Coastal terrain three-dimensional model adaptive construction method, system and device

Abstract

The application discloses a kind of coastal terrain three-dimensional model adaptive construction methods, comprising the following steps: according to initial scanning strategy, target coastal area is scanned, and initial three-dimensional point cloud data is obtained;Real-time tidal level of target coastal area is collected, tidal table is combined, and tidal database is constructed;Initial three-dimensional point cloud data is input into large language model, and the semantic segmentation of initial three-dimensional point cloud data is formed multiple semantic units;The attribute characteristics of each semantic unit are collected, the correlation of semantic unit is analyzed in combination with tidal database, and differentiated scanning strategy is generated;According to differentiated scanning strategy, directional scanning is executed to each semantic unit, and the three-dimensional point cloud data after optimization is obtained;According to the three-dimensional model of three-dimensional point cloud data after optimization of target coastal area is generated.The differentiated scanning strategy of the application can set the best scanning mode for different regions, realize the accurate scanning of target coastal area, and improve the scanning efficiency.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional scanning measurement technology, and in particular to an adaptive construction method, system and equipment for a three-dimensional model of coastal terrain. Background Technology

[0002] Coastal topographic surveying is a fundamental task in coastal engineering planning, design, construction, and operation. Traditional methods—total stations and RTK-GPS ground surveying—offer high accuracy but extremely low efficiency, suitable only for localized surveys of small areas. With the development of unmanned aerial vehicles (UAVs) and 3D scanning technology, airborne depth cameras and LiDAR have gradually become the mainstream methods for coastal topographic surveying: UAVs carrying 3D scanning devices fly over target areas to acquire high-density 3D point cloud data, which is then post-processed to generate digital elevation models (DEMs) and orthophotos.

[0003] However, existing UAV scanning methods generally employ a uniform resolution strategy—covering the entire target area with the same flight altitude, scan density, and flight path spacing. This uniform distribution strategy fails to consider the high heterogeneity of coastal space: reef areas and man-made structures require centimeter-level or even millimeter-level resolution to capture structural details and damage features; flat beaches and vegetated areas only require decimeter-level resolution to meet terrain modeling needs; and water areas do not require 3D modeling at all but need to mark the land-water boundary. The uniform strategy leads to insufficient data in critical areas and redundant data in non-critical areas, resulting in low overall scanning efficiency.

[0004] More importantly, coastal surveys face unique tidal factors—the intertidal zone is only exposed during low tide, with a limited window of time available for scanning. Existing scanning strategies do not take tidal factors into account at all, resulting in insufficient utilization of the low tide window or ineffective scanning of submerged areas during high tide. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology, the purpose of this patent is to provide an adaptive construction method, system and device for a three-dimensional coastal terrain model, which aims to solve the problems that the existing technology cannot adopt differentiated measurement methods for different land cover types and does not consider the influence of tidal factors.

[0006] This patent is achieved using the following technical solution: Firstly, an adaptive method for constructing a 3D model of coastal terrain is provided, including the following steps: Step S1: Perform an initial scan of the target coastal area according to the initial scanning strategy to obtain initial 3D point cloud data; the initial scanning strategy uses a fixed resolution and uniform height to perform a full-coverage scan of the target coastal area; Step S2: Collect real-time tide levels of the target coastal area and construct a tide database by combining them with tide tables; Step S3: Extract various transformation features from the initial 3D point cloud data and map them to the embedding space of the large language model. Perform semantic segmentation on the initial 3D point cloud data to form multiple semantic units. Step S4: Collect the attribute features of each semantic unit, combine them with the tidal database, analyze the correlation between semantic units, and generate a differentiated scanning strategy; Step S5: Perform directional scanning on each semantic unit according to the differentiated scanning strategy to obtain optimized 3D point cloud data; Step S6: Generate a 3D model of the target coastal area based on the optimized 3D point cloud data.

[0007] To optimize the above technical solution, the specific measures also include: Furthermore, multiple transformation features of the initial 3D point cloud data are extracted and mapped to the embedding space of a large language model. Semantic segmentation is then performed on the initial 3D point cloud data to form multiple semantic units, specifically: Extracting multiple transformation features from initial 3D point cloud data; Multiple transformation features are input into the corresponding modality coding branches within the large language model, and semantic segmentation is performed to identify multiple semantic units.

[0008] Furthermore, various transformation features are input into the corresponding modality coding branches within the large language model, and semantic segmentation is performed to identify various semantic units, specifically: The initial 3D point cloud data is projected and converted into a 2D elevation map, which is then input into the visual encoding branch. The initial 3D point cloud data is transformed into a sequence of 2D rendered images according to multiple spatial perspectives and then input into the visual encoding branch. The initial 3D point cloud data is converted into structured text information, and the input text encoding branch is used. Extract the geometric feature vectors from the initial 3D point cloud data and map them to the embedding space of the large language model; The corresponding data are classified and processed using various sub-models. After data fusion analysis and semantic segmentation, the geometric feature vectors mapped to the large language model are combined to identify multiple semantic units.

[0009] Furthermore, the attribute features of each semantic unit are collected, and combined with the tidal database, the relationships between semantic units are analyzed to generate a differentiated scanning strategy, specifically: The semantic unit can be classified into natural features and man-made structures based on its physical attributes. Natural features and man-made structures are ranked according to their importance based on the functional attributes of their semantic units, and a priority sequence list is established. Based on the state attributes of semantic units, the state of each semantic unit is evaluated using standard geometric features, and state anomaly identification is performed to establish a state information table. Based on the tidal level attribute of the semantic units, calculate the exposure time window of each semantic unit within the current tidal cycle, and sort them according to the exposure time window; Based on the sorting of relevant attributes for each semantic unit and the correlation between the relevant attributes, different measurement priorities, scanning modes and scanning resolutions are set to form a differentiated scanning strategy.

[0010] Furthermore, based on the tidal level attribute of the semantic units, the exposure time window of each semantic unit within the current tidal cycle is calculated, and they are sorted according to the exposure time window, specifically as follows: Obtain the elevation information from the physical attributes of the semantic unit, and combine it with the tidal database to calculate the exposure time window of each semantic unit in the current tidal cycle. Before calculating the exposure time window, it is necessary to ensure that the elevation information and the tidal data in the tidal database use the same vertical reference. If different references are used, the elevation reference conversion should be performed first. Based on the exposure time window, semantic units are divided into long-term scannable areas, phased scannable areas, and non-scannable areas. For non-scannable areas (i.e., areas that are always submerged), no active scanning tasks are assigned, and only water level data from the tidal database is used to mark their land-water boundaries. If it is necessary to obtain underwater topographic data, auxiliary detection methods such as multibeam sonar are used to collect it separately. Based on different regional divisions and the exposure time windows of each semantic unit, a scanning time series of each semantic unit is established and sorted.

[0011] Furthermore, based on the order of relevant attributes for each semantic unit and the correlation between these attributes, different measurement priorities, scanning modes, and scanning resolutions are set, specifically as follows: Based on the different attributes of semantic units, the measurement priorities are initially sorted, and the scanning mode and scanning resolution are initially set to generate multiple sets of scanning strategies. Based on the correlation between different attributes of semantic units, the measurement priority sorting, scanning mode and scanning resolution are adjusted a second time; Based on the relationships between semantic units, the measurement priority sorting, scanning mode, and scanning resolution are readjusted to obtain optimized measurement priority sorting, scanning mode, and scanning resolution settings.

[0012] Furthermore, based on different measurement priorities, scanning modes, and scanning resolutions, differentiated scanning strategies are formed, specifically as follows: Based on different measurement priorities and the distribution location of each semantic unit, a scanning path and scanning time window are planned for the target coastal area to form the first scanning strategy; Based on different scanning resolution settings, different scanning modes and scanning resolutions are planned for the target coastal area to form a second scanning strategy; The first and second scanning strategies are combined to form a differentiated scanning strategy.

[0013] Furthermore, after performing directional scanning on each semantic unit according to the differentiated scanning strategy to obtain optimized 3D point cloud data, the process also includes: Determine whether each semantic unit has been scanned completely. If there are semantic units that have not been scanned completely, obtain the uncovered area and perform directional scanning on the uncovered area in the next cycle to form complete 3D point cloud data.

[0014] Secondly, an adaptive construction system for a 3D coastal terrain model is provided, including the following modules: Acquisition Module 1 is used to perform an initial scan of the target coastal area according to the initial scanning strategy to obtain initial 3D point cloud data. The second data acquisition module is used to collect real-time tide levels in the target coastal area and, in conjunction with the tide table, construct a tide database. Processing module one is used to extract various transformation features of the initial 3D point cloud data and map them to the embedding space of the large language model, and to perform semantic segmentation on the initial 3D point cloud data to form multiple semantic units; Processing module two is used to collect the attribute features of each semantic unit, combine them with the tidal database, analyze the correlation between semantic units, and generate differentiated scanning strategies. Processing module three is used to perform directional scanning on each semantic unit according to the differentiated scanning strategy to obtain optimized 3D point cloud data; The modeling module is used to generate a 3D model of the target coastal area based on the optimized 3D point cloud data. The analysis module is used to collect optimized 3D point cloud data and determine whether it is complete, whether all semantic units have been scanned, and send the uncovered areas corresponding to the semantic units that have not been scanned to the processing module 2. The processing module 2 performs directional scanning in the next cycle to form complete 3D point cloud data.

[0015] Thirdly, an apparatus is provided, comprising a processor and a memory, wherein the memory stores a set of program instructions; when the set of program instructions stored in the memory is loaded and executed by the processor, the aforementioned adaptive construction method for a three-dimensional coastal terrain model can be realized.

[0016] The beneficial effects of this invention are: This application presents an adaptive method for constructing 3D coastal topographic models. Based on initially acquired 3D point cloud data, it performs corresponding semantic unit division to identify different terrain features and man-made structures in the target coastal area. Then, it collects and calculates physical, functional, state, and tidal attributes. Compared to traditional adaptive scanning methods that rely solely on geometric features, this approach can identify the differentiated measurement needs of different land cover types. Furthermore, this application considers tidal information, optimizing the temporal scanning strategy by intelligently matching the scanning task of the intertidal zone with the low tide window, maximizing the utilization of the limited available scanning time.

[0017] This application combines spatial regional division with the scannable period of time to generate a differentiated scanning strategy. Finally, based on this differentiated scanning strategy, accurate scanning of the target coastal area can be achieved, realizing the optimal scanning mode for different areas and improving scanning efficiency.

[0018] The differentiated scanning strategy of this application takes into account semantic priority and tidal time constraints, and assigns different resolutions, scanning paths, scanning modes and scanning equipment to different semantic units, thereby improving the overall scanning efficiency while ensuring the data quality of key areas. Attached Figure Description

[0019] Figure 1 This is a flowchart of an adaptive construction method for a three-dimensional coastal terrain model provided in the first embodiment of this application.

[0020] Figure 2 This is a detailed flowchart provided in the first embodiment of this application.

[0021] Figure 3 yes Figure 2 A partial flowchart of the process.

[0022] Figure 4 This is a schematic diagram of the module connections of an adaptive construction system for a three-dimensional coastal terrain model provided in the second embodiment of this application.

[0023] Figure 5 This is a schematic diagram of the structure of an electronic device provided in the third embodiment of this application. Detailed Implementation

[0024] To clarify the technical solution and working principle of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0025] First implementation method: refer to Figures 1-3As shown, this embodiment provides an adaptive construction method for a 3D coastal terrain model, including the following steps: Step S1: Perform an initial scan of the target coastal area according to the initial scan strategy to obtain initial 3D point cloud data; the initial scan strategy uses a fixed resolution and uniform height to perform a full-coverage scan of the target coastal area.

[0026] During the initial scan, the goal is to fully cover the target coastal area. This scan primarily utilizes 3D scanning devices, including depth cameras or LiDAR. These 3D scanning devices can be deployed on drones, which move around the target coastal area to achieve omnidirectional scanning. During the initial scan, there is no need to distinguish between scan resolution and drone flight altitude; the priority is to obtain panoramic 3D data of the target coastal area.

[0027] A 3D scanning device is used to acquire the true 3D geometric information of the surface of the target coastal area, thereby outputting 3D point cloud data containing physical coordinates.

[0028] Step S2: Collect real-time tide levels of the target coastal area and construct a tide database by combining them with tide tables.

[0029] During the initial scan, the acquired point cloud data represents instantaneous sea level data, while sea level height varies continuously with the tides. The tide table is obtained based on astronomical tide forecasts, while real-time tide data incorporates non-astronomical factors such as meteorological rises and falls. The database constructed by combining these two methods can correct systematic biases in the tide table and provide more accurate high-frequency tide level interpolation values ​​for scan moments without measured tide levels. This provides a more accurate tide level correction for subsequent differentiated scanning strategies.

[0030] Step S3: Extract various transformation features from the initial 3D point cloud data and map them to the embedding space of the large language model. Perform semantic segmentation on the initial 3D point cloud data to form multiple semantic units.

[0031] Specifically, the large language model contains multiple modality coding branches, which perform identification and calculation from multiple data dimensions, thereby semantically segmenting the initial 3D point cloud data and ultimately restoring the topographic features of the coastal target area, as well as the corresponding man-made structures.

[0032] When performing initial 3D point cloud data analysis, the first step is to analyze the types of information that can be converted from the initial 3D point cloud data. Then, different information types are extracted to their corresponding modality coding branches, followed by computational inference. Specific modality coding branches include: visual coding branch and text coding branch.

[0033] When performing semantic segmentation, the analysis results of different modality coding branches should be considered comprehensively. This is because each modality coding branch considers a single dimension and needs to take into account factors such as error and the correlation between different dimensions. Therefore, the results inferred from multiple modality coding branches should be comprehensively analyzed to form different semantic units and ensure the accuracy of the inference results.

[0034] S31: Extract various transformation features from the initial 3D point cloud data.

[0035] Before feature transformation, the initial 3D point cloud data needs to be filtered. First, the initial 3D point cloud data is denoised, outlier data points are removed, and downsampling is performed to ensure the accuracy of the initial 3D point cloud data.

[0036] S32: Input the various transformation features into the corresponding modality coding branches within the large language model, perform semantic segmentation, and identify various semantic units.

[0037] Specifically: The initial 3D point cloud data is projected and converted into a 2D elevation map, which is then input into the visual encoding branch. This 2D elevation map uses elevation values ​​to replace RGB color channels and encodes a grayscale or pseudo-color image. The initial 3D point cloud data is converted into a sequence of 2D rendered images according to multiple spatial perspectives, which are then input into the visual encoding branch. These multiple spatial perspectives include different angles such as frontal view, oblique view, and top view. The initial 3D point cloud data is converted into structured text information, which is then input into the text encoding branch. The structured text information includes information such as the 3D coordinate range, statistical elevation, statistical curvature, and slope distribution of each region. The geometric feature vectors of the initial 3D point cloud data are extracted and mapped to the embedding space of the large language model. The geometric feature vectors include information such as local curvature, normal vector direction histogram, and point density distribution.

[0038] Finally, the corresponding data is classified using each modality coding branch. After data fusion analysis and semantic segmentation, the data is combined with the geometric feature vectors mapped to the large language model to identify various semantic units. The identified semantic units include, but are not limited to, the following types: beaches, reef areas, intertidal zones, supratidal zones, subtidal zones, vegetated areas, artificial structures (e.g., breakwaters, wharves), and water surfaces, covering all topographic features and artificial structures of the target coastal area.

[0039] The initial 3D point cloud data in this embodiment is affected by the acquisition time and will have certain differences. For example, under good sunlight conditions, the acquired 3D point cloud data contains the aforementioned features and can be converted and extracted into the corresponding modality coding branch. However, the 3D point cloud data acquired at night does not contain RGB information. In this case, the path from the multi-view 2D rendered image sequence to the visual coding branch cannot obtain color and texture features. However, the other three paths (2D elevation map to visual coding branch, structured text to text coding branch, and geometric feature vector to embedding space) do not rely on RGB information and can still work normally. The only difference is that the semantic segmentation precision is reduced due to the lack of texture assistance. Therefore, the large language model in this application can flexibly combine and utilize different branches during the semantic unit recognition process, depending on the actual acquisition conditions and data completeness. This can be a single branch or a combination of multiple branches, ultimately analyzing different semantic units.

[0040] Step S4: Collect the attribute features of each semantic unit, combine them with the tidal database, analyze the correlation between semantic units, and generate a differentiated scanning strategy.

[0041] In order to set the optimal scanning strategy for different semantic units, multiple attribute features of semantic units are extracted, including physical attributes, functional attributes, state attributes, and tidal attributes. The aforementioned attributes are analyzed separately, and different scanning strategies are set according to the priority ranking of individual attributes and scanning resolution requirements. At the same time, the aforementioned attributes are correlated, and the priority ranking and scanning resolution requirements of different semantic units in the overall attribute planning are analyzed and planned as a whole, ultimately forming a differentiated scanning strategy.

[0042] The classification based on attributes specifically includes: The semantic units are categorized into natural features and man-made structures based on their physical attributes. These physical attributes include geometric shape, surface curvature distribution, average slope, elevation information, and material type estimation. For example, the terrain corresponding to a portion of the 3D point cloud data can be inferred from point cloud roughness and local curvature statistics as either sandy or rocky; the spatial arrangement rules of the points can be used to infer whether it is a man-made structure or a natural feature.

[0043] Based on the distinction between natural features and man-made structures, these features and structures can be ranked according to their semantic unit functional attributes to establish a priority sequence list. The functional attributes are determined by the measured importance of each unit based on its semantic category. For example, breakwaters are critical engineering structures, assigned the highest engineering concern level, and thus have a high priority ranking; reef areas directly affect navigation safety, assigned a high navigation safety relevance, and thus have a high priority ranking; flat beaches are the primary terrain modeling objects, assigned a medium priority ranking; and water surfaces do not require 3D modeling and are assigned the lowest priority. Based on these priority rankings, a final priority sequence list is formed.

[0044] Based on the state attributes of semantic units, the state of each semantic unit is evaluated using standard geometric feature analysis, and state anomalies are identified to establish a state information table. The state attributes are assessed through geometric feature analysis and comparison with a typical reference morphology, which is a pre-collected standard morphology. For example, detecting abnormal depressions in reef areas may indicate erosion; assessing seasonal trends in beach profiles may determine changes in vegetation cover. The corresponding scanning resolution can be set based on the state information table.

[0045] Based on the tidal level attributes of semantic units, the exposure time window of each semantic unit within the current tidal cycle is calculated and sorted according to the exposure time window. Specifically: Elevation information from the physical attributes of the semantic units is obtained, and combined with the tidal database, the exposure time window of each semantic unit within the current tidal cycle is calculated. Before combining the elevation information with the data in the tidal database, it is necessary to ensure that the elevation information and the tidal level data in the tidal database use the same vertical reference. If different references are used, elevation reference conversion must be performed first. Based on the exposure time window, semantic units are divided into long-term scanable areas, periodically scanable areas, and non-scanable areas. For non-scanable areas (i.e., areas that are always submerged), no active scanning tasks are assigned; only the water level data in the tidal database is used to mark their land-water boundaries. If underwater topographic data is required, auxiliary detection methods such as multibeam sonar are used for separate acquisition. Based on the different area divisions and the exposure time window of each semantic unit, a scanning time sequence for each semantic unit is established and sorted.

[0046] For example, intertidal mudflats with elevations between -0.3m and 0.3m are exposed only when the tide level is below 0.3m, corresponding to a time window of approximately 1.5 hours before and after low tide, totaling about 3 hours. Therefore, this area needs to be scanned within this specific time range. Vegetated areas with elevations above 2.0m are always scannable. Areas with elevations below -1.0m are always submerged and therefore cannot be scanned. Thus, it is necessary to distinguish the areas that can only be scanned within the tidal cycle and set these areas as the highest priority within the characteristic time cycle. The remaining areas do not need to consider the time cycle and are sorted according to general priority.

[0047] For each semantic unit, based on the order of related attributes and the correlation between these attributes, different measurement priorities, scanning modes, and scanning resolutions are set, forming a differentiated scanning strategy, specifically: Based on the different attributes of semantic units, the measurement priorities are initially sorted, and the scanning mode and scanning resolution are initially set to generate multiple sets of scanning strategies.

[0048] After sorting all semantic units separately according to physical attributes, functional attributes, state attributes, and tidal attributes, multiple sorting priority combinations will be formed. Then, according to this preliminary priority, the corresponding measurement priority, scanning mode, and scanning resolution will be set to form multiple sets of scanning strategies.

[0049] The scanning modes include: line scan mode, suitable for elongated structures, which uses a reciprocating route along the axis; grid scan mode, suitable for large flat areas, which can traverse the area in a serpentine pattern; spiral scan mode, suitable for point targets, such as isolated reefs, which uses a spiral scanning method from the center outwards; contour scan mode, suitable for sloping areas, which can use contour lines to lay out the scanning route; and waterline tracking mode, suitable for marking water boundaries. The scanning resolution is deployed according to the pre-set scanning resolution for each semantic unit.

[0050] For example: Measurement priority sorting based on functional attributes results in sorting combination A: breakwater, reef area, intertidal mudflat, sandy vegetation area, and water surface area. Measurement priority sorting based on tidal level attributes results in sorting combination B: intertidal mudflat, reef area, breakwater, sandy vegetation area, and water surface area. Measurement priority sorting based on state attributes results in sorting combination C: reef area, breakwater, intertidal mudflat, sandy vegetation area, and water surface area. This generates three different scanning modes. Furthermore, changes in the sorting of each scanning mode will also produce corresponding changes based on the attribute state of the semantic unit, and this is not limited to the examples in this implementation.

[0051] Based on the correlation between different attributes of semantic units, the measurement priority ranking, scanning mode, and scanning resolution are adjusted a second time.

[0052] In the initial sorting combinations, the corresponding sequence positions of each semantic unit may not be exactly the same, because the same semantic unit may have different sorting positions in different attribute sortings. For example, the sorting of a reef area may differ in functional attributes, state attributes, and tide level attributes. Secondary adjustments are made based on these sortings. The principle for adjusting sorting priority at this time is as follows: semantic units with high security associations have higher priority than general semantic units; for example, reef areas and breakwaters have higher priority than other semantic units. Semantic units with time window constraints need to be processed intensively within the window period, and their priority within that time window is higher than general semantic units, such as semantic units within a phased scanable area. Semantic units with abnormal states require higher resolution to accurately quantify the degree of abnormality, i.e., differences in scanning resolution settings. For example, if there is vegetation cover in a beach vegetation area, the scanning resolution needs to be increased to better determine the changing trend of the vegetation area; in this case, the scanning resolution needs to be improved and optimized.

[0053] Based on the relationships between semantic units, the measurement priority sorting, scanning mode, and scanning resolution are readjusted to obtain optimized measurement priority sorting, scanning mode, and scanning resolution settings.

[0054] For example, there may be situations where reef areas overlap with intertidal mudflat areas. If the reef area is only visible during low tide, the priority order for measuring the reef area should be adjusted to reference the priority order of the intertidal mudflat area. Similarly, the foreslope of a breakwater is only visible during low tide; therefore, the priority order for scanning this area also needs to consider the tidal cycle.

[0055] Based on different measurement priorities and the distribution location of each semantic unit, a scanning path and scanning time window are planned for the target coastal area, forming the first scanning strategy. Based on different scanning resolution settings, different scanning modes and resolutions are planned for the target coastal area, forming the second scanning strategy. The first and second scanning strategies are then merged to form a differentiated scanning strategy.

[0056] Once the measurement priority, scanning mode, and scanning resolution of each semantic unit are determined, a scanning path needs to be planned to scan each semantic unit sequentially. At the same time, the scanning time of each semantic unit must be considered to ensure that all semantic units within the tidal cycle can be scanned within the corresponding tidal cycle.

[0057] In this embodiment, the scanning device mainly uses a depth camera or a LiDAR. Different types of depth cameras or LiDARs are selected according to different scanning modes and accuracy requirements. The parameters are configured through the aforementioned differentiated scanning strategy, which can quickly switch the scanning device and ultimately achieve perfect operation of the differentiated scanning strategy. This embodiment provides some selection methods.

[0058] Table 1 Differentiated Scanning Strategies

[0059] If a single low tide window is insufficient to cover all relevant intertidal areas, the uncovered areas are automatically added to the task queue of the next low tide cycle, forming a cross-cycle closed-loop scheduling. After any scanning cycle ends, the large language model updates the scanning completion status of each semantic unit based on the scanning results of that cycle. For areas that are not completed or whose data quality does not meet the requirements, the scanning strategy for the next cycle is automatically replanned.

[0060] Step S5: Perform directional scanning on each semantic unit according to the differentiated scanning strategy to obtain optimized 3D point cloud data.

[0061] Step S6: Generate a 3D model of the target coastal area based on the optimized 3D point cloud data.

[0062] Determine whether each semantic unit has been scanned completely. If there are semantic units that have not been scanned completely, obtain the uncovered area and perform directional scanning on the uncovered area in the next cycle to form complete 3D point cloud data.

[0063] Example 1: This embodiment uses a drone equipped with a TOF depth camera to perform terrain scanning and measurement of a coastal area of ​​a harbor.

[0064] (a) Environmental preparation and initial scan: Acquire tide table data for the target coastal area and determine the scan operation window based on the low tide time of the day. It is preferable to start the operation 1 to 2 hours before low tide to ensure that the core scan task is completed during the low tide period.

[0065] A drone equipped with a TOF depth camera was used to perform a global initial scan of the target coastal space at a first resolution (flight altitude 80 meters, ground sampling interval 5 centimeters). The target coastal space covered a complete coastal profile from supratidal dunes to subtidal shallow waters, including various heterogeneous objects such as beaches, reefs, breakwaters, vegetation areas, and the water surface. Initial 3D point cloud data was acquired using the airborne depth camera, and then denoised and preliminarily filtered to remove flying points and high-frequency noise.

[0066] (II) LLM Semantic Segmentation and Multi-Attribute Reasoning: The initial 3D point cloud data is projected into a 2D elevation map, and local curvature, slope and roughness are extracted and input into the visual coding branch. Here, a single modality coding branch is taken as an example. In the actual extraction process, multiple modality coding branches can be used to fuse and identify semantic units, which has the highest recognition accuracy.

[0067] The visual coding branch identified the following semantic units: beach area (elevation between 0.5m and 2.0m, with a relatively gentle slope), reef area (containing multiple irregular protruding structures with large curvature), intertidal mudflat area (elevation between -0.5m and 0.5m, with a relatively rough surface), breakwater (regular geometric structure, consistent with the characteristics of artificial structures), upper vegetation area (elevation above 2.0m, with rich texture), and water surface (low texture phenomenon, and with high reflectivity).

[0068] Priority of large language model inference measurement: breakwater, reef area, intertidal mudflat, beach and vegetation area, water surface.

[0069] (III) Differentiated Scanning Strategy: Based on scene characteristics, a strategy and preferred sensor are assigned to each semantic unit—currently, under good daytime lighting conditions, all units are assigned a TOF depth camera as the primary sensor: Breakwater: 1 cm resolution, line scan mode, back and forth along the breakwater axis, path covering the breakwater top and the front slope.

[0070] Reef area: 2 cm resolution, spiral scan mode covers the reef protrusion area.

[0071] Intertidal mudflats: resolution of 3 cm, full coverage in raster scan mode, time window of 1.5 hours before and after low tide.

[0072] Beaches and vegetated areas: 5 cm resolution, raster scan mode, scheduled for high tide.

[0073] Water surface area: No scanning task is assigned; only the scan is tracked along the waterline to mark the water-land boundary.

[0074] (iv) Secondary directional scanning and closed-loop planning: The scanning control module controls the UAV to perform directional scanning sequentially according to the differentiated scanning strategy. The system monitors the deviation between the flight trajectory and the planned path in real time, evaluates the coverage integrity online, and automatically inserts waypoints for supplementary scanning when missed areas are found.

[0075] After a complete scanning cycle (from low tide to the next low tide), the completion status of each semantic unit is updated, and the strategy for the next cycle is automatically replanned for areas that are not covered or whose data quality does not meet the requirements.

[0076] The results of using the differential scanning strategy in this embodiment compared to the traditional uniform resolution strategy are as follows: A traditional uniform resolution strategy was used to perform a complete scan of the harbor coastline area (approximately 2.5 km²), which required the highest resolution to cover the entire area and took about 3.5 hours. In contrast, this embodiment uses a differentiated scanning strategy, which uses high resolution (1-2 cm) for fine scanning of breakwaters and reef areas, low resolution (5 cm) for rapid scanning of beaches and vegetation areas, and only tracks the waterline in the water area. The total time taken was approximately 2.0 hours, which is about 43% shorter than the previous method.

[0077] Traditional uniform strategies cover the entire area at a medium resolution (about 3 cm), but the structural details of the breakwater (such as gaps between boulders and undulations in the revetment) cannot be clearly reconstructed. In contrast, this embodiment performs a line scan of the breakwater at a resolution of 1 cm, which preserves the structural details completely, and the recognition accuracy of abnormal depressions in the reef area is improved by about 60% compared with the uniform strategy.

[0078] Traditional uniform strategies do not consider tidal factors and often perform ineffective scans of the intertidal zone after the tide rises, requiring repeated flights in the next low tide cycle. This embodiment, by sorting the exposure time windows, precisely arranges the scanning tasks of the intertidal mudflats and reef areas to be carried out in a concentrated manner within a window period of 1.5 hours before and after low tide. The coverage rate of the intertidal zone in a single cycle reaches more than 95%, which is a significant improvement over the traditional strategy (about 60%).

[0079] Under the traditional uniform strategy, about 40% of the point cloud data is distributed in the beach and water areas where high-precision modeling is not required, resulting in a large amount of redundancy. The differentiated scanning strategy reduces the total amount of point cloud data by about 35% by allocating resolution according to the region, shortens the subsequent modeling and processing time by about 30%, and increases the point cloud density in key areas by 2 to 5 times.

[0080] Example 2: This embodiment is a tide-adaptive scanning system based on multi-sensor fusion. The difference between this embodiment and the first embodiment is that the three-dimensional scanning device is a combination system of a TOF depth camera and a LiDAR. The TOF camera is responsible for high-precision scanning at medium and close distances (within 30 meters), while the LiDAR is responsible for full-coverage scanning at long distances (30 meters to 150 meters).

[0081] When generating differentiated scanning strategies, a preferred sensor type is assigned to each semantic unit based on the average distance between each unit and the scanning platform. For example, a TOF camera is preferred for breakwater structures that are within close range, while a LiDAR sensor is used for beaches with a wide distribution area.

[0082] The combined deployment scheme can meet the needs of coastal scenes with complex terrain, including both close-range fine structures and large-scale natural features. The depth camera provides dense textures in key areas for fine semantic analysis, while LiDAR provides a large-scale geometric framework to ensure full coverage efficiency.

[0083] This embodiment demonstrates that after generating a differentiated scanning strategy, the scanning method can be configured, thereby facilitating rapid switching of scanning devices during actual scanning and ensuring the accuracy of scanning recognition.

[0084] Example 3: This embodiment is a large-scale coastal terrain scanning based on airborne lidar. The difference between this embodiment and the first embodiment is that the three-dimensional scanning device is a lightweight mechanical rotating lidar (LiDAR) mounted on a drone, which can realize rapid scanning of a large-scale coastal terrain.

[0085] (I) LiDAR Technology Configuration: Laser wavelength selection of 1550 nm—this band has an extremely high absorption coefficient in water (penetration depth of only about 1 mm), and the laser energy is almost completely reflected by the water surface, ensuring accurate determination of water surface boundaries; at the same time, it has sufficient diffuse reflectivity for natural features such as sand, rocks, and vegetation. Compared with the commonly used 1064 nm band, 1550 nm has a higher eye safety threshold, allowing for the emission of higher power laser pulses to obtain a longer ranging range.

[0086] (II) Scanning Execution: The lidar performs an initial global scan at a first resolution (flight altitude of 120 meters, point cloud density of approximately 50 points / square meter). After projecting the lidar point cloud into an elevation map, semantic segmentation and multi-attribute reasoning are performed to generate a differentiated scanning strategy. For key areas such as breakwaters and reef areas, the UAV automatically lowers its flight altitude to 50 meters to increase the point cloud density to over 200 points / square meter for a secondary directional scan; the initial altitude and density are maintained for beaches and vegetated areas.

[0087] In this embodiment, the initial point cloud density obtained by the lidar (50 points / square meter) is lower than that of the TOF depth camera scheme in Embodiment 1 (400 points / square meter, converted based on 5 cm GSD). However, the lidar has the advantage of high coverage efficiency (the scanning area per scan is 5-8 times that of the TOF scheme). When performing semantic segmentation using a large language model, a semantic reasoning strategy based on topographic statistical features is additionally applied to the low-density point cloud of the lidar: by analyzing the regional elevation standard deviation, the spatial autocorrelation length of the slope, and the consistency of the point cloud normal vector, natural features and man-made structures are distinguished, thus compensating for the impact of insufficient point cloud density on the accuracy of semantic segmentation.

[0088] This embodiment demonstrates that the differentiated scanning strategy generated using a large language model can be applied to different types of 3D scanning devices. It only requires adaptive adjustment of the preprocessing and inference strategies based on the point cloud features (density, noise level, spatial coverage pattern) output by different sensors.

[0089] Example 4: This embodiment utilizes a combined multibeam sonar and lidar scanning system based on an unmanned surface vessel (USV). The difference between this embodiment and the first embodiment lies in that the 3D scanning device is mounted on an USV platform, and the target coastal space is the nearshore shallow waters and intertidal zone. The USV navigates along a predetermined route at low tide, simultaneously scanning the underwater and above-water topography using a combined shipborne multibeam sonar and LiDAR system.

[0090] The large language model additionally identifies water depth-related semantic units (shoals, deep channels, tidal channels) during semantic segmentation, and combines water depth data to infer the measurement priority and applicable sensor types of each unit, generating differentiated scanning strategies suitable for joint surface and underwater scanning.

[0091] Example 5: This embodiment demonstrates the all-weather operation capability of the invention through autonomous LiDAR scanning at night / inclement weather. A coastal monitoring station needed to conduct an emergency scan and assessment of the coastal topography after a storm during a winter night (when sunshine hours are 0 hours). The system automatically detected the current ambient light conditions (below the depth camera's operating threshold) and instructed the 3D scanning module to switch to LiDAR-only mode.

[0092] A drone equipped with a 1550 nm LiDAR performed an initial global scan at a flight altitude of 100 meters in complete darkness. The large language model performed semantic segmentation from the LiDAR point cloud without texture assistance (nighttime point clouds lack RGB information), relying solely on geometric features: distinguishing between beaches, reefs, and man-made structures by analyzing the statistical features of point cloud elevation, local curvature distribution, and surface roughness in each region. Specifically, without RGB information, the path of converting the initial 3D point cloud data into a 2D rendered image sequence from multiple spatial perspectives was unusable due to the inability to obtain color and texture information. The other three paths—converting point cloud projection into a 2D elevation map input to the visual encoding branch, converting point cloud into structured text information input to the text encoding branch, and extracting point cloud geometric feature vectors and mapping them to the large language model embedding space—all operate without RGB information and still function normally, achieving semantic segmentation. The only drawback was a slight decrease in segmentation precision compared to daytime operations due to the lack of texture assistance. Subsequent differential scanning strategy generation and secondary directional scanning were performed identically to the daytime workflow.

[0093] Post-storm scanning revealed a scour pit of approximately 30 cm on the front slope of the breakwater (not present in the previous daytime scan). The differential scanning strategy automatically marked this area as a high-priority anomaly. This embodiment demonstrates that the differential scanning strategy of this application can select an appropriate scanning device based on the current environment, and ultimately still generate corresponding semantic units through a large language model, facilitating subsequent monitoring and analysis. Furthermore, this embodiment verifies the robustness of the method under conditions of missing RGB information—the system can automatically switch to a pure geometric feature inference mode that does not rely on texture information, ensuring all-weather operational capability.

[0094] The above embodiments demonstrate that the large language model of this application can fuse and analyze various information transformed from 3D point cloud data. The resulting differentiated scanning strategy can achieve optimal scanning device combination, optimal path planning, and optimal scanning accuracy and resolution. Furthermore, the differentiated scanning strategy defined in this application can adapt to different environmental changes, not requiring the 3D point cloud data to be transformed into multiple information sets, but rather allowing for one or more combinations of such sets, changing according to the natural environment corresponding to the acquisition time.

[0095] Second implementation method: refer to Figure 4 As shown, this embodiment provides an adaptive construction system for a 3D coastal terrain model, including the following modules: Acquisition module 101 is used to perform an initial scan of the target coastal area according to the initial scanning strategy to acquire initial three-dimensional point cloud data. Acquisition module 202 is used to collect real-time tide levels of the target coastal area and construct a tide database by combining it with the tide table; Processing module 1203 is used to extract various transformation features of the initial 3D point cloud data and map them to the embedding space of the large language model, and to perform semantic segmentation on the initial 3D point cloud data to form multiple semantic units. Processing module 204 is used to collect the attribute features of each semantic unit, combine them with the tidal database, analyze the correlation between semantic units, and generate a differentiated scanning strategy. Processing module 3 205 is used to perform directional scanning on each semantic unit according to the differentiated scanning strategy to obtain optimized 3D point cloud data; Modeling module 206 is used to generate a 3D model of the target coastal area based on the optimized 3D point cloud data; Analysis module 207 is used to collect optimized 3D point cloud data and determine whether it is complete, whether all semantic units have been scanned, and send the uncovered areas corresponding to the semantic units that have not been scanned to processing module 204. Processing module 204 performs directional scanning in the next cycle to form complete 3D point cloud data.

[0096] It is not difficult to see that this embodiment is a system implementation corresponding to the first embodiment, and this embodiment can be implemented in conjunction with the first embodiment. The relevant technical details mentioned in the first embodiment are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the first embodiment.

[0097] It is worth mentioning that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this invention, this embodiment does not introduce units that are not closely related to solving the technical problem proposed by this invention; however, this does not mean that other units are absent from this embodiment.

[0098] Third implementation method: refer to Figure 5 As shown, this embodiment provides a device including a processor 301 and a memory 302. The memory 302 stores a set of program instructions. When the set of program instructions stored in the memory 302 is loaded and executed by the processor 301, the aforementioned adaptive construction method for a three-dimensional coastal terrain model can be realized.

[0099] The memory 302 and processor 301 are connected via a bus, which can include any number of interconnecting buses and bridges. The bus connects various circuits of one or more processors 301 and memory 302 together. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. A bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 301 is transmitted over a wireless medium via an antenna, which further receives data and transmits it to processor 301.

[0100] Processor 301 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory 302 can be used to store data used by processor 301 during operation.

[0101] The above are merely preferred embodiments of this application. The scope of protection of this application is not limited to the above embodiments. All technical solutions falling within the scope of this application's concept are within the scope of protection of this application. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of this application should be considered within the scope of protection of this application.

Claims

1. An adaptive construction method for a 3D coastal terrain model, characterized in that, Includes the following steps: Step S1: Perform an initial scan of the target coastal area according to the initial scanning strategy to obtain initial 3D point cloud data; the initial scanning strategy uses a fixed resolution and uniform height to perform a full-coverage scan of the target coastal area; Step S2: Collect real-time tide levels of the target coastal area and construct a tide database by combining them with tide tables; Step S3: Extract various transformation features from the initial 3D point cloud data and map them to the embedding space of the large language model. Perform semantic segmentation on the initial 3D point cloud data to form multiple semantic units. Step S4: Collect the attribute features of each semantic unit, combine them with the tidal database, analyze the correlation between semantic units, and generate a differentiated scanning strategy; The semantic unit can be classified into natural features and man-made structures based on its physical attributes. Natural features and man-made structures are ranked according to their importance based on the functional attributes of their semantic units, and a priority sequence list is established. Based on the state attributes of semantic units, the state of each semantic unit is evaluated using standard geometric features, and state anomaly identification is performed to establish a state information table. Based on the tidal level attribute of the semantic units, calculate the exposure time window of each semantic unit within the current tidal cycle, and sort them according to the exposure time window; For each semantic unit, sort it according to the relevant attributes, combine the correlation between the relevant attributes, set different measurement priorities, scanning modes and scanning resolutions, and form a differentiated scanning strategy; Step S5: Perform directional scanning on each semantic unit according to the differentiated scanning strategy to obtain optimized 3D point cloud data; Step S6: Generate a 3D model of the target coastal area based on the optimized 3D point cloud data.

2. The adaptive construction method for a three-dimensional coastal terrain model according to claim 1, characterized in that, Multiple transformation features of the initial 3D point cloud data are extracted and mapped to the embedding space of a large language model. Semantic segmentation is then performed on the initial 3D point cloud data to form multiple semantic units, specifically: Extracting multiple transformation features from initial 3D point cloud data; Multiple transformation features are input into the corresponding modality coding branches within the large language model, and semantic segmentation is performed to identify multiple semantic units.

3. The adaptive construction method for a three-dimensional coastal terrain model according to claim 2, characterized in that, Multiple transformation features are input into the corresponding modality coding branches within the large language model, and semantic segmentation is performed to identify various semantic units, specifically: The initial 3D point cloud data is projected and converted into a 2D elevation map, which is then input into the visual encoding branch. The initial 3D point cloud data is transformed into a sequence of 2D rendered images according to multiple spatial perspectives and then input into the visual encoding branch. The initial 3D point cloud data is converted into structured text information, and the input text encoding branch is used. Extract the geometric feature vectors from the initial 3D point cloud data and map them to the embedding space of the large language model; The corresponding data are classified and processed using various sub-models. After data fusion analysis and semantic segmentation, the geometric feature vectors mapped to the large language model are combined to identify multiple semantic units.

4. The adaptive construction method for a three-dimensional coastal terrain model according to claim 3, characterized in that, Based on the tidal level attribute of the semantic units, the exposure time window of each semantic unit within the current tidal cycle is calculated, and the units are sorted according to their exposure time windows, as follows: Obtain elevation information from the physical attributes of semantic units, and combine it with the tidal database to calculate the exposure time window of each semantic unit within the current tidal cycle; Based on the exposure time window, semantic units are divided into long-term scannable regions, periodically scannable regions, and unscannable regions. Based on different regional divisions and the exposure time windows of each semantic unit, a scanning time series of each semantic unit is established and sorted.

5. The adaptive construction method for a three-dimensional coastal terrain model according to claim 4, characterized in that, For each semantic unit, based on the order of related attributes and the correlation between these attributes, different measurement priorities, scanning modes, and scanning resolutions are set, specifically as follows: Based on the different attributes of semantic units, the measurement priorities are initially sorted, and the scanning mode and scanning resolution are initially set to generate multiple sets of scanning strategies. Based on the correlation between different attributes of semantic units, the measurement priority sorting, scanning mode and scanning resolution are adjusted a second time; Based on the relationships between semantic units, the measurement priority sorting, scanning mode, and scanning resolution are readjusted to obtain optimized measurement priority sorting, scanning mode, and scanning resolution settings.

6. The adaptive construction method for a three-dimensional coastal terrain model according to claim 5, characterized in that, Differentiated scanning strategies are formulated based on different measurement priorities, scanning modes, and scanning resolutions, specifically as follows: Based on different measurement priorities and the distribution location of each semantic unit, a scanning path and scanning time window are planned for the target coastal area to form the first scanning strategy; Based on different scanning resolution settings, different scanning modes and scanning resolutions are planned for the target coastal area to form a second scanning strategy; The first and second scanning strategies are combined to form a differentiated scanning strategy.

7. The adaptive construction method for a three-dimensional coastal terrain model according to claim 6, characterized in that, After performing directional scanning on each semantic unit according to a differentiated scanning strategy to obtain optimized 3D point cloud data, the process also includes: Determine whether each semantic unit has been scanned completely. If there are semantic units that have not been scanned completely, obtain the uncovered area and perform directional scanning on the uncovered area in the next cycle to form complete 3D point cloud data.

8. A system for adaptively constructing a three-dimensional coastal terrain model according to any one of claims 1-7, characterized in that, Includes the following modules: Acquisition Module 1 is used to perform an initial scan of the target coastal area according to the initial scanning strategy to obtain initial 3D point cloud data. The second data acquisition module is used to collect real-time tide levels in the target coastal area and, in conjunction with the tide table, construct a tide database. Processing module one is used to extract various transformation features of the initial 3D point cloud data and map them to the embedding space of the large language model, and to perform semantic segmentation on the initial 3D point cloud data to form multiple semantic units; Processing module two is used to collect the attribute features of each semantic unit, combine them with the tidal database, analyze the correlation between semantic units, and generate differentiated scanning strategies. Processing module three is used to perform directional scanning on each semantic unit according to the differentiated scanning strategy to obtain optimized 3D point cloud data; The modeling module is used to generate a 3D model of the target coastal area based on the optimized 3D point cloud data. The analysis module is used to collect optimized 3D point cloud data and determine whether it is complete, whether all semantic units have been scanned, and send the uncovered areas corresponding to the semantic units that have not been scanned to the processing module 2. The processing module 2 performs directional scanning in the next cycle to form complete 3D point cloud data.

9. An electronic device comprising a processor and a memory, wherein the memory stores a set of program instructions, characterized in that: When the program instruction set stored in the memory is loaded and executed by the processor, the adaptive construction method of the three-dimensional coastal terrain model according to any one of claims 1-7 can be implemented.

Images