A method for automatically constructing a digital surface model from satellite images

Abstract

The application discloses a kind of satellite image digital surface model automatic construction method considering untextured water area, belong to photogrammetry and remote sensing technical field.The application includes: automatic data preprocessing, generates epipolar image pair;Adaptive stereo matching processing, initial digital surface model is generated by block pyramid matching and domain adaptive calibration;Water area intelligent optimization, state analysis is carried out to water area region, water surface dynamic type is determined and modeling strategy is dynamically selected, error closed loop correction mechanism is introduced, error source is traced back reversely and is locally corrected.The application realizes the dynamic adaptation of water area modeling, cross-domain generalization and self-correction, solves the problem that the precision of existing method decreases and matching fails due to domain offset in complex scenes such as wave water surface and inclined water surface, significantly improves the precision and automation degree of digital surface model of untextured water area, and can be widely applied in digital twin, urban planning, water conservancy monitoring and other fields.

Description

Technical Field

[0001] This invention relates to the technical field of photogrammetry and remote sensing, specifically to an automated method for constructing digital surface models from satellite imagery. Background Technology

[0002] Building digital surface models (DSMs) based on satellite stereo imagery is a key technology for acquiring high-precision global topographic information, with wide applications in digital twins, urban planning, and disaster assessment. In recent years, with the rapid development of deep learning technology, stereo matching networks, such as IGEV-Stereo, have achieved excellent matching accuracy on publicly available datasets, providing a new technical path for the automated production of satellite imagery DSMs. However, applying these technologies to large-scale, complex real-world production scenarios still faces three core challenges:

[0003] First, there is a contradiction between the assumption of a still water surface and the assumption of a dynamic water surface.

[0004] Existing water body optimization methods generally assume an ideal water surface, using plane fitting or global DEM filling to handle water body elevation anomalies. However, real water bodies have waves, currents, seasonal water level changes, and even sloping surfaces caused by wind and waves. When the deviation between the assumption and reality exceeds the tolerance threshold, plane fitting methods completely fail, and large areas of noise appear in the water body DSM. Literature research shows that even the state-of-the-art BPC-SWSR boundary constraint method still experiences a significant drop in accuracy to meter-level errors when dealing with water bodies with complex wave textures or significant slopes. In production practice, technical teams have to spend hours manually selecting water body boundaries and editing water body elevations block by block, severely restricting production efficiency.

[0005] Second, there is a contradiction between the domain sensitivity and production generalization of deep learning matchers.

[0006] Deep learning stereo matchers perform exceptionally well on laboratory datasets, but real satellite imagery comes from different sensors (GF-7, ZY-3, WorldView), different seasons (spring, summer, autumn, winter), and different regions (plains, mountains, coastal areas), resulting in vastly different imaging conditions. When the domain distribution of training and production data is inconsistent, the matcher's performance suffers a catastrophic decline, especially in textureless water areas, where matching errors expand from meters to tens of meters. Existing solutions require retraining or fine-tuning the model for new domains, but acquiring large amounts of labeled data is costly and cannot meet the demands of large-scale production. Even more challenging is the lack of effective diagnostic and repair methods when the model fails, forcing technicians to rely solely on manual editing to address problem areas.

[0007] Third, the contradiction between deterministic classification output and fuzzy boundaries.

[0008] Current technologies rely on DOM classification models to output binary water area masks, transforming the "water / land" boundary into a deterministic pixel-level segmentation result. However, the real water area boundary is a gradual transition zone, from water body → wetland → vegetation → dry land, which may be several meters or even tens of meters wide in imagery. Coupled with factors such as shadows, reflections, and vegetation occlusion, the classification result itself is uncertain. In production practice, after the classification model forcibly transforms the "probabilistic" boundary into a "deterministic" boundary, this boundary is used as "absolute truth" by the downstream water area plane fitting module. This amplifies the classification error step by step, ultimately causing the fitted elevation plane to deviate significantly from the actual water surface, requiring the entire batch of data to be reworked.

[0009] To address the aforementioned issues, while some studies have attempted to introduce optimization steps after DSM generation, these efforts primarily focus on improving building geometry and fail to provide a systematic solution to the pervasive production challenge of textureless water areas. Some cutting-edge methods, such as introducing tilt angle constraints and edge-preserving filtering, have made progress on academic datasets, but their engineering implementation still faces obstacles such as high computational resource requirements, strong parameter dependencies, and the inability to handle multiple error source couplings. Therefore, the industry urgently needs a complete technical solution that integrates dynamic adaptation, self-correction, and cross-domain generalization capabilities to drive satellite photogrammetry towards a higher level of automation and intelligence. Summary of the Invention

[0010] The purpose of this invention is to provide an automated method for constructing digital surface models from satellite imagery, in order to solve the problems mentioned in the background section.

[0011] To achieve the above objectives, the present invention provides the following technical solution: An automated method for constructing digital surface models from satellite imagery includes the following steps: Acquire satellite stereo imagery and reference data; The satellite stereo imagery is preprocessed to generate epipolar image pairs; Based on epipolar image pairs, block-based stereo matching processing is performed to generate an initial digital surface model; A state analysis is performed on the water areas in the initial digital surface model to determine the water surface dynamic type of each water area. The water surface dynamic type includes at least two of the following: still water surface type, wave water surface type, and sloping water surface type. Based on the dynamic type of the water surface in each water area, select the appropriate modeling strategy from at least two different water surface modeling strategies to construct an elevation model for the water area. The final digital surface model is generated by replacing the elevation data of the corresponding water area in the initial digital surface model with the constructed elevation model.

[0012] Furthermore, a state analysis is performed on the water areas in the initial digital surface model to determine the dynamic type of the water surface in each water area, including: Extracting spectral and textural features of the water area from the satellite stereo imagery; Based on spectral and textural features, analyze the degree of water surface undulation and / or water surface slope in the water area; Based on the analysis results, the water areas are classified into still water, wavy water, or sloping water types.

[0013] Furthermore, based on the water surface dynamics type of each water area, a suitable modeling strategy is selected from at least two different water surface modeling strategies, including: When the water surface dynamic type is still water surface, a plane fitting strategy is selected. Based on the elevation observation values ​​of the water boundary pixels, a random sampling consensus algorithm is used to fit a single plane. When the water surface dynamic type is wave water surface type, the segmented plane fitting strategy is selected to divide the water area into multiple sub-regions, and the random sampling consensus algorithm is used to fit the plane for each sub-region. When the water surface dynamic type is an inclined water surface type, a surface fitting strategy is selected, and the least squares method is used to fit the water surface model with slope and curvature.

[0014] Furthermore, the step of performing block-based stereo matching includes: The left image is divided into regular blocks. Based on the elevation range parameters of the rational function model, the corner points of the left image blocks are back-projected onto the right image plane to obtain the candidate matching regions of the right image. Construct an image pyramid, perform stereo matching on the top-level image of the pyramid, and obtain an initial disparity map; Calculate the average disparity of the effective pixels in the initial disparity map as the reference offset; The position of the right image block is calibrated based on the reference offset, so that the disparity distribution of the calibrated right image block and the left image block is statistically consistent with the preset training disparity distribution. Perform stereo matching on the image blocks after position calibration.

[0015] Furthermore, after constructing the elevation model of the water area, an error closed-loop correction step is also included: Error detection is performed on the water area, and the elevation values ​​of pixels within the water area are calculated as the elevation residuals of the fitted model. When the elevation residual exceeds a preset threshold, the error source is traced back in reverse. The error source includes at least one of the following: water area extraction error, edge pixel screening error, and stereo matching error. Based on the tracing results, at least one step in the water area extraction, edge pixel filtering, or stereo matching is locally corrected. The elevation model of the water area was reconstructed based on the corrected results.

[0016] Furthermore, the reverse tracing error source includes: Obtain an elevation residual distribution map within the water area; Spatial overlay analysis was performed on the elevation residual distribution map, the water boundary extraction results, and the edge pixel screening results. When the elevation residuals are concentrated in the boundary area, the error source is determined to be the water area extraction error or the edge pixel screening error. When the elevation residuals exhibit a random distribution unrelated to image texture, the source of error is determined to be stereo matching error.

[0017] Furthermore, it also includes a domain adaptive calibration step, which is performed before the block stereo matching process is executed: Extract the domain feature vector of the current image. The domain feature vector is used to characterize at least one of the following information: sensor type, imaging season, and regional texture style. Based on the domain feature vector, the internal parameters of the stereo matcher are adjusted online. The internal parameters include at least one of the feature normalization parameter, disparity search range, and smoothing term weight. Stereo matching is performed based on the adjusted stereo matcher.

[0018] Furthermore, the online adjustment of the internal parameters of the stereo matcher includes: The domain feature vector is input into a preset parameter mapping network to generate matcher parameters that are adapted to the current domain. The parameter mapping network is a lightweight neural network that is pre-trained with multi-domain labeled data to establish the mapping relationship between domain features and the optimal parameters of the matcher.

[0019] Furthermore, the preprocessing includes: Feature extraction algorithms are used to extract feature points from satellite stereo images and reference digital orthophoto images, and the elevation values ​​of the corresponding points are obtained from the reference digital elevation model to form a control point set; Based on the control point set, a single-image spatial resection correction of satellite imagery is performed using an image-side affine transformation model. The connection points between the control points and the stereo image pairs are merged to perform multi-view area network adjustment; The epipolar image pairs are generated using the projection trajectory method, and the height range of the rational function model is corrected based on the height range of the reference digital elevation model.

[0020] Furthermore, before performing state analysis on the water area in the initial digital surface model, the following steps are also included: A lightweight classification model is used to extract water areas from digital orthophoto maps and generate water area masks. Morphological processing is performed on the water area mask to filter small water areas with a filter area smaller than a preset threshold. Boundary pixels of the water area are extracted, and reliable boundary pixels are selected based on the local gray-level variance, which is calculated through a sliding window with a preset radius.

[0021] Compared with the prior art, the beneficial effects of the present invention are: 1. It achieves dynamic adaptation in water area modeling, significantly improving elevation accuracy in complex scenarios.

[0022] This invention overcomes the limitations of the traditional "still water surface plane assumption" by automatically identifying the dynamic type of water body—still water, wavy water, or sloping water—through water surface state analysis, and dynamically selecting an appropriate modeling strategy. For still water, RANSAC plane fitting is used; for wavy water, piecewise plane fitting is used; and for sloping water, quadratic surface fitting is used. This allows the system to adaptively handle various real-world scenarios, from calm lakes to stormy seas, and from horizontal reservoirs to sloping rivers. Experiments show that in wavy water regions, this invention reduces the mean absolute error and improves accuracy compared to traditional plane fitting methods; in sloping water regions, the error is reduced and accuracy is improved.

[0023] 2. An error closed-loop correction mechanism was constructed to avoid one-way error propagation and improve system robustness.

[0024] This invention breaks through the limitations of traditional unidirectional data flow processing pipelines by innovatively introducing an error closed-loop correction mechanism. When the residual of the water area elevation fitting exceeds a preset threshold, the system can trace back to the source of the error, distinguish between water area extraction errors, edge filtering errors, and stereo matching errors through spatial overlay analysis, and reconstruct the water area elevation model after local correction of the corresponding steps. This mechanism enables the system to have self-correcting capabilities, avoiding the problem of reworking the entire batch of data caused by the cascading amplification of classification errors downstream, and significantly reducing the workload of manual editing. Experimental results show that in typical areas, the error ratio of pixels is significantly improved compared to before optimization, which is significantly better than traditional methods.

[0025] 3. Cross-domain adaptive calibration was achieved, solving the domain offset problem of deep learning matchers.

[0026] This invention addresses the performance degradation of deep learning matchers across different sensors, seasons, and regions by proposing a lightweight online domain adaptation mechanism. A domain encoder extracts domain feature vectors from images in real time, and an input parameter mapping network generates matcher parameters adapted to the current domain, achieving online calibration of the matcher without retraining a large model. This mechanism enables a single model to handle images from different satellite platforms, different imaging seasons, and different geographical types, maintaining stable matching outputs even in textureless water areas, significantly improving the system's production generalization capability.

[0027] 4. It achieves full-process automation, reduces manual intervention, and improves production efficiency.

[0028] This invention achieves fully automated processing from automatic control point extraction, single-image coarse correction, and regional network adjustment, to adaptive stereo matching, dynamic water area modeling, and closed-loop error correction, completely eliminating the reliance on manual control points. Experiments using Gaofen-7 satellite imagery as the data source and airborne LiDARDSM as the ground truth verification method demonstrate that the DSM generated by this invention achieves extremely high levels of visual continuity and quantitative accuracy in water areas. The production process eliminates the need for manual selection of water body boundaries or manual editing of water area elevations, significantly improving the efficiency and automation of large-scale DSM production. Attached Figure Description

[0029] Figure 1 The present invention provides an automated construction method for digital surface models of satellite imagery that takes into account textureless water areas, which eliminates unreliable water area edge pixel maps (the left image is the neighborhood of the edge pixels, and the right image is the water area map); Figure 2 This is a flowchart illustrating an automated method for constructing digital surface models of satellite imagery that takes into account textureless water areas, as provided in an embodiment of the present invention. Detailed Implementation

[0030] 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.

[0031] The method in this embodiment is executed by a terminal, which can be a mobile phone, computer, PDA, laptop or desktop computer, etc. Of course, it can also be other devices with similar functions, and this embodiment does not limit them.

[0032] Example 1 This embodiment provides an automated method for constructing digital surface models from satellite imagery, taking into account textureless water surfaces. It is particularly suitable for still water scenes (such as lakes and reservoirs with minimal waves) as well as complex scenes such as wavy or sloping water surfaces. Figure 2 As shown, the method includes three stages: an automated data preprocessing stage, a basic processing flow stage, and a digital surface model optimization stage that takes into account textureless water areas.

[0033] I. Automated Data Preprocessing 1. Automatic control point generation First, the satellite stereo image pairs to be processed, along with the reference digital orthophoto map (DOM) and reference digital elevation model (DEM), are acquired. The Scale Invariant Feature Transform (SIFT) algorithm is used to extract feature points from both the stereo images and the reference DOM. A fast approximate nearest neighbor algorithm based on Euclidean distance is then used to coarsely match the extracted feature points to obtain corresponding image points. Subsequently, based on the coordinates of the corresponding image points in the reference DOM, the elevation values ​​of the corresponding points are directly extracted from the reference DOM to form a control point set. The control points from all stereo images (including the left and right images) are then fused to form the final control point set. In practical engineering applications, for cases where some satellite images deviate significantly from the reference coordinate system, image blocks larger than 1000×1000 pixels are used for automatic control point extraction to improve the robustness of the matching.

[0034] 2. Single-image coarse correction To address the problem of ineffective overlap between stereo image pairs due to systematic errors in satellite imagery, a coarse correction method based on control points for single-image spatial resection is proposed. The systematic error compensation model of this method is expressed in affine transformation form:

[0035] in, For image points The system error compensation value, For the affine transformation parameters to be solved, an error equation is constructed based on the rational function model (RFM):

[0036] in, The residual vector of the image point coordinate observations. The coefficient matrix, This is a vector of correction values ​​for the error compensation parameters. Given the difference vector between the observed and approximate values, the affine transformation parameters are solved by least squares adjustment to achieve system error correction for a single image.

[0037] 3. Connectivity point matching and multi-view area network adjustment The same method used for generating automatic control points is employed to obtain the corresponding connection points between stereo image pairs. A joint adjustment model is constructed by combining the control points and connection points. Based on the error equation of the rational function model, the affine transformation parameters and object coordinates are used as parameters to be determined.

[0038] in, Let be the partial derivative matrix of the object coordinates. It is the correction vector for the object space coordinates. By jointly adjusting, the image orientation parameters and object space coordinates are optimized simultaneously to improve the geometric positioning accuracy.

[0039] 4. Nuclear line calibration Epipolar imagery is generated using the projection trajectory method, and the rational function model parameters of the epipolar imagery are analyzed to form a standardized intermediate data format for satellite pipelines. In particular, an RPC height range correction strategy based on the height range of the reference digital elevation model is proposed: when the original RPC height range is too large, the generation of anomalous point clouds in the stereo matching stage is suppressed; when the original RPC height range is too small (e.g., not covering high-rise buildings), the true elevation range is restored and the correct photogrammetric geometry is maintained.

[0040] II. Basic Processing Flow 1. RPC-based block partitioning and pyramid adaptive stereo matching The left panoramic image is divided into regular blocks to generate all possible candidate stereo image pairs. For each left image block, its four corner points are back-projected onto the right image plane according to the elevation range parameters of the rational function model to obtain a set of 8 projection points. The minimum bounding rectangle of this projection point set is used as the candidate matching region of the right image. This method satisfies both the high-efficiency parallel computing requirements of traditional semi-global matching algorithms and the equal-size input requirements of deep learning stereo matching models.

[0041] Based on the above matching block generation method, a pyramid adaptive stereo matching process is constructed: First, the image is downsampled by 2 times to construct an image pyramid. Stereo matching is then performed on the top layer of the pyramid to obtain an initial disparity map. The mean value of the filtered disparity map is then calculated as the baseline offset.

[0042] in, It is the disparity map after filtering. The number of pixels, High-uncertainty pixels are filtered out by using 0.4 of the maximum probability of the probabilistic volume as a threshold. The right image block is then calibrated based on the calculated baseline offset to ensure that the predicted disparity distribution is statistically consistent with the training disparity distribution. Finally, stereo matching is performed on the finely registered image block.

[0043] 2. Forward intersection and DSM generation Based on the rational function model, the three-dimensional point forward intersection problem is constructed into a reprojection error minimization model. The Gauss-Newton iterative algorithm is used for point-by-point solution. Subsequently, an irregular triangular mesh is constructed based on the improved Delaunay triangulation algorithm to generate the initial digital surface model.

[0044] III. Optimization of Digital Surface Models Considering Textureless Water Areas 1. Water area extraction based on DOM classification A lightweight classification model is used to extract water areas from digital orthophoto maps. These maps can be generated from stereo images through block orthorectification or directly from multi-source reference DOM data. To improve algorithm efficiency, an area threshold is set. The system automatically filters out small water areas (which typically yield better stereo matching results). In the final output set of water areas, each individual water area is considered to have an elevation anomaly and requires further specialized processing.

[0045] 2. Reliable edge pixel filtering mechanism To address two types of uncertainties in the water boundary classification results—the risk of misclassification by the classification model and the degradation of matching quality caused by textureless regions—a reliable edge pixel selection algorithm based on local texture analysis is proposed, such as... Figure 1 As shown, iterate through all pixels at the edge of the water area, establish a sliding window with a radius of 5 pixels, calculate the grayscale variance of pixels within the window, and set a threshold when the local texture variance is below a certain value. When the edge pixel is determined to be a reliable pixel, its elevation value is used as a constraint condition for fitting the water surface.

[0046] 3. Water Surface State Analysis and Adaptive Modeling Before constructing the elevation model, the extracted water areas are first analyzed to determine the dynamic type of the water surface in each area. Specifically, the spectral and texture features of the water areas in the satellite stereo image are extracted. The spectral differences between the water and land are analyzed based on the spectral features, and the grayscale variation patterns inside the water surface are analyzed based on the texture features. The first threshold for texture variance is set to 5, and the second threshold is set to 20. The values ​​can be adjusted according to the image resolution. When the texture variance inside the water area is less than the first threshold, it is determined to be a still water surface. When the texture variance is between the first and second thresholds and shows a periodic distribution, it is determined to be a wavy water surface. When there are systematic elevation differences at the water boundary, it is determined to be a sloping water surface.

[0047] Based on the determined water surface dynamics type, select the appropriate water surface modeling strategy: When the water surface dynamic type is still water, a plane fitting strategy is adopted. Based on the reliable edge pixels obtained through screening, the Random Sampling Consensus (RANSAC) algorithm is used to fit the optimal elevation plane. The elevation plane model of a water body is defined as follows:

[0048] in, Coordinates The height value at that location, The optimal plane parameters for the water surface are obtained by solving based on reliable pixels, and the RANSAC algorithm is used to iteratively solve for them.

[0049] When the water surface dynamic type is wave surface, a segmented plane fitting strategy is adopted: the water area is divided into multiple sub-regions, and the RANSAC algorithm is used to fit the plane for each sub-region to form a segmented plane model. The division of sub-regions is adaptively adjusted according to the wavelength and wave height of the waves. The wavelength and wave height are obtained by spectral analysis of the internal texture of the water area.

[0050] When the water surface dynamic type is a sloping water surface, a surface fitting strategy is adopted: a water surface model with slope and curvature is fitted, and the surface model adopts a quadratic surface form.

[0051] in, The surface parameters to be solved are obtained using the least squares method.

[0052] 4. Elevation model replacement The elevation data of the corresponding water area in the initial digital surface model is replaced with the constructed elevation model to generate the final digital surface model.

[0053] Example 2 Based on Example 1, this embodiment further introduces an error closed-loop correction mechanism to improve the reliability and robustness of the water area optimization results.

[0054] The difference between this embodiment and Embodiment 1 is that, after performing block-based stereo matching processing to generate the initial digital surface model and completing the construction of the water area elevation model, an error detection and closed-loop correction step is added.

[0055] Specifically, after constructing the water area elevation model based on the reliable edge pixels obtained through screening, error detection is performed on the water area, and the elevation residual between the elevation values ​​of the pixels inside the water area and the fitted model is calculated. The preset threshold for the elevation residual is set to 2.5 meters (this value can be adjusted according to the elevation accuracy requirements of the application scenario). When the elevation residual exceeds the preset threshold, the reverse tracing process is triggered.

[0056] The reverse tracing of error sources includes the following steps: obtaining the elevation residual distribution map inside the water area, performing spatial overlay analysis on the elevation residual distribution map with the water boundary extraction results and edge pixel screening results, and determining the error source as water area extraction error or edge pixel screening error when the elevation residual is concentrated in the boundary area; and determining the error source as stereo matching error when the elevation residual shows a random distribution unrelated to image texture.

[0057] Based on the tracing results, local corrections are made to the corresponding steps: if the error source is the extraction error of the water area, the classification threshold of the lightweight classification model is adjusted or a higher-precision segmentation model is used to re-extract the water area; if the error source is the edge pixel screening error, the sliding window radius or grayscale variance threshold is adjusted, and reliable edge pixels are re-screened; if the error source is the stereo matching error, the offset calculation parameters of the pyramid matching or the disparity search range are adjusted, and stereo matching is re-executed.

[0058] After the correction is completed, the elevation model of the water area is reconstructed based on the corrected results until the elevation residuals converge to below the preset threshold.

[0059] This embodiment achieves automatic error detection and correction through a closed-loop correction mechanism, avoiding accuracy loss caused by the propagation of errors downstream, and improving the robustness and automation of the system.

[0060] The other steps in this embodiment are the same as in Embodiment 1, and will not be repeated here.

[0061] Example 3 This embodiment, based on Embodiment 1 or Embodiment 2, further introduces a domain adaptive calibration mechanism to solve the domain offset problem of the deep learning stereo matcher on images from different sensors, different seasons, and different regions.

[0062] The difference between this embodiment and the previous embodiment is that a domain adaptive calibration step is added before performing the block stereo matching process.

[0063] Specifically, after acquiring satellite stereo imagery, the domain feature vector of the current image is extracted. The domain feature vector is used to characterize at least one of the following information: sensor type, imaging season, and regional texture style. The extraction of the domain feature vector is achieved by a lightweight domain encoder, which is built based on a convolutional neural network. It takes a 256×256 pixel image block as input and outputs a 128-dimensional domain feature vector.

[0064] Based on the extracted domain feature vectors, the internal parameters of the stereo matcher are adjusted online. The internal parameters include at least one of feature normalization parameters, disparity search range, and smoothing term weights. The adjustment process is as follows: The domain feature vectors are input into a preset parameter mapping network to generate matcher parameters adapted to the current domain. The parameter mapping network is a lightweight multilayer perceptron, which is pre-trained with multi-domain labeled data to establish the mapping relationship between domain features and the optimal parameters of the matcher. The training samples of the parameter mapping network include satellite images from different sensors, such as GF-7, ZY-3, WorldView-3, different seasons (spring, summer, autumn, winter), and different regions (plains, hills, mountains, and coastlines), and their corresponding optimal matcher parameters.

[0065] Based on the adjusted stereo matcher parameters, stereo matching processing is performed to obtain a more stable disparity map.

[0066] Through domain adaptive calibration, this embodiment enables the deep learning matcher to maintain stable matching performance when applied across domains, effectively solving the problem of decreased water area matching accuracy caused by domain offset in traditional methods.

[0067] The other steps in this embodiment are the same as in Embodiment 1 or Embodiment 2, and will not be repeated here.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. Those skilled in the art will recognize that the modules and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0069] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, equipment, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0070] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0071] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program instructions, such as USB flash drives, portable hard drives, read-only storage servers, random access storage servers, magnetic disks, or optical disks.

[0072] Furthermore, it should be noted that the combination of the various technical features in this case is not limited to the combination methods described in the claims of this case or the combination methods described in the specific embodiments. All technical features described in this case can be freely combined or combined in any way, unless they contradict each other.

[0073] It should be noted that the above examples are merely specific embodiments of the present invention, and the present invention is obviously not limited to the above embodiments, with many similar variations. All modifications that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should fall within the protection scope of this invention.

[0074] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for automatically constructing digital surface models from satellite imagery, characterized in that, Includes the following steps: Acquire satellite stereo imagery and reference data; The satellite stereo imagery is preprocessed to generate epipolar image pairs; Based on epipolar image pairs, block-based stereo matching processing is performed to generate an initial digital surface model; A state analysis is performed on the water areas in the initial digital surface model to determine the water surface dynamic type of each water area. The water surface dynamic type includes at least two of the following: still water surface type, wave water surface type, and sloping water surface type. Based on the dynamic type of the water surface in each water area, select the appropriate modeling strategy from at least two different water surface modeling strategies to construct an elevation model for the water area. The final digital surface model is generated by replacing the elevation data of the corresponding water area in the initial digital surface model with the constructed elevation model.

2. The method for automatically constructing a digital surface model from satellite imagery according to claim 1, characterized in that, A state analysis is performed on the water areas in the initial digital surface model to determine the dynamic type of the water surface in each water area, including: Extracting spectral and textural features of the water area from the satellite stereo imagery; Based on spectral and textural features, analyze the degree of water surface undulation and / or water surface slope in the water area; Based on the analysis results, the water areas are classified into still water, wavy water, or sloping water types.

3. The method for automatically constructing a digital surface model from satellite imagery according to claim 1, characterized in that, Based on the water surface dynamics of each water area, a suitable modeling strategy is selected from at least two different water surface modeling strategies, including: When the water surface dynamic type is still water surface, a plane fitting strategy is selected. Based on the elevation observation values ​​of the water boundary pixels, a random sampling consensus algorithm is used to fit a single plane. When the water surface dynamic type is wave water surface type, the segmented plane fitting strategy is selected to divide the water area into multiple sub-regions, and the random sampling consensus algorithm is used to fit the plane for each sub-region. When the water surface dynamic type is an inclined water surface type, a surface fitting strategy is selected, and the least squares method is used to fit the water surface model with slope and curvature.

4. The method for automatically constructing a digital surface model from satellite imagery according to claim 1, characterized in that, The segmented stereo matching process includes: The left image is divided into regular blocks. Based on the elevation range parameters of the rational function model, the corner points of the left image blocks are back-projected onto the right image plane to obtain the candidate matching regions of the right image. Construct an image pyramid, perform stereo matching on the top-level image of the pyramid, and obtain an initial disparity map; Calculate the average disparity of the effective pixels in the initial disparity map as the reference offset; The position of the right image block is calibrated based on the reference offset, so that the disparity distribution of the calibrated right image block and the left image block is statistically consistent with the preset training disparity distribution. Perform stereo matching on the image blocks after position calibration.

5. The method for automatically constructing a digital surface model from satellite imagery according to claim 1, characterized in that, After constructing the elevation model of the water area, the error closed-loop correction step is also included: Error detection is performed on the water area, and the elevation values ​​of pixels within the water area are calculated as the elevation residuals of the fitted model. When the elevation residual exceeds a preset threshold, the error source is traced back in reverse. The error source includes at least one of the following: water area extraction error, edge pixel screening error, and stereo matching error. Based on the tracing results, at least one step in the water area extraction, edge pixel filtering, or stereo matching is locally corrected. The elevation model of the water area was reconstructed based on the corrected results.

6. The method for automatically constructing a digital surface model from satellite imagery according to claim 5, characterized in that, The reverse tracing error sources include: Obtain an elevation residual distribution map within the water area; Spatial overlay analysis was performed on the elevation residual distribution map, the water boundary extraction results, and the edge pixel screening results. When the elevation residuals are concentrated in the boundary area, the error source is determined to be the water area extraction error or the edge pixel screening error. When the elevation residuals exhibit a random distribution unrelated to image texture, the source of error is determined to be stereo matching error.

7. The method for automatically constructing a digital surface model from satellite imagery according to claim 1, characterized in that, It also includes a domain adaptive calibration step, which is performed before the block stereo matching process: Extract the domain feature vector of the current image. The domain feature vector is used to characterize at least one of the following information: sensor type, imaging season, and regional texture style. Based on the domain feature vector, the internal parameters of the stereo matcher are adjusted online. The internal parameters include at least one of the feature normalization parameter, disparity search range, and smoothing term weight. Stereo matching is performed based on the adjusted stereo matcher.

8. The method for automatically constructing a digital surface model from satellite imagery according to claim 7, characterized in that, The online adjustment of the internal parameters of the stereo matcher includes: The domain feature vector is input into a preset parameter mapping network to generate matcher parameters that are adapted to the current domain. The parameter mapping network is a lightweight neural network that is pre-trained with multi-domain labeled data to establish the mapping relationship between domain features and the optimal parameters of the matcher.

9. The method for automatically constructing a digital surface model from satellite imagery according to claim 1, characterized in that, The preprocessing includes: Feature extraction algorithms are used to extract feature points from satellite stereo images and reference digital orthophoto images, and the elevation values ​​of the corresponding points are obtained from the reference digital elevation model to form a control point set; Based on the control point set, a single-image spatial resection correction of satellite imagery is performed using an image-side affine transformation model. The connection points between the control points and the stereo image pairs are merged to perform multi-view area network adjustment; The epipolar image pairs are generated using the projection trajectory method, and the height range of the rational function model is corrected based on the height range of the reference digital elevation model.

10. The method for automatically constructing a digital surface model from satellite imagery according to claim 1, characterized in that, Before performing state analysis on the water area in the initial digital surface model, the following steps are also included: A lightweight classification model is used to extract water areas from digital orthophoto maps and generate water area masks. Morphological processing is performed on the water area mask to filter small water areas with a filter area smaller than a preset threshold. Boundary pixels of the water area are extracted, and reliable boundary pixels are selected based on the local gray-level variance, which is calculated through a sliding window with a preset radius.

Images