Geomorphologic geometry and physical rule coordinated digital elevation model artifact detection method

Abstract

The application discloses a kind of topography geometry and physical rule coordinated digital elevation model (DEM) artifact detection method.The method is first sampled to target DEM, constructs model training positive and negative sample data set;Then extract sample multidimensional topographic features, and train classification model based on CatBoost algorithm, identify morphological artifacts;While using edge detection filter to process DEM, determine the splicing artifact range by window screening;Subsequently, a difference model is constructed by introducing multi-source reference DEM to obtain elevation error distribution, and a risk threshold is set to delineate the elevation error risk area, complete artifact severity classification, and build a DEM artifact classification detection system.The present application can comprehensively detect DEM artifacts, suitable for large-scale DEM data authenticity testing in large areas, and can provide reference for quality control of DEM products generated by synthetic aperture radar.

Description

Technical Field

[0001] This invention relates to the quality inspection of digital elevation models, and in particular to a method for detecting artifacts in digital elevation models based on the synergy of topographic geometry and physical rules. Background Technology

[0002] Digital elevation models (DEMs), as the fundamental digital representation of landforms, play a crucial supporting role in fields such as hydrological modeling, geomorphological analysis, and surveying engineering. Currently, various methods have been developed for detecting DEM artifacts, mainly including probabilistic statistical methods, fixed slope threshold methods, and multi-topographic feature fusion methods.

[0003] However, existing detection technologies still have limitations. Due to the complex causes of DEM artifacts, they may originate from terrain overlay, shadowing, and phase unwrapping errors during Interferometric Synthetic Aperture Radar (InSAR) imaging, or from systematic errors in data acquisition and processing. This leads to significant differences in morphological characteristics and statistical distribution among different types of artifacts. Therefore, detection methods based on single features cannot comprehensively cover all types of artifacts. While existing multi-topographic feature fusion methods can effectively identify obvious terrain anomalies, they often fail to identify the implicit elevation error propagation phenomenon induced by artifacts, thus failing to meet the needs of high-precision terrain analysis.

[0004] To address the shortcomings of existing technologies, this invention proposes a DEM artifact detection method that coordinates topographic geometry and physical rules. This method constructs a multi-strategy fusion framework that cross-validates morphological differences and vertical accuracy. Specifically, it includes: morphological anomaly detection based on the fusion of multiple topographic geometric features to capture terrain anomalies caused by geometric distortion; elevation anomaly detection guided by a multi-source reference set, which accurately identifies elevation error risk areas caused by artifacts against a real terrain background by introducing external reference data; and stitching edge detection based on physical rules, specifically addressing discontinuities at data stitching boundaries.

[0005] By integrating the above-mentioned multi-dimensional detection evidence, this invention achieves accurate stratification and identification of DEM artifact regions, effectively overcoming the limitations and incompleteness of existing detection methods. It provides a reliable technical means for validating DEM data in specific regions and can also provide important reference for DEM quality evaluation and related research. Summary of the Invention

[0006] The technical problem this invention aims to solve is how to achieve high-precision artifact recognition of Digital Elevation Models (DEMs) by using morphological anomaly detection based on the fusion of multiple geomorphic geometric features, physical rule-guided stitching edge detection, and multi-source reference set-guided elevation anomaly detection, and by generating a complete DEM artifact hierarchical monitoring system based on the above detection methods. To this end, this invention proposes a digital elevation model artifact detection method based on the synergy of geomorphic geometry and physical rules, with the following specific steps:

[0007] 1. DEM Sample Collection: DEM artifact samples (positive samples) and natural terrain samples (negative samples) are extracted from the target area and obtained in advance by visual interpretation. The positive samples cover various artifact types in the target area, including at least abnormal elevation rises around mountains, abnormal undulation noise around water bodies, and irregular steep rises and falls in terrain. The negative samples are natural terrain areas, covering all terrain types in the target area, and the negative samples and positive samples are set to a preset sampling ratio. Oversampling is performed on natural terrain with steep slopes and significant undulations to distinguish artifacts from highly undulating terrain features.

[0008] 2. Machine Learning Training and Artifact Detection of the Artifact Detection Model: Based on selected multidimensional feature parameters, the CatBoost ensemble algorithm is used for model training, and the trained model is used to identify artifacts on the target DEM. The multidimensional feature parameters consist of terrain feature parameters, image filter parameters, and optional DEM auxiliary layer parameters. The terrain feature parameters include: slope, average normal vector angle difference (ANVAD), spherical normal vector standard deviation (SSDoN), elevation standard deviation (SD), slope standard deviation (SDoS), maximum downhill elevation change (MaxDsEC), maximum uphill elevation change (MaxUsEC), terrain roughness, terrain roughness index (TRI), terrain position index (TPI), and total curvature (TotCuv). The image filter parameters include features extracted based on the Roberts operator, Prewitt operator, Laplacian operator, and difference of Gaussian (DoG) operator. The DEM auxiliary layer parameters are the elevation error map (HEM) derived from the target DEM data.

[0009] 3. Lateral stitching artifact detection: The DEM is filtered using a lateral edge detection filter to extract lateral gradient data. Pixels with gradient absolute values ​​exceeding a preset height threshold are identified as potential feature points. Subsequently, a sliding window with a preset number of rows and two columns is used to traverse the gradient data laterally row by row with a step size of 1 pixel. If all pixels within the window are potential feature points (i.e., the window is filled with target pixels), then lateral stitching artifacts are determined to exist in that region.

[0010] 4. Vertical stitching artifact detection: The processing logic in step 3 is followed for detection. The difference is that a vertical edge detection filter is used to obtain vertical gradient data, and a sliding window with two rows of preset columns is used to slide and judge the suspected vertical feature points.

[0011] 5. Integration of stitching artifact results: Perform a union operation on the horizontal stitching artifact detection results and the vertical stitching artifact detection results, and determine the result after the operation as the final stitching artifact detection result;

[0012] 6. Reference DEM preparation: Select at least two high-precision DEMs with the same resolution as the DEM to be detected as reference datasets; wherein, the data source of the reference DEM must be independent of the data source of the DEM to be detected, and it must be obtained based on a heterogeneous observation mechanism, excluding data that have a homogeneous derivation relationship with the DEM to be detected, so as to ensure the accuracy of the difference calculation;

[0013] 7. Geodetic datum harmonization: Harmonize the geodetic datum of all reference DEMs to be consistent with the study DEM, using the following formula:

[0014]

[0015] in This indicates the elevation of the reference DEM after conversion, under the same geodetic datum as the study DEM; Refers to the elevation of the reference DEM under the original geodetic datum; This indicates the geoid difference in the study of the DEM geodetic datum. This indicates the geoid difference relative to the DEM geodetic datum.

[0016] 8. Differential DEM (DoD): The reference DEM is subdivided pixel by pixel using the DEM data source to be inspected, forming a differential DEM map. The differential DEM formula is as follows:

[0017]

[0018] Where DoD represents the absolute vertical deviation of a given pixel. and These represent the elevation values ​​of the target DEM and the reference DEM, respectively.

[0019] 9. Calculate the risk zone threshold: Determine the threshold used to define the elevation risk zone of the DEM; The elevation risk zone is defined as: the surrounding area where, although the terrain morphology appears normal, its absolute elevation error has reached the artifact level due to the influence of artifact elevation fluctuations; The risk zone threshold is calculated according to the following formula:

[0020]

[0021] Where T represents the threshold for identifying risk areas under a certain reference DEM, and N represents the total number of pixels within the confirmed artifact area (the union of machine learning recognition results and edge detection recognition results). This indicates the specific DoD value of a single pixel within a reference DEM in this region;

[0022] 10. Risk Zone Delineation: If a pixel is not an artifact and its DoD result in all reference DEMs is greater than its corresponding T value, then the pixel is defined as a risk zone.

[0023] 11. Delineation of Severe Artifact Regions: If a pixel is an artifact and the DoD results in all reference DEMs are greater than its corresponding T value, then the pixel is defined as a severe artifact.

[0024] 12. Delineation of mild artifact regions: If a pixel is an artifact and the DoD results in all reference DEMs are not all greater than its corresponding T value, then the pixel is defined as a mild artifact. Attached Figure Description

[0025] The invention will now be further described with reference to the accompanying drawings.

[0026] Figure 1 It is the geographical location and elevation distribution of the example area (Alps).

[0027] Figure 2 This is an overall flowchart of a digital elevation model artifact detection method that integrates topographic geometry and physical rules.

[0028] Figure 3 This is a flowchart of a risk area detection and artifact severity classification method based on the risk area determination threshold.

[0029] Figure 4 These are typical detailed satellite images of the implementation area.

[0030] Figure 5 The results show the DEM of typical details in the example area and the detection results of severe artifacts.

[0031] Figure 6 This is the result of DEM showing typical details and minor artifact detection in the example area.

[0032] Figure 7 This is the detection result of typical DEM and detailed risk areas in the instance area. Detailed Implementation

[0033] The following section uses TDX30 as the research DEM, the Alps as the research area, and AW3D30 and SRTM V3 as reference DEMs to illustrate the invention in detail with reference to the accompanying drawings, making the technical route and operation steps of the invention clearer.

[0034] Figure 1 The location of the example area and its elevation data in TDX30 are given. The flowchart of the artifact detection method for digital elevation models based on the synergy of topographic geometry and physical rules proposed in this invention is shown below. Figure 2 As shown, the principles for determining various artifacts and risk areas are as follows: Figure 3 As shown, it includes the following steps:

[0035] The first step, DEM sample collection: Following the principles described in specific step 1 of the invention, the flow polygon tool of GIS software was used to delineate positive samples (artifact areas) and negative samples (natural terrain areas) within the Alps. A total of 398,502 positive samples and 985,013 negative samples were ultimately obtained.

[0036] The second step involves machine learning training and artifact detection of the artifact detection model: (e.g., ...) Figure 2 As shown in the first step, the model was trained using the parameters described in step 2 of the invention (including the HEM layer of TDX30). The training results showed that the model's test accuracy, precision, recall, and F1 scores were 97.51%, 96.77%, 94.49%, and 95.62%, respectively. Furthermore, the model's area under the ROC curve (ROC-AUC) was 99.53%, and its Kappa coefficient (Cohen's Kappa coefficient) was 0.93, demonstrating the model's reliability. Using this model to detect TDX30 data from the Alps, regions with a confidence level of 0.95 or higher were identified as artifact regions.

[0037] The third step is lateral edge detection filtering: (e.g., ...) Figure 2 As shown, a horizontal edge detection filter is used to filter the DEM and extract horizontal gradient data. Pixels with gradient absolute values ​​exceeding a preset height threshold (100m) are identified as suspected feature points. Subsequently, a 10×2 (row × column) sliding window is used to traverse the gradient data horizontally row by row with a step size of 1 pixel. If all pixels within the window are suspected feature points (i.e., the window is filled with target pixels), then the region is determined to have horizontal stitching artifacts.

[0038] Step 4, detection of vertical stitching artifacts: such as Figure 2 As shown, the detection is performed using the same processing logic as in step three. The difference is that a vertical edge detection filter is used to obtain vertical gradient data, and a 2×10 (row × column) sliding window is used to slide and traverse and determine the suspected vertical feature points accordingly.

[0039] Step 5, integration of stitching artifact results: perform a union operation on the horizontal stitching artifact detection results and the vertical stitching artifact detection results, and determine the result after the operation as the final stitching artifact detection result;

[0040] Step 6, Reference DEM Preparation: Select reference DEMs according to the following criteria: Select at least two high-precision DEMs with the same resolution as the DEM to be detected as reference datasets. The data source of the reference DEMs must be independent of the data source of the DEM to be detected, and must be acquired based on a heterogeneous observation mechanism. Data with homologous derivation relationships to the DEM to be detected should be excluded to ensure the accuracy of the differential calculation. Based on the above criteria, this embodiment specifically selects AW3D30 and SRTM V3 as reference DEMs for subsequent differential DEM (DoD) analysis.

[0041] Step 7, Geodetic Datum Coordination: During the DEM comparative analysis, it is necessary to ensure the geodetic consistency of each dataset. The specific processing is as follows: TDX30 uses the EGM08 vertical datum, while AW3D30 and SRTM V3 use the EGM96 vertical datum. To ensure geodetic consistency during the comparative analysis, the reference datasets are unified to the EGM08 datum through the following transformation:

[0042]

[0043] in, This indicates the DEM elevation relative to the EGM08 datum. This indicates the DEM elevation relative to the EGM96 datum. It represents the distance from the EGM96 geoid to the reference ellipsoid (i.e., the geoid difference). This indicates the geoid offset of the EGM08 datum.

[0044] Step 8, Differential DEM (DoD): (e.g.) Figure 3 As shown in Phase 1, each pixel in the DEM is differentially processed to form a DEM difference map. The formula for the differential DEM is as follows:

[0045]

[0046] Where DoD represents the absolute vertical deviation of a given pixel. and These represent the elevation values ​​of the TDX30 DEM and the reference DEM, respectively. Therefore, by using AW3D30 and SRTM V3 as reference DEMs for replacement, two independent sets of difference results can be obtained. Figure 2 , 3 is recorded as and .

[0047] Step 9: Calculate the risk zone threshold: (e.g.) Figure 3 As shown in Stage 1, to combine these vertical indicators with previously identified geometric anomalies, a statistical baseline needs to be established. This is achieved by using the merged set of confirmed artifact regions (derived from the union of machine learning and edge detection results). Figure 3 Chinese record The threshold is calculated, and the discrimination threshold for potential risk areas is calculated as the average vertical error within these confirmed areas:

[0048]

[0049] Where T represents the threshold for identifying risk areas, and N represents the total number of pixels within the confirmed artifact areas. This represents the specific DoD value of a single pixel within the region (the intersection of the artifact region and the DoD result is in...). Figure 2 , 3 is recorded as and ). Parameters derived from AW3D30 and SRTM V3 ( and Substituting into the formula, we obtain the risk region thresholds for the two corresponding DoD datasets: =127.045m, =124.779m.

[0050] Step 10, Risk Zone Delineation: (e.g.) Figure 3 As shown in Phase 2, a pixel is selected as not being an artifact, and the DoD result in all reference DEMs is greater than its corresponding T value to obtain the risk area result. Examples of risk area identification results are provided for comparison. Figure 4 As shown in Figure 7.

[0051] Step 11, Delineation of areas with severe artifacts: (e.g.) Figure 3 As shown in Stage 3, a pixel is selected as an artifact, and the DoD results in all reference DEMs are greater than their corresponding T values, resulting in a severe artifact result. Examples of severe artifact identification results are provided for comparison. Figure 4 As shown in Figure 5.

[0052] Step 12, Delineation of mild artifact areas: (e.g.) Figure 3 As shown in Stage 3, a pixel is selected as an artifact, and the DoD results in all reference DEMs are not greater than their corresponding T values, resulting in a mild artifact. Examples of mild artifact identification results are provided for comparison. Figure 4 As shown in Figure 6.

[0053] This invention is mainly applicable to the comprehensive artifact recognition scenario of DEM. It can accurately, comprehensively and efficiently obtain the coverage of heavy artifacts, light artifacts and risk areas within a region, providing a new method for spatial data reliability assessment and an important reference for DEM artifact research.

[0054] It should be noted that the parameters mentioned in this embodiment (such as sampling ratios of 3:1 to 5:1, gradient threshold of 100m, and window size of 10×2) are preferred parameters for the Alps and TDX30 data characteristics. When performing detection on plains, hills, or other types of DEM data, those skilled in the art can adaptively adjust the above thresholds and window sizes according to the actual terrain undulation and data resolution, which does not deviate from the technical principles of this invention.

[0055] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for artifact detection in a digital elevation model (DEM) that coordinates topographic geometry and physical rules, characterized in that, Includes the following steps: The first step is DEM sample collection: DEM artifact samples (positive samples) and natural terrain samples (negative samples) are extracted from the target area and obtained in advance by visual interpretation. The positive samples cover various artifact types in the target area, including at least abnormal elevation rises around mountains, abnormal undulation noise around water bodies, and irregular steep rises and falls in terrain. The negative samples are natural terrain areas, covering all terrain types in the target area, and the negative samples and positive samples are set to a preset sampling ratio. Oversampling is performed on natural terrain with steep slopes and significant undulations to distinguish artifacts from highly undulating terrain features. The second step involves machine learning training and artifact detection of the artifact detection model: Based on the selected multidimensional feature parameters, the CatBoost ensemble algorithm is used for model training, and the trained model is then used to identify artifacts on the target DEM. The multidimensional feature parameters consist of terrain feature parameters, image filter parameters, and optional DEM auxiliary layer parameters. The terrain feature parameters include: slope, average normal vector angle difference (ANVAD), spherical normal vector standard deviation (SSDoN), elevation standard deviation (SD), slope standard deviation (SDoS), maximum downhill elevation change (MaxDsEC), maximum uphill elevation change (MaxUsEC), terrain roughness, terrain roughness index (TRI), terrain position index (TPI), and total curvature (TotCuv). The image filter parameters include features extracted based on the Roberts operator, Prewitt operator, Laplacian operator, and difference of Gaussian (DoG) operator. The DEM auxiliary layer parameters are the elevation error map (HEM) derived from the target DEM data. The third step is the detection of lateral stitching artifacts: A lateral edge detection filter is used to filter the DEM and extract lateral gradient data. Pixels with gradient absolute values ​​exceeding a preset height threshold are identified as potential feature points. Subsequently, a sliding window with a preset number of rows and two columns is used to traverse the gradient data row by row with a step size of 1 pixel. If all pixels within the window are potential feature points (i.e., the window is filled with target pixels), then a lateral stitching artifact is determined to exist in that area. Step 4, Vertical stitching artifact detection: The processing logic of step 3 is followed for detection. The difference is that a vertical edge detection filter is used to obtain vertical gradient data, and a sliding window with two rows of preset columns is used to slide and judge the suspected vertical feature points. Step 5, integration of stitching artifact results: perform a union operation on the horizontal stitching artifact detection results and the vertical stitching artifact detection results, and determine the result after the operation as the final stitching artifact detection result; Step 6, Reference DEM preparation: Select at least two high-precision DEMs with the same resolution as the DEM to be detected as reference datasets; wherein, the data source of the reference DEM must be independent of the data source of the DEM to be detected, and it must be obtained based on a heterogeneous observation mechanism, excluding data that have a homogeneous derivation relationship with the DEM to be detected, so as to ensure the accuracy of the differential calculation. Step 7, Geodetic datum harmonization: Harmonize the geodetic datum of all reference DEMs to be consistent with the study DEM; Step 8, Differential DEM (DoD): Using the DEM data source to be inspected, perform a pixel-by-pixel differencing operation on the reference DEM to form a DEM difference map. The differential DEM formula is as follows: Where DoD represents the absolute vertical deviation of a given pixel. and These represent the elevation values ​​of the target DEM and the reference DEM, respectively. Step 9: Calculate the risk zone threshold: Determine the threshold used to define the elevation risk zone of the DEM; The elevation risk zone is defined as: the surrounding area where, although the terrain morphology appears normal, the absolute elevation error has reached the artifact level due to the influence of artifact elevation fluctuations; The risk zone threshold is calculated according to the following formula: Where T represents the threshold for identifying risk areas under a certain reference DEM, and N represents the total number of pixels within the confirmed artifact area (the union of machine learning recognition results and edge detection recognition results). This indicates the specific DoD value of a single pixel within a reference DEM in this region; Step 10, Risk Zone Delineation: If a pixel is not an artifact and its DoD result in all reference DEMs is greater than its corresponding T value, then the pixel is defined as a risk zone. Step 11, Delineation of Severe Artifact Regions: If a pixel is an artifact and the DoD results in all reference DEMs are greater than its corresponding T value, then the pixel is defined as a severe artifact. Step 12, Delineation of Slight Artifact Regions: If a pixel is an artifact, and the DoD results in all reference DEMs are not all greater than its corresponding T value, then the pixel is defined as a slight artifact.

2. The method for detecting artifacts in a digital elevation model that coordinates topographic geometry and physical rules as described in claim 1, characterized in that, In the first and second steps, the CatBoost-based ensemble algorithm described above is used for classification, employing features such as slope, average normal vector angle difference (ANVAD), spherical normal vector standard deviation (SSDoN), elevation standard deviation (SD), slope standard deviation (SDoS), maximum downhill elevation change (MaxDsEC), maximum uphill elevation change (MaxUsEC), terrain roughness, terrain roughness index (TRI), terrain location index (TPI), total curvature (TotCuv), Roberts operator, Prewitt operator, Laplacian operator, difference of Gaussian operator (DoG), and DEM auxiliary layer.

3. The method for detecting artifacts in a digital elevation model that coordinates topographic geometry and physical rules as described in claim 1, characterized in that, In both the third and fourth steps, directional filtering is applied to the DEM, and a combination of fixed threshold and sliding window method is used for splicing artifact recognition.

4. The method for detecting artifacts in a digital elevation model that coordinates topographic geometry and physical rules as described in claim 1, characterized in that, In step eight, the DEM is differentially analyzed using the following formula to detect artifacts and risk areas: Where DoD represents the absolute vertical deviation of a given pixel. and These represent the elevation values ​​of the target DEM and the reference DEM, respectively.

5. The method for detecting artifacts in a digital elevation model that coordinates topographic geometry and physical rules as described in claim 1, characterized in that, In step nine, the risk zone determination threshold is calculated using the following formula: Where T represents the threshold for identifying risk areas under a certain reference DEM, and N represents the total number of pixels within the confirmed artifact area (the union of machine learning recognition results and edge detection recognition results). This indicates the specific DoD value of a single pixel under a reference DEM within this region.

6. The method for detecting artifacts in a digital elevation model that coordinates topographic geometry and physical rules as described in claim 1, characterized in that, In step 10, the principle of "if a pixel is not an artifact, and the DoD result in all reference DEMs is greater than its corresponding T value" is used to identify the DEM artifact risk area.

7. The method for detecting artifacts in a digital elevation model that coordinates topographic geometry and physical rules as described in claim 1, characterized in that, In step eleven, severe artifacts are identified using the principle that "if a pixel is an artifact, and the DoD results in all reference DEMs are greater than its corresponding T value".

8. The method for detecting artifacts in a digital elevation model that coordinates topographic geometry and physical rules as described in claim 1, characterized in that, In step 12, mild artifacts are identified using the principle that "if a pixel is an artifact, and the DoD results in all reference DEMs are not all greater than their corresponding T values".

Images