Forest fire risk prediction method based on graph neural network with variable area units

Abstract

The present invention proposes a graph neural network forest fire risk prediction method based on variable area units. By comprehensively considering the attribute similarity and spatial adjacency between geographic units to merge geographic units, the geographic units are modeled as nodes in the graph model, and then connected to the downstream forest fire prediction graph neural network model, providing a basis for accurately predicting the future forest fire risk of the merged geographic units, forming a new method that is different from the regular grid method for dividing the target area geographic units, so as to better understand the law of forest fire occurrence and provide a technical approach to achieve more accurate forest fire risk assessment.

Description

Technical Field

[0001] The invention belongs to the field of computers, and in particular relates to a graph neural network forest fire risk prediction method based on variable area units. Background Art

[0002] Traditional methods include fire indices and mechanism models that consider factors closely related to forest fire occurrence. Salavatiet et al. conducted a study in Sanandaj, a city in western Iran. In this study, they used the "weight of evidence" (WoE) and "statistical index" (SI) models to assess fire risk potential. Chen et al. [2] They believe that precipitation is an important factor affecting the probability of forest fires, and they proposed a method to better represent the impact of precipitation on forest fire prediction. Ge et al. proposed a comprehensive forest fire driver index for forest fire prediction, which is a hierarchical analysis method that takes into account terrain, vegetation, meteorology and human activities.

[0003] Recently, a considerable body of research has shown that machine learning-based methods can discover more complex patterns in forest fire data than traditional mechanism-based or statistical models.

[0004] Jaafari et al. used WoE Bayesian modeling to study the spatial relationships between historical fire events in Chaharmahal-Bakhtiari Province, Iran.

[0005] Sakr et al. proposed a forest fire risk prediction algorithm based on support vector machine (SVM). To predict forest fires, Binh Thai Pham et al. compared the ability of Bayesian network (BN), naive Bayes (NB), decision tree (DT) and multivariate logistic regression (MLP) to develop fire susceptibility maps in Phu Mat National Park, Ninh An Province, Vietnam.

[0006] Ma et al. established a forest fire probability model based on the Logistic model and the random forest (RF) model based on forest heat anomaly data monitored by satellite from 2010 to 2017. The model analyzed the driving factors of forest fires in Shanxi Province, China from four aspects: meteorology, topography, vegetation, and human activities.

[0007] Ge et al. proposed a forest fire prediction method that combines spatiotemporal knowledge graphs with machine learning models, which can integrate multi-source heterogeneous spatiotemporal forest fire data, and proposed a rule-based reasoning method to obtain corresponding data for specific machine learning forest fire prediction methods, thereby significantly improving the prediction performance in actual forest fire prediction scenarios.

[0008] Ge Xingtong et al. proposed a forest fire risk prediction method based on graph convolutional neural network. This method established a feature representation method and graph model suitable for geographic spatial data, and combined with graph convolutional neural network technology, comprehensively considered the impact of meteorological, topographic, and vegetation spatiotemporal data characteristics of the adjacent areas of the prediction target area on the occurrence and spread of fire.

[0009] Some scholars argue that due to the spatial dependence of geographic space, when predicting forest fire risk for a geographic unit, not only its own risk factors should be considered, but also the risk factors of neighboring geographic units. Ge Xingtong et al. proposed a forest fire risk prediction method based on graph convolutional neural network. This method established a feature representation method and graph model suitable for geographic spatial data, and combined with graph convolutional neural network technology, comprehensively considered the impact of meteorological, topographic, and vegetation spatiotemporal data characteristics of neighboring geographic units in the prediction target area on fire occurrence and propagation.

[0010] Existing technologies divide the target area into geographic units in a regular grid manner, without fully considering the important impact that different spatial unit divisions may lead to different analysis results. Spatial heterogeneity points out that geographical phenomena are usually spatially uneven, and different regions may have different characteristics and attributes. Variable area units allow researchers to adjust the size and shape of spatial units according to the actual distribution of geographical phenomena to better capture these characteristics. In addition, the "first law of geography" in geography points out that geographical features tend to be more similar to those of neighboring areas. By considering variable area units, this spatial dependence can be better understood and represented. Summary of the invention

[0011] The purpose of the present invention is to overcome the above shortcomings of the prior art and propose a graph neural network forest fire risk prediction method based on variable area units. By comprehensively considering the attribute similarity and spatial adjacency between geographic units to merge geographic units, the geographic units are modeled as nodes in the graph model, and then connected to the downstream forest fire prediction graph neural network model, providing a basis for accurately predicting the future forest fire risks of the merged geographic units, forming a new method that is different from the regular grid method of dividing the target area geographic units, so as to better understand the laws of forest fire occurrence and provide a technical approach to achieve more accurate forest fire risk assessment.

[0012] The present invention aims to propose a forest fire risk graph model prediction method based on variable area units. The method comprehensively considers the attribute similarity and spatial adjacency between geographic units to merge geographic units, so as to model the geographic units as nodes in the graph model, and then connect to the downstream forest fire prediction graph neural network model, providing a basis for accurately predicting the future forest fire risk of the merged geographic units.

[0013] The method of measuring the attribute similarity and spatial adjacency between geographic units is studied. Based on the spatial adjacency of geographic units, the present invention further measures the attribute similarity between geographic units to merge geographic units, models the merged geographic units as nodes in the graph model, and connects to the downstream forest fire risk prediction graph neural network model.

[0014] The present invention proposes a forest fire risk prediction method based on a graph neural network with variable area units, which specifically includes:

[0015] (1) Graph node construction method based on variable area units: collect meteorological data (temperature, dew point temperature, precipitation, wind component in the east-west direction, wind component in the north-south direction), terrain (slope, slope aspect, elevation) and land cover type that are closely related to the occurrence of forest fires as risk factor data, and design vector representation methods for different types of data. Multi-source risk factor data are spatially superimposed so that the risk prediction target area is divided into irregularly shaped geographic units. Based on the spatial adjacency relationship of geographic units, the attribute similarity between geographic units is further measured to merge geographic units. The merged geographic units are modeled as nodes in the graph model and connected to the downstream forest fire risk prediction graph neural network model.

[0016] (2) Construction of a graph neural network model for forest fire prediction: Graph convolutional neural network (GCN) is a deep learning technology that focuses on graph structure. It can aggregate information from neighboring geographic units and help simulate complex associations in geographic space. Combining feature representation methods and graph convolutional neural networks (GCN), a comprehensive model can be constructed to effectively capture the characteristics of risk factors and risk conditions of the target geographic unit and its neighboring units, and achieve end-to-end prediction of forest fire risks.

[0017] (3) Sample dataset construction: It is necessary to construct a sample dataset containing multi-source spatiotemporal data on risk factors (meteorology, topography, land cover type) and risk conditions (historical fire events) for model training, verification, and testing. These datasets are key resources in the research and development process.

[0018] More specifically, the present invention proposes a graph neural network forest fire risk prediction method based on variable area units, and the method specifically comprises the following steps:

[0019] Step 1: Graph node construction based on variable area units; collect meteorological, topographic and surface cover types that are closely related to the occurrence of forest fires as risk factor data, and design vector representation methods for different types of data, where meteorological includes temperature, dew point temperature, precipitation, and wind component in the east-west direction, and topographic includes slope, slope direction, and elevation; spatially superimpose multi-source risk factor data, divide the risk prediction target area into irregularly shaped geographic units, and further measure the attribute similarity between geographic units on the basis of the spatial adjacency relationship of geographic units to merge geographic units, model the merged geographic units as nodes in the graph model, and connect to the downstream forest fire risk prediction graph neural network model;

[0020] Step 2: Constructing a forest fire prediction graph neural network model to effectively capture the characteristics of risk factors and risk conditions of the target geographic unit and its neighboring geographic units, and to achieve end-to-end prediction of forest fire risks;

[0021] Step 3: Sample dataset construction and model optimization: Construct a sample dataset containing multi-source spatiotemporal data of risk factors including meteorology, topography, and land cover types and historical fire events for model training, verification, and testing.

[0022] The present invention has the following beneficial technical effects:

[0023] The present invention proposes a forest fire risk prediction method based on a graph neural network with variable area units. By comprehensively considering the attribute similarity and spatial adjacency between geographic units to merge geographic units, the geographic units are modeled as nodes in the graph model, and then connected to the downstream forest fire prediction graph neural network model, providing a basis for accurately predicting the future forest fire risk of the merged geographic units, forming a method that is different from dividing the target area geographic units in a regular grid manner. [9] New methods can better understand the laws of forest fire occurrence and provide a technical approach to achieve more accurate forest fire risk assessment.

[0024] The present invention proposes a graph neural network forest fire risk prediction method based on variable area units. By comprehensively considering the attribute similarity and spatial adjacency between geographic units, the geographic units are merged to model the geographic units as nodes in the graph model, and then connected to the downstream forest fire prediction graph neural network model, providing a basis for accurately predicting the future forest fire risks of the merged geographic units, forming a new method that is different from the regular grid method for dividing the target area geographic units, so as to better understand the laws of forest fire occurrence and provide a technical approach to achieve more accurate forest fire risk assessment.

[0025] The present invention proposes a forest fire risk prediction method based on a graph neural network with variable area units. By comprehensively considering the attribute similarity and spatial adjacency between geographic units to merge geographic units, the geographic units are modeled as nodes in the graph model, and then connected to the downstream forest fire prediction graph neural network model, providing a basis for accurately predicting the future forest fire risk of the merged geographic units, forming a method that is different from dividing the target area geographic units in a regular grid manner. [9] New methods can better understand the laws of forest fire occurrence and provide a technical approach to achieve more accurate forest fire risk assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A schematic diagram for dividing the target area for risk prediction into irregularly shaped geographical units;

[0027] Figure 2 Schematic diagram of the forest fire prediction graph neural network model structure. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in each embodiment of the present invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above-mentioned purpose, the present invention adopts the following technical scheme.

[0029] The present invention proposes a forest fire risk prediction method based on a graph neural network with variable area units, which specifically includes the following steps:

[0030] Step 1, graph node construction based on variable area units;

[0031] Meteorological data (temperature, dew point temperature, precipitation, wind component in the east-west direction, wind component in the north-south direction), topography (slope, slope aspect, elevation) and land cover type that are closely related to the occurrence of forest fires are collected as risk factor data.

[0032] Aspect data (°) According to formula (1)-(2), it is decomposed into the slope component in the east-west direction , the slope component in the north-south direction , so as to be subsequently mapped into the vector space.

[0033] (1)

[0034] (2)

[0035] One-Hot Encoding of land cover data. One-Hot Encoding, also known as One-of-Many-Encoding, is a method of converting categorical variables into a form that machine learning algorithms can better handle. This encoding method is particularly suitable for processing categorical data with a finite number of discrete values. Land cover data is usually categorical data, representing the types of different surface features, such as forests, grasslands, water bodies, etc. In one-hot encoding, each category value of the land cover data is converted into a binary vector with all positions being 0 except for one position representing the category, which is 1.

[0036] The data is normalized to the interval [0, 1], as shown in formula (3):

[0037] (3)

[0038] In the formula, is the normalized data, is the original data, is the minimum value in the data, is the maximum value in the data;

[0039] The multi-source data are spatially superimposed so that the target area for risk prediction is divided into irregularly shaped geographic units, such as Figure 1 shown.

[0040] Number the geographic units . Create a matrix representing the spatial adjacency relationship between geographic units .like and There is a spatial adjacency relationship. 1. If and There is no spatial adjacency relationship. 0; 2.

[0041] ;

[0042] Based on the spatial adjacency of geographic units, the attribute similarity between geographic units is further measured. When the value in 1 o'clock, calculate and The cosine similarity of is shown in formula (4). and Indicates Dimensional data In the geographical unit and The number 9 in the formula indicates that, in addition to the land cover type, there are 9 risk factors (elevation, slope, slope component in the east-west direction, slope component in the north-south direction, temperature, dew point temperature, rainfall, wind component in the east-west direction, and wind component in the north-south direction). Indicates the total number of land cover types Species, mapping land cover types to -dimensional binary vector.

[0043] (4)

[0044] Create a matrix representing the merged relationships between geographic units . Set attribute similarity threshold ,when hour, 1 if true; otherwise 0.

[0045] ;

[0046] Matrix-based Merge geographic units. 1, then and Merge into one geographic unit. Mapped as graph nodes.

[0047] Establish edges between spatially adjacent nodes to form an adjacency matrix indicating whether there are edges between nodes .like and There is a spatial adjacency relationship. 1. If and There is no spatial adjacency relationship. 0.

[0048] ;

[0049] Step 2: Construction of forest fire prediction graph neural network model;

[0050] The model structure is as follows Figure 2 As shown in the figure, the model includes an input layer, two graph convolution layers, and two fully connected layers. The input layer is the first layer of the neural network and is used to receive raw input data. In forest fire risk prediction, the input data includes historical risk factors. The input layer does not perform any calculations, but simply passes the data to the next layer. The graph convolution layer is used to process graph structure data. The inputs of the two graph convolution layers are matrices. In forest fire risk prediction, the graph convolution layer is used to capture the correlation between different geographical units. This layer updates the representation of nodes by aggregating and transforming the neighbor information of nodes, thereby learning the mutual influence between geographical units;

[0051] The calculation formulas (5)-(7) for graph convolution in GCN are as follows:

[0052] (5)

[0053] (6)

[0054] (7)

[0055] in, Indicates that the node is The feature vector of the layer, It means that after convolution, the node is The feature vector of the layer, Indicates The parameters of the layer convolution, represents the activation function. And the matrix The component is a Laplacian matrix, In is the identity matrix, that is, a matrix with 1s on the diagonal and 0s elsewhere. is a degree matrix, only the main diagonal elements are not zero; express Main diagonal elements; express Each element in; i represents the row number, j represents the column number.

[0056] The fully connected layer is used to map the output of the previous layer into a high-dimensional space, perform more complex nonlinear transformations, and provide richer information for the final prediction task;

[0057] Step 3: Sample data set construction and model optimization;

[0058] The risk factors and historical fire events (risk conditions) of the target area are collected to form a data set. The graph nodes are constructed according to step 1, and edges are established between spatially adjacent nodes. The model input is: The forest fire risk factors for geographic units; the adjacency matrix indicating whether there are edges between nodes; the model labels are: Forest fire risk in each geographic unit.

[0059] Construct samples based on model input and model label The samples are divided into training set, validation set and test set in a 6:2:2 ratio. The training set is used to train the model; the validation set is used to adjust the model hyperparameters; and the test set is used to test the final accuracy of the model.

[0060] When applying the model, collect data and build geographical unit in the future The feature matrix at time is input into the trained model for forward propagation to obtain geographical unit in the future Forest fire risk prediction results at each moment.

Claims

1. A forest fire risk prediction method based on graph neural network with variable area units, characterized in that: The method specifically comprises the following steps: Step 1: Graph node construction based on variable area units; collect meteorological, topographic and surface cover types that are closely related to the occurrence of forest fires as risk factor data, and design vector representation methods for different types of data, where meteorological includes temperature, dew point temperature, precipitation, and wind component in the east-west direction, and topographic includes slope, slope direction, and elevation; spatially superimpose multi-source risk factor data, divide the risk prediction target area into irregularly shaped geographic units, and further measure the attribute similarity between geographic units on the basis of the spatial adjacency relationship of geographic units to merge geographic units, model the merged geographic units as nodes in the graph model, and connect to the downstream forest fire risk prediction graph neural network model; Step 2: Constructing a forest fire prediction graph neural network model to effectively capture the characteristics of risk factors and risk conditions of the target geographic unit and its neighboring geographic units, and to achieve end-to-end prediction of forest fire risks; Step 3: Sample dataset construction and model optimization: Construct a sample dataset containing multi-source spatiotemporal data of risk factors including meteorology, topography, and land cover types, as well as historical fire events, for model training, verification, and testing. In step 1, the aspect data According to formula (1)-(2), it is decomposed into the slope component in the east-west direction: , the slope component in the north-south direction , so as to be subsequently mapped into the vector space; (1) (2) The land cover data is one-hot encoded. The land cover data is categorical data. The land cover data represents the types of different land features, including forests, grasslands, and water bodies. In one-hot encoding, each category value of the land cover data is converted into a binary vector. Except for one position representing the category, which is 1, the rest of the positions are 0. The data is normalized to the interval [0, 1], as shown in formula (3): (3) In the formula is the normalized data, is the original data, is the minimum value in the data, is the maximum value in the data; The multi-source data are spatially superimposed so that the target area for risk prediction is divided into irregularly shaped geographic units; Number the geographic units , establish a matrix representing the spatial adjacency relationship between geographic units ,like and There is a spatial adjacency relationship.

1. If and There is no spatial adjacency relationship. 0; 2; ； Based on the spatial adjacency of geographic units, we further measure the attribute similarity between geographic units and traverse them row by row and column by column. When the value in 1 o'clock, calculate and The cosine similarity of is shown in formula (4). and Indicates Dimensional data In the geographical unit and The number 9 in formula (4) indicates that, in addition to the land cover type, there are 9 risk factors, including elevation, slope, slope component in the east-west direction, slope component in the north-south direction, temperature, dew point temperature, rainfall, wind component in the east-west direction, and wind component in the north-south direction; Indicates the total number of land cover types Species, mapping land cover types to -dimensional binary vector; (4) Create a matrix representing the merged relationships between geographic units , set the attribute similarity threshold ,when hour, 1, otherwise 0; ； Matrix-based Merge geographic units. 1, then and merged into one geographic unit, and the merged geographic unit Mapped as graph nodes; Establish edges between spatially adjacent nodes to form an adjacency matrix indicating whether there are edges between nodes ,like and There is a spatial adjacency relationship.

1. If and There is no spatial adjacency relationship. 0; 。 2. The method according to claim 1, characterized in that: In step 2, the model structure includes an input layer, two graph convolutional layers, and two fully connected layers; the input layer is the first layer of the neural network, which is used to receive the original input data; in forest fire risk prediction, the input data includes historical risk factors; the input layer does not perform any calculations, but only passes the data to the next layer; the graph convolutional layer is used to process graph structure data, and the inputs of the two graph convolutional layers are matrices In forest fire risk prediction, the graph convolution layer is used to capture the correlation between different geographical units. This layer updates the representation of nodes by aggregating and transforming the neighbor information of nodes, thereby learning the mutual influence between geographical units; The calculation formulas (5)-(7) of graph convolution are as follows: (5) (6) (7) in, Indicates that the node is The feature vector of the layer, It means that after convolution, the node is The feature vector of the layer, Indicates The parameters of the layer convolution, represents the activation function, and the matrix The component is a Laplacian matrix, In is the identity matrix, that is, a matrix with 1 on the diagonal and 0 on the rest. is a degree matrix, only the main diagonal elements are not zero; express Main diagonal elements; express Each element in; i represents the row number, j represents the column number; The fully connected layer is used to map the output of the previous layer into a high-dimensional space, perform more complex nonlinear transformations, and provide richer information for the final prediction task.

3. A computer device comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, characterized in that: When the processor executes the program, the steps in the method according to any one of claims 1 to 2 are implemented.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps in the method according to any one of claims 1 to 2 are implemented.

Images