Hierarchical Mining Method for Land Use Structure Patterns Based on Graph Convolutional Neural Networks

Abstract

The present invention relates to a hierarchical mining method for land use structure patterns based on graph convolutional neural network, which includes: acquiring land use data, constructing a graph structure, generating a label input model, learning the label input model by a graph convolutional neural network model to generate a graph embedding, using a spatially constrained multi-way clustering method to perform partitioning on the graph embedding to obtain a hierarchical partition structure from a partition to each level sub-partition, constructing partition level graph elements of each district with the frequency characteristics of land use types for each year of land use data, respectively, and reflecting the spatiotemporal changes of the land use structure of a specific district according to the changes of the district level graph elements in different years. The present invention can partition according to the difference of land use spatial structure, and can effectively mine hierarchical land use structure patterns and their dynamic characteristics.

Description

Technical Field

[0001] The present invention relates to a hierarchical mining method for land use structure patterns based on graph convolutional neural networks and belongs to the field of data processing technology.

Background Art

[0002] Land use directly reflects the interaction and mutual influence between humans and nature, and the land use structure is the proportional relationship or composition situation of various land use types within a certain range. The land use pattern is the characteristic of land use elements within a region, including the characteristics, composition, spatial distribution, spatial relationship, etc. of the land use structure, and reflects the development of important processes such as social economy. The evolution of the land use structure pattern reflects the development of important processes such as social economy, and mining these structure patterns helps to promote the harmonious development of the relationship between people and land in the region and improve the national land space management ability.

[0003] In the study of conventional land use patterns, usually, various spatial pattern indicators, such as patch density, fractal dimension, patch diversity, and various spatial analysis methods, such as spatial autocorrelation statistics, neighborhood analysis, etc. are used to analyze the land use structure and spatial structure. References can be made to papers such as "Multi-scale Mining Method for Significant Spatial Co-location Patterns" (author: He Zhanjun, etc., Acta Geodaetica et Cartographica Sinica, No. 11, 2016), "Research on Mining Algorithm for Spatial Co-location Patterns" (Yang ▲Chiong▼, Doctoral Dissertation of Kunming University of Science and Technology, 2012), etc. In the above methods, in many cases, the spatial relationship is determined with a certain distance threshold for different geographical regions, and mining-related rules are derived. However, due to spatial heterogeneity, a single fixed geographical proximity threshold is not very suitable for performing local regional pattern mining, and thus its application range is limited.

[0004] In addition, when quantitatively describing the spatial structure at present, the spatial phenomenon is often described using overall statistical values. However, this can only reflect the overall situation of land use and ignores the relationship between the spatial structure layout and geographical objects. That is to say, these indicators only focus on morphological geometric attributes and do not consider the interaction between land use patches, thus underestimating the importance of the land use spatial structure and limiting the ability to perceive changes in land use patterns. Therefore, with the conventional method, when mining regional patterns, different regions with different land use structures cannot be automatically identified, and only the structure of the conventional zoning can be displayed. As a result, the mining display of land use structures at different levels and scales is insufficient, making it difficult for people to accurately and comprehensively recognize the characteristics and dynamic changes of the regional land use structure pattern.

Summary of the Invention

[0005] The technical problem to be solved by the present invention is as follows. To provide a hierarchical mining method for the regional land use structure pattern that can recognize the characteristics and dynamic changes of the regional land use structure pattern, and can assist in the recognition of the spatio-temporal changes of the land use structure.

[0006] To solve the above technical problem, the technical solution proposed by the present invention is as follows. A hierarchical mining method for the land use structure pattern based on a graph convolutional neural network, including the following steps. Step S1: Obtain the land use data of a specific region for N years, N≥2, and integrate the land use data for N years. Step S2: For the land use data after integration in Step S1, construct a graph structure with the center of gravity of the patch as the node and the connection line between adjacent centers of gravity as the edge, and record the land use type, area, and shape of the corresponding patch of each node in the graph structure. Here, when the shortest distance between two patch contours is smaller than the preset proximity threshold, the centers of gravity of the two patches are defined as adjacent centers of gravity. Step S3: Labels are assigned to each node and used to mark the land use type, the nearest dominant type, and the second-nearest dominant type of the node, and a label input model is obtained. The nearest dominant type is the land use type with the largest adjacent area among the first-order neighbors of the node, and the second-nearest dominant type is the land use type with the second-largest adjacent area among the first-order neighbors of the node. Step S4: A graph convolutional neural network model is constructed, and learning is performed on the label input model to generate a graph embedding. Step S5: The graph embedding is partitioned using a spatial constraint multi-class clustering method to generate partitions with different structural features. By iteration, each sub-partition is partitioned again, and so on, until the area of the smallest sub-partition is less than a preset area threshold, at which point the iteration stops. Finally, a hierarchical partition structure from the partition to each level of sub-partitions is obtained. Step S6: All partitions and each level of sub-partitions obtained in Step S5 are collectively referred to as regions. For the land use data of each year, region-level graph elements of each region are constructed according to the frequency characteristics of the land use type, and the spatio-temporal changes of the land use structure of the region are reflected by the changes in the region-level graph elements of different years in a specific region.

[0007] In Step S6, the method for constructing region-level graph elements for a specific region using the land use data of a certain year is as follows. First, the land use type with the highest occurrence frequency within the region is selected as the representative land use type to characterize the central node of the graph element. Then, the land use types of the first-order neighbors of the region are found, and the land use types with an occurrence frequency higher than the first preset frequency and within the top 5 in terms of occurrence frequency are selected as the representative land use types of the first-order neighbors of the graph element. Finally, the land use types of the second-order neighbors of the region are found, and the land use types with an occurrence frequency higher than the second preset frequency and within the top 5 in terms of occurrence frequency are selected as the representative land use types of the second-order neighbors.

[0008] By encoding land use data and its spatial relationships into a graph structure, the spatial, temporal, and relational attributes of land use data can be well abstracted and represented. Therefore, mining can be performed using graph analysis methods to reveal complex land use structure patterns.

[0009] The present invention uses a graph convolutional neural network model to aggregate the neighborhood information of patches. After that, each patch contains its own and its first- and second-order neighborhood land use information. Therefore, patches with similar land use characteristics and neighborhood structure information will ultimately obtain similar graph embeddings, leading to a mixed representation of graph embeddings. For example, a cultivated land may be adjacent to a water area and a village at the same time, and the characteristic values related to rivers, ponds, and villages in its graph embedding are all high. This phenomenon is particularly prominent after passing through a convolutional layer with the function of aggregating multiple neighborhood information. Therefore, a large number of misclassifications occur when predicting patch categories, or the difference in the probabilities assigned to some categories is not large. Therefore, it is inappropriate to focus only on the predicted classification types of individual patches, and it is difficult to introduce the advantages of the graph convolutional neural network model in the prior art into the mining area of the land use model.

[0010] The present invention not only focuses on the classification categories of each patch, but also performs hierarchical partitioning based on the graph embedding generated by the graph convolutional neural network model. The graph convolutional neural network model can capture the complex interactions in land use data, has a very strong modeling ability for the quantitative relationships in the figure, can form graph embedding by considering different components and structures in the data, automatically extract hidden knowledge, and is convenient for mining land use models. The present invention performs regional partitioning on the graph embedding by a spatial constraint multi-class clustering method, and then generates regional level graph elements for each hierarchical partition or sub-partition, so that the overall characteristics of homogeneous regions can be noted, and thus it can be used to mine different hierarchical structural models from complex land use information, that is, to reflect the spatio-temporal changes of the land use structure of a specific region by the changes of regional level graph elements in different years of the region. The method of the present invention can capture land use structure patterns at different scales, promote the understanding of land use in the study area, and thus provide technical support for land resource management and urban planning.

[0011] The present invention constructs a method for hierarchical representation of regional land use structure, constructs land use patches into a graph structure, performs hierarchical partitioning based on the graph embedding formed by the graph convolutional neural network, and constructs a land use structure hierarchical representation model from the regional level, sub-regional level to the patch level. When constructing graph elements, the method considers the influence of multi-hop neighbors and can perform partitioning according to the differences in land use spatial structures, and can effectively mine hierarchical land use structure models and their dynamic characteristics.

Brief Description of the Drawings

[0012]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Embodiments for Carrying out the Invention

[0013] In this embodiment, Xinbei District of Changzhou City (hereinafter abbreviated as the "experimental area") is taken as an example for explanation. However, the land use data is the land use data for six years from 2010 to 2014 and 2016 obtained through land use surveys. The experimental environment is Python 3.7, GDAL 2.3.3, and the deep learning framework Pytorch 1.6.0.

[0014] This embodiment relates to a hierarchical mining method for land use structure patterns based on a graph convolutional neural network, and includes the following steps shown in FIG. 1. Step S1: Obtain the land use data of the experimental area for six years from 2010 to 2014 and 2016, and integrate the land use data for these six years. Each year's land use data includes the type, area, and shape of the patches. When integrating, the patches (also called "land use patches") are sequentially intersected by year. When two patch data are intersected, the non-overlapping parts are divided into two patches.

[0015] First, perform data preprocessing on the annual land use data. According to the relevant national standards, the land use types are divided into ten major types: cultivated land, orchard, forest land, grassland, transportation, river, pond, village, city, and town. Remove the fragmented patches with an area smaller than 1000 square meters. After integrating the data for six years, remove the fragmented patches again to obtain 67,643 land use patches. Each patch contains the land use type, geographical coordinates, and geometric information (including area and shape) for six years.

[0016] Step S2: For the land use data integrated in Step S1, construct a graph structure with the centroid of the patch as the node and the connection line between adjacent centroids as the edge. And in the graph structure, each node records the land use type, area, and shape corresponding to the patch, that is, the centroid of the patch is used as the node for display, and the connection line between adjacent centroids is regarded as the edge. Note that when the shortest distance between two patch contours is smaller than the preset proximity threshold, the centroids of the two patches are defined as adjacent centroids, that is, only the node pairs with a distance smaller than the preset proximity threshold are considered to be edge-connected. The preset proximity threshold can be obtained based on the method of the semivariogram. As a prior art, the paper "Mining regional patterns of land use with adaptive adjacent criteria" (Authors: Tu X, Chen Z, Wang B, et al., Cartography and Geographic Information Science, 2020, 47(5): 418 - 431) can be referred to. In this embodiment, the above-mentioned experimental area data is processed, and the established graph structure contains 920,880 edges and 67,643 nodes. The graph structure can encode the structural information and functional information in the data at the same time, and can well package the space, time, and relational attributes into the abstract representation of the graph structure unit. For irregular structured data, the graph structure can represent various relationships and complex geometric features in the data of non-Euclidean distance, and is an appropriate medium that well considers local correlations.

[0017] When constructing a graph structure, some land use patches may be elongated. Before inputting the land use data into the model, it is necessary to perform special processing on the elongated patches. For example, since the aspect ratios of rivers, roads, and railways are relatively large, the shape indices of these patches become abnormally large, which may cause problems in subsequent processing. To avoid this problem, the following processing is performed on the elongated patches. First, divide the elongated patch into a plurality of patches with small areas, and then construct a graph structure. The elongated patch is a patch whose shape index is larger than a preset index, and the shape index of the patch is as follows. [Number] Here, D is the shape index, P is the perimeter length of the patch, and A is the area of the patch.

[0018] However, since the number of elongated roads and rivers in the experimental area in this embodiment is extremely small, the appearance frequency of graph elements centered on rivers and transportation is significantly lower than that of graph elements centered on many other land use types (such as cultivated land, towns, etc.). These roads and rivers are connected to patches of many other land use types, which means that the elongated roads and rivers in the experimental area are not processed because rivers and transportation land play the role of near the central node within the graph elements. However, when using the method of the present invention for other regions, since the land use data of the region may contain a large number of irregular shapes, special consideration is required for such patches, and they are divided into a plurality of patches with small areas.

[0019] Step S3: Assign labels to each node and use them to mark the land use type, the nearest dominant type, and the second nearest dominant type of the node, and obtain a label input model. The nearest dominant type is the land use type with the largest adjacent area among the first-order neighbors of the node, and the second nearest dominant type is the land use type with the second largest adjacent area among the first-order neighbors of the node.

[0020] The label in this embodiment describes the characteristics of each land use patch node and the structure of its first neighbors, and includes the following parts. The first part is used to represent the land use type that represents the land use patch node itself. The second part represents that in the first neighbors of the node, the land use type with the largest neighboring area is selected as the neighboring dominant type. The third part represents that in the first neighbors of the node, the land use type with the second largest neighboring area is selected as the neighboring sub-dominant type. For example, if the land use type of the node itself, the neighboring dominant type, and the neighboring sub-dominant type are cultivated land, town, and transportation respectively, the node label is displayed as "cultivated land town transportation".

[0021] For the sake of simplifying the calculation, it is preferable to perform integration on the labels in the label input model, that is, classify them as follows. Labels with a quantity greater than 1% are classified into a single category. For labels with a quantity less than 1%, labels with the same node type and neighboring dominant type are integrated into one label. If the quantity of the integrated label is greater than 1%, the label is classified into one category. The remaining labels are integrated into one category with the same node type.

[0022] As shown in Figure 2, in this embodiment, by processing the experimental area data, all combinations of land use types are covered, 990 initial labels are obtained, label integration is performed to simplify the calculation, and finally 27 labels are generated and assigned to each land use patch for model learning.

[0023] Step S4: Construct a graph convolutional neural network model, perform learning on the label input model, and generate a graph embedding. The Graph Convolutional Neural Network model (Graph Neural Network, GCN) is an emerging graph deep learning model and an existing technology that can learn local structural information through convolutional filters. This network aggregates neighboring data of different hops with different weights and can reflect the distance attenuation neighborhood effect. The Graph Convolutional Neural Network model randomly initializes the weights, normalizes the input feature vectors, optimizes the hyperparameters using the training set to obtain an optimized model, can capture the complex interactions in land use data, has a very strong modeling ability for the quantitative relationships in the figure, and can form a graph embedding by considering various different components and structures in the data, aggregate the information of spatial units and their neighboring information, automatically extract hidden knowledge, and facilitate the mining of land use models.

[0024] As shown in Figure 3, the Graph Convolutional Neural Network model of this embodiment is composed of two convolutional layers, two non-linear activation layers and one fully connected layer, and divides the data into a training set and a test set at a ratio of 80:20. Taking the experimental area data as an example, the model adopts a hidden layer of 128 neurons, the dropout ratio is 0.5, the L2 regularization coefficient is 0.001, the learning rate is 0.01, the upper limit of the number of iterations is 200, and finally outputs the graph embedding.

[0025] In the learning process of the graph convolutional neural network model, the number of hidden layer neurons may affect the generation result of the graph embedding. In this embodiment, the number of hidden layer neurons in the graph convolutional neural network model is determined using the grid search method. The mesh search method is a prior art. Referring to the paper "Network Traffic Prediction Based on Grid Search Support Vector Machine" (Author: Liu Daowen, Hu Haina, Computer Applications and Software, 2012, 29(11): 185-186+247), in the grid search process, neurons in the hidden layer are set to different numbers and the corresponding model accuracy is calculated. A numerical value with high model accuracy and short learning time is selected as the numerical value of the number of hidden layer neurons. Due to the randomness of the neural network, the graph embeddings of the land use patches obtained each time may be slightly different. Also, considering the smoothness of the graph neural network, adjacent land use units may ultimately have similar graph embeddings. Due to these properties of the neural network, the boundaries of the partitions and sub-partitions may change slightly each time the graph convolutional network model is executed. Therefore, it is recommended to run this model multiple times and use the average result as the final partition.

[0026] Step S5: Use the spatial constraint multi-class clustering method to partition the graph embedding to generate partitions with different structural features. By iteration, each sub-partition is partitioned again respectively, and so on by analogy until the area of the smallest sub-partition is less than the preset area threshold, and the iteration is stopped. Finally, a hierarchical partition structure from the partition to each level of sub-partition is obtained.

[0027] The spatial constraint multi - class clustering method is a prior art that can refer to the paper "Efficient regionalization techniques for socio - economic geographical units using minimum spanning trees" (Authors: Assuncao R M, Neves M C, Camara G, et al., International Journal of Geographical Information Science, 2006, 20(7): 797 - 811). Its purpose is to make all features within each cluster as similar as possible and amplify the differences between different clusters as much as possible. The spatial constraint multi - class clustering method mines structural models at different hierarchical scales, thereby performing geographical partitioning. Then, it repeatedly divides sub - partitions for each partition. When the area of the smallest sub - partition is below the preset area threshold, the iterative partitioning calculation ends. The division of sub - partitions still has areas with clearly different structural features in each partition. After multiple iterations, hierarchical partitions from the partition to each level of sub - partitions with different structural features are realized.

[0028] Taking the experimental area as an example, in this embodiment, an area threshold of 1000 square meters is preset. After two divisions, a total of 8 partitions and 35 sub - partitions are obtained.

[0029] Step S6: Collectively refer to all the partitions and each level of sub - partitions obtained in Step S5 as regions. For the annual land use data, construct the regional - level graph elements of each region according to the frequency characteristics of each land use type, and reflect the spatio - temporal changes of the land use structure of the corresponding region by the changes of the regional - level graph elements of different years in a specific region, thereby deriving regional patterns at different levels. Construct the hierarchical structure of land use (shown in Figure 4) to reveal the characteristics of land use structures in different - scale regions and different years. For example, the land use structure pattern and dynamic changes at the patch level can be analyzed according to the appearance frequency of patch - level graph elements.

[0030] Here, the method for constructing graph elements at the regional level for a specific area using the land use data of a certain year is as follows. First, select the land use type with the highest frequency of occurrence within the area as the representative land use type to characterize the central node of the graph element. Subsequently, find the land use types of the first-order neighbors of the area, and select the land use types whose occurrence frequencies are greater than the first preset frequency and within the top 5 in terms of occurrence frequency as the representative land use types of the first-order neighbors of the graph element. Being within the top 5 in terms of ranking is a preferred value. Based on the applicant's experience, it is only necessary to pay attention to the land use types within the top 5 in terms of frequency. Finally, find the land use types of the second-order neighbors of the area, and select the land use types whose occurrence frequencies are greater than the second preset frequency and within the top 5 in terms of occurrence frequency as the representative land use types of the second-order neighbors.

[0031] Graph elements are the basic components of complex networks, and the graph elements of each section and sub-section represent the land use structure of that area. This embodiment processes the experimental area data. In each area, the central node of the graph element is characterized by the most frequent land use type within the area. Then, based on the types of the first-order neighbors of all graph elements centered around the central node of the graph element, select the land use types whose frequencies are greater than 10% and within the top 5 as the representative land use types of the first-order neighbors of the graph element. By analogy in this way, determine the representative land use types of the second-order neighbors and form the land use hierarchical structure of the experimental area.

[0032] The applicant has discovered that land use patches with similar land use characteristics and neighboring structure information will ultimately result in similar graph embeddings, leading to a mixed representation of the graph embeddings. For example, one cultivated land may be adjacent to water areas and villages simultaneously. In this case, the feature values regarding rivers, ponds, and villages in its graph embedding will all be high. This phenomenon is particularly prominent after passing through multiple convolutional layers with the function of aggregating neighboring information. Such a problem will lead to many misclassifications when predicting patch categories because the difference in the probabilities assigned to several categories is not large enough. Therefore, it is not appropriate to focus only on the predicted classification type of individual land use patches, and it becomes difficult to introduce the advantages of graph convolutional neural networks in the prior art into the mining area of land use patterns. Instead of only focusing on the classification category of each land use patch, the present invention performs hierarchical partitioning based on the graph embeddings generated by the graph convolutional neural network, generates regional-level graph elements for the partitioning at each hierarchical level, and focuses on the overall features within one homogeneous region. In this way, the graph convolutional neural network can be used to mine the structural patterns of different hierarchies (including partitions, sub-partitions, patches, etc.) from complex land use information.

[0033] The changes in regional-level graph elements of a specific region in different years, that is, the changes in regional-level graph elements from year t - 1 to year t, include appearance, disappearance, branch increase, and branch decrease. As shown in FIG. 6, appearance refers to the case where one graph element did not temporarily exist in year t - 1 but suddenly appeared in year t. Disappearance refers to the case where one graph element existed in year t - 1 but suddenly disappeared in year t. Branch increase refers to the case where a part of one graph element did not temporarily exist in year t - 1 but suddenly appeared in year t. Branch decrease refers to the case where a part of one graph element existed in year t - 1 but suddenly disappeared in year t. These four types of changes in graph elements basically indicate the changes in the land use patterns of each type in each region within the experimental area, and can assist in identifying the spatio-temporal changes in the land use structure.

[0034] From the changes in the graph elements of each region in the experimental area from 2010 to 2016, as shown in Figure 7, it was found that the land use change patterns commonly seen in the urban and rural areas of the experimental area include the expansion of town or rural housing, transportation construction, and the expansion of towns around the city.

[0035] Note that due to the randomness of the neural network, the graph embeddings obtained each time may be slightly different. Also, considering the smoothness of the graph neural network, adjacent land use units may eventually have similar graph embeddings. Due to these properties of the neural network, the boundaries of the compartments and sub-compartments may change slightly each time the model is run. Therefore, it is recommended to run this model multiple times and use the average result as the final compartment result.

Claims

1. A method for hierarchical mining of land use structure patterns based on graph convolutional neural network, comprising: Step S1: acquiring land use data for N years, N≧2, of a specific area, and integrating the land use data for N years; Step S2: for the land use data integrated in step S1, construct a graph structure with the centroids of the patches as nodes and the connecting lines between the adjacent centroids as edges, and record the land use type, area and shape of the corresponding patch of each node in the graph structure, in which if the shortest distance between the contours of two patches is less than a preset proximity threshold, the centroids of the two patches are defined as the proximity centroids; Step S3: assigning labels to each node to mark the land use type, the most dominant type, and the second most dominant type of the node to obtain a label input model, the most dominant type being the land use type with the largest neighborhood area among the first-order neighborhood of the node, and the second most dominant type being the land use type with the second largest neighborhood area among the first-order neighborhood of the node; Step S4: constructing a graph convolutional neural network model, training the label input model, and generating a graph embedding; S5, partitioning the graph embedding using a spatially constrained multi-way clustering method to generate partitions with different structure characteristics, and then re-partitioning each sub-partition by iteration, and by this analogy, stopping the iteration until the area of ​​the smallest sub-partition falls below a preset area threshold, and finally obtaining a hierarchical partition structure from the partition to each level sub-partition; All the divisions and each level subdivision obtained in step S5 are collectively called divisions, and the step S6 includes constructing division-level graph elements of each division according to the frequency characteristics of land use types for each year of land use data, and reflecting the spatiotemporal changes of the land use structure of the division through the changes of division-level graph elements in different years of a specific division; In step S6, the method for constructing area-level graph elements for a particular area using land use data for a given year includes: First, select the land use type with the highest occurrence frequency in the area and use it as the representative land use type. Then, characterize the central node of the graph element. Then, find the land use types of the first-order neighborhood of the area, select the land use type whose occurrence frequency is higher than the first preset frequency and whose occurrence frequency is within the top 5, and set it as the representative land use type of the first-order neighborhood of the graph element; Finally, find the land use types of the secondary neighborhood of the area, and select the land use type whose occurrence frequency is higher than the second preset frequency and whose occurrence frequency is within the top 5, as the representative land use type of the secondary neighborhood. A method for hierarchical mining of land use structure patterns based on graph convolutional neural networks, comprising:

2. In step S6, the changes of the area-level graph elements in different years in a specific area include four kinds of area-level graph elements: appearance, disappearance, branch increase and branch decrease; The method for hierarchical mining of land use structure patterns based on graph convolutional neural network as claimed in claim 1.

3. In step S2, firstly, the elongated patch is divided into a plurality of non-elongated patches, and a graph structure is constructed, the elongated patch is a patch whose shape index is larger than a preset index, and the shape index of the patch is as follows: [0010] where D is the shape exponent, P is the perimeter of the patch, and A is the area of ​​the patch. The method for hierarchical mining of land use structure patterns based on graph convolutional neural network as claimed in claim 1.

4. In step S3, integration is performed on the labels in the label input model, i.e. Labels with a quantity of more than 1% will be classified into a single category. For labels with a quantity less than 1%, the labels with the same node type and the most dominant nearby type are merged into one label. If the quantity of the labels after the merge is greater than 1%, the label is classified into one class, and the remaining labels with the same node type are merged into one class. The method for hierarchical mining of land use structure patterns based on graph convolutional neural network as claimed in claim 1.

5. In step S4, a grid search method is used to determine the number of hidden layer neurons in the graph convolutional neural network model; The method for hierarchical mining of land use structure patterns based on graph convolutional neural network as claimed in claim 1.

Images