A method and system for geographic mapping based on unmanned aerial vehicles

Abstract

This invention discloses a UAV-based geographic mapping method and system. The method generates a layered scanning path for the UAV based on the geographic location information to be collected, and acquires LiDAR UAV point cloud data and UAV image data according to the scanning path. The point cloud data undergoes preprocessing, and an improved MLS algorithm is used to filter the preprocessed point cloud data. A dual-channel geographic information recognition model for terrain feature extraction and semantic segmentation of ground features is established based on a 3D-CNN convolutional neural network. The filtered UAV point cloud data is input into the recognition model for feature fusion. An improved NeRF neural radiation field framework is used to register and model the fused point cloud data and UAV image data to generate a 3D geographic reality model. This significantly improves the model's accuracy in recognizing terrain features and ground feature categories, enabling more accurate extraction of geographic information.

Description

Technical Field

[0001] This invention relates to the field of geographic surveying and mapping technology, and in particular to a geographic surveying and mapping method and system based on unmanned aerial vehicles (UAVs). Background Technology

[0002] Traditional geographic surveying methods, such as terrestrial surveying and satellite remote sensing, have certain limitations. Terrestrial surveying relies on manual field operations, which is inefficient, labor-intensive, and difficult to conduct in complex terrain or dangerous areas; while satellite remote sensing has a wide coverage area, its resolution is limited and cannot meet the needs of high-precision surveying. Summary of the Invention

[0003] The purpose of this invention is to solve the above-mentioned problems by designing a geographic mapping method and system based on unmanned aerial vehicles (UAVs).

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned UAV-based geographic mapping method:

[0005] Acquire geographic location information to be collected, generate a UAV layered scanning path based on the geographic location information to be collected, and collect LiDAR UAV point cloud data and UAV image data according to the UAV layered scanning path;

[0006] The LiDAR UAV point cloud data is preprocessed, and the preprocessed point cloud data is filtered using an improved MLS algorithm to obtain filtered UAV point cloud data.

[0007] A 3D-CNN dual-channel geographic information recognition model for terrain feature extraction and semantic segmentation is established based on a 3D-CNN convolutional neural network. The hyperparameters of the recognition model are optimized using the improved IWOA whale optimization algorithm to obtain the 3D-CNN-IWOA dual-channel geographic information recognition model.

[0008] The filtered UAV point cloud data is input into the 3D-CNN-IWOA dual-channel geographic information recognition model for feature fusion to obtain fused point cloud data.

[0009] The fused point cloud data and UAV image data are registered and modeled using the improved NeRF neural radiation field framework to generate a 3D geographic reality model.

[0010] Furthermore, in the aforementioned UAV-based geographic mapping method, the steps of acquiring the geographic location information to be collected, generating a UAV layered scanning path based on the geographic location information to be collected, and collecting LiDAR UAV point cloud data and UAV image data according to the UAV layered scanning path include:

[0011] The terrain relief of the area to be collected is calculated based on DEM data. The terrain is divided into plains, mountains and high mountains. Different layer heights and densities are set for different terrain types.

[0012] The UAV flight path is designed based on the layer height and density, and the scanning width and overlap rate of the UAV LiDAR sensor are set according to the spacing between adjacent flight paths to obtain the UAV layered scanning path.

[0013] Furthermore, in the aforementioned UAV-based geographic mapping method, the step of preprocessing the LiDAR UAV point cloud data and using an improved MLS algorithm to filter the preprocessed point cloud data to obtain filtered UAV point cloud data includes:

[0014] The local mean and standard deviation of each point in the LiDAR drone point cloud data are calculated using statistical filtering. Points that deviate from the mean by more than a certain multiple of the standard deviation are considered outliers. A radius threshold is set, and the number of points within that radius is counted. Points with fewer than the set threshold are considered outliers.

[0015] Deleting outliers from the LiDAR drone point cloud data yields the initial drone point cloud data;

[0016] The initial UAV point cloud data is converted into standard LAS format to obtain standard UAV point cloud data;

[0017] The POS data in the standard UAV point cloud data is converted into absolute geographic coordinates, error compensation is performed using ground control points, and the coordinate transformation parameters are calculated using the least squares method to obtain the processed UAV point cloud data.

[0018] Furthermore, in the aforementioned UAV-based geographic mapping method, the step of preprocessing the LiDAR UAV point cloud data and using an improved MLS algorithm to filter the preprocessed point cloud data to obtain filtered UAV point cloud data further includes:

[0019] An adaptive weight function is introduced into the MLS algorithm to adjust the weights according to the local density and curvature changes of the point cloud, resulting in an improved MLS algorithm.

[0020] An improved MLS algorithm is used to process UAV point cloud data into blocks, and the block size is determined based on the point cloud density and computing resources.

[0021] Each point cloud data piece is filtered, a local surface is fitted, the distance from each point to the fitted surface is calculated, points whose distance exceeds a set threshold are regarded as ground features, and points whose distance is within the threshold are retained as ground points, thus obtaining filtered UAV point cloud data.

[0022] Furthermore, in the aforementioned UAV-based geographic mapping method, the 3D-CNN dual-channel geographic information recognition model, which establishes terrain feature extraction and semantic segmentation based on a 3D-CNN convolutional neural network, and optimizes the hyperparameters of the recognition model using an improved IWOA whale optimization algorithm to obtain a 3D-CNN-IWOA dual-channel geographic information recognition model, further includes:

[0023] An adaptive step size factor is introduced into the spiral update position strategy of the whale optimization algorithm. The step size is dynamically adjusted according to the current iteration number and the fitness value of the population, resulting in the improved IWOA whale optimization algorithm.

[0024] Obtain the model's hyperparameters, including at least the learning rate, batch size, number of convolutional kernels, number of neurons in fully connected layers, and dropout rate;

[0025] The fitness function is a weighted sum of the model’s classification accuracy and segmentation accuracy on the validation set, and the weights are adjusted according to project requirements.

[0026] The population size is set to 50-100, the maximum number of iterations is 200-300, and the position and fitness value of the whales are updated after each iteration to obtain the optimal combination of hyperparameters.

[0027] Furthermore, in the aforementioned UAV-based geographic mapping method, the step of inputting the filtered UAV point cloud data into the 3D-CNN-IWOA dual-channel geographic information recognition model for feature fusion to obtain fused point cloud data further includes:

[0028] The filtered UAV point cloud data is converted into point cloud feature descriptors, and terrain features and semantic features of ground objects are fused in the intermediate layer of the 3D-CNN model.

[0029] Feature extraction is performed using convolutional and pooling layers of a 3D-CNN model to obtain fused point cloud data.

[0030] Furthermore, in the aforementioned UAV-based geographic mapping method, the step of registering and modeling the fused point cloud data and UAV image data using the improved NeRF neural radiation field framework to generate a 3D geographic reality model further includes:

[0031] Based on the volumetric radiation field representation of NeRF, the reflection intensity and normal vector information of point cloud are introduced as additional input features. The AdamW optimization algorithm is used instead of the Adam algorithm, and weight decay is added during the optimization process.

[0032] The scene is sampled using a hierarchical sampling strategy. Sampling is performed at a coarse-grained level to quickly converge to the scene structure, and dense sampling is performed at a fine-grained level.

[0033] The color and density of each sampling point are predicted by the radiation field function, the loss function with the real image is calculated, and the model parameters are updated by the backpropagation algorithm.

[0034] New images are synthesized from different perspectives using a volumetric rendering algorithm to generate a 3D model of the geographic scene.

[0035] A UAV-based geographic mapping system, comprising the following modules:

[0036] The UAV path planning module is used to acquire geographic location information to be collected, generate a UAV layered scanning path based on the geographic location information to be collected, and collect LiDAR UAV point cloud data and UAV image data according to the UAV layered scanning path;

[0037] The mapping data acquisition module is used to preprocess the LiDAR UAV point cloud data, and to perform point cloud filtering on the preprocessed point cloud data using an improved MLS algorithm to obtain filtered UAV point cloud data.

[0038] The recognition model building module is used to build a 3D-CNN dual-channel geographic information recognition model based on 3D-CNN convolutional neural network for terrain feature extraction and semantic segmentation of ground objects. The hyperparameters of the recognition model are optimized by using the improved IWOA whale optimization algorithm to obtain the 3D-CNN-IWOA dual-channel geographic information recognition model.

[0039] The mapping data fusion module is used to input the filtered UAV point cloud data into the 3D-CNN-IWOA dual-channel geographic information recognition model for feature fusion to obtain fused point cloud data.

[0040] The 3D model building module is used to register and model the fused point cloud data and UAV image data using the improved NeRF neural radiation field framework to generate a geographic reality 3D model.

[0041] Furthermore, in a UAV-based geographic mapping system, the UAV path planning module includes the following sub-modules:

[0042] The sub-module is used to calculate the topographic relief of the area to be collected based on DEM data, and divide the terrain into plains, mountains and high mountains, and set different layer heights and densities for different terrain types;

[0043] A submodule is obtained for designing UAV flight paths based on the layer height and density, setting the scanning width and overlap rate of the UAV LiDAR sensor according to the spacing between adjacent flight paths, and obtaining the UAV layered scanning path.

[0044] Furthermore, in a UAV-based geographic mapping system, the mapping data fusion module includes the following sub-modules:

[0045] The transformation submodule is used to convert the filtered UAV point cloud data into point cloud feature descriptors, and fuse terrain features and semantic features of ground objects in the intermediate layer of the 3D-CNN model;

[0046] The extraction submodule is used to extract features through the convolutional and pooling layers of the 3D-CNN model to obtain fused point cloud data.

[0047] Its beneficial effects are as follows: 1. It acquires geographic location information through multiple methods and formulates layered scanning paths based on terrain classification. Reasonable height layers and scanning routes are set for different terrains; for example, different layering strategies and route designs are used for plains, hills, and mountains. Flight parameters are adjusted in real time considering environmental factors. This not only ensures that the UAV can fly safely and efficiently in complex terrains but also ensures that the collected data has a high overlap rate and completeness, effectively avoiding data loss problems. 2. While removing noise, it better preserves terrain details, making the filtered point cloud data more accurately reflect terrain features. It significantly improves the model's accuracy in recognizing terrain features and land cover categories, enabling more accurate extraction of geographic information. It realistically recreates geographic scenes, providing more reliable 3D model support for geographic information applications. Attached Figure Description

[0048] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0049] Figure 1 This is a schematic diagram of the first embodiment of a UAV-based geographic mapping method according to the present invention;

[0050] Figure 2 This is a schematic diagram of a second embodiment of a UAV-based geographic mapping method according to the present invention;

[0051] Figure 3 This is a schematic diagram of the first embodiment of a UAV-based geographic mapping system according to the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0054] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, a geographic mapping method based on unmanned aerial vehicles (UAVs) includes the following steps:

[0055] Step 101: Obtain the geographic location information to be collected, generate a UAV layered scanning path based on the geographic location information to be collected, and collect LiDAR UAV point cloud data and UAV image data according to the UAV layered scanning path;

[0056] Specifically, in this embodiment, the topographic relief of the area to be collected is calculated based on DEM data, and the terrain is divided into plains, mountains and high mountains. Different layer heights and densities are set for different terrain types.

[0057] The drone flight path is designed based on the layer height and density, and the scanning width and overlap rate of the drone's LiDAR sensor are set according to the spacing between adjacent flight paths to obtain the drone's layered scanning path.

[0058] Specifically,

[0059] I. Generation of Layered Scanning Path for UAVs

[0060] Layered strategy formulation,

[0061] Terrain Classification: Based on the DEM data, the terrain relief of the area to be collected is calculated, and the terrain is divided into four categories: plains (relief ≤ 50m), hills (50m < relief ≤ 200m), mountains (200m < relief ≤ 500m), and high mountains (relief > 500m). Different layer heights and densities are determined for different terrain types.

[0062] Altitude layer settings: For plains areas, set 3-5 altitude layers, with the lowest altitude layer 100-200m above the ground and a layer spacing of 50-100m; for hilly areas, set 5-8 altitude layers, with the lowest altitude layer 150-300m above the ground and a layer spacing of 30-80m; for mountainous and high-mountain areas, set 8-12 altitude layers, with the lowest altitude layer determined based on the actual terrain to ensure the drone flies at a safe altitude, and a layer spacing of 20-50m. The scanning range of each altitude layer must cover the area to be collected, with a 10%-20% edge buffer area reserved to avoid data loss due to drone flight errors.

[0063] Scan path planning,

[0064] Flight path design: A zigzag or grid-like flight path is used for scanning. In a zigzag path, the spacing between adjacent paths is determined based on the scanning width and overlap requirements of the UAV's LiDAR sensor. The overlap rate is set to 60%-80% to ensure sufficient overlap between point cloud data and image data, facilitating subsequent data registration and fusion. For a grid-like flight path, the area to be collected is divided into several square or rectangular grids, and the UAV scans sequentially according to the grid order.

[0065] Flight parameter settings: Based on the UAV's performance parameters and scanning requirements, the flight speed is set to 5-15 m / s, the flight altitude accuracy is controlled within ±1 m, and the deviations of the heading and pitch angles are controlled within ±5°. Simultaneously, considering the impact of environmental factors such as wind and atmospheric visibility on flight, the flight parameters are adjusted in real time to ensure the safe and stable flight of the UAV.

[0066] II. Data Collection

[0067] LiDAR drone point cloud data acquisition:

[0068] A high-precision LiDAR sensor was selected and installed at a suitable location on the drone, ensuring that the sensor's scanning field of view covered the area below the drone's flight path. The sensor's sampling frequency was set to 100kHz-500kHz, adjusted according to the flight speed and scanning accuracy requirements.

[0069] UAV image data acquisition:

[0070] Equipped with a high-resolution digital camera, the camera's focal length is selected based on flight altitude and ground resolution requirements. The ground resolution (GSD) calculation formula is: GSD = sensor pixel size × flight altitude / lens focal length, requiring GSD ≤ 5cm.

[0071] Image acquisition was conducted using a combination of vertical and oblique photography. Vertical photography acquired orthophotos of the ground, while oblique photography acquired oblique images from multiple angles to meet the multi-angle requirements of 3D modeling. The image acquisition interval was determined based on flight speed and overlap requirements, ensuring a forward overlap of ≥80% and a lateral overlap of ≥60% between adjacent images.

[0072] Step 102: Perform data preprocessing on the LiDAR UAV point cloud data, and use the improved MLS algorithm to perform point cloud filtering on the preprocessed point cloud data to obtain filtered UAV point cloud data.

[0073] Specifically, in this embodiment, the local mean and standard deviation of each point in the LiDAR UAV point cloud data are calculated using statistical filtering. Points that deviate from the mean by more than a certain multiple of the standard deviation are considered outliers. A radius threshold is set, and the number of points within that radius is counted. Points with fewer than the set threshold are considered outliers.

[0074] Deleting outliers from the LiDAR drone point cloud data yields the initial drone point cloud data;

[0075] Convert the initial UAV point cloud data to the standard LAS format to obtain standard UAV point cloud data;

[0076] The POS data in the standard UAV point cloud data is converted into absolute geographic coordinates, error compensation is performed using ground control points, and the coordinate transformation parameters are calculated using the least squares method to obtain the processed UAV point cloud data.

[0077] An adaptive weight function is introduced into the MLS algorithm to adjust the weights according to the local density and curvature changes of the point cloud, resulting in an improved MLS algorithm.

[0078] An improved MLS algorithm is used to process UAV point cloud data into blocks, and the block size is determined based on the point cloud density and computing resources.

[0079] Each point cloud data piece is filtered, a local surface is fitted, the distance from each point to the fitted surface is calculated, points whose distance exceeds a set threshold are regarded as ground features, and points whose distance is within the threshold are retained as ground points, thus obtaining filtered UAV point cloud data.

[0080] Specifically,

[0081] I. Data Preprocessing

[0082] Data denoising:

[0083] Outlier detection and removal: A combination of statistical filtering and radius filtering is used to remove obvious outliers. Statistical filtering calculates the local mean and standard deviation of each point, identifying points that deviate from the mean by a certain multiple of the standard deviation as outliers. Radial filtering sets a radius threshold for each point, counting points within that radius; points with fewer than the set threshold are considered outliers. The threshold is adaptively adjusted based on the density and noise level of the point cloud data.

[0084] Outlier handling: Isolated outliers are checked and removed manually to ensure that noise is removed without losing valid data.

[0085] Format conversion: The raw data formats acquired by different LiDAR sensors are uniformly converted into the standard LAS (LiDAR Data Exchange Format) format, using LAS 1.4, to ensure data compatibility and scalability.

[0086] Coordinate calibration: Coordinate calibration is performed using the UAV's POS (Positioning and Orientation System) data and Ground Control Points (GCPs). First, the UAV's POS data is converted into absolute geographic coordinates. Then, error compensation is performed using ground control points. The least squares method is used to calculate the coordinate transformation parameters to ensure that the coordinate system of the point cloud data is consistent with the target system, achieving a calibration accuracy at the centimeter level.

[0087] II. Point Cloud Filtering Process (Improved MLS Algorithm)

[0088] Improvements to the traditional MLS algorithm: Weight function optimization: The traditional MLS algorithm uses a Gaussian weight function. The improved algorithm introduces an adaptive weight function, which adjusts the weights based on local density and curvature changes in the point cloud. For regions with high density and large curvature, the weights of neighboring points are increased to better preserve terrain details; for sparse regions, the weights are decreased to avoid over-smoothing.

[0089] Local Neighborhood Search Optimization: A kd-tree structure is used for fast local neighborhood search, improving search efficiency. Simultaneously, the search radius is dynamically adjusted based on the point cloud density to ensure that the number of neighboring points for each point remains within a reasonable range (generally 10-50), avoiding poor filtering effects due to excessively large or small neighborhoods.

[0090] Filtering process,

[0091] The preprocessed point cloud data is divided into blocks, and the size of each block is determined based on the point cloud density and computing resources, typically 100m × 100m.

[0092] An improved MLS algorithm is applied to each point cloud data to filter the data, fit a local surface, calculate the distance from each point to the fitted surface, and treat points whose distance exceeds a set threshold (usually 0.3-0.5m) as ground features. Points whose distance is within the threshold are retained as ground points, thus obtaining the filtered UAV point cloud data.

[0093] Step 103: Establish a 3D-CNN dual-channel geographic information recognition model based on 3D-CNN convolutional neural network for terrain feature extraction and semantic segmentation of ground features. Optimize the hyperparameters of the recognition model using the improved IWOA whale optimization algorithm to obtain the 3D-CNN-IWOA dual-channel geographic information recognition model.

[0094] Specifically, in this embodiment, an adaptive step size factor is introduced into the spiral update position strategy of the whale optimization algorithm. The step size is dynamically adjusted according to the current iteration number and the fitness value of the population to obtain an improved IWOA whale optimization algorithm.

[0095] Obtain the model's hyperparameters, including at least the learning rate, batch size, number of convolutional kernels, number of neurons in fully connected layers, and dropout rate;

[0096] The fitness function is a weighted sum of the model’s classification accuracy and segmentation accuracy on the validation set, and the weights are adjusted according to project requirements.

[0097] The population size is set to 50-100, the maximum number of iterations is 200-300, and the position and fitness value of the whales are updated after each iteration to obtain the optimal combination of hyperparameters.

[0098] Specifically,

[0099] I. 3D-CNN Dual-Channel Model Architecture Design

[0100] Terrain feature extraction channel

[0101] Network Structure: A multi-layer 3D convolutional neural network is employed, including an input layer, convolutional layers, pooling layers, and fully connected layers. The input layer receives filtered point cloud data and converts it into a voxel grid or point cloud feature vector form. The convolutional layers use 3D convolutional kernels (3×3×3) for feature extraction, progressively extracting local and global terrain features from the point cloud data. The pooling layers employ max pooling or average pooling methods to reduce feature dimensionality and computational cost. Five to eight convolutional layers are used, with the number of channels in each layer gradually increasing from 64 to 512, and the kernel size gradually decreasing to capture terrain features at different scales.

[0102] Geographic feature semantic segmentation channel

[0103] Network Structure: Parallel to the terrain feature extraction channel, it also employs a 3D convolutional neural network, but its design places greater emphasis on extracting detailed features of ground objects. The input layer can combine geometric features (coordinates, normal vectors) and properties such as reflection intensity from the point cloud. Upsampling and deconvolutional layers are added after the convolutional layers to achieve semantic segmentation of each point, outputting category labels for ground objects (buildings, vegetation, roads). Six to nine convolutional layers are used, increasing the number of channels from 32 to 256. A skip connection structure is also introduced to combine shallow detailed features with deep semantic features, improving the accuracy of ground object segmentation.

[0104] Dual-channel fusion: Before the fully connected layer, the output features of the terrain feature extraction channel and the land cover semantic segmentation channel are concatenated and fused to form a comprehensive feature vector containing terrain and land cover information, providing richer input for subsequent feature fusion and geographic information recognition.

[0105] II. Improved IWOA Whale Optimization Algorithm

[0106] Algorithm improvement points

[0107] Initial Population Optimization: Latin hypercube sampling is used instead of traditional random sampling to initialize the whale population, resulting in a more uniform distribution of the population in the solution space and improving the algorithm's global search capability. Latin hypercube sampling divides the solution space into several non-overlapping sub-intervals, and randomly selects a sample from each sub-interval, ensuring better diversity in the initial population.

[0108] Predation strategy adjustment: In the spiral update position strategy of the basic whale optimization algorithm, an adaptive step size factor is introduced to dynamically adjust the step size according to the current iteration number and the fitness value of the population, so as to avoid the algorithm getting stuck in local optima in the later stage.

[0109] Hyperparameter optimization process,

[0110] Determine optimization parameters: The hyperparameters that need to be optimized include the learning rate, batch size, number of convolutional kernels, number of neurons in fully connected layers, dropout rate, etc. of the 3D-CNN model.

[0111] Construct the fitness function: Use the weighted sum of the model's classification accuracy and segmentation accuracy on the validation set as the fitness function. The weights are adjusted according to project requirements. Generally, the classification accuracy weight is 0.6 and the segmentation accuracy weight is 0.4.

[0112] Algorithm execution: The population size is set to 50-100, the maximum number of iterations is 200-300, the hyperparameters are optimized by the improved IWOA algorithm, and the position and fitness value of the whale are updated after each iteration. Finally, the optimal combination of hyperparameters is obtained to build a 3D-CNN-IWOA dual-channel geographic information recognition model.

[0113] Step 104: Input the filtered UAV point cloud data into the 3D-CNN-IWOA dual-channel geographic information recognition model for feature fusion to obtain fused point cloud data;

[0114] Specifically, in this embodiment, the filtered UAV point cloud data is converted into point cloud feature descriptors, and terrain features and semantic features of ground objects are fused in the intermediate layer of the 3D-CNN model;

[0115] Feature extraction is performed using convolutional and pooling layers of a 3D-CNN model to obtain fused point cloud data.

[0116] Specifically,

[0117] I. Input Data Processing

[0118] The filtered UAV point cloud data is converted into a format suitable for input into the 3D-CNN-IWOA model, such as voxel mesh or point cloud feature descriptors. For voxel mesh, the point cloud data is divided into a uniform 3D mesh, and the points within each voxel generate voxel features using statistical methods (mean, variance, maximum, minimum). For point cloud feature descriptors, the local geometric features (normal, curvature, Hessian matrix eigenvalues) of each point are calculated to form the feature vector of the point cloud.

[0119] II. Feature Fusion Methods

[0120] An intermediate fusion approach is adopted, which fuses terrain features and semantic features of ground objects in the intermediate layers of the 3D-CNN model (after the 3rd and 5th convolutional layers). Specifically, the output feature maps of the two channels at the corresponding intermediate layers are added or concatenated element-wise, and then the features are extracted again through subsequent convolutional and pooling layers. This ensures that the fused features retain both the overall structural information of the terrain and the detailed semantic information of the ground objects.

[0121] III. Optimization of the integration process

[0122] Before feature fusion, the features of the two channels are normalized using batch normalization to reduce internal covariate bias and improve model training efficiency and fusion performance. Simultaneously, channel attention is added after the fusion layer, allowing the model to automatically learn the importance of different features and enhance the expressive power of key features.

[0123] Step 105: Using the improved NeRF neural radiation field framework, the fused point cloud data and UAV image data are registered and modeled to generate a 3D geographic reality model.

[0124] Specifically, in this embodiment, based on the volumetric radiation field representation of NeRF, the reflection intensity and normal vector information of the point cloud are introduced as additional input features, and the AdamW optimization algorithm is used instead of the Adam algorithm, with weight decay added during the optimization process;

[0125] The scene is sampled using a hierarchical sampling strategy. Sampling is performed at a coarse-grained level to quickly converge to the scene structure, and dense sampling is performed at a fine-grained level.

[0126] The color and density of each sampling point are predicted by the radiation field function, the loss function with the real image is calculated, and the model parameters are updated by the backpropagation algorithm.

[0127] New images are synthesized from different perspectives using a volumetric rendering algorithm to generate a 3D model of the geographic scene.

[0128] Specifically,

[0129] I. Improvements to the NeRF Framework

[0130] Camera pose estimation optimization

[0131] Initial camera pose estimation is performed using POS data from the UAV and the SFM (Structure from Motion) algorithm, followed by pose optimization using ground control points from the point cloud data. The Bundle Adjustment algorithm is employed to simultaneously optimize the camera's intrinsic and extrinsic parameters and the coordinates of 3D points, improving the accuracy of camera pose estimation and ensuring sub-pixel level spatial alignment between image data and point cloud data.

[0132] Improved representation of radiation field

[0133] Based on the traditional NeRF volumetric radiation field representation, the reflection intensity and normal vector information of point cloud are introduced as additional input features to enhance the ability of the radiation field to describe the surface material and geometry of objects.

[0134] Optimize and improve the algorithm.

[0135] The AdamW optimization algorithm is used instead of the traditional Adam algorithm. Weight decay is added during the optimization process to prevent model overfitting. At the same time, a learning rate scheduling strategy is introduced to dynamically adjust the learning rate according to the number of training epochs. The initial learning rate is set to 0.001, and the learning rate decays to 0.5 times the original value every 50 epochs, which improves the training stability and convergence speed of the model.

[0136] II. Registration and Modeling Process

[0137] Data preprocessing:

[0138] Distortion correction and color normalization are performed on UAV image data. The camera's intrinsic parameter matrix and distortion coefficients are used to correct the image and eliminate the effects of lens distortion. Color normalization employs histogram equalization or linear stretching methods to unify the image's brightness and contrast, improving the accuracy of subsequent registration and modeling.

[0139] The fused point cloud data is converted into point cloud density maps and depth maps, which serve as auxiliary input information for the NeRF model, helping the model to better understand the geometric structure of the scene.

[0140] Scene initialization:

[0141] Based on the camera pose estimation results, the camera's position and orientation are initialized in 3D space, and an initial scene bounding box is constructed. The size of the bounding box is automatically adjusted according to the extent of the point cloud data to ensure that all geographic areas to be modeled are included.

[0142] Radiation field training:

[0143] A hierarchical sampling strategy is employed to sample the scene. First, sampling is performed at a coarse-grained level to quickly converge to the general scene structure. Then, dense sampling is performed at a fine-grained level to optimize details. During training, image data and point cloud auxiliary data are input simultaneously. The color and density of each sampling point are predicted using an improved radiation field function. Loss functions (mean squared error loss and structural similarity loss) are calculated compared to the real image. Model parameters are updated using a backpropagation algorithm. The number of training epochs is set to 1000-2000.

[0144] 3D model generation:

[0145] After training, new images are synthesized from different perspectives using a volumetric rendering algorithm to generate a 3D model of the geographic scene. The MarchingCubes algorithm is used to convert the voxel data into a triangular mesh model for surface reconstruction and texture mapping, ultimately resulting in a 3D model with realistic geographic texture and high-precision geometry.

[0146] III. Precision Control

[0147] During the registration and modeling process, the accuracy of the generated 3D model is periodically evaluated. Ground control points and checkpoints are used to statistically analyze errors, including planar position errors and elevation errors. When the errors exceed the set thresholds (planar error ≤ 5cm, elevation error ≤ 10cm), the camera pose and model parameters are readjusted until the accuracy requirements are met.

[0148] Its beneficial effects are as follows: 1. It acquires geographic location information through multiple methods and formulates layered scanning paths based on terrain classification. Reasonable height layers and scanning routes are set for different terrains; for example, different layering strategies and route designs are used for plains, hills, and mountains. Flight parameters are adjusted in real time considering environmental factors. This not only ensures that the UAV can fly safely and efficiently in complex terrains but also ensures that the collected data has a high overlap rate and completeness, effectively avoiding data loss problems. 2. While removing noise, it better preserves terrain details, making the filtered point cloud data more accurately reflect terrain features. It significantly improves the model's accuracy in recognizing terrain features and land cover categories, enabling more accurate extraction of geographic information. It realistically recreates geographic scenes, providing more reliable 3D model support for geographic information applications.

[0149] Please see Figure 2 In a UAV-based geographic mapping method, the improved NeRF neural radiation field framework is used to register and model fused point cloud data and UAV image data to generate a 3D geographic reality model, including the following steps:

[0150] Step 201: Based on the volumetric radiation field representation of NeRF, the reflection intensity and normal vector information of the point cloud are introduced as additional input features. The AdamW optimization algorithm is used instead of the Adam algorithm, and weight decay is added during the optimization process.

[0151] Step 202: Use a hierarchical sampling strategy to sample the scene. Sampling is performed at the coarse-grained level to quickly converge to the scene structure, and dense sampling is performed at the fine-grained level.

[0152] Step 203: Predict the color and density of each sampling point using the radiation field function, calculate the loss function with respect to the real image, and update the model parameters using the backpropagation algorithm;

[0153] Step 204: Synthesize new images from different perspectives using the volumetric rendering algorithm to generate a 3D model of the geographic scene.

[0154] The above describes an embodiment of the UAV-based geographic mapping method of the present invention. Please refer to [link / reference]. Figure 3 In a UAV-based cadastral mapping system, the geographic mapping system includes the following modules:

[0155] The UAV path planning module is used to acquire the geographic location information to be collected, generate the UAV layered scanning path based on the geographic location information to be collected, and collect LiDAR UAV point cloud data and UAV image data according to the UAV layered scanning path;

[0156] The mapping data acquisition module is used to preprocess the LiDAR UAV point cloud data and use the improved MLS algorithm to filter the preprocessed point cloud data to obtain filtered UAV point cloud data.

[0157] The recognition model building module is used to build a 3D-CNN dual-channel geographic information recognition model based on 3D-CNN convolutional neural network for terrain feature extraction and semantic segmentation of ground objects. The hyperparameters of the recognition model are optimized by using the improved IWOA whale optimization algorithm to obtain the 3D-CNN-IWOA dual-channel geographic information recognition model.

[0158] The mapping data fusion module is used to input filtered UAV point cloud data into the 3D-CNN-IWOA dual-channel geographic information recognition model for feature fusion to obtain fused point cloud data;

[0159] The 3D model building module is used to register and model fused point cloud data and UAV image data using an improved NeRF neural radiation field framework to generate a 3D geographic reality model.

[0160] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for geographic mapping based on unmanned aerial vehicles, characterized in that, The geographic surveying method includes the following steps: Acquire geographic location information to be collected, generate a UAV layered scanning path based on the geographic location information to be collected, and collect LiDAR UAV point cloud data and UAV image data according to the UAV layered scanning path; The LiDAR UAV point cloud data is preprocessed, and the preprocessed point cloud data is filtered using an improved MLS algorithm to obtain filtered UAV point cloud data. A 3D-CNN dual-channel geographic information recognition model for terrain feature extraction and semantic segmentation is established based on a 3D-CNN convolutional neural network. The hyperparameters of the recognition model are optimized using the improved IWOA whale optimization algorithm to obtain the 3D-CNN-IWOA dual-channel geographic information recognition model. The filtered UAV point cloud data is input into the 3D-CNN-IWOA dual-channel geographic information recognition model for feature fusion to obtain fused point cloud data. The fused point cloud data and UAV image data are registered and modeled using the improved NeRF neural radiation field framework to generate a 3D geographic reality model. The 3D-CNN dual-channel geographic information recognition model, which establishes terrain feature extraction and semantic segmentation based on a 3D-CNN convolutional neural network, further includes: The improved IWOA whale optimization algorithm is used to optimize the hyperparameters of the recognition model, resulting in a 3D-CNN-IWOA dual-channel geographic information recognition model. An adaptive step size factor is introduced into the spiral update position strategy of the whale optimization algorithm. The step size is dynamically adjusted according to the current iteration number and the fitness value of the population, resulting in the improved IWOA whale optimization algorithm. Obtain the model's hyperparameters, including at least the learning rate, batch size, number of convolutional kernels, number of neurons in fully connected layers, and dropout rate; The fitness function is a weighted sum of the model’s classification accuracy and segmentation accuracy on the validation set, and the weights are adjusted according to project requirements. The population size was set to 50-100, the maximum number of iterations was 200-300, and the hyperparameters were optimized using the improved IWOA algorithm. After each iteration, the position and fitness value of the whale were updated to obtain the optimal combination of hyperparameters for building a 3D-CNN-IWOA dual-channel geographic information recognition model. The step of inputting the filtered UAV point cloud data into the 3D-CNN-IWOA dual-channel geographic information recognition model for feature fusion to obtain fused point cloud data further includes: The filtered UAV point cloud data is converted into point cloud feature descriptors, and terrain features and semantic features of ground objects are fused in the intermediate layer of the 3D-CNN model. Feature extraction is performed using the convolutional and pooling layers of a 3D-CNN model to obtain fused point cloud data; The method of registering and modeling the fused point cloud data and UAV image data using the improved NeRF neural radiation field framework to generate a 3D geographic reality model also includes: Based on the volumetric radiation field representation of NeRF, the reflection intensity and normal vector information of point cloud are introduced as additional input features. The AdamW optimization algorithm is used instead of the Adam algorithm, and weight decay is added during the optimization process. The scene is sampled using a hierarchical sampling strategy. Sampling is performed at a coarse-grained level to quickly converge to the scene structure, and dense sampling is performed at a fine-grained level. The color and density of each sampling point are predicted by the radiation field function, the loss function with the real image is calculated, and the model parameters are updated by the backpropagation algorithm. New images are synthesized from different perspectives using a volumetric rendering algorithm to generate a 3D model of the geographic scene.

2. The unmanned aerial vehicle based georeferencing method of claim 1, wherein, The process of acquiring the geographic location information to be collected, generating a UAV layered scanning path based on the geographic location information, and collecting LiDAR UAV point cloud data and UAV image data according to the UAV layered scanning path includes: The terrain relief of the area to be collected is calculated based on the DEM data. The terrain is divided into plains with relief ≤ 50m, hills with relief ≤ 200m, mountains with relief ≤ 500m and high mountains with relief > 500m. Different layer heights and densities are set for different terrain types. The UAV flight path is designed based on the layer height and density, and the scanning width and overlap rate of the UAV LiDAR sensor are set according to the spacing between adjacent flight paths to obtain the UAV layered scanning path.

3. The unmanned aerial vehicle based georeferencing method of claim 1, wherein, The step of preprocessing the LiDAR drone point cloud data, and then using an improved MLS algorithm to filter the preprocessed point cloud data to obtain filtered drone point cloud data, includes: A combination of statistical filtering and radius filtering is used to remove obvious outliers. Statistical filtering calculates the local mean and standard deviation of each point and considers points that deviate from the mean by more than a certain multiple of the standard deviation as outliers. Radius filtering sets a radius threshold for each point and counts the number of points within that radius; points with fewer than the set threshold are considered outliers. The threshold setting is adaptively adjusted based on the density and noise level of the point cloud data. Deleting outliers from the LiDAR drone point cloud data yields the initial drone point cloud data; The initial UAV point cloud data is converted into standard LAS format to obtain standard UAV point cloud data; The POS data in the standard UAV point cloud data is converted into absolute geographic coordinates, error compensation is performed using ground control points, and the coordinate transformation parameters are calculated using the least squares method to obtain the processed UAV point cloud data.

4. The UAV-based geographic mapping method as described in claim 3, characterized in that, The step of preprocessing the LiDAR drone point cloud data, and using an improved MLS algorithm to filter the preprocessed point cloud data to obtain filtered drone point cloud data, further includes: An adaptive weight function is introduced into the MLS algorithm to adjust the weights according to the local density and curvature changes of the point cloud, resulting in an improved MLS algorithm. An improved MLS algorithm is used to process UAV point cloud data into blocks, and the block size is determined based on the point cloud density and computing resources. Each point cloud data piece is filtered, a local surface is fitted, the distance from each point to the fitted surface is calculated, points whose distance exceeds a set threshold are regarded as ground features, and points whose distance is within the threshold are retained as ground points, thus obtaining filtered UAV point cloud data.

5. A UAV-based geographic mapping system for executing the UAV-based geographic mapping method as described in claim 1, characterized in that, The geographic mapping system includes the following modules: The UAV path planning module is used to acquire geographic location information to be collected, generate a UAV layered scanning path based on the geographic location information to be collected, and collect LiDAR UAV point cloud data and UAV image data according to the UAV layered scanning path; The mapping data acquisition module is used to preprocess the LiDAR UAV point cloud data, and to perform point cloud filtering on the preprocessed point cloud data using an improved MLS algorithm to obtain filtered UAV point cloud data. The recognition model building module is used to build a 3D-CNN dual-channel geographic information recognition model based on 3D-CNN convolutional neural network for terrain feature extraction and semantic segmentation of ground objects. The hyperparameters of the recognition model are optimized by using the improved IWOA whale optimization algorithm to obtain the 3D-CNN-IWOA dual-channel geographic information recognition model. The mapping data fusion module is used to input the filtered UAV point cloud data into the 3D-CNN-IWOA dual-channel geographic information recognition model for feature fusion to obtain fused point cloud data. The 3D model building module is used to register and model the fused point cloud data and UAV image data using the improved NeRF neural radiation field framework to generate a geographic reality 3D model.

6. A drone-based georeferencing system as claimed in claim 5, wherein, The UAV path planning module includes the following sub-modules: The sub-module is used to calculate the topographic relief of the area to be collected based on the DEM data. The terrain is divided into plains with relief ≤ 50m, hills with relief ≤ 200m, mountains with relief ≤ 500m and high mountains with relief > 500m. Different layer heights and densities are set for different terrain types. A submodule is obtained for designing UAV flight paths based on the layer height and density, setting the scanning width and overlap rate of the UAV LiDAR sensor according to the spacing between adjacent flight paths, and obtaining the UAV layered scanning path.

7. A drone-based georeferencing system as claimed in claim 5, wherein, The mapping data fusion module includes the following sub-modules: The transformation submodule is used to convert the filtered UAV point cloud data into point cloud feature descriptors, and fuse terrain features and semantic features of ground objects in the intermediate layer of the 3D-CNN model; The extraction submodule is used to extract features through the convolutional and pooling layers of the 3D-CNN model to obtain fused point cloud data.

Images