Cross-regional remote sensing image elevation reconstruction method based on domain adaptation
By constructing a domain-adaptive method for cross-regional remote sensing image elevation reconstruction, and utilizing the U-net generator and multi-scale discrimination module, combined with Hann window fusion technology, the problem of scarce measured data and "domain offset" in cross-regional remote sensing image elevation reconstruction is solved, achieving efficient and accurate elevation reconstruction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-13
- Publication Date
- 2026-04-10
AI Technical Summary
In cross-regional remote sensing image elevation reconstruction, there are problems such as scarcity of measured data and "domain offset", which leads to insufficient elevation prediction accuracy. Existing transfer learning methods are difficult to effectively utilize source domain knowledge for target domain elevation reconstruction.
A domain-adaptive approach is adopted, which constructs a cross-regional elevation migration prediction model based on U-net generator, multi-scale discriminant module and domain classifier. Combined with Hann window fusion technology, the model is fine-tuned using a small amount of measured data from the target domain to alleviate "domain offset" and improve accuracy.
It achieves efficient mitigation of accuracy loss without requiring a large amount of measured data in the target domain, improving the efficiency and accuracy of cross-regional elevation reconstruction, and is suitable for engineering applications.
Smart Images
Figure CN121837530A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of remote sensing data processing and relates to remote sensing image elevation reconstruction technology, specifically to a cross-regional remote sensing image elevation reconstruction method based on domain adaptation. Background Technology
[0002] Remote sensing technology, with its advantages of wide coverage, short data acquisition cycle, and low observation cost, has become one of the core means of acquiring coastal zone elevation data. Among them, satellite multispectral data, due to its high spatial resolution, can effectively reflect surface texture and radiation characteristics, providing rich basic information for elevation inversion; while satellite laser altimetry data and UAV field measurement data can provide high-precision local elevation references. The combination of the two has become an important technical approach for coastal zone elevation modeling.
[0003] Elevation data serves as a crucial foundation for coastal zone planning, disaster prevention and mitigation, and engineering construction. The efficiency and coverage of its acquisition directly impact the progress of related work. Traditional elevation modeling methods rely on a large amount of measured elevation data for the target area (such as densely deployed GNSS survey points and full-coverage aerial surveys by UAVs) to ensure model accuracy. However, if measured data is scarce in the target area (such as remote areas or complex coastlines), it can lead to difficulties in constructing elevation models or insufficient accuracy.
[0004] Transfer learning techniques can be used to transfer elevation prediction models trained in a source domain (areas with abundant measured data) to a target domain (areas with scarce measured data), reducing reliance on measured data in the target domain. However, remote sensing images from different regions (source and target domains) exhibit radiometric differences (such as atmospheric conditions and illumination angles) and topographic features (such as shoreline morphology and land cover type). Directly transferring models can easily lead to "domain shift" problems, resulting in decreased elevation prediction accuracy. Furthermore, existing transfer learning methods often fail to optimize for the spatial correlation between remote sensing images and elevation data, making it difficult to efficiently utilize source domain knowledge to reconstruct elevations in the target domain. Therefore, there is an urgent need to design a remote sensing image elevation reconstruction method that is adaptable to cross-regional scenarios and has low data dependency. Summary of the Invention
[0005] Purpose of the invention: In order to overcome the shortcomings of scarce measured elevation data of the target domain in cross-regional scenarios and the "domain offset" defect of direct transfer models, this invention provides a cross-regional remote sensing image elevation reconstruction method based on domain adaptation. This method does not require a large amount of measured data of the target domain and can effectively alleviate the accuracy loss caused by "domain offset".
[0006] Technical Solution: To achieve the above objectives, this invention provides a method for cross-regional remote sensing image elevation reconstruction based on domain adaptation, comprising the following steps:
[0007] S1: Perform data preprocessing on the acquired raw remote sensing image data and elevation data;
[0008] S2: Perform pixel-level alignment on the preprocessed remote sensing image data and elevation data, and resample and slice the data to obtain slice data;
[0009] S3: Construct a cross-regional elevation migration prediction model based on domain adaptation;
[0010] S4: Use slice data to train the cross-regional elevation migration prediction model, obtain the optimal model weights, and output the digital elevation model value of each slice through the trained cross-regional elevation migration prediction model.
[0011] S5: Based on the Hann window, the digital elevation model values of the slices are fused to reconstruct the digital elevation model values of the entire region.
[0012] Furthermore, the data preprocessing in step S1 includes geometric correction, radiometric calibration, and atmospheric correction.
[0013] Further, step S2 includes:
[0014] A1: Based on cubic spline interpolation, remote sensing image data is resampled. The formula for cubic spline interpolation is as follows:
[0015]
[0016] in, Let represent the coordinates of a pixel in the image, where i represents the grid index closest to y, and j represents the grid index closest to x. Represents the entire surrounding pixels; This represents the horizontal offset ratio of the target pixel x relative to the left and right boundaries of the y-th column of the original image. Indicates the vertical offset ratio;
[0017] A2: Resampling of elevation images based on the average interpolation method. The formula for the average interpolation method is as follows:
[0018]
[0019] in, Indicating the original image pixel value of a point This represents the pixel value after resampling using the average interpolation method;
[0020] A3: For resampled band and elevation images, the following formula is used to define the region:
[0021]
[0022] Where img and dem represent remote sensing image and elevation image, respectively; , , , These represent the left, right, upper, and lower boundaries of the overlapping area of the two images, respectively.
[0023] A4: Perform data tiling on the remote sensing imagery and the digital elevation model imagery respectively.
[0024] Furthermore, the cross-regional elevation migration prediction model in step S3 includes a generator based on U-net, a multi-scale discriminant module, and a domain classifier. The multi-scale discriminant module includes two branches: a global discriminant and a local discriminant. The global discriminant is responsible for judging the overall elevation distribution and global consistency, while the local discriminant performs local area authenticity judgment at the downsampled spatial resolution.
[0025] Furthermore, the construction of the U-net-based generator in step S3 includes:
[0026] Add channel attention and frequency attention mechanisms to each layer of the encoder;
[0027] In the channel attention mechanism, given a feature map F, the formula for channel attention is as follows:
[0028]
[0029] in, This represents the feature value located in the i-th row and j-th column of the c-th channel of feature map F;
[0030] H and W represent the height and width of the input feature map F, respectively. W1 represents the channel attention weights calculated by the c-th channel attention mechanism, and W2 represents the first weight and the second weight, respectively.
[0031] In the sixth coding layer of U-net, a frequency attention module is introduced to enhance the perception of texture frequency and spatial structural details; the operation of the frequency attention module includes:
[0032] Targeting bottleneck characteristics Where B represents the batch size, C represents the number of channels, and H and W represent the height and width of the feature map, respectively. After performing a two-dimensional Fourier transform on the feature map and separating the real and imaginary parts, the features are stacked to form a 2C-channel frequency feature map. , and These represent the operations of extracting the real and imaginary parts, respectively, and then the weight relationship between the frequency channels is learned through two pointwise convolutions. Then, the frequency components are weighted and returned to the spatial domain. ,in, This represents the Hadamard product, which is performed element-wise.
[0033] Introduce a cross-scale channel attention mechanism in the decoder:
[0034] The decoder achieves high- and low-level feature fusion through cross-scale channel attention, using the difference between the generated simulated digital elevation model and the real digital elevation model as the core loss. It integrates pixel-level L1 loss, gradient loss, SSIM loss, and adversarial loss to increase generator accuracy. The formulas are as follows:
[0035]
[0036] Where Y represents the true digital elevation model, This represents the elevation model of the predicted data output by the generator network. Represents the mathematical expectation. Describing the L1 norm, and Let represent the gradient operators of the image along the horizontal x-axis and the vertical y-axis, respectively. , , These represent pixel-level loss, gradient loss, and structural similarity loss, respectively.
[0037] Furthermore, the construction of the multi-scale discrimination module in step S3 includes:
[0038] The discrimination result output by the global discriminator is: ;in, This represents the input image for the discriminator. This represents the convolution operation of the global discriminator;
[0039] The discrimination result output by the local discriminator is: ,in, This represents the features of the input image after downsampling.
[0040] Furthermore, in step S3, a spectral normalization constraint weight distribution is introduced into the discriminator to constrain the weights of each convolutional layer. Therefore, the joint loss of the multi-scale discriminator output is:
[0041]
[0042] in, , , These represent the input remote sensing imagery, actual digital elevation imagery, and predicted elevation imagery, respectively. and These represent the scores given by the global and local discriminators to the input sample pairs, respectively, which are used to measure the authenticity of the input samples. This is a gradient penalty term; This represents the gradient penalty coefficient, used to adjust the proportion of the gradient penalty term in the total loss.
[0043] Furthermore, the training of the cross-regional elevation migration prediction model in step S4 includes:
[0044] An improved Wasserstein GAN framework is used for adversarial training: during training, the generator and discriminator alternately update their parameters, forming a stable adversarial relationship; during the discriminator update phase, the generator parameters are fixed, and the adversarial relationship is minimized. ,renew During the generator update phase, the discriminator parameters are fixed, and the minimum value is minimized. ,renew The final loss function objective is
[0045]
[0046] in, , These represent the distributions of real samples and generated samples, respectively. , , , These represent gradient penalty loss, pixel-level L1 loss, gradient loss, and structural similarity loss, respectively. , , , These represent the weight balancing coefficients corresponding to each loss term. By adjusting these coefficients, the optimization weights of the model in terms of adversarial realism, pixel accuracy, edge texture preservation, and structural consistency can be balanced.
[0047] A domain classifier and gradient inversion layer are introduced to implement a domain confusion training mechanism: During generator training, the domain classification loss is calculated based on the output of the gradient inversion layer to drive the generator to learn domain-invariant features; at the same time, the maximum mean difference (MMD) loss is introduced to align the feature distributions of the source domain and the target domain.
[0048] Further, step S5 includes:
[0049] B1: Standardize the digital elevation model values of the slices;
[0050] B2: The slices are stitched together based on their spatial coordinates in the complete image. Within the overlapping area of the slices, a two-dimensional Hann window function is used to perform weighted fusion of the overlapping pixels.
[0051] Furthermore, the expression of the two-dimensional Hann window function in step B2 is as follows:
[0052]
[0053] in, As weight, and These represent the dimensions of the slice on the x-axis and y-axis, respectively.
[0054] Beneficial Effects: Compared with existing technologies, this invention pre-trains a generator model using complete source multispectral remote sensing data, combines domain adaptation techniques to align the feature distributions of the source and target domains, and only uses a small portion of measured elevation data from the target domain for model fine-tuning, ultimately achieving full-area elevation reconstruction of the target domain. This method does not require a large amount of measured data from the target domain, effectively mitigates the accuracy loss caused by "domain offset," and is easy to program and apply in engineering, solving the problems of low efficiency and poor accuracy in cross-regional elevation reconstruction under scarce data scenarios. Attached Figure Description
[0055] Figure 1 This is a flowchart illustrating the method of the present invention;
[0056] Figure 2 This is a schematic diagram of the Doulonggang digital elevation model;
[0057] Figure 3 This is a schematic diagram of the digital elevation model of Dafeng Port;
[0058] Figure 4 This is a schematic diagram of the generator structure in this invention;
[0059] Figure 5 This is a schematic diagram of the discriminator in this invention;
[0060] Figure 6 This is a schematic diagram of the domain classifier in this invention;
[0061] Figure 7 This is a diagram showing the elevation inversion results obtained in this invention. Detailed Implementation
[0062] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.
[0063] This embodiment provides a domain-adaptive method for cross-regional remote sensing image elevation reconstruction. This embodiment selects, for example... Figure 2 The remote sensing image and digital elevation model (DEM) image of Doulonggang, Jiangsu coastal zone, in November 2024, shown below, are used as source domain data. The data selected is as follows: Figure 3The remote sensing imagery and digital elevation model (DEM) imagery of Dafeng Port in November 2024, shown below, serve as the target domain data. Python is used as the tool for data processing, model building, and result visualization. Figure 1 As shown, the reconstruction method in this embodiment specifically includes the following steps:
[0064] S1: Perform data preprocessing on the acquired raw remote sensing image data and elevation data;
[0065] Data preprocessing includes geometric correction, radiometric calibration, and atmospheric correction;
[0066] S2: In order to ensure the consistency of source domain data and target domain data in terms of spatial range and resolution, pixel-level alignment is performed on the preprocessed remote sensing image data and elevation data, and the data is resampled and sliced to obtain slice data.
[0067] Step S2 specifically includes:
[0068] 1) Multispectral remote sensing images of the Jiangsu coastal zone (including four bands: B02, B03, B04, and B08) were acquired using satellites, with an original resolution of 10 m × 10 m. To maintain consistency with elevation data and meet the requirements for high-precision reconstruction, cubic spline interpolation was used to upsample the images, resampling the spatial resolution to 2 m × 2 m. The resulting resampled remote sensing image is stacked_2m.tif. The cubic spline interpolation formula is as follows:
[0069]
[0070] in, Let represent the coordinates of a pixel in the image, where i represents the grid index closest to y, and j represents the grid index closest to x. Represents the entire surrounding pixels; This represents the horizontal offset ratio of the target pixel x relative to the left and right boundaries of the y-th column of the original image. Indicates the vertical offset ratio;
[0071] 2) Elevation data (digital elevation model) collected by UAVs along the Jiangsu coastline was used, with an original resolution of 0.5m x 0.5m. To match the satellite imagery, downsampling was performed using average interpolation to unify the resolution to 2m x 2m, resulting in the resampled digital elevation model image digitized_elevation_model_2m.tif. The average interpolation formula is as follows:
[0072]
[0073] in, Indicating the original image pixel value of a point This represents the pixel value after resampling using the average interpolation method;
[0074] 3) Using the geometry interface provided by the Python rasterio library, read the spatial extent of remote sensing images and digital elevation models respectively, calculate their overlapping areas and extract the common extent of the two; generate the same spatial window through from_bounds() to ensure that the two types of data are completely corresponding in geographic coordinate system and spatial extent;
[0075] For resampled band and elevation images, the region is defined using the following formula:
[0076]
[0077] Where img and dem represent remote sensing image and elevation image, respectively; , , , These represent the left, right, upper, and lower boundaries of the overlapping area of the two images, respectively.
[0078] 4) After completing resolution and region alignment, the image data is divided into patches for subsequent input into the model for training and inference. Each patch is checked for invalid values; if any are found, they are skipped. Valid patches are saved in numerical order to the following folders: image_patches, digital elevation model_dem_patches.
[0079] S3: Construct a cross-regional elevation migration prediction model based on domain adaptation;
[0080] The cross-regional elevation migration prediction model includes a U-net-based generator, a multi-scale discriminant module, and a domain classifier;
[0081] like Figure 4 As shown, the generator construction includes:
[0082] Add channel attention and frequency attention mechanisms to each layer of the encoder;
[0083] In the channel attention mechanism, given a feature map F, channel attention can enhance discriminative channel features and suppress redundant information; the formula is as follows:
[0084]
[0085] in, This represents the feature value located in the i-th row and j-th column of the c-th channel of feature map F;
[0086] H and W represent the height and width of the input feature map F, respectively. W1 represents the channel attention weights calculated by the c-th channel attention mechanism, and W2 represents the first weight and the second weight, respectively.
[0087] In the sixth coding layer of U-net, a frequency attention module is introduced to enhance the perception of texture frequency and spatial structural details; the operation of the frequency attention module includes:
[0088] Targeting bottleneck characteristics Where B represents the batch size, C represents the number of channels, and H and W represent the height and width of the feature map, respectively. After performing a two-dimensional Fourier transform on the feature map and separating the real and imaginary parts, the features are stacked to form a 2C-channel frequency feature map. , and These represent the operations of extracting the real and imaginary parts, respectively, and then the weight relationship between the frequency channels is learned through two pointwise convolutions. Then, the frequency components are weighted and returned to the spatial domain. ,in This represents the Hadamard product, which is performed element-wise.
[0089] Introduce a cross-scale channel attention mechanism in the decoder:
[0090] The decoder achieves high- and low-level feature fusion through cross-scale channel attention, improving the spatial detail fidelity of the reconstructed digital elevation model. Using the difference between the generated simulated digital elevation model and the real digital elevation model as the core loss, it integrates pixel-level L1 loss, gradient loss, SSIM loss, and adversarial loss to increase generator accuracy. The formulas are as follows:
[0091]
[0092] Where Y represents the true digital elevation model, This represents the elevation model of the predicted data output by the generator network. Represents the mathematical expectation. Describing the L1 norm, and Let represent the gradient operators of the image along the horizontal x-axis and the vertical y-axis, respectively. , , These represent pixel-level loss, gradient loss, and structural similarity loss, respectively.
[0093] like Figure 5As shown, the multi-scale discrimination module comprises two branches: a global discriminator and a local discriminator. The global discriminator is responsible for judging the overall elevation distribution and global consistency, while the local discriminator performs local region authenticity discrimination at the downsampled spatial resolution. Both discriminators consist of multiple stacked spectral normalized convolutional layers (SN_Conv2d) and LeakyReLU activation functions, but they differ significantly in network depth, channel configuration, and convolutional kernel size to adapt to different discrimination tasks.
[0094] The global discriminator aims to distinguish between genuine and fake data in the overall distribution of the elevation model. Its input is the original stitched image (4 channels). This branch employs... The convolutional kernels are optimized to achieve a larger receptive field, specifically comprising five convolutional layers: Layer 1: Maps the input features from 4 channels to 32 channels, then connects to the LeakyReLU activation function; Layer 2: Increases the number of feature channels from 32 to 64, extracting low-level spatial features; Layers 3 through 4: The number of channels is doubled layer by layer. As the channel dimension increases, high-level semantic features are extracted and dimensionality-reduced encoding is performed.
[0095] The fifth layer (output layer) maps the 256-channel features to a single-channel output, generating a global discrimination scoring matrix. In this process, each convolutional layer is followed by a LeakyReLU activation function (with leakage correction linear units) to avoid the gradient vanishing problem during training and enhance the model's non-linear expressive power.
[0096] The local discriminator aims to refine the discrimination of downsampled local regions; the input feature map has undergone average pooling. This branch employs... The convolutional kernel focuses on the channel relationships between pixels and local texture details, specifically comprising five convolutional operations: The first layer receives 4-channel input features and compresses them to 2 channels to reduce computational redundancy and focus on key high-frequency information; the second to fourth layers gradually recover and expand the feature dimension through successive convolutions, with the number of channels changing along a specific path. This process gradually constructs a feature representation of the local texture. The fifth layer (output layer) aggregates the 32-channel features into a single-channel output to generate a local discriminant score. Similar to the global discriminator, each convolutional operation (SN_Conv2d) in the local discriminator is followed by a LeakyReLU activation function, ensuring gradient sensitivity when dealing with subtle elevation differences.
[0097] The discrimination result output by the global discriminator is: ;in, This represents the input image for the discriminator. This represents the convolution operation of the global discriminator;
[0098] The discrimination result output by the local discriminator is: ,in, This represents the features of the downsampled input image;
[0099] To ensure training stability, spectral normalization is introduced into the discriminator to constrain the weight distribution, thus restricting the weights of each convolutional layer. Therefore, the joint loss of the multi-scale discriminator output is:
[0100]
[0101] in, , , These represent the input remote sensing imagery, actual digital elevation imagery, and predicted elevation imagery, respectively. and These represent the scores given by the global and local discriminators to the input sample pairs, respectively, which are used to measure the authenticity of the input samples. This is a gradient penalty term; This represents the gradient penalty coefficient, used to adjust the proportion of the gradient penalty term in the total loss.
[0102] S4: Use slice data to train the cross-regional elevation migration prediction model, obtain the optimal model weights, and output the digital elevation model value of each slice through the trained cross-regional elevation migration prediction model.
[0103] Step S4 specifically includes:
[0104] 1) Based on the spatial location of the center of the remote sensing image and digital elevation model data slice, a grid-based spatial segmentation method is adopted to divide the slice samples into training set, validation set and test set in a ratio of 7:2:1 to ensure the spatial independence of sample distribution; the training set is used for model training, the validation set is used for hyperparameter adjustment and early stopping detection, and the test set is used for final model performance evaluation.
[0105] 2) An improved Wasserstein GAN framework is used for adversarial training: During training, the generator and discriminator alternately update their parameters, forming a stable adversarial relationship; during the discriminator update phase, the generator parameters are fixed, and the minimum value is minimized. ,renew During the generator update phase, the discriminator parameters are fixed, and the minimum value is minimized. ,renew The final loss function objective is
[0106]
[0107] in, , These represent the distributions of real samples and generated samples, respectively. , , , These represent gradient penalty loss, pixel-level L1 loss, gradient loss, and structural similarity loss, respectively. , , , These represent the weight balancing coefficients corresponding to each loss term. By adjusting these coefficients, the optimization weights of the model in terms of adversarial realism, pixel accuracy, edge texture preservation, and structural consistency can be balanced.
[0108] Through the above-mentioned multiple loss constraints, the generator not only approximates the real elevation at the numerical level, but also maintains consistency in spatial structure, texture details and terrain undulation, thus achieving the generation of high-quality digital elevation models.
[0109] 3) After each training round, the model performance is evaluated using test-time data augmentation on the validation set: horizontal flipping, vertical flipping, and 180° rotation transformations are performed on the remote sensing images in the validation set to generate four augmented samples, which are then input into the generator. The average of the four prediction results is taken as the final predicted digital elevation model. The mean absolute error (MAE) between the prediction results and the real digital elevation model is calculated. When the validation error does not decrease significantly for several consecutive rounds, an early stopping mechanism is triggered, and the current best model is saved and recorded as best_pretrain_model.pth.
[0110] 4) Extract the digital elevation model data of the target domain (DF Port) and randomly sample it according to the preset sampling ratio (0.1~1.0). Combine the target domain samples with the source domain samples to form a mixed training dataset. By marking the domain attributes in the sample labels (source domain=0, target domain=1), supervised mixed training of cross-domain data is achieved.
[0111] 5) After loading the weights of the pre-trained model in the source domain, calculate the feature cosine similarity between the source and target domains in each encoding layer. The weights of the encoding layers with a similarity higher than the threshold (0.6) are frozen, and only the remaining layers are unfrozen to participate in fine-tuning, thereby achieving feature transfer and efficient adaptive parameter update;
[0112] 6) For example Figure 6 As shown, this invention designs a domain classifier to address the feature distribution difference between the source and target domains. It receives bottleneck layer features from the encoder and high-level encoded features (Enc5, Enc6). Each branch feature first passes through a gradient inversion layer (GRL) for adversarial training, and then passes through a kernel with a size of [missing information]. The convolutional layers reduce the dimensionality of the 512 channels to 128, 64, and 32 channels, respectively. The reduced multi-scale features are flattened and concatenated, then processed by ReLU activation and Dropout regularization before being input into a multilayer perceptron (MLP) for feature mapping. The final output is a domain logit, thus aligning the feature distributions of the source and target domains. An enhanced domain classifier and a gradient inversion layer are introduced to implement a domain confusion training mechanism: during generator training, the domain classification loss is calculated based on the output of the gradient inversion layer, driving the generator to learn domain-invariant features; simultaneously, the maximum mean difference (MMD) loss is introduced to align the feature distributions of the source and target domains, thereby reducing cross-domain shift effects; the domain classifier is independently optimized with the generator fixed to improve domain discrimination accuracy; the MMD loss formula is as follows:
[0113]
[0114] in, Represents the source domain sample set, Let m represent the target domain sample set, and m and n represent the number of samples in the source domain and target domain in a single training batch, respectively.
[0115] 7) Apply the TTA enhancement strategy to the target domain validation set, perform four spatial transformations on the input image and predict the elevation, and take the average value as the final digital elevation model output; calculate the mean absolute error (MAE) to evaluate the model accuracy. If the error does not improve after multiple consecutive validation rounds, trigger the early stopping mechanism, save the current best fine-tuned model, and record it as best_finetune_model.pth.
[0116] S5: Based on the Hann window, the digital elevation model values of the slices are fused to reconstruct the digital elevation model values of the entire region.
[0117] Step S5 includes:
[0118] 1) Standardize the digital elevation model values of the slices;
[0119] 2) The slices are stitched together based on their spatial coordinates in the complete image. Within the overlapping area of the slices, a two-dimensional Hann window function is used to perform weighted fusion of the overlapping pixels.
[0120] The two-dimensional Hann window function is expressed as follows:
[0121]
[0122] in, As weight, and These represent the dimensions of the slices on the x and y axes, respectively. This weighting gradually decreases in the edge regions of the weighting function, effectively smoothing the boundary transitions between slices and avoiding abrupt changes in the digital elevation model. After weighted fusion, the reconstructed entire elevation data is restored to the original geographic coordinate system, and the output is a TIFF format file (visualized as follows). Figure 7 As shown in the figure, it contains spatial reference information and pixel accuracy information, which facilitates subsequent model accuracy evaluation and visualization analysis.
Claims
1. A method for cross-regional remote sensing image elevation reconstruction based on domain adaptation, characterized in that, Includes the following steps: S1: Perform data preprocessing on the acquired raw remote sensing image data and elevation data; S2: Perform pixel-level alignment on the preprocessed remote sensing image data and elevation data, and resample and slice the data to obtain slice data; S3: Construct a cross-regional elevation migration prediction model based on domain adaptation; S4: Use slice data to train the cross-regional elevation migration prediction model, obtain the optimal model weights, and output the digital elevation model value of each slice through the trained cross-regional elevation migration prediction model. S5: Based on the Hann window, the digital elevation model values of the slices are fused to reconstruct the digital elevation model values of the entire region.
2. The method for cross-regional remote sensing image elevation reconstruction based on domain adaptation according to claim 1, characterized in that, The data preprocessing in step S1 includes geometric correction, radiometric calibration, and atmospheric correction.
3. The method for cross-regional remote sensing image elevation reconstruction based on domain adaptation according to claim 1, characterized in that, Step S2 includes: A1: Based on cubic spline interpolation, remote sensing image data is resampled. The formula for cubic spline interpolation is as follows: ； in, Let represent the coordinates of a pixel in the image, where i represents the grid index closest to y, and j represents the grid index closest to x. Represents the entire surrounding pixels; This represents the horizontal offset ratio of the target pixel x relative to the left and right boundaries of the y-th column of the original image. Indicates the vertical offset ratio; A2: Resampling of elevation images based on the average interpolation method. The formula for the average interpolation method is as follows: ； in, Indicating the original image pixel value of a point This represents the pixel value after resampling using the average interpolation method; A3: For resampled band and elevation images, the following formula is used to define the region: ； Where img and dem represent remote sensing image and elevation image, respectively; , , , These represent the left, right, top, and bottom boundaries of the overlapping area of the two images, respectively. A4: Perform data tiling on the remote sensing imagery and the digital elevation model imagery respectively.
4. The method for cross-regional remote sensing image elevation reconstruction based on domain adaptation according to claim 1, characterized in that, The cross-regional elevation migration prediction model in step S3 includes a generator based on U-net, a multi-scale discriminant module, and a domain classifier. The multi-scale discriminant module includes two branches: a global discriminant and a local discriminant. The global discriminant is responsible for judging the overall elevation distribution and global consistency, while the local discriminant performs local area authenticity judgment at the downsampled spatial resolution.
5. The method for cross-regional remote sensing image elevation reconstruction based on domain adaptation according to claim 4, characterized in that, The construction of the U-net-based generator in step S3 includes: Add channel attention and frequency attention mechanisms to each layer of the encoder; In the channel attention mechanism, given a feature map F, the formula for channel attention is as follows: ； in, Let H represent the feature value located at the i-th row and j-th column in the c-th channel of feature map F; H and W represent the height and width of the input feature map F, respectively. W1 represents the channel attention weights calculated by the c-th channel attention mechanism, and W2 represents the first weight and the second weight, respectively. In the sixth coding layer of U-net, a frequency attention module is introduced to enhance the perception of texture frequency and spatial structural details; the operation of the frequency attention module includes: Targeting bottleneck characteristics Where B represents the batch size, C represents the number of channels, and H and W represent the height and width of the feature map, respectively. After performing a two-dimensional Fourier transform on it, the real and imaginary parts are separated and stacked to form a 2C-channel frequency feature map. , and These represent the operations of extracting the real and imaginary parts, respectively, and then the weight relationship between the frequency channels is learned through two pointwise convolutions. Then, the frequency components are weighted and returned to the spatial domain. ,in, This represents the Hadamard product, which is performed element-wise. Introduce a cross-scale channel attention mechanism in the decoder: The decoder achieves high- and low-level feature fusion through cross-scale channel attention, using the difference between the generated simulated digital elevation model and the real digital elevation model as the core loss. It integrates pixel-level L1 loss, gradient loss, SSIM loss, and adversarial loss to increase generator accuracy. The formulas are as follows: ； Where Y represents the true digital elevation model, This represents the elevation model of the predicted data output by the generator network. Represents the mathematical expectation. Describing the L1 norm, and Let represent the gradient operators of the image along the horizontal x-axis and the vertical y-axis, respectively. , , These represent pixel-level loss, gradient loss, and structural similarity loss, respectively.
6. The method for cross-regional remote sensing image elevation reconstruction based on domain adaptation according to claim 5, characterized in that, The construction of the multi-scale discrimination module in step S3 includes: The discrimination result output by the global discriminator is: ;in, This represents the input image for the discriminator. This represents the convolution operation of the global discriminator; The discrimination result output by the local discriminator is: ,in, This represents the features of the input image after downsampling.
7. The method for cross-regional remote sensing image elevation reconstruction based on domain adaptation according to claim 6, characterized in that, In step S3, a spectral normalization constraint weight distribution is introduced into the discriminator to constrain the weights of each convolutional layer. Therefore, the joint loss of the multi-scale discriminator output is: ； in, , , These represent the input remote sensing imagery, actual digital elevation imagery, and predicted elevation imagery, respectively. and These represent the scores given by the global and local discriminators to the input sample pairs, respectively, which are used to measure the authenticity of the input samples. This is a gradient penalty term; This represents the gradient penalty coefficient, used to adjust the proportion of the gradient penalty term in the total loss.
8. The method for cross-regional remote sensing image elevation reconstruction based on domain adaptation according to claim 7, characterized in that, The training of the cross-regional elevation migration prediction model in step S4 includes: An improved Wasserstein GAN framework is used for adversarial training: during training, the generator and discriminator alternately update their parameters, forming a stable adversarial relationship; during the discriminator update phase, the generator parameters are fixed, and the adversarial relationship is minimized. ,renew During the generator update phase, the discriminator parameters are fixed, and the minimum value is minimized. ,renew The final objective of the loss function is: ； in, , These represent the distributions of real samples and generated samples, respectively. , , , These represent gradient penalty loss, pixel-level L1 loss, gradient loss, and structural similarity loss, respectively. , , , These represent the weighting balance coefficients corresponding to each loss term; A domain classifier and gradient inversion layer are introduced to implement a domain confusion training mechanism: During generator training, the domain classification loss is calculated based on the output of the gradient inversion layer to drive the generator to learn domain-invariant features; at the same time, the maximum mean difference (MMD) loss is introduced to align the feature distributions of the source domain and the target domain.
9. The method for cross-regional remote sensing image elevation reconstruction based on domain adaptation according to claim 8, characterized in that, Step S5 includes: B1: Standardize the digital elevation model values of the slices; B2: The slices are stitched together based on their spatial coordinates in the complete image. Within the overlapping area of the slices, a two-dimensional Hann window function is used to perform weighted fusion of the overlapping pixels.
10. The method for cross-regional remote sensing image elevation reconstruction based on domain adaptation according to claim 9, characterized in that, The expression of the two-dimensional Hann window function in step B2 is as follows: ； in, As weight, and These represent the dimensions of the slice on the x-axis and y-axis, respectively.