of the present invention;

[0054] Figure 6 This is a schematic diagram of a task transfer module provided in an embodiment of the present invention;

[0055] Figure 7 This is a schematic diagram of the structure of the digital elevation model reconstruction network provided in an embodiment of the present invention;

[0056] Figure 8 This is a test schematic diagram of the model iterative training process provided in an embodiment of the present invention;

[0057] Figure 9 This is a schematic diagram comparing the reconstruction network model provided in the embodiments of the present invention with the traditional interpolation super-resolution method;

[0058] Figure 10This is a schematic diagram comparing the reconstruction network model provided in the embodiments of the present invention with the classic deep learning super-resolution model;

[0059] Figure 11 This is a schematic diagram of the structure of a digital elevation model reconstruction system provided by an embodiment of the present invention;

[0060] Figure 12 This is a schematic diagram of the structure of a computing device provided in an embodiment of the present invention. Detailed Implementation

[0061] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0062] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 11, 12, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0063] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0064] To address the low accuracy of digital elevation model (DEM) reconstruction caused by sequential spatial interpolation and super-resolution reconstruction operations in existing technologies, this application provides a DEM reconstruction method based on attention and a multi-task mechanism. Inspired by the powerful image processing capabilities and multi-task learning of the Transformer module, this application proposes a DEM reconstruction network based on attention and a multi-task mechanism. It utilizes knowledge of spatial interpolation to optimize the super-resolution reconstruction task, enabling the direct reconstruction of high-precision, high-resolution complete DEM data from randomly discrete, undersampled, low-resolution elevation data, thereby improving the accuracy and completeness of the reconstructed DEM.

[0065] Figure 1This is a flowchart of a digital elevation model reconstruction method provided in an embodiment of this application, such as... Figure 1 As shown, the method includes:

[0066] S11. Obtain low-resolution elevation data of the target terrain region that is randomly undersampled.

[0067] The target terrain area can be a large-scale engineering construction area, an area experiencing extreme weather, or an area with complex terrain (such as deep valleys, steep terrain, etc.). The solution provided in this application embodiment can realize terrain simulation and environmental modeling for large-scale engineering construction areas. In large-scale engineering construction areas, such as high-speed railways, dams, and airports, high-precision DEM data is the foundation for construction planning, earthwork calculation, and environmental impact assessment. Therefore, this application embodiment can first acquire randomly undersampled low-resolution elevation data of the area.

[0068] S12. Input the randomly undersampled low-resolution elevation data corresponding to the target terrain area into the pre-trained reconstruction network model. Through the collaborative operation of the spatial interpolation reconstruction branch and the super-resolution reconstruction branch in the reconstruction network model, obtain the super-resolution elevation data to reconstruct the digital elevation model of the target terrain area.

[0069] The pre-trained reconstruction network model is trained based on the Digital Elevation Model (DEM) reconstruction network, which allows feature representations to be shared between the two branches. This network uses a specially designed task transformation module to embed and synthesize the correlation between the two tasks, leveraging knowledge from the DEM interpolation completion task to accelerate and optimize the learning process of the super-resolution reconstruction task. By executing steps S11-S12, this embodiment can directly reconstruct high-precision, high-resolution complete DEM data from randomly discrete, undersampled, low-resolution elevation data, improving the accuracy and completeness of the reconstructed DEM data.

[0070] It should be noted that the training quality of the reconstruction network model directly affects the reconstruction quality of the digital elevation model (DEM) of the target terrain area during the application phase. Therefore, to improve the reconstruction quality of the DEM of the target terrain area, it is necessary to train the reconstruction network model reasonably and effectively. Furthermore, this application provides a specific example regarding the training process of the reconstruction network model and its subsequent application. Figure 2 As shown:

[0071] Step 21: Obtain the training sample set. Each sample in the training sample set is a random undersampled low-resolution elevation data and the original high-resolution elevation data corresponding to a known local terrain area.

[0072] Optionally, step 21 may include the following process: Step 211, acquiring high-resolution elevation data of the known overall terrain area; Step 212, segmenting the high-resolution elevation data of the known overall terrain area to obtain the original high-resolution elevation data of the known local terrain area; Step 213, downsampling the original high-resolution elevation data of the known local terrain area to obtain the original low-resolution elevation data of the known local terrain area; Step 214, randomly sampling the original low-resolution elevation data of the known local terrain area to obtain the randomly undersampled low-resolution elevation data of the known local terrain area.

[0073] In this embodiment, the training sample set, also known as the dataset, can be understood as a dataset preparation process: First, a DEM dataset for super-resolution reconstruction training and testing is prepared, and outlier removal and gap filling operations are performed on it; then, overlapping segmentation technology is used to divide the dataset into grid blocks of 128×128 pixels in size, such as... Figure 3 The high-resolution DEM shown; the original DEM data of size 128×128 obtained from the first step of random selection. To indicate that it is high resolution, It can also be expressed as The resolution was reduced to one-quarter of the original value using the bicubic interpolation method, resulting in low-resolution DEM data of size 32×32. ,like Figure 3 The low-resolution DEM shown is used as the input for the network model, and then 500 elevation points are randomly sampled. ,like Figure 3 Low-resolution random undersampled DEM in the data.

[0074] Step 22: Construct a digital elevation model reconstruction network based on attention and multi-task mechanisms, including a spatial interpolation reconstruction branch and a super-resolution reconstruction branch. Specifically, Step 22 may include the following process:

[0075] Step 221: Construct a spatial interpolation reconstruction branch. The spatial interpolation reconstruction branch includes multiple spatial interpolation reconstruction modules. Each spatial interpolation reconstruction module can use a dense residual module. The dense residual module includes at least two dense feature extraction units arranged in sequence. Each dense feature extraction unit contains a batch normalization layer, a ReLU activation function layer and a convolutional layer connected in sequence for local feature extraction and terrain feature reconstruction. There are skip connections between each dense feature extraction unit.

[0076] Specifically, such as Figure 4As shown, the spatial interpolation reconstruction module consists of a dense residual module, specifically four dense feature extraction units arranged in sequence. Each dense feature extraction unit contains a batch normalization layer (BN), a ReLU activation function layer, and a convolutional layer (Conv) connected in sequence. The convolutional layer (Conv) can have 128 3×3 convolutional kernels for effective extraction of local features and reconstruction of terrain features. There are skip connections between each dense feature extraction unit.

[0077] Step 222: Construct a super-resolution reconstruction branch. The super-resolution reconstruction branch includes multiple super-resolution reconstruction modules and an upsampling module. Each super-resolution reconstruction module can use a dense residual module. The dense residual module includes at least two dense feature extraction units arranged in sequence. Each dense feature extraction unit contains a batch normalization layer, a ReLU activation function layer and a convolutional layer connected in sequence for local feature extraction and terrain feature reconstruction. There are skip connections between each dense feature extraction unit.

[0078] Specifically, such as Figure 4 As shown, the super-resolution reconstruction module consists of a dense residual module, specifically four dense feature extraction units arranged in sequence. Each dense feature extraction unit contains a batch normalization layer (BN), a ReLU activation function layer, and a convolutional layer (Conv) connected in sequence. The convolutional layer (Conv) can have 128 3×3 convolutional kernels for effective extraction of local features and reconstruction of terrain features. There are skip connections between each dense feature extraction unit.

[0079] Furthermore, the number of super-resolution reconstruction modules and spatial interpolation modules needs to be consistent, for example, four for each. In practical applications, the number of modules can be adjusted according to the data scale, computing resources, or accuracy requirements. For example, when processing higher resolution DEM data, it may be necessary to increase the number of modules to capture more complex terrain features; if resources are limited, the number of modules can be reduced to balance efficiency and accuracy.

[0080] It should be noted that at the end of the last super-resolution reconstruction module, there is an upsampling module, which can be understood as a magnification module. The upsampling module consists of multiple upsampling units connected in sequence and a convolutional layer.

[0081] For example, such as Figure 5 As shown, the upsampling module includes two upsampling units connected in sequence and a convolutional layer. Each upsampling unit includes a convolutional layer Conv with a kernel size of 3×3 and an upsampling block PixelShuffle connected in sequence.

[0082] Step 223: Construct a task transfer module based on the attention mechanism. The number of task transfer modules is the same as the number of super-resolution reconstruction modules. The input of the i-th task transfer module is the output of the i-th super-resolution reconstruction module and the i-th spatial interpolation reconstruction module. The output of the i-th task transfer module is the input of the (i+1)-th super-resolution reconstruction module.

[0083] It should be understood that the attention mechanism specifically refers to the cross-attention mechanism.

[0084] Specifically, such as Figure 6 As shown, the task transfer module mainly consists of three sub-parts: an embedding part for calculating the correlation between feature representations, an attention transfer part for feature transfer, and a soft attention part for feature fusion. This is the output of the i-th super-resolution reconstruction module. V and K To maintain consistency, all outputs are from the i-th spatial interpolation reconstruction module. Conv is a convolutional layer with a 1×1 kernel. Z This is an intermediate output that integrates the dual tasks of spatial interpolation reconstruction and super-resolution reconstruction. This is the final output of the task transfer module.

[0085] Relevance embedding is calculated and K This is achieved through the correlation between them. Specifically, the dot product operation can be used to calculate... and K The similarity. S yes and K The correlation matrix between them, which can also be considered a soft attention feature map or a soft attention weight map, reflects the similarity between the output features of the super-resolution reconstruction module and the output features of the spatial interpolation reconstruction module at different locations. This is achieved through the feature... and K The matrix dot product is obtained. T This represents the attention weight matrix. After obtaining the correlation matrix... S Afterwards, S Normalization yields attention weights, which represent the weights used during feature embedding. K Features at different locations in the middle The importance of each location feature.

[0086] The goal of the attention transfer part is to transfer the reconstruction results features from the spatial interpolation reconstruction branch to the super-resolution reconstruction branch, thereby enabling the super-resolution reconstruction branch to possess spatial interpolation capabilities. This attention transfer part is used to calculate the correlations between input features and generate a new contextual feature matrix. CSpecifically, it adjusts the attention weights in the T matrix accordingly. V The matrix is ​​obtained by weighted summation of the features. This feature representation contains not only local information about the input data but also global contextual information from different tasks. In this way, the task transfer module can capture the potential connections between spatial interpolation reconstruction and super-resolution reconstruction tasks and use these connections to enhance the performance of these two sub-tasks.

[0087] Soft attention modules can adaptively focus on key information in the input data, dynamically adjusting the contribution of different parts to the output, thereby improving the model's ability to process complex information and its performance. Therefore, utilizing the soft attention mechanism for fused features... Z and original input features The aggregation and calculation of the task transfer module's final output are then performed. , represents element-wise addition, used to preserve original input features and enhance task transitions. out is a hyperparameter representing the number of output channels of the convolutional layer Conv, which can be 128 in this application. S is the soft attention weight map.

[0088] In this embodiment, the core function of the task transfer module is to achieve information sharing and feature transfer between the spatial interpolation and super-resolution reconstruction tasks through an attention mechanism. Essentially, it constructs a dynamic knowledge transfer channel, transferring terrain features (such as local continuity and edge structure) learned in the spatial interpolation task to the super-resolution reconstruction task. Simultaneously, it utilizes high-frequency detail information from the super-resolution task to feed back into the interpolation process, forming a closed loop of bidirectional collaborative optimization. The core idea of ​​the task transfer module is to simulate the "whole-part" collaborative mechanism in human cognition: the interpolation task provides the terrain "skeleton," i.e., the overall structure, while the super-resolution task fills in local details. The two interact dynamically through an attention mechanism, ultimately generating DEM data that conforms to both macroscopic terrain patterns and microscopic realism. This design embodies the dialectical relationship between "global constraints and local optimization" in geospatial data reconstruction.

[0089] The application of the task transfer module offers the following technical advantages: 1. Avoiding error accumulation: Traditional methods perform interpolation and super-resolution sequentially, amplifying errors at each stage; the task transfer module achieves "one-step reconstruction" through parallel optimization, significantly reducing accumulated errors. 2. Feature reuse: The global structural information in the interpolation task and the local detail information in the super-resolution task mutually enhance each other, improving the ability to reconstruct complex terrains. 3. Computational efficiency: Compared to training two models independently, the task transfer module reduces the number of parameters and accelerates convergence by sharing feature representations.

[0090] Step 224: Connect a convolutional layer with a kernel size of 1×1 to the head of the spatial interpolation reconstruction branch and the super-resolution reconstruction branch respectively. Both the spatial interpolation reconstruction branch and the super-resolution reconstruction branch have long skip connections that are connected end to end, so as to obtain a digital elevation model reconstruction network based on attention and multi-task mechanism.

[0091] Specifically, such as Figure 7 As shown, the digital elevation model reconstruction network specifically includes two branches: a super-resolution reconstruction branch and a spatial interpolation reconstruction branch. The head of each branch is connected in series with a convolutional layer with a kernel size of 1×1. Figure 7 Since N is 8, the two branches each include 8 super-resolution reconstruction modules and 8 spatial interpolation reconstruction modules, connected by task transfer modules, and both have long skip connections. The super-resolution reconstruction branch outputs the super-resolution DEM reconstructed image at the end. Figure 7 High-resolution DEM in the image.

[0092] For single-channel DEM data, the 1×1 convolutional layer at the beginning can project it into a high-dimensional space and extract initial features, thereby improving the feature extraction capability of subsequent convolutions.

[0093] Step 23: Based on the training sample set, iteratively train the digital elevation model reconstruction network to obtain the reconstruction network model.

[0094] The purpose of this step is to optimize the parameters of the network model based on the constraints of the joint loss function. Specifically, step 23 includes the following process:

[0095] Step 231: A training sample randomly selected from the training sample set is input into the digital elevation model reconstruction network. Spatial interpolation reconstruction and super-resolution reconstruction are performed through the digital elevation model reconstruction network to obtain the reconstructed low-resolution elevation data and the reconstructed super-resolution elevation data.

[0096] Step 232: Based on the original high-resolution elevation data and original low-resolution elevation data of the training samples, calculate the first content loss corresponding to the reconstructed low-resolution elevation data and the second content loss corresponding to the reconstructed super-resolution elevation data using the first sub-loss calculation equation set.

[0097] The expression for the first sub-loss calculation equation set is as follows:

[0098] ;

[0099] ;

[0100] in, The first content loss is represented by W and H, which are the width and height of the input image, respectively. The input image is the image corresponding to randomly undersampled low-resolution elevation data from the training samples, and i and j are the horizontal and vertical coordinates of the pixels in the image. This represents the original low-resolution elevation data of the pixel located at position (i, j). This is the reconstructed low-resolution elevation data for the pixel located at position (i, j). For the second content loss, This is the original high-resolution elevation data for the pixel located at position (i, j). This is the reconstructed high-resolution elevation data for the pixel located at position (i, j).

[0101] The reconstructed low-resolution DEM data is obtained through digital elevation model (DEM) reconstruction network model calculations. and high-resolution DEM data The content loss of DEMs generated at low and high resolutions can be obtained. and .

[0102] Step 233: Calculate the first perceptual loss corresponding to the reconstructed low-resolution elevation data and the second perceptual loss corresponding to the reconstructed super-resolution elevation data using the second sub-loss calculation equation set.

[0103] The expression for the second sub-loss calculation equation set is as follows:

[0104] ;

[0105] ;

[0106] in, For first-perception loss, The feature extraction network can be a pre-trained neural network. For example, a pre-trained VGG16 network is used as the feature extraction network in this embodiment. N This represents the number of convolutional layers in a pre-trained neural network, and can also represent the number of feature vector layers. In this expression, i represents the i-th layer in the pre-trained neural network. This is the original low-resolution elevation data. This represents the feature representation of the original low-resolution elevation data in the i-th layer of a pre-trained neural network. For the reconstructed low-resolution elevation data, This represents the feature representation of the reconstructed low-resolution elevation data in the i-th layer of a pre-trained neural network. For second-sensory loss, This is the original high-resolution elevation data. This represents the feature representation of the original high-resolution elevation data in the i-th layer of a pre-trained neural network. For the reconstructed high-resolution elevation data, This represents the feature representation of the reconstructed high-resolution elevation data in the i-th layer of a pre-trained neural network.

[0107] In this embodiment, the goal of perceptual loss is to minimize the distance between the input image and the target image in the feature space. It focuses more on the reconstruction and restoration of the perceptual quality of the image, that is, the intuitive perception of image quality by the human eye. The target image mentioned above is the image output by the model.

[0108] Step 234: Using the joint loss function, perform a weighted summation of the first content loss, the second content loss, the first perception loss, and the second perception loss to obtain the joint loss.

[0109] The expression for the joint loss function can be:

[0110] ;

[0111] in, For the total joint loss, , , and These are the weights corresponding to each sub-loss. In this embodiment, they can be set to 0.8, 0.8, 0.2, and 0.2, respectively.

[0112] It should be noted that traditional deep learning-based DEM reconstruction methods typically employ a single pixel-level loss function, focusing only on pixel value differences. This can easily lead to blurred DEMs or loss of high-frequency details, such as terrain edges and micro-topography. The joint loss function proposed in this application addresses these issues through a combination of multiple sub-losses. Content loss focuses on reconstruction and restoration at the pixel level, avoiding the generation of blurred or noisy images. Perceptual loss focuses more on the reconstruction and restoration of the image's perceptual quality, i.e., the human eye's intuitive perception of image quality. It constrains the similarity between the generated DEM and the real DEM in the high-level semantic space, improving the perceptual quality of terrain structures.

[0113] Step 235: During training, if the rate of change of the joint loss is greater than the preset rate of change threshold, the parameters of the digital elevation model reconstruction network are updated using the gradient descent algorithm until the rate of change of the joint loss is less than or equal to the preset rate of change threshold, thereby obtaining the optimal network model parameters.

[0114] Specifically, the parameters of the network model are updated using the gradient descent algorithm. The specific expression for the gradient descent algorithm can be found in related technologies; this application does not impose specific limitations on it.

[0115] Step 236: Determine the digital elevation model reconstruction network with the optimal network model parameters as the reconstruction network model.

[0116] Repeat the model training process until the joint loss function converges, thus obtaining the optimal network model parameters. The random undersampled low-resolution DEM is used as input, and the corresponding complete low-resolution DEM data and super-resolution DEM are obtained through calculation by the optimal network model.

[0117] Figure 8 This demonstrates the model evaluation of the Digital Elevation Model (DEM) reconstruction network during training. The super-resolution reconstructed DEM has a pixel size of 128×128, representing a four-fold increase in resolution. The first row of images shows the super-resolution image output by the model after processing the test image, increasing with the number of training iterations. The second row shows a magnified view of the area within the red box in the generated image. The third row shows the elevation value distribution of the generated DEM. Figure 8 As can be seen, the clarity of the generated DEM improves with increasing iterations. After 1-5 iterations, the model can recover the approximate distribution characteristics of the DEM data. After 10-50 iterations, the terrain contour information becomes more distinct. After 50-200 iterations, this application can gradually recover local detailed geographic features, and at the end of training, it has basically achieved accurate super-resolution reconstruction of the DEM.

[0118] like Figure 9 The figure shows a comparison of the super-resolution reconstruction performance of the reconstruction network model provided in this application embodiment with that of four traditional interpolation super-resolution methods. Figure 10 The figure shows a comparison of the super-resolution reconstruction performance of the reconstruction network model provided in this application embodiment with that of four classic deep learning super-resolution models. Combined with... Figure 9 and Figure 10 As can be seen, the super-resolution DEM data generated by the embodiments of this application has higher reconstruction accuracy and richer terrain details. Furthermore, traditional methods based on interpolation and general deep learning models exhibit significant smoothing effects in complex terrain areas (such as steep slopes and gullies), resulting in blurred edges; while the network model provided by the embodiments of this application can clearly restore complex terrain undulations.

[0119] Figure 11 This is a schematic diagram of the structure of a digital elevation model reconstruction system provided in an embodiment of this application, as shown below. Figure 11 As shown, the system includes:

[0120] The acquisition module 111 is used to acquire low-resolution elevation data of random undersampling corresponding to the target terrain area.

[0121] The reconstruction module 112 is used to input the randomly undersampled low-resolution elevation data corresponding to the target terrain area into the pre-trained reconstruction network model. Through the collaborative operation of the spatial interpolation reconstruction branch and the super-resolution reconstruction branch in the reconstruction network model, super-resolution elevation data is obtained to reconstruct the digital elevation model of the target terrain area.

[0122] Figure 11 The aforementioned digital elevation model reconstruction system can perform... Figure 1 The implementation principle and technical effects of the digital elevation model reconstruction method described in the illustrated embodiment will not be repeated here. The specific methods by which each module and unit of the digital elevation model reconstruction system in the above embodiments perform operations have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0123] In one possible design, Figure 11 The digital elevation model reconstruction system of the embodiment shown can be implemented as a computing device, such as... Figure 12 As shown, the computing device may include a storage component 121 and a processing component 122.

[0124] The storage component 121 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 122.

[0125] The processing component 122 is used to perform the following process: acquiring randomly undersampled low-resolution elevation data corresponding to the target terrain region; inputting the randomly undersampled low-resolution elevation data corresponding to the target terrain region into a pre-trained reconstruction network model; and obtaining super-resolution elevation data through the collaborative operation of the spatial interpolation reconstruction branch and the super-resolution reconstruction branch in the reconstruction network model, so as to reconstruct the digital elevation model of the target terrain region.

[0126] The processing component 122 may include one or more processors to execute computer instructions to complete all or part of the steps in the above-described method. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described method.

[0127] Storage component 121 is configured to store various types of data to support operations at the terminal. The storage component can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as Random Access Memory (RAM), Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0128] Of course, computing devices may also include other components, such as input / output interfaces, display components, communication components, etc.

[0129] Input / output interfaces provide interfaces between processing components and peripheral interface modules, which can be output devices, input devices, etc.

[0130] The communication components are configured to facilitate wired or wireless communication between computing devices and other devices.

[0131] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform. In this case, the computing device can refer to a cloud server, and the aforementioned processing components, storage components, etc., can be basic server resources rented or purchased from the cloud computing platform.

[0132] This application also provides a computer storage medium storing a computer program, which, when executed by a computer, can perform the above-described functions. Figure 1 The embodiment shown illustrates a method for reconstructing a digital elevation model.

[0133] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0134] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0135] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for reconstructing a digital elevation model, characterized in that, include: Acquire low-resolution elevation data of randomly undersampled terrain regions corresponding to the target terrain area; Randomly undersampled low-resolution elevation data corresponding to the target terrain region is input into a pre-trained reconstruction network model. Through the collaborative task operations of the spatial interpolation reconstruction branch and the super-resolution reconstruction branch in the reconstruction network model, super-resolution elevation data is obtained to reconstruct the digital elevation model of the target terrain region. The reconstruction network model is trained based on a digital elevation model reconstruction network. The construction of a digital elevation model reconstruction network based on attention and a multi-task mechanism, including a spatial interpolation reconstruction branch and a super-resolution reconstruction branch, includes: A spatial interpolation reconstruction branch is constructed, which includes multiple spatial interpolation reconstruction modules. Each spatial interpolation module includes at least two dense feature extraction units arranged in sequence. Each dense feature extraction unit contains a batch normalization layer, a ReLU activation function layer and a convolutional layer connected in sequence, which are used to perform local feature extraction and terrain feature reconstruction. There are skip connections between each dense feature extraction unit. A super-resolution reconstruction branch is constructed, which includes multiple super-resolution reconstruction modules and an upsampling module. Each super-resolution reconstruction module includes at least two dense feature extraction units arranged in sequence. Each dense feature extraction unit contains a batch normalization layer, a ReLU activation function layer and a convolutional layer connected in sequence, which are used for local feature extraction and terrain feature reconstruction. There are skip connections between each dense feature extraction unit. A task transfer module is constructed based on an attention mechanism. The number of task transfer modules is the same as the number of super-resolution reconstruction modules. The input of the i-th task transfer module is the output of the i-th super-resolution reconstruction module and the i-th spatial interpolation reconstruction module. The output of the i-th task transfer module is the input of the (i+1)-th super-resolution reconstruction module. A convolutional layer with a kernel size of 1×1 is connected in series at the head of the spatial interpolation reconstruction branch and the super-resolution reconstruction branch, and both the spatial interpolation reconstruction branch and the super-resolution reconstruction branch have long skip connections that connect the beginning and the end, thus obtaining a digital elevation model reconstruction network based on attention and multi-task mechanisms.

2. The method according to claim 1, characterized in that, The training process for reconstructing the network model includes: Obtain a training sample set, wherein each sample in the training sample set is a random undersampled low-resolution elevation data and an original high-resolution elevation data corresponding to a known local terrain area; Based on the training sample set, the digital elevation model reconstruction network is iteratively trained to obtain the reconstruction network model.

3. The method according to claim 2, characterized in that, Based on the training sample set, the digital elevation model reconstruction network is iteratively trained to obtain a reconstruction network model, including: A training sample randomly selected from the training sample set is input into the digital elevation model reconstruction network. Spatial interpolation reconstruction and super-resolution reconstruction are performed through the digital elevation model reconstruction network to obtain reconstructed low-resolution elevation data and reconstructed high-resolution elevation data. Based on the original high-resolution elevation data and original low-resolution elevation data of the training samples, the first content loss corresponding to the reconstructed low-resolution elevation data and the second content loss corresponding to the reconstructed super-resolution elevation data are calculated using the first sub-loss calculation equation set. The first perception loss corresponding to the reconstructed low-resolution elevation data and the second perception loss corresponding to the reconstructed super-resolution elevation data are calculated using the second sub-loss calculation equation set. By using the joint loss function, the first content loss, the second content loss, the first perception loss, and the second perception loss are weighted and summed to obtain the joint loss; During training, if the rate of change of the joint loss is greater than a preset rate of change threshold, the parameters of the digital elevation model reconstruction network are updated using the gradient descent algorithm until the rate of change of the joint loss is less than or equal to the preset rate of change threshold, thereby obtaining the optimal network model parameters. The digital elevation model reconstruction network with the optimal network model parameters is determined as the reconstruction network model.

4. The method according to claim 3, characterized in that, The expression for the first sub-loss calculation equation set is: ； ； in, For the loss of the first content, W and H Here, i and j represent the width and height of the input image, respectively. The input image is an image corresponding to randomly undersampled low-resolution elevation data from the training samples, and i and j are the horizontal and vertical coordinates of the pixels in the image. This represents the original low-resolution elevation data of the pixel located at position (i, j). This is the reconstructed low-resolution elevation data for the pixel located at position (i, j). For the second content loss, This is the original high-resolution elevation data for the pixel located at position (i, j). This is the reconstructed high-resolution elevation data for the pixel located at position (i, j).

5. The method according to claim 3, characterized in that, The expression for the second sub-loss calculation equation set is: ； ； in, This is the first perceived loss. The feature extraction network comprises a pre-trained neural network. N Let i be the number of convolutional layers in the pre-trained neural network, where i represents the i-th layer in the pre-trained neural network. This is the original low-resolution elevation data. This represents the feature representation of the original low-resolution elevation data in the i-th layer of a pre-trained neural network. For the reconstructed low-resolution elevation data, This represents the feature representation of the reconstructed low-resolution elevation data in the i-th layer of a pre-trained neural network. For the second perceptual loss, This is the original high-resolution elevation data. This represents the feature representation of the original high-resolution elevation data in the i-th layer of a pre-trained neural network. For the reconstructed high-resolution elevation data, This represents the feature representation of the reconstructed high-resolution elevation data in the i-th layer of a pre-trained neural network.

6. The method according to claim 2, characterized in that, Obtain the training sample set, including: (1) Obtain high-resolution elevation data for a known overall terrain region; (2) The high-resolution elevation data of the known overall terrain region is segmented to obtain the original high-resolution elevation data of the known local terrain region; (3) The original high-resolution elevation data of the known local terrain area is downsampled to obtain the original low-resolution elevation data of the known local terrain area; (4) Random sampling processing is performed on the original low-resolution elevation data of the known local terrain area to obtain random undersampled low-resolution elevation data of the known local terrain area.

7. A digital elevation model reconstruction system, characterized in that, include: The acquisition module is used to acquire randomly undersampled low-resolution elevation data corresponding to the target terrain area; The reconstruction module is used to input randomly undersampled low-resolution elevation data corresponding to the target terrain region into a pre-trained reconstruction network model. Through the collaborative operation of the spatial interpolation reconstruction branch and the super-resolution reconstruction branch in the reconstruction network model, super-resolution elevation data is obtained to reconstruct the digital elevation model of the target terrain region. The reconstruction network model is trained based on the digital elevation model reconstruction network. The construction of the digital elevation model reconstruction network, which includes a spatial interpolation reconstruction branch and a super-resolution reconstruction branch, and is based on attention and a multi-task mechanism, includes: A spatial interpolation reconstruction branch is constructed, which includes multiple spatial interpolation reconstruction modules. Each spatial interpolation module includes at least two dense feature extraction units arranged in sequence. Each dense feature extraction unit contains a batch normalization layer, a ReLU activation function layer and a convolutional layer connected in sequence, which are used to perform local feature extraction and terrain feature reconstruction. There are skip connections between each dense feature extraction unit. A super-resolution reconstruction branch is constructed, which includes multiple super-resolution reconstruction modules and an upsampling module. Each super-resolution reconstruction module includes at least two dense feature extraction units arranged in sequence. Each dense feature extraction unit contains a batch normalization layer, a ReLU activation function layer and a convolutional layer connected in sequence, which are used for local feature extraction and terrain feature reconstruction. There are skip connections between each dense feature extraction unit. A task transfer module is constructed based on an attention mechanism. The number of task transfer modules is the same as the number of super-resolution reconstruction modules. The input of the i-th task transfer module is the output of the i-th super-resolution reconstruction module and the i-th spatial interpolation reconstruction module. The output of the i-th task transfer module is the input of the (i+1)-th super-resolution reconstruction module. A convolutional layer with a kernel size of 1×1 is connected in series at the head of the spatial interpolation reconstruction branch and the super-resolution reconstruction branch, and both the spatial interpolation reconstruction branch and the super-resolution reconstruction branch have long skip connections that are connected end to end, thus obtaining a digital elevation model reconstruction network based on attention and multi-task mechanism.

8. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a digital elevation model reconstruction method as described in any one of claims 1 to 6.

9. A computer storage medium, characterized in that, The device contains a computer program that, when executed by a computer, implements a digital elevation model reconstruction method as described in any one of claims 1 to 6.

Images