A Deep Learning-Based Intelligent Geographic Information Mapping Method

Abstract

This invention relates to the field of geospatial data processing and intelligent mapping technology based on deep learning, and discloses an intelligent mapping method for geographic information based on deep learning. By performing pixel completion and gridding on grayscale images, the problems of incomplete edge data and inconsistent block segmentation are solved. An adjacency matrix is ​​automatically constructed using the relationship between the grid centroid and boundaries, accurately representing spatial connectivity. The distance matrix obtained by combining global shortest paths provides spatial consistency constraints for feature mapping. Multivariate grayscale features are extracted from each block, and mapping parameters are automatically obtained through self-supervised closed-form solving, aligning images with spatial features without extensive annotation. Finally, high-quality continuous mapping maps are generated through regression fine-tuning and bilinear interpolation, improving automation and spatial continuity.

Description

Technical Field

[0001] This invention relates to the field of geospatial data processing and intelligent mapping technology based on deep learning, specifically to an intelligent mapping method for geographic information based on deep learning. Background Technology

[0002] With the rapid development of emerging applications such as Digital Earth and Smart Cities, the acquisition, analysis, and automated processing of geographic information data have become important research directions in the fields of surveying and mapping science and geographic information engineering. Traditional geographic information surveying methods mainly rely on manual field measurements, aerial photogrammetry, and remote sensing image interpretation. These traditional methods often suffer from problems such as long operation cycles, high human involvement, low data processing efficiency, and limited spatial resolution, making it difficult to meet the needs of rapid production and intelligent analysis of large-scale, high-precision geospatial data.

[0003] In recent years, with the popularization of remote sensing technology, UAV aerial surveying, and multi-source high-resolution imagery, geospatial data has exhibited characteristics such as large volume, diversity, and high dimensionality. Faced with such complex and massive geographic information, how to efficiently and automatically perform geospatial feature identification, spatial relationship reconstruction, and terrain and feature extraction has become a key challenge for technological development. Meanwhile, existing deep learning models mostly focus on general visual tasks such as image classification and segmentation, and still have significant limitations in core issues such as self-supervised modeling of geospatial structures, reconstruction of geographic adjacency relationships, and interpolation of continuous spatial mapping information. They also lack customized algorithmic frameworks for spatial topological consistency and physical geographic mapping needs. Currently, some AI-based image mapping methods have attempted to combine neural networks, machine learning, and geographic information processing, but these often rely on large amounts of labeled data, have uninterpretable model structures and parameters, and poor adaptability to spatial continuity and geographic topological constraints, resulting in numerous obstacles to practical promotion and implementation. Especially in scenarios such as large-scale automatic mapping, continuous high-precision modeling of complex terrain, and adaptive reconstruction of spatial structural relationships, existing methods still face significant challenges in terms of accuracy, automation, and interpretability.

[0004] Therefore, this case aims to propose a deep learning-based intelligent geographic information mapping method. By structuring the original geographic images according to a set grid, a self-supervised feature mapping relationship is established based on spatial topology and grayscale features. Then, through methods such as feature embedding, shortest path, and supervised regression, the method can automatically generate high-precision, continuous mapping data from the original images. Summary of the Invention

[0005] This invention provides a deep learning-based intelligent geographic information mapping method, which helps to solve the problems mentioned in the background art.

[0006] This invention provides the following technical solution: a deep learning-based intelligent geographic information mapping method, comprising:

[0007] S1. Acquire and input a grayscale image, and fill the image edges with necessary pixels according to the preset grid side length. Then, divide the image into several equilateral grid blocks according to the grid side length, and assign a unique identifier to each grid block.

[0008] S2. Determine adjacent grid blocks based on the geometric centroid position and shared boundary relationship of the grid blocks, establish adjacency relationships between adjacent grid blocks and generate an adjacency matrix;

[0009] S3. Based on the adjacency matrix, the graph theory shortest path algorithm is used to iteratively calculate all grid block pairs until convergence and the final shortest path distance matrix is ​​output.

[0010] S4. For each grid block, construct a basis vector containing a constant term, an average gray value, and its squared term, and map the basis vector into a self-supervised embedded feature using a linear mapping function;

[0011] S5. Construct a least squares problem based on the basis vector differences and corresponding distances of adjacent edges, and obtain the mapping function parameters through a closed-form analytical solution;

[0012] S6. Apply the obtained mapping function to the basis vectors of each grid block, and output the discrete mapping value of each grid block through linear calculation;

[0013] S7. Construct a regression model based on user-provided labeled samples and perform regression correction on discrete mapping values;

[0014] S8. Based on the discrete mapping values ​​of the grid block and its four neighboring blocks, calculate the mapping value of the point using the bilinear interpolation algorithm.

[0015] Optionally, the process of acquiring and inputting a grayscale image, filling the image edges with necessary pixels according to a preset grid side length, then dividing the image into several equilateral grid blocks according to the grid side length, and assigning a unique identifier to each grid block, specifically includes:

[0016] Acquire and input the grayscale image to be tested, and let it be... ;in, The number of pixel rows in the image; The number of pixel columns in the image; For the image Line number The grayscale value of the column pixels;

[0017] Set minimum grid side length pixels, and select integer grid side length. satisfy ;

[0018] Calculate the number of rows that need to be filled based on the actual number of rows and columns in the image. Number of columns :

[0019] ;

[0020] ;

[0021] in, This represents the number of rows that need to be padded below the image. This represents the number of columns that need to be padded to the right of the image. Modulo operation;

[0022] like Then copy the last line of the image. Fill the bottom row;

[0023] like Then copy the last column of the image. Fill the column to the right;

[0024] Update image size after filling , ;

[0025] Calculate the number of grid rows based on the updated dimensions. Column number and total number of grids :

[0026] , , ;

[0027] Divide the image into OK, The grid blocks of the column, for the first Line number Column block setting number ;

[0028] Build grid blocks pixel set :

[0029] ;

[0030] For each grid block Calculate their average gray values ​​respectively. and centroid coordinates :

[0031] , , .

[0032] Optionally, the step of determining adjacent mesh blocks based on their geometric centroid positions and shared boundary relationships, establishing adjacency relationships between adjacent mesh blocks, and generating an adjacency matrix specifically includes:

[0033] According to grid block number Row and column positions of the inverse calculation block:

[0034] , ;in, For block Row index; For block Column index;

[0035] when At that time, the decision block and If adjacent, then not adjacent; construct an adjacency matrix. ;in, These are the elements of the adjacency matrix.

[0036] Optionally, the step of using a graph-based shortest path algorithm based on the adjacency matrix to iteratively calculate all grid block pairs until convergence and outputting the final shortest path distance matrix specifically includes:

[0037] Let the maximum possible distance be ;

[0038] For all block pairs Construct the initial degree matrix Its elements are:

[0039] , ;in, For block and Initial estimated distance; diagonal elements Explicitly set to zero;

[0040] pass Algorithm Iteration Update:

[0041] ;in, For the first After round iteration, block arrive The current shortest path distance; For iteration rounds;

[0042] After the iteration is complete, output ;in, For the final block arrive The shortest path distance.

[0043] Optionally, the step of constructing a basis vector for each grid block, containing a constant term, an average gray value, and its square, and mapping the basis vector to a self-supervised embedding feature using a linear mapping function, specifically includes:

[0044] For each piece Set three-dimensional basis vectors Among them, the components For constant terms; For block Average gray level; For square terms;

[0045] Let the row vector of the mapping parameters to be determined be... ;in, This is the mapping bias, corresponding to the first dimension of the basis vectors; These are first-order coefficients, corresponding to the second dimension of the basis vectors; These are second-order coefficients, corresponding to the third dimension of the basis vectors;

[0046] Constructing a mapping function using inner product form: ;in, This is a polynomial mapping function, with inputs being basis vectors and parameters being... The output is block embedding features; For block Self-supervised embedding features.

[0047] Optionally, the step of constructing a least-squares problem based on the basis vector differences and corresponding distances of adjacent edges, and obtaining the mapping function parameters through a closed-form analytical solution, specifically includes:

[0048] Construct the set of adjacent edges Number of sides ;

[0049] For the set of Strip edge Calculate the difference row vector :

[0050] ;in, For the first The starting block number of the strip; For the first The termination block number of the edge; The edge number; For the first Feature difference row vectors of the edges; This represents the difference between the grayscale mean values; The sum of the grayscale mean;

[0051] All Stacked row by row to form a feature difference matrix ;

[0052] Set the target vector ;in, For the first The shortest distance between the two blocks at the two ends of the strip;

[0053] structure covariance matrix ;

[0054] like When using the least squares closed-form solution: ;

[0055] like Then let ;like ,make If the grayscale difference is insufficient, return to step S1 to recalculate; otherwise, terminate the process and prompt that the grayscale difference is insufficient and the parameters cannot be determined. Please adjust the resolution or input the parameters.

[0056] Optionally, applying the obtained mapping function to the basis vectors of each grid block and outputting the discrete mapping value of each grid block through linear calculation specifically includes:

[0057] The parameters are obtained by solving the closed-form equation. Apply the mapping function to all blocks:

[0058] ;in, For block Discrete mapping values.

[0059] Optionally, the step of constructing a regression model based on user-provided labeled samples to perform regression correction on discrete mapping values ​​specifically includes:

[0060] Collect labeled sample logs ;

[0061] Construct labeled regression matrices respectively and labeled survey vectors :

[0062] , ;in, For the first The number of each labeled block; For the first Discrete features of each labeled block; No. The actual survey value of each labeled block;

[0063] structure Regression covariance matrix ;

[0064] like Then the closed-form solution yields the regression coefficients. And correct the predicted values. ;

[0065] like If no valid annotations can be added, then maintain .

[0066] Optionally, the step of calculating the mapping value of a point using a bilinear interpolation algorithm based on the discrete mapping values ​​of the grid block and its four neighboring blocks specifically includes:

[0067] For any continuous points Calculate the row number of the corresponding grid. and column number :

[0068] , ;in, The row coordinates of the target space point; The column coordinates of the target space point;

[0069] Get the four neighboring block indices:

[0070] ;

[0071] ;

[0072] ;

[0073] ;

[0074] in, Number the top-left adjacent block; Number the lower left adjacent block; Number the adjacent block in the upper right corner; Number the lower right adjacent block;

[0075] like Then assign the value directly. ;

[0076] Otherwise, calculate the points. Normalized coordinates in the row direction within the block and points Normalized coordinates in the column direction within the block :

[0077] , ;

[0078] The following interpolation formula is used to obtain the mapping results of continuous space:

[0079] ;in, As the interpolation base value, take the fine-tuning value from step S7. ; This represents the final survey results for each point.

[0080] The present invention has the following beneficial effects:

[0081] 1. The original grayscale image is filled and uniformly divided according to a preset grid side length, ensuring that each grid block has a consistent structural basis for subsequent processing. By automatically padding edge pixels, lossless and uniform image segmentation is achieved, guaranteeing the consistency of data structure and feature statistics across all grid blocks. This not only improves the standardization of subsequent spatial topology and feature processing but also lays the foundation for subsequent batch automated processing. Compared to traditional methods that directly process the entire image or irregular segmentation, this method facilitates efficient parallel computation, adapts to different resolutions and image sizes, and reduces accuracy loss and anomalies caused by improper edge processing.

[0082] 2. By automatically determining and constructing a global adjacency matrix based on the geometric centroid and boundary relationships of grid blocks, a structured model of the spatial connectivity of all grid blocks is achieved. This approach breaks through methods that rely solely on pixel neighborhoods or fixed windows, explicitly introducing spatial information into the feature representation process, making the mapping results more spatially interpretable. Compared to most existing geographic information processing methods that only model with pixel-level features, this scheme enhances the consistency between spatial topology and geographic attributes, effectively supporting subsequent path analysis and geospatial modeling, and improving spatial reasoning capabilities and the level of surveying automation.

[0083] 3. Using graph theory's shortest path algorithm, the shortest path distance between all grid block pairs is globally calculated, providing spatial structure constraints. This design allows the scheme to quantify the connectivity distance between any two blocks in space, naturally embedding geospatial structure information into the feature modeling and optimization process. Compared with traditional methods that rely solely on Euclidean distance or ignore global structure, this invention achieves a deep fusion of geographic mapping features and real spatial structure, improving the ability to express spatial consistency and effectively solving the problems of existing methods having a single spatial structure expression and lacking topological constraints.

[0084] 4. For each grid block, multidimensional features including constant terms, average gray level, and their squares are extracted. High-order feature embedding is achieved through linear mapping, fully capturing the nonlinear variation patterns of landform distribution within the block. Compared to traditional methods that only use first-order gray-level mean or simple statistics, this approach enhances feature representation through polynomial expansion, enabling the model to better characterize complex surface reflectance properties and geographical distribution patterns, laying a solid foundation for subsequent self-supervised parameter training and feature alignment.

[0085] 5. This scheme utilizes the basis vector difference of spatially adjacent edges and the actual distance to construct a least-squares problem, and solves for the feature mapping parameters using a closed-form analytical method. This innovative approach aligns spatial information with image grayscale features, improving the model's spatial consistency and automatic modeling capabilities. Compared to traditional mapping methods that rely heavily on manual annotations or empirical features, this self-supervised mechanism reduces the dependence on training data, automates and scales feature learning, and is suitable for efficient processing in large-area, complex scenarios.

[0086] 6. After parameter training is completed, all grid block features are automatically mapped to discrete mapping values, achieving efficient batch generation of mapping features. This ensures strict consistency between discrete mapping features and the original spatial topology and grayscale features, improving the spatial coherence and consistency of the mapping data. Compared to existing solutions that rely on complex post-processing and manual correction, this invention improves automation and batch processing efficiency, facilitating subsequent large-scale data integration and analysis.

[0087] 7. Based on self-supervised learning, the scheme can incorporate a small amount of manual annotation, and fine-tune discrete mapping values ​​through regression to achieve personalized customization and improved accuracy. This design balances automation and high precision requirements, improving the feature matching accuracy of the final mapping even with limited annotation samples, effectively solving the problem of traditional deep learning methods' high dependence on large-scale annotation data. Compared with existing technologies that rely on end-to-end large-sample training, this method is more flexible and practically applicable, especially suitable for scenarios with limited actual mapping resources.

[0088] 8. By employing methods such as bilinear interpolation, discrete mapping features are smoothly transitioned into continuous mapping maps, achieving high-precision reconstruction and natural extension of spatial information. This improves the spatial continuity and visualization quality of the mapping maps and meets the high requirements for smoothness and realism in high-resolution geographic information applications. Compared with traditional simple filling and coarse interpolation methods, this method better maintains the physical consistency and spatial details of the mapping data, enhancing the practical value and scalability of mapping products. Attached Figure Description

[0089] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0090] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0091] Example, refer to Figure 1A deep learning-based intelligent geographic information mapping method includes:

[0092] S1. Acquire and input a grayscale image, and fill the image edges with necessary pixels according to the preset grid side length. Then, divide the image into several equilateral grid blocks according to the grid side length, and assign a unique identifier to each grid block.

[0093] S2. Determine adjacent grid blocks based on the geometric centroid position and shared boundary relationship of the grid blocks, establish adjacency relationships between adjacent grid blocks and generate an adjacency matrix;

[0094] S3. Based on the adjacency matrix, the graph theory shortest path algorithm is used to iteratively calculate all grid block pairs until convergence and the final shortest path distance matrix is ​​output.

[0095] S4. For each grid block, construct a basis vector containing a constant term, an average gray value, and its squared term, and map the basis vector into a self-supervised embedded feature using a linear mapping function;

[0096] S5. Construct a least squares problem based on the basis vector differences and corresponding distances of adjacent edges, and obtain the mapping function parameters through a closed-form analytical solution;

[0097] S6. Apply the obtained mapping function to the basis vectors of each grid block, and output the discrete mapping value of each grid block through linear calculation;

[0098] S7. Construct a regression model based on user-provided labeled samples and perform regression correction on discrete mapping values;

[0099] S8. Based on the discrete mapping values ​​of the grid block and its four neighboring blocks, calculate the mapping value of the point using the bilinear interpolation algorithm.

[0100] First, the original grayscale image is padded with pixels and uniformly divided according to a preset grid side length, solving the problems of data loss and incomplete segmentation caused by the inability to neatly segment image edges, thus providing a consistent basic unit for subsequent processing. Second, a spatial adjacency matrix is ​​automatically constructed based on the grid's geometric centroid and shared boundary relationships, replacing the limitations of traditional methods that rely solely on sliding windows or pixel-level nearest neighbors, clearly expressing the spatial connectivity between regions and facilitating subsequent spatial topology analysis. Then, the shortest path distance matrix is ​​calculated globally through iteration, overcoming the deficiency that using only Euclidean distance or local path search cannot reflect the true terrain connectivity, providing a basis for subsequent feature mapping. Reliable spatial consistency constraints are established. Subsequently, for each grid block, a multivariate feature model containing a constant term, average gray level, and its square term is constructed. Mapping parameters are automatically determined through self-supervised closed-form solving, aligning image gray-level information with spatial distance features without extensive manual annotation. This mapping function is then applied in batches to all grid blocks, rapidly generating discrete mapping values. Regression fine-tuning is performed using a small number of user-annotated samples, improving local accuracy while maintaining the overall model's automation and scalability. Finally, bilinear interpolation smoothly transitions the discrete results into continuous mapping, overcoming the drawbacks of traditional discrete products exhibiting spatial discontinuities or jagged distributions. This approach ensures data consistency through pixel completion and gridding; enhances spatial consistency through explicit adjacency and global path calculation; reduces dependence on labeled data through self-supervised closed-form solving; balances automation and high accuracy through regression fine-tuning; and generates high-quality continuous output through interpolation. Compared to existing technologies, this scheme improves automation, spatial consistency, and result continuity.

[0101] The process involves acquiring and inputting a grayscale image, filling the image edges with necessary pixels according to a preset grid side length, then dividing the image into several equilateral grid blocks according to the grid side length, and assigning a unique identifier to each grid block. Specifically, this includes:

[0102] Acquire and input the grayscale image to be tested, and let it be... ;in, The number of pixel rows in the image; The number of pixel columns in the image; For the image Line number The grayscale values ​​of the columns of pixels; the original grayscale image of the area to be tested is read in and represented in the form of a mathematical function, so that subsequent steps can directly and efficiently access and process each image data point by pixel index, ensuring the integrity and reliability of the basic data throughout the algorithm process;

[0103] Set minimum grid side length pixels, and select integer grid side length. satisfy Set the minimum allowable size and actual partitioning side length of the grid blocks so that the image can be neatly divided according to the actual size, laying the foundation for feature extraction and spatial structure establishment of each block, while providing parameter control for accuracy and computational efficiency.

[0104] Calculate the number of rows that need to be filled based on the actual number of rows and columns in the image. Number of columns :

[0105] ;

[0106] ;

[0107] in, This represents the number of rows that need to be padded below the image. This represents the number of columns that need to be padded to the right of the image. Modular operation; by filling the image boundary, it ensures that the image can be completely and uniformly divided into grid blocks of the same size, avoiding incomplete block division at the edge, so that all subsequent grid blocks have consistency in feature extraction and spatial relationship calculation, improving the stability of the algorithm and the accuracy of the mapping results;

[0108] like Then copy the last line of the image. Fill the bottom row;

[0109] like Then copy the last column of the image. Fill the column to the right;

[0110] Update image size after filling , ;

[0111] Calculate the number of grid rows based on the updated dimensions. Column number and total number of grids :

[0112] , , Based on the size of the filled image, the total number of rows, columns, and blocks after grid division is automatically calculated; this provides a precise index range and logical support for the subsequent numbering of each block, cyclic processing, and adjacency determination.

[0113] Divide the image into OK, The grid blocks of the column, for the first Line number Column block setting number ;

[0114] Build grid blocks pixel set :

[0115] ;

[0116] By defining block numbers and corresponding pixel sets, a unique mapping from two-dimensional grids to one-dimensional indexes is achieved, enabling the algorithm to efficiently locate each block and its pixels, providing a data foundation for subsequent feature statistics and spatial adjacency.

[0117] For each grid block Calculate their average gray values ​​respectively. and centroid coordinates :

[0118] , , ;

[0119] By statistically analyzing the average gray level and pixel coordinate centroid of each grid block, high-dimensional and highly discriminative feature inputs are provided for subsequent self-supervised modeling, reflecting the reflection intensity and spatial location characteristics of objects within the block.

[0120] The grayscale image under test is read in at the pixel level and represented using functions. Combined with an adjustable minimum grid edge length setting and symmetrical pixel copying and filling of image edges, this ensures that the image rows and columns are divisible by a preset grid, avoiding incomplete or abnormal segmentation caused by size mismatches in traditional images. After completion, the image row and column sizes are automatically updated, and the number of grid rows, columns, and the total number of grids are calculated, providing an accurate index range for subsequent batch numbering and iterative processing. Next, the image is neatly divided into standard grid blocks, and each grid block is assigned a unique identifier, achieving efficient mapping from two-dimensional grids to one-dimensional indexes. Finally, the average grayscale value and geometric centroid coordinates are calculated for each grid block, providing highly discriminative initial input for subsequent self-supervised feature mapping based on grayscale and spatial topology. Symmetrical pixel filling solves the problem of inconsistent edge segmentation; automatic size and number updates simplify data management; unified gridding ensures consistency in feature statistics; and accurate extraction of centroids and grayscale values ​​lays a standardized foundation for self-supervised mapping and spatial relationship calculations. Compared to existing methods that arbitrarily segment or process data as a whole, this method has significant advantages in terms of data integrity and subsequent processing efficiency.

[0121] The process of determining adjacent mesh blocks based on their geometric centroid positions and shared boundary relationships, establishing adjacency relationships between adjacent mesh blocks, and generating an adjacency matrix specifically includes:

[0122] According to grid block number Row and column positions of the inverse calculation block:

[0123] , ;in, For block Row index; For block The column index; by restoring its row and column index through the block number, the accurate mapping from one-dimensional block index to two-dimensional grid coordinates is realized, providing a data foundation for spatial adjacency judgment and grid structured management;

[0124] when At that time, the decision block and If adjacent, then not adjacent; construct an adjacency matrix. ;in, These are elements of the adjacency matrix; by determining the direct adjacency relationship between two blocks in the row and column directions, a global adjacency matrix is ​​generated, accurately describing the spatial connectivity between each block and its surrounding blocks, providing a mathematical basis for geospatial path and topological consistency calculations.

[0125] By inversely calculating the row and column positions of grid blocks using their numbers, and combining this with geometric centroid positions and shared boundary determination rules, adjacent grid blocks are automatically identified and an adjacency matrix is ​​generated. Specifically, the system first maps each grid block number back to its corresponding row and column coordinates, then determines whether two blocks are adjacent in the row direction and aligned in the column direction, thus recording this relationship in the adjacency matrix. Simultaneously, index positions are cached to reduce the complexity of subsequent queries and improve computational efficiency. Through precise row and column mapping and adjacency determination, a refined spatial topology model is established, overcoming the limitations of traditional methods that rely solely on simple pixel neighborhood determination, and achieving global management of grid-level spatial connectivity. Furthermore, the caching mechanism reduces the time overhead of adjacency calculations, providing high-quality spatial structure input for global paths and self-supervised mappings. Compared with existing technologies, this method significantly improves both the accuracy of spatial relationship representation and computational efficiency.

[0126] The process, based on the adjacency matrix, uses a graph-theoretic shortest path algorithm to iteratively calculate all grid block pairs until convergence, outputting the final shortest path distance matrix. Specifically, this includes:

[0127] Let the maximum possible distance be Set an upper limit for the initial distance between non-adjacent blocks to ensure that the algorithm can distinguish between non-connectable block pairs and connectable block pairs during initialization, thus preventing misjudgment during path search.

[0128] For all block pairs Construct the initial degree matrix Its elements are:

[0129] , ;in, For block and Initial estimated distance; diagonal elements Explicitly set to zero; initialize the path weights between all block pairs according to the spatial Euclidean distance to provide a reasonable starting point for the shortest path algorithm, so that subsequent iterations can converge to the true spatial shortest distance;

[0130] pass Algorithm Iteration Update:

[0131] ;in, For the first After round iteration, block arrive The current shortest path distance; The iteration cycle is determined by dynamic programming, which, at each step, considers the possible iterations that have already passed the first iteration. The path of each block is compared with the original path to obtain the global shortest path between any pair of blocks, providing a precise quantitative basis for geospatial consistency modeling.

[0132] After the iteration is complete, output ;in, For the final block arrive The shortest path distance is calculated; the final spatial shortest path distance matrix between all blocks is output, providing a supervision signal for subsequent self-supervised parameter solving, and realizing the mapping and alignment of geospatial structure and image grayscale features.

[0133] Based on the constructed spatial adjacency matrix, a graph-theoretic shortest path iterative algorithm is applied to calculate global path distances. The system first assigns reasonable initial distance estimates to adjacent block pairs and assigns large weights to indirect adjacent blocks to distinguish between reachable and inaccessible regions, avoiding misjudgments in the early stages of iteration. Then, in each iteration, paths that may pass through any grid block are compared and updated until all path distances converge to the minimum, ultimately outputting the global shortest spatial distance matrix between all grid block pairs. By distinguishing between the initial weights of direct and indirect adjacencies, the accuracy of path search is ensured; dynamic iterative updates achieve precise quantification of connectivity in complex terrain, overcoming the biases caused by existing methods relying solely on straight-line distances or local window searches; the resulting distance matrix provides realistic spatial consistency constraints for subsequent feature mapping and self-supervised parameter solving, improving the model's adaptability to complex scenes and the robustness of the mapping results.

[0134] The process of constructing a basis vector for each grid block, containing a constant term, an average gray value, and its square, and then mapping the basis vector to self-supervised embedded features using a linear mapping function, specifically includes:

[0135] For each piece Set three-dimensional basis vectors Among them, the components For constant terms; For block Average gray level; The term is squared; each gray-level feature is expanded into a high-dimensional feature containing constant terms, first-order terms, and second-order terms, providing stronger expressive power for self-supervised mapping in order to capture the nonlinear characteristics of ground feature distribution;

[0136] Let the row vector of the mapping parameters to be determined be... ;in, This is the mapping bias, corresponding to the first dimension of the basis vectors; These are first-order coefficients, corresponding to the second dimension of the basis vectors; The coefficients are second-order coefficients, corresponding to the third dimension of the basis vector; the weight parameters of the linear mapping are defined, providing an adjustment space for the adaptive training of the polynomial basis function in the mapping from actual features to surveying features, and determining the geometric structure of the final embedding space;

[0137] Constructing a mapping function using inner product form: ;in, This is a polynomial mapping function, with inputs being basis vectors and parameters being... The output is block embedding features; For block The self-supervised embedding feature; this function realizes the adaptive mapping from block grayscale features to mapping embedding features, and is the core tool for aligning image information with spatial distance in the scheme, providing basic values ​​for subsequent feature optimization and fine-tuning.

[0138] For each grid block, a multivariate feature set consisting of a constant term, an average gray level term, and its squared term is constructed. This set is then linearly combined with mapping parameters to generate self-supervised embedding features. This method enhances the expressive power for nonlinear variations in gray level distribution by introducing an additional gray level squared term on top of the traditional average gray level. The mapping parameters adaptively assign weights to different features, achieving fine-grained capture of ground feature reflectivity and texture details. In this process, the system can acquire the embedding features of all grid blocks at once, eliminating the need for repeated calculations for each region. The multivariate feature extension improves sensitivity to complex surface reflectivity variations, making the embedding features more discriminative in expressing differences in ground feature types. The adaptive weight mechanism optimizes for gray level feature variations in different regions, providing richer and more accurate input information for subsequent self-supervised parameter solving. Compared to using only first-order gray levels or simple statistical methods, this method improves both feature expressive power and model adaptability, laying a solid foundation for the accuracy and interpretability of surveying results.

[0139] The process of constructing a least-squares problem based on the basis vector differences and corresponding distances of adjacent edges, and obtaining the mapping function parameters through a closed-form analytical solution, specifically includes:

[0140] Construct the set of adjacent edges Number of sides Collect all directly adjacent block pairs to provide the most realistic spatial constraint data set for subsequent least squares parameter solving, so that self-supervised training depends only on geographically connected sample pairs.

[0141] For the set of Strip edge Calculate the difference row vector :

[0142] ;in, For the first The starting block number of the strip; For the first The termination block number of the edge; The edge number; For the first Feature difference row vectors of the edges; This represents the difference between the grayscale mean values; The grayscale mean is calculated; the feature difference between adjacent block pairs is calculated to form the input features of the least squares equation system, so that the mapping parameter learning can explicitly align the relationship between spatial distance and feature distance;

[0143] All Stacked row by row to form a feature difference matrix ;

[0144] Set the target vector ;in, For the first Find the shortest distance between the two ends of the edge; collect the actual spatial distance between all adjacent block pairs, which is the target output for parameter training, ensuring that the training converges to the optimal solution with geospatial consistency;

[0145] structure covariance matrix ;

[0146] like When using the least squares closed-form solution: ;

[0147] like Then let ;like ,make If the grayscale difference is insufficient, return to step S1 to recalculate; otherwise, terminate this process and prompt that the grayscale difference is insufficient and the parameters cannot be determined. Please adjust the resolution or input the parameters.

[0148] The optimal mapping parameters are uniquely determined by the least squares closed-form solution, achieving optimal alignment between the mapping space and the geographic space. If matrix degradation occurs, the grid size is automatically adjusted to ensure parameter solvability.

[0149] The system collects the feature differences of all adjacent grid block pairs and their corresponding spatial distances, organizes them into least-squares constraint equations, and obtains the mapping parameters through closed-form analytical solutions. The system first calculates the difference in grayscale features of each pair of adjacent blocks and pairs them with their true shortest spatial distances. Then, it constructs explicit geometric-grayscale correspondence constraints globally, determining the optimal parameters in one go through closed-form solutions, without requiring multiple rounds of iterative optimization or relying on large-scale labeled samples. In cases of unstable parameters or insufficient data, the system can automatically adjust the grid resolution or prompt for re-collection to ensure the stability of the solution process. The least-squares closed-form solution achieves parameter uniqueness and interpretability, avoiding the risks of overfitting and training non-convergence that may occur in iterative optimization. High-quality mapping parameters can be quickly obtained relying only on a small amount of adjacent block information, reducing computational complexity and data requirements. Compared with existing methods that rely on deep networks or large amounts of labeled samples for parameter learning, it has advantages in efficiency and traceability, facilitating flexible deployment in multiple scenarios and at multiple resolutions.

[0150] The step of applying the obtained mapping function to the basis vectors of each grid block and outputting the discrete mapping value of each grid block through linear calculation specifically includes:

[0151] The parameters are obtained by solving the closed-form equation. Apply the mapping function to all blocks:

[0152] ;in, For block The discrete mapping values ​​are obtained; the grayscale features of all blocks are batch mapped to discrete mapping features, and a consistent embedding space is constructed for the entire mapping area, laying a solid foundation for subsequent interpolation and continuous mapping generation.

[0153] This design applies the mapping function obtained from closed-loop solving to all grid block basis vectors in batches, generating discrete mapping values ​​for each grid block through a single linear mapping. This approach unifies grayscale and spatial topology information under the same mapping function, eliminating the need for repeated training or manual parameter tuning for different regions, thus improving overall computational efficiency while maintaining mapping consistency. The mapping process can be completed in a parallel computing environment, avoiding the overhead of multiple loading and iterations required in traditional pixel-level deep network inference. By applying the mapping function in batches, the generation time of mapping values ​​is shortened and computational resource consumption is reduced. Globally consistent mapping parameters ensure that discrete mapping values ​​follow the same rules across different surface types, improving product comparability and stability. Compared to existing solutions that require separate processing of different local regions or rely on multiple network inferences, this design offers significant advantages in efficiency and consistency, making it suitable for large-scale or edge computing deployments.

[0154] The process of constructing a regression model based on user-provided labeled samples to perform regression correction on discrete mapping values ​​specifically includes:

[0155] Collect labeled sample logs Set the minimum number of labeled samples required for fine-tuning to ensure the invertibility of the subsequent regression matrix and achieve the improvement in accuracy through supervised fine-tuning.

[0156] Construct labeled regression matrices respectively and labeled survey vectors :

[0157] , ;in, For the first The number of each labeled block; For the first Discrete features of each labeled block; No. The actual mapping values ​​of each labeled block; the manually labeled values ​​and block features are assembled into a standard linear regression problem, providing a mathematical basis for supervised correction of discrete features and improving the accuracy of ground feature matching in the mapping results;

[0158] structure Regression covariance matrix ;

[0159] like Then the closed-form solution yields the regression coefficients. And correct the predicted values. ;

[0160] like If no valid annotations can be added, then maintain ;

[0161] By using least squares regression correction parameters, the residuals between discrete features and actual mapping annotations are corrected, improving the accuracy and interpretability of the final mapping; branching ensures that the process remains stable even if the regression matrix degenerates.

[0162] By collecting real mapping annotation samples provided by users, a regression model is constructed to perform supervised fine-tuning of discrete mapping values. When the number of annotation samples reaches a minimum threshold, the system automatically calculates and applies correction parameters using a closed-loop regression solution. If the number of annotation samples is insufficient, the original mapping parameters are maintained to avoid over-adjustment due to insufficient samples. This mechanism allows for the flexible introduction of a small number of manual annotations in highly automated mapping processes, improving local accuracy while ensuring the robustness and stability of the overall model. A small number of high-quality annotation samples improve the accuracy and feature matching of discrete mapping values, meeting the personalized accuracy requirements of different application scenarios. The closed-loop regression solution for correction parameters offers both fast computation speed and ease of interpretation. Compared with deep learning methods that rely on large-scale annotation data or multiple rounds of iterative fine-tuning, it has advantages in engineering implementation costs and deployment efficiency, facilitating rapid deployment and iterative optimization.

[0163] The step of calculating the mapping value of a point using a bilinear interpolation algorithm based on the discrete mapping values ​​of the grid block it is located in and its four neighboring blocks specifically includes:

[0164] For any continuous points Calculate the row number of the corresponding grid. and column number :

[0165] , ;in, The row coordinates of the target space point; The column coordinates of the target space point;

[0166] Get the four neighboring block indices:

[0167] ;

[0168] ;

[0169] ;

[0170] ;

[0171] in, Number the top-left adjacent block; Number the lower left adjacent block; Number the adjacent block in the upper right corner; Number the lower right adjacent block;

[0172] like Then assign the value directly. ;

[0173] Otherwise, calculate the points. Normalized coordinates in the row direction within the block and points Normalized coordinates in the column direction within the block :

[0174] , ;

[0175] It enables rapid location of the grid and local normalized coordinates of any continuous spatial point, ensuring that the weights of the interpolation calculation accurately reflect the physical distance between the spatial point and its neighboring blocks, thus achieving accuracy and smoothness in spatial continuity.

[0176] The following interpolation formula is used to obtain the mapping results of continuous space:

[0177] ;in, As the interpolation base value, take the fine-tuning value from step S7. ; The final mapping result for each point is obtained; a linear weighted combination of four neighboring block features is adopted, and a natural transition from discrete mapping values ​​to continuous mapping is achieved based on local spatial coordinates, thereby improving the continuity and smoothness of the mapping results in the actual geographic scene and meeting the high-precision application requirements of geographic information products.

[0178] This method converts discrete grid-based survey values ​​into continuous survey values ​​using bilinear interpolation. First, the system quickly determines the grid block to which the target point belongs and its four adjacent grid blocks based on the target point's spatial coordinates, ensuring that the interpolation weights accurately reflect the physical distance between the spatial point and the center of each block. If the target point is located at the center of a block, the block's survey value is used directly; otherwise, the weights are calculated based on the target point's normalized row and column position within the block, achieving a smooth transition through linear weighting. This method ensures numerical continuity at the boundaries while preserving local differences. By using spatially based bilinear interpolation, it solves the problem of jagged or discontinuous distributions caused by traditional nearest-neighbor or simple average interpolation. Through weighted fusion of the survey values ​​from four adjacent blocks, it achieves a natural transition and high-precision fidelity at grid boundaries. Compared to interpolation based on a single interpolation kernel or complex frequency domain interpolation, it balances computational efficiency with smoothness and local accuracy, providing reliable support for high-quality visualization and subsequent spatial analysis of continuous survey values.

[0179] It should be noted that, in this document, relational terms such as "first" and "second" are used only 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 process, method, article, or apparatus.

[0180] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent geographic information mapping method, characterized in that, The method comprises the following steps: S1, collecting and inputting a gray image, and filling necessary pixels to the image edge according to a preset grid side length, then dividing the image into a plurality of equal-side grid blocks according to the grid side length, and assigning a unique identification to each grid block; S2, determining adjacent grid blocks according to the geometric barycentric position and shared boundary relationship of the grid blocks, establishing an adjacency relationship between the adjacent grid blocks, and generating an adjacency matrix; S3, based on the adjacency matrix, iteratively calculating all grid block pairs by using a graph theory shortest path algorithm until convergence, and outputting a final shortest path distance matrix; S4, constructing a basis vector containing a constant term, an average gray value and a square term of the average gray value for each grid block, and mapping the basis vector to a self-supervised embedding feature through a linear mapping function; S5, constructing a least squares problem according to the basis vector difference of the adjacent edges and the corresponding distance, and obtaining the mapping function parameters through a closed-form analytical solution; S6, applying the obtained mapping function to the basis vectors of the grid blocks, and outputting the discrete mapping values of each grid block through linear calculation; S7, constructing a regression model based on the labeled samples provided by a user, and performing regression correction on the discrete mapping values; S8, for any continuous point, calculating the mapping value of the continuous point by using a bilinear interpolation algorithm according to the discrete mapping values of the grid block where the continuous point is located and its four adjacent blocks. 2.The method of claim 1, wherein, The collecting and inputting a gray image, and filling necessary pixels to the image edge according to a preset grid side length, then dividing the image into a plurality of equal-side grid blocks according to the grid side length, and assigning a unique identification to each grid block, specifically comprises: Acquire and input the grayscale image to be tested, and let it be... ;in, The number of pixel rows in the image; The number of pixel columns in the image; For the image Line number The grayscale value of the column pixels; Setting a minimum grid edge length pixels, and selecting an integer grid edge length satisfies ; According to the actual number of rows and columns of the image, the number of rows to be filled in is calculated and the number of columns : ； ； wherein, is the number of rows that need to be padded below the image; is the number of columns that need to be padded to the right of the image; is a modulo operation; If then copy the last row of the image to fill the bottom of the row; If then copy the last column of the image perform right padding on the column; The image size after filling is updated as , ; According to the updated size, the number of grid rows , columns and the total number of grids are calculated. ， ， ； Divide the image into row, column grid blocks, set the number of the block in the first row and the first column ; Constructing a grid tile of pixels : ； For each grid block the average gray value and the barycentric coordinates are calculated separately ， ， 。 3.The method of claim 2, wherein, The determining adjacent grid blocks according to the geometric barycentric position and shared boundary relationship of the grid blocks, establishing an adjacency relationship between the adjacent grid blocks, and generating an adjacency matrix, specifically comprises: According to the grid block number Calculate the row and column positions of the block , ; wherein, is a row index of the block ; is a column index of the block ; When the decision block is adjacent to , otherwise not adjacent, constructs the adjacency matrix ; wherein is the adjacency matrix element. 4.The method of claim 3, wherein, The based on the adjacency matrix, iteratively calculating all grid block pairs by using a graph theory shortest path algorithm until convergence, and outputting a final shortest path distance matrix, specifically comprises: Let the maximum possible distance be ; for all pairs of blocks , an initial degree matrix whose elements are: , ; wherein, is a block with an initial estimate of the distance; the diagonal elements are explicitly zeroed; By Algorithmic iterative update: ; wherein, is the after the i-th iteration, the block to is the current shortest path distance from the source node to the node is the iteration round. After the iteration is completed, output ; wherein, is the shortest path distance from the final block to the root node. 5.The intelligent geographic information mapping method based on deep learning according to claim 4, characterized in that, The constructing a basis vector containing a constant term, an average gray value and a square term of the average gray value for each grid block, and mapping the basis vector to a self-supervised embedding feature through a linear mapping function, specifically comprises: For each block , a three-dimensional base vector is set; wherein the component is a constant term; is the average gray scale of the block ; and is a square term; Let the mapping parameter row vector to be solved be ; wherein, is a mapping bias, corresponding to the first dimension of the basis vector; is a first-order coefficient, corresponding to the second dimension of the basis vector; is a second-order coefficient, corresponding to the third dimension of the basis vector; The mapping function is constructed in the form of inner product: ; wherein, is a polynomial mapping function, the input is a basis vector, and the parameter is , the output is a block embedding feature; is a self-supervised embedding feature of the block . 6.The method of claim 5, wherein, The constructing a least squares problem according to the basis vector difference of the adjacent edges and the corresponding distance, and obtaining the mapping function parameters through a closed-form analytical solution, specifically comprises: Constructing the set of adjacent edges , number of edges ; for each row in the set bar edge , compute the difference row vector : ; wherein, is the start block number of the th edge; is the end block number of the th edge; is the edge number; is the feature difference vector of the th edge; is the gray level mean difference; is the gray level mean sum; stacking all stacking rows into feature difference matrix ; Setting target vector ; wherein, is the shortest distance between the two end blocks of the th side. construction covariance matrix ; If the least square closed-form solution is used: ; If , let ; if , let , return to step S1 to recalculate; otherwise, terminate the process and prompt that the gray difference is insufficient and the parameters cannot be determined, please adjust the resolution or input.

7. The geographic information intelligent mapping method based on deep learning according to claim 6, characterized in that, The applying the obtained mapping function to the basis vectors of the grid blocks, and outputting the discrete mapping values of each grid block through linear calculation, specifically comprises: Solve for parameters in closed form Apply the mapping function to all blocks: ; wherein, is a discrete mapping value of a block . 8.The method of claim 7, wherein, The constructing a regression model based on the labeled samples provided by a user, and performing regression correction on the discrete mapping values, specifically comprises: Number of collected labeled samples ; constructing a labeling regression matrix and a labeling mapping vector respectively , ; wherein, is the number of the th annotation block; is the discrete feature of the th annotation block; the true survey value of the th annotation block; constructing regression covariance matrix ; If then the closed-form solution for the regression coefficients is obtained and the prediction is revised ; If , then keep in case the valid annotation cannot be replenished. 9.The method of claim 8, wherein, The for any continuous point, calculating the mapping value of the continuous point by using a bilinear interpolation algorithm according to the discrete mapping values of the grid block where the continuous point is located and its four adjacent blocks, specifically comprises: For any consecutive point , the row number and column number of the grid to which it belongs are calculated: , ; wherein, is a row coordinate of the target spatial point; is a column coordinate of the target spatial point; Obtaining four adjacent block indexes: ； ； ； ； wherein, is a left upper neighboring block number; is a left lower neighboring block number; is a right upper neighboring block number; is a right lower neighboring block number; If , then directly assign ; Otherwise, compute point In-row normalized coordinates within the block And point In-column normalized coordinates within the block : ， ； The interpolation formula is as follows: The interpolation formula is as follows: ; wherein, is an interpolated base value, taking the fine value of step S7 ; is the final mapping result for each point sought.

Images