A method and system for simulating complex forest fire scenes based on VR technology

Abstract

The present invention discloses a method and system for deducing complex forest fire scenes based on VR technology, the method comprising the following steps: S1, building a virtual three-dimensional geographic scene according to the basic geographic information data of the target area; S2, dynamically displaying the three-dimensional dynamic model of flames in different time domains in the virtual three-dimensional geographic scene according to the output result of the fire spread simulation calculation analysis of the target area according to the complex scene fire spread analysis model, and dynamically simulating the spread of virtual fire; S3, using the virtual reality device interactive tool to perform combat mapping, mapping the constructed combat command mapping three-dimensional static model to the virtual three-dimensional geographic scene, and forming a mapping combat map. The method of the present invention supports immersive viewing of the forest fire spread process in complex scenes and supports combat mapping that conforms to human operating habits.

Description

Technical Field

[0001] The present invention belongs to the application of virtual reality (VR) technology in the field of fire prediction technology, and in particular to a complex forest fire scene deduction method and system based on virtual reality VR technology. Background Art

[0002] With the rapid development of urbanization, the interaction between human activities and the natural environment has become more frequent, and the frequency of fires at the junction of urban and forest areas has gradually increased, often causing huge losses of life and property and damage to the natural ecology. The fire spread prediction model can provide fire spread trend predictions, providing rich disaster predictions and scientific decision-making support for emergency command.

[0003] However, the existing fire spread prediction models are generally displayed based on a two-dimensional grid. This display method is not conducive to users to view the terrain more intuitively, and thus is not conducive to commanders making command decisions. At the same time, the traditional command system's combat mapping method is to map on a two-dimensional plane map, which is not intuitive enough. Summary of the invention

[0004] In order to make up for the deficiencies of the above-mentioned prior art, the present invention proposes a complex forest fire scene simulation method and system based on virtual reality VR technology.

[0005] The technical problem of the present invention is solved by the following technical solutions:

[0006] A complex forest fire scene simulation method based on VR technology includes the following steps:

[0007] S1. Build a virtual three-dimensional geographic scene based on the basic geographic information data of the target area;

[0008] S2, according to the complex scene fire spread analysis model, the output result of the fire spread simulation calculation analysis of the target area is dynamically displayed in the virtual three-dimensional geographic scene in different time domains. The three-dimensional dynamic model of the flame is dynamically simulated to dynamically simulate the virtual fire spread;

[0009] S3. Use the virtual reality device interactive tool to carry out combat mapping, map the constructed combat command mapping three-dimensional static model to the virtual three-dimensional geographic scene, and form a mapped combat map.

[0010] Preferably, the basic geographic information data of the target area includes a digital elevation model (DEM), remote sensing satellite images and surface structure data; the surface structure data is one or more of ground vegetation type data, river data, and building data.

[0011] Preferably, step S1 specifically includes:

[0012] Constructing a three-dimensional model of surface structures in the target area according to the basic geographic information data of the target area;

[0013] A basic three-dimensional elevation surface is constructed according to the basic geographic information data of the target area, and each pixel value of the remote sensing satellite image is mapped onto the three-dimensional elevation surface to generate a basic three-dimensional geographic scene with ups and downs, and the constructed three-dimensional model of the surface structure is placed in the corresponding surface area, thereby constructing a complete virtual three-dimensional geographic scene; preferably, before placing the surface structure in the corresponding surface area, the geographic information data of the surface structure is converted into coordinates so that it is in the same coordinate reference system as the DEM elevation data and the remote sensing image data.

[0014] Preferably, the three-dimensional model of the surface structures in the target area includes three-dimensional static models of ground vegetation and buildings and a three-dimensional dynamic model of rivers.

[0015] Preferably, the three-dimensional dynamic model of flame includes one or more of bonfire, building burning fire, chemical burning fire, dark fire, booming fire, forest fire, oil ring fire, and smoke fire.

[0016] Preferably, step S2 specifically includes: determining the combustible material data existing in the target area according to the basic geographic information data of the target area, analyzing the possible fire types in the target area according to the combustible material data, and constructing a three-dimensional dynamic model of flame that conforms to the fire characteristics of the target area; selecting a complex scene fire spread analysis model that can perform fire spread simulation calculation and analysis on the target area, and dynamically displaying the three-dimensional dynamic model of flame in different time domains in the virtual three-dimensional geographic scene according to the output result of the complex scene fire spread analysis model and in combination with the three-dimensional dynamic model of flame; preferably, the complex scene fire spread analysis model is a complex scene fire spread analysis model based on the principle of cellular automaton.

[0017] Preferably, in step S3, a model library of the combat command mapping three-dimensional static model is constructed according to the emergency forces, emergency materials and combat identifier lists required for use in the fire emergency command process.

[0018] Preferably, in step S3, the use of a virtual reality device interactive tool for combat mapping includes adding, moving, enlarging, reducing, and deleting a designated combat command mapping three-dimensional static model in the virtual three-dimensional geographic scene through human-computer interaction instructions.

[0019] Preferably, in step S3, the mapped combat map is an orthographically projected two-dimensional combat map.

[0020] A complex forest fire scene simulation system based on VR technology includes a processor and a computer-readable storage medium. When a computer program stored in the computer-readable storage medium is executed by the processor, the complex forest fire scene simulation method is implemented.

[0021] The beneficial effects of the present invention compared with the prior art include: the complex forest fire scene simulation method based on VR technology of the present invention includes three parts: virtual three-dimensional geographical scene construction, virtual fire spread dynamic simulation and virtual forest fire combat command mapping. It supports immersive viewing of the forest fire spread process in complex scenes and supports combat mapping that conforms to human operating habits. That is, through the collaboration of these three parts, people can view or experience the spread process of forest fires in complex scenes in a more immersive way, and can perform three-dimensional combat mapping in virtual scenes, providing rich disaster prediction information in a more immersive way and a more ergonomic combat mapping method to scientifically assist emergency command decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flowchart of a complex forest fire scenario simulation method according to a specific implementation mode of the present invention;

[0023] Figure 2 It is a schematic diagram of the process of building a virtual three-dimensional geographic scene in a specific example of the present invention;

[0024] Figure 3 It is a schematic diagram of the construction process of a virtual fire spread dynamic simulation in a specific example of the present invention;

[0025] Figure 4 It is a schematic diagram of the construction process of virtual forest fire combat command and mapping in a specific example of the present invention;

[0026] Figure 5 Schematic diagrams of eight typical three-dimensional dynamic models of flames in specific examples of the present invention;

[0027] Figure 6 It is a schematic diagram of a complex scenario fire spread simulation in a specific example of the present invention;

[0028] Figure 7 It is a virtual forest fire combat command plotting map in a specific example of the present invention;

[0029] Figure 8 It is a flow chart of a forest fire spread prediction method based on deep learning in a specific embodiment of the present invention;

[0030] Fig. 9 is a flow chart for generating a simulated forest fire spread data set in a specific example of the present invention;

[0031] Fig.10It is a comparison diagram of the WSSTF and the binary classification forest fire spread state representation mode in the specific example of the present invention;

[0032] Fig.11 is a structural diagram of a convolutional neural network for forest fire spread prediction in a specific example of the present invention;

[0033] Fig.12 is a simulation result diagram of the network application output in a specific example of the present invention;

[0034] Fig.13 It is a flow chart of a forest fire spread prediction method based on an irregular graph network in a specific implementation manner of the present invention;

[0035] Fig.14 is a schematic diagram of an initialized irregular graph network in a specific example of the present invention;

[0036] Fig.15 is a schematic diagram of adaptive optimization of irregular graph networks in a specific example of the present invention;

[0037] Fig.16 is a schematic diagram of effective spreading nodes VSNs and effective spreading edges VSEs in a specific example of the present invention;

[0038] Fig.17 is a schematic diagram of a graph-grid mapping method in a specific example of the present invention;

[0039] Fig.18 It is a schematic diagram for visualizing simulation results based on an irregular graph network in a specific example of the present invention. DETAILED DESCRIPTION

[0040] The present invention is further described below with reference to the accompanying drawings and in combination with preferred embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0041] The specific implementation of the present invention provides a complex forest fire (i.e., forest fire) scene simulation method based on VR technology, such as Figure 1 As shown, it includes three parts: virtual three-dimensional geographic scene construction, virtual fire spread dynamic simulation, and virtual forest fire combat command mapping, which specifically include the following steps:

[0042] S1. Construction of virtual three-dimensional geographic scene: Build a virtual three-dimensional geographic scene based on the basic geographic information data of the target area.

[0043] Step 1.1: Geographic scene data collection

[0044] Collect basic geographic information data of the target area, including: digital elevation model (DEM), remote sensing satellite images, and data on ground vegetation types, river data, building data and other surface structures. Among them, digital elevation model (DEM) and remote sensing satellite images should have a high resolution so that they can be used to build a virtual three-dimensional geographic scene with high definition; ground vegetation type data, river data, building data and other surface structure data should have geographic coordinate information.

[0045] Step 1.2: Construction of virtual 3D geographic scene of target area

[0046] Constructing a three-dimensional model of surface structures in the target area according to the basic geographic information data of the target area; wherein the three-dimensional model of surface structures in the target area includes three-dimensional static models of ground vegetation and buildings and three-dimensional dynamic models of rivers;

[0047] A basic three-dimensional elevation surface is constructed based on the basic geographic information data of the target area, and each pixel value of the remote sensing satellite image is mapped onto the three-dimensional elevation surface to generate a basic three-dimensional geographic scene with ups and downs, and the constructed three-dimensional model of the surface structure is placed in the corresponding surface area, thereby constructing a complete virtual three-dimensional geographic scene; it is more preferred that the geographic information data of the surface structure is converted into coordinates before placing the surface structure in the corresponding surface area, so that it is in the same coordinate reference system as the DEM elevation data and the remote sensing image data.

[0048] Specifically, if Figure 2 As shown in the figure, in one example, to build a virtual 3D geographic scene of the target area, the DEM elevation data and remote sensing satellite image data are first preprocessed, such as image radiation correction and geometric correction, and the coordinate reference system of the data is unified (i.e., the unified reference geographic coordinate system), such as the WGS 1984 coordinate system, and then the basic 3D elevation surface is constructed through the DEM elevation data, and each pixel value of the remote sensing satellite image is mapped to the 3D elevation surface, thereby generating a basic 3D geographic scene with ups and downs. At present, most 3DGIS software supports the construction function of virtual 3D geographic scenes, such as Skyline, ArcGIS, Unity3D, etc., and does not require manual programming to realize data mapping and 3D display functions. This example uses Unity3D as an example.

[0049] Construct three-dimensional static models of ground vegetation and buildings and three-dimensional dynamic models of rivers, perform pre-processing operations such as projection coordinate conversion, resampling, and clipping on the collected data, and unify the coordinate reference system of the data, that is, use three-dimensional modeling tools, such as MAYA, 3Dmax and other software to perform three-dimensional modeling on surface structures such as vegetation, rivers, and buildings, and perform coordinate conversion on geographic information data of surface structures such as vegetation type data, river data, and building data, so that they are in the same coordinate reference system as DEM elevation data and remote sensing satellite image data; place the constructed three-dimensional models of surface structures (such as ground vegetation, three-dimensional static models of buildings, and three-dimensional dynamic models of rivers) in the corresponding surface area to form a complete virtual three-dimensional geographic scene.

[0050] S2. Dynamic simulation of virtual fire spread: Based on the output results of fire spread simulation calculation and analysis of the target area according to the complex scene fire spread analysis model, the three-dimensional dynamic model of flames in different time domains is dynamically displayed in the virtual three-dimensional geographic scene to dynamically simulate the virtual fire spread.

[0051] Step 2.1: Construction of flame 3D dynamic model

[0052] According to the basic geographic information data of the target area, the combustible data in the target area is determined, and the possible fire types in the target area are analyzed based on the combustible data, and a three-dimensional dynamic model of the flame that conforms to the fire characteristics of the target area is constructed.

[0053] Step 2.2: Construction of dynamic simulation scene of fire spread in complex scenarios

[0054] A complex scene fire spread analysis model that can perform fire spread simulation and calculation analysis on the target area is selected. Based on the output results of the complex scene fire spread analysis model and the flame three-dimensional dynamic model constructed in step 2.1, the flame three-dimensional dynamic model in different time domains is dynamically displayed in the virtual three-dimensional geographic scene, thereby completing the construction of the complex scene fire spread dynamic simulation scene.

[0055] Among them, the better one is the complex scene fire spread analysis model based on the principle of cellular automaton. This type of model is based on the principle of cellular automaton. It divides the vegetation type raster data according to the concept of cells, and then performs dynamic simulation based on data information such as wind speed, humidity, temperature, vegetation type, etc., and outputs offline data including the fire point and spread status.

[0056] like Figure 3As shown in FIG. 1 , in a specific example, based on the geographic scene data collection, the types of ground combustibles that exist are determined, and based on the types of ground combustibles, the types of fires that may exist in the target area (such as forest fires, residential building fires, chemical plant fires, etc.) are analyzed, and a three-dimensional dynamic model of flames that conforms to the characteristics of the fires in the target area is constructed. Figure 5 As shown, the three-dimensional dynamic model of flame includes one or more of bonfire, building burning fire, chemical burning fire, dark fire, roaring fire, forest fire, oil ring fire, and smoke fire. Specifically, based on the basic principle of particle system, many three-dimensional software have developed particle system modules to simulate the three-dimensional effect of flame. For example, Unity3D, as a powerful three-dimensional application development platform, supports detailed editing of particle effects, including the number of particles, initial particle speed, particle shape change, color gradient and particle period, etc. Users can fully customize all properties of particles in the main interface of Unity3D; by analyzing the geographical scene data, analyze the possible fire types in the target area, such as forest fire, residential building fire, chemical plant fire, etc., and build a three-dimensional dynamic model of flame that meets the characteristics of fire in the target area based on Unity3D's particle system.

[0057] like Figure 3 As shown, a complex scene fire spread analysis model based on the principle of cellular automata that can perform fire spread simulation and analysis on the target area is selected, and the calculation result data of the complex scene fire spread analysis model is connected. According to its output results, combined with the flame three-dimensional dynamic model constructed in the previous step, the flame three-dimensional dynamic model in different time domains is dynamically displayed in the virtual three-dimensional geographic scene. Specifically, first, the calculation result file of the complex scene fire spread analysis model is read through the Unity3D software, and the fire point information, time information, and fire status information in the file are read to verify that its coordinate system is consistent with the virtual three-dimensional geographic scene in which the target area is located; then, the base map in the virtual three-dimensional geographic scene is divided into grids of equal size according to the size of the grid in the calculation result file; finally, the flame three-dimensional dynamic model is placed in different areas at different times in the virtual three-dimensional geographic scene to form a dynamic flame model display animation in the virtual three-dimensional space, as shown in FIG. Figure 6 shown.

[0058] S3. Virtual forest fire combat command mapping: Use virtual reality device interactive tools to perform combat mapping, map the constructed combat command mapping three-dimensional static model to the virtual three-dimensional geographic scene, and form a mapping combat map.

[0059] Step 3.1: Build a 3D static model library for combat command and plotting

[0060] According to the list of emergency forces, emergency materials and combat identifiers needed in the fire emergency command process, a model library of combat command mapping three-dimensional static models is constructed.

[0061] Step 3.2: Design dynamic interaction of virtual reality combat command and plotting

[0062] Use virtual reality device interactive tools for combat mapping, including adding, moving, enlarging, reducing, and deleting specified combat command mapping three-dimensional static models in the virtual three-dimensional geographic scene through human-computer interaction commands. After completing the combat command mapping, the version can be saved as shown, thereby completing the dynamic interactive design of virtual reality combat command mapping.

[0063] Step 3.3: Output the combat command and plotted combat map (the plotted combat map is an orthographic projection and plotted two-dimensional combat map)

[0064] According to the actual needs of emergency command work, based on the different versions of combat command mapping saved, the positions of different combat command mapping three-dimensional static models are mapped to the virtual three-dimensional geographic scene to form an orthographic projection two-dimensional combat map and export the offline file.

[0065] like Figure 4 and Figure 7 As shown, in one example, emergency force data, emergency material data and combat mapping identifiers required for the emergency command process of complex fire scenes are collected and sorted, including firefighters, fire trucks, assembly points, etc., to build a combat command mapping three-dimensional static model library. For example, 3Dmax software can be used for modeling to form a three-dimensional model file that can be used in Unity3D. Through Unity3D software development, virtual reality devices such as HTC vivePro helmets and handles can be used to develop interactive mapping systems, which support adding, moving, enlarging, reducing, and deleting specified combat command mapping three-dimensional static models in virtual three-dimensional geographic scenes by dragging, clicking, etc. After completing the combat command mapping, the version can be saved. According to the actual emergency command needs, based on the different versions of combat command mapping saved, the positions of different combat command mapping three-dimensional static models are mapped to the virtual three-dimensional geographic scene to form an orthographic projection mapping two-dimensional combat map and export offline files. In this example, the emergency command combat mapping based on virtual reality technology can be carried out by directly dragging in three-dimensional space, which is more in line with the user's usage habits.

[0066] A specific embodiment of the present invention also provides a complex forest fire scene simulation system based on VR technology, including a processor and a computer-readable storage medium. When the computer program stored in the computer-readable storage medium is executed by the processor, the complex forest fire scene simulation method in the above-mentioned specific embodiment or specific example is implemented.

[0067] Forest fire spread prediction method and prediction device based on deep learning

[0068] Most existing forest fire spread prediction models are based on small-scale combustion experiments to build derivation models and generate model parameters, and the model application has high requirements for boundary conditions such as the ignition point. However, the scene diversity of the combustion experiment is limited, the constructed model has high computational complexity, and it is difficult to obtain accurate model boundary conditions at the forest fire site, resulting in the existing models being difficult to meet the actual application requirements in terms of scene stability, simulation accuracy and computational complexity.

[0069] In order to make up for the deficiencies of the above-mentioned prior art, the present invention also proposes a forest fire spread prediction method and prediction device based on deep learning. The aforementioned complex scene fire spread analysis model of the present invention can be a model using the forest fire spread prediction method and prediction device based on deep learning.

[0070] A forest fire spread prediction method based on deep learning includes the following steps:

[0071] Train a convolutional neural network using the constructed forest fire spread dataset;

[0072] The convolutional neural network predicts the ignition time of each grid point in the grid-based forest fire space according to input data; wherein the input data includes forest fire spread status, geographic data and meteorological data, and the convolutional neural network encodes and decodes the forest fire spread spatiotemporal distribution field in the prediction area according to the input data, and uses the forest fire spread spatiotemporal distribution field to predict the ignition time of each grid point in the forest fire space.

[0073] Preferably, the spatiotemporal distribution field of forest fire spread is defined as:

[0074] WSSTF=Net(θ;Fs,Geo,Meo)

[0075] Among them, WSSTF represents the spatiotemporal distribution field of forest fire spread, describing the ignition time of each grid point, Fs represents the burning and spreading state of forest fire, Geo represents geographic data, Meo represents meteorological data, Net represents convolutional neural network, θ is the parameter set of the network, and θ is generated after the network is trained and converged:

[0076]

[0077] in, Represents the loss function.

[0078] Preferably, the pytorch framework is used to build the network and the L1 loss function is used to train the network until convergence.

[0079] Preferably, when performing network prediction on the forest fire spread boundary at a specific time T, a threshold filter is used to filter out the area where the ignition time value in WSSTF is less than T, and the forest fire spread boundary wsstf at time T is generated according to the area where the ignition time value is less than T. T .

[0080] Preferably, the forest fire spread dataset includes forest fire thematic data, basic geographic data and meteorological data; the forest fire thematic data includes vector type fire point and spread boundary data, the basic geographic data includes raster type digital elevation model DEM and vegetation category data, and the meteorological data includes temperature, humidity, wind speed and wind direction.

[0081] Preferably, the construction of the forest fire spread dataset includes: converting the data into a unified projection coordinate system, and performing resampling, clipping and filtering.

[0082] Preferably, the forest fire spread dataset includes a historical forest fire dataset and a simulation dataset generated by simulating using a traditional forest fire spread engine.

[0083] Preferably, the simulation data set is obtained in the following manner: first, the prediction area is cut and cropped to construct a set of geographical cuts; for the combination of geographical cuts, a random function is used to generate multiple groups of fire points and meteorological parameters and the forest fire spread engine is used for simulation to generate multiple groups of forest fire spread simulation results, which constitute the key part of the forest fire spread data set.

[0084] Preferably, the encoding is a hierarchical progressive encoding of the input data through a state-condition mechanism; wherein the input data is divided into a state branch and a conditional branch, the state branch is the state data of the forest fire spread, and the conditional branch is the environmental parameters acting on the forest fire spread state, including a geographic coding sub-branch and a meteorological coding sub-branch; each branch is first independently encoded and then deeply fused to form a hybrid feature, and then a backbone network coding of the encoding-decoding structure is adopted to further extract higher-order features of the hybrid feature, and decoded into the spatiotemporal distribution field of the forest fire spread, thereby realizing end-to-end rapid inference and prediction of forest fire spread.

[0085] Preferably, the input of the state branch is the spreading state value at a certain moment, and a series of Convolution layers, Batch Normalization layers, Relu layers, i.e., a series of CBR layers are used to extract the primary state features;

[0086] The geocoding sub-branch is composed of a DEM coding layer and a vegetation type coding layer; the DEM coding layer is composed of the series of CBR layers and can generate a primary feature tensor of the DEM; the vegetation type coding layer first uses a one-hot coding layer to convert the vegetation category code into a category vector, and then generates a vegetation type tensor from the series of CBR layers;

[0087] The meteorological coding sub-branch is composed of a wind vector coding layer and a temperature and humidity coding layer; the wind vector coding layer first uses a Linear layer to map the wind vector to a high dimension to generate a wind vector tensor, and upsamples it to the DEM primary tensor scale; the temperature and humidity coding layer uses a Linear layer to map the temperature and humidity vector to a high dimension, and upsamples it to the DEM primary tensor scale;

[0088] The features extracted by the conditional branch are fused to form a conditional coding feature tensor, and are merged with the state feature tensor of the state branch;

[0089] The merged features are input into the backbone network for deep encoding and decoding to output the predicted spatiotemporal distribution field of forest fire spread WSTTF.

[0090] A forest fire spread prediction device based on deep learning includes a processor and a computer-readable storage medium. When a computer program stored in the computer-readable storage medium is executed by the processor, the forest fire spread prediction device is implemented.

[0091] The beneficial effects of the present invention compared with the prior art include: the forest fire spread prediction method based on deep learning of the present invention transforms the forest fire spread prediction problem into a spatiotemporal distribution field prediction problem, adopts a deep learning convolutional neural network to realize the prediction of forest fire spread behavior in a complex spatial environment, realizes the end-to-end rapid deduction of forest fire spread and solves the restrictions of the model on constraints such as the ignition point. Specifically, the convolutional neural network designed based on deep learning theory can learn model parameters from the forest fire spread data set, reduce the model complexity and reduce the model boundary condition constraints, realize efficient end-to-end prediction of the spatiotemporal dynamic evolution of forest fire spread, and provide scientific auxiliary decision support for emergency management departments to carry out forest fire rescue.

[0092] The present invention is further described below with reference to the accompanying drawings and in combination with preferred embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0093] like Figure 8 As shown, a specific embodiment of the present invention provides a forest fire (i.e., forest fire) spread prediction method based on deep learning, comprising the following steps:

[0094] Train a convolutional neural network using the constructed forest fire spread dataset;

[0095] The convolutional neural network predicts the ignition time of each grid point in the grid-based forest fire space according to the input data; wherein the input data includes the forest fire spread status, geographic data and meteorological data, and the convolutional neural network encodes and decodes the forest fire spread spatiotemporal distribution field in the prediction area according to the input data, and uses the forest fire spread spatiotemporal distribution field to predict the ignition time of each grid point in the forest fire space.

[0096] Specifically, the forest fire spread prediction method includes two parts: forest fire spread dataset construction and convolutional neural network design and application, including:

[0097] 1. Construction of forest fire spread dataset: including converting data into a unified projection coordinate system, and performing resampling, cropping, and filtering.

[0098] The spread of forest fires is related to many factors such as geographical environment and meteorological conditions. The forest fire spread dataset is a dataset that records the entire process of forest fire occurrence and spread and related elements.

[0099] In some embodiments, the forest fire spread dataset includes forest fire thematic data, basic geographic data and meteorological data, wherein the forest fire thematic data includes vector type fire point and spread boundary data, etc., the basic geographic data includes raster type digital elevation model DEM (digital elevation model) and vegetation category data, etc., and the meteorological data includes temperature, humidity, wind speed and wind direction, etc.

[0100] The existing publicly available historical data on the spread of forest fires can provide some real data for the construction of the data set, but at the same time, considering the limited amount of real sample data, the traditional forest fire spread engine can also be used to batch generate large-scale simulated spread data. The fire points and meteorological data in the simulation process can be randomly generated, and the simulation area is not limited to areas where forest fires have occurred in history, so as to increase the diversity of the forest fire spread data set. Therefore, in other embodiments, the forest fire spread data set includes a historical forest fire data set and a simulated data set generated by simulating using a traditional forest fire spread engine; wherein the simulated data set is obtained in the following manner: first, the prediction area is cut and cropped to form a set of geographic cuts; for the combination of geographic cuts selected from the set of geographic cuts, a random function is used to generate multiple groups of fire points and meteorological parameters, and the forest fire spread engine is used for simulation to generate multiple groups of forest fire spread simulation results, which constitute the key part of the forest fire spread data set.

[0101] The collected forest fire spread data were converted into a unified projection coordinate system, and then resampled, clipped, and filtered to construct a forest fire spread dataset.

[0102] 2. Convolutional neural network design and application: The convolutional neural network predicts the ignition time of each grid point in the grid-based forest fire space according to the input data. The convolutional neural network encodes and decodes the spatiotemporal distribution field of forest fire spread in the predicted area based on the input data, and uses the spatiotemporal distribution field of forest fire spread to predict the ignition time of each grid point in the forest fire space.

[0103] 2.1 Network Task Design

[0104] Network task design is the key to determining the performance of convolutional neural networks, focusing on the definition of network input and prediction output. Considering that forest fire spread prediction is to predict the time when each grid point in the forest fire space is ignited, the input data of this network includes forest fire spread status data (such as fire point), geographic data (such as DEM, vegetation type VT, etc.) and meteorological data (temperature, humidity, wind speed, wind direction, etc.). Network output is the key to network task design. Compared with setting a binary prediction task, the embodiment of the present invention defines the concept of wildfire spread spatio-temporal field (WSSTF), and the convolutional neural network directly predicts the ignition time of each point in the space, where the forest fire spread spatio-temporal field is defined as:

[0105] WSSTF=Net(θ;Fs,Geo,Meo)

[0106] Among them, WSSTF represents the spatiotemporal distribution field of forest fire spread, describing the ignition time of each grid point, Fs represents the burning and spreading state of forest fire, Geo represents geographic data, Meo represents meteorological data, Net represents convolutional neural network, and θ is the parameter set of Net. When performing network prediction on the forest fire spread boundary at a specific time T, the threshold filtering is used to filter out the areas with ignition time values ​​less than T in WSSTF, and the forest fire spread boundary wsstf at time T is generated according to the areas with ignition time values ​​less than T. T .

[0107] 2.2 Network structure design

[0108] The encoding mode of the input data determines the performance of the network to a large extent. Simple stacked encoding input parameters are prone to feature confusion, making it difficult for the network to converge during training. Therefore, an embodiment of the present invention proposes a state-condition mechanism to achieve hierarchical progressive encoding of the input data. The method first divides the input data of the network into a state branch and a conditional branch; wherein the state branch refers to the state data of the forest fire spread, and the conditional branch refers to the environmental parameters acting on the forest fire spread state, including a geographic coding sub-branch and a meteorological coding sub-branch; each branch is first independently encoded and then deeply fused to form a hybrid feature, and then the backbone network coding of the encoding-decoding structure is adopted to further extract higher-order features of the hybrid feature, and decoded into the forest fire spread spatiotemporal distribution field WSSTF, thereby achieving end-to-end rapid inference and prediction of forest fire spread. Furthermore, the input of the state branch is the spread state value at a certain moment, and a series of Convolution layers, Batch Normalization layers, Relu layers, i.e., a series of CBR layers, are used to extract the primary state features; the geographic coding sub-branch is composed of a DEM coding layer and a vegetation type coding layer; the DEM coding layer composed of a series of CBR layers can generate the primary feature tensor of the DEM; the vegetation type coding layer first uses a one-hot coding layer to convert the vegetation category code into a category vector, and then generates a vegetation type tensor by a series of CBR layers; the meteorological coding sub-branch is composed of a wind vector coding layer and a temperature and humidity coding layer; the wind vector coding layer first uses a Linear layer to map the wind vector to a high dimension to generate a wind vector tensor, and upsamples it to the DEM primary tensor scale; the temperature and humidity coding layer uses a Linear layer to map the temperature and humidity vector to a high dimension, and upsamples it to the DEM primary tensor scale; the features extracted by the conditional branch are fused to form a conditional coding feature tensor, which is merged with the state feature tensor of the state branch; the merged features are input into the backbone network for deep encoding and decoding to output the predicted spatiotemporal distribution field of forest fire spread WSTTF.

[0109] 2.3 Network training and application

[0110] The label data in the network training process comes from the forest fire spread dataset. After training convergence, the network parameters θ can be generated. The better method is to use the pytorch framework to build the network and use the L1 loss function to train the network until convergence.

[0111]

[0112] When used on the Internet, the burning and spreading state of the forest fire represented by Fs can be the state at the ignition point, or any other burning and spreading state at any time. This can solve the limitation of the traditional spread model on the requirement of accurate ignition point. At the same time, the model deduction process does not need to be iterated step by step, and can realize end-to-end one-time deduction of forest fire spread spatiotemporal dynamic data.

[0113] The present invention is further described below in conjunction with specific examples.

[0114] In this example, the forest fire spread prediction method based on deep learning includes the following steps:

[0115] 1. Construction of forest fire spread dataset

[0116] 1.1 Fire spread data collection and processing

[0117] Taking California as an example, we collected basic data for forest fire simulation and historical forest fire data in the area, including raster-type DEM, raster-type vegetation type data, vector-type fire point and spread boundary data, etc. Then, we used spatial data processing tools to convert the data to a unified projection coordinate system, and performed data processing operations such as resampling, clipping, and filtering.

[0118] 1.2. Construction of forest fire spread dataset

[0119] The forest fire spread dataset includes forest fire thematic data (fire point, spread boundary data, etc.), basic geographic data (DEM, vegetation type, etc.), and meteorological data (temperature, humidity, wind speed, wind direction, etc.). Real forest fire data is very important for dataset construction. For example, the GeoMAC dataset that can be downloaded from the USGS official website contains forest fire spread data across the United States for many years. However, the scene diversity and data integrity of real forest fire data are defective, so it is better to use simulation data to enhance the usability of the dataset.

[0120] like Fig. 9 As shown in the figure, it is a flow chart for generating a simulated forest fire spread dataset. For the simulated dataset, the DEM, vegetation type and other raster data of the study area are firstly cut, blocked and filtered to construct a set of geographic blocks. For each combination of geographic blocks, a random function is used to generate multiple groups of fire points and meteorological parameters, and the traditional forest fire spread engine is used for simulation to generate multiple groups of forest fire spread simulation results, which constitute the key part of the forest fire (wildfire) spread dataset.

[0121] 2. Convolutional Neural Network Design and Application

[0122] 2.1 Network Task Design

[0123] like Fig.10The figure shows a comparison of the forest fire spread state representation mode of WSSTF and binary classification. Compared with the common binary classification task (i.e. predicting the burned 1 and unburned 0 in the forest fire scene), this example sets the network task as a dense regression task. The network will encode and decode the spatiotemporal distribution field WSSTF of forest fire spread in the prediction area based on the input forest fire burning spread state Fs, geographic data Geo, and meteorological data Meo. Among them, WSSTF describes the ignition time of each grid point. The forest fire spread boundary at any time can be obtained through simple threshold screening. For example, when the forest fire spread boundary at a specific time T is to be obtained, a simple threshold filter can be used to filter out the area with the ignition time value less than T in wsstf, and the corresponding forest fire spread boundary wsstf at time T can be generated. T , while the traditional two-classification model contains less data information, has difficulty in network convergence, and requires continuous iteration to obtain the forest fire spread boundaries at different times.

[0124] 2.2 Network structure design

[0125] like Fig.11 The figure shows the structure of the convolutional neural network for forest fire spread prediction. Based on the end-to-end dense regression network task, the state-condition mechanism is used to implement hierarchical progressive encoding of the input data to solve the problem that simple stacking encoding leads to feature confusion and network convergence difficulties. The mechanism classifies the input data into state branches and conditional branches by category.

[0126] The input of the state branch is the spreading state value at a certain moment (such as 0 / 1 burning binary data), and a series of Convolution layers + BatchNormalization layers + Relu layers (CBR layers) are used to extract the primary state features. The conditional branch contains geographic coding sub-branches and meteorological coding sub-branches. The geographic coding sub-branch consists of a DEM coding layer and a vegetation type coding layer, and a series of CBR layers constitute the DEM coding layer, which can generate the primary feature tensor of the DEM. The vegetation type coding layer first uses a one-hot coding layer to convert the vegetation category code into a category vector, and then generates a vegetation type tensor by a series of CBR layers. The meteorological coding sub-branch consists of a wind vector coding layer and a temperature and humidity coding layer. The wind vector coding layer first uses a Linear layer to map the wind vector to a high dimension to generate a wind vector tensor, and then upsamples it to the DEM primary tensor scale. The temperature and humidity coding layer directly uses a Linear layer to map the temperature and humidity vector to a high dimension, and upsamples it to the DEM primary tensor scale. The features extracted by the conditional sub-branch are fused to form a conditional coding feature tensor and merged with the state feature tensor. For example, the shape (scale) of the DEM tensor is NxNxC, while the wind feature and temperature and humidity feature are outputted in a scale of 1x1xC after the Linear layer. Therefore, they are upsampled and expanded to NxN dimensions. The scale of the upsampled tensors is consistent to achieve the splicing of NxNx (4C) fusion features. The merged features are input into the encoding-decoding backbone network for deep encoding and decoding to output the predicted spatiotemporal distribution field of forest fire spread WSTTF.

[0127] 2.3 Network training and application

[0128] The labeled data for training the network comes from the forest fire spread dataset. For example, the pytorch framework can be used to build the network and the L1 loss function can be used to train the network until convergence, and the trained network parameters θ can be saved.

[0129]

[0130] When the network is used, the burning and spreading state Fs, geographic data Geo, and meteorological data Meo of the forest fire in the prediction area are input into the Net network, and the network infers the spatiotemporal distribution field WSSTF of forest fire spread based on the design framework and training parameters θ.

[0131] WSSTF=Net(θ;Fs,Geo,Meo)

[0132] The network-predicted WSSTF describes the dynamic process of forest fire spread over the entire period. When obtaining the forest fire spread boundary at a specific time T, a simple threshold filter can be used to filter out the area in the WSSTF where the ignition time value is less than T, and the corresponding forest fire spread boundary wsstf at time T can be generated. T .

[0133] Fig.12 The simulation prediction results of the network application output in the specific example of the present invention are shown.

[0134] A specific embodiment of the present invention also provides a forest fire spread prediction device based on deep learning, including a processor and a computer-readable storage medium. When a computer program stored in the computer-readable storage medium is executed by the processor, the forest fire spread prediction device is implemented.

[0135] Forest fire spread prediction method and device based on irregular graph network

[0136] Existing forest fire spread prediction models mostly use regular grid patterns to represent geographic space, and calculate the ignition time of each grid based on forest fire spread rules. However, in forest fire scenes, combustible vegetation and non-combustible urban facilities are usually closely intertwined. It is usually difficult to set a suitable spatial resolution for forest fire spread models in regular grid patterns. Too high a resolution often results in a waste of computing resources and a long simulation time; too low a resolution easily leads to large simulation errors.

[0137] In order to make up for the deficiencies of the above-mentioned prior art, the present invention proposes a forest fire spread prediction method and device based on irregular graph network. The aforementioned complex scene fire spread analysis model of the present invention can be a model using the forest fire spread prediction method and device based on irregular graph network.

[0138] A forest fire spread prediction method based on irregular graph network includes the following steps:

[0139] S1. Generate an irregular graph network that characterizes a forest fire scene in a variable scale mode, wherein the irregular graph network includes a series of graph nodes and graph edges, and the graph nodes and the graph edges have a set of attributes set in combination with the characteristics of the forest fire scene; wherein, after initializing the irregular graph network, the number and spatial distribution of the graph nodes are adaptively determined according to the complexity of the forest fire scene;

[0140] S2, using a fitting function to approximate the forest fire spread function of the regular grid of the forest fire scene, to construct the forest fire spread function under the irregular graph network;

[0141] S3. Starting from the graph node corresponding to the fire starting point, the forest fire spread duration of the graph edge on the irregular graph network is calculated according to the forest fire spread function, and the ignition time of the graph node on the irregular graph network is calculated in combination with the minimum travel time principle, and the forest fire spread simulation result under the irregular graph network is iteratively generated.

[0142] Preferably, in the irregular graph network, each graph node has an attribute set A node :

[0143] A node ={ID,X,Y,LT,EL,IT}

[0144] Among them, ID is the unique number of the graph node, (X,γ) is the geographic coordinates of the graph node, LT is the vegetation type, EL is the elevation value, and IT is the relative time when the graph node is ignited;

[0145] Each graph edge has a set of attributes A describing the line segments between two adjacent nodes. edge :

[0146] A edge ={SID,TID,LT,SL,L,DA}

[0147] Among them, SID and TID are the starting point and target point of the edge respectively, LT is the vegetation type, SL is the slope from SID to TID, L is the geographic length of the edge, and DA is the edge angle, which describes the angle between the vector edge and the true north direction.

[0148] Preferably, the initialization of the irregular graph network includes:

[0149] Set the resolution range of the irregular graph network, including the maximum graph resolution threshold GR max and the minimum image resolution threshold GR min ;

[0150] By uniformly distributing samples, the graph nodes in the forest fire scene are initialized. The maximum distance between the graph nodes is GR min ;

[0151] Build graph edges on the initialized graph node set so that the graph edges can completely cover the forest fire space and there is no overlap between the graph edges;

[0152] Generate a set of graph node attributes and a set of graph edge attributes based on the geographic locations of the graph nodes.

[0153] Preferably, based on the initialized irregular graph network, the density of graph nodes is adaptively adjusted according to the complexity of the forest fire scene to perform adaptive optimization of the irregular graph network.

[0154] Preferably, the irregular graph network adaptive optimization includes:

[0155] For any graph edge in the irregular graph network, a candidate node set for pre-interpolating graph nodes for the graph edge is generated according to an interpolation interval determined by a maximum graph resolution of the irregular graph network.

[0156] Perform graph edge consistency check. When the candidate nodes in the candidate node set If the vegetation type is different from that of the previous node, insert the candidate node; and calculate the elevation difference change between the candidate node and the previous node when When the set threshold T is exceeded, the candidate node is inserted;

[0157] Reconstruct the irregular graph network based on the updated graph nodes;

[0158] Repeat the above process until no new candidate nodes are inserted or the preset maximum number of iterations is reached;

[0159] Delete the graph nodes whose adjacent nodes are all non-flammable attributes to form an optimized irregular graph network.

[0160] Preferably, the fitting function is used To approximate the spreading function f of a regular grid:

[0161] t=f(LT,SL,L,Meteo)

[0162]

[0163] Among them, LT, SL, L and Meteo are vegetation type, slope, grid length and meteorological parameters respectively, and t is the spreading time under the regular grid representation. is the spreading time under the representation of irregular graph network, It is calculated The loss function between and f.

[0164] Preferably, the fitting function is a multivariate linear equation, a multivariate nonlinear equation or a deep neural network; the model parameter θ of the fitting function is:

[0165]

[0166] By fitting the function Implement the calculation of ignition time in graph edge mode:

[0167]

[0168] Preferably, step S2 comprises: converting the forest fire spread result generated by the regular grid model into a label result under the irregular graph network by a graph-grid mapping method, and generating the parameter θ of the fitting function;

[0169] Wherein, the graph-grid mapping method comprises:

[0170] A regular grid model is used to generate forest fire spread simulation results, with each grid point recording the time when it was ignited;

[0171] Spatially align the irregular graph network corresponding to the forest fire area with the regular grid, and copy and map the ignition time of the grid point to the irregular graph node;

[0172] Retrieve all adjacent nodes for each graph node And get the ignition time of all adjacent nodes Then determine the adjacent node whose ignition time is the smallest relative to the ignition time of the graph node, and set the adjacent node as the effective spreading node VSN of the graph node, and the corresponding connecting graph edge as the effective spreading edge VSEs;

[0173] The parameters θ of the fitting function are generated by training using the label data of the effective spreading edges VSEs.

[0174] Preferably, step S3 comprises:

[0175] Search for the nearest graph node based on the ignition point coordinates and set it as the starting ignition time;

[0176] Search for the adjacent graph nodes N of the fire graph node i With the adjacency graph edge E i , according to the fitting function Calculate the time from the fire graph node to the adjacent graph node Will Add to candidate ignition set T edge ;

[0177] According to the minimum travel time principle, from the set T edge Select the graph node with the smallest ignition time as the next fitting function The calculated node ignition time is updated to T edge middle;

[0178] The above process is iterated to generate the forest fire spread simulation results under the irregular graph network.

[0179] A forest fire spread prediction device based on an irregular graph network comprises a processor and a computer-readable storage medium. When a computer program stored in the computer-readable storage medium is executed by the processor, the forest fire spread prediction method is implemented.

[0180] The beneficial effects of the present invention compared with the prior art include: the forest fire spread prediction method based on irregular graph network of the present invention generates irregular graph network with adaptive variable scale, and can dynamically adjust the spatial distribution density of graph nodes according to the complexity of regional scenes. The variable scale spatial representation mode can better balance the computational complexity and simulation accuracy, simulate the forest fire spread behavior in a complex spatial environment, and visualize the spatiotemporal process of forest fire spread with irregular graph network mode. At the same time, compared with the grid mode, the irregular graph network mode describes the dynamic process of forest fire spread with the ignition time of graph nodes and the connecting edges, and indicates the ignition relationship between different graph nodes. It can more significantly visualize the spatiotemporal distribution characteristics of forest fire spread, and provide a novel and effective forest fire spread situation analysis method for government emergency departments, and provide auxiliary decision support for key links such as forest fire prevention, rescue deployment, and personnel evacuation of emergency management departments, which is conducive to achieving efficient rescue decision-making and reducing forest fire disaster losses.

[0181] The present invention is further described below with reference to the accompanying drawings and in combination with preferred embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0182] like Fig.13 As shown, a specific embodiment of the present invention provides a forest fire spread prediction method based on an irregular graph network, comprising the following steps:

[0183] S1. Generate an irregular graph network that represents a forest fire scene in a variable scale mode, wherein the irregular graph network includes a series of graph nodes and graph edges, and the graph nodes and graph edges have a set of attributes set in combination with the characteristics of the forest fire scene; after initializing the irregular graph network, adaptively determine the number and spatial distribution of the graph nodes according to the complexity of the forest fire scene;

[0184] S2, constructing a forest fire spread function under an irregular graph network by using a fitting function to approximate the forest fire spread function of a regular grid of a forest fire scene;

[0185] S3. Starting from the graph node corresponding to the fire starting point, the forest fire spread function calculates the forest fire spread duration of the graph edge on the irregular graph network, and calculates the ignition time of the graph node on the irregular graph network in combination with the minimum travel time principle, and iteratively generates the forest fire spread simulation result under the irregular graph network.

[0186] The following further describes the invention through specific examples.

[0187] 1. Variable-scale graph network space representation

[0188] 1.1 Definition of irregular graph network in forest fire scenario

[0189] The irregular graph network G = (V, E) consists of a series of graph nodes V and graph edges E. The graph nodes and graph edges have a set of attributes A set according to the characteristics of forest fire scenes. node and A edge , the attribute values ​​of graph nodes and graph edges are key data for calculating forest fire spread, including coordinate position, azimuth, vegetation type, terrain slope, graph edge length, etc. Specifically, in the forest fire scene, each graph node i ∈V has the attribute set A node :

[0190] A node ={ID,X,Y,LT,EL,IT}

[0191] Among them, ID is the unique number of the graph node, (X, Y) is the geographic coordinates of the graph node, LT is the vegetation type, EL is the elevation value, and IT is the relative time when the graph node is ignited (the relative time of the fire point is 0). Each graph edge describes the attribute set A of the line segment between two adjacent nodes. edge :

[0192] A edge ={SID,TID,LT,SL,L,DA}

[0193] Among them, SID and TID are the starting point and target point of the edge, LT is the vegetation type, SL is the slope from SID to TID, L is the geographical length of the edge, and DA is the edge angle, which describes the angle between the vector edge and the true north direction. When a forest fire spreads, starting from the node where the fire started, the diffusion speed and ignition time along the edge to the surrounding nodes are calculated.

[0194] 1.2. Initialize irregular graph network

[0195] like Fig.14 As shown in the figure, it is a schematic diagram of initializing an irregular graph network. The complexity of the forest fire area scene determines the spatial distribution of the nodes of the irregular graph network. The irregular graph network adopts a variable scale mode to characterize the forest fire scene. After initializing the irregular graph network, the number and spatial distribution of graph nodes are adaptively determined according to the complexity of the forest fire scene. The initialization of the irregular graph network includes: first, setting the resolution range of the irregular graph network, including the maximum graph resolution threshold GR max and the minimum image resolution threshold GR min , where GR max It is also the shortest graph edge length allowed in an irregular graph network, GR min It is also the longest graph edge length allowed in the irregular graph network; secondly, the graph nodes in the forest fire scene are initialized by uniformly distributed sampling, and the maximum distance between graph nodes is GR min; Again, graph edges are constructed on the initialized graph node set so that the graph edges can completely cover the forest fire space and there is no overlap between the graph edges. The attribute set of the graph nodes and the attribute set of the graph edges are generated according to the geographical location of the graph nodes. Specifically, the graph edges can be constructed on the initialized graph node set according to the Delaunay algorithm, and the attribute set A of the corresponding graph nodes is generated. node and the attribute set A of the graph edge edge .

[0196] In a specific example, GR can be set according to application requirements. max 5 meters, GR min is 30 meters, which means that the shortest and longest graph edge lengths allowed in the irregular graph network are 5 meters and 30 meters respectively. Then, according to GR min The sampling points are evenly distributed in the forest fire space, and the connecting edges between the graph nodes are generated based on the Delaunay algorithm to ensure that the graph edges can completely cover the forest fire space and there is no overlap between the graph edges. Finally, according to the geographical location of the graph nodes, the corresponding graph node attribute set and graph edge attribute set are generated.

[0197] 1.3. Adaptive Optimization of Irregular Graph Networks

[0198] like Fig.15 The figure shows a schematic diagram of the irregular graph network adaptive optimization. Based on the initialized irregular graph network, the density of the graph nodes is adaptively adjusted according to the complexity of the forest fire scene to perform irregular graph network adaptive optimization. Graph network adaptive optimization is the key to ensure the consistency of graph edge attributes in complex areas. The complexity of the forest fire scene is mainly determined by the complexity of vegetation types and terrain slopes. For simple areas, the forest fire scene is relatively simple, and GR is adopted. min It can more accurately represent the geographical environment. However, for complex areas, the heterogeneity of forest fire scenes is high, and there are often multiple vegetation types or drastically changing slopes on the edges. Therefore, it is necessary to increase the node density on the edges to reduce the attribute uncertainty of the edges. Among them, the adaptive optimization of irregular graph networks includes: for any edge in the graph network, according to the interpolation interval determined by the maximum graph resolution of the graph network, generate a set of candidate nodes for pre-interpolation of graph nodes on the edge Perform graph edge consistency check. When the candidate nodes in the candidate node set If the vegetation type is different from that of the previous node, insert the candidate node; and calculate the elevation difference change between the candidate node and the previous node when When the threshold T is exceeded, the candidate point is inserted; the irregular graph network is reconstructed based on the updated graph nodes; the above process is repeated until no new candidate nodes are inserted or the preset maximum number of iterations is reached; the graph nodes whose adjacent nodes are all non-flammable attributes are deleted to form an optimized irregular graph network. More specifically, the irregular graph network adaptive optimization includes:

[0199] First, for any edge in the irregular graph network, use GR max Determine the interpolation interval and generate a candidate node set Edge for pre-interpolation of graph nodes on the graph edge i :

[0200]

[0201] in, and is the starting point SID and target point TID of the i-th graph edge, the number of interpolated nodes is M+1, and each node in the candidate node set has a temporary attribute set

[0202] The purpose of considering the starting point SID and the target point TID as candidate nodes is to facilitate calculation. In the judgment rule of node insertion, first (i.e. SID starting point) starts iteration and first compares and If the interpolation condition is met, insert (Notice It is only used for comparison and will not be inserted). Iterate according to this rule to and when If the insertion condition is met, the node will be inserted, but the duplicate nodes in the entire graph will be deleted. Of course, it is also possible that the starting point SID and the target point TID are not considered as candidate nodes.

[0203] Then, the graph edge consistency check is performed. When the candidate nodes in the candidate node set With the previous node If the vegetation types of the edges are different, it indicates that the vegetation types of the edges are heterogeneous, and the original resolution is difficult to accurately describe the complex regional scene. Therefore, it is necessary to insert the candidate node to refine the regional representation capability. That is, in order to more accurately represent the area, the candidate node The node is inserted into the graph node set; Represents a candidate node With the previous node The elevation difference between Representation Node With the previous node Calculate the elevation difference between the candidate node and the previous node when When the set threshold T is exceeded, the candidate node is inserted. Specifically, the elevation difference change between the candidate nodes can be calculated by the quadratic difference method. when (Right now and When the absolute difference between the two nodes exceeds the set threshold T, it indicates that the terrain of the area is relatively complex (specifically, there is a large slope change on the edge of the graph). Therefore, it is better to set the candidate node Inserted into the graph node set to reduce the uncertainty of terrain changes.

[0204] Then, after inserting the new node, the structure of the original irregular graph network will be reconstructed based on the updated graph nodes. Specifically, based on the updated graph nodes, the Delaunay algorithm is used again to generate graph edges, and the graph node attributes and graph edge attributes are updated.

[0205] Finally, iterate the above steps (i.e., the reconstructed irregular graph network will continue to perform pre-interpolation and graph edge attribute consistency checks) until no new graph nodes are inserted or the maximum number of iterations is reached. For the irregular graph network after iteration, delete the graph nodes whose adjacent nodes are all non-flammable, reduce the redundancy of the irregular graph network, and generate the final optimized irregular graph network.

[0206] 2. Forest fire spread modeling based on irregular graph networks

[0207] 2.1. Definition of forest fire spread function

[0208] Existing forest fire spread models are mainly developed based on the spatial representation mode of regular grids, and there are few studies on designing combustion experiments and building spread models based on irregular graph network representation modes. Therefore, the specific embodiment of the present invention uses a fitting function To approximate the spread function f of the regular grid of the forest fire scene, construct the forest fire spread function under the irregular graph network:

[0209] t=f(LT,SL,L,Meteo)

[0210]

[0211] Among them, LT, SL, L and Meteo are vegetation type, slope, grid length and meteorological parameters (such as temperature, humidity, wind speed, wind direction, etc.), and t is the spreading time under the regular grid representation. It is the spreading time under the representation of irregular graph network. It is calculated The loss function between and f. If the number of hours is sufficient, it indicates It can achieve the simulation accuracy in regular grid mode.

[0212] The equation form of the fitting function can be various, such as multivariate linear equations, multivariate nonlinear equations, deep neural networks, etc. Taking deep neural networks as an example, a deep convolutional network composed of multiple Linear layers + Batch Normalization layers + Relu layers can be constructed to predict the spread time of forest fires on the edge of the image:

[0213]

[0214] Among them, θ is the network training parameter, is the training loss function, LT, SL, L, DA and Meteo are vegetation type, slope, grid length and meteorological parameters respectively, and t is the label data of spreading time.

[0215] 2.2. Graph-grid mapping

[0216] like Fig.17 The figure shows a schematic diagram of the graph-grid mapping method. Converting the simulation results of the regular grid into the simulation results of the irregular graph network is the key to achieving function fitting. The label data is the key to generating the fitting function parameters. Therefore, it is better to convert the forest fire spread results generated by the regular grid model into the label results under the irregular graph network through the graph-grid mapping method to generate the parameters θ of the fitting function. The graph-grid mapping method includes:

[0217] A regular grid model (such as a regular grid model based on the principle of cellular automation) is used to generate forest fire spread simulation results, and each grid point records the time when it is ignited.

[0218] The irregular graph network corresponding to the forest fire area is spatially aligned with the regular grid, and the ignition time of the grid point is copied and mapped to the irregular graph node.

[0219] Retrieve all adjacent nodes for each graph node And get the ignition time of all adjacent nodes Then determine the adjacent node whose ignition time is the smallest relative to the ignition time of the graph node, and set the adjacent node as the effective spreading node VSN of the graph node, and the corresponding connecting graph edge as the effective spreading edge VSEs.

[0220] The parameters θ of the fitting function are generated by training using the label data of the effective spreading edges VSEs.

[0221] Specifically:

[0222] Firstly, a regular grid model based on the principle of cellular automation is used to generate forest fire spread simulation results, and each grid point records the moment when it is ignited.

[0223] Then, based on a unified spatial coordinate system, the regular grid space is aligned with the irregular graph network space. Each graph node can correspond to a grid on the regular grid, and the ignition attributes on the grid can be mapped and copied to the corresponding graph node.

[0224] Then, according to the mapping relationship, the ignition time IT and other ignition attributes of each graph node can be obtained, and then the ignition graph node n can be searched. i The set of adjacent nodes and the corresponding ignition time Where M is the number of adjacent nodes. However, a graph node can only be ignited by adjacent nodes, but not all adjacent nodes are ignited by node n. i Therefore, it is necessary to determine the node that first ignites node n^i and define it as a valid spread node. That is, valid spread nodes (VSNs) are defined as the nodes that first ignite node n^i. i The ignition node, Ignition time ratio node n i Smaller and with the largest time difference, the effective burning node and node n i The graph edges formed are called valid spread edges (VSEs). That is, first search for the adjacent nodes of each graph node And get the ignition time of all adjacent nodes Then determine the adjacent node with the smallest ignition time relative to the graph node in the ignition time, that is, And set the graph node to node n i The corresponding connection graph edges of VSN are defined as effective spreading edges VSEs, such as Fig.16 FIG. 4 is a schematic diagram of effective spreading nodes VSNs and effective spreading edges VSEs.

[0225] 2.3. Spreading Function of Irregular Graph Networks

[0226] The VSEs generated by the graph-grid mapping method are labeled with the ignition time and other attributes. The attribute set of a VSE is Fitting function Then the ignition time calculation in the edge mode is realized:

[0227]

[0228] Fitting function There are many forms, such as linear function (such as least squares fitting equation), nonlinear function (such as deep learning fitting equation), and the parameter θ of the fitting function can be trained using label data to calculate the forest fire spread time under different environmental conditions. Specifically, the spread function is composed of the fitting function Taking the deep neural network fitting function as an example, the VSEs label data can be used to train the model and generate the model parameters θ. The trained converged model can calculate the forest fire spread time under different environmental conditions:

[0229]

[0230] 2.4 Forest fire spread simulation

[0231] The fitting function f^ can calculate the spreading time of a certain graph edge, so the minimum travel time algorithm is combined to realize the calculation of the ignition time of the graph nodes on the global irregular graph network, that is, starting from the graph node corresponding to the fire point, the forest fire spreading time of the graph edge on the irregular graph network is calculated according to the forest fire spreading function, and the ignition time of the graph node on the irregular graph network is calculated in combination with the minimum travel time principle, and the forest fire spreading simulation results under the irregular graph network are iteratively generated. Specifically:

[0232] First, according to the coordinates (X, Y) of the fire point, the node closest to the graph node in space is retrieved and set as the fire graph node N. ign , the starting ignition time is 0. Then, the ignition graph node N ign Set as the burning spread node, search for the adjacent graph node N of the fire graph node i With the adjacency graph edge E i , and use the fitting function Calculate the ignition time from the fire graph node to the adjacent graph node Will Add to candidate ignition set T edge ; Then, according to the minimum travel time principle, from the set T edge The graph node corresponding to the minimum ignition time value is selected (the smaller the value, the node will be ignited first) as the next combustion propagation node, that is, the next fitting function The calculated node ignition time is updated to T edge In the above process, when a graph node is ignited by multiple adjacent nodes, the minimum ignition time is used as the standard. Finally, the above process is iterated to calculate the combustion time IT of all graph nodes in the irregular graph network and the corresponding graph edge propagation rate, propagation duration and other simulation results, such as Fig.18 shown.

[0233] The specific implementation of the present invention also provides a forest fire spread prediction device based on an irregular graph network, which includes a processor and a computer-readable storage medium. When the computer program stored in the computer-readable storage medium is executed by the processor, the above-mentioned forest fire spread prediction method is implemented.

[0234] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art of the present invention, several equivalent substitutions or obvious variations can be made without departing from the concept of the present invention, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A complex forest fire scene simulation method based on VR technology, characterized in that: The steps include: S1. Build a virtual three-dimensional geographic scene based on the basic geographic information data of the target area; S2. Output results of fire spread simulation calculation and analysis of the target area according to the complex scene fire spread analysis model, dynamically displaying the three-dimensional dynamic model of flames in different time domains in the virtual three-dimensional geographic scene, and dynamically simulating the virtual fire spread; the complex scene fire spread analysis model is a forest fire spread prediction method based on deep learning, the method comprising the following steps: training a convolutional neural network using a constructed forest fire spread dataset; the convolutional neural network predicts the ignition time of each grid point in a grid-based forest fire space based on input data; wherein the input data includes forest fire spread status, geographic data and meteorological data, the convolutional neural network encodes and decodes the spatiotemporal distribution field of forest fire spread in the prediction area based on the input data, and uses the spatiotemporal distribution field of forest fire spread to predict the ignition time of each grid point in the forest fire space; the forest fire spread dataset includes a historical forest fire dataset and a simulation generated by using a traditional forest fire spread engine Data set; the simulation data set is obtained in the following manner: first, the prediction area is cut and cropped to form a set of geographical cuts; for the combination of geographical cuts, a random function is used to generate multiple groups of fire points and meteorological parameters, and the forest fire spread engine is used for simulation to generate multiple groups of forest fire spread simulation results, which constitute the key part of the forest fire spread data set; the encoding is to realize hierarchical progressive encoding of the input data through a state-condition mechanism; wherein, the input data is divided into a state branch and a condition branch, the state branch is the state data of the forest fire spread, and the condition branch is the environmental parameter acting on the forest fire spread state, including a geographical coding sub-branch and a meteorological coding sub-branch; each branch is first independently encoded and then deeply fused to form a hybrid feature, and then the backbone network coding of the encoding-decoding structure is adopted to further extract higher-order features of the hybrid feature, and decoded into the spatiotemporal distribution field of the forest fire spread, so as to realize end-to-end rapid reasoning and prediction of the forest fire spread; S3. Use the virtual reality device interactive tool to carry out combat mapping, map the constructed combat command mapping three-dimensional static model to the virtual three-dimensional geographic scene, and form a mapped combat map.

2. The complex forest fire scenario simulation method according to claim 1, characterized in that: The basic geographic information data of the target area includes a digital elevation model DEM, remote sensing satellite images and surface structure data; the surface structure data is one or more of ground vegetation type data, river data, and building data.

3. The complex forest fire scenario simulation method according to claim 2, characterized in that: Step S1 specifically includes: Constructing a three-dimensional model of surface structures in the target area according to the basic geographic information data of the target area; A basic three-dimensional elevation surface is constructed according to the basic geographic information data of the target area, and each pixel value of the remote sensing satellite image is mapped onto the three-dimensional elevation surface to generate a basic three-dimensional geographic scene with ups and downs, and the constructed three-dimensional model of the surface structure is placed in the corresponding surface area, thereby constructing a complete virtual three-dimensional geographic scene; before placing the surface structure in the corresponding surface area, the geographic information data of the surface structure is converted into coordinates so that it is in the same coordinate reference system as the DEM elevation data and the remote sensing image data.

4. The complex forest fire scenario simulation method according to claim 3, characterized in that: The three-dimensional model of the surface structures in the target area includes three-dimensional static models of ground vegetation and buildings and a three-dimensional dynamic model of rivers.

5. The complex forest fire scenario simulation method according to any one of claims 1 to 4, characterized in that: The three-dimensional dynamic model of flame includes one or more of bonfire, building burning fire, chemical burning fire, dark fire, loud fire, forest fire, oil ring fire, and smoke fire.

6. The complex forest fire scenario simulation method according to any one of claims 1 to 4, characterized in that: Step S2 specifically includes: Determine the combustible data existing in the target area according to the basic geographic information data of the target area, analyze the possible fire types in the target area according to the combustible data, and construct a three-dimensional dynamic model of flame that conforms to the fire characteristics of the target area; A complex scene fire spread analysis model that can perform fire spread simulation and calculation analysis on the target area is selected, and based on the output results of the complex scene fire spread analysis model and combined with the flame three-dimensional dynamic model, the flame three-dimensional dynamic model in different time domains is dynamically displayed in the virtual three-dimensional geographic scene.

7. The complex forest fire scenario simulation method according to any one of claims 1 to 4, characterized in that: In step S3, a model library of the combat command mapping three-dimensional static model is constructed according to the emergency forces, emergency materials and combat identifier lists required for use in the fire emergency command process.

8. The complex forest fire scenario simulation method according to any one of claims 1 to 4, characterized in that: In step S3, the use of the virtual reality device interactive tool for combat mapping includes adding, moving, enlarging, reducing, and deleting a specified combat command mapping three-dimensional static model in the virtual three-dimensional geographic scene through human-computer interaction instructions.

9. The complex forest fire scenario simulation method according to any one of claims 1 to 4, characterized in that: In step S3, the marked operation map is a two-dimensional operation map drawn by orthographic projection.

10. A complex forest fire scene simulation system based on VR technology, characterized in that: The method comprises a processor and a computer-readable storage medium. When a computer program stored in the computer-readable storage medium is executed by the processor, the method for simulating a complex forest fire scene as described in any one of claims 1 to 8 is implemented.

Images