Mountain shadow guided DEM super-resolution reconstruction method and system

Abstract

The invention relates to the technical field of topographic mapping, in particular to a mountain shadow-guided DEM super-resolution reconstruction method and system, and the method comprises the steps: constructing a Sichuan data set which comprises DEM pairing data of different sources and resolutions and corresponding mountain shadow maps; designing an attention-based layered terrain fusion model, performing deep fusion on DEM elevation information and mountain shadow terrain features through an attention-based feature fusion module, and outputting a high-resolution DEM image according to the deep fusion features through a reconstruction module; and training an attention-based layered terrain fusion model by using the public data set, the Sichuan data set and a collaborative optimization-based loss module. According to the method, DEM elevation information and landform features of a mountain shadow are efficiently fused through the attention-based feature fusion module, so that detailed reconstruction of a complex landform is realized, and the precision of DEM super-resolution reconstruction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of terrain surveying and mapping, and in particular to a mountain shadow guided DEM super-resolution reconstruction method and system. Background Art

[0002] DEM plays an important role in hydrology, ecology, and geomorphology. Obtaining high-resolution DEM data is a challenging task. For DEM super-resolution, the traditional methods used in the past are based on interpolation, such as bilateral interpolation (Dunlop, 1980). These methods are usually simple in calculation and easy to use, but the high-resolution images they generate may have problems such as insufficient sharpness and distortion. Later, studies tried to introduce learning-based methods to solve this problem. The first method is based on CNN, such as the simple SRCNN (Chao et al., 2015) model, the deep residual network (Jiao et al., 2020), and the EDSR model (Xu et al., 2020), and its results are better than traditional interpolation methods. The network structure used for DEM super-resolution is gradually complex. In order to reduce the depth of the model, Zhang et al. proposed the RSPCN ​​model based on the idea of ​​recursion. In addition, Han et al. proposed a global information constrained DEM super-resolution reconstruction method (GISR) to solve the problem that the multi-layer aggregation process in deep learning is difficult to fully capture the underlying feature dependencies. The latest research (Yao et al., 2024) uses implicit neural representation models for DEM super-resolution. The second method is to use GAN methods to solve DEM super-resolution tasks, such as D-SRGAN (Demiray et al., 2021) and ESRGAN models (Wu et al., 2021). The third method is based on Transformer for large-scale DEM super-resolution tasks (Wang et al., 2024), which has been proven to achieve better reconstruction results.

[0003] In the above studies, LR DEMs were obtained from HR DEMs through a certain degradation process (e.g., bicubic downsampling), and paired DEM datasets in real scenes were not used. Wu et al. first discussed the problem of DEM super-resolution reconstruction in real scenes. In addition, the research usually focuses on the elevation features of DEM. Zhou et al. proposed the EDEM-SR model, which uses two parallel convolution kernels with different receptive fields to form a dual convolution kernel residual network, which can better extract and fuse the features of DEM data. However, it ignores the multi-scale spatial and terrain features of DEM. Recent studies have used deeper networks to transmit its terrain detail features, which have been proven to improve the accuracy of DEM reconstruction. Zhang et al. (2022) considered the unique spatial characteristics of DEM and proposed a terrain feature-aware super-resolution (TfaSR) combined with a deformable convolution network (DCN). Jiang et al. (2023) took into account the terrain parameters and constrained the learning and convergence of the network by constructing loss functions for slope and aspect. Zhou et al. (2023) proposed an MTF-SR method that combines terrain features, integrating raster terrain features in the input and loss functions, and integrating vector terrain features in the optimization of model output. These studies can utilize terrain features and have been proven to be effective in optimizing DEM super-resolution models. Summary of the invention

[0004] The present invention aims to solve the shortcomings of the prior art that the DEM space and terrain parameter information are not fully considered and the model training data lacks real scene pairing. A hill shadow guided DEM super-resolution reconstruction method and system are proposed. The DEM elevation information is efficiently fused with the terrain features of the hill shadow through an attention-based feature fusion module, thereby realizing detailed reconstruction of complex terrain and improving the accuracy of DEM super-resolution reconstruction.

[0005] To achieve the above purpose, the technical solution adopted is:

[0006] A hillshade-guided DEM super-resolution reconstruction method, comprising:

[0007] Construct the Sichuan dataset, which contains DEM paired data from different sources and resolutions and the corresponding hillshade map;

[0008] Design an attention-based hierarchical terrain fusion model, deeply fuse DEM elevation information with mountain shadow terrain features through an attention-based feature fusion module, and then output a high-resolution DEM image based on the deep fusion features through a reconstruction module;

[0009] The attention-based hierarchical terrain fusion model is trained using public datasets, Sichuan dataset, and a loss module based on collaborative optimization.

[0010] According to the mountain shadow guided DEM super-resolution reconstruction method of the present invention, further, when constructing the Sichuan data set, the calculation formula for generating the mountain shadow map is as follows:

[0011] Hillshade=254.0*((cos(Zenith_rad)*cos(Slope_rad))+(sin(Zenith_rad)*sin(Slope_rad)*cos(Azimuth_rad-Aspect_rad)))

[0012] Where Slope_rad, Aspect_rad, Zenith_rad and Azimuth_rad represent the slope, aspect, solar incidence angle and solar azimuth, respectively.

[0013] According to the mountain shadow guided DEM super-resolution reconstruction method of the present invention, further, the attention-based hierarchical terrain fusion model also includes a hierarchical terrain feature extraction module, which uses a hierarchical convolutional architecture to extract features of the elevation information of the DEM and the terrain features of the mountain shadow, specifically including:

[0014] The low-resolution DEM image and its corresponding mountain shadow map are upsampled and then enter their respective encoders for feature extraction. The encoder includes multiple 3×3 convolution and PReLU activation function submodules. The multi-level feature extraction expression is as follows:

[0015]

[0016] Where x∈{DEM,Shade}, is the convolution kernel of the i-th layer, which is used for DEM elevation feature extraction and mountain shadow terrain feature extraction respectively. is the bias term, i is the number of layers used for feature extraction, σ represents the PReLU activation function, represents the features extracted by the first layer of convolution, Represents the features extracted by the i-1th convolution layer.

[0017] According to the hill shadow guided DEM super-resolution reconstruction method of the present invention, further, the number of the attention-based feature fusion modules is equal to that of the encoder submodules, multiple feature fusion modules are connected in series, and receive output results of different levels of the encoder; it includes a feature enhancement submodule and a feature recalibration submodule:

[0018] Feature enhancement submodule: The encoder obtains the hierarchical feature maps of DEM and hill shadow, and concatenates the two feature maps of the same layer in the channel dimension as one input of the feature enhancement submodule. The other input of the feature enhancement submodule comes from the output of the feature fusion module based on attention at the previous level. Each input feature Figure 1 One branch outputs a feature map through Conv+Sigmoid, and the other branch outputs a feature map through Conv+ReLU. The output feature maps of these two parts are multiplied.

[0019] Feature recalibration submodule: The multiplication results of the feature enhancement submodule DEM and the mountain shadow are spliced ​​in the channel dimension, and after a convolution layer, global average pooling and global maximum pooling operations are performed respectively. The two feature maps obtained by the pooling operation are added, and the addition results are respectively passed through a Conv+Sigmoid to obtain two channel attention weights. The two channel attention weights are multiplied with the two results output by the feature enhancement submodule, and finally the two multiplication results are added and output to obtain the deep fusion feature.

[0020] According to the hill shadow guided DEM super-resolution reconstruction method of the present invention, further, for the attention-based feature fusion module corresponding to the last layer of the encoder, its input comes from the DEM feature map and the hill shadow feature map output by the last submodule of the encoder.

[0021] According to the mountain shadow guided DEM super-resolution reconstruction method of the present invention, further, the reconstruction module includes upsampling and a multi-layer variable convolution, which upsamples and reconstructs the deep fusion features output by the attention-based feature fusion module to output a high-resolution DEM image.

[0022] According to the mountain shadow guided DEM super-resolution reconstruction method of the present invention, further, the loss module based on collaborative optimization includes elevation loss L1, slope loss RMSE and terrain feature loss MSE, and the influence weight of each loss is adjusted through a dynamic weighting strategy. The total loss function expression is as follows:

[0023]

[0024] in, is the weighted parameter of self-learning, L τ (W) represents various losses.

[0025] Furthermore, the present invention also provides a mountain shadow guided DEM super-resolution reconstruction system, comprising a data set construction module, a model building module and a model training module, wherein:

[0026] The dataset construction module is used to construct the Sichuan dataset, which contains DEM paired data from different sources and resolutions and the corresponding hillshade map;

[0027] The model building module is used to design an attention-based hierarchical terrain fusion model. The DEM elevation information is deeply integrated with the mountain shadow terrain features through the attention-based feature fusion module, and then the reconstruction module outputs a high-resolution DEM image based on the deep fusion features;

[0028] Model training module, used to train the attention-based hierarchical terrain fusion model using public datasets, Sichuan dataset, and loss module based on collaborative optimization.

[0029] The beneficial effects achieved by adopting the above technical solution are:

[0030] 1. Terrain feature guidance method: The present invention proposes to use terrain feature images represented by mountain shadows to guide the DEM super-resolution reconstruction task, which significantly improves the ability to restore local terrain features of DEM and provides a new idea for multi-source data fusion for DEM reconstruction.

[0031] 2. Challenging dataset: A high-quality and challenging Sichuan dataset was constructed. This dataset not only contains DEM paired data from different sources and resolutions, but also combines the corresponding hillshade maps, covering areas with significant terrain undulations, providing a new experimental basis for research and practice in the field of DEM super-resolution.

[0032] 3. Multi-feature fusion module: An attention-based hierarchical terrain fusion framework (AHTF) is proposed. Through the attention-based feature fusion module (AFFM), the deep fusion of DEM elevation information and mountain shadow terrain features is realized, which significantly improves the performance of DEM super-resolution. Through the collaborative loss function of self-learning weight optimization, the refinement and efficient reconstruction of terrain features are further achieved.

[0033] 4. Excellent experimental performance: System experiments were conducted on three datasets of different landform types and resolutions. AHTF outperformed traditional Bicubic interpolation methods and the most advanced deep learning models (EDSR, TfaSR, and SRFormer) in quantitative analysis of terrain indicators such as elevation, slope, and aspect, as well as qualitative analysis of terrain elements such as rivers. The experimental results show that the proposed method has higher accuracy and stronger generalization ability.

[0034] The present invention not only provides an effective solution for DEM super-resolution reconstruction in complex terrain areas, but also opens up a new direction for research and application in the field of geographic information and remote sensing, and has important scientific significance and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings of the embodiments of the present invention, wherein the drawings are only used to illustrate some embodiments of the present invention, but not to limit all embodiments of the present invention thereto.

[0036] Figure 1 is a framework diagram of an attention-based hierarchical terrain fusion model (AHTF) according to an embodiment of the present invention;

[0037] Figure 2 It is a schematic diagram of a process in which a DEM represented by discrete data in an embodiment of the present invention is expressed as a hill shadow;

[0038] Figure 3 is a structural diagram of an attention-based feature fusion module (AFFM) according to an embodiment of the present invention;

[0039] Figure 4 is a graph of RMSE evaluation results of different methods of the embodiments of the present invention on the test set;

[0040] Figure 5 Detailed diagram of the training of the AHTF method and other super-resolution methods in the embodiment of the present invention on three data sets;

[0041] Figure 6 is a comparison diagram of error distribution of different DEM super-resolution models according to an embodiment of the present invention;

[0042] Figure 7 It is a DEM visualization result and error distribution diagram generated by different methods of the embodiments of the present invention. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings of specific embodiments of the present invention to clearly and completely describe the exemplary scheme of the embodiment of the present invention. Unless otherwise defined, the technical terms or scientific terms used in the present invention should be the common meanings understood by people with ordinary skills in the field.

[0044] This embodiment discloses a mountain shadow guided DEM super-resolution reconstruction method, which aims to improve the accuracy of converting low-resolution DEM to high-resolution DEM and make up for the shortcomings of existing methods in preserving local terrain features. The method specifically includes:

[0045] Step S101: Use two public DEM datasets (Austria and TFASR30) and a constructed real-scene dataset Sichuan dataset.

[0046] Public datasets: Austria and TFASR30 datasets are reconstructed from 30m resolution LR DEM to 10m resolution HR DEM. Since the LR DEMs of these public datasets are generated from HR DEM through a degradation process, the model mainly fits the inverse operation of the degradation process during training, which limits the generalization ability of the model in actual scenarios.

[0047] Sichuan dataset: Based on the analysis of DEM data from different sources and resolutions in real scenes, a new DEM super-resolution dataset, the Sichuan dataset, was constructed to be closer to real-world application scenarios. This dataset consists of DEM data from different sources in real scenes, covering DEM paired data with resolutions from 90m to 30m and corresponding mountain shadow maps. The study area has large undulating terrain, complex and diverse terrain, and is more challenging.

[0048] In the data preprocessing stage, the hillshade map is generated using the formula, which is expressed as follows:

[0049] Hillshade=254.0*((cos(Zenith_rad)*cos(Slope_rad))+(sin(Zenith_rad)*sin(Slope_rad)*cos(Azimuth_rad-Aspect_rad)))

[0050] In the formula, Slope_rad, Aspect_rad, Zenith_rad and Azimuth_rad represent slope, aspect, solar incidence angle and solar azimuth respectively. For the convenience of calculation, Zenith_rad and Azimuth_rad are set to 45 and 314 respectively. The process of expressing DEM represented by discrete data into hill shadow is as follows: Figure 2 shown.

[0051] In addition, since the value range of DEM is large, in order to make the model better learn and optimize, the input DEM and hillshade are normalized separately, and all data values ​​are scaled between [-1,1]. The normalization process is as follows:

[0052]

[0053] Among them, X∈{low-resolution DEM data, low-resolution hillshade data}, X max and X minRespectively represent the maximum and minimum values ​​of each data.

[0054] The Sichuan dataset covers areas in Sichuan with significant elevation differences and terrain undulations, and contains samples of various landform types, providing high-quality benchmark data for DEM super-resolution research. Hill shadow maps can intuitively reflect terrain features by simulating the distribution of light and shadow on the surface, thus making up for the shortcomings of single DEM data in capturing complex terrain details.

[0055] Step S102, design an attention-based hierarchical terrain fusion model, deeply fuse the DEM elevation information with the mountain shadow terrain features through the attention-based feature fusion module, and then output a high-resolution DEM image according to the deep fusion features through the reconstruction module. The attention-based hierarchical terrain fusion model makes full use of the DEM elevation information and the terrain feature information of the mountain shadow to achieve efficient fusion of multiple features and improve the effect of DEM super-resolution reconstruction. The attention-based hierarchical terrain fusion model includes a hierarchical terrain feature extraction module, an attention-based feature fusion module and a reconstruction module. The three modules are introduced as follows.

[0056] ① Hierarchical terrain feature extraction module, which extracts meaningful features from low-resolution DEM (LR DEM) and its corresponding mountain shadow map (shaded relief) images in a hierarchical manner.

[0057] The low-resolution DEM image and its corresponding mountain shadow map are upsampled and then enter their respective encoders for feature extraction. The encoder consists of multiple 3×3 convolution and Parametric Rectified Linear Unit (PReLU) activation function submodules. The number of submodules in this embodiment is 4. The PReLU activation function can improve the convergence speed of the model and effectively alleviate the overfitting problem. The multi-level feature extraction expression is as follows:

[0058]

[0059] Where x∈{DEM,Shade}, is the convolution kernel of the i-th layer, which is used for DEM elevation feature extraction and mountain shadow terrain feature extraction respectively. is the bias term, i is the number of layers used for feature extraction, σ represents the PReLU activation function, represents the features extracted by the first layer of convolution, Represents the features extracted by the i-1th convolution layer.

[0060] ② Attention-based feature fusion module

[0061] The number of attention-based feature fusion modules is equal to that of encoder submodules, which is 4. The four feature fusion modules are connected in series and receive the output results of different levels of the encoder. It includes feature enhancement submodule and feature recalibration submodule. The overall structure is as follows: Figure 3 shown.

[0062] Feature Enhancement Submodule: To enhance the feature representation capability, the gating unit is used as FEB to enhance the features of the intermediate feature maps of each encoder stream separately. The feature maps of the two feature maps are fused based on their interdependence between different channels.

[0063] The encoder obtains the hierarchical feature maps of DEM and hillshade. For the first to third layers of the encoder, the two feature maps of the same layer are spliced ​​in the channel dimension as one input of the feature enhancement submodule (the second, third or fourth). The other input of the feature enhancement submodule comes from the output of the feature fusion module based on attention at the previous level. Each input feature Figure 1 One branch outputs a feature map through Conv+Sigmoid, and the other branch outputs a feature map through Conv+ReLU. The output feature maps of these two parts are multiplied. For the fourth layer of the encoder, it corresponds to the first feature enhancement submodule, which inputs the DEM feature map and the hill shadow feature map output from the last submodule of the encoder.

[0064] Feature recalibration submodule: The multiplication results of the feature enhancement submodule DEM and mountain shadow are spliced ​​in the channel dimension, and after a convolution layer, global average pooling and global maximum pooling operations are performed respectively. The two feature maps obtained by the pooling operation are added, and the addition results are respectively passed through a Conv+Sigmoid to obtain two channel attention weights. The two channel attention weights are multiplied with the two results output by the feature enhancement submodule to establish a nonlinear relationship between the DEM and the mountain shadow intermediate feature map; finally, the two multiplication results are added and output to obtain the deep fusion feature.

[0065] To address the problem of similar feature redundancy caused by simple splicing or superposition of features, an attention-based feature fusion module is proposed. This module selects effective features through a feature enhancement sub-module and unifies different feature patterns with a feature recalibration sub-module, thereby achieving more accurate fusion of elevation and terrain information.

[0066] ③Reconstruction module

[0067] The reconstruction module includes upsampling and a multi-layer variable convolution. The reconstruction module performs upsampling and feature reconstruction based on the deep fusion features to output a high-resolution DEM image.

[0068] Step S103: train an attention-based hierarchical terrain fusion model using a public dataset, a Sichuan dataset, and a loss module based on collaborative optimization.

[0069] In order to ensure the refinement of terrain features in the process of DEM super-resolution reconstruction, a loss module based on collaborative optimization is proposed, including elevation loss, slope loss and terrain feature loss, and the influence weight of each loss is adjusted through a dynamic weighting strategy.

[0070] (1) Elevation loss (L1): It is used to measure the pixel error between the generated super-resolution DEM and the true high-resolution DEM. Compared with the mean square error (MSE), L1 loss is more robust to outliers and can prevent the generated results from being too smooth.

[0071]

[0072] Where m is the number of samples, DEM real It is a real high-resolution DEM, DEM fake It is to generate high-resolution DEM.

[0073] (2) Slope loss (RMSE): The root mean square error of the DEM slope is calculated and used to optimize local terrain features.

[0074]

[0075] Where N is the number of samples, O i represents the slope derived from the real high-resolution DEM, S i Represents the slope derived from generating a high-resolution DEM.

[0076] (3) Terrain feature loss: The visual performance difference of DEM is measured by calculating the MSE of hillshade.

[0077]

[0078] Where n is the number of samples, Indicates the generation of high-resolution DEM-derived hillshade, y i Represents a hillshade derived from a true high-resolution DEM.

[0079] (4) Collaborative loss: Optimize the numerical stability of model training by dynamically adjusting the weights of each loss based on uncertainty in the same variance.

[0080]

[0081] in, is the weighted parameter of self-learning, L τ (W) represents various losses.

[0082] The present invention can effectively restore the local details of the high-resolution DEM and can also make the visual effect closer to the real terrain.

[0083] Corresponding to the above method, this embodiment also discloses a mountain shadow guided DEM super-resolution reconstruction system, comprising:

[0084] The dataset construction module is used to construct the Sichuan dataset, which contains DEM paired data from different sources and resolutions and the corresponding hillshade map;

[0085] The model building module is used to design an attention-based hierarchical terrain fusion model. The DEM elevation information is deeply integrated with the mountain shadow terrain features through the attention-based feature fusion module, and then the reconstruction module outputs a high-resolution DEM image based on the deep fusion features;

[0086] Model training module, used to train the attention-based hierarchical terrain fusion model using public datasets, Sichuan dataset, and loss module based on collaborative optimization.

[0087] In order to verify the effectiveness of this scheme, further explanation is given below in combination with experimental data.

[0088] (1) Experimental setup

[0089] The following four methods were selected as comparison models:

[0090] ①Interpolation method: BiCubic.

[0091] ② Deep learning methods: EDSR model based on residual blocks, TfaSR model combined with terrain features, and SRFormer model based on Transformer.

[0092] All deep learning methods are implemented based on the PyTorch framework and trained on an NVIDIA Quadro P6000 24GB graphics card. During the training process, to ensure fair comparison, all deep learning models use the same hyperparameters (except batch size), including: Optimizer: Adam. Learning rate: 0.0001. Training phase: 200 epochs. Loss function: Collaborative Loss proposed in this paper

[0093] The batch size is adjusted according to the input image size: 10 for 100×100 input, 128 for 21×21 input, and 12 for 85×85 input. In addition, to reduce computational complexity, SRFormer uses a lightweight version (SRFormer-light), which can ensure that the model accuracy remains basically unchanged while significantly reducing the amount of computation.

[0094] ③Evaluation indicators:

[0095] The experiment uses the following three quantitative terrain indicators to evaluate the DEM reconstruction performance:

[0096] RMSE elevation: measures the elevation error of DEM reconstruction. The lower the RMSE value, the better the model performance.

[0097] RMSE slope: measures the error of the reconstruction result on the slope.

[0098] RMSE Aspect: Measures the error of the reconstruction result in aspect.

[0099] (2) Quantitative comparison results

[0100] All trained models are applied to three test data respectively, and the comparison results are listed in Figure 4 . From the statistical results, it can be seen that the proposed AHTF outperforms all other methods in terms of the three terrain evaluation indicators, not only the Austria and TfaSR30 datasets for the task of restoring 30m to 10m, but also the Sichuan dataset for the task of restoring 90m resolution to 30m. However, it should be pointed out that the five models do not perform as well in the Sichuan dataset as in the other two datasets. This is because the LR and HR DEMs of the Sichuan dataset are derived from two different real measurement data, and the experimental area has large ups and downs, which adds challenges to the DEM super-resolution task.

[0101] In addition, it can be observed that compared with the traditional interpolation method, namely BiCubic, the deep learning method generally has more satisfactory performance and greater stability, which highlights the superiority of deep learning methods for DEM super-resolution tasks. SRFormer, as the latest transformer-based super-resolution model, can achieve better results than EDSR for DEM super-resolution tasks. It is worth mentioning that from Figure 5(b) It can be seen that the training curve of SRFormer on the Austria dataset is slightly lower than that of TfaSR, but it performs worse than TfaSR on the test set. This may be because SRFormer is a super-resolution model based on Transformer, and it overfits in the DEM super-resolution task due to the small amount of data. EDSR and SRFormer do not take into account the influence of terrain features, and the statistical results show that TfaSR performs better, second only to the proposed AHTF model. This shows that TfaSR combined with Terrain feature-aware can play a positive role in the DEM super-resolution task. It not only integrates the deep residual module and the DCN module to extract depth and adaptive terrain features, but also improves the perception of terrain features through its loss function. On this basis, AHTF guides the DEM super-resolution network with mountain shadows to achieve multi-feature fusion of terrain and elevation. In addition, AHTF optimizes by adjusting the loss function of TfaSR and adding terrain visual loss. Therefore, AHTF provides a more accurate and reliable solution for DEM super-resolution.

[0102] (3) Error analysis

[0103] To better understand the results, the performance of the five models was quantitatively analyzed by presenting overall error box plots. Figure 6 The overall error distribution of each model on the three datasets is shown. By comparing the upper and lower boundaries of the box of each model, the discreteness of the model results can be evaluated. It is observed that the boxes of the AHTF, SRFormer, and TfaSR models are narrower, indicating that their prediction results are relatively stable and the models are more robust. The boxes of the EDSR model and BiCubic are wider, which means that the data is more discrete and the uncertainty of the prediction results is higher. In addition, it can be observed that the box width span on the Sichuan test set is larger, indicating that the Sichuan dataset will pose a greater challenge to the DEM super-resolution task due to its own data source and terrain characteristics. Finally, by calculating the average value of each RMSE, the prediction performance curve of each model is obtained. Considering the above analysis results, the proposed method can recover most of the DEM with high accuracy.

[0104] (4) Visual evaluation results

[0105] For each of the three datasets, an example DEM was selected to demonstrate the visual assessment of terrain features. Figure 7(a)(c)(e) show the high-resolution DEMs generated by different methods. At first glance, all methods can restore the low-resolution DEM to a higher-resolution version. However, there are still many differences in the details of each generated DEM. In order to intuitively display the difference between the original high-resolution DEM and the super-resolution result, the original HR DEM is subtracted from the DEM reconstructed by different methods to obtain the elevation error map, see sub-figures (b)(d)(f), and the distribution of the difference results is represented by a line graph. It can be seen that compared with other models, the error of the AHTF proposed in the present invention is more concentrated near 0, and is less distributed in the interval with larger errors, which is consistent with the Figure 4 The evaluation index results of RMSE-elevation are consistent. On the contrary, the elevation error map of BiCubic shows that the blue and red areas are large, and its error distribution is in a larger range of values. In addition, the error map obtained by the SRFormer model in the Austria dataset test is redder, and the error map obtained by the TfaSR model in the Sichuan dataset test is bluer, which is reflected in the right or left deviation of the curve on the error distribution map, indicating that the elevation features learned by the model are biased. In addition, it can be observed from the error map that most errors occur in areas with large terrain fluctuations such as rivers and valleys, and the error map is shown as the redder or bluer the color. However, the AHTF is lighter in these areas, which can show that the designed AHTF can more accurately maintain the terrain trend.

[0106] Unless otherwise specifically stated, the components, steps, numerical expressions and values ​​set forth in these embodiments do not limit the scope of the present invention.

[0107] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0108] The units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation is not considered to be beyond the scope of the present invention.

[0109] Those skilled in the art will appreciate that all or part of the steps in the above method can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk or an optical disk. Optionally, all or part of the steps in the above embodiment can also be implemented using one or more integrated circuits, and accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or in the form of software function modules. The present invention is not limited to any specific form of combination of hardware and software.

[0110] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A hill shadow guided DEM super-resolution reconstruction method, characterized in that: include: Construct the Sichuan dataset, which contains DEM paired data from different sources and resolutions and the corresponding hillshade map; Design an attention-based hierarchical terrain fusion model, deeply fuse DEM elevation information with mountain shadow terrain features through an attention-based feature fusion module, and then output a high-resolution DEM image based on the deep fusion features through a reconstruction module; The attention-based hierarchical terrain fusion model is trained using public datasets, Sichuan dataset, and a loss module based on collaborative optimization.

2. The method for DEM super-resolution reconstruction guided by mountain shadow according to claim 1, characterized in that: When constructing the Sichuan dataset, the calculation formula for generating the hillshade map is as follows: Hillshade=254.0*((cos(Zenith_rad)*cos(Slope_rad))+(sin(Zenith_rad)*sin(Slope_rad)*cos(Azimuth_rad-Aspect_rad))) In the formula, Sl ope_ra d 、 A spect_ra d 、 Zenith_rad and A z i mut h _ra d They represent slope, aspect, solar incidence angle and solar azimuth respectively.

3. The method for DEM super-resolution reconstruction guided by mountain shadow according to claim 1, characterized in that: The attention-based hierarchical terrain fusion model also includes a hierarchical terrain feature extraction module, which uses a hierarchical convolutional architecture to extract the elevation information of the DEM and the terrain features of the mountain shadow, specifically including: The low-resolution DEM image and its corresponding mountain shadow map are upsampled and then enter their respective encoders for feature extraction. The encoder includes multiple 3×3 convolution and PReLU activation function submodules. The multi-level feature extraction expression is as follows: Where x∈{DEM,Shade}, is the convolution kernel of the i-th layer, which is used for DEM elevation feature extraction and mountain shadow terrain feature extraction respectively. is the bias term, i is the number of layers used for feature extraction, σ represents the PReLU activation function, represents the features extracted by the first layer of convolution, Represents the features extracted by the i-1th convolution layer.

4. The method for DEM super-resolution reconstruction guided by mountain shadow according to claim 3 is characterized in that: The number of the attention-based feature fusion modules is equal to that of the encoder submodules, and multiple feature fusion modules are connected in series and receive output results of different levels of the encoder; That Including feature enhancement submodule and feature recalibration submodule: Feature enhancement submodule: The encoder is used to obtain the hierarchical feature maps of DEM and hill shadow, and the two feature maps of the same layer are spliced ​​in the channel dimension as one input of the feature enhancement submodule. The other input of the feature enhancement submodule comes from the output of the attention-based feature fusion module at the previous level. One branch of each input feature map outputs a feature map through Conv+Sigmoid, and the other branch outputs a feature map through Conv+ReLU. The output feature maps of these two parts are multiplied. Feature recalibration submodule: The multiplication results of the feature enhancement submodule DEM and the mountain shadow are spliced ​​in the channel dimension, and after a convolution layer, global average pooling and global maximum pooling operations are performed respectively. The two feature maps obtained by the pooling operation are added, and the addition results are respectively passed through a Conv+Sigmoid to obtain two channel attention weights. The two channel attention weights are multiplied with the two results output by the feature enhancement submodule, and finally the two multiplication results are added and output to obtain the deep fusion feature.

5. The method for DEM super-resolution reconstruction guided by mountain shadow according to claim 4, characterized in that: For the attention-based feature fusion module corresponding to the last layer of the encoder, its input comes from the DEM feature map and hill shadow feature map output by the last sub-module of the encoder.

6. The method for DEM super-resolution reconstruction guided by mountain shadow according to claim 1, characterized in that: The reconstruction module includes upsampling and a multi-layer variable convolution, which upsamples and reconstructs the deep fusion features output by the attention-based feature fusion module to output a high-resolution DEM image.

7. The method for DEM super-resolution reconstruction guided by mountain shadow according to claim 1, characterized in that: The loss module based on collaborative optimization includes elevation loss L1, slope loss RMSE and terrain feature loss MSE, and the influence weight of each loss is adjusted through a dynamic weighting strategy. The total loss function expression is as follows: in, is the weighted parameter of self-learning, L τ (W) represents various losses.

8. A mountain shadow guided DEM super-resolution reconstruction system, characterized in that: It includes dataset construction module, model building module and model training module, among which: The dataset construction module is used to construct the Sichuan dataset, which contains DEM paired data from different sources and resolutions and the corresponding hillshade map; The model building module is used to design an attention-based hierarchical terrain fusion model. The DEM elevation information is deeply integrated with the mountain shadow terrain features through the attention-based feature fusion module, and then the reconstruction module outputs a high-resolution DEM image based on the deep fusion features; Model training module, used to train the attention-based hierarchical terrain fusion model using public datasets, Sichuan dataset, and loss module based on collaborative optimization.

9. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images