A 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 network, which belongs to the field of data processing technology. [Background technology]

[0002] Land use directly reflects the interaction and mutual influence between people and nature, and land use structure is the proportional relationship or composition of various land types within a certain area. Land use pattern refers to the characteristics of land use elements within an area, including the characteristics, composition, spatial distribution and spatial relationships of land use structure, and reflects the development of important processes such as socio-economy. The evolution of land use structure patterns reflects the development of important processes such as socio-economy, and mining these structural patterns is helpful in promoting the harmonious development of the relationship between people and land in the region and enhancing the spatial management capacity of the country.

[0003] In traditional land use pattern research, 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 usually used to analyze land use structure and spatial structure. The papers "Multi-scale mining method of salient spatial association patterns" (authors: He Zhanjun et al., Journal of Map Surveying and Mapping, No. 11, 2016) and "Research on mining algorithm of spatial association patterns" (Yang ▲Chong ▼, Kunming University of Science and Technology PhD dissertation, 2012) can be referred to. In the above-mentioned methods, spatial relationships are often determined with a certain distance threshold for different geographical areas, and mining association rules are derived. However, due to spatial heterogeneity, a certain geographical proximity threshold is not very suitable for carrying out local area pattern mining, which limits its scope of application.

[0004] In addition, currently, when making quantitative descriptions of spatial structures, overall statistics are often used to describe spatial phenomena, which can only reflect the overall situation of land use and ignore the relationship between spatial structure layout and geographic objects. That is, these indicators only focus on morphological and geometric attributes and do not consider the interactions between land use patches, which underestimates the importance of land use spatial structure and limits the ability to sense changes in land use patterns. Therefore, when mining regional patterns, traditional methods cannot automatically identify regions with different land use structures, and can only display the structures of traditional plots, so the mining display of land use structures at different levels and different scales is lacking, which makes it difficult for people to accurately and comprehensively recognize the characteristics and dynamic changes of regional land use structure patterns. 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 regional land use structure patterns, which can recognize the characteristics and dynamic changes of regional land use structure patterns, and can assist in recognizing the spatiotemporal changes of land use structure.

[0006] In order to solve the above technical problems, the technical solutions put forward by the present invention are as follows: A hierarchical mining method for land use structure patterns based on graph convolutional neural network, which includes the following steps: Step S1: Obtain N-year land use data, where N≧2, for a specific area, and integrate the N-year land use data. 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. Here, if the shortest distance between the contours of two patches is smaller than the preset proximity threshold, the centroids of the two patches are defined as the proximity centroids. Step S3: A label is assigned to each node, which is used 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, where the most dominant type is the land use type with the largest neighboring area among the first-order neighbors of the node, and the second most dominant type is the land use type with the second largest neighboring area among the first-order neighbors of the node. Step S4: Construct a graph convolutional neural network model and train it on the label input model to generate a graph embedding. Step S5: Use a spatially constrained multi-way clustering method to partition the graph embedding, and generate partitions with different structure characteristics. By iteration, partition each sub-partition again, and by analogy, stop iterating until the area of ​​the smallest sub-partition is below a preset area threshold. Finally, obtain a hierarchical partition structure from the partition to each level sub-partition. Step S6: All the divisions and each level subdivision obtained in step S5 are collectively called districts, and for each year's land use data, district-level graph elements are constructed for each district according to the frequency characteristics of land use types, and the changes in district-level graph elements in different years for a specific district reflect the spatiotemporal changes of the land use structure of the district.

[0007] In step S6, the method for constructing area-level graph elements for a particular area using land use data for a given year is as follows. First, the land use type that appears most frequently in the area is selected as a representative land use type, and the central node of the graph element is characterized. Then, find the land use types in the first-order neighborhood of the area, and select the land use type whose occurrence frequency is higher than the first preset frequency and whose occurrence frequency is within the top 5, as the representative land use type in 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.

[0008] By encoding land use data and its spatial relationships into a graph structure, the present invention can provide a good abstraction of the spatial, temporal and relationship attributes of land use data, so that complex land use structure patterns can be revealed by mining using graph analysis methods.

[0009] After the present invention uses the graph convolutional neural network model to aggregate the neighborhood information of the patch, each patch contains the land use information of itself and its first and second neighbors. Therefore, patches with similar land use characteristics and neighborhood structure information will finally obtain similar graph embeddings, resulting in a mixed representation of graph embeddings. For example, a cultivated land may be adjacent to a water body and a village at the same time, and the feature values ​​of the river, the reservoir, and the village 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, and therefore a large amount of misclassification occurs when predicting patch classification, because the difference in the probability of being assigned to some classes is not large. Therefore, it is inappropriate to focus only on the predicted classification type of each individual patch, and it is difficult to introduce the advantages of the graph convolutional neural network model in the prior art into the mining area of ​​land use models.

[0010] The present invention does not only focus on the classification of each patch, but 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 the land use data, has a very strong modeling ability for the quantitative relationships in the diagram, and can form a graph embedding by considering the different components and structures of each class in the data, which can automatically extract hidden knowledge and is convenient for mining land use models. The present invention performs area partitioning on the graph embedding by the spatially constrained multi-way clustering method, and then generates area-level graph elements for each level of area or sub-area, thus focusing on the overall characteristics in homogeneous areas, which can be used to mine structural models of different levels from complex land use information, that is, the changes in area-level graph elements in different years of a specific area reflect the spatiotemporal changes of the land use structure of the area. The method of the present invention can capture the land use structure patterns of different scales, which promotes the understanding of the land use of the research area, and thus provides technical support for land resource management and urban planning.

[0011] The present invention establishes 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 a graph convolutional neural network, and establishes a hierarchical representation model of land use structure from the region level, sub-region level to the patch level. The method considers the influence of multi-order neighbors when constructing graph elements, and can perform partitioning according to the difference of land use spatial structure, and can effectively mine the hierarchical land use structure model and its dynamic characteristics. [Brief description of the drawings]

[0012] [Figure 1] 1 is a flowchart of an embodiment of the present invention. [Diagram 2] FIG. 2 is a schematic diagram of a label input model generated according to an embodiment of the present invention; [Diagram 3]FIG. 2 is a schematic diagram of a graph convolutional neural network model used by an embodiment of the present invention. [Figure 4] FIG. 1 is a schematic diagram of a land use hierarchy pattern. [Diagram 5] FIG. 2 is a schematic diagram of a district level graph element. [Figure 6] FIG. 2 is a schematic diagram of changes in graph elements. [Figure 7] Figure 1 shows a schematic diagram of the land use structure pattern of each plot in the experimental area from 2010 to 2016. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] In this embodiment, Xinbei District, Changzhou City (hereinafter referred to as the "experimental district") is used as an example, and the land use data is obtained from a land use survey for six years from 2010 to 2014 and 2016. 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 graph convolutional neural network, which includes the following steps, as shown in FIG. 1: Step S1: Obtain land use data for 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 patch. When integrating, patches (also called "land use patches") are intersected sequentially by year. When two patch data are intersected, the non-overlapping parts are divided into two patches.

[0015] First, data pre-processing was carried out on the annual land use data. In accordance with relevant national standards, the land use types were divided into ten major types, including cultivated land, orchards, forestland, grassland, transportation, rivers, reservoirs, villages, cities and towns. Fragmentary patches with an area of ​​less than 1,000 square meters were removed. After combining the six-year data, the fragmentary patches were removed again, obtaining 67,643 land use patches, each of which contains six-year land use type, geographical coordinates and geometric information (including area and shape).

[0016] Step S2: For the land use data after integration in step S1, construct a graph structure with the centroids of the patches as nodes and the connecting lines between adjacent centroids as edges, and in the graph structure, each node records the land use type, area and shape corresponding to the patch, i.e., display the centroids of the patches as nodes and regard the connecting lines between adjacent centroids as edges. Note that if the shortest distance between the contours of two patches is smaller than a preset proximity threshold, the centroids of the two patches are defined as adjacent centroids, i.e., only node pairs whose distance is smaller than the preset proximity threshold are considered to be edge-connected. The preset proximity threshold can be obtained based on the method of semi-mutation function, and 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) is a reference for the prior art. In this embodiment, the above-mentioned experimental data is processed, and the established graph structure includes 920880 edges and 67643 nodes. The graph structure can simultaneously encode the structural information and functional information in the data, and can well package the spatial, temporal and relational attributes into the abstract representation of the graph structure unit. For irregular structure data, the graph structure is an appropriate medium that can express various relationships and complex geometric features in the data of non-Euclidean distances, and well considers local correlation.

[0017] When constructing the graph structure, some land use patches may be elongated, and special processing is required for elongated patches before inputting the land use data into the model. For example, the aspect ratio of rivers, roads and railways is relatively large, so the shape index of these patches will be abnormally large, which may cause problems in subsequent processing. To avoid this problem, the following processing is performed on the elongated patches. First, the elongated patch is divided into multiple patches with small areas, 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:

number

[0018] However, since the number of narrow roads and rivers in the experimental area in this embodiment is very small, the frequency of graph elements centered on rivers and traffic is significantly lower than that of graph elements centered on many other land use types (e.g., farmland, towns, etc.). These roads and rivers are connected to patches of many other land use types, which is why the rivers and traffic lands play the role of neighbors of the central nodes in the graph elements, and so the narrow roads and rivers in the experimental area are not processed. However, when using the method of the present invention for other areas, such patches need special consideration because the land use data for the area may contain a large amount of irregular shapes, and they are divided into multiple patches with small areas.

[0019] Step S3: A label is assigned to each node, which is used 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, where the most dominant type is the land use type with the largest neighboring area among the first-order neighbors of the node, and the second most dominant type is the land use type with the second largest neighboring area among the first-order neighbors of the node.

[0020] The label in this embodiment describes the characteristics and first-order neighborhood structure of each land use patch node, 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-order neighborhood of the node, the land use type with the largest neighborhood area is selected as the most dominant type; The third part represents that in the first-order neighborhood of the node, the land use type with the second largest neighborhood area is selected as the second most dominant type. For example, if the land use type, the most dominant type, and the second most dominant type of the node itself are cultivated land, town, and traffic, respectively, the node label is displayed as "cultivated land town traffic".

[0021] To simplify the computations, it is preferable to perform a consolidation on the labels in the label input model, ie, classify them as follows: Labels with a quantity greater 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, and if the quantity of the labels after the merge is greater than 1%, the label is classified into one class. For the remaining labels, the labels with the same node type are merged into one class.

[0022] As shown in Figure 2, in this embodiment, by processing the experimental area data, all combinations of land use types are covered, and an initial 990 labels are obtained. To simplify the calculation, the labels are integrated, 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 and train it on the label input model to generate a graph embedding. Graph Neural Network (GCN) is an emerging graph deep learning model and an existing technology that can learn local structure information through convolution filters. This network can aggregate neighboring data of different hops with different weights to reflect the neighborhood effect of distance decay. The graph convolution neural network model randomly initializes the weights, normalizes the input feature vector, and uses the learning set to perform super-parameter optimization to obtain the optimized model, which can capture the complex interactions in the land use data, has very strong modeling ability for quantitative relationships in the map, and can form graph embedding by considering the different components and structures of each type in the data, aggregate the information of spatial units and their neighboring information, automatically extract hidden knowledge, and facilitate land use model mining.

[0024] The graph convolutional neural network model of this embodiment is composed of two convolutional layers, two nonlinear activation layers and one fully connected layer, as shown in Figure 3, and divides the data into a training set and a test set in a ratio of 80:20. Taking the experimental data as an example, the model adopts a hidden layer of 128 neurons, the drop out ratio is 0.5, the L2 normalization coefficient is 0.001, the learning rate is 0.01, the upper limit of the number of iterations is 200, and the final graph embedding is output.

[0025] In the learning process of the graph convolutional neural network model, the number of hidden layer neurons may affect the graph embedding generation result. In this embodiment, the number of hidden layer neurons in the graph convolutional neural network model is determined by using a grid search method. The mesh search method is a conventional technology, and in the grid search process, referring to the paper "Network Traffic Prediction Based on Grid Search Support Vector Machine" (authors: Liu Daowen, Hu Haina, Computer Applications and Software, 2012, 29(11):185-186+247), we try to set different numbers of neurons in the hidden layer to calculate the corresponding model accuracy, and select the quantity value with high model accuracy and short learning time as the quantity value of the hidden layer neurons. Due to the randomness of the neural network, the graph embedding of the land use patch obtained each time may be slightly different. In addition, 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 partitions and sub-partitions may change slightly each time the graph convolutional network model is run. Therefore, we recommend running the model multiple times and taking the average result as the final partition.

[0026] Step S5: Use a spatially constrained multi-way clustering method to partition the graph embedding, and generate partitions with different structure characteristics. By iteration, partition each sub-partition again, and by analogy, stop iterating until the area of ​​the smallest sub-partition is below a preset area threshold. Finally, obtain a hierarchical partition structure from the partition to each level sub-partition.

[0027] The spatially constrained multi-way clustering method is a prior art that can be referred to in the paper "Efficient regionalization techniques for socio-economic geographical units using minimum spanning trees" (Authors: Assuncao RM, Neves MC, Camara G, et al., International Journal of Geographical Information Science, 2006, 20(7): 797-811), which aims to make all the characteristics in each cluster as similar as possible and amplify the differences between different clusters as much as possible. The spatially constrained multi-way clustering method mines structural models at different scales hierarchically, thereby performing geographical division, and then iteratively divides sub-divisions for each division, and when the area of ​​the smallest sub-division falls below the preset area threshold, the iterative division calculation ends. The division of sub-divisions is still a process in which each division has areas with obviously different structural characteristics, and after multiple iterations, the structural characteristics are different, realizing a hierarchical division from divisions to each level sub-division.

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

[0029] Step S6: All the plots and each level sub-plot obtained in step S5 are collectively called zones. For each year's land use data, zone-level graph elements are constructed for each zone according to the frequency characteristics of the land use types. The changes in zone-level graph elements in different years of a specific zone reflect the spatiotemporal changes in the land use structure of the zone, thereby deriving zone patterns at different levels. A land use hierarchical structure (shown in Figure 4) is constructed to display the characteristics of land use structure in different scale zones and different years. For example: The occurrence frequency of patch-level graph elements can be used to analyze the land use structure pattern and dynamic changes at the patch level.

[0030] Here is how to construct an area-level graph element for a particular area using land use data for a given year: First, the land use type that appears most frequently in the area is selected as a representative land use type, and the central node of the graph element is characterized. Then, find the land use types of the first-order neighborhood of the area, and select the land use types whose occurrence frequency is greater than the first preset frequency and whose occurrence frequency is within the top 5, as the representative land use types of the first-order neighborhood of the graph element. The top 5 ranks are preferable values, and based on the applicant's experience, it is sufficient to pay attention to the land use types whose frequency is within the top 5. Finally, find the land use types of the secondary neighborhood of the area, and select the land use type whose occurrence frequency is greater than the second preset frequency and whose occurrence frequency is within the top 5, as the representative land use type of the secondary neighborhood.

[0031] Graph elements are the basic building blocks of complex networks, and the graph elements of each district and subdivision represent the land use structure of the district. This embodiment processes the experimental district data, and in each district, characterizes the central node of the graph element with the most frequent land use type in the district, and then selects the land use type with a frequency greater than 10% and within the top 5 according to the first-order neighbor types of all graph elements centered on the central node of the graph element as the representative land use type of the first-order neighbor of the graph element, and thus determines the representative land use type of the second-order neighbor to form the land use hierarchical structure of the experimental district.

[0032] The applicant has found that land use patches with similar land use features and neighborhood structure information will eventually obtain similar graph embeddings, leading to mixed graph embedding representations. For example, a piece of cultivated land may be adjacent to a water body and a village at the same time, in which case the feature values ​​for rivers, reservoirs, and villages in its graph embedding will all be high. This phenomenon is particularly evident after passing through multiple convolutional layers with the function of aggregating neighborhood information. This problem results in a lot of misclassifications in predicting patch categories, since the difference in the probabilities assigned to some categories is not very large. Therefore, it is not appropriate to focus only on the predicted classification type of individual land use patches, which makes it difficult to introduce the advantages of graph convolutional neural networks in the current technology into land use pattern mining areas. The present invention does not only focus on the classification category of each land use patch, but performs hierarchical partitioning based on the graph embedding generated by the graph convolutional neural network, generates area-level graph elements for each hierarchical level of partitions, and focuses on the overall features in one homogeneous area. In this way, graph convolutional neural networks can be used to mine structural patterns at different hierarchies (including parcels, sub-parcels, patches, etc.) from complex land use information.

[0033] The changes of the area-level graph elements in different years in a specific area, i.e., the changes of the area-level graph elements from year t-1 to year t, include appearance, disappearance, branch increase and branch decrease. As shown in Figure 6, appearance refers to the case where a graph element was temporarily absent in year t-1 but suddenly appeared in year t. Disappearance refers to the case where a graph element was present in year t-1 but suddenly disappeared in year t. Branch increase refers to the case where a part of a graph element was temporarily absent in year t-1 but suddenly appeared in year t. Branch decrease refers to the case where a part of a graph element was present in year t-1 but suddenly disappeared in year t. The changes of these four types of graph elements basically indicate the changes in the land use patterns of each type in each area in the experimental area, which can assist in identifying the spatiotemporal changes in land use structure.

[0034] Based on the changes in the graphical elements of each area in the experimental area from 2010 to 2016, as shown in Figure 7, it can be seen that the land use change patterns commonly seen in urban and rural areas in the experimental area include residential expansion in towns or rural areas, transportation construction, and expansion of towns around cities.

[0035] Please note that due to the randomness of the neural network, the graph embedding obtained each time may be slightly different. Also, considering the smoothness of the graph neural network, adjacent land use units may end up having similar graph embeddings. Due to these properties of the neural network, the boundaries of the parcels and subparcels may change slightly each time the model is run. Therefore, we recommend running this model multiple times and taking the average result as the final parcel 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 a label to each node, which is used to mark the land use type, the neighborhood most dominant type, and the neighborhood second most dominant type of the node, to obtain a label input model, where the neighborhood most dominant type is the land use type with the largest neighborhood area among the first-order neighborhood of the node, and the neighborhood second most dominant type is 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 number of non-elongated patches with small areas, and then 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