Machine vision-based surveying and mapping geographic information analysis method and system, and medium

Abstract

The application relates to the technical field of machine vision, and discloses a surveying and mapping geographic information analysis method and system based on machine vision and a medium. The method comprises the following steps: preprocessing multi-source remote sensing data, obtaining a standardized surveying and mapping data set, extracting ground object edges and texture information, obtaining multi-scale spatial features and semantic segmentation results, integrating multi-source data, obtaining a unified geographic feature representation map, identifying the unified geographic feature representation map, obtaining a surveying and mapping geographic feature set, performing Bayesian inference of spatial uncertainty quantization and parameter uncertainty quantization on the surveying and mapping geographic feature set, obtaining comprehensive uncertainty evaluation results, and constructing vectorized geographic information based on the surveying and mapping geographic feature set and the comprehensive uncertainty evaluation results. The application effectively retains the edge information of geographic features, improves the recognition accuracy of the geographic features, makes the geographic information expression more structured, and realizes automatic updating.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine vision, in particular to a surveying and mapping geographic information analysis method and system based on machine vision and a medium. BACKGROUND

[0002] With the rapid development of remote sensing technology, the acquisition of multi-source remote sensing data (optical images, radar data, point cloud data, etc.) has become increasingly convenient, providing a rich data basis for surveying and mapping geographic information analysis. However, these multi-source remote sensing data have strong heterogeneity, high dimensionality, and complex noise, and traditional geographic feature extraction methods mainly rely on manually designed features and rules, making it difficult to effectively utilize the complementary information of multi-source data and resulting in low recognition accuracy in complex terrain conditions. In particular, in complex scenes such as shadow areas, cloud coverage, and urban-rural junctions, existing methods often misjudge or miss, failing to meet the needs of high-precision mapping.

[0003] Another key challenge facing current surveying and mapping geographic information analysis is the problem of uncertainty assessment. Traditional geographic information extraction methods only output deterministic results, lacking quantitative evaluation of the reliability of the recognition results, which makes subsequent decision-making lack risk assessment basis. Existing uncertainty assessment methods are mostly based on simple statistical models, which cannot accurately describe the spatial correlation and parameter uncertainty in complex geographic scenes, especially when environmental factors such as lighting conditions, seasonal changes, and image resolution change, the comprehensive uncertainty assessment results are often unreliable, limiting the promotion and use of surveying and mapping results in key application fields. SUMMARY

[0004] The present application provides a surveying and mapping geographic information analysis method and system based on machine vision, which effectively preserves the edge information of geographic features, improves the recognition accuracy of geographic features, and makes the geographic information expression more structured and realizes automatic updating.

[0005] In a first aspect, the present application provides a surveying and mapping geographic information analysis method based on machine vision, which comprises:

[0006] Preprocessing multi-source remote sensing data to obtain a standardized surveying and mapping data set, and extracting ground object edge and texture information from the standardized surveying and mapping data set to obtain multi-scale spatial features and semantic segmentation results;

[0007] Integrating multi-source data based on the multi-scale spatial features and the semantic segmentation results to obtain a unified geographic feature representation map, and identifying geographic features from the unified geographic feature representation map to obtain a surveying and mapping geographic feature set;

[0008] perform Bayesian inference on the surveying geographic element set to obtain spatial uncertainty quantification and parameter uncertainty quantification, and obtain a comprehensive uncertainty evaluation result;

[0009] construct vectorized geographic information based on the surveying geographic element set and the comprehensive uncertainty evaluation result.

[0010] In a second aspect, the present application provides a surveying geographic information analysis system based on machine vision, comprising:

[0011] a preprocessing module configured to preprocess multi-source remote sensing data to obtain a standardized surveying data set, and extract ground object edges and texture information from the standardized surveying data set to obtain multi-scale spatial features and semantic segmentation results;

[0012] a geographic element recognition module configured to integrate multi-source data based on the multi-scale spatial features and the semantic segmentation results to obtain a unified geographic element representation map, and recognize geographic elements from the unified geographic element representation map to obtain a surveying geographic element set;

[0013] a Bayesian inference module configured to perform Bayesian inference on the surveying geographic element set to obtain spatial uncertainty quantification and parameter uncertainty quantification, and obtain a comprehensive uncertainty evaluation result;

[0014] a construction module configured to construct vectorized geographic information based on the surveying geographic element set and the comprehensive uncertainty evaluation result.

[0015] In a third aspect, a computer readable storage medium is provided, which stores instructions, when executed on a computer, causes the computer to perform the above-mentioned surveying geographic information analysis method based on machine vision.

[0016] In the technical solutions provided in the application, through multi-source remote sensing data adaptive preprocessing and multi-scale feature extraction, technical problems such as low recognition accuracy, missing uncertainty evaluation and information updating difficulty of traditional methods in a complex scene are effectively solved. The dynamic parameter adjustment denoising algorithm and the deep learning-based image enhancement technology introduced in the application effectively preserve the edge information of geographic features; the multi-scale feature extraction network with a five-layer pyramid structure combined with the dual-branch processing mechanism of direction sensitivity and regional aggregation significantly enhances the recognition ability of different types of geographic features; the feature transfer network of the "feature-space-semantic" three-layer transfer framework effectively solves the feature expression imbalance problem in multi-modal data fusion; the geographic feature perception residual block and the three-branch network architecture improve the recognition accuracy of geographic features; the Bayesian deep learning method combined with variational inference realizes accurate quantification of the uncertainty of the recognition result; the geographic feature relationship reasoning based on the graph neural network and the change detection technology of the deep similarity measurement network make the geographic information expression more structured and realize automatic updating. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0018] Figure 1 An embodiment schematic diagram of a surveying and mapping geographic information analysis method based on machine vision in the present application;

[0019] Figure 2 An embodiment schematic diagram of a surveying and mapping geographic information analysis system based on machine vision in the present application. DETAILED DESCRIPTION

[0020] The embodiment of the present application provides a surveying and mapping geographic information analysis method, system and medium based on machine vision. The terms "first", "second", "third", "fourth" and the like (if any) in the specification and claims of the present application and the above drawings are used to distinguish similar objects, and do not necessarily mean a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the term "includes" or "has" and any variation thereof is intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0021] For ease of understanding, the specific process of the embodiments of the present application is described below. Please refer to Figure 1 An embodiment of the method for analyzing geographic information based on machine vision in the embodiments of the present application includes the following steps:

[0022] Step S101, pre-processing the multi-source remote sensing data to obtain a standardized surveying and mapping data set, and extracting the ground object edge and texture information of the standardized surveying and mapping data set to obtain multi-scale spatial features and semantic segmentation results;

[0023] It can be understood that the execution subject of the present application can be a surveying and mapping geographic information analysis system based on machine vision, and can also be a terminal or a server, which is not limited here. The embodiments of the present application take the server as the execution subject for example.

[0024] Specifically, the automatic sensor parameter identification of multi-source remote sensing data is performed by automatically analyzing data meta-information and combining a deep learning model to classify data source types and extract feature parameters. These data include high-resolution optical images, radar data, and laser point cloud data. The spatial resolution, band range, and imaging characteristics of each data source are calibrated to generate parameter calibration results for each data source. According to the calibration results, color distribution histograms of optical images are calculated. By analyzing the color distribution characteristics of pixels in the image, a corresponding histogram mapping relationship is established, and image characteristic analysis data is obtained. These analysis data are used for noise reduction processing of optical images. The noise reduction process uses an adaptive filtering algorithm combining Gaussian filtering and bilateral filtering. The filtering parameters are automatically adjusted according to the image complexity to achieve edge-preserving noise reduction effect, thereby maximizing the suppression of noise interference while maintaining the clarity of feature boundaries. At the same time, point cloud data is subjected to outlier detection and removal processing. An outlier detection algorithm based on density clustering is used to remove noise and isolated points in low-density areas to obtain more accurate terrain feature expression. The noise-reduced optical image is subjected to cloud coverage and shadow area compensation processing. A generative adversarial network based on deep learning is used for image enhancement. By identifying cloud coverage and shadow areas, the enhanced image is automatically generated. The enhanced optical image, point cloud data set after outlier removal, and radar data are subjected to spatial resolution resampling and registration processing. An image registration algorithm based on a multi-scale pyramid is used to calibrate the spatial position in combination with the geographic coordinate system to ensure accurate correspondence of multi-source data in spatial position, and to obtain a standardized surveying and mapping data set. The standardized surveying and mapping data set is input into a multi-scale feature extractor to extract feature edges and texture information. The feature extractor uses a five-layer pyramid feature extraction network with different sizes of convolution kernels (such as 3x3, 5x5, 7x7, 9x9, and 11x11) for convolution operation at different levels to capture multi-scale spatial features and optimize the extraction effect of different scale features through an adaptive receptive field adjustment mechanism. During the feature extraction process, specific direction-sensitive convolution modules and regional aggregation convolution modules are used to detect the edge information of linear features (such as roads and rivers) and regional features (such as buildings and water bodies), respectively. Local response normalization operation is introduced to enhance high-frequency information and obtain fine feature edge and texture information.

[0025] The standardized mapping data set is input into a multi-scale feature extractor for convolution operation, a multi-scale convolution network with a five-layer pyramid structure is adopted, different sizes of convolution kernels (for example, 3x3, 5x5, 7x7, 9x9, 11x11) are used at each layer to perform convolution calculation on the input data, so as to extract initial feature maps of different scales. Due to the significant difference in the distribution of features of ground objects of different scales, the multi-scale convolution mechanism can effectively capture the spatial characteristics of linear ground objects (such as roads and rivers) and regional ground objects (such as buildings and water bodies). The initial feature maps are respectively input into a direction-sensitive convolution unit and a regional aggregation convolution unit for extraction of specific features. The direction-sensitive convolution unit adopts a convolution kernel with direction selectivity, which effectively captures the edge information and direction characteristics of linear ground objects by performing convolution operations in different directions such as horizontal, vertical and diagonal, and highlights the boundary features of linear ground objects such as roads and rivers in the feature maps. The regional aggregation convolution unit realizes the aggregation of the overall features of regional ground objects through a larger convolution receptive field combined with an adaptive weight distribution mechanism, enhances the feature expression ability of regional ground objects such as buildings and water bodies, and the two convolution units work together to strengthen and finely express the linear ground object features and the regional ground object features respectively. The extracted linear ground object features and regional ground object features are input into a local response normalization layer for high-frequency information enhancement processing. Local response normalization effectively highlights the ground object boundary and texture features by suppressing low-frequency components and enhancing high-frequency details in the feature map, improves the resolution of edge details, and generates enhanced edge and texture feature maps. On the basis of local response normalization, channel attention weighting processing is performed on the enhanced edge and texture feature maps. The channel attention mechanism calculates the weight contribution of different spectral channels in feature extraction, adaptively adjusts the influence degree of different channel information, so that the identification of different ground object materials is more targeted. The multi-scale spatial features weighted by channel attention are input into a full convolution neural network together with the enhanced edge and texture feature maps for semantic segmentation. The semantic segmentation network is based on a pre-trained U-Net architecture, combines multi-scale features for pixel-by-pixel classification, realizes accurate identification of ground object categories, and obtains a semantic segmentation result.

[0026] In step S102, multi-source data integration is performed based on the multi-scale spatial features and the semantic segmentation result to obtain a unified geographic feature representation map, and geographic feature recognition is performed on the unified geographic feature representation map to obtain a mapping geographic feature set.

[0027] Specifically, the multi-scale spatial features and semantic segmentation results are encoded, and the encoding process uses a multi-layer convolutional neural network to represent the features in a high dimension, and combines batch normalization and ReLU activation function to enhance the non-linear expression ability of the features, to obtain the encoded feature representation. The encoded feature representation is separated into core representation transmitted through an identity mapping and variable representation adjusted through an adaptive residual connection, wherein the core representation retains the inherent characteristics of geographic elements and stably transmits basic information, while the variable representation dynamically adjusts the features according to the environmental changes of multi-source data by introducing an adaptive residual connection, to adapt to different data characteristics, and the two feature forms jointly form a double-path feature flow. The double-path feature flow is subjected to cross-modal feature interaction, the information flow between different data modalities is controlled through a bidirectional gate mechanism, the correlation between different modal features is calculated, and the correlated features are integrated to generate preliminary fusion features. On the basis of the preliminary fusion features, spatial attention maps and channel attention maps are calculated to strengthen feature expression, the spatial attention maps are used to capture the importance information of geographic elements in spatial position, and the channel attention maps are used to measure the weight contribution of different spectral channels or data modalities in the feature extraction process, and a feature attention weight matrix is constructed through the two attention maps. According to the feature attention weight matrix, the preliminary fusion features are adaptively reweighted, in this process, the semantic segmentation results are input into a local context aggregation unit as prior knowledge for feature guidance, through the guidance of context information, the expression of features in the semantic related area is more accurate, and context enhanced features are generated. The context enhanced features are subjected to cross-scale residual fusion operation, through the introduction of a multi-scale feature fusion mechanism, the context enhanced features of different scales are subjected to residual connection, to realize adaptive supplement and information enhancement of cross-scale features, and a unified geographic element representation map is generated. The unified geographic element representation map is input into a geographic element recognition model, and a convolutional neural network based on an improved ResNet architecture is used for fine recognition, the recognition model combines a feature selection gate mechanism and a multi-task branch network, to realize simultaneous recognition and regression of geographic element boundaries, categories and attributes, so as to obtain a surveying and mapping geographic element set, including accurate geographic element boundaries, detailed category labels and attribute parameters (such as height, width, area, etc.), to form a structured geographic information representation.

[0028] The unified geographic feature representation map is input into a trunk network adopting an improved ResNet architecture, which combines a geographic feature perception residual block for feature extraction. The geographic feature perception residual block expands the receptive field by introducing a sequence of dilated convolutions, with the dilation rates set to 1, 2, 4, and 8 in sequence, so that the network captures feature information of geographic features at different scales to obtain multi-scale receptive field mapping features. The multi-scale receptive field mapping features are subjected to geographic feature contour feature extraction, which captures the boundary information of geographic features through a depth separable convolution network combined with a boundary attention mechanism, and generates a boundary mask of the geographic features. The boundary mask can effectively identify the contours of different geographic features, such as building contours, road boundaries, and water system boundaries. At the same time, the multi-scale receptive field mapping features are subjected to spatial distribution relationship analysis of geographic features, which captures the spatial correlation of geographic features through a context attention module, and generates a class probability map of the geographic features combined with a fully convolutional neural network. The probability map can predict the probability of each pixel belonging to different feature classes. On the basis of the multi-scale receptive field mapping features, attribute information of the geographic features is extracted, and the geometric and physical attributes of the features are calculated through a fully connected layer combined with an attention mechanism to obtain attribute parameter values of the geographic features, such as the height of a building, the width of a road, and the area of a water body. The boundary mask of the geographic features is subjected to boundary refinement processing, and a boundary optimization algorithm based on conditional random fields is adopted to optimize the rough boundary at the pixel level. The class label analysis module is combined to finely label the class information of the boundary region, and accurate boundary and class information of the geographic features are generated. This process can effectively avoid the boundary blur problem caused by resolution difference or noise interference, thereby improving the accuracy of feature recognition. After boundary refinement and class label analysis, the accurate boundary, class information, class probability map, and attribute parameter values of the geographic features are associated and integrated, and the multi-dimensional feature fusion module is used to cross-verify and weight-process the information from different sources to ensure the integrity and consistency of the recognition results, and finally a surveying and mapping geographic feature set containing building contours, road networks, and water system boundaries is formed.

[0029] Step S103, performing Bayesian inference of spatial uncertainty quantification and parameter uncertainty quantification on the surveying and mapping geographic feature set to obtain a comprehensive uncertainty evaluation result;

[0030] Specifically, the probability distribution data of the boundary position and attribute parameters of the geographic feature is calculated based on the surveyed geographic feature set, wherein the probability distribution of the boundary position is obtained by Monte Carlo rejection sampling and multiple forward propagation in the network inference stage, and the probability distribution of the attribute parameters is estimated by introducing the variational inference method into the Bayesian neural network, which provides basic information for subsequent uncertainty quantification. Based on the probability distribution data, the coefficient of variation of the boundary position of the geographic feature is calculated, the variance and mean of the boundary position are obtained by statistical analysis of the boundary position probability distribution at different spatial positions, and the coefficient of variation is calculated using these data, which reflects the uncertainty degree of the boundary position. According to the coefficient of variation, adaptive kernel density estimation is performed, and by dynamically adjusting the kernel function parameters, the kernel density function can more accurately fit the boundary position distribution of different types of geographic features, and more refined spatial uncertainty quantification results are obtained. At the same time, for the attribute parameters of the geographic feature, the Bayesian neural network is used to estimate the probability of the attribute parameters of the geographic feature, the variational inference method is used to approximate the posterior distribution, and the Bayesian loss function is constructed by combining the negative log-likelihood and the KL divergence term, which can effectively balance the prediction accuracy and distribution fitting ability of the model, and obtain the parameter uncertainty quantification result. Based on the spatial uncertainty quantification result and the parameter uncertainty quantification result, the two are jointly modeled, the Bayesian loss function is constructed for spatial correlation modeling, the variational graph autoencoder is introduced to model the spatial correlation of the geographic feature distribution, the implicit spatial correlation between different geographic features is captured, and the accuracy of uncertainty evaluation is strengthened at the spatial feature level, and the uncertainty evaluation index with enhanced spatial correlation is obtained. The spatial correlation enhanced uncertainty evaluation index is subjected to environmental factor sensitivity analysis, and the sensitivity analysis module simulates the influence of different environmental factors on the recognition accuracy of the geographic feature by introducing environmental factor parameters (such as light conditions, seasonal changes, image resolution, etc.), combines the spatial correlation enhanced uncertainty evaluation index, quantifies the influence degree of environmental factor changes on the recognition result, and generates a sensitivity atlas and an uncertainty change curve, forming a comprehensive uncertainty evaluation result including the class probability distribution, position uncertainty and attribute parameter probability distribution of the geographic feature.

[0031] Step S104, constructing vector geographic information based on the surveyed geographic feature set and the comprehensive uncertainty evaluation result.

[0032] Specifically, a multi-level geographic feature data structure is created based on the surveyed geographic feature set, which includes four levels of point feature layer, line feature layer, area feature layer and composite feature layer. The point feature layer records the key positions of inflection points and intersection points in geographic space, the line feature layer contains linear feature information such as roads and rivers, the area feature layer stores regional features such as buildings and water bodies, and the composite feature layer is used to express complex geographic units, such as functional areas such as residential areas and industrial areas. Based on the comprehensive uncertainty evaluation results, the threshold parameters of the elements in the line feature layer are adjusted, the position variation coefficient of the line feature boundary in the uncertainty quantification results is calculated, the low-precision boundary segments are automatically removed, and the adaptive Douglas-Peucker algorithm is used to simplify the boundary to obtain the simplified line feature data. After generating the multi-level geographic feature data structure, the geographic features are regarded as network nodes, and the spatial relationships between the features (such as proximity relationship, connectivity relationship, inclusion relationship, etc.) are regarded as network edges. Each node and edge is assigned a corresponding attribute feature vector, and a geographic feature relationship network is constructed in this way. This relationship network can capture the spatial topological characteristics between geographic features, and through the assignment of feature vectors, the functional attributes and spatial relationships between different features are effectively associated. On this basis, the geographic feature relationship network is learned for element relationship, graph neural networks are introduced for network embedding, the convolution operation is used to capture the feature interaction relationship between nodes, and the element group and its functional attributes are automatically identified to form the clustering characteristics and functional patterns of different features in spatial relationship. The identification of the element group can realize the automatic clustering of the same type of features and infer the area with specific functional characteristics, such as identifying residential areas through the arrangement of buildings and the structure of road networks, and identifying industrial areas through the distribution of industrial plants and logistics channels. By comparing the element group and its functional attributes with historical data, combining the comprehensive uncertainty evaluation results, and calculating the degree of change between the surveyed geographic feature set and the historical data through deep similarity measurement network, the change information of the geographic features is identified to generate the change detection result. According to the geographic feature change information, the simplified line feature data and the functional attributes of the element group, combined with the spatial relationship and time dimension change characteristics, the topological correction and data fusion mechanism are used to generate the vectorized geographic information, and the boundary information of the basic geographic features including building outline, road network and water system boundary is output, including feature attributes, functional characteristics and element change state.

[0033] In the embodiments of the present application, through adaptive preprocessing of multi-source remote sensing data and multi-scale feature extraction, the technical problems such as low recognition accuracy, lack of uncertainty evaluation and difficulty in information updating of traditional methods in complex scenes are effectively solved. The dynamic parameter adjustment denoising algorithm and the deep learning-based image enhancement technology introduced in the present application effectively preserve the edge information of geographic features; the multi-scale feature extraction network with a five-layer pyramid structure combined with the dual-branch processing mechanism of direction sensitivity and regional aggregation significantly enhances the recognition ability of different types of geographic features; the feature transfer network of the "feature-space-semantic" three-layer transfer framework effectively solves the feature expression imbalance problem in multi-modal data fusion; the geographic feature perception residual block and the three-branch network architecture improve the recognition accuracy of geographic features; the Bayesian deep learning method combined with variational inference realizes accurate quantification of the uncertainty of the recognition result; the geographic feature relationship reasoning based on graph neural network and the change detection technology of deep similarity measurement network make the geographic information expression more structured and realize automatic updating.

[0034] In a specific embodiment, the process of performing step S101 can specifically include the following steps:

[0035] Sensor parameter automatic identification is performed on the multi-source remote sensing data to obtain parameter calibration results of each data source, and the multi-source remote sensing data includes optical images, radar data and point cloud data;

[0036] Color distribution histogram calculation is performed on the optical images according to the parameter calibration results to obtain image characteristic analysis data;

[0037] Noise reduction processing is performed on the optical images based on the image characteristic analysis data to obtain edge-retained denoised images, and abnormal point detection is performed on the point cloud data to obtain a point cloud data set after removing abnormal points;

[0038] Cloud coverage and shadow area compensation processing are performed on the edge-retained denoised images to obtain enhanced processing images;

[0039] Spatial resolution resampling and registration processing are performed on the enhanced processing images, the radar data and the point cloud data set after removing abnormal points to obtain a standardized surveying and mapping data set;

[0040] The standardized surveying and mapping data set is input into a multi-scale feature extractor to extract geographic feature edge and texture information, and multi-scale spatial features and semantic segmentation results are obtained.

[0041] Specifically, a multi-source data automatic identification and calibration module is constructed, which automatically identifies different data sources based on the combination of metadata analysis, deep learning classification model and data feature analysis, and calibrates parameters according to data types and sensor characteristics. For optical images, the sensor model, shooting angle, spectral band information and resolution parameters are obtained by analyzing the image metadata, and the consistency of the metadata and the image content is ensured by combining the convolutional neural network for image content classification verification. For radar data, the synthetic aperture radar metadata analysis and feature matching algorithm are used to automatically calibrate parameters such as image polarization mode, incident angle, echo intensity, etc., and the accuracy of the parameters is verified by feature extraction of different polarization channels. For point cloud data, the scanning mode, point cloud density, scanning angle and distance resolution of the laser radar are automatically judged by combining density distribution analysis with point cloud metadata analysis, and accurate point cloud data calibration is achieved. After automatic identification and parameter calibration, the parameter calibration results of each data source are formatted to obtain the parameter calibration results of each data source. According to the parameter calibration results, the color distribution histogram of the optical image is calculated, the optical image is grayed and the RGB channel is split, the pixel intensity histogram of each channel is calculated, and the normalized color distribution histogram is obtained through normalization processing, reflecting the color characteristics, brightness distribution and image contrast characteristics of the image. Based on the image characteristic analysis data, noise reduction processing is performed on the optical image, and an edge-preserving noise reduction method combining adaptive Gaussian filtering and bilateral filtering is used. Gaussian filtering can effectively remove random noise in the image, while bilateral filtering can preserve the edge details of the image by combining spatial domain and pixel intensity domain information, avoiding the occurrence of boundary blur, and obtaining an edge-preserving denoised image. At the same time of noise reduction processing, abnormal point detection is performed on the point cloud data, the spatial density distribution of the point cloud data is analyzed by density clustering algorithm, the abnormal points in low density area are identified and removed, and the point cloud data set after removing abnormal points is obtained, which effectively avoids the distortion of terrain features caused by isolated points or noise points in the point cloud data. The edge-preserving denoised image is subjected to cloud layer covering and shadow area compensation processing, and the image inpainting technology based on generative adversarial network is used to fill the semantic of the cloud covering area and perform adaptive pixel compensation combined with the texture features of the adjacent area. At the same time, for the shadow area, the spectral features of the occluded area are restored by brightness correction and color transfer algorithm, and the enhanced image is obtained. The enhanced optical image, radar data and point cloud data set after removing abnormal points are subjected to spatial resolution resampling and registration processing.The nearest neighbor interpolation method and the bicubic interpolation method are used to match the spatial resolution of the optical image. The radar data and the point cloud data are matched with the optical image in spatial position by the multi-scale pyramid matching algorithm. In this process, a similarity measure function (such as mutual information or mean square error) is introduced to optimize and adjust the registration result, so as to ensure the accurate correspondence of multi-source data in spatial position, and generate a standardized surveying and mapping data set. In the process of inputting the standardized surveying and mapping data set into the multi-scale feature extractor for extracting the edge and texture information of the ground object, a multi-scale feature extraction network with a five-layer pyramid structure is constructed. Each layer uses different sizes of convolution kernels (such as 3x3, 5x5, 7x7, 9x9, and 11x11) for convolution operation to capture the features of geographical elements at different scales. The feature extractor combines the direction-sensitive convolution and the region-aggregated convolution mechanisms. The direction-sensitive convolution is used to extract the features of linear ground objects (such as roads and rivers) by horizontal, vertical and diagonal convolution operations to capture the boundary characteristics of linear ground objects. The region-aggregated convolution is used to aggregate the features of regional ground objects (such as buildings and water bodies) by an adaptive receptive field adjustment mechanism to optimize the extraction of regional features. After multi-scale feature extraction, the feature map is input into the local response normalization layer for high-frequency information enhancement processing to further preserve the edge and texture features of the ground object. The channel attention module is used to weight and optimize the extracted feature map, and the features are adaptively reweighted according to the importance of different data channels to highlight the key features. Then, the enhanced feature map is input into the fully convolutional neural network for semantic segmentation. The network is based on the U-Net architecture, which restores the spatial resolution layer by layer through the encoding-decoding structure and realizes pixel-by-pixel semantic classification to obtain multi-scale spatial features and semantic segmentation results, and finally forms a complete feature representation containing ground object categories, boundary information and texture features.

[0042] In a specific embodiment, the step of inputting the standardized surveying and mapping data set into the multi-scale feature extractor to extract the edge and texture information of the ground object and obtain the multi-scale spatial features and semantic segmentation results can specifically include the following steps:

[0043] Inputting the standardized surveying and mapping data set into the multi-scale feature extractor for convolution operation to obtain a plurality of initial feature maps;

[0044] Inputting the plurality of initial feature maps into the direction-sensitive convolution unit and the region-aggregated convolution unit respectively for processing to obtain linear ground object features and regional ground object features;

[0045] Inputting the linear ground object features and the regional ground object features into the local response normalization layer for high-frequency information enhancement processing to obtain enhanced edge and texture feature maps;

[0046] The enhanced edge and texture feature maps are subjected to channel attention weighting to obtain multi-scale spatial features, and the enhanced edge and texture feature maps are input into a full convolutional neural network for semantic segmentation to obtain a semantic segmentation result.

[0047] Specifically, a feature extractor with multi-scale feature extraction capability is constructed, which adopts a five-layer pyramid structure, and each layer uses different sizes of convolution kernels for convolution operation. The sizes of the convolution kernels are 3x3, 5x5, 7x7, 9x9 and 11x11 respectively. Each convolution kernel performs convolution calculation on the input data in different scale spatial receptive fields, thereby capturing the edge, texture and spatial distribution characteristics of geographic features at different resolutions. The standardized mapping data set is taken as the input, and after the layer-by-layer convolution operation of the feature extractor, multiple initial feature maps are formed, which contain spatial information at different scales and preserve the spatial consistency between multi-source data. Since different geographic features have different spatial scale characteristics in remote sensing images, for example, linear features such as roads and rivers can be captured in a small receptive field, while regional features such as buildings and water bodies need a larger receptive field to be completely identified, so multi-scale convolution can effectively cover the features of different scale geographic features, improving the completeness and accuracy of feature expression. The multiple initial feature maps are input into the direction-sensitive convolution unit and the region-aggregated convolution unit for feature processing. The direction-sensitive convolution unit uses multi-directional convolution kernels to capture the direction information of linear features, extract the boundary features of linear features such as roads and rivers, and perform direction-aware feature enhancement on these boundaries. At the same time, the region-aggregated convolution unit introduces an adaptive aggregation mechanism to spatially aggregate regional features at different scales, capturing the overall features of regional features such as the shape, area and texture characteristics of buildings and water bodies in a larger receptive field. The combination of the direction-sensitive convolution unit and the region-aggregated convolution unit allows linear feature characteristics and regional feature characteristics to be enhanced in the feature map. The linear feature characteristics and the regional feature characteristics are input into the local response normalization layer for high-frequency information enhancement processing. The local response normalization layer normalizes the activation response of each convolution kernel, effectively suppressing the influence of low-frequency components in the feature map and enhancing the expression of high-frequency information, thereby highlighting the edge and texture characteristics of the features. By adaptively adjusting the normalization parameters, local response normalization dynamically adjusts the feature weights in feature maps at different scales, ensuring balanced expression of boundary information and regional features at different spatial scales, thereby generating enhanced edge and texture feature maps. The enhanced edge and texture feature maps are subjected to channel attention weighting processing. The channel attention mechanism calculates the importance of different feature channels to dynamically adjust the weight contribution of each channel in feature expression, thereby highlighting key feature information and improving the feature selection capability of the model. In the channel attention weighting process, the spatial information of the feature map is mapped to a global feature vector through global average pooling for spatial compression, and then the weights of each channel are calculated through two fully connected networks and remapped to the original feature map to achieve adaptive weighting at the channel level.Through the channel attention mechanism, the attention to important feature of ground objects is enhanced, and the interference of redundant features is suppressed, obtaining multi-scale spatial features with better feature expression ability. The enhanced edge and texture feature maps are input into a full convolutional neural network for semantic segmentation. The network uses an improved version of the U-Net architecture, decodes the feature maps layer by layer through the encoding-decoding structure, gradually restores the spatial resolution, and combines multi-scale features to realize pixel-by-pixel classification, so as to identify the ground object categories. In the encoding stage, multi-scale convolution is used to capture ground object features of different scales, and residual connection is used to retain key boundary information and texture features; in the decoding stage, spatial information is restored through deconvolution operation, and the features in the encoding stage are introduced into the decoding path through the jump connection, so as to ensure the complete fusion of semantic information and spatial information. Through the semantic segmentation process of the full convolutional neural network, basic ground object categories such as roads, rivers, buildings and water bodies are identified, and different types of ground objects in complex scenes are accurately distinguished, realizing fine expression of multi-scale spatial features.

[0048] In a specific embodiment, the process of performing step S102 can specifically include the following steps:

[0049] Performing encoding processing on the multi-scale spatial features and the semantic segmentation result to obtain an encoded feature representation;

[0050] Separating the encoded feature representation into a core representation transmitted through an identity mapping and a variable representation adjusted through an adaptive residual connection, and taking the core representation and the variable representation as a dual-path feature flow;

[0051] Performing cross-modal feature interaction on the dual-path feature flow to obtain a preliminary fusion feature;

[0052] Based on the preliminary fusion feature, a spatial attention map and a channel attention map are calculated, and a feature attention weight matrix of the spatial attention map and the channel attention map is constructed;

[0053] According to the feature attention weight matrix, the preliminary fusion feature is adaptively reweighted, and the semantic segmentation result is input into a local context aggregation unit as prior knowledge for feature guidance to obtain a context-enhanced feature;

[0054] Performing cross-scale residual fusion operation on the context-enhanced feature to obtain a unified geographic element representation map;

[0055] Performing geographic element recognition on the unified geographic element representation map to obtain a surveying and mapping geographic element set.

[0056] Specifically, the multi-scale spatial features and semantic segmentation results are taken as inputs, and a feature encoder based on a convolutional neural network is used to compress and map the high-dimensional features. The feature encoder is composed of multiple convolutional units and batch normalization modules. Through layer-by-layer convolution, the spatial dimensions of the feature map are gradually compressed, while the feature dimensions are gradually increased, and the non-linear expression ability of the features is enhanced by combining the ReLU activation function. The convolutional unit captures local information of ground features through convolution kernels of different sizes and gradually forms a high-dimensional mapping result of multi-scale spatial features. The semantic segmentation results are used as auxiliary information in the feature encoding process and participate in feature remodeling together with multi-scale features. After this encoding process, the encoded feature representation is obtained. The encoded feature representation is separated into core representation transmitted through an identity mapping and variable representation adjusted through an adaptive residual connection. The core representation preserves the basic spatial and semantic information of the features through a direct identity mapping path. These information remains stable throughout the feature processing process and is not affected by external environmental changes, ensuring that the model can maintain the consistency of features at different scales. The variable representation is adjusted through an adaptive residual connection. The residual connection uses feature mapping of different scales to weight and fuse the feature information of different layers through a dynamic weight adjustment mechanism, enabling the variable representation to adaptively adapt to the changes of ground feature in different modal data, thereby improving the generalization ability of the features. The core representation and the variable representation form a feature expression structure as a double-path feature flow. One path stably transmits core information, and the other path dynamically adjusts feature weights through a residual mechanism. The two paths are cross-modality feature interaction. A bidirectional gating unit is used to control the information interaction between different modal feature flows. The bidirectional gating unit contains a parameterized gating mechanism that adaptively controls the information flow of different modal features during the interaction process. By calculating the feature attention weight of each path, the transmission proportion of the features is dynamically adjusted, effectively maintaining the information integrity of different modal features during feature fusion. Through cross-modality feature interaction, information fusion from different data sources and different perception scales is realized, and preliminary fusion features are formed. Based on the preliminary fusion features, a spatial attention map and a channel attention map are calculated, and a feature attention weight matrix is constructed. The spatial attention map calculates the feature importance of each pixel position through an adaptive convolution kernel, dynamically adjusts the feature weight according to the feature complexity of different spatial regions, and enhances the expression ability of key ground feature. The channel attention map sorts the importance of different feature channels through global pooling and weight calculation, and highlights the contribution of different channels in feature fusion through an adaptive weighting mechanism, thereby optimizing the channel information of feature expression. Through joint feature weighting calculation of the spatial attention map and the channel attention map, a feature attention weight matrix is constructed. This matrix can adaptively weight and adjust the preliminary fusion features, so that key features are enhanced at different scales and modalities, thereby improving the accuracy of feature expression.According to the characteristic attention weight matrix, the preliminary fused features are adaptively reweighted, and the semantic segmentation result is input as prior knowledge into the local context aggregation unit for feature guidance. The local context aggregation unit captures local context information in the feature space, models the spatial correlation of the features, and uses the semantic segmentation result as guidance information to guide the feature fusion direction of different spatial regions, thereby optimizing the feature discrimination degree between different ground object categories. Through the context feature guidance mechanism, the accuracy of the ground object boundary feature is effectively enhanced, and the aggregation ability of the semantic related region feature is improved, and the context enhanced feature is obtained. The cross-scale residual fusion operation is performed on the context enhanced feature, and the residual connection mechanism of different scale features is introduced to fuse the context enhanced features of different scales. The cross-scale residual fusion adjusts the fusion ratio of features of different scales through the superposition and weighting of multi-scale feature maps, thereby ensuring the consistency and integrity of the ground object features at different spatial scales. Through cross-scale feature fusion, the expression ability of ground object features at different scales is effectively enhanced, and the robustness of the features is maintained in complex ground object scenes, and a unified geographic element representation map is generated. The unified geographic element representation map is input into the geographic element recognition model for accurate recognition of the ground object category, boundary and attribute. The geographic element recognition model adopts an improved ResNet architecture, combines a sequence of atrous convolutions and a multi-scale feature perception module, captures the boundary information of the ground object in different receptive fields, and extracts the boundary, category and attribute features of the ground object through multiple task branches respectively. The boundary recognition branch realizes accurate positioning of the ground object contour through a depth separable convolution network, the category recognition branch performs category mapping of spatial features through a global attention mechanism, and the attribute regression branch estimates the parameters of the ground object attribute through a fully connected layer. After feature recognition and fusion processing, a set of surveying and mapping geographic elements is finally obtained, which includes structured information such as building contours, road networks and water system boundaries, and forms a complete geographic element expression in combination with attribute parameters and spatial relationships.

[0057] In a specific embodiment, the process of performing step of performing geographic element recognition on the unified geographic element representation map to obtain a set of surveying and mapping geographic elements can specifically include the following steps:

[0058] The unified geographic element representation map is input into the geographic feature perception residual block of the backbone network for processing to obtain multi-scale receptive field mapping features;

[0059] The multi-scale receptive field mapping features are subjected to geographic element contour feature extraction to obtain a boundary mask of the geographic element;

[0060] The multi-scale receptive field mapping features are subjected to spatial distribution relationship analysis of the geographic element to obtain a category probability map of the geographic element;

[0061] Attribute information of the geographic element is extracted from the multi-scale receptive field mapping feature to obtain an attribute parameter value of the geographic element.

[0062] Boundary refinement and category label analysis are performed on the boundary mask of the geographic element to obtain an accurate boundary of the geographic element and category information of the geographic element.

[0063] The accurate boundary of the geographic element, the category information of the geographic element, the category probability map of the geographic element, and the attribute parameter value of the geographic element are associated and integrated to obtain a surveying and mapping geographic element set containing building contours, road networks, and water system boundaries.

[0064] Specifically, the unified geographic feature representation map is input into a backbone network with multi-scale feature capture capability. The network adopts an improved ResNet architecture and introduces a geographic feature-aware residual block to enhance the network's multi-scale perception of geographic feature. The geographic feature-aware residual block expands the receptive field by combining dilated convolution with convolution kernels of different scales and dynamically adjusts the weight distribution of features at different scales by introducing an attention mechanism, achieving accurate capture of feature information at different spatial scales. The dilated rate of the dilated convolution is set to 1, 2, 4, and 8 respectively to ensure that the network captures boundary, texture, and spatial distribution information at different scales within different receptive field ranges, generating multi-scale receptive field mapping features. The multi-scale receptive field mapping features are subjected to geographic feature contour feature extraction. The boundary extraction network uses depth separable convolution combined with a boundary attention mechanism to capture boundary information by layer-by-layer convolution of feature maps at different scales, and optimizes the weights of pixels in the boundary region through the boundary attention mechanism to generate a boundary mask of geographic features. At the same time, the boundary extraction network automatically suppresses background noise and interference from non-related areas during the acquisition of the boundary mask, thereby improving the accuracy and robustness of boundary detection. Meanwhile, the multi-scale receptive field mapping features are subjected to spatial distribution relationship analysis of geographic features to generate a class probability map of geographic features. The semantic feature distribution network is introduced, which adopts a fully convolutional neural network architecture and combines with a spatial attention mechanism to dynamically weight the feature responses at different spatial locations, achieving accurate modeling of the spatial distribution of features. During spatial distribution relationship analysis, the multi-scale feature information is used to model the spatial correlation of different feature classes, and the context aggregation module is used to capture the spatial relationship between adjacent pixels to generate an accurate class probability map that reflects the probability distribution of each pixel belonging to different feature classes. The multi-scale receptive field mapping features are subjected to attribute information extraction of geographic features to obtain attribute parameter values of geographic features. This is achieved through an attribute regression network that combines fully connected layers and feature pooling modules to globally pool the spatial dimensions of the feature map and combine the feature channel weighting mechanism to regress different types of feature attributes and obtain attribute parameter values of geographic features, such as building height, road width, water area, etc. The boundary mask of geographic features is subjected to boundary refinement and class label analysis to obtain accurate boundary and class information of geographic features. Boundary refinement introduces a conditional random field for pixel-level boundary optimization, which models the spatial consistency of adjacent pixels to fine-tune the classification results in the boundary region, thereby eliminating the ambiguity in the boundary region and improving the accuracy of boundary positioning. Class label analysis uses a multi-task learning mechanism to jointly input the class probability map and the boundary mask into the class recognition network, and uses an attention-guided feature weighting mechanism to achieve accurate classification of feature classes and generate accurate class information of geographic features.The accurate boundary of the geographic element, the category information of the geographic element, the category probability map of the geographic element and the attribute parameter value of the geographic element are associated and integrated, and these information are spatially aligned and fused through the feature fusion network to form a surveying and mapping geographic element set. The feature fusion network constructs a spatial relationship graph of geographic elements by introducing a graph neural network, regards different elements as nodes and the spatial relationship between elements as edges, and learns and optimizes the features of nodes and edges through a graph convolution network to model the spatial correlation between different elements. Under the comprehensive calculation of the feature fusion network, a surveying and mapping geographic element set containing building contours, road networks and water system boundaries is generated.

[0065] In a specific embodiment, the process of performing step S103 can specifically include the following steps:

[0066] Based on the surveying and mapping geographic element set, the probability distribution data of the geographic element boundary position and the geographic element attribute parameter are calculated;

[0067] Based on the probability distribution data, the coefficient of variation of the geographic element boundary position is calculated, and adaptive kernel density estimation is performed according to the coefficient of variation to obtain a spatial uncertainty quantification result;

[0068] The probability estimation of the geographic element attribute parameter is calculated using a Bayesian neural network to obtain a parameter uncertainty quantification result, and a Bayesian loss function is constructed based on the spatial uncertainty quantification result and the parameter uncertainty quantification result;

[0069] Based on the Bayesian loss function, the surveying and mapping geographic element set is modeled for spatial correlation to obtain a spatial correlation enhanced uncertainty evaluation index;

[0070] The spatial correlation enhanced uncertainty evaluation index is subjected to environmental factor sensitivity analysis to generate a comprehensive uncertainty evaluation result including the category probability distribution of the geographic element, the position uncertainty and the attribute parameter probability distribution.

[0071] Specifically, the set of surveyed geographic features is taken as input, and the Monte Carlo dropout technique with multiple forward passes is used to model the uncertainty of the boundary positions and attribute parameters of the geographic features. By introducing a stochastic dropout layer during inference, a different feature representation is obtained each time the forward pass is performed, thereby forming a probability distribution data for the boundary positions and attribute parameters. After multiple forward passes, the mean and variance distribution of each geographic feature boundary position is obtained, and the regression results of the geographic feature attribute parameters (such as building height, road width, water area, etc.) are statistically analyzed to form a probability distribution matrix. Based on the probability distribution data, the coefficient of variation of the geographic feature boundary position is calculated, and the uncertainty is measured by calculating the ratio of the standard deviation to the mean. For each geographic feature boundary point, the position change during multiple forward passes is calculated, and the uncertainty level of the boundary is measured by the ratio of the variance to the mean. The larger the boundary coefficient of variation, the higher the uncertainty of the boundary position, and vice versa. According to the calculated coefficient of variation, adaptive kernel density estimation is performed, and the kernel density estimation is used to fit the probability distribution of the boundary position to form a spatial probability density function. This function dynamically adjusts the width of the kernel function according to the size of the coefficient of variation, so as to more accurately fit the uncertainty distribution of the boundary position and generate a spatial uncertainty quantification result. The key of adaptive kernel density estimation is to adjust the smoothing parameter of the kernel function according to the variation characteristics of different geographic features, so that the kernel density function can accurately reflect the variation law of different geographic feature boundaries in different spatial regions, and provide more reliable uncertainty quantification results for subsequent spatial correlation modeling. At the same time, in order to quantify the uncertainty of the attribute parameters of the geographic features, Bayesian neural network is used to calculate the probability estimation of the attribute parameters of the geographic features. The Bayesian neural network models the uncertainty of the parameters by introducing a probability distribution on the network weights. During the training process, the posterior distribution is approximated as a set of parameterized distributions through variational inference, and the model is optimized by maximizing the lower bound of evidence. During inference, the mean and variance of the attribute parameters are obtained through multiple sampling to form a probability distribution of the attribute parameters, which reflects the fluctuation of the attribute parameters of the geographic features under different environmental conditions. Based on the spatial uncertainty quantification result and the parameter uncertainty quantification result, a Bayesian loss function is constructed, which consists of a negative log-likelihood term and a KL divergence term. The former ensures the fitting accuracy of the model to the observed data, and the latter prevents the model from overfitting by limiting the distance between the posterior distribution and the prior distribution. The goal of the Bayesian loss function is to optimize the uncertainty quantification result and improve the prediction reliability and generalization ability of the model. Based on the Bayesian loss function, the spatial correlation of the set of surveyed geographic features is modeled to obtain the uncertainty evaluation index enhanced by spatial correlation.The spatial correlation modeling is completed by using a variational graph autoencoder, which constructs geographic features and their spatial relationships into an undirected graph, with nodes representing geographic features and edges representing spatial relationships between features. The node features are encoded by a graph convolution network, and the uncertainty distribution of nodes and edges is captured by variational inference to generate a spatial correlation enhanced uncertainty evaluation index. By modeling with a variational graph autoencoder, the spatial correlation of geographic features is captured, and the potential association between different geographic features is identified through graph structure learning, thereby significantly improving the accuracy of uncertainty evaluation. The spatial correlation enhanced uncertainty evaluation index is subjected to environmental factor sensitivity analysis. The sensitivity analysis simulates the multi-dimensional disturbance of the uncertainty evaluation results by introducing environmental factor variables (such as lighting conditions, seasonal changes, resolution changes, etc.), and quantifies the influence of different environmental factors on the recognition results of geographic features through variational sensitivity analysis (VSA). During the sensitivity analysis, the parameters of different environmental factor variables are disturbed, and the changes in the uncertainty evaluation results are observed. By calculating the correlation coefficient between the uncertainty change and the environmental factor change, an environmental factor sensitivity map is generated, reflecting the sensitivity of different feature categories, boundary positions and attribute parameters to environmental factor changes. Through the results generated by the sensitivity analysis, comprehensive uncertainty evaluation results including the category probability distribution of geographic features, position uncertainty and attribute parameter probability distribution are obtained. The evaluation results reflect the recognition accuracy of geographic features under different scales and environmental conditions, and quantify the uncertainty level of boundary positions and attribute parameters under different conditions.

[0072] In a specific embodiment, the process of performing step S104 can specifically include the following steps:

[0073] Based on the set of surveying and mapping geographic features, a multi-level geographic feature data structure including a point feature layer, a line feature layer, a surface feature layer and a composite feature layer is created. The point feature layer contains inflection points and intersection points, the line feature layer contains roads and rivers, the surface feature layer contains buildings and water bodies, and the composite feature layer contains residential areas and industrial areas.

[0074] Based on the comprehensive uncertainty evaluation results, the elements in the line feature layer are adjusted by threshold parameters to obtain a simplified line feature data;

[0075] Based on the multi-level geographic feature data structure, the geographic features are taken as nodes, the spatial relationships between features are taken as edges, and attribute feature vectors are assigned to each node and edge to obtain a geographic feature relationship network;

[0076] The geographic feature relationship network is subjected to feature relationship learning to obtain feature groups and their functional attributes;

[0077] Based on the element group, the function attribute and the comprehensive uncertainty evaluation result, the surveying and mapping geographic element set is compared with historical data to obtain geographic element change information.

[0078] According to the geographic element change information, the simplified line element data and the element group and the function attribute, vectorized geographic information is generated.

[0079] Specifically, the survey geographic feature set is processed hierarchically, and different types of geographic features are hierarchically divided according to spatial attributes, morphological characteristics and functional attributes to create a four-layer multi-level geographic feature data model. The point feature layer is used to represent the key point information in space, including the inflection point and the intersection point. These point features are located at road intersections, river junctions or building corner inflection points, and are important nodes of the spatial structure of geographic features. The line feature layer contains linear geographic features such as roads and rivers, which are represented by connecting point features to form complete line features. Road features record the spatial path of the transportation network, while river features record the spatial flow direction of the water system. The area feature layer is used to represent geographic features with area characteristics, including buildings and water bodies. The boundary and spatial range of the area feature are expressed by a closed polygon, and the building feature provides spatial occupation information of the geographic feature, while the water feature records the distribution and area characteristics of the water area. The composite feature layer is used to represent regions with complex functions and containing multiple basic features, including residential areas and industrial areas. These regions are composed of multiple buildings, roads and infrastructure, and the basic features are spatially aggregated and functionally associated to form a composite feature layer with specific functional attributes. Based on the comprehensive uncertainty evaluation results, the threshold parameters of the line feature layer are adjusted to obtain the simplified line feature data. Threshold parameter adjustment is mainly for linear geographic features such as roads and rivers. According to the spatial uncertainty quantification results and the coefficient of variation, the retention threshold of different line features is dynamically adjusted. Through adaptive optimization of the Douglas-Peucker algorithm, the geometric accuracy of line features is ensured while redundant points are removed and the expression accuracy of the boundary is optimized. When simplifying line features, the algorithm focuses on retaining areas with high variation and simplifying areas with low variation or noise interference, thereby effectively reducing data volume and improving the expression efficiency of line features. After threshold parameter adjustment, the simplified line feature data can maintain the spatial structure characteristics of the line feature and accurately express the boundary of the geographic feature while significantly reducing the data volume. Based on the multi-level geographic feature data structure, geographic features are taken as nodes, and the spatial relationship between features is taken as edges. Each node and edge is assigned an attribute feature vector to generate a geographic feature relationship network. Point features, line features, area features and composite features are nodes in the network, and the spatial adjacency relationship, topological relationship and functional dependence between different features are edges in the network. The attribute information, spatial coordinates and functional characteristics of each feature are mapped to a high-dimensional feature vector through the node feature embedding mechanism, and the attribute feature vector assigned to the edge includes spatial distance, topological adjacency weight and spatial direction information, forming a multi-attribute and multi-dimensional geographic feature relationship network. Based on the geographic feature relationship network, the feature relationship network is learned to obtain the feature group and its functional attributes.The element relationship learning adopts a graph neural network to model the spatial topological structure, and iteratively updates the features of nodes and edges through graph convolution and attention mechanism, so as to capture the high-order spatial correlation and functional attribute relationship between elements. In the element relationship learning process, the feature embedding of nodes and edges is updated continuously, so that the correlation between elements gradually emerges, and element groups with similar spatial features and functional attributes are automatically formed. These element groups can reflect the topological relationship of spatially adjacent elements, and reveal the functional attribute patterns between different elements. After completing the element relationship learning, the change information of geographic elements is identified by comparing the set of surveying and mapping geographic elements with historical data based on the element groups and their functional attributes, and combining the comprehensive uncertainty evaluation results. The change detection compares the features of the current data and the historical data through a deep similarity measurement network, performs multi-scale feature matching in the spatial, semantic and attribute dimensions, and calculates the change matrix of geographic elements. Through threshold analysis, the change matrix is classified and judged to identify the added, disappeared and changed element categories and attribute information. In the change detection process, the functional attribute information of the element groups is used to optimize the change detection results. By combining the functional features of the element groups, false changes caused by noise interference or environmental changes are effectively excluded, thereby improving the accuracy and reliability of the change detection. According to the geographic element change information, the simplified line element data, and the element groups and their functional attributes, complete vector geographic information is generated. By fusing the change information with the simplified line element data, the boundary and attribute information of the elements are accurately updated. At the same time, the functional attribute information of the element groups is used to reconstruct the composite element layer, so as to update the spatial structure and functional attributes of complex areas such as residential areas and industrial areas. The generated vector geographic information contains the accurate boundary information of points, lines, surfaces and composite elements, and integrates attribute parameters, spatial relationships and functional attributes.

[0080] The above describes the method for analyzing surveying and mapping geographic information based on machine vision in the embodiments of the present application. The following describes a system for analyzing surveying and mapping geographic information based on machine vision in the embodiments of the present application. Please refer to Figure 2 The system for analyzing surveying and mapping geographic information based on machine vision in the embodiments of the present application includes one embodiment as follows:

[0081] The preprocessing module 201 is configured to preprocess the multi-source remote sensing data to obtain a standardized surveying and mapping data set, and extract feature information of ground objects and textures from the standardized surveying and mapping data set to obtain multi-scale spatial features and semantic segmentation results.

[0082] The geographic element recognition module 202 is configured to integrate the multi-source data based on the multi-scale spatial features and semantic segmentation results to obtain a unified geographic element representation map, and recognize geographic elements from the unified geographic element representation map to obtain a set of surveying and mapping geographic elements.

[0083] a Bayesian inference module 203 configured to perform Bayesian inference of spatial uncertainty quantification and parameter uncertainty quantification on the surveyed geographic feature set to obtain a comprehensive uncertainty evaluation result;

[0084] a construction module 204 configured to construct vectorized geographic information based on the surveyed geographic feature set and the comprehensive uncertainty evaluation result.

[0085] Through the cooperation of the above components, through multi-source remote sensing data adaptive preprocessing and multi-scale feature extraction, the technical problems such as low recognition accuracy, lack of uncertainty evaluation and difficulty in information updating of traditional methods in complex scenes are effectively solved. The dynamic parameter adjustment denoising algorithm and the image enhancement technology based on deep learning introduced in the present application realize the effective preservation of the edge information of geographic features; the multi-scale feature extraction network with a five-layer pyramid structure combined with the dual-branch processing mechanism of direction sensitivity and regional aggregation significantly enhances the recognition ability of different types of geographic features; the feature transfer network of the "feature-space-semantic" three-layer transfer framework effectively solves the feature expression imbalance problem in multi-modal data fusion; the geographic feature perception residual block and the three-branch network architecture improve the recognition accuracy of geographic features; the Bayesian deep learning method combined with variational inference realizes the accurate quantification of the uncertainty of the recognition result; the geographic feature relationship reasoning based on graph neural network and the change detection technology of deep similarity measurement network make the geographic information expression more structured and realize automatic update.

[0086] The application further provides a computer readable storage medium, which can be a non-volatile computer readable storage medium or a volatile computer readable storage medium, and instructions are stored in the computer readable storage medium, and when the instructions are run on a computer, the computer is caused to execute the steps of the machine vision-based surveyed geographic information analysis method.

[0087] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, system and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be described herein.

[0088] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application or the entire or part of the technical solutions that essentially contribute to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a machine vision-based surveying geographic information analysis device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0089] The above-described embodiments are merely used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not cause the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A machine vision-based surveying geographic information analysis method, characterized by, The method comprises the following steps: Preprocessing multi-source remote sensing data to obtain a standardized surveying and mapping data set, and extracting ground object edge and texture information from the standardized surveying and mapping data set to obtain multi-scale spatial features and semantic segmentation results; Based on the multi-scale spatial features and the semantic segmentation results, multi-source data integration is performed to obtain a unified geographic feature representation map, and geographic feature recognition is performed on the unified geographic feature representation map to obtain a surveying and mapping geographic feature set; Performing Bayesian inference of spatial uncertainty quantization and parameter uncertainty quantization on the surveying and mapping geographic feature set to obtain comprehensive uncertainty evaluation results, including: calculating the probability distribution data of the geographic feature boundary position and the geographic feature attribute parameter based on the surveying and mapping geographic feature set; calculating the coefficient of variation of the geographic feature boundary position based on the probability distribution data, and performing adaptive kernel density estimation according to the coefficient of variation to obtain a spatial uncertainty quantization result; calculating the probability estimate of the geographic feature attribute parameter using a Bayesian neural network to obtain a parameter uncertainty quantization result, and constructing a Bayesian loss function based on the spatial uncertainty quantization result and the parameter uncertainty quantization result; based on the Bayesian loss function, modeling the spatial correlation of the surveying and mapping geographic feature set to obtain a spatial correlation enhanced uncertainty evaluation index; performing environmental factor sensitivity analysis on the spatial correlation enhanced uncertainty evaluation index to generate comprehensive uncertainty evaluation results including category probability distribution, position uncertainty and attribute parameter probability distribution of geographic features; Based on the surveying and mapping geographic feature set and the comprehensive uncertainty evaluation results, vectorized geographic information is constructed.

2. The machine vision-based surveying geographic information analysis method according to claim 1, wherein, The preprocessing of multi-source remote sensing data to obtain a standardized surveying and mapping data set, and the extraction of ground object edge and texture information from the standardized surveying and mapping data set to obtain multi-scale spatial features and semantic segmentation results comprises: Automatic sensor parameter identification of multi-source remote sensing data to obtain parameter calibration results of each data source, the multi-source remote sensing data comprising optical images, radar data and point cloud data; Color distribution histogram calculation of the optical images based on the parameter calibration results to obtain image characteristic analysis data; Based on the image characteristic analysis data, performing noise reduction processing on the optical images to obtain edge-preserving denoised images, and simultaneously performing outlier detection on the point cloud data to obtain a point cloud data set after removing outliers; Compensation processing of cloud coverage and shadow area on the edge-preserving denoised images to obtain enhanced processing images; Performing spatial resolution resampling and registration processing on the enhanced processing images, the radar data and the point cloud data set after removing outliers to obtain a standardized surveying and mapping data set; Inputting the standardized surveying and mapping data set into a multi-scale feature extractor to extract ground object edge and texture information to obtain multi-scale spatial features and semantic segmentation results. 3.The machine vision-based surveying and mapping geographic information analysis method according to claim 2, characterized in that, The inputting of the standardized surveying and mapping data set into a multi-scale feature extractor to extract ground object edge and texture information to obtain multi-scale spatial features and semantic segmentation results comprises: The standardized mapping data set is input into a multi-scale feature extractor for convolution operation to obtain a plurality of initial feature maps; The plurality of initial feature maps are respectively input into a direction-sensitive convolution unit and a region-aggregated convolution unit for processing to obtain linear linear feature and regional feature of linear feature and regional feature of linear feature and regional feature; The linear feature and the regional feature are input into a local response normalization layer for high-frequency information enhancement processing to obtain an enhanced edge and texture feature map; The enhanced edge and texture feature map is subjected to channel attention weighting to obtain a multi-scale spatial feature, and the enhanced edge and texture feature map is input into a fully convolutional neural network for semantic segmentation to obtain a semantic segmentation result. 4.The machine vision-based surveying and mapping geographic information analysis method according to claim 1, wherein, The multi-source data integration based on the multi-scale spatial feature and the semantic segmentation result obtains a unified geographic feature representation map, and the unified geographic feature representation map is subjected to geographic feature recognition to obtain a mapping geographic feature set, including: The multi-scale spatial feature and the semantic segmentation result are subjected to encoding processing to obtain an encoded feature representation; The encoded feature representation is separated into a core representation transmitted through an identity mapping and a variable representation adjusted through an adaptive residual connection, and the core representation and the variable representation are taken as a double-path feature flow; The double-path feature flow is subjected to cross-modal feature interaction to obtain a preliminary fusion feature; a spatial attention map and a channel attention map are calculated based on the preliminary fusion feature, and a feature attention weight matrix of the spatial attention map and the channel attention map is constructed; The preliminary fusion feature is adaptively re-weighted according to the feature attention weight matrix, and the semantic segmentation result is input into a local context aggregation unit as prior knowledge for feature guidance to obtain a context-enhanced feature; The context-enhanced feature is subjected to cross-scale residual fusion operation to obtain a unified geographic feature representation map; The unified geographic feature representation map is subjected to geographic feature recognition to obtain a mapping geographic feature set. 5.The machine vision-based surveying and mapping geographic information analysis method according to claim 4, characterized in that, The unified geographic feature representation map is subjected to geographic feature recognition to obtain a mapping geographic feature set, including: The unified geographic feature representation map is input into a geographic feature perception residual block of a backbone network for processing to obtain a multi-scale receptive field mapping feature; The multi-scale receptive field mapping feature is subjected to geographic feature contour feature extraction to obtain a boundary mask of a geographic feature; The multi-scale receptive field mapping feature is subjected to spatial distribution relationship analysis of a geographic feature to obtain a class probability map of the geographic feature; The multi-scale receptive field mapping feature is subjected to attribute information extraction of the geographic feature to obtain attribute parameter values of the geographic feature; The boundary mask of the geographic feature is subjected to boundary refinement and class label analysis to obtain accurate boundaries of the geographic feature and class information of the geographic feature; The accurate boundaries of the geographic feature, the class information of the geographic feature, the class probability map of the geographic feature and the attribute parameter values of the geographic feature are associated and integrated to obtain a mapping geographic feature set containing building contours, road networks and water system boundaries. 6.The machine vision-based surveying and mapping geographic information analysis method according to claim 1, wherein, constructing a point, line, surface, and composite four-layer element expression model based on the surveying and mapping geographic element set and the comprehensive uncertainty evaluation result, and outputting vectorized geographic information, including: creating a multi-level geographic element data structure including a point element layer, a line element layer, a surface element layer, and a composite element layer based on the surveying and mapping geographic element set, the point element layer including inflection points and intersection points, the line element layer including roads and rivers, the surface element layer including buildings and water bodies, and the composite element layer including residential areas and industrial areas; adjusting the threshold parameters of the elements in the line element layer based on the comprehensive uncertainty evaluation result to obtain simplified line element data; based on the multi-level geographic element data structure, taking geographic elements as nodes and spatial relationships between elements as edges, and assigning attribute feature vectors to each node and edge to obtain a geographic element relationship network; performing element relationship learning on the geographic element relationship network to obtain element groups and their functional attributes; comparing the surveying and mapping geographic element set with historical data based on the element groups and their functional attributes and the comprehensive uncertainty evaluation result to obtain geographic element change information; generating vectorized geographic information according to the geographic element change information, the simplified line element data, and the element groups and their functional attributes.

7. A machine vision-based surveying geographic information analysis system, characterized by, A machine vision-based surveying and mapping geographic information analysis system for implementing the machine vision-based surveying and mapping geographic information analysis method according to any one of claims 1 to 6, the machine vision-based surveying and mapping geographic information analysis system comprising: a preprocessing module configured to preprocess multi-source remote sensing data to obtain a standardized surveying and mapping data set, and extract ground object edges and texture information from the standardized surveying and mapping data set to obtain multi-scale spatial features and semantic segmentation results; a geographic element recognition module configured to integrate multi-source data based on the multi-scale spatial features and the semantic segmentation results to obtain a unified geographic element representation map, and recognize geographic elements from the unified geographic element representation map to obtain a surveying and mapping geographic element set; The Bayesian inference module is configured to perform Bayesian inference of spatial uncertainty quantification and parameter uncertainty quantification on the surveyed geographic feature set to obtain a comprehensive uncertainty evaluation result, including: calculating probability distribution data of geographic feature boundary positions and geographic feature attribute parameters based on the surveyed geographic feature set; calculating a coefficient of variation of the geographic feature boundary positions based on the probability distribution data, and performing adaptive kernel density estimation according to the coefficient of variation to obtain a spatial uncertainty quantification result; calculating a probability estimate of the geographic feature attribute parameters by using a Bayesian neural network to obtain a parameter uncertainty quantification result, and constructing a Bayesian loss function based on the spatial uncertainty quantification result and the parameter uncertainty quantification result; modeling spatial correlation based on the Bayesian loss function to obtain a spatial correlation enhanced uncertainty evaluation index; performing environmental factor sensitivity analysis on the spatial correlation enhanced uncertainty evaluation index to generate a comprehensive uncertainty evaluation result including a category probability distribution of geographic features, a position uncertainty, and an attribute parameter probability distribution. The construction module is configured to construct vectorized geographic information based on the surveyed geographic feature set and the comprehensive uncertainty evaluation result.

8. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, causes the processor to perform the machine vision based surveyed geographic information analysis method according to any one of claims 1 to 6.

Images