Deep learning based method and apparatus for super-resolution reconstruction of united states coastal digital elevation model (dem)

Abstract

This invention specifically relates to a deep learning-based super-resolution reconstruction method for the U.S. coastal digital elevation model (DEM), comprising: proposing an improved deep residual network architecture specifically for the topographic features of the U.S. coast; designing a multi-stage data preprocessing workflow to ensure data consistency; and achieving high-precision conversion of DEM data across resolutions. This method can be widely applied in fields such as coastal zone management, storm surge simulation, and sea-level rise prediction, providing high-precision basic data support for marine geographic information systems. Experiments show that this invention significantly improves objective indicators such as RMSE, PSNR, and SSIM compared to traditional interpolation methods, and the generated topographic feature preservation is significantly better than existing super-resolution algorithms, while also showing a significant improvement in visual quality.

Description

Technical Field

[0001] This invention relates to the field of geographic information processing and artificial intelligence, specifically to a method and apparatus for super-resolution reconstruction of the digital elevation model (DEM) of the U.S. coast based on deep learning. Background Technology

[0002] Digital Elevation Models (DEMs) are crucial foundational data for Geographic Information Systems (GIS) and remote sensing applications. However, the latest publicly available datasets (CUDEMs) have a resolution of only 1 / 3 arcsecond for nearshore DEMs, which often falls short of the high-precision requirements in practical applications. While traditional interpolation methods are simple and easy to implement, they can lead to loss of terrain detail and blurred edges. For example, bicubic interpolation.

[0003] Currently, deep learning technology has made significant progress in the field of image super-resolution, but existing methods have the following problems in DEM data processing: (1) Insufficient ability to represent topographic features; (2) It is difficult to maintain the continuity and accuracy of elevation data; (3) The reconstruction effect in the edge areas is not good; (4) Insufficient fusion of multi-scale features.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] This invention provides a method and apparatus for super-resolution reconstruction of the digital elevation model (DEM) of the U.S. coast based on deep learning, a computer-readable storage medium, and a computer program product, which can effectively overcome the defects existing in the prior art.

[0006] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0007] According to a first aspect of the present invention, a deep learning-based super-resolution reconstruction method for a digital elevation model (DEM) of the U.S. coast is provided, the method comprising: The shallow feature extraction module based on the super-resolution reconstruction model maps the preprocessed initial image to the feature space, extracts shallow features, and obtains the initial feature map; where the pixels of the initial image are the elevation values ​​of a preset region. The deep feature enhancement module is used to perform long-range dependency modeling and multi-scale feature extraction on the initial feature map to obtain a fused feature map. The deep feature enhancement module includes a first long-range dependency modeling unit, a multi-scale feature extraction unit, a second long-range dependency modeling unit, and a feature fusion unit connected in series. Based on the feature reconstruction module, the fused feature map is upsampled and reconstructed to obtain the target image; wherein the resolution of the target image is higher than that of the initial image.

[0008] In some exemplary embodiments, the step of using a deep feature enhancement module to perform long-range dependency modeling and multi-scale feature extraction on the initial feature map to obtain a fused feature map includes: The first long-range dependency modeling unit is used to perform long-range spatial correlation modeling on the initial feature map to obtain a shallow long-range dependency feature map. The shallow long-range dependency feature map is input into the multi-scale feature extraction unit to perform multi-scale feature extraction and enhancement to obtain a multi-scale feature map. The multi-scale feature extraction unit includes at least one residual multi-scale feature extraction layer and at least one residual multi-scale expansion feature layer, and the residual multi-scale feature extraction layer and the residual multi-scale expansion feature layer are connected in series. The second long-range dependency modeling unit is used to perform long-range spatial correlation modeling on the enhanced feature map to obtain a deep long-range dependency feature map. Based on the feature fusion unit, the preprocessed initial image, initial feature map, shallow long-range dependency feature map, multi-scale feature map, deep long-range dependency feature map, and feature maps output by each feature layer in the multi-scale feature extraction unit are fused to obtain a fused feature map; wherein, the feature fusion unit is connected to the long-range dependency modeling unit, the residual multi-scale feature extraction layer, and the residual multi-scale extended feature extraction layer, respectively.

[0009] In some exemplary embodiments, the step of inputting the shallow long-range dependency feature map into the multi-scale feature extraction unit for multi-scale feature extraction and enhancement to obtain a multi-scale feature map includes: Using at least one residual multi-scale feature extraction layer, multi-branch scale transformation processing is performed on the shallow long-range dependency feature map to extract multi-scale terrain detail features. The terrain detail features output by the multi-branch layer are then aggregated and residually superimposed with the input shallow long-range dependency features to obtain the detail feature map. Based on at least one residual multi-scale dilated feature extraction layer, multi-branch scale transformation and dilated convolution are performed on the detail feature map and shallow long-range dependency feature map to extract larger-scale terrain structure features. The terrain structure features are then residually superimposed with the input detail feature map and shallow long-range dependency feature map to obtain a multi-scale feature map.

[0010] In some exemplary embodiments, the feature fusion unit performs feature fusion on the preprocessed initial image, initial feature map, shallow long-range dependency feature map, multi-scale feature map, deep long-range dependency feature map, and feature maps output by each feature layer in the multi-scale feature extraction unit to obtain a fused feature map, including: Based on the feature refinement and fusion layer in the feature fusion unit, the initial feature map, shallow long-range dependency feature map, multi-scale feature map, deep long-range dependency feature map, and feature maps output by each feature layer in the multi-scale feature extraction unit are refined and fused through cross-layer connections to output an enhanced feature map. The preprocessed initial image is passed to the fusion node through a direct connection, and the preprocessed initial image and the enhanced feature map are added and fused at the fusion node to obtain the fused feature map.

[0011] In some exemplary embodiments, the feature reconstruction module upsamples and reconstructs the fused feature map to obtain the target image, including: The feature transformation convolutional layer in the feature reconstruction module is used to perform convolution operations on the fused feature map, and the number of channels of the enhanced feature map is adjusted to the preset number of channels corresponding to the target magnification, so as to obtain an intermediate feature map for upsampling. The upsampled feature map is obtained by periodically rearranging the intermediate feature map based on the pixel rearrangement upsampling layer. The upsampled feature map is reconstructed by using the output convolutional reconstruction layer, which maps the upsampled feature map to the pixel elevation values ​​of the target image, thereby generating the target image.

[0012] In some exemplary embodiments, the shallow feature extraction module includes at least one convolutional layer.

[0013] In some exemplary embodiments, the method further includes: Obtain a sample image dataset; wherein the resolution of the sample images in the sample image dataset is equal to the resolution of the target image; The sample image dataset is preprocessed to construct the model training dataset, and the training and test datasets are input into the initial super-resolution reconstruction model to obtain the pixel loss, perceptual loss and adversarial loss of the initial super-resolution reconstruction model. The pixel loss term, perceptual loss term, and adversarial loss term are weighted and the comprehensive loss function is calculated. The initial super-resolution reconstruction model is trained with the goal of minimizing the comprehensive loss function until the preset termination condition is met, thus obtaining the trained super-resolution reconstruction model.

[0014] In some exemplary embodiments, the preprocessing of the sample image dataset to construct the model training dataset includes: The sample images are standardized to obtain standardized sample images; The standardized sample image is downsampled using an interpolation method to obtain a low-resolution sample image; wherein the resolution of the low-resolution sample image is equal to the resolution of the original image. The low-resolution sample image is cropped according to a preset grid size to obtain sample image blocks, and the edges of the sample image blocks are filled with pixels using a symmetrical mapping method. The filled sample image blocks are divided according to a preset ratio to generate training and test datasets.

[0015] According to a second aspect of the present invention, a deep learning-based super-resolution reconstruction apparatus for a digital elevation model (DEM) of the U.S. coast is provided, the apparatus comprising: The shallow feature extraction module is used for shallow feature extraction based on the super-resolution reconstruction model. It maps the preprocessed initial image to the feature space, extracts shallow features, and obtains the initial feature map; where the pixels of the initial image are elevation values. The deep feature enhancement module is used to perform long-range dependency modeling and multi-scale feature extraction on the initial feature map to obtain a fused feature map. The deep feature enhancement module includes a first long-range dependency modeling unit, a multi-scale feature extraction unit, a second long-range dependency modeling unit, and a feature fusion unit connected in series. The feature reconstruction module is used to upsample and reconstruct the fused feature map based on the feature reconstruction module to obtain the target image; wherein the resolution of the target image is higher than the resolution of the initial image.

[0016] According to a third aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is executed, the device on which the storage medium is located executes the above-described method for super-resolution reconstruction of a deep learning-based digital elevation model (DEM) of the U.S. coast.

[0017] According to a fourth aspect of the present invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the above-described deep learning-based super-resolution reconstruction method for a digital elevation model (DEM) of the U.S. coast.

[0018] According to a fifth aspect of the present invention, an electronic device is provided, comprising: Processor; and Memory for storing the executable instructions of the processor; The processor is configured to implement the aforementioned deep learning-based super-resolution reconstruction method for the U.S. coastal digital elevation model (DEM) via executing the executable instructions.

[0019] The embodiments of this invention provide a deep learning-based super-resolution reconstruction method for the U.S. coastal digital elevation model (DEM). By constructing a super-resolution reconstruction model for elevation raster data, the method performs feature extraction, long-range dependency modeling, multi-scale feature enhancement, feature refinement and fusion, and upsampling reconstruction on low-resolution digital elevation model data. This enables the reconstruction from low-resolution elevation raster data to high-resolution elevation raster data, which can improve the spatial resolution of the digital elevation model while better maintaining the terrain continuity, edge structure, and detailed features, resulting in good reconstruction effect and application value.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0022] Figure 1 The flowchart illustrates an exemplary embodiment of the present invention: a deep learning-based super-resolution reconstruction method for a digital elevation model (DEM) of the U.S. coast. Figure 2 This schematic diagram illustrates the structure of an exemplary embodiment of the present invention, the RMSB. Figure 3 This schematic diagram illustrates the structure of the MBRB, an exemplary embodiment of the present invention. Figure 4 This schematic diagram illustrates the structure of an exemplary embodiment of the present invention, the RMSB. Figure 5 This schematic diagram illustrates the sampling principle of PixelShuffle, an exemplary embodiment of the present invention. Figure 6 This diagram schematically illustrates the network structure of a super-resolution reconstruction model according to an exemplary embodiment of the present invention. Figure 7 This illustration schematically shows an overall process diagram of a deep learning-based super-resolution reconstruction method for a digital elevation model (DEM) of the U.S. coast, as exemplified by an embodiment of the present invention. Figure 8The diagram illustrates a visual effect comparison of exemplary embodiments of the present invention. Figure 9 This illustration schematically shows another visual effect comparison diagram of an exemplary embodiment of the present invention; Figure 10 The illustration shows a schematic diagram of the architecture of a deep learning-based super-resolution reconstruction device for the digital elevation model (DEM) of the U.S. coast, as an exemplary embodiment of the present invention. Detailed Implementation

[0023] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0024] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0025] To address the shortcomings and deficiencies of existing technologies, this example implementation provides a deep learning-based super-resolution reconstruction method for the digital elevation model (DEM) of the U.S. coast. (Reference) Figure 1 As shown, it can specifically include: Step S10: The shallow feature extraction module based on the super-resolution reconstruction model maps the preprocessed initial image to the feature space, extracts shallow features, and obtains the initial feature map; wherein, the pixels of the initial image are the elevation values ​​of the preset region. Step S12: Using the deep feature enhancement module, long-range dependency modeling and multi-scale feature extraction are performed on the initial feature map to obtain a fused feature map; wherein, the deep feature enhancement module includes a first long-range dependency modeling unit, a multi-scale feature extraction unit, a second long-range dependency modeling unit, and a feature fusion unit connected in series. Step S14: Based on the feature reconstruction module, the fused feature map is upsampled and reconstructed to obtain the target image; wherein the resolution of the target image is higher than that of the initial image.

[0026] Based on steps S10 to S14 above, a super-resolution reconstruction model for elevation raster data is constructed. This model performs feature extraction, long-range dependency modeling, multi-scale feature enhancement, feature refinement and fusion, and upsampling reconstruction on low-resolution digital elevation model data, achieving reconstruction from low-resolution elevation raster data to high-resolution elevation raster data. Specifically, a shallow feature extraction module maps the elevation information of the initial image to the feature space, effectively encoding basic terrain information and providing stable input for subsequent deep feature learning. The deep feature enhancement module utilizes long-range dependency modeling and multi-scale feature extraction units... This method employs long-range spatial correlation modeling and multi-scale feature extraction enhancement to improve the joint representation of global terrain structure and local detail information. A feature fusion unit performs cross-layer convergence and refinement fusion of multi-level features, and adds these features at fusion nodes using directly connected paths from the initial image. This preserves the overall elevation trend while enhancing high-frequency detail information, improving the structural continuity and stability of the reconstruction results. A feature reconstruction module performs convolutional transformation, upsampling, and reconstruction on the fused features to improve resolution and generate high-resolution digital elevation model data, enhancing terrain detail recovery and reconstruction accuracy. Therefore, it can improve the spatial resolution of the digital elevation model while effectively maintaining terrain continuity, edge structure, and detail features, resulting in good reconstruction performance and application value.

[0027] The following will describe in more detail each step of a deep learning-based super-resolution reconstruction method for the U.S. coastal digital elevation model (DEM) in this exemplary embodiment, with reference to the accompanying drawings and embodiments.

[0028] For example, in step S10, the shallow feature extraction module is a convolutional layer with a 3×3 kernel, and is non-linearly activated by LeakyReLU to extract shallow feature information and obtain an initial feature map.

[0029] The initial image described above is the image to be reconstructed at a low resolution. For example, the initial image is raw CUDEM 1 / 3rd CONUS data with a resolution of 1 / 3 arcsecond. Here, CUDEM represents the Continuously Updated Digital Elevation Model (CEM) data system (a coastal elevation data product system developed by relevant agencies of the U.S. National Oceanic and Atmospheric Administration), used for coastal topography / depth representation; 1 / 3rd indicates a spatial resolution of one-third arcsecond, serving as a low-resolution input data source; and CONUS indicates the coverage area is the contiguous United States.

[0030] Specifically, the preprocessing of the initial image is as follows: acquire the original CUDEM 1 / 3rdCONUS data with a resolution of 1 / 3 arcsecond, standardize the data using the Z-Score normalization method, perform 32×32 regular grid cropping, and use symmetric mapping for edge filling.

[0031] For example, in step S12, the use of the deep feature enhancement module to perform long-range dependency modeling and multi-scale feature extraction on the initial feature map to obtain a fused feature map includes: Step S121: Use the first long-range dependency modeling unit to perform long-range spatial correlation modeling on the initial feature map to obtain a shallow long-range dependency feature map. Step S122: Input the shallow long-range dependency feature map into the multi-scale feature extraction unit to perform multi-scale feature extraction and enhancement to obtain a multi-scale feature map; wherein, the multi-scale feature extraction unit includes: at least one residual multi-scale feature extraction layer and at least one residual multi-scale expansion feature layer, and the residual multi-scale feature extraction layer and the residual multi-scale expansion feature layer are connected in series. Step S123: Use the second long-range dependency modeling unit to perform long-range spatial correlation modeling on the enhanced feature map to obtain a deep long-range dependency feature map. Step S124: Based on the feature fusion unit, feature fusion is performed on the preprocessed initial image, initial feature map, shallow long-range dependency feature map, multi-scale feature map, deep long-range dependency feature map, and feature maps output by each feature layer in the multi-scale feature extraction unit to obtain a fused feature map; wherein, the feature fusion unit is connected to the long-range dependency modeling unit, the residual multi-scale feature extraction layer, and the residual multi-scale extended feature extraction layer, respectively.

[0032] The first and second long-range dependency modeling units mentioned above are RL-NL model structures.

[0033] The above residual multi-scale feature extraction layer is an RMSB structure.

[0034] The aforementioned residual multiscale expansion feature layer is an RMSDB structure.

[0035] The aforementioned feature fusion unit includes a feature refinement fusion layer (FRF) and fusion points.

[0036] Specifically, the deep feature enhancement module is responsible for processing the initial feature map through multiple units to capture long-range dependencies and multi-scale features, ultimately obtaining a fused feature map. This process is executed sequentially through multiple sub-steps, and finally, the feature fusion module completes the fusion and enhancement of multi-layer features.

[0037] Specifically, the first long-range dependency modeling unit is responsible for long-range spatial correlation modeling of the initial feature map. By capturing the dependencies between distant regions in the image through the RL-NL model structure, it ensures that the network can better understand global structural information when processing images, such as preserving a large range of terrain change trends in complex terrain. The output retains preliminary global information.

[0038] A multi-scale feature extraction unit is used to extract and enhance shallow long-range dependent feature maps at multiple scales, extracting more feature details from different scales and improving the ability to capture local details. Specifically, the RMSB structure in this unit is responsible for extracting local terrain features at different scales, further enhancing image details. The RMSDB structure can extract features over a wider range through dilated convolution to capture terrain structures at larger scales. The output of this step is a multi-scale feature map, which combines information from multiple scales and preserves terrain features at different scales.

[0039] The second long-range dependency modeling unit performs long-range spatial correlation modeling on multi-scale feature maps to obtain deep long-range dependency feature maps. This further enhances the understanding of the global structure, especially after multi-scale feature extraction, by combining deep features to enhance the spatial dependencies of the entire image, thereby improving the network's global understanding and long-distance relationship modeling.

[0040] In the feature fusion unit, features from different modules are fused together to obtain the final fused feature map. Input features are fused and enhanced using a Functional Randomization (FRF) and fusion points to ultimately obtain the fused feature map. The FRF optimizes the fusion process, ensuring that different features achieve optimal results during fusion, minimizing information loss, and enhancing high-quality details. The fusion points further help integrate multi-level features according to weights to ultimately generate a fused feature map with high resolution and fine structure.

[0041] For example, in step S122, the step of inputting the shallow long-range dependency feature map into the multi-scale feature extraction unit for multi-scale feature extraction and enhancement to obtain a multi-scale feature map includes: Step S21: Using at least one residual multi-scale feature extraction layer, perform multi-branch scale transformation processing on the shallow long-range dependency feature map to extract multi-scale terrain detail features. Then, after pooling the terrain detail features output by the multi-branch layer, perform residual superposition with the input shallow long-range dependency features to obtain the detail feature map. For details, please refer to Figure 2 As shown, Figure 2This is a schematic diagram of the RMSB structure. The RMSB employs a complex multi-branch design with parallel multi-branch residual blocks (MBRBs). These parallel-processed features are merged with the original input through residual connections to preserve key low-level information. To maximize feature utilization, all MBRB outputs are concatenated and processed through a channel attention mechanism. This mechanism dynamically recalibrates the importance of channel features, effectively amplifying discriminative features while suppressing less informative ones. The attention-weighted features are then refined through 1×1 convolutional layers and fused with the module's initial input via additional residual connections, completing the complex feature enhancement pipeline of the RMSB.

[0042] Further, refer to Figure 3 As shown, Figure 3 This is a schematic diagram of the MBRB structure. Each MBRB processes input features through two complementary paths: (1) the upper branch has two 3×3 convolutional layers for hierarchical feature extraction; (2) the lower branch has a single 3×3 convolutional layer for efficiently capturing local features. After the features from the upper and lower parts are extracted and merged, they are connected to the input via residuals to further improve efficiency.

[0043] Step S22: Based on at least one residual multi-scale dilated feature extraction layer, perform multi-branch scale transformation and dilated convolution on the detail feature map and shallow long-range dependency feature map to extract larger-scale terrain structure features, and perform residual superposition of the terrain structure features with the input detail feature map and shallow long-range dependency feature map to obtain a multi-scale feature map.

[0044] For details, please refer to Figure 4 As shown, Figure 4 This is a schematic diagram of the RMSB structure. RMSDB is a high-level extension of the RMSB structure. It combines a set of stacked MBRBs with parallel dilated convolutional branches utilizing multiple dilation rates. This hybrid configuration effectively expands the receptive field, enabling each output pixel to capture rich multi-scale background information. After summing the features of all MBRBs, the merged output is passed through dilated convolutional branches and then refined using a channel attention (CA) mechanism, which adaptively emphasizes semantically formed channels. Dimensionality reduction is then applied using 1×1 convolutions, and the result is fused with the original input of the module via residual connections to generate an intermediate RMSDB output.

[0045] Furthermore, the input first passes through MBRB, with each block consisting of multiple convolutional layers and activation functions (such as PreLU) responsible for extracting local details of the input features. Within these modules, the input features are processed through multiple branches, each extracting different types of features, and finally converged within the block to retain important information and eliminate redundancy.

[0046] Then, the output of MBRB is aggregated, which combines features from different sources to enhance the network's ability to express diverse features.

[0047] The aggregated features are then passed through multiple 3×3 convolutional layers (Conv3×3) and the PreLU activation function for further feature extraction and nonlinear transformation. These convolutional layers are responsible for capturing more detailed local information, while the activation function increases the network's expressive power.

[0048] Following the convolutional and activation layers, a channel attention module is applied. This module aims to weight the features of each channel, enabling the network to adaptively focus on more important features and further enhance the effectiveness of feature representation. After channel attention processing, the features are concatenated with the previous features, merging features from different stages to form a richer feature representation.

[0049] Finally, the feature map is shaped by a 1×1 convolutional layer (Conv1×1) to adjust the number of channels and then passed to the RL-NL layer for long-range spatial dependency modeling, ultimately outputting an enhanced feature map.

[0050] By employing various operations such as multi-branch residual blocks, convolutional layers, activation functions, and channel attention mechanisms, the expressive power of features is gradually improved. At the same time, the global dependencies of the image are further modeled through RL-NL layers, providing more accurate feature representations for subsequent image reconstruction or processing.

[0051] For example, in step S124, the feature fusion unit performs feature fusion on the preprocessed initial image, initial feature map, shallow long-range dependency feature map, multi-scale feature map, deep long-range dependency feature map, and feature maps output by each feature layer in the multi-scale feature extraction unit to obtain a fused feature map, including: Step S31: Based on the feature refinement fusion layer in the feature fusion unit, the initial feature map, shallow long-range dependency feature map, multi-scale feature map, deep long-range dependency feature map, and feature maps output by each feature layer in the multi-scale feature extraction unit are refined and fused through cross-layer connections to output an enhanced feature map. Step S32: The preprocessed initial image is passed to the fusion node through a direct connection, and the preprocessed initial image and the enhanced feature map are added and fused at the fusion node to obtain the fused feature map.

[0052] Specifically, based on the feature refinement fusion layer in the feature fusion unit, the initial feature map, shallow long-range dependency feature map, multi-scale feature map, deep long-range dependency feature map, and feature maps output from each feature layer in the multi-scale feature extraction unit are refined and fused through cross-layer connections. This step further enhances the expressive power of features by fusing feature information from different stages and scales, outputting an enhanced feature map. The feature refinement fusion layer adjusts the fused features from each layer through adaptive weight adjustment to ensure that important features are strengthened while removing invalid information.

[0053] Based on this, the preprocessed initial image is passed to the fusion node through a direct connection, where it is added to and fused with the enhanced feature map. Through residual connections, low-frequency basic image information is combined with the deeply processed enhanced feature map to obtain the final fused feature map. This fused feature map preserves the topographical trends of the original image and, combined with the details and structural information extracted from deep layers, further enhances the image reconstruction quality.

[0054] For example, in step S14, the feature reconstruction module upsamples and reconstructs the fused feature map to obtain the target image, including: Step S141: Using the feature transformation convolutional layer in the feature reconstruction module, perform convolution operation on the fused feature map, adjust the number of channels of the enhanced feature map to the preset number of channels corresponding to the target magnification, and obtain the intermediate feature map for upsampling. Step S142: Based on the pixel rearrangement upsampling layer, the intermediate feature map is periodically rearranged to obtain the upsampled feature map; Step S143: Using the output convolutional reconstruction layer, the upsampled feature map is reconstructed by convolution, and the upsampled feature map is mapped to the pixel elevation values ​​of the target image to generate the target image.

[0055] The aforementioned feature transformation convolutional layer is Conv3x3.

[0056] The pixel rearrangement upsampling layer mentioned above is PixelShuffle.

[0057] The above output convolutional reconstruction layer is Conv3x3.

[0058] The target image has a higher resolution than the initial image. For example, the initial image is the raw CUDEM1 / 3rd CONUS data with a resolution of 1 / 3 arcsecond. The preprocessed initial image is input into the super-resolution reconstruction model, and the corresponding target image with a resolution of 1 / 9 arcsecond is output.

[0059] For details, please refer to Figure 5 As shown, Figure 5This is a schematic diagram illustrating the sampling principle of PixelShuffle. Utilizing advanced PixelShuffle operations, high-quality resolution enhancement is achieved through intelligent channel-to-spatial transformation. This sophisticated technique fundamentally reorganizes multi-channel feature information into an expanded spatial dimension while preserving key topographic relationships in elevation data. Mathematically, this operation transforms the input tensor into a spatial dimension through the periodic rearrangement of channel elements. Convert to higher resolution output Here, b represents the batch size, c represents the number of output channels, h and w represent the spatial dimensions of the input feature map, and r is an integer boosting factor. The PixelShuffle operation has three main advantages: (1) it preserves all original feature information while improving spatial resolution; (2) it facilitates learnable upsampling when used in conjunction with preceding convolutional layers; and (3) it effectively reduces checkerboard artifacts associated with transposed convolution operations. This mechanism has proven particularly advantageous for DEM super-resolution tasks because it maintains the integrity of fine-grained elevation gradients and ensures terrain continuity throughout the upsampling process.

[0060] For example, the shallow feature extraction module includes at least one convolutional layer.

[0061] For details, please refer to Figure 6 As shown, Figure 6 This is a schematic diagram of the network structure of the super-resolution reconstruction model (DEM-SRADSGAN) in this invention. The dashed lines in the diagram represent N identical solid lines. Figure 6 The model structure in this paper is based on ResNet, with embedded multi-scale feature extraction units and feature fusion units to construct an improved deep residual network, resulting in DEM-SRADSGAN.

[0062] Specifically, Figure 6 The model architecture has three core modules: (1) a shallow feature extraction module (yellow), used for initial data acquisition; (2) a deep feature enhancement module (green), which contains two RL-NL units to capture long-distance spatial dependencies, and combines residual multiscale blocks (RMSB) and residual multiscale dilated blocks (RMSDB) for multi-level feature enhancement; and (3) a feature reconstruction module (blue), which uses dual 3×3 convolutional layers and pixel convolutional layers to generate high-quality target images.

[0063] Furthermore, the input image first passes through a convolutional layer with a 3×3 kernel and undergoes non-linear activation using LeakyReLU to extract shallow feature information. Subsequently, 12 residual groups (RGs) are input to further extract deeper features. Simultaneously, a multi-scale feature extraction module (MSB) performs multi-scale convolution operations on the input image to extract local and global information, enhancing the model's ability to capture features at different scales. The output of the MSB is added to the deep features and then sequentially passes through a channel-wise global attention mechanism (CGAM) and a spatial global attention mechanism (SGAM) to strengthen long-distance feature dependencies and improve feature representation. Subsequently, the features enter an upsampling layer and are upsampled using the PixelShuffle method. Finally, a super-resolution image is generated through the output convolutional layer. Benchmark tests confirm that DEM-SRADSGAN outperforms existing state-of-the-art methods, significantly improving perception quality and terrain classification accuracy in various remote sensing applications.

[0064] Specifically, the low-resolution image (LR Image) is input into DEM-SRADSGAN. The image first passes through a 3×3 convolutional layer to extract preliminary features and reduce the spatial dimensionality of the image. Next, the image passes through an RL-N (Residual-Negative ...

[0065] The image is processed through a Feature Refinement and Fusion (FRF) layer, which refines and fuses features from different levels to obtain an enhanced feature map. Through residual connections, low-frequency information from the input image is directly passed to the fusion node and added to the enhanced feature map, ensuring that the output high-resolution image retains its basic structure while effectively enhancing details. The output fused feature map then enters a 3×3 convolutional layer to further adjust the number of feature channels. Next, a pixel shuffle operation is performed for upsampling to improve image resolution. Finally, another 3×3 convolutional layer generates the final high-resolution image (HR Image).

[0066] For example, the method further includes: Step S161: Obtain the sample image dataset; wherein the resolution of the sample images in the sample image dataset is equal to the resolution of the target image; Step S162: Preprocess the sample image dataset to construct the model training dataset, and input the training test dataset into the initial super-resolution reconstruction model to obtain the pixel loss, perceptual loss and adversarial loss of the initial super-resolution reconstruction model. Step S163: Weight the pixel loss term, perceptual loss term, and adversarial loss term, and calculate the comprehensive loss function; Step S164: With minimizing the comprehensive loss function as the optimization objective, train the initial super-resolution reconstruction model until the preset termination condition is met, and obtain the trained super-resolution reconstruction model.

[0067] In step S163, the comprehensive loss function is:

[0068] in, For pixel loss terms, For the perceived loss term, To counteract the loss.

[0069] Furthermore, pixel loss is used to directly measure the difference between the generated image and the real image in pixel space, employing L1 loss to calculate the absolute error between them. Compared to L2 loss, L1 loss is more sensitive to outliers (such as edges, texture details, etc.), helping to generate smoother and more realistic super-resolution images, and it converges more easily. Pixel loss primarily ensures the accuracy of the generated image in low-frequency information, thereby guaranteeing the consistency of the overall image structure and basic details.

[0070] The perceptual loss is calculated in the feature space by using a pre-trained 19-layer Visual Geometric Group (VGG) network to compute the perceptual difference between the generated image and the real image. Specifically, the model uses the VGG network to extract feature representations and calculates the Euclidean distance between the generated image and the real image in the feature space, with the loss form shown below:

[0071] Where l represents the VGG network The perceptual loss layer is the l-th layer. Because it compares from deeper feature layers, it can better capture the texture information and semantic content of the image, compensating for the shortcomings of pixel loss in reconstructing complex visual features. Therefore, the perceptual loss helps improve the subjective visual quality of the image, ensuring the model performs well in restoring fine edges and complex textures.

[0072] Adversarial loss, drawing inspiration from Jolicoeur-Martineau et al., is based on the GAN framework and uses adversarial learning to make generated images more realistic. The mathematical expression of adversarial loss is shown below:

[0073] in, This represents the expected value of a small batch of data during training. The introduction of adversarial loss can effectively improve the sensory quality of the generated image, especially in terms of local texture and edge sharpness, making the reconstructed image more realistic.

[0074] In one alternative embodiment, the weights of pixel loss, perceptual loss, and adversarial loss are set to α=100, β=0.1, and γ=1, respectively. This collaborative optimization strategy of multiple loss functions has significant advantages in balancing reconstruction accuracy and visual effects, and can effectively improve the super-resolution reconstruction effect of complex nearshore terrain.

[0075] For example, in step S162, the preprocessing of the sample image dataset to construct the model training dataset includes: Step S41: Standardize the sample images to obtain standardized sample images; Step S42: Downsample the standardized sample image based on interpolation to obtain a low-resolution sample image; wherein the resolution of the low-resolution sample image is equal to the resolution of the initial image; Step S43: The low-resolution sample image is cropped according to the preset grid size to obtain sample image blocks, and the edges of the sample image blocks are filled with pixels using a symmetrical mapping method. Step S44: Divide the filled sample image blocks according to a preset ratio to generate training dataset and test dataset.

[0076] Specifically, the raw CUDEM 1 / 9rd CONUS data at 1 / 9 arcsecond resolution is obtained, and the data is standardized using the Z-Score normalization method to obtain standardized sample images. Cubic interpolation is used to downsample to 1 / 3 arcsecond resolution to obtain low-resolution sample images. The low-resolution sample images are cropped into 96×96 regular grids to obtain sample image blocks, and edge filling is performed using symmetric mapping. The filled sample image blocks are divided into training and test sets according to the ratio of 80% and 20%, respectively.

[0077] For example, this example implementation provides a deep learning-based super-resolution reconstruction method for the digital elevation model (DEM) of the U.S. coast. (See references.) Figure 7As shown in the figure, this diagram illustrates the overall flow of the digital elevation model super-resolution reconstruction method in this embodiment of the invention, including model construction, data preprocessing, model training, inference output, and result evaluation, as detailed below: First, model building and data preprocessing are performed: on the one hand, a super-resolution reconstruction model is constructed; on the other hand, the digital elevation model super-resolution dataset is preprocessed to form the data foundation for training. Then, the model is trained based on the constructed model architecture and the preprocessed dataset to obtain the trained super-resolution reconstruction model.

[0078] During the inference phase, the low-resolution image is input into the trained super-resolution reconstruction model, which outputs the corresponding high-resolution image. Simultaneously, a real high-resolution image corresponding to the output high-resolution image is obtained as a reference.

[0079] Finally, the results are evaluated based on the output high-resolution image and the real high-resolution image, including the calculation of the root mean square error index (RMSE), peak signal-to-noise ratio index (PSNR), and structural similarity index (SSIM), and the visual effects are compared and analyzed to evaluate the super-resolution reconstruction performance of the method of the present invention.

[0080] Specifically, during the training phase, the weights of the L1 regularization term, perceptual loss term, and adversarial loss term were 100, 1, and 1, respectively. An Adam optimizer with a learning rate of 1e-5 was used to train the model on a large dataset. An exponential decay formula was employed to adjust the learning rate. The batch size and momentum rate were set to 64 and 0.9, respectively. To improve model training performance, the learning rate was decreased by 0.9 when the model's performance on the validation set did not improve for three consecutive batches. The total number of model epochs was set to 150.

[0081] As shown in Table 1, the DEM-SRADSGAN model exhibits the lowest reconstruction error, with an average RMSE of 0.01185, significantly outperforming the traditional bicubic interpolation method (average RMSE: 0.02541) and all corresponding GAN-based models. Notably, DEM-SRADSGAN demonstrates superior robustness in highly complex terrain scenarios such as TEST3, TEST5, and TEST10, where traditional models typically exhibit higher reconstruction errors. In these challenging test cases, compared to the suboptimal model (NDSRGAN, RMSE: 0.01198), DEM-SRADSGAN reduces the RMSE by 38.7% to 62.3%, highlighting its superior ability to preserve complex terrain features.

[0082] Table 1. RMSE test results of the US coastal DEM super-resolution model and other super-resolution models at a magnification of 3.

[0083] The PSNR results (Table 2) further confirm this advantage, with DEM-SRADSGAN averaging 53.30 dB, approximately 1.2 dB higher than NDSRGAN and 4.7 dB higher than bicubic baselines. This significant improvement in signal fidelity is particularly important for elevation data, as even small errors can have a major impact on the accuracy of hydrological modeling and flood forecasting.

[0084] Table 2. Comparison of PSNR results between the US coastal DEM super-resolution model and other super-resolution models at a magnification of 3.

[0085] Table 3, showing the structural preservation results evaluated using SSIM, further highlights the advantages of our proposed method. DEM-SRADSGAN achieves near-perfect reconstruction quality with an average SSIM of 0.995056, demonstrating consistently excellent performance across different terrain types, particularly in the following aspects: (1) topographic map continuity in coastal transition zones (TEST3: 0.99512), (2) water depth pattern preservation in shallow marine environments (TEST5: 0.99508), and (3) fine-scale topographic texture reconstruction in urban coastal areas (TEST10: 0.99501). Although NDSRGAN's SSIM performance is competitive (average: 0.995043), careful analysis reveals that DEM-SRADSGAN can more effectively preserve high-frequency features, achieving an SSIM 0.4%-1.2% higher in areas with rapid elevation changes. In contrast, traditional GAN-based architectures (such as SRGAN and SRAGAN) exhibit significant limitations in preserving complex structural relationships, especially in areas where natural terrain and artificial features are integrated, leading to a decrease in terrain consistency.

[0086] Table 3. Comparison of SSIM results between the US coastal DEM super-resolution model and other super-resolution models at a magnification of 3.

[0087] Further reference Figure 8 As shown, Figure 8The images present a visual comparison of the reconstruction results. (a) shows a low-resolution DEM at 1 / 3 arcsecond, (b) a high-resolution DEM at 1 / 9 arcsecond, (c) the BICUBIC interpolation result, (d) the EDSR reconstruction result, (e) the SRGAN reconstruction result, (f) the SRGAN reconstruction result, (g) the NDSRGAN reconstruction result, and (h) the DEM-SRADSGAN reconstruction result. Each sub-image includes a magnified area in the lower right corner, marked with a red rectangle in the original image, highlighting key details for a more intuitive comparison of the methods' effectiveness in detail reconstruction.

[0088] refer to Figure 9 As shown, Figure 9 The images present a visual comparison of the effects. (a) shows a low-resolution DEM at 1 / 3 arcsecond, (b) a high-resolution DEM at 1 / 9 arcsecond, (c) the BICUBIC interpolation result, (d) the EDSR reconstruction result, (e) the SRGAN reconstruction result, (f) the SRGAN reconstruction result, (g) the NDSRGAN reconstruction result, and (h) the DEM-SRADSGAN reconstruction result. Each sub-image includes a magnified area in the lower right corner, marked with a red rectangle in the original image, highlighting key details for a more intuitive comparison of the effectiveness of each method in detail reconstruction.

[0089] The comparison results show that traditional bicubic interpolation struggles to accurately reconstruct image details, leading to a loss of smoothness and sharpness in DEM images. SRGAN, SRAGAN, and NDSRGAN are not precise enough in detail processing, resulting in discontinuities when reconstructing man-made structures such as roads and railways. Among these, the DEM-SRADSGAN model significantly outperforms all of these models. It excels not only in contour reconstruction but also in preserving and enhancing high-frequency details, exhibiting a clear advantage in detail compared to other models.

[0090] For example, this example embodiment provides a deep learning-based super-resolution reconstruction apparatus for the digital elevation model (DEM) of the U.S. coast. Figure 10 As shown, it can specifically include: The shallow feature extraction module 1001 is used for shallow feature extraction based on the super-resolution reconstruction model. It maps the preprocessed initial image to the feature space, extracts shallow features, and obtains the initial feature map; where the pixels of the initial image are elevation values. The deep feature enhancement module 1002 is used to perform long-range dependency modeling and multi-scale feature extraction on the initial feature map to obtain a fused feature map; wherein, the deep feature enhancement module includes a first long-range dependency modeling unit, a multi-scale feature extraction unit, a second long-range dependency modeling unit, and a feature fusion unit connected in series. The feature reconstruction module 1003 is used to upsample and reconstruct the fused feature map based on the feature reconstruction module to obtain the target image; wherein the resolution of the target image is higher than the resolution of the initial image.

[0091] The beneficial effects of this invention are as follows: 1) An improved deep residual network architecture specifically designed for the topographic features of the US coastal region was proposed; 2) A multi-stage data preprocessing workflow was designed to ensure data consistency; 3) High-precision conversion of DEM data across resolutions has been achieved.

[0092] This invention can be widely applied to fields such as coastal zone management, storm surge simulation, and sea level rise prediction, providing high-precision basic data support for marine geographic information systems.

[0093] Experiments show that the present invention significantly improves objective indicators such as RMSE, PSNR and SSIM compared with traditional interpolation methods, and the generated terrain feature preservation is significantly better than existing super-resolution algorithms, while also showing a significant improvement in visual effect.

[0094] It should be noted that the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Furthermore, it is readily understood that these processes may, for example, be executed synchronously or asynchronously in multiple modules.

[0095] It should be noted that although several modules or units of the device for performing actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0096] It should be noted that the storage medium shown in the embodiments of the present invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, wherein computer-readable program code is carried. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any storage medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0097] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0098] The units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0099] It should be noted that, as another aspect, this application also provides a storage medium, which may be included in an electronic device or may exist independently without being assembled into the electronic device. The aforementioned storage medium carries one or more programs, which, when executed by an electronic device, cause the electronic device to perform the methods described in the following embodiments. For example, the electronic device may perform... Figure 1 The steps of the method shown.

[0100] In one embodiment, this application provides a computer program product including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0101] Furthermore, the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0102] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0103] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A deep learning-based super-resolution reconstruction method for the digital elevation model (DEM) of the U.S. coast, characterized in that, The method includes: The shallow feature extraction module based on the super-resolution reconstruction model maps the preprocessed initial image to the feature space, extracts shallow features, and obtains the initial feature map; where the pixels of the initial image are the elevation values ​​of a preset region. The deep feature enhancement module is used to perform long-range dependency modeling and multi-scale feature extraction on the initial feature map to obtain a fused feature map. The deep feature enhancement module includes a first long-range dependency modeling unit, a multi-scale feature extraction unit, a second long-range dependency modeling unit, and a feature fusion unit connected in series. Based on the feature reconstruction module, the fused feature map is upsampled and reconstructed to obtain the target image; wherein the resolution of the target image is higher than that of the initial image.

2. The method according to claim 1, characterized in that, The deep feature enhancement module performs long-range dependency modeling and multi-scale feature extraction on the initial feature map to obtain a fused feature map, including: The first long-range dependency modeling unit is used to perform long-range spatial correlation modeling on the initial feature map to obtain a shallow long-range dependency feature map. The shallow long-range dependency feature map is input into the multi-scale feature extraction unit to perform multi-scale feature extraction and enhancement to obtain a multi-scale feature map. The multi-scale feature extraction unit includes at least one residual multi-scale feature extraction layer and at least one residual multi-scale expansion feature layer, and the residual multi-scale feature extraction layer and the residual multi-scale expansion feature layer are connected in series. The second long-range dependency modeling unit is used to perform long-range spatial correlation modeling on the enhanced feature map to obtain a deep long-range dependency feature map. Based on the feature fusion unit, the preprocessed initial image, initial feature map, shallow long-range dependency feature map, multi-scale feature map, deep long-range dependency feature map, and feature maps output by each feature layer in the multi-scale feature extraction unit are fused to obtain a fused feature map; wherein, the feature fusion unit is connected to the long-range dependency modeling unit, the residual multi-scale feature extraction layer, and the residual multi-scale extended feature extraction layer, respectively.

3. The method according to claim 2, characterized in that, The step of inputting the shallow long-range dependency feature map into the multi-scale feature extraction unit for multi-scale feature extraction and enhancement to obtain a multi-scale feature map includes: Using at least one residual multi-scale feature extraction layer, multi-branch scale transformation processing is performed on the shallow long-range dependency feature map to extract multi-scale terrain detail features. The terrain detail features output by the multi-branch layer are then aggregated and residually superimposed with the input shallow long-range dependency features to obtain the detail feature map. Based on at least one residual multi-scale dilated feature extraction layer, multi-branch scale transformation and dilated convolution are performed on the detail feature map and shallow long-range dependency feature map to extract larger-scale terrain structure features. The terrain structure features are then residually superimposed with the input detail feature map and shallow long-range dependency feature map to obtain a multi-scale feature map.

4. The method according to claim 2, characterized in that, The feature fusion unit performs feature fusion on the preprocessed initial image, initial feature map, shallow long-range dependency feature map, multi-scale feature map, deep long-range dependency feature map, and feature maps output from each feature layer in the multi-scale feature extraction unit to obtain a fused feature map, which includes: Based on the feature refinement and fusion layer in the feature fusion unit, the initial feature map, shallow long-range dependency feature map, multi-scale feature map, deep long-range dependency feature map, and feature maps output by each feature layer in the multi-scale feature extraction unit are refined and fused through cross-layer connections to output an enhanced feature map. The preprocessed initial image is passed to the fusion node through a direct connection, and the preprocessed initial image and the enhanced feature map are added and fused at the fusion node to obtain the fused feature map.

5. The method according to claim 1, characterized in that, The feature reconstruction module upsamples and reconstructs the fused feature map to obtain the target image, including: The feature transformation convolutional layer in the feature reconstruction module is used to perform convolution operations on the fused feature map, and the number of channels of the enhanced feature map is adjusted to the preset number of channels corresponding to the target magnification, so as to obtain an intermediate feature map for upsampling. The upsampled feature map is obtained by periodically rearranging the intermediate feature map based on the pixel rearrangement upsampling layer. The upsampled feature map is reconstructed by using the output convolutional reconstruction layer, which maps the upsampled feature map to the pixel elevation values ​​of the target image, thereby generating the target image.

6. The method according to claim 1, characterized in that, The shallow feature extraction module contains at least one convolutional layer.

7. The method according to any one of claims 1 to 6, characterized in that, The method further includes: Obtain a sample image dataset; wherein the resolution of the sample images in the sample image dataset is equal to the resolution of the target image; The sample image dataset is preprocessed to construct the model training dataset, and the training and test datasets are input into the initial super-resolution reconstruction model to obtain the pixel loss, perceptual loss and adversarial loss of the initial super-resolution reconstruction model. The pixel loss term, perceptual loss term, and adversarial loss term are weighted and the comprehensive loss function is calculated. The initial super-resolution reconstruction model is trained with the goal of minimizing the comprehensive loss function until the preset termination condition is met, thus obtaining the trained super-resolution reconstruction model.

8. The method according to claim 1, characterized in that, The preprocessing of the sample image dataset to construct the model training dataset includes: The sample images are standardized to obtain standardized sample images; The standardized sample image is downsampled using an interpolation method to obtain a low-resolution sample image; wherein the resolution of the low-resolution sample image is equal to the resolution of the original image. The low-resolution sample image is cropped according to a preset grid size to obtain sample image blocks, and the edges of the sample image blocks are filled with pixels using a symmetrical mapping method. The filled sample image blocks are divided according to a preset ratio to generate training and test datasets.

9. A deep learning-based super-resolution reconstruction device for the digital elevation model (DEM) of the U.S. coast, characterized in that, The device includes: The shallow feature extraction module is used for shallow feature extraction based on the super-resolution reconstruction model. It maps the preprocessed initial image to the feature space, extracts shallow features, and obtains the initial feature map; where the pixels of the initial image are elevation values. The deep feature enhancement module is used to perform long-range dependency modeling and multi-scale feature extraction on the initial feature map to obtain a fused feature map. The deep feature enhancement module includes a first long-range dependency modeling unit, a multi-scale feature extraction unit, a second long-range dependency modeling unit, and a feature fusion unit connected in series. The feature reconstruction module is used to upsample and reconstruct the fused feature map based on the feature reconstruction module to obtain the target image; wherein the resolution of the target image is higher than the resolution of the initial image.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 6.

Images