Intelligent fault-tolerant and self-repairing control method and system for DEM production process

Abstract

The application discloses an intelligent fault-tolerant and self-repairing control method and system for a DEM production process, relates to the technical field of self-repairing control, and is used for solving the problem that DEM data may be affected by multi-temporal spectral differences, microwave and laser backscattering differences, leading to misjudgment of water body regions and further affecting the accuracy of terrain data; the accuracy of water body identification is improved, and misjudgment of a single data source is reduced by combining multi-source data such as spectrum, microwave backscattering and laser intensity; through multi-dimensional consistency verification and self-learning of a Bayesian inference self-learning model, the error threshold can be flexibly adjusted, high-confidence regions can be automatically corrected, the system can be continuously optimized and adjusted according to quality inspection feedback, and through accurate identification and marking of abnormal regions in DEM data, comprehensive monitoring and correction of height mutations are performed by using connected domain analysis, height anomalies caused by terrain mutations are reduced, and the accuracy and reliability of the final model are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of self-repairing control, more particularly, to an intelligent fault-tolerant and self-repairing control method and system for a DEM production process. BACKGROUND

[0002] The existing digital elevation model generation process is widely applied in many fields, such as urban planning, environmental monitoring, natural disaster warning, etc. The traditional DEM generation method relies on a single remote sensing data source, usually using optical images or laser radar data, to generate terrain information through point cloud or image analysis.

[0003] The prior art has the following disadvantages:

[0004] Due to different resolutions, accuracies and information coverage of different remote sensing data sources, the generated DEM data may be affected by multi-temporal spectral differences, microwave and laser backscatter differences, resulting in misjudgment of water body areas, such as parking lots and roofs being misjudged as lakes and being flattened, thereby affecting the accuracy of the terrain data. The traditional method fails to fully consider the differences between water bodies and non-water bodies, and has imperfect places in the processing of height mutations.

[0005] In view of the above problems, the present application provides a solution. SUMMARY

[0006] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide an intelligent fault-tolerant and self-repairing control method and system for a DEM production process to solve the problems raised in the background art.

[0007] To achieve the above object, the present application provides the following technical scheme:

[0008] The intelligent fault-tolerant and self-repairing control method for the DEM production process comprises the following steps:

[0009] Step S1, using LiDAR to obtain DSM, combining cloth simulation filtering and RandLA-Net to extract the elevation value and semantic representation of the DSM pixel, labeling the ground mark and weak water mark, and generating an initial DEM;

[0010] Step S2, for the weak water mark area of the initial DEM, using spectral index, microwave backscatter and laser intensity to perform three-domain consistency verification, identifying high-confidence water body area and performing elevation flattening processing;

[0011] Step S3, monitoring the conflict area of the three-domain consistency verification result, marking it as a conflict tile area, self-learning and adjusting the threshold value based on historical data through a Bayesian inference self-learning model, locally rescheduling the conflict tile area and replacing the corresponding DEM data to generate a revised DEM;

[0012] Step S4, calculate the slope grid, curvature grid and neighborhood height difference grid of various terrains in the revised DEM, identify the height mutation area through threshold segmentation, feed back to the Bayesian inference self-learning model of step S3 based on the artificial review result, and output the final DEM.

[0013] In a preferred embodiment, step S1 includes the following contents:

[0014] Obtain the latitude and longitude information point cloud data of LiDAR scanning from the target area to generate a smooth and continuous DSM grid map;

[0015] Obtain the pixel height measured by LiDAR, and calculate the time taken by the laser sensor to emit a pulse back and forth multiplied by the speed of light and then divided by two, which is recorded as the slant range from the ground to the sensor. The height is subtracted from the component of the slant range in the vertical direction to obtain the elevation value of the pixel;

[0016] Divide the ground and non-ground point cloud data by using cloth simulation filtering algorithm, set the maximum distance parameter to control the grid accuracy, and set the cloth hardness coefficient parameter to adapt to the terrain, and generate a cloth simulation DEM;

[0017] Use RandLA-Net for point cloud semantic segmentation to output ground, vegetation, building and weak water body identification;

[0018] Obtain an initial DEM in which each pixel has corresponding semantic identification and corresponding elevation value.

[0019] In a preferred embodiment, step S2 includes the following contents:

[0020] Take the weak water body area labeled in the initial DEM as the detection object, and perform absolute positioning and shape matching of the multi-source external data including multispectral remote sensing image, platform radar image and laser intensity map formed by LiDAR with the initial DEM, and calculate the corresponding three characteristic index values including the difference water body index NDWI of the multispectral remote sensing image, the backscattering intensity of the platform radar image and the laser reflection intensity of the laser intensity map;

[0021] The difference water body index NDWI of the multispectral remote sensing image is expressed as: ; wherein Green is the reflection value of the green light band, and NIR is the reflection value of the near-infrared band;

[0022] The backscattering intensity is expressed as the ratio of the received reflection signal intensity to the transmitted signal intensity;

[0023] The laser reflection intensity is expressed as the ratio of the received reflection laser signal intensity to the transmitted laser signal intensity;

[0024] Three-domain consistency verification is performed on weak water body areas. Corresponding thresholds are set for the parameters of differential water body index, backscattering intensity and laser reflection intensity, which are represented as spectral threshold, microwave threshold and laser threshold. A water body area with high confidence is identified only when the differential water body index is greater than the spectral threshold, the backscattering intensity is less than the microwave threshold and the laser reflection intensity is less than the laser threshold.

[0025] For all high-confidence water bodies, the elevation values ​​are flattened, and the elevation values ​​of the contact pixels between the boundary of the water body and the shoreline are calculated. The average elevation value of the boundary pixels is selected as the estimated water level, and this value is assigned to the elevation values ​​of all pixels in the area.

[0026] In a preferred embodiment, step S3 includes the following:

[0027] For areas in the weak water body identification area where the three-domain consistency verification fails, a conflict tile area label will be added;

[0028] Historical confirmation data of water areas in conflict tile regions are obtained, conflict tile region samples are extracted and three feature index values ​​of each extracted conflict tile region pixel are extracted to form a conflict training set.

[0029] Using the aforementioned conflict training set as input, a Bayesian inference self-learning model is established. The spectral threshold, microwave threshold, and laser threshold are used as random variables. Based on the statistical analysis of conflict tile area samples, the distribution of corresponding feature values ​​in water and non-water bodies is obtained. The posterior probability distribution of the threshold in the historical confirmed data of the water body area in the conflict tile area is calculated using the Bayesian formula. The corresponding threshold with the highest posterior probability is selected as the updated threshold.

[0030] The spectral threshold, microwave threshold, and laser threshold are updated, the three-domain consistency verification is called again, the weak water body area is re-identified with the updated threshold, and the new high-confidence water body area is leveled after identification.

[0031] The DEM data of the high-confidence water body area and the conflict tile area were replaced in the initial DEM and merged into a revised DEM.

[0032] In a preferred embodiment, step S4 includes the following:

[0033] Calculate the degree raster, curvature raster, and neighborhood elevation difference raster indices for the revised DEM;

[0034] The slope raster calculates the degree of rapid change in the horizontal and vertical directions from the elevation values ​​of the DEM, the curvature raster calculates it by the quadratic deviation between the elevation value of each cell and the values ​​of the neighboring cells, and the neighborhood elevation difference raster calculates the difference between the height of the center cell and the average or nearest neighbor value of the nearest neighbor.

[0035] A threshold segmentation method is set up. For slope rasters, if the value of each cell exceeds the preset upper limit threshold of slope, it is considered that there is an unnatural drop. For curvature rasters, if the curvature value exceeds the range between the preset upper limit threshold of curvature and the preset lower limit threshold, it is considered that there is a cell with a rapid change in width. For neighborhood elevation difference rasters, if a cell exceeds the preset upper limit threshold of neighborhood elevation difference, it is marked as a very abnormal high or low point.

[0036] Mark all the above-mentioned abnormal pixels as regions of elevation abrupt change;

[0037] For regions with abrupt elevation changes, feedback is provided for manual verification. Based on the results of the manual verification, feedback is fed back to the Bayesian inference self-learning model, and its corresponding parameters are readjusted to iterate and obtain the final DEM.

[0038] The intelligent fault-tolerant and self-healing control system for the DEM production process includes: an initial DEM construction module, a high-confidence water body module, a conflict tile monitoring module, a verification feedback module, and signal connections between the modules.

[0039] Initial DEM construction module: Use LiDAR to acquire DSM, combine cloth simulation filtering and RandLA-Net to extract the elevation values ​​and semantic representations of DSM pixels, label ground markers and weak water body markers, and generate initial DEM;

[0040] High-confidence water body module: For the weak water body identification areas in the initial DEM, the three-domain consistency verification is performed using spectral index, microwave backscattering and laser intensity to identify high-confidence water body areas and perform elevation leveling processing;

[0041] Conflict Tile Monitoring Module: Monitors conflict areas in the three-domain consistency verification results, marks them as conflict tile areas, adjusts the threshold based on historical data using a Bayesian inference self-learning model, locally reschedules conflict tile areas and replaces the corresponding DEM data, and generates a revised DEM;

[0042] Verification Feedback Module: Calculates slope raster, curvature raster, and neighborhood elevation difference raster for various terrain features in the revised DEM, identifies areas of abrupt elevation changes through threshold segmentation, and feeds back the manual verification results to the Bayesian inference self-learning model in step S3 to output the final DEM.

[0043] The technical effects and advantages of the intelligent fault-tolerant and self-repairing control method for the DEM production process of this invention are as follows:

[0044] By combining multi-source data such as spectral data, microwave backscattering data, and laser intensity data, the accuracy of water body identification is improved, misjudgments from single data sources are reduced, and the precision of water body elevation is enhanced. Through multi-dimensional consistency verification and self-learning of the Bayesian inference model, the error threshold can be flexibly adjusted, and high-confidence areas can be automatically corrected. Furthermore, the system can continuously optimize and adjust based on quality inspection feedback, enhancing the intelligence and automation of the algorithm. By accurately identifying and marking abnormal areas in DEM data and using connected component analysis to comprehensively monitor and correct elevation abrupt changes, elevation anomalies caused by terrain changes are reduced, ensuring the accuracy and reliability of the final model. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the intelligent fault-tolerant and self-repairing control method for the DEM production process of the present invention.

[0046] Figure 2 This is a schematic diagram of the intelligent fault-tolerant and self-repairing control system module for the DEM production process of this invention. Detailed Implementation

[0047] 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.

[0048] Example 1

[0049] Please see Figure 1 As shown, this invention discloses an intelligent fault-tolerant and self-repairing control method for DEM production processes, including the following steps:

[0050] Step S1: Use LiDAR to acquire DSM, and use a combination of cloth simulation filtering and RandLA-Net to extract the elevation values ​​and semantic representations of each pixel in the DSM. In particular, mark the ground and water bodies to make ground labels and weak water body labels, and establish the initial DEM.

[0051] Step S2: For the weak water body identification area in the initial DEM, the three-domain consistency verification is performed using spectral index, microwave backscattering and laser intensity information to identify reliable water bodies and determine high-confidence water body areas. Elevation leveling is then performed on the high-confidence water body areas, and a single elevation is assigned to each high-confidence water body area.

[0052] Step S3: Monitor conflict areas in the three-domain consistency verification results, label conflict tile areas, and adjust the threshold through Bayesian inference self-learning model based on historical confirmation data feedback. Perform local return to the previous step scheduling for conflict tile areas, replace only the DEM data of conflict tile areas, and update the initial DEM to the revised DEM.

[0053] Step S4: Calculate and revise various terrain indicator grids in the DEM, including slope grid, curvature grid and neighborhood elevation difference grid. Perform threshold segmentation based on the corresponding preset upper and lower thresholds. Identify elevation change areas based on the threshold segmentation results. Feedback is provided for manual review. After completing the judgment and correction of elevation changes, the results are written back to the self-learning model in step three for adjustment to obtain the final DEM.

[0054] In step S1, a DSM is acquired using LiDAR. A combination of cloth simulation filtering and RandLA-Net is employed to extract the elevation values ​​and semantic representations of each pixel in the DSM. Ground and water bodies are specifically labeled with ground markers and weak water body markers to establish an initial DEM. Specific details include:

[0055] Acquire latitude and longitude point cloud data from the target area using LiDAR scanning, and generate a smooth and continuous DSM raster map by performing grid transformation based on the shift of the three-dimensional Earth coordinate system and neighborhood moving average interpolation based on the triangular irregular network.

[0056] The pixel height measured by LiDAR is obtained, and the time taken for the laser sensor to emit a pulse and travel to and from the sensor is multiplied by the speed of light and divided by two. This is recorded as the slant distance from the ground to the sensor. The difference between the height and the vertical component of the slant distance is used to obtain the elevation value of the pixel.

[0057] A cloth simulation filtering algorithm is used to distinguish between ground and non-ground points in the point cloud data. A physical simulation method is used to invert the point cloud data and place a virtual cloth at the highest point of the inverted point cloud data surface. The cloth is allowed to fall freely under the action of gravity. As the cloth is subjected to gravity and intersects with the terrain of the point cloud data, its position deforms and gradually adapts to fit the ground. The cloth stays at the highest point of the point cloud data, simulating the undulation of the terrain.

[0058] The CSF algorithm's settings include maximum distance and fabric stiffness coefficient. The former controls the accuracy and smoothness of the fabric grid, while the latter controls the internal strength of the fabric, i.e., the softness and stiffness of the fabric. Stiff fabric is usually used in gentle plains and soft fabric is used in steep mountains to make the fabric better cover the terrain.

[0059] It should be noted that the maximum distance of the CSF can be set to 2-5 times the average point spacing. With a typical LiDAR density of 4 pts / m², the maximum distance is approximately 0.8–2 m, which is used to control the smoothing scale of the grid. The fabric stiffness coefficient of the soft fabric in mountainous areas is taken as 20 N / m, and that of the hard fabric in plains is taken as 100 N / m, to ensure that the fabric can fit tightly against the undulating terrain without excessively penetrating the steep cliff surface. The above value range is based on the statistical results of playback experiments of typical landform samples, and the optimal value can be determined by those skilled in the art within the corresponding range.

[0060] CSF calculations can divide point cloud data into two categories: ground point cloud data and non-ground point cloud data. By interpolating the ground point cloud data, a cloth simulation ground elevation model, i.e., a cloth simulation DEM, can be generated.

[0061] The RandLA-Net point cloud data semantic segmentation algorithm is used to perform more refined category segmentation on the point cloud data of the fabric simulation DEM.

[0062] RandLA-Net is an efficient semantic segmentation network for large-scale point cloud data. It accelerates computation through random downscaling and local feature aggregation, and can efficiently perform semantic prediction and classification for each pixel.

[0063] Pre-trained on a large-scale dataset, the network can output a semantic label for each new pixel. The basic labels include ground, vegetation, buildings, and weak water bodies. Weak water bodies are labeled as possible water surface areas. Since the echo intensity of LiDAR is often very low in water surface areas, RandLA-Net can calculate the specific features of the pixel cloud data in water surface areas and classify and label them as weak water bodies to obtain the initial DEM. Each pixel has a corresponding semantic label and a corresponding elevation value.

[0064] In step S2, for the weak water body identification areas in the initial DEM, reliable water bodies are identified by three-domain consistency verification using spectral index, microwave backscattering, and laser intensity information. High-confidence water body areas are then determined, and elevation leveling is performed. A single elevation is assigned to each high-confidence water body area. Specific details include:

[0065] Using the weak water areas marked in the initial DEM as the detection targets, the external data from multiple sources were absolutely located and morphologically matched with the initial DEM to ensure that all data were aligned under the same latitude and longitude reference. The unified data included: multispectral remote sensing imagery, platform radar imagery, and laser intensity maps generated by LiDAR;

[0066] For each pixel within a weak water body region, three characteristic index values ​​are calculated: the Differential Water Index (NDWI) from the multispectral remote sensing image; the backscattering intensity from the platform radar image; and the laser reflection intensity from the laser intensity map.

[0067] The differential water index NDWI of multispectral remote sensing images is represented as: Among them, Green is the reflectance value of the green light band, NIR is the reflectance value of the near-infrared band, and NDWI utilizes the characteristic that water bodies have strong absorption and almost no reflection in the near-infrared band, while vegetation has a very high reflectance. The NDWI value is usually a positive number in water bodies, and an extremely high value indicates a large area of ​​water.

[0068] The backscatter intensity of a platform radar image is the signal strength reflected back to the radar receiver after the electromagnetic waves emitted by the radar system interact with the ground or objects. In calculation, the backscatter intensity is expressed as the ratio of the received reflected signal intensity to the emitted signal intensity. Specifically, this ratio reflects the reflectivity of ground objects to radar waves and is affected by factors such as the material, shape, and surface characteristics of the ground objects.

[0069] The laser reflection intensity in the laser intensity map represents the signal intensity reflected back to the lidar receiver after a laser pulse hits the ground or an object. In calculation, the laser reflection intensity is the ratio of the received reflected laser signal intensity to the emitted laser signal intensity, reflecting the reflection characteristics of the ground surface to the laser. It is affected by factors such as the material, roughness, and reflection angle of the object's surface.

[0070] Since airborne lidar typically uses near-infrared lasers, the water surface reflects this wavelength very weakly, resulting in a lack of echo points in the water surface area in LiDAR data. This makes the water area appear extremely dark or without echoes in LiDAR intensity images.

[0071] Three-domain consistency verification is performed on the weak water body area. The corresponding thresholds for the above three parameters are set as spectral threshold, microwave threshold and laser threshold. The water body area with high confidence is identified only when the difference water body index is greater than the spectral threshold, the backscattering intensity is less than the microwave threshold and the laser reflection intensity is less than the laser threshold.

[0072] It should be noted that the preset value for the spectral threshold can be 0.20, with an allowable adjustment range of 0.15–0.30; the preset value for the microwave threshold can be -15dB, which is typical for C-band side-looking radar, with an allowable adjustment range of -18dB to -10dB; and the preset value for the laser threshold can be 30DN, with an allowable adjustment range of 20–40DN. These can be adjusted by experts in the field based on their experience.

[0073] For all water bodies identified as having high confidence, the DEM elevation values ​​of the area are flattened, the contact pixel elevation values ​​between the boundary of the water body area and the coastal ground plane are calculated, the average elevation value of the boundary pixels is selected as the estimated water level, and this value is assigned to the elevation values ​​of all pixels in the area, and their elevations are flattened.

[0074] If necessary, the lowest elevation value in the region can also be used as the water level benchmark. By ensuring the water surface area is consistently leveled, the water surface elevation layer in the final DEM will be smooth and consistent, which also conforms to the characteristics of the actual topography of the water body.

[0075] In step S3, conflict areas in the three-domain consistency verification results are monitored, labeled as conflict tile areas, and thresholds are adjusted through a Bayesian inference self-learning model based on historical confirmation data feedback. The conflict tile areas are then locally returned to the previous step's scheduling, replacing only the DEM data of the conflict tile areas and updating the initial DEM to a revised DEM. Specific details include:

[0076] For areas in the weak water body identification area where the three-domain consistency verification failed, the elevation leveling process will not be carried out for the time being. Instead, the conflict tile area label will be applied to the area, and the entire DEM area will be divided into several rectangular tiles. Each tile will be assigned a unique number as a tracking identifier.

[0077] Conflicting tile areas may include several different situations: First, some areas actually contain water, but one of the three pieces of information is not fully supported; second, some areas are not actually water, but optical or LiDAR information is offset, causing them to be incorrectly marked as water. In this case, the threshold for identifying water needs to be automatically readjusted and the conflicting tile areas need to be locally rewritten.

[0078] If the conflict tile area is confirmed to actually contain water, the previous discrimination criteria may be too strict, and the corresponding threshold needs to be lowered; conversely, if the conflict tile area is not actually water but is marked as water, the previous discrimination criteria are too lenient, and the threshold needs to be increased to reduce false detections. A sample comparison model based on Bayesian statistics was constructed to automatically optimize the new threshold setting.

[0079] Historical confirmation data of water areas in conflict tile regions are obtained, and conflict tile region samples are extracted for judgment. Three feature index values ​​of each extracted conflict tile region pixel are extracted to form a conflict training set for the next step of threshold update.

[0080] Using the aforementioned conflict training set as input, a Bayesian inference self-learning model is established. The spectral threshold, microwave threshold, and laser threshold are treated as random variables with uncertainty. At the same time, the distribution of corresponding feature values ​​in water and non-water bodies is obtained based on the statistical analysis of conflict tile area samples. The posterior probability distribution of the threshold in the historical confirmed data of the water body area in the conflict tile area is calculated using the Bayesian formula, and the corresponding threshold with the maximum posterior probability is selected as the updated threshold.

[0081] It should be noted that the mathematical expression for maximum a posteriori estimation is: ;in, It is an ordered triplet composed of spectral threshold, microwave threshold, and laser threshold, that is, the overall representation of the three discrimination thresholds; D is the conflict training set. For based on The likelihood function for the training dataset D, The threshold prior probability;

[0082] When the spectral threshold, microwave threshold, and laser threshold are independent within the same category and all follow a Gaussian distribution, the likelihood function of the water body can be expressed as follows: The likelihood functions for non-aqueous bodies can be expressed as follows: ;in, and These represent the mean and variance of this feature for water and non-water samples, respectively. The result is obtained by maximizing the aforementioned posterior probability. And update the three discrimination thresholds accordingly;

[0083] Based on the calculation results of the previous step, the spectral threshold, microwave threshold, and laser threshold are updated. For the spectral threshold, the new value may be smaller than the initial threshold to cover some real water bodies with slightly lower NDWI. Conversely, if there may be cases where the NDWI is high but it is not actually water, the NDWI threshold may need to be increased to resolve this. Similarly, the microwave threshold may be adjusted upward or downward depending on whether there are missed water body signals. The laser threshold may also be adjusted according to the weakness of the point cloud of nearby water bodies, resulting in a new set of well-structured thresholds.

[0084] After setting a new threshold, a local execution method is adopted. For each area marked as a conflict tile, the three-domain consistency verification is re-invoked, the weak water body area is re-identified with the updated threshold, and the new high-confidence water body area is subjected to elevation leveling.

[0085] The DEM data of the high-confidence water body area and the conflict tile area are replaced in the initial DEM. The replacement method can be the overall replacement of the tile elevation. The DEM data has maintained consistency in the three characteristic index values. Except for a few special cases, they are merged into the revised DEM.

[0086] Most water surface areas have been flattened, and incorrect water body labels have been corrected. The quality of DEM data has been significantly improved. Since only conflict tiles are reconstructed locally, this self-repair process is more efficient than global recalculation. If new conflict tile areas still appear, the system can repeat the above steps to continuously learn and correct the threshold, and finally obtain the optimal DEM dataset.

[0087] In step S4, various terrain indicator grids in the revised DEM are calculated, including slope grids, curvature grids, and neighborhood elevation difference grids. Threshold segmentation is performed based on corresponding preset upper and lower thresholds. Areas with abrupt elevation changes are identified based on the threshold segmentation results. Feedback is provided for manual review. After determining and correcting elevation changes, the results are written back to the self-learning model in step three for adjustment, resulting in the final DEM. Specific details include:

[0088] The revised DEM is used to calculate various topographic indicators and generate corresponding indicator raster maps, including: slope raster, curvature raster and neighborhood elevation difference raster.

[0089] The slope raster calculates the degree of rapid change in the horizontal and vertical directions from the elevation values ​​of the DEM. The curvature raster calculates the curvature by the quadratic deviation between the elevation value of each cell and the values ​​of the neighboring cells. The neighborhood elevation difference raster calculates the difference between the height of the center cell and the average or nearest neighbor value. These three indicators can reflect the smoothness of the DEM terrain from different perspectives and are used to detect unnatural abrupt changes in pixels.

[0090] To detect areas with common anomalies, upper and / or lower thresholds were set for each indicator. For slope rasters, if a cell's value exceeds the upper slope threshold, it is considered to have an unnatural steepness or abrupt drop. For curvature rasters, if a cell has an impactful curvature value that exceeds the range of the upper and lower curvature thresholds, it is considered to have a cell with a rapidly changing width. For neighborhood elevation difference rasters, if a cell exceeds the upper neighborhood elevation difference threshold, it is marked as a highly anomalous high or low point.

[0091] It should be noted that the upper limit threshold for slope can be set to 35°, which conforms to the recommended upper limit of 30–40° for grid DEMs; the upper limit threshold for curvature can be set to ±0.08 to ±0.15. The lower threshold range for capturing true elevation changes can be set to: ±0.03~±0.05. It can filter out fine noise; the upper limit threshold for neighborhood elevation difference can be set to 2m, i.e. 5× grid resolution. In case of extreme terrain, the threshold can be relaxed or tightened by ±20%, and the specific setting should be made by a person skilled in the art according to the actual situation.

[0092] Based on the threshold segmentation method described above, all potentially abnormal pixels are marked, and then a connected component algorithm is used for clustering to aggregate adjacent abnormal pixels into continuous regions. The boundary of each continuous region is drawn, and the region is marked as an elevation change region, indicating that there is an unreasonable elevation change in the region.

[0093] For areas with abrupt elevation changes, feedback is provided for manual verification. The results of the manual verification are then fed back to the Bayesian inference self-learning model to adjust its corresponding parameters. By analyzing and monitoring the connection feedback information, the final DEM is obtained.

[0094] Example 2

[0095] Please see Figure 2 As shown, this invention discloses an intelligent fault-tolerant and self-healing control system for the DEM production process, including: an initial DEM construction module, a high-confidence water body module, a conflict tile monitoring module, and a verification feedback module, with signal connections between each module;

[0096] Initial DEM construction module: LiDAR is used to acquire the DSM, and a combination of cloth simulation filtering and RandLA-Net is used to extract the elevation values ​​and semantic representations of each pixel in the DSM. Ground and water bodies are specially marked with ground labels and weak water body labels to build the initial DEM.

[0097] High-confidence water body module: For the weak water body identification areas in the initial DEM, the three-domain consistency verification is performed using spectral index, microwave backscattering and laser intensity information to identify reliable water bodies and determine high-confidence water body areas. Elevation leveling is then performed on the high-confidence water body areas, and a single elevation is assigned to each high-confidence water body area.

[0098] Conflict Tile Monitoring Module: Monitors conflict areas in the three-domain consistency verification results, labels conflict tile areas, and adjusts thresholds through a Bayesian inference self-learning model based on historical confirmation data feedback. It performs local return to the previous step scheduling for conflict tile areas, only replacing the DEM data of the conflict tile areas and updating the initial DEM to the revised DEM.

[0099] Verification and feedback module: Calculates and revises various terrain indicator grids in the DEM, including slope grid, curvature grid and neighborhood elevation difference grid, performs threshold segmentation based on the corresponding preset upper and lower thresholds, identifies areas of elevation change based on the threshold segmentation results, provides feedback for manual verification, completes the judgment and correction of elevation changes, and writes the results back to the self-learning model in step three for adjustment to obtain the final DEM.

[0100] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0101] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0102] Those skilled in the art will recognize that the modules and algorithm 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 inventive 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 implementation should not be considered beyond the scope of this application.

[0103] In addition, the functional modules in the various embodiments of this application 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.

[0104] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0105] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent fault-tolerant and self-repairing control method for DEM production processes, characterized in that, Including the following steps: Step S1: Use LiDAR to acquire DSM, combine cloth simulation filtering and RandLA-Net to extract the elevation values ​​and semantic representations of DSM pixels, label ground markers and weak water body markers, and generate the initial DEM; Step S2: For the weak water body identification areas in the initial DEM, calculate the three characteristic index values: Differential Water Index (NDWI), microwave backscattering intensity, and laser reflection intensity; perform three-domain consistency verification on the weak water body identification areas, and set spectral thresholds, microwave thresholds, and laser thresholds. If and only if the Differential Water Index is greater than the spectral threshold, the microwave backscattering intensity is less than the microwave threshold, and the laser reflection intensity is less than the laser threshold, a high-confidence water body area is identified and elevation leveling is performed. Step S3: Monitor conflict areas in the three-domain consistency verification results, mark them as conflict tile areas, adjust the threshold based on historical data using a Bayesian inference self-learning model, locally reschedule conflict tile areas and replace the corresponding DEM data, and generate a revised DEM; Step S4: Calculate the slope raster, curvature raster, and neighborhood elevation difference raster for various terrains in the revised DEM. Identify areas of abrupt elevation changes through threshold segmentation. Feedback the results of manual verification to the Bayesian inference self-learning model in step S3. With the conflict training set as input, output the updated spectral threshold, microwave threshold, and laser threshold.

2. The intelligent fault-tolerant and self-repairing control method for the DEM production process according to claim 1, characterized in that: Acquire latitude and longitude point cloud data from the target area using LiDAR scanning to generate a smooth and continuous DSM raster map; The pixel height measured by LiDAR is obtained, and the time taken for the laser sensor to emit a pulse and travel to and from the sensor is multiplied by the speed of light and divided by two. This is recorded as the slant distance from the ground to the sensor. The difference between the height and the vertical component of the slant distance is used to obtain the elevation value of the pixel. A cloth simulation filtering algorithm is used to segment ground and non-ground point cloud data. The maximum distance parameter is set to control the grid accuracy, and the cloth stiffness coefficient parameter is adapted to the terrain to generate a cloth simulation DEM. RandLA-Net was used to perform semantic segmentation of the point cloud on the fabric simulation DEM, and the ground, vegetation, building and weak water body labels were output. The initial DEM containing the corresponding semantic identifier and the corresponding elevation value is obtained.

3. The intelligent fault-tolerant and self-repairing control method for the DEM production process according to claim 1, characterized in that: Using the weak water body areas marked in the initial DEM as the detection objects, the multi-source external data, including multispectral remote sensing images, platform radar images, and laser intensity maps generated by LiDAR, are absolutely located and morphologically matched with the initial DEM. The corresponding three feature index values ​​are calculated, including the differential water index NDWI of the multispectral remote sensing images; the backscattering intensity of the platform radar images; and the laser reflection intensity of the laser intensity map. The differential water index NDWI of multispectral remote sensing images is represented as: Where Green is the reflectance value in the green light band, and NIR is the reflectance value in the near-infrared band; Backscattering intensity is expressed as the ratio of the intensity of the received reflected signal to the intensity of the transmitted signal; Laser reflection intensity is expressed as the ratio of the intensity of the received reflected laser signal to the intensity of the emitted laser signal.

4. The intelligent fault-tolerant and self-repairing control method for the DEM production process according to claim 3, characterized in that: Three-domain consistency verification is performed on weak water body areas. Corresponding thresholds are set for the parameters of differential water body index, backscattering intensity and laser reflection intensity, which are represented as spectral threshold, microwave threshold and laser threshold. A water body area with high confidence is identified only when the differential water body index is greater than the spectral threshold, the backscattering intensity is less than the microwave threshold and the laser reflection intensity is less than the laser threshold. For all high-confidence water bodies, the elevation values ​​are flattened, and the elevation values ​​of the contact pixels between the boundary of the water body and the shoreline are calculated. The average elevation value of the boundary pixels is selected as the estimated water level, and this value is assigned to the elevation values ​​of all pixels in the area.

5. The intelligent fault-tolerant and self-repairing control method for the DEM production process according to claim 1, characterized in that: For areas in the weak water body identification area where the three-domain consistency verification fails, a conflict tile area label will be added; Historical confirmation data of water areas in conflict tile regions are obtained. Stratified random sampling of conflict tile regions is performed based on confidence scores. Three feature index values ​​of each sampled conflict tile region pixel are determined and extracted to form a conflict training set. Using the aforementioned conflict training set as input, a Bayesian inference self-learning model is established. The spectral threshold, microwave threshold, and laser threshold are used as random variables. Based on the statistical analysis of conflict tile area samples, the distribution of corresponding feature values ​​in water and non-water bodies is obtained. The posterior probability distribution of the threshold in the historical confirmed data of the water body area in the conflict tile area is calculated using the Bayesian formula. The corresponding threshold with the highest posterior probability is selected as the updated threshold.

6. The intelligent fault-tolerant and self-repairing control method for the DEM production process according to claim 5, characterized in that: The spectral threshold, microwave threshold, and laser threshold are updated, the three-domain consistency verification is called again, the weak water body area is re-identified with the updated threshold, and the new high-confidence water body area is leveled after identification. The DEM data of the high-confidence water body area and the conflict tile area were replaced in the initial DEM and merged into a revised DEM.

7. The intelligent fault-tolerant and self-repairing control method for the DEM production process according to claim 1, characterized in that: The revised DEM is used to calculate the slope raster, curvature raster, and neighborhood elevation difference raster indices. The slope raster is calculated from the elevation values ​​of the revised DEM to determine the degree of rapid change in the horizontal and vertical directions. The curvature raster is calculated by the quadratic deviation between the elevation value of each cell and the values ​​of the neighboring cells. The neighborhood elevation difference raster is calculated by the average height of the center cell and the nearest neighbor.

8. The intelligent fault-tolerant and self-repairing control method for the DEM production process according to claim 7, characterized in that: A threshold segmentation method is set up. For slope rasters, if the slope exceeds the preset upper limit threshold, it is considered that there is an unnatural drop. For curvature rasters, if the curvature value exceeds the range between the preset upper limit threshold and the preset lower limit threshold, it is considered that there is a cell with a rapidly changing width. For neighborhood elevation difference rasters, if the elevation difference exceeds the preset upper limit threshold, it is marked as a very abnormal high or low point.

9. The intelligent fault-tolerant and self-repairing control method for the DEM production process according to claim 7, characterized in that: All pixels exhibiting anomalies in the threshold segmentation method are marked as elevation change regions; For regions with abrupt elevation changes, feedback is provided for manual verification. Based on the results of the manual verification, feedback is fed back to the Bayesian inference self-learning model, and its corresponding parameters are readjusted to iterate and obtain the final DEM.

10. An intelligent fault-tolerant and self-repairing control system for a DEM production process, used to implement the intelligent fault-tolerant and self-repairing control method for the DEM production process as described in any one of claims 1-9, characterized in that: The initial DEM construction module uses LiDAR to acquire the DSM, combines cloth simulation filtering and RandLA-Net to extract the elevation values ​​and semantic representations of DSM pixels, labels ground markers and weak water body markers, and generates the initial DEM. The high-confidence water body module calculates three characteristic index values—differential water body index, microwave backscattering intensity, and laser reflection intensity—for the weak water body identification areas in the initial DEM. It then performs three-domain consistency verification on the weak water body identification areas, setting spectral thresholds, microwave thresholds, and laser thresholds. A high-confidence water body area is identified and elevation leveling is performed only when the differential water body index is greater than the spectral threshold, the microwave backscattering intensity is less than the microwave threshold, and the laser reflection intensity is less than the laser threshold, all of which are simultaneously satisfied. The conflict tile monitoring module monitors conflict areas in the three-domain consistency verification results, marks them as conflict tile areas, adjusts the threshold based on historical data through a Bayesian inference self-learning model, locally reschedules conflict tile areas and replaces the corresponding DEM data, and generates a revised DEM. The review and feedback module calculates the slope raster, curvature raster, and neighborhood elevation difference raster for various terrains in the revised DEM, identifies areas of abrupt elevation changes through threshold segmentation, and feeds back the results of manual review to the Bayesian inference self-learning model in step S3 to output the final DEM.

Images