Multi-source DEM fusion and error correction method based on neural network

Abstract

This invention relates to the field of neural network multi-source technology, and in particular to a method for multi-source DEM fusion and error correction based on neural networks, comprising the following steps: data preprocessing, resampling multi-source DEM data, including InSAR DEM and optical DEM, to a unified resolution and a unified coordinate system; PEEs detection and preliminary error estimation, using the gradient direction consistency (GDC) detection method to locate error pixels (PEEs) in the InSAR DEM and perform preliminary error estimation to obtain a pre-corrected InSAR DEM. This invention introduces an integrated processing flow for multi-source DEM fusion and error correction, unifying multiple data such as InSAR DEM and optical DEM to the same coordinate system and resolution during the data preprocessing stage, thereby eliminating systematic errors caused by differences in data sources.

Description

Technical Field

[0001] This invention relates to the field of neural network multi-source technology, specifically to a method for fusion and error correction of multi-source DEMs based on neural networks. Background Technology

[0002] With the intensification of global environmental change and the rapid evolution of land use patterns, traditional digital elevation model (DEM) data, such as SRTM, are difficult to match with current high-resolution synthetic aperture radar (SAR) data due to their early data acquisition time and low spatial resolution. This results in complex DEM error correction processes and limited accuracy in applications such as differential interferometric synthetic aperture radar (DInSAR). Especially in complex terrain areas such as mountainous and urban areas, outdated DEM data cannot accurately reflect real terrain changes, seriously affecting the reliability of surface deformation monitoring and terrain modeling. Existing DEM update methods mostly rely on a single data source or simple interpolation fusion, lacking a mechanism for identifying and partitioning systematic errors in multi-source DEMs. This leads to significant deviations in the fusion results in areas with abrupt terrain changes, limiting their application in large-scale, high-precision terrain analysis.

[0003] Traditional multi-temporal InSAR (MT-InSAR) methods, such as PSInSAR and SBAS, heavily rely on phase unwrapping techniques and pre-defined deformation models when performing DEM error correction. This not only results in low computational efficiency but also easily introduces systematic biases due to unwrapping errors or model mismatches. Especially in complex terrain areas, the reliability of phase unwrapping decreases significantly, leading to poor DEM error correction results. Existing methods typically adopt a globally uniform correction strategy, failing to differentiate the regional differences in error distribution and effectively distinguish between actual terrain changes and sensor-induced errors, thus limiting their applicability and accuracy in multi-source DEM fusion. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for multi-source DEM fusion and error correction based on neural networks, which solves the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for multi-source DEM fusion and error correction based on neural networks, comprising the following steps:

[0006] Data preprocessing involves resampling multi-source DEM data, including InSAR DEM and optical DEM, to a uniform resolution and unifying the coordinate system;

[0007] PEE detection and preliminary error estimation: The gradient direction consistency (GDC) detection method is used to locate the error pixels (PEEs) in the InSAR DEM and perform preliminary error estimation to obtain the preliminarily corrected InSAR DEM.

[0008] Terrain feature extraction and sample segmentation: Terrain features are extracted from the corrected InSAR DEM and optical DEM. The residuals are calculated based on the reference DEM, outliers are screened out, and the data is divided into PEEs region samples and error-free region samples. The PEEs region samples include the error feature δh, while the error-free region samples do not include the error feature δh. The samples are also divided into training set, validation set and test set according to the proportion.

[0009] The neural network construction and training steps are as follows: a two-branch neural network is constructed, the TRAINLM training algorithm is adopted, RMSE and NMAD are used as loss functions, the network weights are iteratively optimized, and two sub-models are obtained for the PEEs region and the error-free region.

[0010] Partition fusion involves inputting the DEM data to be fused into the corresponding sub-model to obtain the fusion results of the PEEs region and the fusion results of the error-free region, and then stitching the two together to form a complete fused DEM.

[0011] Post-processing involves smoothing the stitched and merged DEM to eliminate boundary errors;

[0012] The verification step involves calculating the RMSE and NMAD indices of the fused DEM using the reference DEM to evaluate the accuracy of the fused DEM.

[0013] Preferably, in the PEEs detection and preliminary error estimation, the gradient direction consistency (GDC) detection calculates the gradient direction matrix of the InSARDEM, compares it pixel by pixel with the gradient direction matrix of the DEM, and when the direction difference exceeds a preset threshold, the pixel is marked as a PEEs. A preliminary error estimate is then performed based on the magnitude of the gradient difference to generate a preliminarily corrected InSARDEM.

[0014] Preferably, in the neural network construction and training steps, the dual-branch neural network is used to process PEEs region samples, and the branch is used to process error-free region samples. The branch consists of an input layer, multiple hidden layers, and an output layer. The hidden layers use the ReLU activation function, and the output layer uses the linear activation function. The network training adopts the TRAINLM algorithm, with the weighted sum of RMSE and NMAD as the loss function.

[0015] Preferably, in the terrain feature extraction and sample segmentation, the terrain features include slope, aspect, curvature and elevation variation coefficient, the reference DEM is LiDAR DEM, the residual is calculated as the difference between the DEM value and the LiDAR DEM value, and outlier screening adopts the 3x NMAD criterion, that is, samples with residuals greater than 3x NMAD value are removed.

[0016] Preferably, in the partition fusion step, for the PEEs region, the sub-model input includes terrain features and error features δh, and the output is the corrected elevation value; for the error-free region, the sub-model input only includes terrain features, and the output is the fused elevation value; when stitching the fusion results, the overlapping area weighted average method is used to ensure a smooth transition.

[0017] Preferably, in the post-processing step, the Gaussian smoothing process uses a two-dimensional Gaussian kernel function to convolve the stitching edges. The kernel size is adaptively adjusted according to the DEM resolution, the kernel standard deviation is set to the resolution, and smoothing is only applied to the area near the stitching boundary to avoid over-smoothing terrain details.

[0018] Preferably, in the verification step, the reference DEM is a LiDAR DEM. When calculating the RMSE and NMAD indices of the fused DEM, a pixel-by-pixel comparison method is used, and the accuracy is compared with that of the initial InSAR DEM and the optical DEM to evaluate the accuracy improvement effect of the fusion method.

[0019] Preferably, the method further includes the utilization of InSAR temporal phase information, introducing InSAR temporal phase data as an additional feature to calculate temporal deformation rate and coherence, and using it as input features of the neural network to optimize the fusion weights.

[0020] Preferably, in the data preprocessing step, the multi-source DEM data includes InSAR DEM, optical DEM and radar DEM, the resampling adopts bilinear interpolation, and the coordinate system is UTM projection system.

[0021] Preferably, the method is applicable to various terrain scenarios such as mountains, cities, and hills, wherein the neural network training adjusts the feature set and loss function weights for different scenarios to enhance scenario adaptability.

[0022] Compared with the prior art, the beneficial effects of the present invention are:

[0023] This invention provides a method for multi-source DEM fusion and error correction based on neural networks:

[0024] 1. By introducing an integrated processing flow for multi-source DEM fusion and error correction, in the data preprocessing stage, multiple data such as InSAR DEM and optical DEM are unified to the same coordinate system and resolution, eliminating systematic errors caused by differences in data sources. The gradient direction consistency (GDC) method is used to accurately identify error pixels (PEEs) in InSAR DEM and perform preliminary correction, significantly reducing the processing burden of subsequent neural networks. By constructing a dual-branch neural network, optimization training is performed separately for PEEs regions and error-free regions. Combined with terrain features and error features, partitioned fusion is performed, which not only improves the adaptability to complex terrain regions, but also ensures the overall consistency of the fusion results, improves the accuracy and efficiency of DEM fusion, and has strong scene generalization ability, which can be widely applied to various complex terrain environments such as mountains, cities, and hills.

[0025] 2. By rapidly locating error pixels (PEEs) in InSAR DEMs using GDC technology, the cumulative errors caused by the unwrapping process are avoided, significantly improving the accuracy and efficiency of error identification. A dual-branch neural network structure is constructed to process the PEEs region and the error-free region separately. The TRAINLM optimization algorithm and the RMSE+NMAD hybrid loss function are used to achieve accurate modeling and compensation for different error characteristics. In the extraction stage, InSAR temporal phase information is introduced to extract deformation rate and coherence, etc., enhancing the model's adaptability in temporally changing regions, effectively eliminating stitching boundary errors, and ensuring the spatial continuity and detail fidelity of DEM data. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0029] In the description of this invention, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0030] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "configuration" should be interpreted broadly. For example, they can refer to a fixed connection or configuration, a detachable connection or configuration, or an integral connection or configuration. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0031] like Figure 1 As shown, the present invention proposes a method for multi-source DEM fusion and error correction based on neural networks, which includes the following steps:

[0032] Step 1: Data preprocessing. Multi-source DEM data, including InSAR DEM and optical DEM, are resampled to a uniform resolution and coordinate system is unified.

[0033] Step 2, PEEs detection and preliminary error estimation: The gradient orientation consistency (GDC) detection method is used to locate the error pixels (PEEs) in the InSAR DEM and perform preliminary error estimation to obtain the preliminarily corrected InSAR DEM.

[0034] Step 3: Terrain feature extraction and sample division. Terrain features are extracted from the corrected InSAR DEM and optical DEM. The residuals are calculated based on the reference DEM, outliers are removed, and the data is divided into PEEs region samples and error-free region samples. The PEEs region samples include the error feature δh, while the error-free region samples do not include the error feature δh. The samples are then divided into training set, validation set, and test set according to the proportions.

[0035] Step 4, Neural Network Construction and Training: Construct a two-branch neural network, use the TRAINLM training algorithm, with RMSE and NMAD as loss functions, iteratively optimize the network weights, and obtain two sub-models for the PEEs region and the error-free region.

[0036] Step 5: Partition fusion. Input the DEM data to be fused into the corresponding sub-model to obtain the fusion results of the PEEs region and the fusion results of the error-free region, and then stitch the two together to form a complete fused DEM.

[0037] Step 6: Post-processing involves smoothing the stitched and merged DEM to eliminate boundary errors;

[0038] Step 7, Verification Step: Calculate the RMSE and NMAD indices of the fused DEM using the reference DEM to evaluate the accuracy of the fused DEM.

[0039] In practical application, during the data preprocessing stage, multi-source DEM data (such as InSAR DEM and optical DEM) are resampled and their coordinate systems are unified to ensure spatial consistency of the data, laying the foundation for subsequent processing. The DEM generated by the stereo image pairs is resampled to the InSAR DEM resolution to avoid errors caused by scale differences. In the PEEs detection and preliminary error estimation stage, the gradient direction consistency (GDC) method is used to detect error pixels (PEEs) in the InSAR DEM. GDC calculates the gradient direction of the DEM and compares it with the reference data to identify inconsistent regions as PEEs, thereby performing preliminary error correction and generating an InSAR_corr DEM. In the terrain feature extraction and sample segmentation stage, terrain features (such as slope and aspect) are extracted from the corrected DEM, and residuals are calculated with the LiDAR DEM as a reference. The samples are divided into PEEs regions (including δh error features) and error-free regions to ensure the generalization ability of the model training. In the neural network construction and training stage, a dual-branch neural network is designed to process PEEs and error-free region samples respectively. The TRAINLM (Levenberg-Marquardt) algorithm is used to optimize network weights. In the partition fusion stage, the data to be fused is input into the corresponding sub-model. The PEEs region sub-model uses the δh feature for error compensation, while the error-free region sub-model directly fuses terrain features, ultimately stitching together a complete DEM. In the post-processing stage, Gaussian smoothing is used to process the stitching edges. Through staged processing, efficient and high-precision DEM fusion and error correction are achieved, making it suitable for various scenarios such as mountainous and urban areas.

[0040] In the PEEs detection and preliminary error estimation, the gradient direction consistency (GDC) detection calculates the gradient direction matrix of the InSAR DEM, compares it pixel by pixel with reference to the gradient direction matrix of the DEM, and marks the pixel as a PEE when the direction difference exceeds a preset threshold. Based on the magnitude of the gradient difference, a preliminary error estimate is performed to generate a preliminarily corrected InSAR DEM.

[0041] In practical applications, during PEEs detection and preliminary error estimation, the GDC method locates error pixels by analyzing the gradient direction consistency of the DEM, calculates the gradient direction matrix of the InSAR DEM, typically using Sobel or Prewitt operators to extract horizontal and vertical gradients, and then calculates the gradient direction angle for each pixel. The same processing is applied to a reference DEM (such as a high-precision LiDAR DEM) to obtain a reference gradient direction matrix. The two gradient direction matrices are compared pixel by pixel, and by establishing a linear or nonlinear relationship between gradient differences and elevation errors, the error value δh for each PEE is estimated. The InSAR DEM is then preliminarily corrected to generate an InSAR_corr DEM. By utilizing the sensitivity of gradient direction to terrain changes, it can effectively detect errors caused by phase unwrapping errors or abrupt terrain changes. In urban areas, building shadows may cause gradient anomalies in the InSAR DEM, and GDC can accurately identify these areas. GDC detection and preliminary correction are performed in tandem, avoiding the bias of relying on phase unwrapping in traditional methods and improving computational efficiency.

[0042] In the neural network construction and training steps, the dual-branch neural network is used to process PEEs region samples, and the branch is used to process error-free region samples. The branch consists of an input layer, multiple hidden layers, and an output layer. The hidden layers use the ReLU activation function, and the output layer uses the linear activation function. The network training adopts the TRAINLM algorithm, with the weighted sum of RMSE and NMAD as the loss function.

[0043] In practical use, the dual-branch neural network is designed as two independent sub-networks, which optimize the fusion of the PEEs region and the error-free region respectively. Each branch includes an input layer that receives terrain feature data, while the PEEs branch additionally receives error features δh. The hidden layer uses the ReLU activation function to avoid the gradient vanishing problem and enhance nonlinear fitting ability. The output layer uses a linear activation function to directly output the fused elevation value. The network weights are optimized iteratively, and the validation set is used for early stopping to prevent overfitting. The PEEs branch learns the error compensation pattern, while the error-free branch learns the smooth fusion of terrain. The dual-branch structure achieves error partitioning, with the PEEs branch focusing on complex error correction and the error-free branch maintaining high-precision terrain details. Combined with the TRAINLM algorithm and a hybrid loss function, the model's generalization ability and accuracy under various terrains are improved.

[0044] In terrain feature extraction and sample segmentation, terrain features include slope, aspect, curvature and elevation variation coefficient. The reference DEM is LiDAR DEM. The residual is calculated as the difference between the DEM value and the LiDAR DEM value. Outlier removal adopts the 3-times NMAD criterion, that is, samples with residuals greater than 3 times the NMAD value are removed.

[0045] In practical application, during the terrain feature extraction stage, the spatial variation of the terrain is captured through neighborhood window calculation. Using the LiDAR DEM as a reference, the residual of each pixel is calculated. The residual distribution is used to identify outliers. Samples with residuals greater than 3 times the NMAD are considered outliers and are removed to avoid training bias. The samples are then divided into PEEs regions and error-free regions: PEEs region samples contain error features δh (estimated by G detection), while error-free region samples contain only terrain features. The division ratio is based on the error distribution, with a higher proportion of PEEs in mountainous areas. Through multi-feature extraction and strict outlier removal, data quality is enhanced, enabling the neural network to learn the relationship between terrain and error more accurately, thereby improving the reliability and accuracy of the fused DEM.

[0046] In the partition fusion step, for the PEEs region, the sub-model input includes terrain features and error features δh, and the output is the corrected elevation value. For the error-free region, the sub-model input only includes terrain features, and the output is the fused elevation value. When stitching the fusion results, the overlapping area weighted average method is used to ensure a smooth transition.

[0047] In practical application, during partitioned fusion, the DEM data to be fused is divided into PEEs regions and error-free regions based on GDC detection results. The input features of the sub-model include terrain features (such as slope and aspect) and error features δh (obtained from preliminary estimation). The sub-model learns the relationship between δh and terrain through a neural network and outputs corrected elevation values ​​to achieve error compensation. In mountainous areas, δh may be associated with phase unwrapping errors, and the sub-model learns its correction mode. When stitching the fusion results, the overlapping area weighted average method is used: an overlapping band (e.g., 2-3 pixels wide) is set at the boundary between PEEs and error-free regions. The elevation values ​​within the overlapping area are calculated by weighting the outputs of the two sub-models. The weights change linearly with distance to ensure a smooth transition. After stitching, a complete fused DEM is generated. The PEEs region focuses on correcting systematic errors, while the error-free region focuses on preserving details. Combined with weighted stitching, the overall consistency and visual smoothness of the DEM are improved.

[0048] In the post-processing step, Gaussian smoothing uses a two-dimensional Gaussian kernel function to convolve the stitching edges. The kernel size is adaptively adjusted according to the DEM resolution, and the kernel standard deviation is set to the resolution. Smoothing is only applied to the area near the stitching boundary to avoid over-smoothing terrain details.

[0049] In practical applications, after partition fusion, the stitching edges may be discontinuous due to model differences. Gaussian smoothing is used to eliminate these boundary errors. Smoothing is only applied to the area near the stitching boundary. New elevation values ​​are calculated through convolution, and the weights are determined by the Gaussian kernel. In urban areas, building edges may be jagged due to stitching. Gaussian smoothing can restore smooth contours. In mountainous areas, it can eliminate false terrain undulations. Through local adaptive smoothing, it reduces spatial inconsistencies introduced by model stitching, while avoiding the loss of terrain information caused by global smoothing, ensuring that the fused DEM is both smooth and faithful.

[0050] In the verification step, the reference DEM is a LiDAR DEM. When calculating the RMSE and NMAD indices of the fused DEM, a pixel-by-pixel comparison method is used. At the same time, the accuracy is compared with that of the initial InSAR DEM and optical DEM to evaluate the accuracy improvement effect of the fusion method.

[0051] The method also includes the utilization of InSAR temporal phase information, introducing InSAR temporal phase data as an additional feature to calculate temporal deformation rate and coherence, and using it as input features for neural networks to optimize fusion weights.

[0052] In practical application, during the feature extraction step, deformation rate and coherence features are extracted from the temporal phase: deformation rate reflects the slow movement of the ground surface, and coherence measures the stability of radar signals. During training, the neural network learns the correlation between temporal features and elevation error, optimizes the fusion weights, and in the PEEs branch, δh is combined with deformation rate to enhance the accuracy of error estimation. By capturing dynamic changes through temporal information, the shortcomings of a single temporal phase DEM are made up for, and the accuracy and reliability of the fused DEM in temporally changing regions are improved.

[0053] In the data preprocessing step, the multi-source DEM data includes InSAR DEM, optical DEM and radar DEM. The resampling adopts the bilinear interpolation method, and the coordinate system is the UTM projection system.

[0054] In practical use, data preprocessing covers multi-source DEM data, such as InSAR DEM (from DInSAR processing), optical DEM (from stereo pairs), and radar DEM (such as SRTM). Resampling uses bilinear interpolation to unify the data to the target resolution (such as the resolution of InSAR DEM), avoiding the jagged effect caused by nearest neighbor interpolation. The coordinate system is UTM (Universal Transverse Mercator Projection) to ensure spatial reference consistency. Through standardized preprocessing, system errors are reduced, ensuring seamless integration of multi-source data and improving the accuracy and efficiency of subsequent processing.

[0055] The method is applicable to various terrain scenarios, including mountains, cities, and hills. The neural network training adjusts the feature set and loss function weights for different scenarios to enhance scenario adaptability.

[0056] In practical applications, the neural network training is adaptively adjusted for different terrains such as mountains, cities, and hills. In mountainous scenarios, the feature set focuses on complex terrain features such as slope and curvature, and the NMAD weight in the loss function is increased to handle abnormal undulations. In urban scenarios, building height and temporal deformation features are introduced, and the RMSE weight is increased to ensure planar accuracy. The training data comes from samples of multiple scenarios to ensure the model's generalization ability. The mountain training set includes steep slope and valley samples, while the urban training set includes building clusters and road samples. Through scenario-specific optimization, the method can flexibly cope with different terrain challenges, eliminating phase unwrapping errors in mountains and correcting building shadow effects in cities to achieve high-precision fusion.

[0057] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0058] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for multi-source DEM fusion and error correction based on neural networks, characterized in that, Includes the following steps: Step 1: Data preprocessing, resampling multi-source DEM data, including InSAR DEM and optical DEM, to a uniform resolution and unifying the coordinate system; Step 2: PEEs detection and preliminary error estimation. The gradient orientation consistency (GDC) detection method is used to locate the error pixels (PEEs) in the InSAR DEM and perform preliminary error estimation to obtain the preliminarily corrected InSAR DEM. Step 3: Terrain feature extraction and sample division. Terrain features are extracted from the corrected InSAR DEM and optical DEM. The residuals are calculated based on the reference DEM, outliers are removed, and the data is divided into PEEs region samples and error-free region samples. The PEEs region samples include the error feature δh, while the error-free region samples do not include the error feature δh. The samples are then divided into training set, validation set, and test set according to the proportions. Step 4: Neural Network Construction and Training Steps. Construct a two-branch neural network, use the TRAINLM training algorithm, and use RMSE and NMAD as loss functions to iteratively optimize the network weights to obtain two sub-models for the PEEs region and the error-free region. Step 5: Partition fusion. Input the DEM data to be fused into the corresponding sub-model to obtain the fusion results of the PEEs region and the error-free region, and then stitch the two together to form a complete fused DEM. Step Six: Post-processing. The stitched and merged DEM is smoothed to eliminate boundary errors. The RMSE and NMAD indices of the merged DEM are calculated by referring to the DEM to evaluate the accuracy of the merged DEM.

2. The method for multi-source DEM fusion and error correction based on neural networks according to claim 1, characterized in that: In the PEEs detection and preliminary error estimation, the gradient direction consistency (GDC) detection calculates the gradient direction matrix of the InSAR DEM, compares it pixel by pixel with reference to the gradient direction matrix of the DEM, and marks the pixel as a PEE when the direction difference exceeds a preset threshold. Based on the magnitude of the gradient difference, a preliminary error estimate is performed to generate a preliminarily corrected InSAR DEM.

3. The method for multi-source DEM fusion and error correction based on neural networks according to claim 1, characterized in that: In the neural network construction and training steps, the dual-branch neural network is used to process PEEs region samples, and the branch is used to process error-free region samples. The branch consists of an input layer, multiple hidden layers, and an output layer. The hidden layers use the ReLU activation function, and the output layer uses the linear activation function. The network training adopts the TRAINLM algorithm, with the weighted sum of RMSE and NMAD as the loss function.

4. The method for multi-source DEM fusion and error correction based on neural networks according to claim 1, characterized in that: In the terrain feature extraction and sample segmentation, the terrain features include slope, aspect, curvature and elevation variation coefficient. The reference DEM is LiDAR DEM. The residual is calculated as the difference between the DEM value and the LiDAR DEM value. Outlier removal adopts the 3x NMAD criterion, that is, samples with residuals greater than 3x NMAD value are removed.

5. The method for multi-source DEM fusion and error correction based on neural networks according to claim 1, characterized in that: In the partition fusion step, for the PEEs region, the sub-model input includes terrain features and error features δh, and the output is the corrected elevation value. For the error-free region, the sub-model input only includes terrain features, and the output is the fused elevation value. When stitching the fusion results, the overlapping area weighted average method is used to ensure a smooth transition.

6. The method for multi-source DEM fusion and error correction based on neural networks according to claim 1, characterized in that: In the post-processing step, Gaussian smoothing uses a two-dimensional Gaussian kernel function to convolve the stitching edges. The kernel size is adaptively adjusted according to the DEM resolution, and the kernel standard deviation is set to the resolution. Smoothing is only applied to the area near the stitching boundary to avoid over-smoothing terrain details.

7. The method for multi-source DEM fusion and error correction based on neural networks according to claim 1, characterized in that: In the verification step, the reference DEM is a LiDAR DEM. When calculating the RMSE and NMAD indices of the fused DEM, a pixel-by-pixel comparison method is used. At the same time, the accuracy is compared with that of the initial InSAR DEM and the optical DEM to evaluate the accuracy improvement effect of the fusion method.

8. The method for multi-source DEM fusion and error correction based on neural networks according to claim 1, characterized in that: The method also includes the utilization of InSAR temporal phase information, introducing InSAR temporal phase data as an additional feature to calculate temporal deformation rate and coherence, and using it as input features for neural networks to optimize fusion weights.

9. The method for multi-source DEM fusion and error correction based on neural networks according to claim 1, characterized in that: In the data preprocessing step, the multi-source DEM data includes InSAR DEM, optical DEM and radar DEM. The resampling adopts the bilinear interpolation method, and the coordinate system is the UTM projection system.

10. The method for multi-source DEM fusion and error correction based on neural networks according to claim 1, characterized in that: The method is applicable to various terrain scenarios, including mountains, cities, and hills. The neural network training adjusts the feature set and loss function weights for different scenarios to enhance scenario adaptability.

Images