CN-122024033-A - Multi-scale planar residential land matching method based on graph self-encoder

Abstract

The invention discloses a multi-scale planar residential land matching method based on a graph self-encoder. Firstly, constructing a residential matching probability prediction model GAE-MLP of a fusion map self-encoder (GAE) and a multi-layer perceptron (MLP), and then dividing a matching flow into primary matching and secondary matching based on a separation idea, and respectively utilizing the GAE-MLP model in the two matching to realize 1:1, 1:N and M:1 and M:N types of matching. The planar residential land matching method provided by the invention overcomes the limitation that the traditional geometric method relies on manual setting of the matching threshold and the matching factor weight, simultaneously solves the problem of poor feature extraction capability of the traditional machine learning method, can more effectively cope with the problems of position deviation and shape homogeneity and is more accurate for matching complex relations of 1:N, M:N and the like compared with the traditional geometric method and the traditional machine learning method.

Inventors

• WANG ZHONGHUI
• LIANG CE
• CUI JIE
• LI JINGZHONG

Assignees

• 兰州交通大学

Dates

Publication Date

20260512

Application Date

20251020

Claims (4)

1. 1. A multi-scale planar residential land matching method based on a graph self-encoder comprises two parts of graph construction and feature definition, matching probability prediction model construction and residential land secondary matching method: the graph construction and feature definition steps are as follows: S1, constructing the graph structure by taking the vertexes on the natural contour of the residential land as nodes of the graph structure and the edges on the natural contour as edges of the graph structure. S2, comprehensively considering three dimensional information of local shape, size and direction of a residential area, defining 6 local features for each graph structure node P i , wherein the local features are respectively a corner alpha calculated by two sides of ①P i connection clockwise, an included angle gamma formed by connecting a ② building center point with a P i neighbor node, a distance Dis between an area S T ;④ of a graph surrounded by the adjacent three points of the ③ building center point and an area S C ;⑤P i neighbor node surrounded by connecting the P i neighbor node, and a length Len of connecting the ⑥ building center point with the P i . And S3, comprehensively considering two dimensional information of the image level size and the direction of the residential area, and defining 2 image level characteristics, namely ① the area of the building and the direction beta of the building S A ;②. S4, defining 2 neighborhood features from two dimensions of the neighborhood size and the neighborhood direction, wherein the two dimensions are ① neighborhood size features, namely the sum of the areas of the neighborhood objects, and ② neighborhood direction features, namely the direction unit vector sum of the neighborhood objects. The construction steps of the matching probability prediction model are as follows: S5, building a graph self-encoder (GAE). Respectively constructing an encoder and a decoder of the graph self-encoder by using two graph roll layers, and introducing a ReLU function between the graph roll layers to enhance the nonlinear expression capacity so as to realize node-level encoding of large and small scale data; And S6, building a characteristic enhancement module. And carrying out feature aggregation on node level codes of large and small scale data respectively through global maximization pooling to obtain picture level codes of the large and small scale data, and fusing the picture level codes with picture level features in a tensor splicing mode to realize feature enhancement. S7, constructing a classifier (MLP). An MLP classifier is built by adopting a three-layer full-connection network, a ReLU function and a Dropout layer are sequentially embedded between layers, and the tail end outputs matching probability through normalization of a Softmax function. And S8, integrating three parts of a graph self-encoder, a characteristic enhancement module and a classifier, and building an integrated model GAE-MLP. The method comprises the steps of measuring the reconstruction Loss of a graph from an encoder by means of mean square error, measuring the classification Loss of a classifier by means of cross entropy Loss, obtaining total Loss by adding the two losses, and carrying out iterative updating on all parameters of the model by means of Loss, so that joint optimization and construction of the integrated model are achieved. Secondary matching strategy of planar residential land: S9, data preprocessing. Data format conversion, coordinate system projection, topology checking and redundant point deletion are performed on the data to be matched. S10, primary matching. The method comprises the steps of ① searching candidate matching pairs, namely constructing an external buffer area aiming at elements S i in a small-scale data set S, adding large-scale elements intersected with the external buffer area into a set A 1 , and then combining the elements in the set A 1 to obtain candidate matching pairs of S i . ② And screening the candidate matching pairs, namely setting an area ratio threshold sigma to screen the candidate matching pairs, namely calculating the area ratio of the candidate matching pairs, and discarding the candidate matching pairs if the area ratio of the candidate matching pairs is smaller than the threshold sigma. ③ And ④, performing matching and post-processing, namely further comparing the matching probability values of the candidate matching pairs according to the conflict situation that the same surface element is judged to be matched in a plurality of candidate matching pairs, and selecting the candidate matching pair with the largest matching probability value as a final matching result. S11, secondary matching. The method comprises the following steps of ① candidate matching pair searching, namely, selecting a certain element s i in a small-scale building dataset, constructing an external buffer zone, searching large-scale elements intersected with the element s i , adding a large-scale candidate set A 2 , II, performing reverse superposition searching on the large-scale elements newly added into the B 2 , searching the small-scale elements intersected with the large-scale elements, adding a small-scale candidate set B, and III, repeating the steps I-III aiming at the small-scale elements newly added into the set B until no new elements are added. IV, combining elements in the candidate sets B and A 2 to obtain secondary matched candidate matching pairs, screening ② candidate matching pairs, namely screening the primary matched candidate matching pairs, ③ secondary matching judgment, namely constructing a graph structure on large and small scale data in the candidate matching pairs, calculating local features and graph level features, inputting a GAE-MLP model to carry out matching judgment, and ④ matching post-processing, namely carrying out matching post-processing on the primary matching pairs. S12, ending.
2. 2. The multi-scale planar residential land matching method based on the graph self-encoder according to claim 1, wherein in the steps S1 to S4, local, graph level and neighborhood characteristics of the residential land are comprehensively considered, and a multi-scale planar residential land characteristic system is constructed.
3. 3. The multi-scale planar residential land matching method based on the graph self-encoder as claimed in claim 1, wherein the GAE-MLP integrated matching probability prediction model is built in the steps S5 to S8.
4. 4. A multi-scale planar residential land matching method based on a graph self-encoder as claimed in claim 1 or claim 2, wherein the secondary matching strategy, candidate matching pair searching and screening strategy, matching post-processing strategy in steps S10 to S11.

Description

Multi-scale planar residential land matching method based on graph self-encoder Technical Field The invention relates to the field of map data matching, in particular to a multi-scale planar residential land matching method based on a graphic self-encoder (GAE). Background Spatial entity matching, namely, establishing association relations among entities with the same name among different scales through entity feature measurement, is an important precondition for realizing multi-source spatial information fusion, spatial object change detection and dynamic update. With the continuous acceleration of the urbanization process, buildings, which are one of the main contents of large and medium scale maps, are changing more and more frequently. In order to maintain the situation and accuracy of the geographic database, the changing building information in the map database needs to be updated in time, and building matching is a key technology for achieving the goal. At present, the matching method of the planar building is mainly divided into two major categories of a traditional method and a machine learning method. The traditional method mainly matches three aspects of topology, geometry and semantics. The matching method based on the geometry is used for identifying the homonymous entities in different data sets by measuring the geometrical similarity between the surface entities, is the most commonly used matching means at present, and semanteme is one of important attributes of space entities, but because different data source attribute information can have larger difference or attribute integrity is difficult to ensure, the semanteme information is often used as an auxiliary feature and is combined with the geometrical and topological features. Although the traditional method achieves a certain effect in matching, the traditional method generally relies on manual experience to set a matching judgment threshold and weights of all matching factors, and the matching result is greatly influenced by subjective factors. The machine learning method mainly uses the traditional BP neural network and CatBoost integrated learning network for matching. However, the existing machine learning method is mostly dependent on manually defined macroscopic structural features as a matching index, and it is difficult to fully utilize the graph structure to fuse multidimensional information of building local and graph-level features, so that feature expression capability is limited, and matching accuracy is affected. Compared with traditional machine learning, the deep learning of the graph shows remarkable advantages by virtue of stronger feature recognition and feature extraction capability. In the matching task of the building, the graphic neural network can excavate deeper features for the matching task by carrying out message transmission and feature aggregation on the graphic structure of the building, so that the matching accuracy and reliability are improved. Disclosure of Invention For this purpose, the graph neural network is introduced into a planar building matching study, and a multi-scale planar building matching method based on a graph self-encoder is provided, and the basic flow of the method is as follows: S1, constructing a graph structure of a residential area; S2, defining local characteristics, picture level characteristics and neighborhood characteristics of residents; s3, building and training an integrated matching judgment model GAE-MLP by fusing the graph self-encoder GAE, the feature enhancement module and the MLP classifier; and S4, dividing the matching process into two stages of primary matching and secondary matching based on a separation idea, and respectively applying a GAE-MLP model in the two matching to realize the matching of the planar building. According to the multi-scale planar residential land matching method, end-to-end matching judgment is achieved by constructing the integrated GAE-MLP model, the model comprehensively considers the characteristics of three different angles of a part, a picture level and a neighborhood, multi-dimensional information is extracted for matching judgment, matching precision is effectively improved, and the problem of position deviation and non-1:1 type matching can be effectively solved through the algorithm design selected by the candidate matching pairs in two times of matching. Drawings In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following description will briefly introduce the drawings required to be used in the embodiments or the prior art descriptions, it is obvious that the drawings in the following description are only schematic views of the present invention, and other drawings can be obtained according to the provided drawings without inventive effort for a person skilled in the art. Fig. 1 is a flowchart of a planar residential land matching method based on a graph