Comprehensive surveying and mapping method for complex terrain and data processing method

Abstract

The application relates to a comprehensive surveying and mapping and data processing method for complex terrains, and relates to the technical field of surveying and mapping geographic information. The method comprises the following steps: based on a terrain complexity index, a to-be-surveyed region is adaptively divided into sub-regions, and a multi-platform surveying and mapping system is dynamically scheduled to generate a job parameter set; a unified navigation module and a reference point are utilized to construct a global reference framework, real-time quality sensing data are returned, and re-acquisition is automatically triggered; multi-source data spatial alignment is realized by extracting multiple types of geometric features and utilizing a two-level registration mechanism and a water-land junction constraint; point cloud classification and denoising are performed based on a deep learning model, and a filling strategy is automatically selected according to a cavity feature; a five-dimensional quality evaluation model is constructed to generate a quality inspection report and output multi-format achievements. The application can effectively solve the problems of incomplete surveying and mapping coverage, low multi-source data registration accuracy and fragmented operation process under complex terrains, and guarantee the integrity, consistency and high precision of surveying and mapping achievements.

Description

Technical Field

[0001] This disclosure generally relates to the field of surveying and mapping geographic information technology. More specifically, this disclosure relates to methods for integrated surveying and data processing for complex terrain.

[0002] Terminology definition: 1. Topographic Complexity Index: A comprehensive quantitative evaluation of the topographic relief, vegetation uniformity, and water coverage ratio of the area under test. 2. Two-stage registration mechanism: a spatial alignment method comprising two stages: coarse registration with strong control points as constraints and fine registration based on multiple geometric features. 3. Iterative Nearest Geometric Feature Algorithm: An improvement on the traditional ICP algorithm, this algorithm integrates multiple geometric features (points, lines, and surfaces) for registration optimization. 4. Multidimensional Quality-Aware Data: A set of indicators reflecting data quality generated in real time during the data collection process. Background Technology

[0003] Complex terrain areas are key targets for infrastructure construction, natural resource surveys, and geological disaster monitoring. Their environmental characteristics are usually characterized by dramatic landform undulations, dense vegetation cover, or a mix of land and water.

[0004] In existing surveying and mapping operations, geospatial data is typically acquired using technologies such as UAV aerial surveying, terrestrial 3D laser scanning, RTK surveying, or underwater sonar detection. In practical applications, these technologies often follow a pre-defined work plan, using data acquisition equipment equipped with navigation and positioning systems to collect data from the target area and establishing a coordinate reference frame by setting up ground control points. During the data processing phase, technicians perform geometric corrections on the acquired imagery and point cloud data, using feature matching algorithms to stitch data from different sources into a unified coordinate system. Subsequently, noise is removed and data gaps are filled using manual or semi-automatic methods, ultimately generating surveying and mapping results such as digital elevation models and digital orthophotos to meet the needs of engineering planning and management.

[0005] However, in the face of highly complex environments such as mountains, canyons, dense forests and water bodies, existing technologies cannot achieve effective coverage of all elements when using single or simple combination of surveying and mapping methods. Moreover, in the absence of a unified intelligent scheduling and real-time quality feedback mechanism, multi-source heterogeneous data are prone to local data loss, insufficient registration accuracy and fragmented work processes, making it difficult to fully guarantee the integrity and consistency of surveying and mapping results. Summary of the Invention

[0006] This application provides a comprehensive surveying and data processing method for complex terrain, which can solve the technical problems of incomplete surveying methods, low accuracy of multi-source data registration, low degree of automation in data processing, and fragmented work processes in complex terrain.

[0007] In some embodiments, a comprehensive mapping and data processing method for complex terrain includes the following steps: obtaining a terrain complexity index based on basic geographic data of the area to be surveyed, wherein the terrain complexity index is a preset weighted quantitative index that integrates slope variability, vegetation cover homogeneity, and water area proportion; adaptively dividing the area into sub-regions based on the index, and dynamically scheduling one or more of the following systems—UAV aerial survey system, ground 3D laser scanning system, backpack laser scanning system, and underwater sonar detection system—based on the terrain attribute characteristics of each sub-region, generating a set of operational parameters for each system, wherein the set of operational parameters includes at least spatial coverage overlap rate, sensor working height / depth, point cloud sampling resolution, and image exposure timing constraints; integrating a unified high-precision GNSS / IMU tightly coupled navigation module and synchronizing timing for all scheduling systems, and deploying permanent reference points with known 3D coordinates and rapidly deployable temporary markers within the survey area to form a global spatial reference frame; transmitting multi-dimensional quality sensing data in real time during field data collection, wherein the multi-dimensional quality sensing data includes point cloud density distribution map, image modulation transfer function value, positioning solution confidence, and IMU data. The attitude angle jitter is measured, and when any indicator falls below a preset threshold, a local reacquisition command is automatically triggered for the corresponding system. After error modeling and geometric correction of the raw observation data output by each system, point and line geometric features in the image data and area and edge geometric features in the point cloud data are extracted respectively. Spatial alignment of multi-source data is achieved through a two-level registration mechanism: coarse registration uses permanent reference points and temporary marker points as strong constraints, and completes global coordinate system normalization by combining cross-modal feature matching results; fine registration criteria construct a weighted objective function that integrates point-to-point distance, point-to-line distance, and point-to-plane distance. Local deformation optimization was performed using an iterative nearest geometric feature algorithm, and a shoreline spatial projection consistency constraint equation was introduced in the land-water interface area to force the edge features of the land image and the underwater isobath features to coincide in the tidal-corrected 2D projection space. A lightweight deep learning semantic segmentation model was used to classify the corrected point cloud point by point, identifying ground points, vegetation points, building points, and dynamic noise points, and denoising operations were performed differentially based on the classification results. Further analysis of the spatial scale, topological morphology, and neighborhood topographic gradient characteristics of point cloud holes was conducted, and hole-filling strategies were automatically selected according to preset decision rules: for holes with an area less than 0.5 m² and a neighborhood slope change rate less than 5%, neighbor point weighted interpolation was used; for holes that are long and narrow and cross significant slope abrupt change zones, surface fitting based on a quadratic polynomial trend surface was used; and for holes that cross the land-water boundary and have a depth gradient greater than 0...A 15 m / m cavity was filled using multi-source collaborative filling, integrating texture information from UAV orthophotos and underwater sonar profile data. A quality evaluation model was constructed, encompassing five dimensions: planar position accuracy, elevation accuracy, point cloud spatial integrity, feature topological correctness, and image radiometric consistency. This model automatically performed quality checks on the entire dataset, generating a structured quality inspection report containing spatial location data of non-compliant areas, error cause analysis, and correction suggestions. The processed results can be exported to industry-standard vector, point cloud, raster, and engineering drawing formats. The platform scheduling driven by the terrain complexity index, the two-level registration through multi-geometric feature fusion, and the data processing involving semantic recognition and joint decision-making based on cavity features constitute a closed-loop feedback technology chain, jointly ensuring the integrity, consistency, and high accuracy of surveying and mapping results in complex terrain environments.

[0008] In some embodiments, the Terrain Complexity Index (TCI) is calculated using the following formula: in, For the standard deviation of slope, The mean of the normalized vegetation index. The percentage of pixels in the water area. , , These are preset weighting coefficients.

[0009] In some embodiments, the dynamic weight coefficients of point features, line features, and surface features in the iterative nearest geometric feature algorithm. Update according to the following rules: , , , in For the first Residual for feature matching at each point; For the first The deviation of the included angle of the characteristic direction of each line For the first The deviation of the included angle of the normal of each surface feature This is the preset attenuation factor.

[0010] In some embodiments, when the cavity crosses the land-water boundary and the absolute value of the neighborhood water depth gradient is... When this occurs, multi-source fusion filling is forcibly enabled; the multi-source fusion filling specifically involves: using the shoreline extracted by the UAV DOM as a constraint, projecting the sonar profile data along the normal direction onto the DOM plane, and constructing a Poisson reconstruction equation with terrain constraints to solve the elevation field of the cavity region.

[0011] In some embodiments, after the automatic triggering of the local re-acquisition command, the system synchronously pushes the coordinates of the re-acquisition area, the original quality problem type, and historical quality inspection records to the task planning module, which is used to update the terrain complexity index calculation parameters of the sub-area and the subsequent platform scheduling strategy.

[0012] In some embodiments, in the unified high-precision GNSS / IMU tightly coupled navigation module, the GNSS receiver positioning accuracy is better than ±1 cm (RTK mode), the IMU gyroscope zero-bias instability is ≤0.5° / h, and the time synchronization error between the two is <1ms.

[0013] In some embodiments, the lightweight deep learning semantic segmentation model is an improved PointNet++ architecture, whose backbone network introduces a channel attention module (SE Block), and the input features include point cloud 3D coordinates, laser reflection intensity, local curvature, and corresponding image RGB texture values.

[0014] In some embodiments, the preset thresholds include: point cloud density < 80 pts / m², image modulation transfer function MTF50 < 80 lp / mm, positioning solution confidence < 95%, and IMU pitch jitter > 0.3°. Exceeding any of these limits will trigger a retest.

[0015] In some embodiments, the preset weighting coefficient =0.4、 =0.35、 =0.25, obtained through regression calibration using measured data from the southwestern mountainous region.

[0016] In some embodiments, this scheme constructs a terrain complexity index that integrates slope variability, vegetation cover homogeneity, and water area proportion. This index drives adaptive sub-region division of the surveying area and dynamic scheduling of multi-platform surveying systems, enabling the operational parameter set to accurately match the terrain attribute characteristics of each sub-region. This ensures the targeted nature and coverage integrity of data collection from the source. Based on this, a unified high-precision GNSS / IMU tightly coupled navigation module and synchronization timing mechanism, combined with a global spatial reference frame composed of permanent reference points and temporary markers, achieves strict uniformity of multi-source data in spatiotemporal reference. By real-time transmission of multi-dimensional quality-sensing data including point cloud density, image modulation transfer function values, positioning solution confidence, and IMU attitude angle jitter, and automatically triggering local re-acquisition commands when the indicators fall below preset thresholds, a closed-loop control of simultaneous measurement and inspection is formed, effectively preventing low-quality data from flowing into subsequent stages. Furthermore, by performing error modeling and geometric correction on the original data, multiple geometric features were extracted from the images and point clouds. A two-tiered mechanism was employed, utilizing coarse registration with strong control points as constraints and fine registration with a weighted objective function that integrates point, line, and surface distances. In particular, the introduction of a shoreline spatial projection consistency constraint equation in the land-water boundary region eliminated local deformation and stitching errors between heterogeneous data, significantly improving the spatial alignment accuracy of multi-source data. Subsequently, a lightweight deep learning semantic segmentation model was used to classify and differentially denoise the point clouds point by point. Based on the spatial scale, topological morphology, and neighborhood terrain gradient features of the holes, neighboring point weighted interpolation, trend surface fitting, or multi-source collaborative filling strategies were automatically selected, solving the problem of data hole repair under complex terrain. Ultimately, by constructing a five-dimensional quality evaluation model to automatically generate structured quality inspection reports and support multi-format output, the high-quality delivery of surveying and mapping results in terms of planar location, elevation, integrity, topology, and radiation consistency is ensured. This effectively solves the problems of incomplete surveying and mapping coverage, low registration accuracy, low processing efficiency, and fragmented processes in complex terrain environments, and realizes intelligent, integrated, and highly reliable surveying and mapping operations.

[0017] In some embodiments, this application constructs a rigorous closed-loop feedback technology chain through intelligent scheduling driven by terrain complexity index, two-level registration through multi-geometric feature fusion, and data processing involving semantic recognition and joint decision-making based on cavity features. This not only improves the technical performance of individual links but also achieves end-to-end collaborative optimization at the system level, from task planning to output, ensuring the integrity, consistency, and high accuracy of surveying and mapping results in complex terrain environments. In typical complex terrain surveying in southwestern mountains, compared with traditional single UAV aerial surveying methods, this method can improve overall operational efficiency by 3-5 times and significantly shorten project cycles. Attached Figure Description

[0018] Figure 1 is a flowchart illustrating a comprehensive mapping and data processing method for complex terrain according to a disclosed embodiment. Detailed Implementation

[0019] The present application will now be described in further detail with reference to embodiments. It is to be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the scope of the application.

[0020] Exemplary embodiments of the integrated mapping and data processing method for complex terrain according to this disclosure will now be described more fully below with reference to the accompanying drawings. However, such systems and methods may be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will convey certain exemplary aspects to those skilled in the art.

[0021] As used herein, the terrain complexity index or similar term can refer to a comprehensive quantitative evaluation of the degree of terrain undulation, vegetation distribution evenness, and water coverage ratio of the area to be surveyed. It should be understood that comprehensive mapping and data processing methods for complex terrain can include platform scheduling driven by the terrain complexity index, two-level registration through multi-geometric feature fusion, and data processing involving joint decision-making based on semantic recognition and void features. While this disclosure references certain embodiments, numerous modifications, alterations, and variations of the described embodiments are possible without departing from the spirit and scope of this disclosure as defined in the appended claims. Therefore, this disclosure is not limited to the described embodiments but has the full scope defined by the language of the following claims and their equivalents.

[0022] The embodiments disclosed herein include systems and methods for comprehensive mapping and data processing in complex terrain areas. For example, the method may include obtaining a terrain complexity index based on basic geographic data of the area to be surveyed, adaptively subdividing the area according to the index, and dynamically scheduling multiple mapping systems to generate operational parameter sets. Specifically, all scheduling systems integrate a unified high-precision GNSS / IMU tightly coupled navigation module and synchronize timing, transmitting multi-dimensional quality-sensing data in real time during field data acquisition and automatically triggering local re-acquisition commands. In this regard, after error modeling and geometric correction of the raw observation data output by each system, spatial alignment of multi-source data is achieved through a two-level registration mechanism. Additionally or alternatively, in some embodiments, the corrected point cloud is classified point-by-point based on a lightweight deep learning semantic segmentation model, and a hole-filling strategy is automatically selected according to preset decision rules. In some embodiments, a five-dimensional quality assessment model is constructed to automatically perform full dataset quality detection and generate a structured quality inspection report.

[0023] First, we will discuss the details of adaptive sub-region division and dynamic scheduling based on the terrain complexity index. The terrain complexity index is obtained based on the basic geographic data of the area to be measured. This index is a weighted quantitative indicator that integrates slope variability, vegetation cover homogeneity, and water space proportion. The terrain complexity index is a comprehensive quantitative evaluation of the terrain undulation, vegetation distribution uniformity, and water coverage ratio of the area to be measured. It is calculated by integrating basic geographic data such as slope standard deviation, normalized vegetation index mean, and water cell proportion, and is used to characterize the ease or difficulty of conducting surveying operations in the area. This index serves as the core driving factor for intelligent scheduling, guiding the subsequent optimal allocation of surveying resources. Specifically, the system first reads the digital elevation model and digital orthophoto of the area to be measured, extracts slope variability data to reflect the steepness and fragmentation of the terrain, extracts vegetation cover homogeneity data to assess visibility conditions, and extracts water space proportion data to identify the land-water boundary characteristics. Based on the calculated terrain complexity index, the system adaptively divides the area to be measured into several sub-regions with similar terrain attributes. For example, in sub-regions with high terrain complexity and dense vegetation cover, classified as dense forest areas, the system automatically schedules a combination of ground-based 3D laser scanning systems and backpack-mounted laser scanning systems to penetrate the vegetation and obtain the true ground features. In sub-regions with medium terrain complexity and large areas of water, classified as land-water interface areas, the system dynamically schedules UAV aerial surveying systems and underwater sonar detection systems to work collaboratively. Based on this, the system generates specific operational parameter sets for each scheduled system. These parameter sets include spatial coverage overlap rate, sensor operating altitude or depth, point cloud sampling resolution, and image exposure timing constraints. For example, for multi-rotor UAVs used in mountainous and canyon areas, the system sets a forward overlap rate of 85%, a lateral overlap rate of 75%, a flight altitude of 60 meters relative to the ground, a point cloud sampling resolution of 2 cm / pixel, and constrains the image exposure timing to avoid shadow areas. Through this adaptive classification and dynamic scheduling based on terrain complexity index, accurate matching of surveying resources and terrain features can be achieved, solving the problem of incomplete coverage by a single method.

[0024] Secondly, the details of unified high-precision GNSS / IMU tightly integrated navigation and real-time monitoring of multi-dimensional quality perception data will be discussed. All scheduling systems integrate a unified high-precision GNSS / IMU tightly integrated navigation module and synchronize timing. Permanent reference points with known three-dimensional coordinates and rapidly deployable temporary markers are deployed within the surveying area to form a global spatial reference frame. The high-precision GNSS / IMU tightly integrated navigation module can refer to a hardware unit integrating a global navigation satellite system receiver and an inertial measurement unit. It fuses satellite signals and inertial data through a tightly integrated algorithm to provide a unified spatial position and attitude reference for all scheduling systems. Synchronized timing can refer to using network time protocols or BeiDou timing signals to control the timestamp error of all acquisition devices within milliseconds, ensuring strict alignment of multi-source data in the time dimension. A global spatial reference frame is a spatial control network composed of permanent reference points with known three-dimensional coordinates and rapidly deployable temporary markers. Permanent reference points are typically measurement markers buried in stable bedrock, while temporary markers are marker boards coated with highly reflective paint or corner reflectors. The combination of these two types is used to establish high-precision coordinate transformation parameters in complex terrain. During field data acquisition, multi-dimensional quality-sensing data is transmitted back in real time. This multi-dimensional quality-sensing data includes point cloud density distribution maps, image modulation transfer function values, positioning solution confidence levels, and IMU attitude angle jitter. When any of these indicators falls below a preset threshold, a local reacquisition command is automatically triggered for the corresponding system. Multi-dimensional quality-sensing data can refer to a set of indicators reflecting data quality generated in real time during the acquisition process, including point cloud density distribution maps reflecting point cloud sparsity, image modulation transfer function values ​​reflecting image sharpness, positioning solution confidence levels reflecting positioning reliability, and IMU attitude angle jitter reflecting platform stability. The purpose of this data is to monitor the quality of field data acquisition in real time, forming a closed-loop feedback loop. Specifically, during field data acquisition, each system transmits multi-dimensional quality-sensing data back to the command center or edge computing nodes in real time via communication links. The system has preset thresholds for point cloud density, MTF value, positioning reliability, and IMU jitter. For example, the point cloud density threshold is set to 80 pts / m², and the MTF50 threshold is set to 80 lp / mm. When the transmitted data shows that the density of a local area in the point cloud density distribution map of a certain region is lower than the preset threshold, or the image MTF value indicates image blurring, or the IMU attitude angle jitter exceeds 0.3°, the system automatically determines that the data in that area is unqualified and immediately triggers a local re-acquisition command for the corresponding system, directing the UAV or ground equipment to return to that area for rework. This on-the-fly testing and inspection mechanism effectively avoids large-scale rework after quality problems are discovered during the office processing stage, significantly improving operational efficiency.

[0025] Next, we will discuss the details of the two-level registration mechanism for multi-geometric feature fusion. After error modeling and geometric correction of the raw observation data output by each system, point and linear geometric features from the image data and area and edge geometric features from the point cloud data are extracted respectively. The spatial alignment of multi-source data is then achieved through a two-level registration mechanism. Error modeling and geometric correction refer to the process of mathematically modeling and eliminating sensor system errors, atmospheric delay errors, and installation deviations present in the raw observation data, aiming to improve the geometric accuracy of the raw data. Point and linear geometric features refer to discrete point features such as corner points and center points extracted from the image data, as well as linear features such as road edges and building outlines. Area and edge geometric features refer to planar regions and sharp edges extracted from the point cloud data. The two-level registration mechanism includes two stages: coarse registration and fine registration. Its function is to unify data from different sources and coordinate systems into the same global coordinate system and eliminate local deformations. Coarse registration uses permanent reference points and temporary marker points as strong constraints, and combines cross-modal feature matching results to complete global coordinate system normalization. In the coarse registration stage, the established permanent reference points and temporary marker points serve as strong constraints. The initial transformation matrix is ​​calculated using these control points with known coordinates, and the initial normalization of the global coordinate system is completed by combining the cross-modal feature matching results. The fine registration criterion constructs a weighted objective function that integrates point-to-point distance, point-to-line distance, and point-to-plane distance. An iterative nearest geometric feature algorithm is used for local deformation optimization. The iteration termination condition of this algorithm is: the change in the objective function value between two consecutive iterations is less than 10⁻⁻⁶. 6 The maximum number of iterations can reach 50. In the fine registration stage, a weighted objective function is constructed that integrates point-to-point distance, point-to-line distance, and point-to-plane distance. This function balances the impact of various errors by assigning different weights to different geometric features. The system uses an iterative nearest geometric feature algorithm to iteratively optimize this objective function, continuously fine-tuning the transformation parameters to achieve the best fit for local deformations. Specifically, a shoreline spatial projection consistency constraint equation is introduced in the land-water boundary area, forcing the edge features of the land image to coincide with the underwater isobath features in the tidal-corrected 2D projection space. In the land-water boundary area, the system introduces a shoreline spatial projection consistency constraint equation, which considers the impact of tidal changes on the land-water boundary. After tidal correction, it forces the edge features extracted from the land image to strictly coincide with the isobath features detected by underwater sonar in the 2D projection space. For example, in the coastal area of ​​a lake, this constraint equation eliminates the stepped misalignment at the junction of the land DEM and the underwater terrain model caused by water level fluctuations, achieving seamless stitching of the land and water terrain. Two-level registration through the fusion of multiple geometric features significantly improves the spatial consistency and registration accuracy of multi-source data under complex terrain.

[0026] Subsequently, the details of point cloud classification and hole-filling strategies based on a lightweight deep learning semantic segmentation model will be discussed. The lightweight deep learning semantic segmentation model is used to classify the corrected point cloud point by point, identifying ground points, vegetation points, building points, and dynamic noise points, and then performing differentiated denoising operations based on the classification results. The lightweight deep learning semantic segmentation model can refer to a pruned and quantized neural network architecture that can run efficiently on computationally limited devices, used to perform point-by-point semantic classification of the corrected point cloud data, identifying ground points, vegetation points, building points, and dynamic noise points. Differentiated denoising operations can refer to applying different filtering strategies based on the identified point categories, such as directly removing points marked as dynamic noise, while retaining the canopy structure of vegetation points and removing only outliers. Further analysis of the spatial scale, topological morphology, and neighborhood terrain gradient characteristics of point cloud holes will be conducted, and a hole-filling strategy will be automatically selected according to preset decision rules. Point cloud holes can refer to areas where point cloud data is missing due to occlusion or low reflectivity. The preset decision rules are a logical judgment system based on the spatial scale, topological morphology, and gradient characteristics of the surrounding terrain of the cavity, used to automatically select the optimal cavity filling strategy. Specifically, the system first analyzes the geometric attributes of each cavity. If a cavity area is less than 0.5 m² and its neighborhood slope change rate is less than 5%, it indicates that the terrain in this area is gentle and the missing area is small. The system automatically uses the neighbor point weighted interpolation method to fill the cavity using the weighted average of the coordinates of surrounding points. If a cavity is detected as a narrow strip that crosses a significant abrupt change in slope aspect, it indicates that the terrain in this area changes drastically. The system uses a surface fitting method based on a quadratic polynomial trend surface to construct a surface that conforms to the terrain trend for filling. If a cavity is detected as crossing the boundary between land and water and the absolute value of the depth gradient is greater than 0.15 m / m, it indicates that this area is a steep waterfront boundary, and a single data source is difficult to accurately reconstruct. The system then integrates the texture information of UAV orthophotos and the profile data of underwater sonar for multi-source collaborative filling, using image texture to constrain the planar position and sonar data to constrain the underwater elevation. By combining semantic recognition and void feature-based decision-making, a high degree of automation in data processing is achieved, effectively solving the problems of difficult noise removal and inaccurate void filling in complex terrain, and ensuring the integrity of terrain reconstruction.

[0027] Finally, the five-dimensional quality evaluation model and output details will be discussed. A quality evaluation model covering five dimensions—planar position accuracy, elevation accuracy, point cloud spatial integrity, feature topological correctness, and image radiometric consistency—is constructed. This model automatically performs quality checks on the entire dataset, generating a structured quality inspection report containing spatial location data of non-compliant areas, error cause analysis, and correction suggestions. It also supports exporting the processed results to industry-standard vector, point cloud, raster, and engineering drawing formats. The quality evaluation model refers to a multi-dimensional quantitative evaluation index system covering the five dimensions: planar position accuracy, elevation accuracy, point cloud spatial integrity, feature topological correctness, and image radiometric consistency. The model's role is to comprehensively assess the quality and reliability of the final surveying and mapping results. The system automatically performs full dataset quality checks based on this model, comparing measured data with theoretical standards pixel by pixel or point by point. Specifically, planar position accuracy is calculated by comparing the coordinates of measured checkpoints with the coordinates of the results, with a pass rate of ≤±5cm; elevation accuracy is calculated by the elevation difference between RTK measured ground points and corresponding points in the digital elevation model, with a pass rate of ≤±3cm; point cloud spatial integrity is the ratio of the effective point cloud coverage area to the total survey area, with a pass rate of ≥99%; element topology correctness is determined by checking whether the spatial relationships between adjacent elements conform to geographic logic; and image radiometric consistency is determined by statistically analyzing the average grayscale difference in overlapping areas of adjacent images, with a pass rate of ≤5 grayscale levels. Once a non-conformity is detected, the system automatically generates a structured quality inspection report. The report not only includes the precise spatial location of the non-conforming area but also the cause of the error based on historical data and algorithm analysis, and provides targeted correction suggestions. Furthermore, the system supports exporting the processed final results into various industry-standard formats, including vector formats such as SHP, point cloud formats such as LAS, raster formats such as TIFF, and engineering drawing formats such as DWG, to meet the needs of different industry applications. Through integrated result quality checking and output, the standardization and usability of surveying and mapping results are ensured.

[0028] This application constructs a closed-loop feedback technology chain from task planning to output through the synergistic effect of the aforementioned technical solutions. The platform scheduling mechanism driven by the terrain complexity index enables UAVs, ground-based lasers, backpack systems, and underwater sonar to intelligently combine according to terrain features, solving the problem of incomplete coverage under complex terrain using a single method and achieving all-terrain mapping without blind spots. A unified high-precision GNSS / IMU navigation and multi-level quality control mechanism, combined with real-time data transmission and automatic resampling strategies, ensures high-quality source data for fieldwork and significantly reduces rework rates. A two-level registration mechanism integrating multiple geometric features, especially the shoreline projection consistency constraint at the land-water interface, effectively eliminates splicing errors between heterogeneous data, significantly improving the spatial consistency and registration accuracy of multi-source data. The data processing flow of semantic recognition and joint decision-making based on void features utilizes deep learning and multi-source fusion technology to achieve precise noise removal and adaptive filling of complex voids, improving the completeness and automation level of terrain reconstruction. Finally, a five-dimensional quality evaluation model ensures the credibility of the results. The above-mentioned links are closely coupled, which together ensure the integrity, consistency and high accuracy of the surveying and mapping results in complex terrain environments. In three typical southwestern mountain surveying and mapping projects, with a total area of ​​about 120 km², the traditional method took an average of 45 days, while this method took an average of 10 days, improving the work efficiency by about 3.5 times and significantly shortening the project cycle.

[0029] In some embodiments, this application also provides a specific method for calculating the Terrain Complexity Index (TCI). The Terrain Complexity Index (TCI) is calculated using the following formula: The Terrain Complexity Index (TCI) is a weighted index used to quantify the overall terrain complexity of the area under survey. Its value directly determines the granularity of subsequent sub-region division and the aggressiveness of multi-platform scheduling strategies. This index is calculated by integrating feature data from three dimensions: slope variability, vegetation cover homogeneity, and the proportion of water space. Specifically, The slope standard deviation represents the undulation of terrain. The larger the value, the steeper and more drastic the terrain, and the higher the requirements for the attitude stability of ground equipment and low-altitude aircraft. The normalized mean vegetation index reflects the density and homogeneity of vegetation cover within the region. The formula uses... The term "topography perception" means that the lower the vegetation cover, the greater the contribution value of this term, indicating that the difficulty of terrain perception increases when there is no vegetation or sparse vegetation, or when high-density forests cause severe signal obstruction. The percentage of pixels representing water areas is used to identify the proportion of water bodies distributed within the measured area. The presence of water areas usually necessitates the introduction of specialized detection methods such as underwater sonar, thus increasing the complexity of the operational system. Weighting coefficients. , , These are used to adjust the weights of the three factors mentioned above on the final complexity index. In this embodiment, the weight coefficients... , , It is the empirically optimal solution obtained through regression calibration using measured data from the southwestern mountainous region. For example, in a typical southwestern karst landform area, if the standard deviation of the slope of a certain sub-region... 15°, mean normalized vegetation index The percentage is 0.6, representing the proportion of pixels in the water area. If the value is 0.1, then substituting it into the formula yields: This weighted calculation method, based on measured data calibration, transforms multidimensional and heterogeneous geographical environmental features into unified numerical scalars, eliminating the incomparability between data of different dimensions. This step aims to provide an objective and quantifiable basis for subsequent adaptive sub-region division, avoiding subjective biases caused by manual judgment of terrain difficulty based on experience, and ensuring that higher-precision mapping platform scheduling can be automatically triggered in areas with abrupt terrain changes. This application constructs a closed-loop feedback technical link through platform scheduling driven by terrain complexity index, two-level registration fusion of multiple geometric features, and data processing involving semantic recognition and joint decision-making based on void features. Specifically, by introducing weight coefficients calibrated by measured data, slope variability, vegetation shading effect, and water distribution characteristics are reasonably reflected in the complexity assessment. This not only solves the problem that a single indicator cannot fully reflect the difficulty of mapping complex terrain, but also significantly improves the adaptability of the task planning module to typical complex scenarios such as the southwestern mountains. Based on this, the high-precision TCI calculation results directly guide the dynamic scheduling of multi-source sensors such as UAVs and ground laser scanning, ensuring that the set of operational parameters can be automatically optimized in areas with dense vegetation or vast water areas. This lays a solid data foundation for subsequent high-precision registration of multi-source data and intelligent cavity filling, jointly ensuring the integrity, consistency and high precision of surveying and mapping results in complex terrain environments.

[0030] In some embodiments, the method further includes a step of dynamically adjusting geometric feature weights during fine registration of multi-source data. In the iterative nearest geometric feature algorithm, matching error parameters for point features, line features, and surface features are obtained. These matching error parameters are the basic input variables for calculating the dynamic weight coefficients, specifically including the first... Residual of feature matching at each point , No. The deviation of the included angle of the characteristic direction of each line and the Individual feature normal angle deviation Point feature matching residual This can refer to the Euclidean distance between a keypoint in the source point cloud and its nearest corresponding point in the target point cloud during the current iteration step. This value reflects the spatial overlap of point features. Line feature direction angle deviation. This can refer to the angle between the extracted line segment feature vector and the corresponding reference line segment vector, used to characterize the consistency of the orientation of linear structures such as road edges and building outlines. (Surface feature normal angle deviation) This can refer to the angle between the normal vector of the locally fitted plane and the normal vector of the reference plane, used to measure the degree of shape fit of a large area of ​​terrain surface or building facade. These parameters are derived from the geometric feature matching results of the previous coarse registration or the previous ICP iteration, obtained by calculating the geometric differences between the source data and the target data in the local neighborhood. For example, when processing laser point clouds in a canyon area, if the point cloud at a rock wall crack has an excessively large matching distance due to noise, then its corresponding... The value is significantly higher than that of the surrounding smooth area. If the extracted river shoreline segment deviates from the actual shoreline by 5 degrees, then... This is the angle value. By acquiring these three types of error parameters in real time, the reliability of different geometric features in the current registration state can be quantified, providing data support for subsequent adaptive weight allocation. Based on the preset attenuation factor and matching error parameters, the dynamic weight coefficients of point features, line features, and surface features are updated according to the exponential decay rule. Among them, dynamic weighting coefficients It is calculated using an exponential function model and is used to balance the contributions of different types of geometric features when constructing a weighted objective function. The specific calculation formula is as follows: , , In the formula, , , These are preset attenuation factors for point, line, and area features, respectively. Their values ​​can be set according to terrain complexity and data noise levels, typically ranging from 0.1 to 2.0. These factors control the rate at which the weights decrease as the error increases. This mechanism enables adaptive adjustment of feature weights: when the matching error of a feature is small, the exponential term approaches 0, and the weight coefficient approaches 1, indicating that the feature has high matching quality and plays a dominant role in the optimization process. Conversely, when the matching error is large, the weight coefficient rapidly decays to near 0, thereby suppressing the interference of low-quality features or outliers on the overall registration result. For example, setting attenuation factors... If a certain point characteristic residual m, then its weight The data is fully preserved. If another area experiences a mismatch due to vegetation shading, resulting in a residual... m, then its weight The influence is significantly reduced. If the residual further expands to 2.0m, the weight will drop below 0.14, almost eliminating it from the optimization calculation. Through this dynamic update mechanism, the three features of points, lines, and surfaces automatically adjust their weights in each iteration based on their current matching performance, allowing the fine registration process to prioritize high-precision geometric constraints and effectively avoiding the problem of traditional fixed-weight ICP algorithms being susceptible to outliers and getting trapped in local optima. This application achieves collaborative optimization of point features, line features, and surface features in the fine registration process by introducing a dynamic weight update mechanism based on the exponential decay rule. , , It reflects the matching status of various geometric elements in real time and combines them with... , , The attenuation factor flexibly adjusts the weight distribution, allowing high-confidence geometric features to play a dominant role in constructing the weighted objective function, while the influence of low-confidence or noisy features is automatically suppressed. This adaptive adjustment strategy not only improves the robustness of the iterative nearest geometric feature algorithm in complex terrain environments and effectively overcomes the limitations of a single feature type in specific scenarios, but also significantly accelerates the convergence speed of the algorithm, ensuring that multi-source mapping data can still achieve millimeter-level high-precision spatial alignment in challenging areas such as land-water junctions, steep slopes, and dense forests.

[0031] In some embodiments, this application also provides a method for filling specific cavities in water-land interface areas. When the cavity crosses the water-land boundary and the absolute value of the adjacent water depth gradient is... When this occurs, multi-source fusion filling is forcibly enabled. "Crossing the land-water boundary" refers to a point cloud cavity region to be filled that spatially covers both the land surface and underwater topography, making it impossible for a single data source to fully describe the topographic continuity of the region. The neighborhood depth gradient refers to the rate of change of water depth per unit horizontal distance within the region adjacent to the cavity edge; its calculation formula is the ratio of the depth difference to the horizontal distance, used to quantify the steepness of the land-water transition zone. (Settings are not specified in the original text.) The threshold is an empirical critical value derived from a large amount of measured data. When the water depth gradient exceeds this value, it indicates that there are drastic topographic changes in the area, such as steep banks, embankments, or scour channels. If conventional neighbor-point weighted interpolation or trend surface fitting is used, it is easy to produce step effects or geometric distortions caused by smooth transitions in the terrain. Therefore, the system calculates the gradient characteristics of the cavity's neighborhood in real time. Once the above two conditions are met, the system automatically locks the area and forces a switch to multi-source fusion filling mode, no longer executing the conventional interpolation strategy. For example, in a river embankment repair surveying project, there is a narrow cavity at the toe of the embankment formed by water erosion. The water depth in its neighborhood changes drastically from 0.2m to 1.8m, with a horizontal span of only 8m. The calculated water depth gradient is about 0.2m / m, exceeding the preset threshold of 0.15m / m. The system then determines that multi-source fusion filling needs to be activated. Through this dynamic decision-making mechanism based on topographic gradient, high-risk and complex areas can be accurately identified, avoiding systematic errors introduced by faulty algorithms. Multi-source fusion filling specifically involves: using the shoreline extracted by the UAV DOM as a constraint, projecting sonar profile data along the normal direction onto the DOM plane, and constructing a Poisson reconstruction equation with terrain constraints to solve for the elevation field of the void region. Here, the UAV DOM can refer to an orthorectified digital orthophoto map, which has a unified planar coordinate system and high-precision geographic positioning information. The shoreline can refer to the land-water boundary vector data automatically extracted from the DOM using image processing algorithms, representing the precise geometric boundary of the land-water medium on a two-dimensional plane. In practice, the high-resolution image acquired by the UAV is first semantically segmented or edge-detected to extract shoreline features with sub-meter precision, which are then used as strong geometric constraints for subsequent data fusion to correct for potential lateral positional deviations in the underwater data. The sonar profile data can refer to a discrete set of bottom depth points collected by an underwater sonar detection system, typically distributed on a profile perpendicular to the navigation path. Projection along the normal direction refers to calculating the perpendicular line from the extracted shoreline to the sonar data point and mapping the three-dimensional coordinates of the sonar point to the two-dimensional reference plane of the DOM. This eliminates nonlinear distortions caused by water flow refraction, attitude jitter, or coordinate system transformation, ensuring strict alignment between underwater topographic data and land image data in planar position. Based on this, the Poisson reconstruction equation with topographic constraints is a mathematical model for surface reconstruction based on partial differential equations. Its core idea is to transform the problem of restoring the topographic elevation field into solving the Dirichlet boundary value problem of the Poisson equation. Specifically, the elevations of the projected and aligned shoreline points and sonar profile points are used as known boundary conditions and internal constraint points of the equation. A gradient field under the Laplace operator is constructed, and the elevation values ​​of unknown points within the cavity region are solved by minimizing the error between the reconstructed surface and the observed gradient.The advantage of this method lies in its utilization of discrete observation point data and the introduction of the shoreline as a continuous geometric constraint. This ensures that the reconstructed terrain surface maintains the continuity of its first derivative when crossing the land-water boundary, effectively avoiding the breaks or unnatural undulations caused by abrupt changes in the medium in traditional interpolation methods. For example, when dealing with the aforementioned erosion cavities in the dam, the system sets the extracted shoreline elevation as a fixed boundary value and the projected sonar depth data as an internal constraint. The filling surface generated after solving the Poisson equation not only perfectly matches the shape of the steep slope at the dam toe but also smoothly extends into the deep water area, eliminating obvious splicing marks. Through the synergistic effect of shoreline geometric constraints and sonar depth data, as well as the global optimization characteristics of the Poisson equation, high-fidelity reconstruction of complex terrain at the land-water boundary is achieved. This application significantly improves the geometric consistency and integrity of the mapping results in the land-water boundary area through the synergistic effect of the above-mentioned technical features. Specifically, by setting a water depth gradient threshold as a trigger condition, the system can intelligently identify areas of drastic terrain change that are difficult for conventional algorithms to handle, preventing terrain distortion caused by blind interpolation. Furthermore, by utilizing the high-precision shoreline extracted by UAV DOM as a spatial reference, the coordinate system misalignment problem caused by sensor differences and different media between underwater sonar data and land image data was resolved. The projection strategy along the normal direction ensured the physical rationality of the data mapping. Finally, by employing a Poisson reconstruction equation with terrain constraints, discrete sonar points were organically integrated with continuous shoreline constraints, mathematically guaranteeing the smoothness and continuity of the reconstructed surface at the boundaries. This deep fusion mechanism of multi-source data effectively overcomes the limitations of single data sources in the water-land transition zone, generating a unified elevation field that conforms to both terrestrial geomorphological characteristics and underwater topographical patterns. This provides high-precision, seamless basic geographic data support for water conservancy and flood control, waterway management, and ecological environment monitoring.

[0032] In some embodiments, the method further includes a feedback update step after automatically triggering a local reacquisition command. After automatically triggering the local reacquisition command, the system synchronously pushes the reacquisition area coordinates, the original quality problem type, and historical quality inspection records to the task planning module. The reacquisition area coordinates refer to the geographic spatial range marked as unqualified during field acquisition because any indicator in the multidimensional quality perception data is below a preset threshold. It is typically represented by a latitude and longitude bounding box or a polygon vector, used to accurately locate the area requiring rework. The original quality problem type is the specific reason classification that triggered the reacquisition command. For example, when the point cloud density distribution map shows that the number of points in a certain area is less than 80 pts / m², the problem type is identified as sparse point cloud. When the image modulation transfer function MTF50 is less than 80 lp / mm, it is identified as image blurring. When the positioning solution confidence is less than 95%, it is identified as positioning loss. Historical quality inspection records contain the quality fluctuation trend of the sub-area in past work cycles, the types of defects that have occurred, and the evaluation data of the effectiveness of corrective measures. The system packages the above three types of information into structured data packets via an internal communication bus or wireless network and transmits them to the task planning module in real time. For example, if a UAV operating in a canyon area experiences IMU pitch angle jitter exceeding 0.3° due to signal obstruction, triggering a retest, the system immediately pushes the coordinates of that canyon segment, the problem label of the IMU attitude anomaly, and the attitude jitter statistics from the past three times in that area to the planning end. This real-time push mechanism ensures that quality anomalies in the field can be transformed into input for office decision-making without delay, breaking down the information barriers between field and office operations in traditional operations. This is used to update the terrain complexity index calculation parameters for that sub-area and subsequent platform scheduling strategies. Specifically, the task planning module dynamically adjusts the weight coefficients in the TCI calculation formula using a weighted moving average method based on the frequency of various quality problems in historical quality inspection records. The single adjustment step size does not exceed 0.05, and the sum of the weight coefficients always remains at 1. A quality evaluation model covering five dimensions—planar position accuracy, elevation accuracy, point cloud spatial integrity, feature topology correctness, and image radiometric consistency—is constructed.

[0033] Updating the terrain complexity index calculation parameters for this sub-region can refer to the task planning module dynamically adjusting the weighting coefficients or threshold settings when calculating the terrain complexity index after receiving feedback information. Specifically, if historical quality inspection records show that a certain sub-region frequently experiences point cloud gaps due to vegetation cover, the system can automatically increase the weighting coefficient of vegetation cover homogeneity in the TCI calculation formula. Alternatively, the NDVI threshold triggering high-density scanning may be lowered, thus identifying the area as having a higher complexity level in subsequent assessments. Subsequent platform scheduling strategies can refer to a regenerated mapping resource configuration scheme based on the updated complexity index and the current quality problem type. This includes changing the mapping platform combination, adjusting flight altitude, increasing flight path overlap, or switching sensor operating modes. For example, regarding the aforementioned attitude jitter problem caused by signal obstruction, the mission planning module, when updating its strategy, might change the scheduling scheme for this sub-area from a single fixed-wing UAV to a multi-rotor UAV combined with ground-based 3D laser scanning, and mandate the deployment of more temporary markers in the area to enhance the constraints of tightly integrated GNSS / IMU navigation. Alternatively, the spatial coverage overlap rate could be increased from the usual 60% to 80% to compensate for quality fluctuations on individual flight paths through redundant data. This linkage between parameter and strategy updates allows the system to adaptively optimize its response logic to complex terrain based on real-time feedback learning results. This result provides a validated optimization path for subsequent repetitive operations, effectively preventing the recurrence of similar quality problems, significantly improving the accuracy and foresight of resource scheduling, and demonstrating the closed-loop feedback characteristics of continuous learning and self-evolution in surveying and mapping systems. This application achieves a complete technical closed loop of problem identification, local correction, model learning, and strategy optimization through an automatic triggering mechanism for local re-acquisition commands and a dynamic update mechanism for the task planning module. Specifically, the system uses the coordinates of the re-acquisition area, the original quality problem type, and historical quality inspection records as key feedback variables, not only resolving local defects in a single operation but also transforming these discrete quality events into training samples for the global model. By updating the terrain complexity index calculation parameters, the system can more accurately quantify the actual operational difficulty of a specific area, correcting the bias of relying solely on static evaluation based on basic geographic data. Based on this, the updated parameters guide subsequent platform scheduling strategies, enabling the allocation of surveying and mapping resources to shift from passive response to proactive defense, i.e., mitigating potential risks before problems occur by adjusting platform combinations or operational parameters. This synergy enables the surveying and mapping process to have memory and evolution capabilities. As operational data accumulates, the system's adaptation strategies to complex terrain will become increasingly precise, thereby significantly reducing the field rework rate in long-term operation and ensuring the consistency and high accuracy of surveying and mapping results.

[0034] In some embodiments, the method further includes specifically defining the performance parameters of the unified high-precision GNSS / IMU tightly integrated navigation module. In the unified high-precision GNSS / IMU tightly integrated navigation module, the GNSS receiver positioning accuracy is better than ±1 cm (RTK mode), the IMU gyroscope zero-bias instability is ≤0.5° / h, and the time synchronization error between the two is <1 ms. The unified high-precision GNSS / IMU tightly integrated navigation module can refer to a core positioning and attitude determination unit integrated into an UAV aerial survey system, a ground-based 3D laser scanning system, a backpack laser scanning system, and an underwater sonar detection system, used to provide a unified spatial position reference and attitude angle information for all field data acquisition equipment. The GNSS receiver in this module is specifically configured as a multi-frequency, multi-system receiver supporting RTK mode. Its positioning accuracy is constrained to within ±1 cm using carrier phase differential technology. This accuracy is achieved by accessing permanent reference points or network CORS stations with known coordinates, calculating integer ambiguity in real time, and eliminating atmospheric delay errors. Its function is to provide the absolute position truth in open areas and environments with good signal, serving as a global anchor point for spatial alignment of multi-source data. For example, when operating in the southwestern mountainous surveying sub-region, when the UAV flies over open valleys, and the GNSS receiver locks onto at least 5 satellites with a PDOP value less than 2.0, its real-time output latitude, longitude, and elevation coordinate errors remain within ±0.8 cm, thus ensuring that the generated orthophoto plane position accuracy meets the requirements of 1:500 mapping. The IMU specifically comprises a three-axis fiber optic gyroscope and a three-axis quartz accelerometer. Its core performance indicator, gyroscope bias instability, is limited to ≤0.5° / h. This indicator characterizes the degree of drift of the gyroscope's output angular velocity over time under isothermal and static conditions. It is obtained through long-term temperature drift testing and Allan variance analysis calibration before shipment. Its function is to maintain high-precision attitude angle calculations through inertial estimation during the short period of positioning failure caused by GNSS signal obstruction, preventing trajectory divergence. For example, when a ground 3D laser scanning system enters a dense jungle area and the GNSS signal is lost for approximately 30 seconds, thanks to the highly stable IMU with a bias instability of 0.3° / h, the system's heading angle drift in pure inertial navigation mode is controlled within 0.004°, effectively avoiding layering or distortion of point cloud data caused by attitude estimation errors. The time synchronization error between the GNSS receiver and the IMU can refer to the alignment deviation of the internal clock stamps of the two. This error is strictly controlled to <1 ms through a hardware pulse triggering and serial time message joint calibration mechanism. Its function is to ensure that the position information and attitude information acquired by the high-speed motion platform in an instant correspond strictly in the spatiotemporal domain, and to eliminate the lever arm effect error caused by time asynchrony.For example, when a multi-rotor UAV flies at 10 m / s and performs oblique photography, if the time synchronization error is 0.5 ms, the resulting position projection error is only 5 mm, far less than the system's allowable positioning tolerance, thus ensuring accurate matching between image exposure time and POS data recording. The high-precision absolute positioning capability of the GNSS receiver and the high-stability relative attitude estimation capability of the IMU work together to form the core complementary mechanism of tightly integrated navigation: GNSS is used to correct the cumulative drift of the IMU, and the IMU is used to fill the signal blind spots of GNSS and smooth high-frequency vibration noise. The microsecond-level time synchronization is the data foundation for the deep integration of the two, ensuring the timing consistency of the Kalman filter algorithm when processing measurement updates. Through the synergistic constraints of this high-precision hardware specification, a global spatial reference frame with centimeter-level accuracy can be constructed in complex terrain environments, enabling subsequent coarse registration based on permanent reference points and temporary markers, as well as fine registration based on multiple geometric features, to have reliable initial convergence conditions, significantly reducing the overall stitching deviation caused by underlying positioning errors. This application establishes a highly robust underlying spatiotemporal reference by limiting GNSS receiver positioning accuracy to better than ±1 cm, IMU gyroscope zero-bias instability to ≤0.5° / h, and time synchronization error to <1 ms. Based on this, high-precision GNSS provides an irreplaceable absolute position correction source for scenarios common in complex terrain, such as signal obstruction and severe vibration, effectively suppressing accumulated errors in long-distance operations. The highly stable IMU acts as an inertial bridge in weak signal environments, ensuring the continuity and temporal integrity of data acquisition. The stringent time synchronization parameters further eliminate time misalignment between multiple sensors, making multi-source data fusion possible under high-speed dynamic conditions. The combined effect of these performance parameters not only solves the problem of difficult multi-platform data registration caused by positioning drift and time asynchrony in traditional surveying, but also provides a precise input data foundation for subsequent two-level registration mechanisms, semantic segmentation denoising, and hole filling, thus fundamentally ensuring the high accuracy and consistency of the final surveying results in terms of planar position, elevation, and topological relationships.

[0035] In some embodiments, the method further includes a specific implementation of point-by-point classification of the corrected point cloud. An improved PointNet++ architecture is adopted as a lightweight deep learning semantic segmentation model. The improved PointNet++ architecture refers to a neural network model that is specifically optimized based on the original PointNet++ network structure. It retains the ability of the original architecture to extract local and global features of the point cloud layer by layer, while reducing the number of parameters and computational complexity through pruning, quantization, or structural reorganization to meet the real-time or near-real-time processing requirements in complex terrain mapping scenarios. The model specifically includes an input layer, several sets of ensemble abstraction modules, a feature propagation module, and an output classification layer. A channel attention module, SE Block, is introduced into the backbone network. SE Block is a squeeze-and-excitation mechanism used to adaptively recalibrate the channel feature responses. Specifically, SE Block first compresses spatial dimensionality information through global average pooling to obtain a global descriptor for each channel. Then, it learns the nonlinear interaction relationship between channels through a bottleneck structure composed of two fully connected layers, generating weight coefficients for each channel. Finally, the weight coefficients are applied to the original feature map to enhance the response of key feature channels and suppress redundant or noisy channels. For example, when processing point cloud data containing a mixture of dense vegetation and exposed rocks, SE Block can automatically increase the weights of local curvature and RGB texture channels, as these two types of features have higher discriminative power in distinguishing leaf surfaces from rock surfaces, while reducing the weight interference of the laser reflection intensity channel under specific lighting conditions. Through this channel attention mechanism and deep fusion with PointNet++ hierarchical features, the model can more sensitively capture subtle differences in ground features within complex terrain. The model's input features are constructed to include point cloud 3D coordinates, laser reflection intensity, local curvature, and corresponding image RGB texture values. The input features are multi-dimensional data vectors that drive the semantic segmentation model to make classification decisions. The point cloud 3D coordinates provide spatial geometric information about ground features, serving as the basis for distinguishing ground undulations, building vertical surfaces, and vegetation canopy height. Laser reflection intensity records the energy value returned by the laser pulse, reflecting the material properties and roughness of the ground surface; for example, water bodies typically exhibit low reflection intensity, while metal or smooth surfaces exhibit high reflection intensity. Local curvature describes the degree of surface curvature within the point cloud neighborhood and is used to identify edges, corners, and smooth areas. It is crucial for extracting linear features such as road boundaries and roof ridges. The corresponding RGB texture values ​​map the color information of the high-resolution optical image onto each point in the point cloud, supplementing the discriminative information in the spectral dimension and effectively solving the problem that a single geometric feature cannot distinguish between similar-colored but different-category ground features. These four types of features undergo normalization and alignment before being input into the model, forming a unified multimodal feature tensor.For example, for a point located at the boundary between land and water, its 3D coordinates show abrupt elevation changes, reflectance intensity shows low energy absorption, local curvature shows gentle water surface characteristics, and RGB texture values ​​exhibit a deep blue hue. The joint input of multi-source features enables the model to accurately classify it as a water point rather than a shadow noise point. The point cloud 3D coordinates, laser reflectance intensity, local curvature, and corresponding image RGB texture values ​​are used in combination. Through the complementarity and enhancement of multimodal information, the model's classification robustness under complex conditions such as occlusion, shadows, and heterogeneous spectrophotometry is significantly improved. This application achieves a balance between lightweight design and high accuracy by adopting an improved PointNet++ architecture and combining it with a channel attention module. On this basis, a comprehensive point cloud feature representation system is constructed by using multi-source fusion input of point cloud 3D coordinates, laser reflectance intensity, local curvature, and corresponding image RGB texture values. The SE Block dynamically adjusts the contribution of each feature channel, enabling the model to adaptively focus on the most discriminative feature combination when facing diverse land cover types in complex terrain. The synergy between this architectural design and feature engineering not only effectively overcomes the shortcomings of traditional methods that rely on manual rules and have poor generalization ability, but also significantly improves the accuracy of identifying ground points, vegetation points, building points and dynamic noise points. This provides a reliable data foundation for subsequent differentiated denoising operations and intelligent decision-making on hole filling strategies, thereby ensuring the integrity and consistency of the final surveying and mapping results.

[0036] In some embodiments, the method further includes setting specific quantitative judgment criteria for multidimensional quality-sensing data to achieve real-time automated control of surveying and mapping operation quality. Point cloud density < 80 pts / m², image modulation transfer function MTF50 < 80 lp / mm, positioning solution confidence < 95%, and IMU pitch angle jitter > 0.3° are all considered thresholds; exceeding any of these limits triggers a retest. The preset thresholds are a set of critical values ​​used by the system to determine whether field-collected data is qualified, covering four core dimensions: spatial coverage, imaging clarity, positioning reliability, and platform stability. Specifically, the point cloud density threshold measures the number of effective laser echo points per unit area. When the real-time transmitted point cloud density distribution map shows that the number of points in a certain area is less than 80 pts / m², it indicates that the terrain details in that area are insufficiently sampled, making it impossible to construct a high-precision digital surface model. The image modulation transfer function threshold is used to evaluate the imaging sharpness of the optical system. The MTF50 value reflects the spatial frequency at which the contrast in the image drops to 50%. If this value is below 80 lp / mm, it indicates that the image is severely blurred or out of focus, which will lead to the failure of subsequent feature point extraction. The positioning solution confidence threshold characterizes the reliability of the GNSS / IMU tightly integrated navigation solution results. When the confidence is below 95%, it means that the current satellite geometry is poor or the signal is blocked, and the generated coordinate data has a large risk of error. The IMU pitch angle jitter threshold is used to monitor the attitude stability of the UAV or mobile measurement platform. If the pitch angle jitter exceeds 0.3°, it usually indicates that the platform is under strong wind disturbance or mechanical vibration, which can easily cause point cloud layering or image motion blur. For example, when operating in the mountainous canyons of Southwest China, if the UAV encounters a sudden gust of wind that causes the IMU pitch angle jitter to reach 0.45°, the system immediately determines that the data frame is unqualified and marks it as a re-sampling area. For example, in underwater sonar detection, if the echo is sparse due to water turbidity, the calculated equivalent point cloud density is only 65 pts / m², triggering a re-measurement command. Through parallel monitoring and logical OR judgment of the above four indicators, the system can accurately identify various inferior data sources and prevent invalid data from entering the subsequent processing flow. This step aims to establish an objective and quantifiable quality access mechanism. Based on the aforementioned multi-dimensional quality perception data, the detection system generates a local re-acquisition command by comparing the deviation of each indicator with the preset threshold in real time. This significantly improves the first-time success rate of field data acquisition, effectively avoids rework in the office due to data quality problems, and ensures the integrity and consistency of surveying and mapping results in complex terrain environments. This application constructs a comprehensive quality monitoring closed loop by setting preset thresholds for four key indicators: point cloud density, image MTF50, positioning solution confidence, and IMU jitter. The point cloud density threshold and the image MTF50 threshold work together to ensure the geometric and texture quality of the original data from the perspectives of spatial sampling rate and frequency domain sharpness, respectively, and to prevent information loss caused by terrain occlusion or equipment vibration.The positioning confidence threshold and IMU jitter threshold work synergistically. The former ensures positional accuracy from the perspective of global coordinate system constraints, while the latter avoids motion distortion from the perspective of local attitude stability. Together, they ensure the uniformity and reliability of the spatiotemporal reference of multi-source data. Based on this, a linkage mechanism that triggers retesting when any index exceeds its limit allows the system to detect and correct quality problems in real time at the data acquisition site. This shifts traditional post-event quality inspection to in-process control, significantly reducing the labor intensity of manually screening for inferior data and eliminating interference from low-quality data to subsequent two-level registration and semantic segmentation processing. Ultimately, this achieves high efficiency and intelligence in complex terrain mapping operations.

Claims

1. A comprehensive surveying and data processing method for complex terrain, characterized in that, Includes the following steps: The terrain complexity index is obtained based on the basic geographic data of the area to be measured. The terrain complexity index is a preset weighted quantitative index that integrates slope variability, vegetation cover homogeneity and water space ratio. Based on the index, the area is adaptively divided into sub-regions. For the terrain attribute characteristics of each sub-region, one or more of the following systems are dynamically scheduled: UAV aerial survey system, ground 3D laser scanning system, backpack laser scanning system and underwater sonar detection system. The corresponding operation parameter set for each system is generated. The operation parameter set includes at least spatial coverage overlap rate, sensor working height / depth, point cloud sampling resolution and image exposure time sequence constraints. All scheduling systems integrate a unified high-precision GNSS / IMU tightly coupled navigation module and synchronize timing. Permanent reference points with known three-dimensional coordinates and rapidly deployable temporary marker points are deployed within the survey area to form a global spatial reference frame. During field data acquisition, multi-dimensional quality sensing data is transmitted back in real time. The multi-dimensional quality sensing data includes point cloud density distribution map, image modulation transfer function value, positioning solution confidence, and IMU attitude angle jitter. When any indicator falls below a preset threshold, a local reacquisition command is automatically triggered for the corresponding system. After error modeling and geometric correction of the raw observation data output by each system, point and line geometric features in the image data and area and edge geometric features in the point cloud data are extracted respectively. Spatial alignment of multi-source data is achieved through a two-level registration mechanism: coarse registration uses permanent reference points and temporary marker points as strong constraints, and completes global coordinate system normalization by combining cross-modal feature matching results; fine registration criteria construct a weighted objective function that integrates point-to-point distance, point-to-line distance and point-to-plane distance, and uses an iterative nearest geometric feature algorithm to optimize local deformation. In the water-land interface area, a shoreline spatial projection consistency constraint equation is introduced to force the edge features of the land image and the underwater isobath features to coincide in the two-dimensional projection space after tide level correction. A lightweight deep learning semantic segmentation model is used to classify the corrected point cloud point by point, identifying ground points, vegetation points, building points, and dynamic noise points. Denoising operations are performed differently based on the classification results. The spatial scale, topological morphology, and neighborhood terrain gradient characteristics of point cloud holes are further analyzed, and hole filling strategies are automatically selected according to preset decision rules: for holes with an area of ​​less than 0.5 m² and a neighborhood slope change rate of less than 5%, neighbor point weighted interpolation is used; for holes that are long and narrow and cross significant slope aspect abrupt change zones, surface fitting based on quadratic polynomial trend surfaces is used; for holes that cross the boundary between land and water media and have a depth gradient greater than 0.15 m / m, multi-source collaborative filling is performed by fusing texture information from UAV orthophoto images and underwater sonar profile data. Construct a quality evaluation model covering five dimensions: planar position accuracy, elevation accuracy, point cloud spatial integrity, feature topology correctness, and image radiometric consistency. Automatically perform quality inspection on the entire dataset, generate a structured quality inspection report containing spatial positioning of unqualified areas, error cause analysis, and correction suggestions, and support exporting the processed results into vector format, point cloud format, raster format, and engineering drawing format that conform to industry standards. Among them, the platform scheduling driven by the terrain complexity index, the two-level registration of multi-geometric feature fusion, and the data processing of semantic recognition and joint decision-making of void features constitute a closed-loop feedback technical link, which together ensures the integrity, consistency and high accuracy of surveying and mapping results in complex terrain environments.

2. The integrated mapping and data processing method for complex terrain according to claim 1, wherein, The Terrain Complexity Index (TCI) is calculated using the following formula: in, For the standard deviation of slope, The mean of the normalized vegetation index. The percentage of pixels in the water area. , , These are preset weighting coefficients.

3. The integrated surveying and data processing method for complex terrain according to claim 1, characterized in that, In the iterative nearest geometric feature algorithm, the dynamic weight coefficients of point features, line features, and surface features... Update according to the following rules: ， ， ， wherein is the point feature matching residual, is the line feature direction angle deviation, is the face feature normal angle deviation, is a preset attenuation factor.

4. The integrated surveying and data processing method for complex terrain according to claim 1, characterized in that, When the cavity crosses the boundary between land and water and the absolute value of the water depth gradient in the vicinity is... When this occurs, multi-source fusion filling is forcibly enabled; the multi-source fusion filling specifically involves: using the shoreline extracted by the UAV DOM as a constraint, projecting the sonar profile data along the normal direction onto the DOM plane, and constructing a Poisson reconstruction equation with terrain constraints to solve the elevation field of the cavity region.

5. The integrated mapping and data processing method for complex terrain according to claim 1, wherein, After the automatic triggering of the local re-acquisition command, the system synchronously pushes the coordinates of the re-acquisition area, the original quality problem type, and historical quality inspection records to the task planning module, which is used to update the terrain complexity index calculation parameters of the sub-area and the subsequent platform scheduling strategy.

6. The method of claim 1, wherein, In the unified high-precision GNSS / IMU tightly coupled navigation module, the GNSS receiver positioning accuracy is better than ±1cm, the IMU gyroscope zero-bias instability is ≤0.5° / h, and the time synchronization error between the two is <1ms.

7. The method of claim 1, wherein, The lightweight deep learning semantic segmentation model is an improved PointNet++ architecture, whose backbone network introduces a channel attention module. The input features include point cloud 3D coordinates, laser reflection intensity, local curvature, and corresponding image RGB texture values.

8. The method of claim 1, wherein, The preset thresholds include: point cloud density < 80 pts / m², image modulation transfer function MTF50 < 80 lp / mm, positioning solution confidence < 95%, and IMU pitch jitter > 0.3°. Exceeding any of these limits will trigger a retest.

9. The integrated mapping and data processing method for complex terrain according to claim 2, wherein, The preset weighting coefficient =0.4、 =0.35、 =0.25, obtained through regression calibration using measured data from the southwestern mountainous region.

Images