A land survey intelligent analysis method and system based on deep learning

Abstract

This invention discloses a deep learning-based intelligent analysis method and system for land surveys, belonging to the fields of computer vision and remote sensing mapping technology. The method includes: acquiring multi-source high-resolution remote sensing images and extracting features using a super-resolution semantic segmentation network; extracting sub-pixel-level features of land cover edges using a boundary refinement module and completing sub-pixel feature point offset calibration; performing adaptive topological vectorization of patch boundaries based on edge probability response maps and offset vector fields, and using spatial geometric mechanisms to perform high-precision fine-tuning of vector boundary lines; performing coordinate space transformation and topological verification on the fine-tuned boundaries, and encapsulating the data to complete database construction and data mapping. This invention effectively overcomes the defects of sub-pixel-level boundary blurring and topological distortion in remote sensing images, achieving sub-meter-level high-precision patch boundary extraction, and meeting the needs of accurate database construction in land surveys.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and remote sensing mapping technology, and in particular to a deep learning-based intelligent analysis method and system for land surveys. Background Technology

[0002] Against the backdrop of the construction of the national spatial planning system and the refined management of natural resources, the national land survey business is transforming towards a micro-level and precise scale, and the extraction accuracy of land feature boundaries such as high-standard farmland and homesteads has been raised to the stringent requirements of sub-meter or even centimeter level.

[0003] Current remote sensing image semantic segmentation techniques are limited by the inherent physical limits of image spatial resolution. Extracted patch boundaries often exhibit jagged edges accompanied by a grid effect, leading to severe sub-pixel-level boundary blurring and topological distortion. This low-precision result easily causes topological overlap and area calculation errors during vector polygon conversion and data entry. Furthermore, traditional edge detection operators lack high-level semantic understanding capabilities, making them prone to misidentifying false edges or breaking genuine edges in complex remote sensing scenes, thus failing to maintain the global topological integrity of feature boundaries.

[0004] Currently, there is a lack of a good technology that can effectively solve the above problems.

[0005] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based intelligent analysis method and system for land surveys, in order to solve the problems of sub-pixel-level boundary ambiguity and topological distortion in the semantic segmentation of remote sensing images in the prior art.

[0007] The technical solution of the present invention is as follows: On one hand, this invention provides a deep learning-based intelligent analysis method for land surveys. The method includes: acquiring multi-source high-resolution remote sensing images and extracting features using a preset super-resolution semantic segmentation network to output an initial super-resolution semantic segmentation map and a corresponding initial multi-class edge probability response map; processing the initial super-resolution semantic segmentation map and the initial multi-class edge probability response map using a preset boundary refinement module to extract sub-pixel-level features of land cover edges and complete sub-pixel feature point offset calibration, outputting a high-precision edge probability response map and a two-dimensional sub-pixel offset vector field; performing adaptive topological vectorization of patch boundaries based on the high-precision edge probability response map and the two-dimensional sub-pixel offset vector field, and using a preset spatial geometric fine-tuning mechanism to perform high-precision fine-tuning of the vectorized patch boundary lines; performing coordinate space transformation and topological verification on the fine-tuned high-precision patch boundaries, and encapsulating the verified boundary data to complete database construction and land parcel assignment.

[0008] Optionally, the step of acquiring multi-source high-resolution remote sensing images and extracting features using a preset super-resolution semantic segmentation network includes: inputting the remote sensing images into a backbone feature extraction network to obtain a multi-scale feature pyramid; introducing an attention mechanism into the multi-scale feature pyramid to adaptively suppress background noise features and activate target salient features; applying sub-pixel convolution operations to the aggregated multi-scale feature map for feature reconstruction, and combining the edge attention mechanism to output the initial super-resolution semantic segmentation map and the corresponding initial multi-class edge probability response map.

[0009] Optionally, the step of applying sub-pixel convolution operations to the aggregated multi-scale feature maps for feature reconstruction and combining an edge attention mechanism to output the initial super-resolution semantic segmentation map includes: expanding the number of channels of the feature maps using convolutional layers, then applying a periodic shuffling mechanism to rearrange the expanded feature maps in the spatial dimension to convert them into high-resolution feature maps; introducing the rearranged high-resolution feature maps into auxiliary convolutional branches to predict a coarse semantic boundary distribution, and multiplying this boundary distribution back into the backbone feature map as a spatial gating signal to generate the initial super-resolution semantic segmentation map after edge enhancement.

[0010] Optionally, the process of using a preset boundary refinement module to process the initial super-resolution semantic segmentation map and the initial multi-class edge probability response map includes: applying edge gradient differential enhancement operations to amplify the real physical boundary response and generate a high signal-to-noise ratio edge feature map; calculating the information entropy of each spatial grid position based on the normalized probability distribution of the initial multi-class edge probability response map to determine the boundary ambiguity region, and deploying a two-branch orthogonal decoupling network in this region to extract pure semantic category features and pure spatial boundary features with inner products approaching zero; inputting the enhanced edge feature map and the pure spatial boundary features into the offset regression prediction sub-network, outputting the two-dimensional sub-pixel offset vector field, and using it to perform geometric deformation and coordinate remapping on discrete pixel feature points to complete adaptive adsorption and calibration.

[0011] Optionally, the application of edge gradient differentiation enhancement amplifies the true physical boundary response to generate a high signal-to-noise ratio edge feature map, including: Calculate the first-order feature gradient magnitude scalar using depthwise separable convolution operations With the second-order Laplace response scalar And an enhanced boundary feature map is generated using a nonlinear differential enhancement mapping formula:

[0012] in, Represents an enhanced boundary feature map; A high-resolution feature map representing the input; Represents the global scaling factor of the differential features, which is dynamically set based on the physical resolution of the image. This represents the hyperbolic tangent nonlinear activation function; Represents the weight of the first-order gradient response contribution; This represents the contribution weight of the second-order boundary zero-crossing localization.

[0013] Optionally, after completing the coordinate remapping, the method further includes implementing edge continuity constraints, the steps of which are: applying a non-maximum suppression algorithm to the edge probability response map after offset calibration and resampling to extract sub-pixel boundary skeleton lines with a single pixel width; tracing the connected domain nodes on the boundary skeleton lines to construct an edge continuity topological constraint loss function; and incorporating the edge continuity penalty loss value into the global network training to penalize ground feature boundaries with abnormal jagged edges or unnatural sharp angle abrupt changes during backpropagation. The formula for calculating the edge continuity topological constraint loss function, which involves tracking connected nodes on the boundary skeleton line, is as follows:

[0014] in, This represents the edge continuity penalty loss value; This represents the total number of discrete points of the extracted boundary skeleton. , Representing the first The boundary skeleton point and its adjacent points. Local topological tangent direction vectors of each boundary skeleton point; Representative bestowed upon the first The adaptive continuous dynamic weights of each vertex, and ,in For the first The pixel coordinate vectors of discrete points of the boundary skeleton on the two-dimensional plane of the raster image. This represents the information entropy value of the uncertainty metric matrix corresponding to the pixel coordinate vector.

[0015] Optionally, the adaptive topological vectorization of the patch boundary based on the high-precision edge probability response map and the two-dimensional sub-pixel offset vector field, and the high-precision fine-tuning of the vectorized patch boundary lines using a preset spatial geometric fine-tuning mechanism, includes: constructing a local offset statistical covariance matrix by combining the terrain undulation measurement and local spatial context of the digital elevation model data, and calculating the spatial error confidence of the predicted position of each vector polygon vertex; performing initial topological vectorization using a contour extraction algorithm combined with the two-dimensional sub-pixel offset vector field, and repairing the continuity of the boundary lines by injecting prior geometric constraints of ground features according to semantic categories; calculating the discrete average curvature of the polygon vertices for smoothness verification; constructing an image gravity field model in the physical grayscale space of the remote sensing image, and reprojecting the vector polygon nodes into the model to drive the nodes to perform physical gravity adsorption fine-tuning.

[0016] Optionally, the step of constructing an image gravitational field model within the physical grayscale space of the remote sensing image, and reprojecting the vector polygon nodes into this model to drive the nodes to perform physical gravitational adsorption fine-tuning, includes: processing the basic grayscale map of the remote sensing image using a Gaussian filter, calculating the square of the spatial gradient magnitude, and constructing a Gaussian image gravitational field potential energy model; projecting each spatial node of the vector polygon into this model, and driving the vector node movement based on the displacement iterative update fine-tuning formula:

[0017] in, and These represent the coordinate vectors of the nodes in the vector polygon space before and after the iteration; The two-dimensional spatial gradient vector representing the gravitational field of the Gaussian image at the current node position; This represents the resistance of the energy partial derivative gradient within the geometry used to maintain the topological shape of polygons without distortion. This represents the basic iteration step size coefficient, which is dynamically set based on the image spatial resolution. The spatial stiffness damping coefficient of a polygonal structure representing the balance of internal and external forces.

[0018] Optionally, the step of performing coordinate space transformation and topological verification on the fine-tuned high-precision map boundary, and encapsulating the verified boundary data to complete the database construction and mapping, includes: using a parametric model or affine transformation matrix to project the fine-tuned vector polygon set from the image pixel coordinate system to the national geodetic absolute coordinate system; running a spatial topology tolerance check algorithm to identify and repair illegal overlapping or self-intersecting nodes; using boundary error weight allocation technology to determine whether the accuracy meets the business assessment requirements for fine land division; if not, using an error distribution scale decomposition method to drive the fine-tuning steps for closed-loop iterative optimization; if compliant, encapsulating the results into a geographic information data format with extended attributes and pushing it to the national land spatial information platform through a standard interface.

[0019] On the other hand, this invention also provides a deep learning-based intelligent analysis system for land surveys. The system includes: a feature extraction module for connecting to the raw data stream from aerial cameras, reading multi-source images, and performing depth normalization preprocessing for geometric deformation and scale; driving high-performance GPU computing power to perform convolution calculations containing spatial hole dilation rates to construct a global feature pyramid with macroscopic perception capabilities; and directly performing high-dimensional tensor channel expansion and rearrangement shuffling operations within the GPU memory to achieve initial physical enhancement and improvement of feature map resolution, outputting an initial super-resolution semantic segmentation map and corresponding initial multi-class edges. The system includes: a probability response map; a boundary refinement module for nonlinear enhancement by calculating and amplifying gradient signals of weak real edges through depthwise separable convolution; a Softmax probability dynamic calculation matrix information entropy to determine fuzzy boundary regions, and an orthogonalization calculation to force the inner product of semantic feature flow and boundary high-frequency physical feature flow to approach zero for dual-branch decoupling; extraction of map skeleton lines, and imposing topological deformation constraints based on orientation angle change penalty mechanism on the deep network during backpropagation; and continuous output of a two-dimensional offset vector field based on the regression network model, using deformable convolution operators to drive the mesh. Discrete pixel feature points are drifted and aligned to the physical ground truth boundary to complete deep edge feature extraction and calibration; a geometric fine-tuning module is used to calculate the local covariance matrix and Mahalanobis distance confidence score distribution map in real time by combining terrain undulation and heterogeneity to assist in judging error clustering hotspots; a contour extraction algorithm is called and combined with mathematical tools such as collinear fitting of straight line segments to drive polygons to conform to specific prior artificial geometric feature specifications for adaptive vectorization; the discrete curvature of the polygon vertex array is calculated in real time, and Bézier curves or B-spline mathematical models are intelligently triggered and implanted in areas where curvature exceeds the limit for smooth transition processing; and an external high-resolution model is constructed. The system balances the gravitational potential field of the stretched image with the internal topological defense holding force, and performs sub-pixel-level lossless adsorption within a limited step size to complete adaptive iterative fine-tuning of boundary points. The database construction and mapping module is used to enforce the mathematical projection transformation from the pixel image coordinate system to the national geodetic absolute coordinate system, and removes all illegal topological intersections that may cause database entry failure for joint verification. After confirming that the boundary meets the high-precision assessment indicators, it automatically attaches geographical attribute business fields such as category, perimeter, and area, and directly streams and distributes them to the cloud-based basic information big data array center through network transmission protocols to complete accurate database construction and mapping.

[0020] The beneficial effects of this invention are as follows: This invention acquires multi-source high-resolution remote sensing images and extracts features using a pre-defined super-resolution semantic segmentation network. It combines channel attention and spatial attention mechanisms to activate the salient features of the target, effectively overcoming the bottleneck of traditional remote sensing images being limited by physical spatial resolution. This improves the basic resolution and edge sensitivity of the images during the feature extraction stage, providing rich deep and high-dimensional feature support for subsequent high-precision boundary extraction.

[0021] Meanwhile, this invention processes the initial super-resolution semantic segmentation map by using a preset boundary refinement module, applies edge gradient differentiation enhancement and constructs a dual-branch orthogonal decoupling network, thereby achieving effective separation of macroscopic semantic category features and high-frequency spatial boundary features. Then, it uses the predicted two-dimensional sub-pixel offset vector field to perform coordinate remapping calibration on discrete feature points, enabling the network to overcome the physical limitations of grid pixels and significantly improve the sub-pixel level boundary positioning accuracy at the junction of heterogeneous regions.

[0022] Furthermore, this invention also performs adaptive topological vectorization of patch boundaries based on spatial error confidence assessment, and constructs an image gravity field model in the physical gray space of remote sensing images to drive vector polygon nodes to perform controlled physical gravity adsorption fine-tuning, so that the extracted vector lines can adaptively fit the real texture abrupt changes of actual ground objects in the physical world, thereby effectively alleviating the jagged or stepped distortion problems that are common in traditional segmentation results accompanied by grid effects.

[0023] Finally, this invention also performs precise coordinate space transformation and topological tolerance joint verification on the fine-tuned high-precision map patch boundaries, and uses the error distribution scale decomposition method to drive the fine-tuning steps for closed-loop iterative optimization. This ensures that the output boundary description data has high geometric regularity and topological connectivity, thereby meeting the business assessment requirements of the national land spatial information platform for sub-meter level map patch data, significantly reducing the cost of manual post-editing intervention, and improving the overall efficiency of database construction and parcel assignment. Attached Figure Description

[0024] Figure 1 A flowchart illustrating a deep learning-based intelligent analysis method for land surveys provided in this embodiment of the invention. Figure 1 ; Figure 2 A flowchart illustrating a deep learning-based intelligent analysis method for land surveys, provided as another embodiment of the present invention. Figure 2 ; Figure 3 This is a framework architecture diagram of a deep learning-based intelligent analysis system for land surveys provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0026] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0027] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0028] As mentioned earlier, against the backdrop of comprehensively advancing the construction of the national spatial planning system and the refined management of natural resources, land survey operations are undergoing a profound transformation from a macro-scale to a micro- and precise scale. Current remote sensing image semantic segmentation technology is limited by the inherent physical limits of image spatial resolution, often resulting in jagged edges with grid effects on extracted patch boundaries, leading to severe "sub-pixel-level boundary blurring and topological distortion." When these low-precision raster segmentation results are forcibly converted into vector polygons and attempted for data entry into the database, it often causes significant topological overlap, gaps, and area calculation errors, making it impossible for operational data to directly meet the precise database construction requirements of the national spatial information platform. Simultaneously, traditional edge detection operators lack high-level semantic cognition capabilities, making them prone to misidentifying false edges or causing breaks in real edges in complex remote sensing scenes, thus failing to maintain the global topological integrity of feature boundaries.

[0029] In response to this, the present invention provides a deep learning-based intelligent analysis method and system for land surveys, which solves the problems in the prior art in the following ways.

[0030] Example 1: like Figure 1 As shown, Embodiment 1 of the present invention provides a land survey intelligent analysis method based on deep learning. The executing entity can be a controller, and the method mainly includes the following steps: S100: Acquire multi-source high-resolution remote sensing images and extract features using a preset super-resolution semantic segmentation network to output an initial super-resolution semantic segmentation map and the corresponding initial multi-class edge probability response map. S200. The initial super-resolution semantic segmentation map and the initial multi-class edge probability response map are processed using a preset boundary refinement module to extract sub-pixel-level features of ground object edges and complete sub-pixel feature point offset calibration, and output a high-precision edge probability response map and a two-dimensional sub-pixel offset vector field. S300. Based on the high-precision edge probability response map and the two-dimensional sub-pixel offset vector field, adaptive topological vectorization of the patch boundary is performed, and a preset spatial geometric fine-tuning mechanism is used to perform high-precision fine-tuning of the vectorized patch boundary lines. S400 performs coordinate space transformation and topology verification on the fine-tuned high-precision map boundary, and encapsulates the verified boundary data to complete the database construction and mapping.

[0031] Based on the above steps, the present invention provides a deep learning-based intelligent analysis method for land surveys. This method acquires multi-source high-resolution remote sensing images and extracts features using a pre-defined super-resolution semantic segmentation network. It combines channel attention and spatial attention mechanisms to activate salient feature characteristics, effectively overcoming the bottleneck of traditional remote sensing images being limited by physical spatial resolution. This improves the basic resolution and edge sensitivity of the images during feature extraction, providing rich deep, high-dimensional feature support for subsequent high-precision boundary extraction. Simultaneously, by using a pre-defined boundary refinement module to process the initial super-resolution semantic segmentation map, the present invention achieves effective separation of macroscopic semantic category features and high-frequency spatial boundary features. This allows the network to overcome the physical limitations of raster pixels, significantly improving sub-pixel-level boundary localization accuracy at heterogeneous region boundaries.

[0032] Example 2: Based on the above embodiments, in order to provide a clearer and more complete explanation of the technical solutions therein, the present invention also provides an embodiment two. For example... Figure 2 As shown in this second embodiment, in another embodiment of the present invention, in step S100, acquiring multi-source high-resolution remote sensing images and extracting features using a preset super-resolution semantic segmentation network may include: Step S110: Input the remote sensing image into the backbone feature extraction network to obtain a multi-scale feature pyramid; Step S120: Introduce an attention mechanism into the multi-scale feature pyramid to adaptively suppress background noise features and activate target salient features; Step S130: Apply sub-pixel convolution operation to the aggregated multi-scale feature map for feature reconstruction, and combine the edge attention mechanism to output the initial super-resolution semantic segmentation map and the corresponding initial multi-class edge probability response map.

[0033] Step S130 further includes: Step S131: Expand the number of channels of the feature map using convolutional layers, and then apply a periodic shuffling mechanism to rearrange the expanded feature map in the spatial dimension and convert it into a high-resolution feature map. Step S132: Introduce the rearranged high-resolution feature map into the auxiliary convolution branch to predict the coarse semantic boundary distribution, and multiply this boundary distribution back into the backbone feature map as a spatial gating signal to generate the initial super-resolution semantic segmentation map after edge enhancement.

[0034] For example, steps S110 to S132 may further include the following steps: The system uses a pre-set input stream interface to receive high-resolution remote sensing images that have undergone orthorectification and feeds them into a backbone feature extraction network based on a combination of deep residual structure and void spatial pyramid pooling.

[0035] The network obtains receptive fields of different scales through convolutional kernels with different dilation rates, and extracts a multi-scale feature pyramid that contains both large-scale low-frequency contextual semantic information and preserves local high-frequency texture details.

[0036] Subsequently, the system injects channel attention and spatial attention mechanisms into the top and middle layers of the feature pyramid, adaptively calculates the weight distribution of feature maps on each channel, actively suppresses background noise features that are irrelevant to the current land survey task, and selectively activates salient feature responses related to targets such as high-standard farmland boundaries and homestead building outlines.

[0037] Next, sub-pixel convolution operations are applied to the aggregated multi-scale feature maps, expanding the number of channels of the feature maps to [number missing] through a regular convolutional layer. (in In this embodiment, the super-resolution spatial magnification factor is set as follows: ), and apply a periodic mixing and washing mechanism to this The pixels of each channel are rearranged to convert the channel count to a number of channels. Spatial resolution is magnified in both horizontal and vertical directions. High-resolution feature maps.

[0038] Finally, the rearranged high-resolution feature map is introduced into the prior edge attention mechanism module. The coarse semantic boundary distribution is predicted by the auxiliary convolution branch and multiplied back into the backbone feature map as a spatial gating signal, thereby outputting the initial super-resolution semantic segmentation map and the corresponding initial multi-class edge probability response map.

[0039] In another embodiment of the present invention, step S200, which involves processing the initial super-resolution semantic segmentation map and the initial multi-class edge probability response map using a preset boundary refinement module, may further include: Step S210: Apply edge gradient differentiation enhancement operation to amplify the real physical boundary response and generate a high signal-to-noise ratio edge feature map; Step S210 may further include: Step S211: Calculate the first-order feature gradient magnitude scalar and the second-order Laplacian response scalar using depthwise separable convolution operations, and generate the enhanced boundary feature map using the nonlinear differential enhancement mapping formula; Step S220: Calculate the information entropy of each spatial grid position based on the normalized probability distribution of the initial multi-class edge probability response map to determine the boundary ambiguity area, and deploy a two-branch orthogonal decoupling network in the area to extract pure semantic category features and pure spatial boundary features with inner product approaching zero respectively. Step S230: Input the enhanced edge feature map and the pure spatial boundary feature into the offset regression prediction subnetwork, output the two-dimensional sub-pixel offset vector field, and use it to perform geometric deformation and coordinate remapping on discrete pixel feature points to complete adaptive adsorption and calibration. Step S240: Apply the non-maximum suppression algorithm to the edge probability response map after offset calibration and resampling to extract the sub-pixel boundary skeleton line of single pixel width; Step S250: Trace the connected nodes on the boundary skeleton line and construct the edge continuity topological constraint loss function; Step S260: Incorporate the edge continuity penalty loss value into the global network training, and penalize the boundaries of ground features with abnormal jagged edges or unnatural sharp angle abrupt changes during backpropagation.

[0040] For example, steps S210 to S250 can further include the following steps: First, apply a learnable depthwise separable convolution operation to each channel to calculate the first-order feature gradient magnitude scalar and the second-order Laplacian response scalar. Then, generate the enhanced boundary feature map using the nonlinear differential enhancement mapping formula, as follows:

[0041] in, Represents an enhanced boundary feature map; This represents a high-resolution super-resolution feature map output by a sub-pixel convolution mechanism; This represents a scalar of the magnitude of the first-order feature gradient calculated in the network feature space, which is obtained by applying predefined weights to the input tensor. Depthwise separable convolution (horizontal and vertical Sobel operators) is used to calculate the square root of the sum of squares of partial derivatives, which characterizes the degree of drastic change in local regions. The second-order Laplace response scalar represents the interior of the feature space, which uses the standard that the center weights are negative and the periphery weights are positive. Laplacian kernel convolution extraction, taking the absolute value to indicate the extreme points of image feature change intensity, i.e., the zero-crossing points of the physical real boundary; Represents the global scaling factor for dynamic differential features. Its value is forcibly bound to the physical ground resolution (GSD) of the current input remote sensing image. The calculation formula is defined as follows:

[0042] in The metric value corresponding to the actual physical resolution of the input image; The standard hyperbolic tangent nonlinear activation function is used to smoothly and monotonically normalize and compress the gradient-enhanced joint response value to a value that is... Within the range.

[0043] Represents the weights contributing to the first-order gradient response. This represents the contribution weight of the second-order boundary zero-crossing localization. and These are all learnable parameters in deep networks. To accelerate network convergence, based on prior statistics of the proportion of first-order and second-order gradient energy distributions in the land survey sample set, [the following parameters are used]. The network initial weights are set to 0.70. The network initial weights are set to 0.30. During the model training phase, the system calculates the gradient of the loss function using the backpropagation algorithm and adaptively updates the weights dynamically. and The value of is selected to find the optimal local solution.

[0044] Next, the information entropy of each grid position is calculated based on the normalized probability distribution of the initial semantic segmentation map, and the global mean of the information entropy of all grid positions in the current input feature map is calculated. The system will As a dynamically set threshold, regions with entropy values ​​greater than this threshold are defined as boundary ambiguity regions. A two-branch orthogonal decoupling network is deployed within these ambiguity regions, introducing an orthogonality loss function to force the inner product of the pure semantic category features extracted by the left branch and the pure spatial boundary features extracted by the right branch to approach zero in the latent space. Then, the enhanced boundary feature map is concatenated with the separated high-frequency boundary features and input into a shift regression prediction sub-network consisting of three consecutive convolutional layers. The output is a tensor with two channels, representing a two-dimensional sub-pixel shift vector field in the horizontal and vertical directions. This vector field is then used for coordinate remapping and adaptive snapping calibration using bilinear spatial resampling techniques or deformable convolution.

[0045] To ensure physical continuity of the boundaries, a non-maximum suppression algorithm is applied to the response map to remove redundant responses and extract boundary skeleton lines with a width of one pixel. The system tracks the connected nodes of the skeleton lines and constructs the following discretized edge continuity topological constraint loss function for penalty:

[0046] in, This represents the edge continuity penalty loss value; This represents the total number of discrete points of the extracted boundary skeleton. , Representing the first The boundary skeleton point and its adjacent points. The local topological tangent direction vector of each boundary skeleton point; obtained by the two-dimensional plane direction vector formed by two adjacent points; The ratio of the standard inner product to the product of the magnitudes of two adjacent local topological tangent direction vectors is used to calculate the cosine of the included angle. Representative bestowed upon the first The adaptive continuous dynamic weights of each vertex, and ,in For the first The pixel coordinate vectors of discrete points of the boundary skeleton on the two-dimensional plane of the raster image. This represents the information entropy value of the uncertainty metric matrix corresponding to the pixel coordinate vector; regions with higher uncertainty will adaptively reduce the hard constraint weight of geometric continuity.

[0047] In this second embodiment, in step S300, the adaptive topological vectorization of the patch boundary based on the high-precision edge probability response map and the two-dimensional sub-pixel offset vector field, and the high-precision fine-tuning of the vectorized patch boundary lines using a preset spatial geometric fine-tuning mechanism, may include: Step S310: Combining the terrain relief measurement and local spatial context of the digital elevation model data, construct the local offset statistical covariance matrix and calculate the spatial error confidence of the predicted position of each vector polygon vertex. Step S320: Initial topology vectorization is performed using a contour extraction algorithm combined with a two-dimensional sub-pixel offset vector field, and the continuity of boundary lines is repaired by injecting prior geometric constraints of ground features according to semantic categories. Step S330: Calculate the discrete average curvature of the polygon vertices to perform smoothness verification; Step S340: Construct an image gravity field model in the physical gray space of the remote sensing image, and reproject the vector polygon nodes into the model to drive the nodes to perform physical gravity adsorption fine-tuning.

[0048] Step S340 may further include: Step S341: Use a Gaussian filter (e.g., standard deviation preset to 1). The two-dimensional Gaussian filter is used to process the basic grayscale map of the remote sensing image and calculate the square of the spatial gradient magnitude to construct a Gaussian image gravitational field potential energy model. Step S342: Project each spatial node of the vector polygon into the model, and drive the vector node to move based on the displacement iterative update fine-tuning formula to implement physical gravity adsorption fine-tuning.

[0049] For example, steps S310 to S342 can further include the following steps: Based on the Mahalanobis distance theory in local spatial statistics, the system does not blindly trust the deterministic output of a single network, but instead calculates the spatial error confidence level using the following formula:

[0050] in, The first one represents the calculated output. The spatial error confidence score scalar of the vertices of a vector polygon, whose range is constrained by an exponential function, is within a certain range. between; The representation is the prediction obtained through the offset regression network module for the first... Independent two-dimensional sub-pixel offset vectors of vertex positions of a vector polygon; The mean eigenvector representing the local offset of the micro-environment of the vector polygon's vertices is obtained by using the vertices of the vector polygon... Accurately extract the periphery for the geometric center The arithmetic mean of the offset vectors of all boundary points within the pixel neighborhood window is obtained; Represents the The two-dimensional local offset statistical covariance matrix of the distribution of all spatial offset vectors within the pixel neighborhood window; The algebraic inverse matrix representing the aforementioned local offset statistical covariance matrix; This represents the Gaussian decay constant of the spatial confidence level. To adapt to specific sub-meter tolerance evaluation standards, The calibration method is as follows: ,in To verify the variance of the sub-pixel average spatial regression error obtained from centralized statistics, when the assessment tolerance requirement is adjusted, the statistics are re-statistically analyzed. This allows for dynamic acquisition of the appropriate settings. Numerical value.

[0051] The system employs the MarchingSquares contour extraction algorithm for initial adaptive topology vectorization and injects geometric constraints based on semantic categories (e.g., right-angle alignment constraints for residential land or least-squares fitting line collinearity constraints for high-standard farmland). The system further calculates the discrete mean curvature of nodes. When the calculated curvature is lower than a preset smoothness verification threshold and the spatial error confidence level is low, it triggers local offset fine-tuning using cubic B-splines or Bézier curves and removes outlier nodes through DBSCAN clustering. Finally, to further push the accuracy limits, the system constructs a Gaussian gravitational field model, where the Gaussian filter is the standard deviation. The system uses a two-dimensional Gaussian smoothing kernel and drives the movement of vector polygon vertices through the following displacement iteration update fine-tuning formula:

[0052] in, and Others represent the first The second iteration and the first After several fine-tuning calculation iterations, the coordinate vectors of the vertices of the vector polygon in the two-dimensional coordinate system of the image are obtained. The two-dimensional spatial gradient vector representing the gravitational field of the Gaussian image at the current vertex coordinates of the vector polygon; The energy resistance within the geometry used to maintain the reasonable topological shape of a polygon without distortion is defined by the computational logic as follows: This represents the gradient of the partial derivative of the aforementioned internal resistance energy with respect to the spatial coordinates of the current vector polygon vertex; This represents the basic iteration step size coefficient dynamically set based on image spatial resolution; its calculation formula is: ,in The input image is given a metric value corresponding to its actual physical resolution to ensure that the fine-tuning displacement is strictly locked at the sub-pixel level.

[0053] The spatial stiffness damping coefficient of a polygonal structure representing the equilibrium of internal and external forces. Its acquisition depends on the semantic category probabilities output by the semantic segmentation network, and is obtained through continuous mapping using the following formula: .in, The probability confidence level that the image patch output by the network belongs to a regular geometric configuration category (such as a homestead); The lower limit of the foundation damping for irregular natural features is set to 0.2. This is the upper limit of the basic damping for regular features (set to 0.8). Using this formula, polygons with higher shape regularity will be assigned a higher damping value to resist minor deformations.

[0054] In this second embodiment, step S400, which involves performing coordinate space transformation and topological verification on the fine-tuned high-precision patch boundaries and encapsulating the verified boundary data to complete the database construction and mapping, may include: Step S410: Using a parametric model or affine transformation matrix, project the fine-tuned vector polygon set from the image pixel coordinate system to the national geodetic absolute coordinate system. Step S420: Run the space topology tolerance check algorithm to identify and repair illegal overlapping or self-intersecting nodes; Step S430: Use boundary error weighting technique to determine whether the accuracy meets the performance evaluation requirements for fine land division. If step S440 does not meet the requirements, the error distribution scaling method is used to drive the fine-tuning step for closed-loop iterative optimization. Step S450: If the conditions are met, the results will be packaged into a geographic information data format with extended attributes and pushed to the national land and space basic information platform through a standard interface.

[0055] For example, steps S410 to S450 may further include the following steps: The vector polygon set is precisely projected onto the national geodetic absolute coordinate system (such as CGCS2000) using the RPC parameter model or high-precision affine transformation matrix attached to the remote sensing image. Then the system automatically runs a spatial topology tolerance check algorithm to identify and repair residual dangling nodes or illegal minor overlaps.

[0056] When updating the resource protection database using the above results, the boundary error weight allocation technique is used to automatically determine whether the accuracy fully meets the business assessment requirements. If the local accuracy does not meet the standard, discrete wavelet transform is used to decompose the error to different spatial frequency scales, driving the preceding fine-tuning steps to perform closed-loop iterations until complete convergence.

[0057] The final boundary delineation description that meets the requirements will be encapsulated in GeoJSON or structured Shapefile format and pushed directly to the national land and space information platform through a standard API interface, completing the entire chain of unmanned, high-precision intelligent business from original image input to database construction and land parcel establishment.

[0058] Example 3: Based on the same overall inventive concept, this invention also provides a deep learning-based intelligent analysis system for land surveys. This system shares a highly consistent logical flow with the aforementioned method embodiments and features a deeply integrated hardware and software computing power architecture, enabling it to physically execute and perfectly support the operation of the aforementioned series of high-precision method processes. For example... Figure 3 As shown, the system specifically includes: The multi-source remote sensing image low-level acquisition and spatial feature deep extraction central module is used to interface with the low-level data input and perform deep-dimensional semantic extraction on the basic image. In order to achieve normalization and feature reconstruction of images of different scales and deformations, this central module is further subdivided into image parsing, correction and scale preprocessing scheduling sub-modules, and global multi-dimensional semantic super-resolution cross-scale mapping sub-module.

[0059] Specifically, the image analysis, correction and scale preprocessing scheduling submodule is equipped with a multi-source heterogeneous data joint normalization control unit. This control unit is responsible for interfacing with the raw data stream of the aerial camera, reading multi-source images such as visible light and near-infrared, and performing depth normalization preprocessing operations for geometric deformation and scale.

[0060] As the core support for deep feature extraction, the global multidimensional semantic super-resolution cross-scale mapping submodule deploys a deep residual dilated multi-scale feature pyramid construction unit and a sub-pixel convolutional feature reconstruction unit based on a channel rearrangement mechanism. The pyramid construction unit drives high-performance GPU computing power to perform convolutional calculations with spatial dilation rates, constructing a global feature pyramid with macroscopic perception capabilities. The feature reconstruction unit directly performs high-dimensional tensor channel expansion and rearrangement shuffling operations (PixelShuffle) within the GPU memory, achieving an initial physical improvement and enhancement of the feature map resolution.

[0061] The boundary refinement and sub-pixel-level deep edge feature extraction network processing module takes the high-dimensional features output by the aforementioned central module and uses them to thoroughly extract sub-pixel-level features of ground object edges in remote sensing images. This processing module is further divided into a sub-pixel convolutional high-dimensional feature capture and extremum enhancement sub-module, and a feature orthogonal decoupling and geometric constraint sub-module for complex boundary ambiguity areas.

[0062] Furthermore, the sub-pixel convolutional high-dimensional feature capture and extremum enhancement submodule is configured with a feature space edge gradient differential calculation and nonlinear enhancement unit, as well as a depth tensor sub-pixel feature point spatial offset calibration operation unit. The former internally embeds nonlinear differential enhancement logic, which quickly calculates and amplifies the gradient signal of weak real edges through depthwise separable convolution; the latter relies on a continuous regression network model to continuously output a two-dimensional offset vector field, and uses deformable convolution operators to drive discrete pixel feature points within the grid to drift and align towards the physical ground truth boundary.

[0063] To address the severe pixel confusion at the boundaries of heterogeneous regions, the complex boundary fuzzy region feature orthogonal decoupling and geometric constraint submodule embeds an information entropy calculation and evaluation unit and a dual-branch orthogonal decoupling separation unit for fuzzy regions, as well as a physical morphology edge topological continuity orientation angle constraint loss calculation unit. The separation unit uses Softmax probability to dynamically calculate matrix information entropy to determine fuzzy boundary regions and performs orthogonalization calculations to force the inner product of the semantic feature flow and the boundary high-frequency physical feature flow to approach zero. The loss calculation unit systematically extracts the skeleton lines of the image patches and, during backpropagation, forcibly applies a topological deformation constraint based on an orientation angle change penalty mechanism to the deep network.

[0064] The raster-to-vector cross-domain high-precision patch boundary polygonization and geometric fine-tuning correction module is used to transform the raster-form probability field inferred from the network layer into structured vector patch polygons and perform adaptive fine-tuning. This correction module is deconstructed into a spatial polygon boundary adaptive vectorization and local error distribution quantization evaluation submodule, and a discrete spatial vector boundary curve adaptive smoothing and gravitational field feedback fine-tuning submodule.

[0065] As a preprocessing step in this stage, the evaluation submodule includes a unit for evaluating the spatial distribution density of map boundary errors that combines terrain undulations and heterogeneity. This unit deeply nests a submodule for real-time calculation of the local covariance matrix of spatial errors, specifically designed for rapidly calculating the Mahalanobis distance confidence score distribution map to aid in identifying error clustering hotspots. Simultaneously, this evaluation submodule also includes a collaborative optimization unit that integrates prior knowledge of land cover categories and boundary line geometric rule constraints. This unit invokes contour line extraction algorithms and combines mathematical tools such as collinear fitting of straight line segments to drive polygons to conform to specific prior artificial geometric feature specifications.

[0066] After initial optimization, the discrete spatial vector boundary curve adaptive smoothing and gravitational field feedback fine-tuning submodule is activated. Its built-in boundary smoothness multidimensional quantization verification and adaptive curve interpolation correction unit calculates the discrete curvature of the polygon vertex array in real time, and intelligently triggers and implants cubic Bézier curves or B-spline mathematical models for smooth transition processing in the curvature excess region. More importantly, the boundary point adaptive iterative adsorption fine-tuning unit based on Gaussian gradient potential energy gravitational field in this submodule constructs a balance equation between the external Gaussian image stretching gravitational potential field and the internal topological defense holding force, and performs sub-pixel level lossless adsorption action within a limited step size.

[0067] The national geographic information spatial data standard interface docking and precise database construction and data entry terminal module, as the system's business closed-loop execution component, is used to realize the conversion of high-precision boundary coordinates to the national standard coordinate system and the final data entry into the database. This terminal module is composed of a multi-source spatial geographic coordinate system topological joint verification and affine transformation submodule, and a closed boundary business structured data attribute encapsulation and high-speed streaming data entry into the database submodule.

[0068] The transformation submodule is equipped with a coordinate system high-dimensional inversion projection and patch self-intersection and gap topological geometry joint verification unit, which forces the mathematical projection transformation of the pixel image coordinate system to the national geodetic absolute coordinate system such as CGCS2000, and removes all illegal topological intersections that may cause the database to fail.

[0069] Finally, the data entry submodule is processed by an automated GeoJSON / Shapefile standard encapsulation unit with weight allocation and scale decomposition optimization. After confirming that the boundary meets the high-precision assessment indicators for fine land division, it automatically attaches geographic attribute business fields such as land feature category, perimeter, and area, and directly packages them in a streaming manner through network transmission protocol, distributing them to the cloud-based basic information big data array center of the Ministry of Natural Resources, completing the unmanned data entry storage of the entire business chain.

[0070] It should be understood that, in the embodiments of the present invention, "B corresponding to A" means that B is associated with A, and B can be determined based on A. However, it should also be understood that determining B based on A does not mean that B is determined solely based on A; B can also be determined based on A and / or other information.

[0071] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0072] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0073] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not performed. The coupling or direct coupling or communication connection between the shown or discussed units may be an indirect coupling or communication connection through some interfaces, apparatuses, or units, or it may be an electrical, mechanical, or other form of connection.

[0074] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention, depending on actual needs.

[0075] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0076] From the above description of the embodiments, those skilled in the art will clearly understand that the present invention can be implemented in hardware, firmware, or a combination thereof. When implemented in software, the above-described functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of a computer program from one place to another. Storage media can be any available medium accessible to a computer. For example, but not limited to, computer-readable media can include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having the form of instructions or data structures and accessible to a computer. Furthermore, any connection can suitably be a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of the medium. As used in this invention, disk and disc include compressed optical discs (CDs), laser discs, optical discs, digital versatile discs (DVDs), floppy disks, and Blu-ray discs, wherein disks typically magnetically copy data, while discs optically copy data using lasers. The combinations described above should also be included within the scope of protection for computer-readable media.

[0077] In summary, the above description is merely a preferred embodiment of the technical solution of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent analysis method for land surveys, characterized in that, The method includes: Acquire multi-source high-resolution remote sensing images and extract features using a pre-defined super-resolution semantic segmentation network to output an initial super-resolution semantic segmentation map and the corresponding initial multi-class edge probability response map. The initial super-resolution semantic segmentation map and the initial multi-class edge probability response map are processed by a preset boundary refinement module to extract sub-pixel-level features of ground object edges and complete sub-pixel feature point offset calibration, outputting a high-precision edge probability response map and a two-dimensional sub-pixel offset vector field. Based on the high-precision edge probability response map and the two-dimensional sub-pixel offset vector field, adaptive topological vectorization of the patch boundary is performed, and a preset spatial geometric fine-tuning mechanism is used to perform high-precision fine-tuning of the vectorized patch boundary lines. The high-precision map boundary after fine-tuning is subjected to coordinate space transformation and topological verification, and the boundary data that passes the verification is encapsulated to complete the database construction and map assignment.

2. The intelligent analysis method for land survey based on deep learning according to claim 1, characterized in that, The acquisition of multi-source high-resolution remote sensing images and the extraction of features using a pre-defined super-resolution semantic segmentation network include: The remote sensing images are input into a backbone feature extraction network to obtain a multi-scale feature pyramid. An attention mechanism is introduced into the multi-scale feature pyramid to adaptively suppress background noise features and activate target salient features; Subpixel convolution is applied to the aggregated multi-scale feature map for feature reconstruction, and the initial super-resolution semantic segmentation map and the corresponding initial multi-class edge probability response map are output by combining the edge attention mechanism.

3. The intelligent analysis method for land survey based on deep learning according to claim 2, characterized in that, The process of applying sub-pixel convolution operations to the aggregated multi-scale feature maps for feature reconstruction, and combining this with an edge attention mechanism to output the initial super-resolution semantic segmentation map, includes: The number of channels in the feature map is expanded by using convolutional layers, and then a periodic shuffling mechanism is applied to rearrange the expanded feature map in spatial dimensions, transforming it into a high-resolution feature map. The rearranged high-resolution feature map is introduced into an auxiliary convolutional branch to predict a coarse semantic boundary distribution, and this boundary distribution is multiplied back into the backbone feature map as a spatial gating signal to generate the initial super-resolution semantic segmentation map with edge enhancement.

4. The intelligent analysis method for land survey based on deep learning according to claim 1, characterized in that, The process of using a preset boundary refinement module to process the initial super-resolution semantic segmentation map and the initial multi-class edge probability response map includes: The edge gradient differentiation enhancement operation is applied to amplify the real physical boundary response and generate edge feature maps with high signal-to-noise ratio. Based on the normalized probability distribution of the initial multi-class edge probability response map, the information entropy of each spatial grid position is calculated to determine the boundary ambiguity area, and a two-branch orthogonal decoupling network is deployed in the area to extract pure semantic category features and pure spatial boundary features with inner product approaching zero, respectively. The enhanced edge feature map and the pure spatial boundary feature are input into the offset regression prediction subnetwork, which outputs the two-dimensional sub-pixel offset vector field. This field is then used to perform geometric deformation and coordinate remapping on discrete pixel feature points to complete adaptive adsorption and calibration.

5. The intelligent analysis method for land survey based on deep learning according to claim 4, characterized in that, The applied edge gradient differentiation enhancement operation amplifies the true physical boundary response, generating a high signal-to-noise ratio edge feature map, including: Calculate the first-order feature gradient magnitude scalar using depthwise separable convolution operations. With the second-order Laplace response scalar And an enhanced boundary feature map is generated using a nonlinear differential enhancement mapping formula: in, Represents an enhanced boundary feature map; A high-resolution feature map representing the input; Represents the global scaling factor of the differential features, which is dynamically set based on the physical resolution of the image. This represents the hyperbolic tangent nonlinear activation function; Represents the weight of the first-order gradient response contribution; This represents the contribution weight of the second-order boundary zero-crossing localization.

6. The intelligent analysis method for land survey based on deep learning according to claim 4, characterized in that, After completing the coordinate remapping, the method also includes implementing edge continuity constraints, the steps of which are as follows: The non-maximum suppression algorithm is applied to the edge probability response map after offset calibration and resampling to extract the sub-pixel boundary skeleton line with a single pixel width; Tracing the connected nodes on the boundary skeleton line, constructing an edge continuity topological constraint loss function; The edge continuity penalty loss value is incorporated into the global network training to penalize ground feature boundaries with abnormal jagged edges or unnatural sharp angle abrupt changes during backpropagation. The formula for calculating the edge continuity topological constraint loss function, which involves tracking connected nodes on the boundary skeleton line, is as follows: in, This represents the edge continuity penalty loss value; This represents the total number of discrete points of the extracted boundary skeleton. , Representing the first The boundary skeleton point and its adjacent points. Local topological tangent direction vectors of each boundary skeleton point; Representative bestowed upon the first The adaptive continuous dynamic weights of each vertex, and ,in For the first The pixel coordinate vectors of discrete points of the boundary skeleton on the two-dimensional plane of the raster image. This represents the information entropy value of the uncertainty metric matrix corresponding to the pixel coordinate vector.

7. The intelligent analysis method for land survey based on deep learning according to claim 1, characterized in that, The adaptive topological vectorization of the patch boundary based on the high-precision edge probability response map and the two-dimensional sub-pixel offset vector field, and the high-precision fine-tuning of the vectorized patch boundary lines using a preset spatial geometric fine-tuning mechanism, includes: By combining the terrain relief measurement and local spatial context of digital elevation model data, a local offset statistical covariance matrix is ​​constructed, and the spatial error confidence of the predicted position of each vector polygon vertex is calculated. An contour line extraction algorithm combined with a two-dimensional sub-pixel offset vector field is used for initial topology vectorization, and prior geometric constraints of ground features are injected according to semantic categories to restore the continuity of boundary lines. Calculate the discrete average curvature of the polygon vertices to verify smoothness; An image gravity field model is constructed within the physical grayscale space of remote sensing images. Vector polygon nodes are then reprojected into this model to drive the nodes to perform physical gravity adsorption fine-tuning.

8. The intelligent analysis method for land survey based on deep learning according to claim 7, characterized in that, The step of constructing an image gravity field model within the physical grayscale space of remote sensing images, and reprojecting vector polygon nodes into this model to drive the nodes to perform physical gravity adsorption fine-tuning, includes: The basic grayscale map of the remote sensing image is processed using a Gaussian filter, and the square of the spatial gradient magnitude is calculated to construct a Gaussian image gravitational field potential energy model. Each spatial node of the vector polygon is projected into the model, and the vector node movement is driven by an iterative update fine-tuning formula based on displacement: in, and These represent the coordinate vectors of the nodes in the vector polygon space before and after the iteration; The two-dimensional spatial gradient vector representing the gravitational field of the Gaussian image at the current node position; This represents the resistance of the energy partial derivative gradient within the geometry used to maintain the topological shape of polygons without distortion. This represents the basic iteration step size coefficient, which is dynamically set based on the image spatial resolution. The spatial stiffness damping coefficient of a polygonal structure representing the balance of internal and external forces.

9. The intelligent analysis method for land survey based on deep learning according to claim 1, characterized in that, The process of performing coordinate space transformation and topological verification on the fine-tuned high-precision map boundaries, and encapsulating the verified boundary data to complete the database construction and mapping, includes: Using a parametric model or affine transformation matrix, the fine-tuned set of vector polygons is projected from the image pixel coordinate system to the national geodetic absolute coordinate system; The runtime topology tolerance check algorithm identifies and repairs illegally overlapping or self-intersecting nodes; The boundary error weighting technique is used to determine whether the accuracy meets the performance evaluation requirements for the detailed land delineation business. If it does not meet the requirements, the error distribution scaling method is used to drive the fine-tuning step for closed-loop iterative optimization. If the criteria are met, the results will be packaged into a geographic information data format with extended attributes and pushed to the national land and space basic information platform through a standard interface.

10. A deep learning-based intelligent analysis system for land surveys, characterized in that, The system includes: The feature extraction module is used to interface with the raw data stream of the aerial camera, read multi-source images and perform depth normalization preprocessing for geometric deformation and scale; drive high-performance GPU computing power to perform convolution calculations with spatial hole expansion rate to construct a global feature pyramid with macroscopic perception capability; and directly perform channel expansion and rearrangement shuffling operations of high-dimensional tensors in the video memory to achieve initial physical improvement and enhancement of feature map resolution, and output initial super-resolution semantic segmentation map and corresponding initial multi-class edge probability response map. The boundary refinement module is used to nonlinearly enhance the gradient signal of weak real edges by calculating and amplifying it through depthwise separable convolution; it uses Softmax probability to dynamically calculate the matrix information entropy to determine the blurred boundary region, and performs orthogonalization calculation to force the inner product of the semantic feature flow and the high-frequency physical feature flow of the boundary to approach zero in order to decouple and separate the two branches; it extracts the skeleton line of the image patch, and forces the topological deformation constraint based on the orientation angle change penalty mechanism to be applied to the deep network during backpropagation; and it continuously outputs a two-dimensional offset vector field based on the regression network model, and uses deformable convolution operators to drive the discrete pixel feature points in the grid to drift and align to the physical ground value boundary, thus completing the deep edge feature extraction and calibration. The geometric fine-tuning module is used to calculate the local covariance matrix and Mahalanobis distance confidence score distribution map in real time by combining terrain undulation and heterogeneity, and to help identify error clustering hotspots; it calls the contour extraction algorithm and combines mathematical tools such as collinear fitting of straight line segments to drive polygons to adaptively vectorize according to specific prior artificial geometric feature specifications; it calculates the discrete curvature of the polygon vertex array in real time, and intelligently triggers and implants Bézier curves or B-spline mathematical models for smooth transition processing in areas where curvature exceeds the limit; and it constructs the balance equation between the external Gaussian image stretching gravitational potential field and the internal topological defense holding force, and performs sub-pixel level lossless adsorption within the limited step size to complete the adaptive iterative fine-tuning of boundary points; The database creation and mapping module is used to enforce the mathematical projection transformation from the pixel image coordinate system to the national geodetic absolute coordinate system, and to remove all illegal topological intersections that may cause database entry failures for joint verification. After confirming that the boundary meets the high-precision assessment indicators, it automatically attaches geographical attribute business fields such as category, perimeter, and area, and directly streams and distributes them to the cloud-based basic information big data array center through network transmission protocols to complete accurate database creation and mapping.

Images