Geographic surveying and mapping method and system based on remote sensing of unmanned aerial vehicle

Abstract

The invention relates to the technical field of geographic surveying and mapping, and discloses a geographic surveying and mapping method and system based on unmanned aerial vehicle remote sensing, and the geographic surveying and mapping method based on unmanned aerial vehicle remote sensing comprises the following steps: processing source domain topographic data and target domain scarce samples through a feature selective migration network, and obtaining a feature adaptability model parameter set; processing a low-resolution remote sensing image through a physical constraint generative adversarial network to obtain a high-resolution geographic feature enhanced image data set; processing the high-resolution geographic feature enhanced image data set and the feature adaptability model parameter set through a hierarchical attention network to obtain a fine-grained ground feature classification result matrix; through a small sample learning strategy of feature selective migration, a super-resolution reconstruction technology of physical constraint, a fine-grained classification method of hierarchical attention and a standardized geographic data product generation process, the dependency of surveying and mapping data acquisition on the environment is reduced, the surveying and mapping efficiency is improved, and the surveying and mapping precision is improved.

Description

Technical Field

[0001] This invention relates to the field of geographic surveying and mapping technology, and more specifically, to a geographic surveying and mapping method and system based on unmanned aerial vehicle (UAV) remote sensing. Background Technology

[0002] With the rapid development of remote sensing technology, UAV remote sensing has become an important tool in the field of geographic surveying. Traditional geographic surveying methods mainly rely on satellite remote sensing and manual field measurement, but these methods have obvious limitations in the application of special geographic areas (such as polar regions, high mountains, deserts, volcanoes, etc.).

[0003] Existing surveying methods typically require a large number of labeled samples to ensure model performance. When samples are scarce, accuracy drops significantly. Furthermore, while traditional super-resolution methods can improve image resolution, they often ignore the physical characteristics of geospatial data, resulting in a lack of geometric consistency and geographic accuracy in the reconstruction results. At the same time, existing classification methods have low accuracy when dealing with land features that are similar in shape but different in category, and lack the ability to selectively focus on key areas.

[0004] Therefore, existing surveying methods in special environments have difficulty obtaining labeled samples, making it difficult to meet the needs of deep learning models for large amounts of labeled data. Furthermore, the resolution of the raw remote sensing images obtained is limited by the flight conditions, payload capacity, and imaging equipment performance of UAVs. At the same time, the types of land features in special geographical areas are complex and have significant subtle differences, resulting in low fine-grained recognition capabilities, which reduces surveying efficiency and accuracy. Summary of the Invention

[0005] This invention provides a geographic mapping method and system based on UAV remote sensing, which solves the problems of low efficiency and accuracy of existing mapping methods in special environments.

[0006] The first aspect of this invention provides a geographic mapping method based on unmanned aerial vehicle (UAV) remote sensing, comprising the following steps: The source domain topographic data and target domain scarce samples are processed by a feature selective transfer network to obtain a feature adaptation model parameter set; low-resolution remote sensing images are processed by a physical constraint generative adversarial network to obtain a high-resolution geographic feature enhanced image dataset; the high-resolution geographic feature enhanced image dataset and the feature adaptation model parameter set are processed by a hierarchical attention network to obtain a fine-grained land cover classification result matrix; and the fine-grained land cover classification result matrix and the high-resolution geographic feature enhanced image dataset are processed by a mapping data product generator to obtain a fine mapping geographic data result set.

[0007] As a further optimization of the present invention, the step of processing source domain terrain data and target domain scarce samples through a feature-selective transfer network to obtain a feature-adaptive model parameter set includes: The source domain remote sensing image data is processed by a multi-scale feature extraction module to obtain a hierarchical terrain feature representation: ; in, Represents source domain remote sensing imagery. This represents a hierarchical representation of terrain features. This represents the multi-scale feature extraction module; The hierarchical terrain feature representation and scarce samples from the target domain are processed by the domain-invariant feature learning module to obtain the inter-domain mapping matrix: ; in, Represents the features of the target domain. Represents the mapping relationship matrix between domains. Representation domain invariant feature learning module; The feature-selective attention module processes the inter-domain mapping matrix and target domain features to obtain the feature-adaptive model parameter set. ; in, For the feature-adaptive model parameter set, This is a feature-selective attention module.

[0008] As a further optimization of the present invention, the calculation process of the multi-scale feature extraction module is as follows: ; ; ; in, This indicates that convolutional layers are extracted using initial features. Remote sensing images of the source region The initial feature map obtained after processing. This represents the initial feature extraction convolutional layer. Indicates the first The feature map of a layer is derived from the feature map of the previous layer. Add to This is obtained by performing convolution operations with different kernel sizes and then weighting the sum. Indicates the number of feature levels. Represents a set of different convolution kernel sizes. These represent the weight coefficients of the corresponding convolution kernel. Indicates the kernel size as Convolution operation, This indicates a feature cascade operation. This is the final hierarchical representation of terrain features. This represents the feature maps extracted by the multi-scale feature extraction module at different levels.

[0009] As a further optimization of the present invention, the step of processing low-resolution remote sensing images through physical constraint generative adversarial networks to obtain a high-resolution geographic feature-enhanced image dataset includes: The original low-resolution image is processed by a degradation modeling analyzer to obtain an image degradation parameter vector: ; in, This indicates low-resolution remote sensing imagery. This represents the image degradation parameter vector. This represents a degradation modeling analyzer; The initial super-resolution image tensor is obtained by processing the low-resolution image and the image degradation parameter vector through an adaptive reconstruction generator: ; in, This represents the initial super-resolution image tensor. This represents an adaptive reconstruction generator; The initial super-resolution image tensor is processed by a physical constraint discriminator to obtain a physical consistency score matrix: ; in, Represents the physical consistency scoring matrix. Represents a physical constraint discriminator; The initial super-resolution image tensor and the physical consistency score matrix are processed by a multi-objective optimization controller to obtain a high-resolution geographic feature enhancement image dataset: ; in, This represents a high-resolution geographic feature enhancement image dataset. Indicates reference high-resolution imagery, This represents a multi-objective optimization controller.

[0010] As a further optimization of the present invention, the optimization process of the multi-objective optimization controller is based on the following loss function: ; in, Represents the total loss function. Indicating resistance to loss, Indicates the losses incurred during reconstruction. Indicates perceived loss. Represents physical constraint loss. , and It is the weighting coefficient that balances the various loss terms.

[0011] As a further optimization of the present invention, the step of processing the high-resolution geographic feature enhancement image dataset and the feature adaptation model parameter set through a hierarchical attention network to obtain a fine-grained land cover classification result matrix includes: By processing the parameter set of the feature-adaptive model and the pre-trained deep network using a feature-adaptive regulator, a domain-adaptive feature extraction network is obtained. ; in, Representation domain adaptive feature extraction network parameters, This represents the parameters of the pre-trained deep network. Indicates a characteristic adaptive regulator; The high-resolution geographic feature enhancement image dataset is processed by a multi-scale feature extractor to obtain a multi-scale feature tensor set: ; in, Represents a multi-scale feature tensor set. This represents a multi-scale feature extractor. Represents a single image in a high-resolution geographic feature-enhanced imagery dataset; The multi-scale feature tensor set is processed by a hierarchical spatial attention module to obtain the attention-enhanced feature tensor: ; in, This represents the attention-enhanced feature tensor. This represents a hierarchical spatial attention module; The attention-enhanced feature tensor is processed by a multi-level semantic classifier to obtain a fine-grained land cover classification result matrix: ; in, This represents the matrix of fine-grained land cover classification results. This represents a multi-level semantic classifier.

[0012] As a further optimization of the present invention, the calculation expression of the hierarchical spatial attention module is: ; ; in, , , These represent the query, key, and value matrices, respectively. , , This represents the corresponding weight matrix. This represents the transpose of the key matrix. The dimension of the key vector. This represents the activation function. This indicates self-attention output.

[0013] As a further optimization of the present invention, the step of processing the fine-grained land cover classification result matrix and the high-resolution geographic feature enhancement image dataset through the mapping data product generator to obtain a fine mapping geographic data result set includes: processing the fine-grained land cover classification result matrix through a vectorization processing module to obtain a geographic feature geometric representation dataset. ; in, This represents a dataset that represents the geometric representation of geographic features. This indicates the vectorization processing module. This represents the matrix of fine-grained land cover classification results; The attribute information extraction module processes the high-resolution geographic feature enhancement image dataset and the geographic feature geometric representation dataset to obtain a geographic feature attribute information table: ; in, A table representing geographic element attribute information. This indicates the attribute information extraction module. This represents a high-resolution geographic feature-enhanced image dataset; The geographic feature geometric representation dataset and the geographic feature attribute information table are processed by the geographic data standardization module to obtain a set of detailed surveying and mapping geographic data results. ; in, This represents a set of detailed surveying and mapping geographic data. This indicates the geographic data standardization module.

[0014] As a further optimization of the present invention, the geometric construction calculation expression of the vectorization processing module is as follows: ; in, Indicates the first Geometric representation of a geographic element Represents the constructor function for point features. Indicates the constructor function for line features. Represents the constructor of face features. Represents the coordinates of a point. Represents a sequence of contour points. Represents the area of ​​the outline. Indicates the outline width. and These are the threshold parameters for point and line features, respectively.

[0015] A second aspect of this invention provides a geographic mapping system based on UAV remote sensing, used to execute the aforementioned geographic mapping method based on UAV remote sensing, comprising: a feature-selective transfer network for processing source domain terrain data and target domain scarce samples to obtain a feature adaptation model parameter set; a physical constraint generative adversarial network for processing low-resolution remote sensing images to obtain a high-resolution geographic feature enhancement image dataset; a hierarchical attention network for processing the high-resolution geographic feature enhancement image dataset and the feature adaptation model parameter set to obtain a fine-grained land cover classification result matrix; and a mapping data product generator for processing the fine-grained land cover classification result matrix and the high-resolution geographic feature enhancement image dataset to obtain a refined mapping geographic data result set.

[0016] The beneficial effects of this invention are as follows: This invention achieves a reduction in the dependence of surveying and mapping data acquisition on the environment, an increase in surveying and mapping efficiency, and an improvement in surveying and mapping accuracy through a feature-selective transfer few-sample learning strategy, a physically constrained super-resolution reconstruction technology, a hierarchical attention fine-grained classification method, and a standardized geographic data product generation process. Attached Figure Description

[0017] Figure 1 This is a flowchart of a geographic mapping method based on UAV remote sensing according to the present invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a geographic mapping method based on UAV remote sensing, which adopts a "knowledge transfer-super-resolution reconstruction-refined recognition" technical route, integrating transfer learning, physically constrained super-resolution reconstruction, and multi-level attention mechanisms to achieve high-precision geographic feature extraction and mapping under low-resource conditions. Figure 1 As shown, it includes the following steps: Step 100: Process source domain terrain data and scarce target domain samples using a Feature Selective Transfer Network (FSTN) to obtain the parameter set for the adaptive feature model. The FSTN transfers common terrain knowledge to specific terrain types. Through three modules—multi-scale feature extraction, domain-invariant feature learning, and feature-selective attention—it addresses the problem of scarce target domain data and generates the parameter set for the adaptive model. The specific steps are as follows: Step 101: Process source domain remote sensing image data using the multi-scale feature extraction module to obtain hierarchical terrain feature representations: This sub-step utilizes the Multi-Scale Feature Extraction Module (MSFEM) to extract features from remote sensing images of common terrain areas in the source domain through cascaded multi-scale convolutional units (containing three different sizes of convolutional kernels and residual connections). The input to MSFEM is the source domain remote sensing image. The output is a hierarchical representation of terrain features. ;in, , and These represent the image's height, width, and number of channels, respectively. Represents the set of real numbers. Indicates the number of feature levels. For feature dimension, It is the set of real numbers.

[0020] The mathematical expression for MSFEM is: The calculation process is as follows: ; ; ; in, This indicates that convolutional layers are extracted using initial features. Remote sensing images of the source region The initial feature map obtained after processing. This represents the initial feature extraction convolutional layer. Indicates the first Feature map of the layer This is the feature map of the previous layer. Indicates the number of feature levels. Represents a set of different convolution kernel sizes. These represent the weight coefficients of the corresponding convolution kernel. Indicates the kernel size as Convolution operation, This indicates a feature cascade operation. This is the final hierarchical representation of terrain features. This represents the feature maps extracted by the multi-scale feature extraction module at different levels.

[0021] Step 102: Process source domain features and target domain scarce samples using the Domain-Invariant Feature Learning Module (DIFLM) to obtain the inter-domain mapping matrix. The DIFLM uses a feature encoder and domain discriminator to perform adversarial learning on source domain features and target domain scarce samples, establishing cross-domain feature mapping relationships and achieving feature alignment. The input to the DIFLM includes the source domain feature representation. and scarce samples in the target domain The output is a matrix of mapping relationships between different domains. This matrix represents the transformation relationship from the source domain feature space to the target domain feature space, where the two dimensions... Both are related to the feature dimensions in the input source domain feature representation. Correspondingly, , and These represent the height, width, and number of channels of the target domain sample, respectively. It represents the set of real numbers. The mathematical expression for DIFLM is: ; ; in, Through feature encoder ( ) for target domain samples The target domain feature representation obtained after encoding. Through the domain-invariant feature learning module ( ) Representation of source domain features and target domain feature representation The result after processing; DIFLM is trained using the following objective function: ; in, The total loss function of the domain-invariant feature learning module. This represents the loss function of the main task. Indicating resistance to loss, This represents the feature alignment loss. and These are the weighting coefficients that balance the various loss terms. The feature alignment loss is defined as: ; in, For feature alignment loss, and Let these represent the feature vectors of the source domain and the target domain, respectively. and These represent the corresponding tags. This is a label similarity matrix, where 1 represents a match of the same label and 0 represents the match of the same label. and These represent the number of samples in the source and target domains, respectively. The square of the Euclidean distance. Indicates the index of the source domain sample, with values ​​ranging from 1 to... , This represents the index of the target domain sample, with values ​​ranging from 1 to... .

[0022] Inter-domain mapping matrix ( The following is calculated using features from the source and target domains: ; in, This represents the set of sample pairs with the same semantic labels in the source and target domains. Indicates the number of sample pairs. Indicates matrix transpose. This indicates an outer product operation that generates a mapping matrix.

[0023] Step 103: Process the mapping matrix and target domain features through the Feature Selective Attention Module to obtain the parameter set of the adaptive feature model: The Feature Selective Attention Module (FSAM) automatically evaluates the applicability of source domain features in the target domain and generates the parameter set of the adaptive model through an attention generation network and a feature modulation network. The input of FSAM includes the inter-domain mapping matrix. and target domain features The output is the feature-adaptive model parameter set. It includes modulation parameters and network weights.

[0024] The mathematical expression for FSAM is: ; The attention generation network calculates the feature importance weights: ; in, This represents the feature importance weights obtained through specific calculations. It is an activation function. Representing a multilayer perceptron, it is a type of feedforward artificial neural network. Indicates a connection operation. This represents the source domain features after mapping; Feature modulation networks generate adaptive features: ; in, This represents the generated adaptive features. This represents the feature importance weights obtained through specific calculations. This indicates an element-wise multiplication operation. This represents the weighted mapping source domain features. Indicates the weighted target domain features; Ultimately, the parameter set of the feature-adaptive model consists of the following: ; in, For the adjusted network weights, and These are the scaling and offset parameters used for feature normalization: ; ; ; in, Represents the weights of the source domain pre-trained model. This represents the weight update function. and These represent the fully connected layers that generate scaling and offset parameters, respectively.

[0025] Through steps 101 to 103, the Feature Selective Transfer Network generates a set of model parameters suitable for the specific terrain of the target domain, effectively handling scarce samples and laying the foundation for subsequent tasks.

[0026] Step 200: Process low-resolution remote sensing images using a Physically Constrained Generative Adversarial Network (PCGAN) to obtain a high-resolution geographic feature-enhanced image dataset: This step utilizes a PCGAN to perform super-resolution reconstruction of low-resolution remote sensing images through four key components (degradation modeling analyzer, adaptive reconstruction generator, physical constraint discriminator, and multi-objective optimization controller), generating a high-resolution geographic feature-enhanced image dataset. The specific steps are as follows: Step 201: Process the original low-resolution image using the Degradation Modeling Analyzer (DMA) to obtain the image degradation parameter vector: The DMA performs degradation characteristic analysis on the low-resolution image through a shared feature extraction backbone network and three dedicated task head networks, estimating various degradation parameters during the imaging process. The input to the DMA is the low-resolution remote sensing image. ,in , and These represent the image's height, width, and number of channels, respectively; the output is a vector of image degradation parameters. It includes information such as fuzzy kernel parameters, downsampling rate, and noise intensity. This represents the dimension of the parameter vector. The mathematical expression for DMA is: The specific calculation process includes: Feature extraction: ; in, This represents the extracted low-resolution image features. A deep convolutional network composed of residual blocks Fuzzy kernel estimation: ; in, This represents the estimated fuzzy kernel parameters. It is a task head network used for estimating fuzzy kernels. Downsampling rate estimation: ; in, This represents the estimated downsampling rate parameter. It is a task head network used to estimate the downsampling rate. Noise characteristic estimation: , in, Estimated noise intensity and type parameters, It is a task head network used to estimate noise characteristics. Parameter aggregation: The parameters are combined into a degenerate parameter vector. in, This represents the final degenerate parameter vector. This indicates a join operation.

[0027] Feature extraction backbone network It is composed of multiple stacked residual blocks, and its expression is: ; ; ; in, Represents the initial feature map. Indicates the initial convolutional layer. Indicates the first Output feature maps of each residual block and Indicates a convolutional layer. This represents the activation function. Indicates the index of the residual block. Indicates the number of residual blocks. Indicates the first Output feature maps of each residual block This represents the output feature map of the nth residual block.

[0028] Step 202: Process the low-resolution image and image degradation parameter vector using the Adaptive Reconstruction Generator to obtain the initial super-resolution image tensor: The Adaptive Reconstruction Generator (ARG) dynamically adjusts its internal structure (including a feature extraction module, multiple adaptive convolutional modules, and an image reconstruction module) to generate the initial super-resolution image based on the degradation parameters. The input to the ARG includes low-resolution remote sensing images. and image degradation parameter vector The output is the initial super-resolution image tensor. ,in, , indicating the height of the output image It is the height of the input image. of times, , indicating the width of the output image It is the input image width of times, Indicates the magnification factor.

[0029] The mathematical expression for ARG is: The calculation process is as follows: Feature extraction: ; in, This represents the initial feature map extracted. Indicates from low-resolution remote sensing imagery Functions for extracting features; Adaptive feature processing: ; in, This represents the feature map after adaptive processing. Based on image degradation parameter vector For input features A function that performs adaptive processing; Upsampling: ; in, This represents the feature map after upsampling. It is used for features after convolution. A function that performs upsampling operations; Image reconstruction: ; in, It is an activation function used here to upsample the features after convolution. Processing is performed to reconstruct the initial super-resolution image tensor. ; The adaptive feature processing is implemented based on multiple adaptive convolutional modules (ACMs): ; This formula represents the input feature map. and image degradation parameter vector Passing through multiple adaptive convolutional modules in sequence arrive The process ultimately yields adaptively processed features. .

[0030] The calculation expression for each adaptive convolutional module is: ; in, This represents the output features of the adaptive convolution module. Indicates the input feature map, Indicates the first A predefined convolution operation, The weighting coefficients are generated based on the degradation parameters: ; in, Indicates the first The weight coefficients of each convolution operation, This represents a multilayer perceptron. The function is used to convert the input numerical value into a probability distribution such that all output values ​​are between 0 and 1 and sum to 1.

[0031] Step 203: Process the initial super-resolution image tensor using the Physics-Constrained Discriminator (PCD) to obtain the physical consistency score matrix. The PCD performs a comprehensive evaluation of the authenticity and geophysical consistency of the generated image using a shared encoder and a dual-branch network (authenticity discrimination branch and physical consistency evaluation branch). The input to the PCD is the initial super-resolution image tensor. The output is a physical consistency score matrix. In this matrix, the first row of values ​​represents the authenticity score, and the second row of values ​​represents the physical consistency score. This indicates the number of feature regions being evaluated.

[0032] The mathematical expression for PCD is: The calculation process is as follows: Feature encoding: ; Authenticity rating: ; Physical consistency score: ; Rating compilation: ; in, This represents the features extracted by the shared encoder. It is a shared encoder in the Physical Constraint Discriminator (PCD). Indicates the authenticity rating. It is the network of the authenticity discrimination branch in PCD. Indicates the physical consistency score. It is the network of the physical consistency assessment branch in PCD. This represents the final physical consistency score matrix; The physical consistency score consists of three geographical constraints: ; in, The edge preservation score is calculated using the following formula: ; The structure preservation score is calculated using the following formula: ; The spectral consistency score is calculated using the following formula: ; in, For activation function, , and It balances the weighting coefficients of each score, which are automatically optimized through network training. To evaluate the convolutional layer at the edge, For gradient extraction function, For structural evaluation of convolutional layers, For structural feature extraction function, To evaluate the convolutional layer using spectral analysis, This is the spectral feature extraction function.

[0033] Step 204: Process the initial super-resolution image tensor and physical consistency score matrix using a multi-objective optimization controller to obtain a high-resolution geographic feature-enhanced image dataset: The Multi-Objective Optimization Controller (MOOC) generates a high-resolution image dataset with enhanced geographic features by comprehensively calculating multiple loss functions and optimizing generator parameters. The input to the MOOC includes the initial super-resolution image tensor. Physical consistency scoring matrix Original low-resolution image and reference high-resolution image (If a reference high-resolution image is available, it is used for supervised learning; otherwise, unsupervised learning is used, relying solely on low-resolution imagery and physical constraints for optimization.) The output is a high-resolution geographic feature enhancement image dataset. ,in, Indicates the number of images processed. , , These represent 1, 2, and M high-resolution geographic feature enhancement images obtained after processing by a multi-objective optimization controller.

[0034] The mathematical expression for MOOC is: ; in, This indicates that the multi-objective optimization controller ( The resulting high-resolution geographic feature-enhanced image dataset after processing; Its optimization process is based on the following loss function: ; in, Represents the total loss function. , and These are the weighting coefficients that balance the various loss terms, each defined as follows: Combating loss ( The adversarial loss between the generator and the discriminator is represented by the formula: ; in, It is the score given by the discriminator for the realism of the generated image. The mathematical expectation is represented by the loss, which prompts the generator to produce realistic images that can fool the discriminator. Reconstruction losses ( ): This measures the pixel-level difference between the generated image and the target image, and its formula is: When a reference high-resolution image is available: ; When there is no reference image: ; in, Describing the L1 norm, Indicates a downsampling operation; Perceived loss ( ): This measures the difference between the generated image and the target image in the high-level feature space, and its formula is: ; in, This represents the feature extractor of a pre-trained VGG network. This represents the square of the L2 norm, and this loss helps generate results that are visually closer to the real image. Physical constraint loss ( ): To ensure the generated super-resolution imagery conforms to geophysical characteristics, it consists of three sub-losses: ; The components of the physical constraint loss are defined as follows: edge preservation loss To ensure the accuracy of the generated image edge features, the formula is as follows: When a reference image is available: ; When there is no reference image: ; in, Describing the L1 norm, Represents the gradient operator, Represents total variation; Structural retention loss ( To ensure that the structural information of the generated image is consistent with that of the reference image, the formula is: When reference images are available: ; No reference image: ; in, Indicators representing structural similarity This represents a function that measures structural regularity. Spectral uniformity loss To ensure that the spectral characteristics of the generated image are consistent with those of the reference image, the calculation formula is as follows: When reference images are available: ; When there is no reference image: ; in, Describing the L2 norm, This represents the histogram feature extraction function. This represents the spectral constraint function.

[0035] Through steps 201 to 204, the Physically Constrained Generative Adversarial Network (GAN) generates a high-quality, high-resolution geographic feature-enhanced image dataset. These images not only possess high visual quality but also maintain the geometric accuracy and spatial relationships of geographic features, laying the foundation for subsequent fine-grained classification.

[0036] Step 300: Process the high-resolution geographic feature enhancement image dataset and feature adaptation model parameter set using a hierarchical attention network to obtain a fine-grained land cover classification result matrix: The hierarchical attention network (HAN) performs fine-grained land cover classification on the high-resolution geographic feature enhancement image. Through four main components—a multi-scale feature extractor, a feature adaptation regulator, a hierarchical spatial attention module, and a multi-level semantic classifier—it achieves the adjustment of feature adaptation parameters to highlight key regional features, ultimately outputting an accurate land cover classification result matrix. The specific steps are as follows: Step 301: Process the feature adaptation model parameter set and the pre-trained deep network using the feature adaptation regulator to obtain the domain-adaptive feature extraction network: The Feature Adaptivity Regulator (FAR) adjusts the parameters of the pre-trained deep network using the feature adaptation model parameter set from Step 100, constructing a domain-adaptive feature extraction network capable of efficiently processing high-resolution images in the target domain. The input to FAR includes the feature adaptation model parameter set. (Output of step 100) and pre-trained deep network parameters The output is the parameters of the domain-adaptive feature extraction network. .

[0037] The mathematical expression for FAR is: ; The calculation process is as follows: Parameter mapping: ; in, This represents the mapped parameters. It is a parameter mapping network that maps feature adaptation parameters to a parameter space compatible with pre-trained networks; Parameter fusion: , in, Represents element-wise multiplication. It is a coefficient that adjusts the fusion strength; The computational expression for the parameter mapping network is: ; in, and Indicates a fully connected layer. This represents the activation function.

[0038] Step 302: Process the high-resolution geographic feature enhancement image dataset using a multi-scale feature extractor to obtain a multi-scale feature tensor set: In this step, the multi-scale feature extractor (MFE) processes the high-resolution geographic feature enhancement image dataset using domain-adaptive feature extraction network parameters to extract multi-scale feature representations. The MFE employs a feature pyramid network structure, which can capture geographic features at different spatial scales.

[0039] The input to MFE is a single image from a high-resolution geospatial enhancement image dataset. (Output of step 200) and parameters of the domain-adaptive feature extraction network The output is a multi-scale feature tensor set. ,in, , , express , , Hierarchical feature maps Indicates the number of feature levels.

[0040] The mathematical expression for MFE is: The calculation process is as follows: Basic feature extraction: ; Feature pyramid construction: ; in, It is a backbone network that uses domain adaptability parameters. This represents the extracted basic feature tensor. The feature pyramid is constructed using a top-down and lateral connection approach. Feature maps representing different scale levels; The computational expression for the feature pyramid network is: ; ; ; in, This is the top-level feature map of the feature pyramid. The first feature pyramid Layer feature map, Indicates the final number Layer feature map, The first feature pyramid Layer feature map, These are the feature maps of the deepest layer of the backbone network. Main backbone network Feature map of the layer and for and Convolution operation, Indicates an upsampling operation. This indicates the number of layers in the feature pyramid. Indicates the first The final feature map output of the layer, This represents the hierarchical index of the feature pyramid.

[0041] Step 303: Process the multi-scale feature tensor set through the Hierarchical Spatial Attention Module to obtain attention-enhanced feature tensors: In this step, the Hierarchical Spatial Attention Module (HSAM) analyzes the multi-scale feature tensor set, generates attention maps at different scales, highlights key areas in geographic features, and integrates multi-scale features.

[0042] The input to HSAM is a multi-scale feature tensor set. ,in, , , They represent , , The layer's feature tensor outputs an attention-enhanced feature tensor. ,in, , , These represent the height, width, and number of channels of the output feature tensor, respectively.

[0043] The mathematical expression for HSAM is: The calculation process is as follows: Intra-scale attention: ; Inter-scale attention: ; in, Indicates the first Attention graph of the layer This indicates an attention-weighted operation. It is a self-attention mechanism. Indicates the attention-weighted number of... Layer feature map, Indicates the first Feature map of the layer Index representing the scale level. Indicates the total number of feature levels; Hierarchical integration: ; in, This represents the attention-enhanced feature tensor after fusion. This represents a hierarchical feature fusion function. This represents the feature tensor after attention weighting at each scale level. This indicates the total number of feature levels.

[0044] The computational expression for the self-attention mechanism is as follows: ; ; in, , , These represent the query, key, and value matrices, respectively. , , This represents the corresponding weight matrix. This represents the transpose of the key matrix. The dimension of the key vector. This represents the activation function. Indicates self-attention output; The calculation expression for hierarchical fusion is: ; ; ; in, Represents a temporary feature map. This represents the highest-level feature tensor after attention weighting. This indicates a channel-level join operation. This represents the convolution operation. Indicates an upsampling operation. For the first The feature tensor after attention weighting at each level, Indicates from arrive Hierarchical index.

[0045] Step 304: Process the attention-enhanced feature tensor through a multi-level semantic classifier to obtain a fine-grained land cover classification result matrix: In this step, the multi-level semantic classifier (MSC) processes the attention-enhanced feature tensor to generate hierarchical land cover classification results. The MSC adopts a multi-level classification structure, refining the classification results step by step from coarse-grained categories to fine-grained categories.

[0046] The input to MSC is the attention-enhanced feature tensor. The output is a fine-grained land cover classification result matrix. ,in, This indicates the number of land cover categories at the finest level.

[0047] The mathematical expression for MSC is: The calculation process is as follows: Coarse-grained classification: ; Medium particle size classification: ; Fine-grained classification: ; Results integration: ; in, This indicates the coarse-grained classification result. This indicates the results of medium-grained classification. This indicates the results of fine-grained classification. This represents a coarse-grained classifier. This represents a medium-granularity classifier. This represents a fine-grained classifier. This indicates the result integration function. This represents the final fine-grained land cover classification result matrix. The matrix contains the probability distribution of each pixel belonging to each fine-grained land cover category, with dimensions of [dimensionality missing]. ,in, and These represent the height and width of the image, respectively. This represents the total number of fine-grained land cover categories.

[0048] The calculation expression for each classifier is: ; in, Represents the classifier function. Represents a convolutional neural network. Indicates global average pooling. Indicates a fully connected layer. This represents the activation function. This represents the input feature tensor.

[0049] The calculation expression for the result integration is: ; in, , , This is the weight matrix for the corresponding level. This is a weighted operation.

[0050] Through steps 301 to 304, the hierarchical attention network generates a high-precision, fine-grained land cover classification matrix. This matrix fully utilizes the rich details in high-resolution imagery and the domain transfer capabilities provided by feature adaptive parameters, achieving fine classification of complex geographic features and providing a foundation for subsequent geographic mapping data generation.

[0051] Step 400: Process the fine-grained land cover classification result matrix and high-resolution geographic feature enhancement image dataset using the Mapping Data Product Generator to obtain a refined mapping geographic data result set: This step utilizes the Mapping Data Product Generator (MDPG) to transform the results of the previous steps into mapping data products conforming to Geographic Information System (GIS) standards. MDPG comprises three key components: vectorization processing, attribute information extraction, and geographic data standardization. Its workflow is as follows: the vectorization processing module first converts the classification result matrix into a geometric representation of geographic features; then, the attribute information extraction module extracts attribute information from the high-resolution imagery; finally, the geographic data standardization module integrates the information from both to generate a refined mapping geographic data result set conforming to industry standards. The specific steps are as follows: Step 401: Process the fine-grained land cover classification result matrix through the vectorization processing module to obtain a geographic feature geometric representation dataset: In this step, the vectorization processing module (VPM) converts the raster-format fine-grained land cover classification result matrix into a vector-format geographic feature geometric representation dataset. VPM employs deep learning-based boundary extraction and topology preservation algorithms to ensure that the generated vector data has high geometric accuracy.

[0052] The input to VPM is a fine-grained land cover classification result matrix. (Output of step 300), where, , The height and width of the feature map, The output is a dataset of geometric representations of geographic features, representing the number of land cover categories. ;in, , , The first , , A geometric representation of a geographic feature, where each geometric representation can be a vector form such as a point, line, or polygon. It represents the total number of geographical elements.

[0053] The mathematical expression for VPM is: The calculation process is as follows: Category probability mapping: ; Boundary extraction: ; Contour tracking: ; Vector simplification: ; Geometric representation generation: ; in, This represents the category label for each pixel. Indicates the boundary extraction operator, Represents the boundary binary graph. This indicates that the matrix elements take values ​​of 0 or 1, and the dimension is... 0 represents a non-boundary pixel, and 1 represents a boundary pixel. This represents the set of extracted contours. , , They represent , , A collection of outlines Indicates the first ( A sequence of points for a contour. This represents the contour tracing operator. Represents the simplified set of contours. This represents a simplified threshold. This represents the simplified Douglas-Peucker algorithm. This represents the final generated geometric representation dataset. Represents the geometry construction operator; The boundary extractor uses a deep learning model, and its expression is: ; in, It is a deep convolutional network. It is a threshold parameter.

[0054] The calculation expression for the geometry construction is: ; in, Indicates the first Geometric representation of a geographic element Represents the constructor function for point features. Indicates the constructor function for line features. Represents the constructor of face features. Represents the coordinates of a point. Represents a sequence of contour points. Represents the area of ​​the outline. Indicates the outline width. and These are the threshold parameters for point and line features, respectively.

[0055] Step 402: Process the high-resolution geographic feature enhancement image dataset and geographic feature geometric representation dataset through the attribute information extraction module to obtain the geographic feature attribute information table: In step 402, the Attribute Information Extraction Module (AIEM) extracts attribute information of geographic features, such as height, area, and texture features, from the high-resolution geographic feature enhancement image dataset.

[0056] AIEM uses region-based feature analysis and statistical calculation methods to generate attribute information for each geographic feature.

[0057] AIEM's input includes high-resolution geospatial augmentation image datasets. (Output of step 200) and geographic feature geometric representation dataset (Output of step 401) The output is a geographic feature attribute information table. ,in, , , They represent the first , , A vector of attribute information for a geographic element. It represents the total number of geographical elements.

[0058] The mathematical expression for AIEM is: The calculation process is as follows: Image cropping: ; Feature extraction: ; Attribute calculation: ; in, Indicates the first Image patches of the region where each geographic feature is located It's a cropping operation. Indicates the first Feature vectors of geographical elements It is a feature extractor. It's an attribute calculator; The feature extractor contains multiple feature calculation modules, whose expressions are as follows: ; in, The feature join operation is represented, and the feature calculation modules are defined as follows: Spectral characteristics: ; in, For image blocks The average value of each band reflects the overall spectral reflectance characteristics of the ground object. For image blocks The standard deviation of each band reflects the internal variability of the ground object's spectrum. Represents image block The minimum value for each band reflects the lowest reflectance of the ground object. Represents image block The maximum value of each band reflects the highest reflectivity of the ground object. Other spectral statistical characteristics, such as median, skewness, and kurtosis, are also represented. Texture features: ; in, This represents the GrayLevel Co-occurrence Matrix feature extraction function, which calculates the feature of an image patch. Texture features, including homogeneity, contrast, entropy, correlation, etc.; Morphological characteristics: ; in, Representing geographical elements The area (in square meters). Representing geographical elements The perimeter (in meters). Representing geographical elements Compactness, Other morphological features, such as aspect ratio, directionality, and convexity ratio, are also indicated. Contextual features: ; in, This represents a contextual analysis function that analyzes the relationship between geographic features and their surrounding environment. The function considers geographic features... Information such as the distribution of surrounding land features, their proximity relationships, and spatial patterns; The attribute calculator calculates various attributes based on feature vectors and geographic feature categories: ; in, Indicates the first A vector of attribute information for a geographic feature, containing multiple attribute values. Indicates the first The height attribute value of each geographic feature, in meters. Indicates the first The area attribute value of each geographic element, in square meters. Indicates the first The volume attribute value of a geographic feature, in cubic meters. Indicates the first The material attribute value of a geographic feature is usually a classification code. Indicates the first The functional attribute values ​​of a geographic element are usually classified codes. It can represent other attribute values, such as age, status, ownership, etc. The calculation methods for different attributes vary, for example: Height attribute: ; Area attribute: ; Volume attribute: ; in, This represents the height estimation function. This represents the area calculation function. This represents a volume estimation function; Step 403: Process the geographic feature geometric representation dataset and geographic feature attribute information table through the Geographic Data Standardization Module to obtain a detailed mapping geographic data output set: The Geographic Data Standardization Module (GDSM) integrates the geometric representation and attribute information of geographic features into a data product that conforms to the Geographic Information System (GIS) standard. GDSM handles tasks such as coordinate system transformation, topological relationship checking and repair, and format conversion to ensure that the generated data product can be directly used in a GIS system. The input to GDSM includes the geographic feature geometric representation dataset. (Output of step 401) and geographic feature attribute information table (Output of step 402) The output is a set of detailed surveying and mapping geographic data. ,in, , , They represent the first , , Types of geographic data products, This represents the total number of types of geographic data products.

[0059] The mathematical expression for GDSM is: The calculation process is as follows: Coordinate transformation: ; Transform the geometric representation of geographic features into the target coordinate system; where, Represents the coordinate transformation function. This represents the transformed geometric representation dataset of geographic features. Topology processing: Check and fix topology errors; among which, This represents a function for handling topological relationships. A dataset representing the geometric representation of geographic features after topological processing; Data association: ; associate geometric representation and attribute information into a feature set; where, Represents the feature correlation function. This represents the set of elements after association; Format conversion: ; Convert the feature set into various standard formats; among which, Represents format conversion functions; Topology processing includes various operations, such as: ; in, This function repairs self-intersections and handles topological errors such as polygon self-intersections. This refers to the vertex snapping function, which merges vertices that are too close together into a single vertex. This function removes duplicate features, deleting duplicate geographic feature records. Data association integrates geometric representations and attribute information into feature objects: ; ; in, Indicates the first A complete geographic feature, containing both geometric and attribute information. The feature construction function combines geometric representation and attribute information into a complete geographic feature. Indicates the first [item] after topological processing Geometric representation of a geographic element Indicates the first A vector of attribute information for a geographic element. Indicates the total number of geographical elements. This represents a set of elements consisting of all geographic features. Format conversion to generate data products in various standard formats: ; ; ; ; in, Shapefile format represents a data product, a commonly used vector GIS data format. This function represents the conversion to Shapefile format. GeoJSON format represents data products, an open geographic data exchange format based on JSON. This function represents the conversion to GeoJSON format. GeoTIFF format data products are a raster image format with georeferenced information. This function represents the conversion to GeoTIFF format. This represents a vector rasterization function that converts vector data into raster data. DXF format data products are a type of CAD data exchange format. This function represents the conversion to DXF format. This represents the final deliverables set containing data products in all formats.

[0060] Through steps 401 to 403, the mapping data product generator transforms fine-grained land cover classification results and high-resolution geographic feature enhanced imagery into a professional, detailed mapping geographic data set. These data sets not only possess high geometric accuracy and rich attribute information but also conform to the standard format of geographic information systems, making them directly usable in various mapping application scenarios.

[0061] It should be noted that this implementation method innovatively integrates multiple technologies, successfully overcoming the challenge of detailed mapping in special geographical areas when samples are scarce and resolution is limited. Compared with existing technologies, this method reduces the dependence of mapping on the environment, improves efficiency and accuracy, and provides a highly efficient and precise technical solution.

[0062] Example Application: Application Scenario Description: This example was tested in the high mountain and canyon area (altitude 2000-4500 meters) on the eastern edge of the Qinghai-Tibet Plateau. The geographical environment in this area is complex and diverse, sample data is scarce, the original image resolution is low (about 0.8 meters / pixel), and the land cover categories are diverse (16 categories) and highly similar. The aim is to verify whether the method of the present invention can construct high-precision geographic mapping data products based on limited labeled samples and low-resolution images.

[0063] The data used in this embodiment includes: 5000 source domain data images (plains and hilly areas, 0.5m / pixel resolution, 1024×1024 pixels, including 12 types of land cover labels); 300 target domain labeled samples (mountain and canyon areas, 0.8m / pixel resolution, same size, including 16 types of land cover labels); 2000 unlabeled target domain data images (same area and size as the labeled samples, no labels); and 500 test data images (same area and size as the target domain labeled samples, including complete labels).

[0064] Land feature categories include: mountain forests, alpine meadows, shrublands, bare rocks, bare soil, snow-covered areas, rivers, lakes, seasonal wetlands, farmland, roads, settlements, mining areas, nature reserves, geological disaster sites, and other man-made structures.

[0065] Implementation examples of each step Step 100 Implementation Example: Feature Selective Transfer Network; its implementation uses the following specific parameters and configurations: Network structure and parameter configuration: The network architecture consists of three key modules: multi-scale feature extraction (5-layer cascaded structure, multi-size convolutional kernels), domain-invariant feature learning (3-layer fully connected network), and feature-selective attention (a combination of channel and spatial attention).

[0066] Data processing and training strategy: Data and training configuration: 4000 source domain plain and hilly images and 240 target domain high mountain and canyon images were used, combined with data augmentation techniques such as random cropping. The strategy of pre-training in the source domain for 100 rounds and then adapting to the target domain for 50 rounds was adopted. The initial learning rate was 0.001 and cosine annealing was used for adjustment.

[0067] Experimental results show that the Feature Selective Transfer Network (FSTN) of this invention achieves higher average F1 scores (0.832 and 0.876, respectively) compared to direct transfer and domain adversarial transfer methods under different target domain sample ratios (10% and 30%), and its computational efficiency is superior to that of the domain adversarial transfer method. The experimental results indicate that, using only 10% of the target domain samples, the Feature Selective Transfer Network of this invention improves the F1 score by approximately 15% compared to the direct transfer method and by approximately 8% compared to the domain adversarial transfer method, while also achieving superior computational efficiency.

[0068] Step 200 Implementation Example: Physically Constrained Generative Adversarial Network; its implementation uses the following specific parameters and configurations: Network Structure and Parameter Configuration: The physically constrained generative adversarial network consists of an 8-layer residual network degradation modeling analyzer (outputting 15-dimensional degradation parameters), a 24-layer deep residual network adaptive reconstruction generator (64-channel features), a PatchGAN structure discriminator with three physically constrained branches, and loss weights set to... , , It consists of a multi-objective optimization controller.

[0069] Data processing and training strategy: In this embodiment, 300 target domain labeled samples and 500 source domain images processed in step 100 are used. The original 0.8 m / pixel image is downsampled to 3.2 m / pixel as input, and an optimization strategy of alternating generator and discriminator ratio of 1:1 is adopted to train for 200 rounds with a batch size of 16.

[0070] Experimental results show that the Physically Constrained Generative Adversarial Network (PCGAN) of this invention outperforms traditional methods in all evaluation metrics. PCGAN achieves a Peak Signal-to-Noise Ratio (PSNR) of 30.21 dB, a Structural Similarity (SSIM) of 0.8697, and Edge Preservation (EPR) and Geometric Fidelity (GFI) of 0.7832 and 0.7645, respectively, comprehensively surpassing comparable methods such as bicubic interpolation, SRCNN, and SRGAN. The advantages of PCGAN are particularly evident in the two key metrics of edge preservation and geometric fidelity. The experimental results demonstrate that the Physically Constrained Generative Adversarial Network of this invention significantly improves upon traditional methods in both PSNR and SSIM, especially in edge preservation and geometric fidelity, which is crucial for subsequent fine-grained ground feature classification.

[0071] Step 300 Implementation Example: Hierarchical Attention Network; its implementation uses the following specific parameters and configurations: Network structure and parameter configuration: The hierarchical attention network consists of a feature adaptive regulator with an MLP structure, a multi-scale feature extractor with a 5-level FPN structure, a hierarchical spatial attention module with a three-level attention mechanism (channel / space / class), and a multi-level semantic classifier with three levels of coarse (4 classes), medium (8 classes), and fine (16 classes).

[0072] Data processing and training strategy: This embodiment uses the high-resolution image dataset generated in step 200, solves the imbalance problem through a mixed sample strategy and class balanced sampling, and adopts a three-stage training strategy (coarse, medium and fine) to train the classifier sequentially, with the learning rate set to 0.0005.

[0073] Experiments show that the hierarchical attention network (HAN) of this invention outperforms existing methods such as ResNet-50, DeepLabv3+ and SegFormer in terms of overall accuracy (0.8723), average F1 score (0.8431) and Kappa coefficient (0.8356), and the number of confused categories (12) is significantly reduced, indicating that the classification accuracy and stability are significantly improved.

[0074] Under various terrain conditions, the method of this invention outperforms the comparative methods, especially in steep areas. The highest accuracy is achieved in flat areas (0.89 for this invention), with accuracy decreasing slightly as terrain complexity increases. However, this invention maintains an accuracy above 0.83 in steep areas, while other methods show a significant decrease in accuracy under complex terrain.

[0075] Experimental results show that the hierarchical attention network of this invention outperforms existing methods in all evaluation metrics, especially in reducing easily confused categories and improving the classification accuracy of complex terrain.

[0076] Step 400 Implementation Example: Surveying and Mapping Data Product Generator; its implementation uses the following specific parameters and configurations: Module Implementation and Parameter Configuration: The surveying and mapping data product generator includes a vectorization processing module (using a deep learning boundary extraction algorithm and setting key threshold parameters), an attribute information extraction module (constructing a feature library containing multiple features), and a geographic data standardization module (using the CGCS2000 coordinate system and supporting multiple output formats).

[0077] Data processing and verification strategy: The system uses the classification result matrix from step 300 and the high-resolution image from step 200 for processing, automatically corrects topological errors, and verifies accuracy by comparing with manually vectorized results.

[0078] Experimental results show that the Mapping Data Product Generator (MDPG) of this invention outperforms traditional vectorization methods and deep learning-based methods in key indicators such as geometric accuracy, attribute accuracy, and topology error rate. MDPG achieves a geometric accuracy of 0.86 meters, an attribute accuracy of 92.56%, and a topology error rate of only 1.24%, while significantly improving processing efficiency to only 87 seconds per square kilometer, far superior to traditional methods (376 seconds) and deep learning-based methods (146 seconds).

[0079] Detailed data on the accuracy of feature extraction: The method of this invention shows significant extraction effects on various types of features, with detection rates generally exceeding 90%. Lakes showed the highest detection rate (98.32%), while mountainous forests exhibited the best location accuracy (0.94m) and the highest attribute completeness (97.35%). The detection rates for linear features (such as rivers and roads) and point features (such as settlements) both reached over 93%, with location accuracy between 0.6 and 0.8 meters and attribute completeness exceeding 93%, indicating that this method demonstrates excellent performance in extracting various types of features.

[0080] The experimental results show that the mapping data product generator of this invention outperforms traditional methods and deep learning-based methods in three key indicators: geometric accuracy, attribute accuracy, and topological error rate, while significantly improving processing efficiency. It also demonstrates a relatively balanced high accuracy across different land cover types, particularly in the extraction accuracy of linear features (such as rivers and roads), which is higher than existing methods.

[0081] Technical effectiveness verification: To comprehensively evaluate the technical effectiveness of this invention, we conducted field verification in five typical survey areas (total area of ​​approximately 150 square kilometers) in the high mountain and canyon region on the eastern edge of the Qinghai-Tibet Plateau, and compared it with traditional methods and existing advanced methods. The verification employed manual inspection of 500 random sampling points and detailed evaluation of 10 1-square-kilometer sample areas.

[0082] (1) Overall system effect evaluation: The method of the present invention surpasses traditional and advanced methods in key indicators such as comprehensive mapping accuracy (91.78%), sample utilization efficiency (only 12% of the samples required by traditional methods), resolution improvement (6.5 times), processing speed (6.4 minutes per square kilometer), types of ground features identified (16 types), and geometric accuracy (0.75 meters), which fully demonstrates its technical advantages in the field of geographic surveying and mapping.

[0083] 1. Application scenario adaptability assessment: This invention shows significant advantages in various challenging environmental conditions such as high altitude, cloudy and shadowy conditions, seasonal changes and complex terrain boundaries, with an accuracy improvement of 23-41%, and always maintains a high accuracy of over 80%, while traditional methods generally have an accuracy of less than 70% under the same conditions.

[0084] (3) Detailed evaluation of key technical indicators: The four key technical modules of this invention (feature selective transfer network, physical constraint generative adversarial network, hierarchical attention network and mapping data product generator) all perform well in terms of technical indicators such as sample utilization, resolution improvement, classification accuracy and data standardization, and are significantly better than existing methods.

[0085] (4) Overall technical effects: Based on the above verification results, the key technical effects of the present invention are summarized as follows: First, it improves data efficiency by significantly reducing reliance on labeled target domain data through feature selective transfer networks. Second, it increases resolution by enhancing image details using physical constraint generative adversarial networks. Third, it improves classification accuracy by achieving fine classification of similar features in complex scenes through hierarchical attention networks. Fourth, it improves computational efficiency with lightweight component design adapted to UAV computing environments. Fifth, it expands the application scope, applicable to various complex geographical environments and multi-source remote sensing data.

[0086] The above verification results demonstrate that the present invention has achieved a significant breakthrough in technical effectiveness, especially in terms of mapping accuracy and efficiency under small sample conditions, providing an innovative solution for the field of UAV remote sensing geographic mapping.

[0087] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A geographic mapping method based on unmanned aerial vehicle remote sensing, characterized in that, The method comprises the following steps: a feature selective migration network is used to process source domain terrain data and target domain scarce samples to obtain a feature adaptive model parameter set; a physically constrained generative adversarial network is used to process low-resolution remote sensing images to obtain a high-resolution geographic feature enhanced image data set; a hierarchical attention network is used to process the high-resolution geographic feature enhanced image data set and the feature adaptive model parameter set to obtain a fine-grained feature classification result matrix; a surveying and mapping data product generator is used to process the fine-grained feature classification result matrix and the high-resolution geographic feature enhanced image data set to obtain a fine surveying and mapping geographic data result set. 2.The geographic mapping method based on unmanned aerial vehicle remote sensing of claim 1, wherein, The step of processing source domain terrain data and target domain scarce samples by a feature selective migration network to obtain a feature adaptive model parameter set comprises: a multi-scale feature extraction module is used to process source domain remote sensing image data to obtain hierarchical terrain feature representations: ； wherein, denotes a source domain remote sensing image, denotes a hierarchical terrain feature representation, denotes a multi-scale feature extraction module; a domain invariant feature learning module is used to process the hierarchical terrain feature representations and target domain scarce samples to obtain an inter-domain mapping relationship matrix: ； wherein, denotes a target domain feature, denotes an inter-domain mapping relationship matrix, denotes a domain-invariant feature learning module; a feature selective attention module is used to process the inter-domain mapping relationship matrix and target domain features to obtain a feature adaptive model parameter set: ； wherein, is a set of feature-adaptive model parameters, is a feature-selective attention module. 3.The geographic mapping method based on UAV remote sensing of claim 1, wherein, The calculation process of the multi-scale feature extraction module is as follows: ； ； ； wherein, denotes the initial feature extraction convolutional layer processing the source domain remote sensing image to obtain the initial feature map, denotes the initial feature extraction convolutional layer, denotes the feature map of the layer, which is obtained by adding the feature map of the previous layer to the weighted sum obtained by using different convolution kernel sizes for convolution operation, denotes the number of feature levels, denotes a set of different convolution kernel sizes, denotes the weight coefficient of the corresponding convolution kernel, denotes the convolution operation with a kernel size of denotes the feature concatenation operation, is the final obtained hierarchical terrain feature representation, denotes the feature map extracted by the multi-scale feature extraction module at different levels.​​ 4. The geographic mapping method based on unmanned aerial vehicle remote sensing according to claim 1, characterized in that, The step of processing low-resolution remote sensing images by a physically constrained generative adversarial network to obtain a high-resolution geographic feature enhanced image data set comprises: an image degradation parameter vector is obtained by processing an original low-resolution image through a degradation modeling analyzer: ； wherein, denotes a low resolution remote sensing image, denotes an image degradation parameter vector, denotes a degradation modeling analyzer; an initial super-resolution image tensor is obtained by processing the low-resolution image and the image degradation parameter vector through an adaptive reconstruction generator: ； wherein, denotes the initial super-resolution image tensor, denotes the adaptive reconstruction generator; a physical consistency score matrix is obtained by processing the initial super-resolution image tensor through a physically constrained discriminator: ； wherein, denotes a physical consistency score matrix, denotes a physical constraint discriminator; a high-resolution geographic feature enhanced image data set is obtained by processing the initial super-resolution image tensor and the physical consistency score matrix through a multi-objective optimization controller: ； wherein, denotes a high resolution geographic feature enhanced imagery dataset, denotes a reference high resolution imagery, denotes a multi-objective optimization controller.

5. The geographic mapping method based on unmanned aerial vehicle remote sensing according to claim 1, characterized in that, The optimization process of the multi-objective optimization controller is based on the following loss function: ； wherein, represents a total loss function, represents an adversarial loss, represents a reconstruction loss, represents a perceptual loss, represents a physical constraint loss, , and are weight coefficients balancing the respective loss terms. 6.The geographic mapping method based on UAV remote sensing of claim 1, wherein, The step of processing the high-resolution geographic feature enhanced image data set and the feature adaptive model parameter set by a hierarchical attention network to obtain a fine-grained feature classification result matrix comprises: a domain adaptive feature extraction network is obtained by processing the feature adaptive model parameter set and a pre-trained deep network through a feature adaptive adjuster: ； wherein, denotes a domain adaptation feature extraction network parameter, denotes a pre-trained deep network parameter, denotes a feature adaptation conditioner; a multi-scale feature tensor set is obtained by processing the high-resolution geographic feature enhanced image data set through a multi-scale feature extractor: ； wherein, denotes a set of multi-scale feature tensors, denotes a multi-scale feature extractor, denotes a single image in the high-resolution geofeature augmented imagery dataset; an attention enhanced feature tensor is obtained by processing the multi-scale feature tensor set through a hierarchical spatial attention module: ； wherein, denotes an attention-enhanced feature tensor, denotes a hierarchical spatial attention module; a fine-grained feature classification result matrix is obtained by processing the attention enhanced feature tensor through a multi-level semantic classifier: ； wherein, denotes a fine-grained object classification result matrix, denotes a multi-level semantic classifier.

7. The geographic mapping method based on unmanned aerial vehicle remote sensing according to claim 1, characterized in that, The calculation expression of the hierarchical spatial attention module is as follows: ； ； wherein, , , denote the query, key, value matrices, respectively, , , denote the corresponding weight matrices, denotes the transpose of the key matrix, denotes the dimension of the key vector, denotes the activation function, denotes the self-attention output. 8.The geographic mapping method based on UAV remote sensing of claim 1, wherein, The step of processing the fine-grained feature classification result matrix and the high-resolution geographic feature enhanced image data set by a surveying and mapping data product generator to obtain a fine surveying and mapping geographic data result set comprises: a geographic feature geometric representation data set is obtained by processing the fine-grained feature classification result matrix through a vectorization processing module: ； wherein, represents a geographic feature geometry representation dataset, represents a vectorization processing module, represents a fine-grained geographic feature classification result matrix; The attribute information extraction module processes the high-resolution geographic feature enhanced image dataset and the geographic feature geometric representation dataset to obtain a geographic feature attribute information table: ； wherein, represents a table of geographic feature attribute information, represents an attribute information extraction module, represents a high-resolution geographic feature enhanced imagery dataset; The geographic data standardization module processes the geographic feature geometric representation dataset and the geographic feature attribute information table to obtain a fine surveying and mapping geographic data product set: ； wherein, is a fine mapping geographic data result set, is a geographic data standardization module. 9.The geographic mapping method based on UAV remote sensing of claim 1, wherein, The geometric construction calculation expression of the vectorization processing module is: ； wherein, represents a geometric representation of the thgeographical feature, represents a point feature constructor, represents a line feature constructor, represents a face feature constructor, represents a point coordinate, represents a sequence of contour points, represents a contour area, represents a contour width, and are threshold parameters for point and line features, respectively.

10. A geographic mapping system based on unmanned aerial vehicle remote sensing, configured to perform the geographic mapping method based on unmanned aerial vehicle remote sensing of claims 1-9, characterized in that, Comprise: A feature selective migration network is used to process source domain terrain data and target domain scarce samples to obtain a feature adaptability model parameter set; A physical constraint generation adversarial network is used to process low-resolution remote sensing images to obtain a high-resolution geographic feature enhanced image dataset; A hierarchical attention network is used to process the high-resolution geographic feature enhanced image dataset and the feature adaptability model parameter set to obtain a fine-grained ground object classification result matrix; A surveying and mapping data product generator is used to process the fine-grained ground object classification result matrix and the high-resolution geographic feature enhanced image dataset to obtain a fine surveying and mapping geographic data product set.

Images