A cloud-based surveying and mapping data processing method and system

Abstract

The present invention discloses a cloud-based surveying and mapping data processing method and system. By collecting multi-source surveying and mapping data sets, the multi-source surveying and mapping data sets are fused based on an improved hierarchical Transformer model in a cloud platform. An MTL-DNN surveying and mapping data analysis model is established based on an MTL-DNN multi-task deep learning network. The model is pruned using a fine-grained pruner. The fused multi-source surveying and mapping data sets are input into the target MTL-DNN surveying and mapping data analysis model to perform terrain modeling, object classification, and risk prediction. This method can reduce model parameters and computational complexity, and improve deployment efficiency. The method provides accurate and efficient data support and decision-making basis for surveying and mapping projects, disaster warning, and other fields.

Description

Technical Field

[0001] The present invention relates to the field of data analysis technology, and in particular to a cloud-based surveying and mapping data processing method and system thereof. Background Art

[0002] In the field of surveying and mapping, with the continuous development of technologies such as drones, satellite remote sensing, and ground-based measurement equipment, the methods for acquiring surveying and mapping data are becoming increasingly diverse, and the amount of data is also exploding. Traditional surveying and mapping data processing methods are usually performed locally and rely on high-performance local computers or servers. This not only places high demands on hardware equipment and increases costs, but also has significant limitations in data storage, computing resource sharing, and collaborative processing. Although existing cloud-based surveying and mapping data processing methods have solved the problems of data storage and resource sharing to a certain extent, the efficiency, accuracy, and intelligence of data processing still need to be improved. In the data preprocessing stage, the identification and processing of noise data are not accurate enough; in the data feature extraction and analysis stage, there is a lack of efficient algorithm support, resulting in low accuracy and reliability of the processing results. Summary of the Invention

[0003] The purpose of the present invention is to solve the above problems and design a cloud-based surveying and mapping data processing method and system.

[0004] To achieve the above-mentioned purpose, the technical solution of the present invention is as follows: further, in the above-mentioned cloud-based surveying and mapping data processing method, the cloud-based surveying and mapping data processing method includes the following steps:

[0005] Collect satellite remote sensing data, UAV LiDAR data, and IoT sensor data, pre-process the collected data to obtain a multi-source mapping dataset, and transmit the multi-source mapping dataset to a cloud platform;

[0006] Based on the improved hierarchical Transformer model in the cloud platform, data fusion is performed on the multi-source mapping dataset, and a spatiotemporal attention mechanism is introduced to dynamically associate data with different resolutions and timestamps to obtain a fused multi-source mapping dataset;

[0007] Based on the MTL-DNN multi-task deep learning network, an MTL-DNN surveying and mapping data analysis model is established. The model is pruned using a fine-grained pruner to obtain the target MTL-DNN surveying and mapping data analysis model.

[0008] The fused multi-source mapping data set is input into the target MTL-DNN mapping data analysis model to perform terrain modeling, object classification and risk prediction to obtain mapping data analysis results.

[0009] Furthermore, in the above-mentioned cloud-based mapping data processing method, collecting satellite remote sensing data, UAV LiDAR data, and IoT sensor data, preprocessing the collected data to obtain a multi-source mapping dataset, and transmitting the multi-source mapping dataset to a cloud platform includes:

[0010] Collecting satellite remote sensing data, and performing radiation correction and geometric correction on the satellite remote sensing data based on a model of atmospheric transmission theory to obtain first mapping data;

[0011] Acquire UAV LiDAR data, perform point cloud denoising on the UAV LiDAR data using a statistical filtering method, and perform coordinate transformation and point cloud resampling on the denoised data using a linear interpolation method to obtain second surveying and mapping data;

[0012] collecting IoT sensor data, performing outlier processing on the IoT sensor data using a polynomial interpolation method, and performing time synchronization on the outlier-processed data to obtain third surveying and mapping data;

[0013] Normalizing the first surveying and mapping data, the second surveying and mapping data, and the third surveying and mapping data, and integrating the normalized data to obtain a multi-source surveying and mapping data set;

[0014] The multi-source surveying and mapping dataset is encrypted, and the encrypted multi-source surveying and mapping dataset is transmitted to the cloud platform based on 5G communication.

[0015] Furthermore, in the above-mentioned cloud-based surveying and mapping data processing method, the improved hierarchical Transformer model in the cloud platform performs data fusion on the multi-source surveying and mapping dataset, introduces a spatiotemporal attention mechanism to dynamically associate data with different resolutions and timestamps, and obtains a fused multi-source surveying and mapping dataset, including:

[0016] A hierarchical Transformer model is built on the cloud platform using the hierarchical structure of multi-layer Transformer encoders and decoders;

[0017] In the model, each encoder layer consists of a multi-head self-attention mechanism and a feedforward neural network, and the decoder layer consists of a multi-head self-attention mechanism and a feedforward neural network. At the same time, the encoder-decoder attention mechanism is introduced to obtain an improved hierarchical Transformer model.

[0018] A spatiotemporal attention module is added to the encoder and decoder of the improved hierarchical Transformer model, and the spatiotemporal attention module is used to process the temporal dimension and spatial dimension information of the data respectively.

[0019] Furthermore, in the above-mentioned cloud-based surveying and mapping data processing method, the improved hierarchical Transformer model in the cloud platform performs data fusion on the multi-source surveying and mapping dataset, introduces a spatiotemporal attention mechanism to dynamically associate data of different resolutions and timestamps, and obtains a fused multi-source surveying and mapping dataset, further comprising:

[0020] Obtaining a multi-source surveying and mapping dataset in a cloud platform, and inputting the multi-source surveying and mapping dataset into an improved hierarchical Transformer model;

[0021] The model's underlying encoder performs preliminary feature extraction on the input multi-source data, and dynamically associates data of different resolutions and timestamps through the spatiotemporal attention module;

[0022] The spatiotemporal attention mechanism is used to calculate the attention weights according to the spatial position and distance, and the high-resolution detail features and low-resolution global features are fused to obtain a fused multi-source mapping dataset.

[0023] Furthermore, in the above-mentioned cloud-based surveying and mapping data processing method, the MTL-DNN surveying and mapping data analysis model is established based on the MTL-DNN multi-task deep learning network, and the model is pruned using a fine-grained pruner to obtain a target MTL-DNN surveying and mapping data analysis model, including:

[0024] An MTL-DNN mapping data analysis model is established based on the MTL-DNN multi-task deep learning network. The shared layer is located at the bottom layer of the MTL-DNN model and is used to extract common features of the data.

[0025] Design task-specific layers based on the different requirements of terrain modeling, feature classification, and risk prediction;

[0026] The output of the shared layer is used as the input of the three task-specific layers at the same time. The task-specific layers are independent of each other, and features are only shared in the shared layer.

[0027] Furthermore, in the above-mentioned cloud-based surveying and mapping data processing method, the MTL-DNN surveying and mapping data analysis model is established based on the MTL-DNN multi-task deep learning network, and the model is pruned using a fine-grained pruner to obtain a target MTL-DNN surveying and mapping data analysis model, further comprising:

[0028] Using the Adam optimizer, set the learning rate, momentum, and weight decay parameters of the MTL-DNN mapping data analysis model;

[0029] Based on the fine-grained pruner, the absolute value of each weight in the network is calculated, a pruning threshold is set, and the weights with absolute values ​​less than the threshold are set to zero, and unimportant connections are removed to the target MTL-DNN mapping data analysis model.

[0030] Furthermore, in the above-mentioned cloud-based surveying and mapping data processing method, the fused multi-source surveying and mapping data set is input into the target MTL-DNN surveying and mapping data analysis model to perform terrain modeling, object classification, and risk prediction to obtain surveying and mapping data analysis results, including:

[0031] By extracting spatial and temporal features from the shared layer and combining them with the processing of the task-specific layer, the three-dimensional coordinates and height information of the terrain are output to obtain terrain modeling.

[0032] The spectrum, texture, shape and other features extracted by the shared layer are combined with the classifier of the task-specific layer to output the category probability distribution of the ground object and obtain the ground object classification;

[0033] By integrating the time series features, spatial features and land feature features in the data, combining historical risk data and relevant risk indicators, and using classification algorithms, the risk level or probability of risk occurrence is output to obtain risk prediction.

[0034] Furthermore, in a cloud-based surveying and mapping data processing system, the cloud-based surveying and mapping data processing system includes the following modules:

[0035] A data acquisition and transmission module is used to collect satellite remote sensing data, UAV LiDAR data, and IoT sensor data, pre-process the collected data to obtain a multi-source mapping data set, and transmit the multi-source mapping data set to the cloud platform;

[0036] A surveying and mapping data fusion module is used to fuse the multi-source surveying and mapping dataset based on the improved hierarchical Transformer model in the cloud platform, introduce a spatiotemporal attention mechanism to dynamically associate data with different resolutions and timestamps, and obtain a fused multi-source surveying and mapping dataset;

[0037] The analysis model building module is used to build an MTL-DNN surveying and mapping data analysis model based on the MTL-DNN multi-task deep learning network, and use the pruner to prune the model to obtain the target MTL-DNN surveying and mapping data analysis model;

[0038] The surveying and mapping data analysis module is used to input the fused multi-source surveying and mapping data set into the target MTL-DNN surveying and mapping data analysis model to perform terrain modeling, object classification and risk prediction, and obtain surveying and mapping data analysis results.

[0039] Furthermore, in a cloud-based surveying and mapping data processing system, the surveying and mapping data fusion module includes the following submodules:

[0040] Establish a submodule for building a hierarchical Transformer model in the cloud platform using the hierarchical structure of multi-layer Transformer encoders and decoders;

[0041] A submodule is introduced, where each encoder layer in the model consists of a multi-head self-attention mechanism and a feedforward neural network, and the decoder layer consists of a multi-head self-attention mechanism and a feedforward neural network. At the same time, the encoder-decoder attention mechanism is introduced to obtain an improved hierarchical Transformer model;

[0042] A submodule is added to add a spatiotemporal attention module to the encoder and decoder of the improved hierarchical Transformer model, and the spatiotemporal attention module is used to process the temporal dimension and spatial dimension information of the data respectively.

[0043] Furthermore, in a cloud-based surveying and mapping data processing system, the surveying and mapping data fusion module further includes the following submodules:

[0044] An input submodule, configured to obtain a multi-source mapping dataset in the cloud platform and input the multi-source mapping dataset into the improved hierarchical Transformer model;

[0045] The extraction submodule is used to perform preliminary feature extraction on the input multi-source data using the model's underlying encoder, and dynamically associate data of different resolutions and timestamps through the spatiotemporal attention module;

[0046] The fusion submodule is used to use the spatiotemporal attention mechanism to calculate the attention weight according to the spatial position and distance, and to fuse the high-resolution detail features and low-resolution global features to obtain a fused multi-source mapping dataset.

[0047] The beneficial effects of this approach include collecting satellite remote sensing data, drone LiDAR data, and IoT sensor data, preprocessing the collected data to obtain a multi-source mapping dataset, and transmitting the multi-source mapping dataset to a cloud platform. The multi-source mapping dataset is then fused using an improved hierarchical Transformer model in the cloud platform, introducing a spatiotemporal attention mechanism to dynamically associate data of different resolutions and timestamps to obtain a fused multi-source mapping dataset. An MTL-DNN mapping data analysis model is established based on the MTL-DNN multi-task deep learning network, and pruned using a fine-grained pruner to obtain a target MTL-DNN mapping data analysis model. The fused multi-source mapping dataset is then input into the target MTL-DNN mapping data analysis model to perform terrain modeling, feature classification, and risk prediction, resulting in mapping data analysis results. This reduces model parameters and computational complexity, improving deployment efficiency. Ultimately, terrain modeling, feature classification, and risk prediction are achieved, with visualization output and optimization verification, providing accurate and efficient data support and decision-making basis for fields such as mapping engineering and disaster warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Various other advantages and benefits will become apparent to those skilled in the art by reading the following detailed description of the preferred embodiment.The accompanying drawings are only for the purpose of illustrating the preferred embodiment and are not to be considered as limiting the present invention.

[0049] Figure 1 Schematic diagram of a first embodiment of a cloud-based surveying and mapping data processing method in an embodiment of the present invention;

[0050] Figure 2 Schematic diagram of a second embodiment of a cloud-based surveying and mapping data processing method according to an embodiment of the present invention;

[0051] Figure 3 Schematic diagram of a first embodiment of a cloud-based surveying and mapping data processing system in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present 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 only used to explain the present invention and are not intended to limit the present invention.

[0053] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a," "an," "said," and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the stated features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0054] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 As shown, a cloud-based surveying and mapping data processing method includes the following steps:

[0055] Step 101: Collect satellite remote sensing data, UAV LiDAR data, and IoT sensor data, pre-process the collected data to obtain a multi-source mapping dataset, and transmit the multi-source mapping dataset to a cloud platform;

[0056] Specifically, in this embodiment, satellite remote sensing data is collected, and radiation correction and geometric correction are performed on the satellite remote sensing data based on a model of atmospheric transmission theory to obtain first mapping data;

[0057] Obtain UAV LiDAR data, perform point cloud denoising on the UAV LiDAR data using a statistical filtering method, and perform coordinate transformation and point cloud resampling on the denoised data using a linear interpolation method to obtain the second mapping data;

[0058] Collect IoT sensor data, process outliers on the IoT sensor data using a polynomial interpolation method, and synchronize the processed data with time to obtain third mapping data;

[0059] Normalizing the first surveying and mapping data, the second surveying and mapping data, and the third surveying and mapping data, and integrating the normalized data to obtain a multi-source surveying and mapping data set;

[0060] The multi-source surveying and mapping dataset is encrypted and transmitted to the cloud platform based on 5G communication.

[0061] Step 102: Based on the improved hierarchical Transformer model in the cloud platform, data fusion is performed on the multi-source surveying and mapping dataset, and a spatiotemporal attention mechanism is introduced to dynamically associate data with different resolutions and timestamps to obtain a fused multi-source surveying and mapping dataset.

[0062] Specifically, in this embodiment, a hierarchical Transformer model is established in the cloud platform using a hierarchical structure of multi-layer Transformer encoders and decoders;

[0063] In the model, each encoder layer consists of a multi-head self-attention mechanism and a feedforward neural network, and the decoder layer consists of a multi-head self-attention mechanism and a feedforward neural network. At the same time, the encoder-decoder attention mechanism is introduced to obtain an improved hierarchical Transformer model.

[0064] A spatiotemporal attention module is added to the encoder and decoder of the improved hierarchical Transformer model, and the spatiotemporal attention module is used to process the temporal and spatial dimension information of the data respectively.

[0065] Obtain multi-source mapping datasets from the cloud platform and input them into the improved hierarchical Transformer model;

[0066] The model's underlying encoder performs preliminary feature extraction on the input multi-source data, and dynamically associates data of different resolutions and timestamps through the spatiotemporal attention module;

[0067] The spatiotemporal attention mechanism is used to calculate the attention weights according to the spatial position and distance, and the high-resolution detail features and low-resolution global features are fused to obtain a fused multi-source mapping dataset.

[0068] Step 103: Establish an MTL-DNN surveying and mapping data analysis model based on the MTL-DNN multi-task deep learning network, and use a fine-grained pruner to prune the model to obtain a target MTL-DNN surveying and mapping data analysis model;

[0069] Specifically, in this embodiment, an MTL-DNN surveying and mapping data analysis model is established based on the MTL-DNN multi-task deep learning network. The shared layer is located at the bottom layer of the MTL-DNN model and is used to extract common features of the data.

[0070] Design task-specific layers based on the different requirements of terrain modeling, feature classification, and risk prediction;

[0071] The output of the shared layer is used as the input of the three task-specific layers at the same time. The task-specific layers are independent of each other, and features are only shared in the shared layer.

[0072] Use the Adam optimizer to set the learning rate, momentum, and weight decay parameters of the MTL-DNN mapping data analysis model;

[0073] Based on the fine-grained pruner, the absolute value of each weight in the network is calculated, a pruning threshold is set, and the weights with absolute values ​​less than the threshold are set to zero, and unimportant connections are removed to the target MTL-DNN mapping data analysis model.

[0074] Step 104: Input the fused multi-source surveying and mapping data set into the target MTL-DNN surveying and mapping data analysis model to perform terrain modeling, object classification, and risk prediction to obtain surveying and mapping data analysis results.

[0075] Specifically, in this embodiment, the spatial and temporal features extracted by the shared layer are combined with the processing of the task-specific layer to output the three-dimensional coordinates and height information of the terrain to obtain terrain modeling;

[0076] The spectrum, texture, shape and other features extracted by the shared layer are combined with the classifier of the task-specific layer to output the category probability distribution of the ground object and obtain the ground object classification;

[0077] By integrating the time series features, spatial features and land feature features in the data, combining historical risk data and relevant risk indicators, and using classification algorithms, the risk level or probability of risk occurrence is output to obtain risk prediction.

[0078] Its beneficial effect is that it can reduce model parameters and calculation amount, improve deployment efficiency; ultimately realize terrain modeling, land feature classification and risk prediction, output visualization results and verify optimization, and provide accurate and efficient data support and decision-making basis for surveying and mapping engineering, disaster warning and other fields.

[0079] See also Figure 2 In a cloud-based mapping data processing method, a multi-source mapping dataset is fused based on an improved hierarchical Transformer model in the cloud platform. A spatiotemporal attention mechanism is introduced to dynamically associate data with different resolutions and timestamps. The fused multi-source mapping dataset includes the following steps:

[0080] Step 201: Establish a hierarchical Transformer model in the cloud platform using a hierarchical structure of multi-layer Transformer encoders and decoders;

[0081] Step 202: Each encoder layer in the model is composed of a multi-head self-attention mechanism and a feedforward neural network, and the decoder is composed of a multi-head self-attention mechanism and a feedforward neural network. At the same time, the encoder-decoder attention mechanism is introduced to obtain an improved hierarchical Transformer model;

[0082] Step 203: Add a spatiotemporal attention module to the encoder and decoder of the improved hierarchical Transformer model, and use the spatiotemporal attention module to process the temporal dimension and spatial dimension information of the data respectively.

[0083] The above is an introduction to an embodiment of a cloud-based surveying and mapping data processing method of the present invention. Figure 3 In a cloud-based surveying and mapping data processing system, the cloud-based surveying and mapping data processing system includes the following modules:

[0084] The data acquisition and transmission module is used to collect satellite remote sensing data, UAV LiDAR data, and IoT sensor data, pre-process the collected data to obtain multi-source mapping data sets, and transmit the multi-source mapping data sets to the cloud platform;

[0085] The surveying and mapping data fusion module is used to fuse multi-source surveying and mapping datasets based on the improved hierarchical Transformer model in the cloud platform. It introduces a spatiotemporal attention mechanism to dynamically associate data with different resolutions and timestamps to obtain a fused multi-source surveying and mapping dataset.

[0086] The analysis model building module is used to build an MTL-DNN surveying and mapping data analysis model based on the MTL-DNN multi-task deep learning network, and use the pruner to prune the model to obtain the target MTL-DNN surveying and mapping data analysis model;

[0087] The surveying and mapping data analysis module is used to input the fused multi-source surveying and mapping data set into the target MTL-DNN surveying and mapping data analysis model to perform terrain modeling, object classification and risk prediction, and obtain the surveying and mapping data analysis results.

[0088] Specifically, the present invention can also be implemented through the following steps:

[0089] 1. Data collection and preprocessing;

[0090] (1) Data collection;

[0091] Satellite remote sensing data collection;

[0092] Clarify the surveying and mapping area scope, accuracy requirements, and required spectral information of the project, and use this as a basis to select appropriate satellite data sources, such as the Gaofen series of satellites and Sentinel satellites.

[0093] Within the predetermined collection time window, the satellite data receiving device is started to receive and store satellite remote sensing data in real time.

[0094] UAV LiDAR data collection;

[0095] Control the drone to fly according to the planned route, and start the LiDAR sensor to collect data at the same time, recording flight attitude data (longitude, latitude, altitude, heading angle, pitch angle, roll angle, etc.) and LiDAR point cloud data in real time.

[0096] IoT sensor data collection;

[0097] Set the data collection frequency and transmission time interval of the sensor, start the sensor to collect data, and the data collection terminal receives and stores the data sent by the sensor in real time.

[0098] (2) Data preprocessing;

[0099] Satellite remote sensing data preprocessing;

[0100] Radiometric correction: Satellite remote sensing data contains radiometric errors due to factors such as atmospheric scattering and sensor response characteristics, necessitating radiometric correction. Radiometric correction models, such as those based on atmospheric transmission theory, are used to correct the data, eliminating atmospheric and sensor effects and restoring the true radiometric brightness of the ground objects.

[0101] Geometric correction: During the acquisition process of satellite remote sensing data, geometric distortion occurs due to factors such as satellite attitude changes, Earth curvature, and terrain fluctuations, necessitating geometric correction. Appropriate ground control points are selected and geometric correction models, such as polynomial models and collinearity equation models, are established to correct the data and convert the image to its correct position in the geographic coordinate system.

[0102] Data cropping and stitching: Based on the scope of the survey area, the calibrated satellite remote sensing data is cropped to remove data from irrelevant areas. If the survey area is large and covered by multiple satellite images, image stitching is required to ensure that the stitched image is seamless and geometrically consistent.

[0103] Noise removal: Use filtering algorithms, such as median filtering and Gaussian filtering, to remove noise from satellite remote sensing data and improve image quality.

[0104] UAV LiDAR data preprocessing;

[0105] Point cloud denoising: UAV LiDAR data may contain noise points, such as those caused by swaying vegetation and reflections from ground debris, which require denoising. Methods such as statistical filtering, radius filtering, and voxel grid filtering are used to remove outliers and noise points, retaining valid ground and object point cloud data.

[0106] Coordinate conversion: Convert the coordinate system of the drone LiDAR data to the unified coordinate system required by the project, such as the WGS84 coordinate system or the National Geodetic Coordinate System. First, obtain the attitude data and GNSS positioning data of the drone during flight. Using the coordinate conversion model, convert the LiDAR point cloud data from the sensor coordinate system to the world coordinate system, and then convert it to the unified target coordinate system.

[0107] Point cloud resampling: To improve the uniformity and resolution of point cloud data, LiDAR point cloud data is resampled. Interpolation algorithms such as nearest neighbor interpolation and linear interpolation are used to increase the point cloud density in sparse areas and appropriately reduce the number of point clouds in dense areas to make the point cloud data more uniform.

[0108] Object separation: Utilizes point cloud height information and density features to separate ground points from object points (buildings, trees, vehicles, etc.). Use slope and curvature-based ground point extraction algorithms, such as the Progressive Enhancement Triangulation (PEARL) algorithm, to extract the ground point cloud and generate a digital elevation model (DEM), while also separating the object point cloud data.

[0109] IoT sensor data preprocessing;

[0110] Outlier detection and repair: IoT sensor data may contain outliers due to sensor failure, communication interference, and other reasons, requiring detection and repair. Statistical methods (Z-score, IQR) or machine learning algorithms (Isolation Forest, One-Class Support Vector Machine) are used to detect outliers. Interpolation methods (linear, polynomial) can be used to repair small numbers of outliers. Data with a large number of outliers or sensor failures should be marked as invalid and processed.

[0111] Standardize data formats: Different types of IoT sensors may output data in different formats. This data needs to be converted to a unified format, such as CSV or JSON, to facilitate subsequent data integration and processing. Define unified data fields and data types to ensure that each sensor data contains necessary information, such as timestamps, location information, and monitoring parameters.

[0112] Time synchronization: Because IoT sensors may have different collection times, data synchronization is required. Using the collection times of satellite remote sensing data and drone LiDAR data as a benchmark, the timestamps of IoT sensor data are calibrated to ensure consistency across the time dimension of multi-source data.

[0113] (3) Obtaining multi-source surveying and mapping datasets;

[0114] Integrate preprocessed satellite remote sensing data, drone LiDAR data, and IoT sensor data using a unified data format and coordinate system to create a multi-source mapping dataset. During the data integration process, perform a quality check to ensure data integrity, accuracy, and consistency. Record metadata such as data source, acquisition time, and preprocessing process to facilitate subsequent data management and traceability.

[0115] 2. Data fusion;

[0116] (1) Improved hierarchical Transformer model construction;

[0117] Model architecture design;

[0118] Construction of a basic hierarchical Transformer structure: A hierarchical structure of multi-layer Transformer encoders and decoders is adopted. Each encoder layer consists of a multi-head self-attention mechanism and a feedforward neural network, which is used to extract features and encode input data; the decoder also consists of a multi-head self-attention mechanism and a feedforward neural network. At the same time, the encoder-decoder attention mechanism is introduced to process contextual information during the decoding process.

[0119] Introduction of the spatiotemporal attention mechanism: A spatiotemporal attention module is added to the encoder and decoder of the hierarchical Transformer model. The spatiotemporal attention module processes the temporal and spatial dimensions of the data separately. For the temporal dimension, it leverages the correlation of time series to calculate attention weights between data at different timestamps, dynamically associating data at different times. For the spatial dimension, it considers the spatial position and geometric relationships of the data and calculates attention weights between data at different spatial locations, effectively fusing spatial data of different resolutions.

[0120] Multi-source data input interface design: To accommodate the diverse formats and characteristics of satellite remote sensing data, drone LiDAR data, and IoT sensor data, a multi-source data input interface was designed. Satellite remote sensing data was converted into image feature vectors, drone LiDAR data into point cloud feature vectors, and IoT sensor data into numerical feature vectors. An embedding layer was used to map these different feature vectors into a feature space of the same dimensionality for input into the hierarchical Transformer model for processing.

[0121] Model parameter initialization;

[0122] Pre-trained Transformer model parameters, such as those from BERT and ViT, were used as initial values. The model's embedding layer, attention mechanism parameters, and feedforward neural network parameters were fine-tuned based on the characteristics of the surveying and mapping data. The parameters of the spatiotemporal attention module were initialized based on the characteristics of the temporal and spatial dimensions. For example, the weight matrix for temporal attention was initialized to a uniform distribution, while the weight matrix for spatial attention was initialized to a Gaussian distribution related to spatial distance.

[0123] (2) Data fusion process;

[0124] Multi-source data input: The pre-processed satellite remote sensing data, UAV LiDAR data, and IoT sensor data are converted into feature vectors of uniform dimensions through the multi-source data input interface and input into the improved hierarchical Transformer model.

[0125] Hierarchical feature extraction and spatiotemporal association: In the model's underlying encoder, preliminary feature extraction is performed on the input multi-source data, and the spatiotemporal attention module dynamically associates data of different resolutions and timestamps. For spatial data of different resolutions, the spatiotemporal attention mechanism calculates attention weights based on spatial position and distance, fusing high-resolution detail features with low-resolution global features. For data of different timestamps, the spatiotemporal attention mechanism calculates attention weights based on temporal sequence and time interval, capturing the changing trends and correlations of the data over time. In the hierarchical encoder and decoder, features are gradually abstracted and fused. The spatiotemporal attention module at each layer further performs spatiotemporal association processing on the features output by the upper layer, ultimately obtaining a fused feature vector at the model's output layer.

[0126] Fusion data generation: The fused feature vectors output by the model are converted into a format that matches the input data, generating a fused multi-source mapping dataset. This fused dataset incorporates the spatiotemporal information and features of the multi-source data, providing a more comprehensive reflection of the actual conditions in the mapping area.

[0127] 3. Establishment and pruning of MTL-DNN mapping data analysis model;

[0128] (1) MTL-DNN model construction;

[0129] Network structure design;

[0130] Shared Layer Design: Located at the bottom layer of the MTL-DNN model, the shared layer is used to extract common features from multi-source mapping data, such as spatial, temporal, and spectral features. The shared layer can employ structures such as fully connected layers, convolutional neural networks (CNNs), or recurrent neural networks (RNNs), depending on the characteristics of the multi-source data. For satellite remote sensing imagery and drone LiDAR point cloud data, CNNs can be used for spatial feature extraction; for time series data from IoT sensors, RNNs or LSTMs can be used for temporal feature extraction.

[0131] Task-Specific Layer Design: Task-specific layers are designed based on the different requirements of the three tasks: terrain modeling, feature classification, and risk prediction. These layers are layered above the shared layers, with each task corresponding to an independent subnetwork to process the specific features of that task and output the task results. The subnetwork for the terrain modeling task can use fully connected layers or 3D convolutional layers to generate a three-dimensional model of the terrain; the subnetwork for the feature classification task can use a softmax layer to output feature class probabilities; and the subnetwork for the risk prediction task can use a regression layer to output a continuous value for the risk level.

[0132] Network connection method: The output of the shared layer serves as the input of the three task-specific layers at the same time. The task-specific layers are independent of each other, and features are only shared in the shared layer to achieve parameter sharing and feature reuse in multi-task learning.

[0133] Activation function and loss function selection;

[0134] Activation function: Use appropriate activation functions in shared layers and task-specific layers, such as ReLU, Sigmoid, and Tanh. The ReLU function can alleviate the vanishing gradient problem and increase model training speed, and is often used in hidden layers. The Sigmoid function and Softmax function are often used in the output layer for probabilistic output of classification tasks.

[0135] Loss Function: Choose the appropriate loss function based on the task type. Terrain modeling is a regression task and uses the mean squared error (MSE) loss function; feature classification is a classification task and uses the cross-entropy loss function; risk prediction can use either the mean squared error loss function or the cross-entropy loss function (when risk levels are divided into discrete categories), depending on the prediction objective. The total loss function for multi-task learning is the weighted sum of the loss functions for each task, with the weights adjusted based on the importance of each task and the size of the data.

[0136] (2) Model pruning;

[0137] Pruning strategy selection: Weight-based pruning: Calculate the absolute value of each weight in the network, set a pruning threshold, and set weights with absolute values ​​less than the threshold to zero, removing unimportant connections. This method is simple and effective, reducing the number of model parameters without significantly affecting model performance.

[0138] Setting the pruning threshold: Determine the appropriate pruning threshold through experimentation and verification. During the pruning process, gradually increase the pruning percentage. After each pruning, evaluate the model performance using the validation set. Stop pruning when the model performance drops below an acceptable level. For example, first prune 5% of the weights and evaluate the model performance. If the performance drop does not exceed 5%, continue increasing the pruning percentage until the performance drops below the set threshold.

[0139] Post-pruning model fine-tuning: Fine-tune the pruned model to restore performance lost due to pruning. During fine-tuning, retrain the model using the training data, setting the learning rate to 1 / 10 or lower than the pre-pruning value, and training for 10-20 epochs. This allows the model to adapt to the pruned network structure while retaining important connections and neurons, improving model accuracy and stability.

[0140] The target MTL-DNN mapping data analysis model is obtained: After pruning and fine-tuning, the model is the target MTL-DNN mapping data analysis model. While maintaining high performance, this model has fewer parameters and computational complexity, making it more suitable for deployment and operation in the cloud or on edge devices.

[0141] 4. Surveying and mapping data analysis;

[0142] (1) Model execution tasks;

[0143] Terrain Modeling: The model's terrain modeling task subnetwork receives the fused feature vectors and, by combining spatial and temporal features extracted from the shared layer with processing from the task-specific layer, outputs the terrain's 3D coordinates and elevation information. Using this information, 3D modeling algorithms, such as triangulation and surface fitting, are employed to generate a 3D terrain model of the surveyed area. During the modeling process, a digital elevation model (DEM) and digital surface model (DSM) can be combined to improve the accuracy and detail of the terrain modeling.

[0144] Object Classification: The object classification task subnetwork analyzes the fused features and outputs a probability distribution of object categories by combining spectral, texture, and shape features extracted from the shared layer with classifiers in the task-specific layer. Based on the preset object category labels, the category with the highest probability is selected as the object classification result, enabling the recognition and classification of different object types such as buildings, vegetation, water bodies, and roads. Post-processing techniques such as morphological filtering and region growing can be used to optimize the classification results and improve classification accuracy and consistency.

[0145] Risk Prediction: The risk prediction subnetwork utilizes the time series, spatial, and feature characteristics of the fused data, combined with historical risk data and relevant risk indicators (such as terrain slope, feature type, and rainfall). Using regression or classification algorithms, it outputs a risk level or probability of occurrence. For geological disaster risk prediction (landslides and debris flows), a risk prediction model can be established that comprehensively considers multiple factors, including topography, vegetation cover, and hydrological conditions, providing a basis for risk assessment and management in the surveyed area.

[0146] The above shows and describes 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 above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A cloud-based surveying and mapping data processing method, characterized in that: The cloud-based surveying and mapping data processing method comprises the following steps: Collect satellite remote sensing data, UAV LiDAR data, and IoT sensor data, pre-process the collected data to obtain a multi-source mapping dataset, and transmit the multi-source mapping dataset to a cloud platform; A hierarchical Transformer model is built on the cloud platform using the hierarchical structure of multi-layer Transformer encoders and decoders; In the model, each encoder layer consists of a multi-head self-attention mechanism and a feedforward neural network, and the decoder layer consists of a multi-head self-attention mechanism and a feedforward neural network. At the same time, the encoder-decoder attention mechanism is introduced to obtain an improved hierarchical Transformer model. Adding a spatiotemporal attention module to the encoder and decoder of the improved hierarchical Transformer model, and using the spatiotemporal attention module to process the temporal and spatial dimension information of the data respectively; Obtaining a multi-source surveying and mapping dataset in a cloud platform, and inputting the multi-source surveying and mapping dataset into an improved hierarchical Transformer model; The model's underlying encoder performs preliminary feature extraction on the input multi-source data, and dynamically associates data of different resolutions and timestamps through the spatiotemporal attention module; The spatiotemporal attention mechanism is used to calculate the attention weight according to the spatial position and distance, and the high-resolution detail features and low-resolution global features are fused to obtain a fused multi-source mapping dataset. Based on the MTL-DNN multi-task deep learning network, an MTL-DNN surveying and mapping data analysis model is established. The model is pruned using a fine-grained pruner to obtain the target MTL-DNN surveying and mapping data analysis model. The fused multi-source mapping data set is input into the target MTL-DNN mapping data analysis model to perform terrain modeling, object classification and risk prediction to obtain mapping data analysis results.

2. A cloud-based surveying and mapping data processing method according to claim 1, characterized in that: The collecting of satellite remote sensing data, UAV LiDAR data, and IoT sensor data, preprocessing the collected data to obtain a multi-source mapping data set, and transmitting the multi-source mapping data set to a cloud platform includes: Collecting satellite remote sensing data, and performing radiation correction and geometric correction on the satellite remote sensing data based on a model of atmospheric transmission theory to obtain first mapping data; Acquire UAV LiDAR data, perform point cloud denoising on the UAV LiDAR data using a statistical filtering method, and perform coordinate transformation and point cloud resampling on the denoised data using a linear interpolation method to obtain second surveying and mapping data; collecting IoT sensor data, performing outlier processing on the IoT sensor data using a polynomial interpolation method, and performing time synchronization on the outlier-processed data to obtain third surveying and mapping data; Normalizing the first surveying and mapping data, the second surveying and mapping data, and the third surveying and mapping data, and integrating the normalized data to obtain a multi-source surveying and mapping data set; The multi-source surveying and mapping dataset is encrypted, and the encrypted multi-source surveying and mapping dataset is transmitted to the cloud platform based on 5G communication.

3. The cloud-based surveying and mapping data processing method according to claim 1, wherein: The method of establishing an MTL-DNN surveying and mapping data analysis model based on the MTL-DNN multi-task deep learning network and pruning the model using a fine-grained pruner to obtain a target MTL-DNN surveying and mapping data analysis model includes: An MTL-DNN mapping data analysis model is established based on the MTL-DNN multi-task deep learning network. The shared layer is located at the bottom layer of the MTL-DNN model and is used to extract common features of the data. Design task-specific layers based on the different requirements of terrain modeling, feature classification, and risk prediction; The output of the shared layer is used as the input of the three task-specific layers at the same time. The task-specific layers are independent of each other, and features are only shared in the shared layer.

4. A cloud-based surveying and mapping data processing method according to claim 3, characterized in that: The method further includes: establishing an MTL-DNN surveying and mapping data analysis model based on the MTL-DNN multi-task deep learning network, and performing a pruning operation on the model using a fine-grained pruner to obtain a target MTL-DNN surveying and mapping data analysis model. Using the Adam optimizer, set the learning rate, momentum, and weight decay parameters of the MTL-DNN mapping data analysis model; Based on the fine-grained pruner, the absolute value of each weight in the network is calculated, a pruning threshold is set, and the weights with absolute values ​​less than the threshold are set to zero, and unimportant connections are removed to the target MTL-DNN mapping data analysis model.

5. The cloud-based surveying and mapping data processing method according to claim 1, wherein: The step of inputting the fused multi-source surveying and mapping data set into the target MTL-DNN surveying and mapping data analysis model to perform terrain modeling, object classification, and risk prediction to obtain surveying and mapping data analysis results includes: By extracting spatial and temporal features from the shared layer and combining them with the processing of the task-specific layer, the three-dimensional coordinates and height information of the terrain are output to obtain terrain modeling. The spectrum, texture, shape and other features extracted by the shared layer are combined with the classifier of the task-specific layer to output the category probability distribution of the ground object and obtain the ground object classification; By integrating the time series features, spatial features and land feature features in the data, combining historical risk data and relevant risk indicators, and using classification algorithms, the risk level or probability of risk occurrence is output to obtain risk prediction.

6. A cloud-based surveying and mapping data processing system, characterized in that: The cloud-based surveying and mapping data processing system includes the following modules: A data acquisition and transmission module is used to collect satellite remote sensing data, UAV LiDAR data, and IoT sensor data, pre-process the collected data to obtain a multi-source mapping data set, and transmit the multi-source mapping data set to the cloud platform; The mapping data fusion module is used to build a hierarchical Transformer model using the hierarchical structure of multi-layer Transformer encoders and decoders in the cloud platform; Each encoder layer in the model consists of a multi-head self-attention mechanism and a feedforward neural network, and the decoder consists of a multi-head self-attention mechanism and a feedforward neural network. At the same time, the encoder-decoder attention mechanism is introduced to obtain an improved hierarchical Transformer model; a spatiotemporal attention module is added to the encoder and decoder of the improved hierarchical Transformer model, and the spatiotemporal attention module is used to process the temporal and spatial dimension information of the data respectively; a multi-source mapping dataset is obtained from the cloud platform and input into the improved hierarchical Transformer model; the bottom-level encoder of the model is used to perform preliminary feature extraction on the input multi-source data, and the spatiotemporal attention module is used to dynamically associate data of different resolutions and timestamps; the spatiotemporal attention mechanism is used to calculate attention weights according to spatial position and distance, and the high-resolution detail features and low-resolution global features are fused to obtain a fused multi-source mapping dataset; The analysis model building module is used to build an MTL-DNN surveying and mapping data analysis model based on the MTL-DNN multi-task deep learning network, and use the pruner to prune the model to obtain the target MTL-DNN surveying and mapping data analysis model; The surveying and mapping data analysis module is used to input the fused multi-source surveying and mapping data set into the target MTL-DNN surveying and mapping data analysis model to perform terrain modeling, object classification and risk prediction, and obtain surveying and mapping data analysis results.

7. A cloud-based surveying and mapping data processing system according to claim 6, characterized in that: The surveying and mapping data fusion module includes the following submodules: Establish a submodule for building a hierarchical Transformer model in the cloud platform using the hierarchical structure of multi-layer Transformer encoders and decoders; A submodule is introduced, where each encoder layer in the model consists of a multi-head self-attention mechanism and a feedforward neural network, and the decoder layer consists of a multi-head self-attention mechanism and a feedforward neural network. At the same time, the encoder-decoder attention mechanism is introduced to obtain an improved hierarchical Transformer model; A submodule is added to add a spatiotemporal attention module to the encoder and decoder of the improved hierarchical Transformer model, and the spatiotemporal attention module is used to process the temporal dimension and spatial dimension information of the data respectively.

8. The cloud-based surveying and mapping data processing system according to claim 6, wherein: The surveying and mapping data fusion module also Includes the following submodules: An input submodule, configured to obtain a multi-source mapping dataset in the cloud platform and input the multi-source mapping dataset into the improved hierarchical Transformer model; The extraction submodule is used to perform preliminary feature extraction on the input multi-source data using the model's underlying encoder, and dynamically associate data of different resolutions and timestamps through the spatiotemporal attention module; The fusion submodule is used to use the spatiotemporal attention mechanism to calculate the attention weight according to the spatial position and distance, and to fuse the high-resolution detail features and low-resolution global features to obtain a fused multi-source mapping dataset.

Images