Forest fire spreading prediction method based on joint prediction model

Abstract

The invention provides a forest fire spreading prediction method based on a joint prediction model, and the method comprises the steps: discretizing a forest region into cells, and associating forest fire data with environment data; according to the forest fire data, it is determined that the initial state of the cells is unburned, combusted, burnt out or incombustible; on the basis of a cellular automaton model, according to the environment data and the cellular neighborhood, predicting a cellular state sequence of cellular states changing along with time; extracting a cellular state sequence according to a time step length, organizing a fire spread space-time matrix, fusing the fire spread space-time matrix with external environment features to form a joint feature matrix, and inputting the joint feature matrix into a space-time convolutional neural network; and the space-time convolutional neural network generates a real-time fire spreading state prediction result. According to the joint prediction model, the physical law of fire spreading is simulated through the cellular automaton, and the space-time dependency relationship of multi-dimensional environment data is learned in combination with the space-time convolutional neural network, so that the adaptability and accuracy of the model are improved, and the dynamic prediction of the forest fire can be effectively and quickly completed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of forest fire prediction, and particularly relates to a forest fire spreading prediction method combining a cellular automaton (CA) and a space-time convolutional neural network (STCNN). BACKGROUND

[0002] Forest fire is one of the natural disasters that frequently occur worldwide, which is strong in suddenness and destructive, and poses a serious threat to the ecological environment, social economy and human safety. According to statistics, the forest area lost due to forest fires worldwide each year exceeds several million hectares, directly leading to a sharp decline in biodiversity and releasing a large amount of carbon dioxide to aggravate climate change. In terms of emergency response, the rapid spreading characteristics of fire make it difficult for traditional monitoring methods (such as satellite remote sensing and ground patrol) to timely predict the trend of fire, resulting in low efficiency of fire fighting.

[0003] Current forest fire prediction mainly relies on two types of models, one is a physical model, which simulates the spread of fire based on rules, which can simulate the basic law of fire propagation, but it is difficult to handle the nonlinear influence of complex environmental variables (such as dynamic meteorology and terrain slope), and has limited adaptability; the other is a data-driven model, which relies on historical data to learn patterns, but lacks explanation of the physical mechanism of fire and has insufficient generalization ability, such as pure data-driven neural network models, which are prone to misjudgment when there is insufficient training data or environmental mutation. Therefore, the present application combines the rule-driven characteristics of cellular automata with the data-driven capabilities of space-time convolutional neural networks to propose a hybrid prediction model that combines physical law expression and complex environmental adaptability. SUMMARY

[0004] Therefore, the present application provides a forest fire spreading prediction method based on a joint prediction model, which combines the rule-driven characteristics of cellular automata with the data-driven capabilities of space-time convolutional neural networks to propose a hybrid prediction model that combines physical law expression and complex environmental adaptability, and can effectively and quickly complete forest fire prediction.

[0005] To solve the above technical problems, the present application is implemented as follows.

[0006] A forest fire spreading prediction method based on a joint prediction model, comprising:

[0007] Step 1, collecting forest fire data and environmental data;

[0008] Step 2, dividing the forest area into two-dimensional grid cells, associating forest fire data and environmental data with each cell; according to the forest fire data, determining the initial state of the cell as unburned, burning, burned out or non-flammable;

[0009] Step 3, based on the cellular automaton model, according to the environmental data and the neighborhood of the cell, the sequence of cell state changing with time is predicted;

[0010] Wherein, for the unburned state of the cell, the neighborhood cell state and the environmental data are comprehensively considered to predict the forest fire spread trend, and the cell state at the next time step is determined according to the forest fire spread trend; for the burning state of the cell, whether the cell state at the next time step changes to the burned out is determined according to the combustible consumption model;

[0011] Step 4, the cell state sequence is extracted according to the time step, and is organized into a fire spread space-time matrix to reflect the preliminary development trend of the fire in the space-time dimension; the fire spread space-time matrix and the environmental data are aligned according to the time step and the spatial position, and are fused to form a joint feature matrix, which is input into the space-time convolutional neural network;

[0012] Step 5, the space-time convolutional neural network generates a real-time fire spread state prediction map; the cell state and the environmental data are updated according to the set time interval, so as to obtain the fire prediction result changing with time.

[0013] Preferably, the environmental data includes meteorological data, terrain data and vegetation data; the forest fire data includes the time, location and fire situation of the fire.

[0014] Preferably, in step 3, for the unburned state of the cell, the neighborhood cell state and the environmental data are comprehensively considered to predict the forest fire spread trend, and the cell state at the next time step is determined according to the forest fire spread trend as follows:

[0015] For the unburned state of the cell A, according to the number of burning neighborhood cells, the wind speed, temperature, humidity and terrain slope associated with the cell A, the spread probability is calculated, and when the spread probability value is greater than a set threshold, the cell state at the next time step is determined to switch to the burning state.

[0016] Preferably, the calculation formula of the spread probability is as follows:

[0017]

[0018] Wherein, P spread (i) represents the spread probability of the i-th cell; p j is a value between 0.1 and 1 set, representing the influence degree of the neighborhood cell, and n is the total number of the burning neighborhood cells; k v is the wind speed coefficient, v is the current wind speed; k T is the temperature coefficient, T is the current temperature; k H is the humidity coefficient, H is the current humidity; k S is the slope coefficient, S is the size of the slope.

[0019] Preferably, the neighborhood cells are determined by a Moore neighborhood, each cell is adjacent to the cells in eight directions around it.

[0020] Preferably, in step 3, for the cells in the burning state, it is determined whether the cell state at the next time step is transformed into the burned state according to the combustible consumption model:

[0021] The combustible mass in the cell grid is determined according to the vegetation type, and the burning rate R is determined according to the wind speed;

[0022] The combustible consumption amount in each time step Δt is calculated as ΔM = R·Δt;

[0023] The remaining combustible mass M is updated as: remain ←M remain -ΔM;

[0024] When the remaining combustible mass M remain ≤0, it is determined that the combustible is exhausted, and the corresponding burning cell state is updated to the burned state.

[0025] Preferably, the spatio-temporal convolutional neural network is composed of a spatial convolutional layer, a temporal convolutional layer, a stride convolutional layer, and a 1×1 convolutional layer; the working steps of the spatio-temporal convolutional neural network include:

[0026] Step 5.1, the spatio-temporal convolutional neural network performs two-dimensional convolution operation on the input joint feature matrix through the spatial convolutional layer, extracts the first feature representing the propagation characteristics of the fire in the two-dimensional space, and inputs the temporal convolutional layer;

[0027] Step 5.2, the first feature is subjected to one-dimensional convolution operation through the temporal convolutional layer to obtain the second feature representing the evolution trend and propagation characteristics of the fire in the time dimension, and the stride convolutional layer is inputted;

[0028] Step 5.3, the stride convolutional layer is used to reduce the dimension of the second feature to obtain the third feature;

[0029] Step 5.4, the 1×1 convolutional layer is used to map the high-dimensional third feature to the output prediction value, and the fire spread state prediction map is outputted; the cell state and environmental data are dynamically updated, and the real-time changing fire prediction result is generated by steps 4.1-4.4.

[0030] Preferably, the method further includes training the spatio-temporal convolutional neural network:

[0031] The cross-entropy loss function is defined to measure the difference between the prediction value outputted by the 1×1 convolutional layer and the true value, and the network parameters of the spatio-temporal convolutional neural network are updated through the Adam optimizer.

[0032] Preferably, in step 4, the fire spread space-time matrix is aligned with the environmental data according to the time step and the spatial position, and is fused to form a joint feature matrix as follows: The joint feature matrix includes static feature encoding and dynamic feature encoding.

[0033] The static feature describes the inherent properties of the space, is bound to each spatial position, and keeps a fixed value after encoding, including terrain data and vegetation data.

[0034] The dynamic feature describes parameters that evolve over time, is updated according to the time step and is aligned with the spatial position, including meteorological data and the fire spread space-time matrix.

[0035] The static feature and the dynamic feature are aligned according to the time step and the spatial position to construct a multi-dimensional joint input tensor Where T is the time step, C is the number of feature channels, and N×N is the spatial two-dimensional grid resolution.

[0036] Advantages:

[0037] (1) The present application proposes a combined physical model and data-driven joint prediction model for the complexity problem of the influence of fuel, environment (meteorology, terrain, etc.) on the dynamic propagation process of forest fires, which has physical law expression and adaptability to complex environment. The joint prediction model simulates the physical law of fire spread through cellular automata, learns the spatio-temporal dependence of multi-dimensional environmental data through spatio-temporal convolutional neural network, improves the adaptability and accuracy of the model, and can effectively and quickly complete the dynamic prediction of forest fires.

[0038] (2) The cellular automata in the joint prediction model can effectively simulate the basic process of fire spread through spatial discretization, state definition, neighborhood rule definition, state transition rule definition and spread probability calculation, and the spread probability considers multiple factors, making the simulation more realistic.

[0039] (3) The spatio-temporal convolutional neural network in the joint prediction model efficiently extracts spatio-temporal features, reduces model complexity and improves prediction accuracy through data preprocessing, multi-modal feature fusion, spatial and temporal convolution layers, stride convolution and direct use of convolution layer output prediction value design. Among them, the stride convolution method replaces the traditional pooling layer, reduces the resolution of the feature map by adjusting the moving step of the convolution kernel, and realizes the purpose of dimension reduction. At the same time, through 1×1 convolution, the high-dimensional feature channel is integrated into the dimension of the target output, so that the high-dimensional feature is mapped to the final output prediction value.

[0040] (4) The environmental information associated with the cell of the present application includes weather, terrain, and vegetation information, and plays a role in cell updating, so the present application can integrate multiple factor variables such as wind speed, temperature, terrain, and vegetation, optimize fire spread path prediction, and significantly improve adaptability in complex environments, providing a scientific basis for forest fire prevention and control.

[0041] (5) The selection of the loss function and the optimizer of the present application further improves the performance of the model.

[0042] (6) The method has both computational efficiency and prediction accuracy, effectively reducing the lag and misjudgment of traditional models, and is suitable for real-time monitoring and intelligent decision support of forest fires. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a flowchart of the forest fire spread prediction method based on the joint prediction model of the present application;

[0044] Figure 2 is a schematic diagram for defining the state of the cell;

[0045] Figure 3 is a schematic diagram of the STCNN method. DETAILED DESCRIPTION

[0046] The technical implementation process of the forest fire spread prediction model based on the cellular automaton (CA) and the spatio-temporal convolutional neural network (STCNN) of the present application will be described in detail below in combination with the drawings and examples. The present method can dynamically simulate the fire propagation process and optimize the prediction results by fusing physical models and data-driven techniques.

[0047] As shown in Figure 1 , the present application provides a forest fire spread joint prediction method, comprising the following steps:

[0048] Step 1: Collect forest fire data and environmental data.

[0049] The forest fire data includes the time, location, and fire situation of the fire, and can also include the fire spread path, etc. The environmental data includes meteorological data, terrain data, and vegetation data. Among them, the meteorological data comes from the weather station, including wind speed, wind direction, temperature, etc.; the terrain data comes from remote sensing satellites, including slope, terrain, etc.; the vegetation data mainly includes vegetation types, etc.

[0050] The collected data is sorted, cleaned, and normalized, and abnormal values and missing values are eliminated to eliminate the influence of different dimensions and ensure data quality.

[0051] Step 2: Discretize the forest area into a two-dimensional grid cell, and associate forest fire data and environmental data with each cell; according to the forest fire data, determine the initial state of the cell as unburned, burning, burned out, or non-flammable.

[0052] In this step, the continuous forest area is first divided into a regular two-dimensional grid to form a cellular automata model. The grid resolution is adjusted according to the computing resources and precision requirements, such as 5m x 5m or 10m x 10m.

[0053] The forest fire data and environmental information, including fire intensity, meteorological, vegetation, and terrain data, are associated with each cell.

[0054] According to the burning state of the fire intensity in the forest fire data, an initial state is assigned to each cell. The cell state is defined as an unburned state (0), a burning state (1), a burned-out state (2), or an unburnable state (3), and one of them is assigned.

[0055] Step 3: Based on the cellular automata model, the cell state sequence over time is predicted according to the environmental data and the cell neighborhood.

[0056] As shown in Figure 2 , a selected molar neighborhood (8-neighborhood) is selected as the neighborhood selection rule.

[0057] The present application needs to predict the subsequent time step state based on the initial state of the cell to form a time sequence. The present application predicts the state for cells in the unburned state and the burning state. For cells in the unburned state, the neighborhood cell state and environmental data are considered to predict the forest fire spread trend, and the cell state at the next time step is determined according to the forest fire spread trend; for cells in the burning state, the duration of the burning and the combustible consumption model are used to determine whether the cell state at the next time step changes to the burned-out state.

[0058] In a preferred embodiment, the state prediction for a cell A in the unburned state is as follows: according to the number of neighboring cells in the burning state, the wind speed, temperature, humidity, and terrain slope associated with the cell A, the spread probability is calculated, and when the spread probability value is greater than a set threshold, the cell state at the next time step is determined to switch to the burning state.

[0059] A preferred spread probability calculation formula is:

[0060]

[0061] where P spread (i) represents the spread probability of the i-th cell; p j is the influence probability of the j-th burning state cell in the neighborhood on the target cell, p j is a set value, selected from a value between 0.1 and 1, representing the influence degree of the neighborhood cell; when the neighborhood is in the burning state, the influence degree on the target cell is p j , and n is the number of burning cells. The influence of each burning neighborhood is the same pj Value. k v is the wind speed coefficient, v is the current wind speed; k T is the temperature coefficient, T is the current temperature; k H is the humidity coefficient, H is the current humidity; k S is the slope coefficient, S is the size of the slope.

[0062] The state prediction of the cell B in the burning state is specifically: for the cell in the burning state, according to the combustible consumption model, it is judged whether it is changed to the burned-out state. Specifically:

[0063] First, the combustible mass and the burning rate of the grid are determined; wherein the combustible mass is determined according to the vegetation type and the grid area; for example, the vegetation of grassland type is dry and easy to burn, but the total amount is small, and the combustible mass per unit area can be set to 0.5 kg / m 2 , and the vegetation of coniferous forest type burns for a long time, and the combustible mass per unit area can be set to 2.0 kg / m 2 . The burning rate R is positively correlated with the wind speed v, and the calculation expression is R=k·v, k is the wind speed correlation coefficient, which is a certain value. The wind speed v is obtained from the meteorological data;

[0064] In each time step Δt, the consumed combustible mass is:

[0065] ΔM=R·Δt (2)

[0066] The remaining combustible mass is updated as:

[0067] M remain ←M remain -ΔM (3)

[0068] When the remaining combustible mass M remain ≤0, it is determined that the combustible is consumed, and the state of the burning cell is updated to the burned-out state.

[0069] Step 4: according to the state sequence of each cell, extracting according to the time step, obtaining all the cell state matrices of each time step, organizing into a fire spread space-time matrix to reflect the preliminary development trend of the fire in the space-time dimension. Align the fire spread space-time matrix and the external environment characteristics according to the time step and the spatial position, and fuse to form a joint feature matrix, which is input into the space-time convolutional neural network.

[0070] In this step, the cell state is calculated iteratively by time step to simulate the initial spread of the fire in the forest area, and a sequence of cell state changes over time is generated to provide basic data for subsequent STCNN analysis. The CA outputs the sequence of cell state changes over time, extracts the cell state sequence by time step, organizes it into a fire spread space-time matrix, and uses it as a basic feature of fire spread to reflect the initial development trend of the fire in the time-space dimension.

[0071] The space-time matrix of fire spread is combined with meteorological, terrain and vegetation and other external environmental characteristics, strictly aligned by time step and spatial location, and fused to form a joint feature matrix, which provides comprehensive and accurate data for STCNN to ensure that the model can fully capture the spatio-temporal characteristics and environmental factor influences in the fire spread process.

[0072] Here, the joint feature matrix includes static feature encoding and dynamic feature encoding. The static feature describes the inherent properties of the space and needs to be bound to each spatial location, and the encoded value remains fixed, including terrain slope and vegetation type; the dynamic feature describes parameters that evolve over time and needs to be updated by time step and aligned with the spatial location, including meteorological data and the space-time matrix of fire spread. Align the static features and dynamic features by time step and spatial location to construct a multi-dimensional joint input tensor Where T is the time step, C is the number of feature channels, and N x N is the spatial grid resolution.

[0073] Step 5, the space-time convolutional neural network generates a real-time fire spread state prediction map.

[0074] Further add a dynamic updating mechanism, according to the set time interval, iteratively update the environment data and cell state, so as to obtain the fire prediction result map changing over time.

[0075] The space-time convolutional neural network in the application is composed of a spatial convolution layer, a time convolution layer, a stride convolution layer and a 1 x 1 convolution layer. The working steps of the space-time convolutional neural network are as shown in Figure 3 , including:

[0076] Step 5.1, the space-time convolutional neural network performs two-dimensional convolution operation on the input joint feature matrix through the spatial convolution layer, extracts the first feature representing the propagation characteristics of the fire in the two-dimensional space, and inputs the time convolution layer.

[0077] In this step, the spatial convolution layer is used to extract the spatial features in the grid during the fire spread process, and the propagation characteristics of the fire in the two-dimensional space are captured. The spatial convolution kernel slides in the local k x k area of the input feature map, extracts the neighborhood features and generates the output feature map. For the position (i, j) of the output feature , the calculation formula is:

[0078]

[0079] where N' is the spatial dimension of the output feature map; c and c' represent different feature channels; C is the total number of input feature channels, and C' is the total number of output feature channels; is the weight of the convolution kernel at the position (u, v) between the channel c and the channel c', which controls the importance of each neighborhood point to the output feature; X c is the value of the input feature channel c at the position (i+u-[k / 2], j+v-[k / 2]); b c' is the bias term of the channel c', which adjusts the reference value of the feature and is used to improve the expression ability of the model; f(·) is an activation function, and commonly used activation functions include ReLU function and Sigmoid function; k is the size of the convolution kernel.

[0080] The input feature is subjected to a two-dimensional convolution operation by the spatial convolution layer of the STCNN. The convolution kernel slides in the local area of the input feature map, extracts the propagation characteristics of the fire in the two-dimensional space, and simultaneously processes multiple spatial information by setting a multi-channel convolution kernel, thereby enriching the expression ability of the model to the spatial characteristics. In addition, the expression ability of the model to the complex fire propagation mode is enhanced by introducing a nonlinear operation to the activation function.

[0081] Step 5.2, performing one-dimensional convolution operation on the first feature by the time convolution layer to obtain the second feature representing the evolution trend and propagation characteristics of the fire in the time dimension, and inputting the stride convolution layer.

[0082] This step performs one-dimensional convolution operation on the input feature by the time convolution layer, and focuses on capturing the evolution trend and dependence relationship of the fire in the time dimension.

[0083] In this step, the time convolution layer is used to extract the time characteristics of the grid in the fire spreading process, and capture the propagation characteristics of the fire in the time dimension. The time convolution kernel performs one-dimensional convolution operation on the input feature, and k t is the length of the time convolution kernel, and for the position (i, j) of the output feature , the calculation formula is:

[0084]

[0085] where T' is the time dimension of the output feature map; W t (c',c) (u) is the weight of the time convolution kernel between the channel c and the channel c', which performs weighting operation across the time dimension.

[0086] Step 5.3, performing dimension reduction on the second feature by the stride convolution layer to obtain the third feature.

[0087] In this step, stride convolution is used instead of the traditional pooling layer to reduce the resolution of the feature map by adjusting the moving step of the convolution kernel, achieving the purpose of dimension reduction.

[0088] In this step, the feature matrix is reduced in dimension by stride convolution, and the features are extracted at intervals of s in each operation. The operation can be represented as:

[0089]

[0090] where Y out (i,j) is the value of the output feature map at position (i,j) after the stride convolution operation; W c (m,n) is the weight of the convolution kernel at position (m,n) corresponding to the input feature channel c; X c (i·s+m,j·s+n) is the value of the input feature channel c at position (i·s+m,j·s+n); b is the bias term of the convolution operation.

[0091] Step 5.4, use a 1×1 convolution layer to map high-dimensional third features to output prediction values, output fire spread state prediction map; dynamically update cell state and environment data, and generate real-time changing fire prediction results using steps 4.1-4.4.

[0092] This step directly uses a 1×1 convolution layer to map high-dimensional features to the final output prediction value. By reasonably setting the weights and biases of the convolution layer, the high-dimensional feature channels are integrated into the target output dimension to predict the spread range, intensity change, and other key information of the fire in the future time step.

[0093] The output calculation formula of the 1×1 convolution layer is where is the weight of the 1×1 convolution, X f,c (i,j) is the value of the feature channel c at position (i,j) after spatial and temporal convolution, b out is the independent bias term of each output channel in the convolution operation. Through 1×1 convolution, the high-dimensional feature channels are integrated into the dimension of the target output.

[0094] The STCNN needs to be trained. During training, historical forest fire data, weather data, terrain data, and vegetation data are widely collected to form training samples. After processing by CA, the joint feature matrix is input into STCNN, and STCNN outputs the prediction value. The cross-entropy loss function is selected to improve the prediction performance of the model, and the function calculation formula is where Z represents the number of samples, y i represents the true value of the i-th sample, represents the prediction value of the i-th sample.

[0095] Update network parameters by Adam optimizer, input data, output predicted value by convolutional layer Use loss function L CE Measure prediction error, calculate gradient according to loss, update network parameters.

[0096] The above specific embodiments only describe the design principles of the present application, and the shapes and names of the components in the description can be different and are not limited. Therefore, those skilled in the art of the present application can modify or equivalently replace the technical solutions described in the foregoing embodiments; and these modifications and replacements do not deviate from the purpose and technical solutions of the present application, and should all belong to the protection scope of the present application.

Claims

1. A method for predicting forest fire spread based on a joint prediction model, characterized in that, include: Step 1: Collect forest fire data and environmental data; Step 2: Divide the forest area into two-dimensional grid cells, and associate each cell with forest fire data and environmental data; based on the forest fire data, determine the initial state of the cell as unburned, burning, burned out, or non-burnable; Step 3: Based on the cellular automata model, predict the cellular state sequence of changes over time according to environmental data and the cellular neighborhood. Specifically, for cells in the unburned state, the spread trend of forest fires is predicted by comprehensively considering the states of neighboring cells and environmental data, and the cell state of the next time step is determined based on the spread trend of forest fires; for cells in the burning state, the cell state of the next time step is determined based on the combustible material consumption model to see if it has changed to burnout. Step 4: Extract the cell state sequence according to the time step and organize it into a fire spread spatiotemporal matrix to reflect the initial development trend of the fire in the spatiotemporal dimension; align the fire spread spatiotemporal matrix with the environmental data according to the time step and spatial location, fuse them to form a joint feature matrix, and input it into the spatiotemporal convolutional neural network. Step 5: The spatiotemporal convolutional neural network generates a real-time fire spread prediction map; the cell state and environmental data are updated according to the set time interval to obtain the fire prediction results that change over time.

2. The method as described in claim 1, characterized in that, The environmental data includes meteorological data, topographic data, and vegetation data; the forest fire data includes the time, location, and intensity of the fire.

3. The method as described in claim 1, characterized in that, In step 3, for cells in the unburned state, the fire spread trend is predicted by comprehensively considering the states of neighboring cells and environmental data. Based on the fire spread trend, the cell state for the next time step is determined as follows: For cell A in the unburned state, the spread probability is calculated based on the number of neighboring cells in the burning state, the wind speed, temperature, humidity and terrain slope associated with cell A. When the spread probability value is greater than a set threshold, the cell state in the next time step is switched to the burning state.

4. The method as described in claim 3, characterized in that, The formula for calculating the propagation probability is: Among them, P spread (i) represents the propagation probability of the i-th cell; p j The value between 0.1 and 1 represents the degree of influence of the neighboring cells, where n is the total number of neighboring cells during combustion; k v This is the wind speed coefficient, where v is the current wind speed; k T It is a temperature coefficient, where T is the current temperature; k H It is the humidity coefficient, where H is the current humidity; k S It is the slope coefficient, and S is the magnitude of the slope.

5. The method as described in claim 1, 3, or 4, characterized in that, The neighborhood cells are determined using Mohs neighborhood, and each cell is adjacent to cells in eight surrounding directions.

6. The method as described in claim 1, characterized in that, In step 3, for cells in the combustion state, based on the combustible material consumption model, it is determined whether the cell state in the next time step will change to burnout: The mass of combustible material within the cell grid is determined based on vegetation type, and the combustion rate R is determined based on wind speed. The amount of combustible material consumed in each time step Δt is calculated as ΔM = R·Δt; Update remaining combustible mass: M remain ←M remain -ΔM; When the remaining combustible mass M remain When the value is ≤0, the combustible material is determined to be exhausted, and the corresponding burning cell state is updated to "burned out".

7. The method as described in claim 1, characterized in that, The spatiotemporal convolutional neural network consists of spatial convolutional layers, temporal convolutional layers, stride convolutional layers, and 1×1 convolutional layers; the working steps of the spatiotemporal convolutional neural network include: Step 5.1: The spatiotemporal convolutional neural network performs a two-dimensional convolution operation on the input joint feature matrix through a spatial convolutional layer to extract the first feature representing the characteristics of fire propagation in two-dimensional space, and inputs it into the temporal convolutional layer; Step 5.2: Perform a one-dimensional convolution operation on the first feature through a temporal convolutional layer to obtain the second feature representing the evolution trend and propagation characteristics of the fire in the time dimension, and input it into the stride convolutional layer; Step 5.3: Use stride convolutional layers to reduce the dimensionality of the second feature to obtain the third feature; Step 5.4: Use a 1×1 convolutional layer to map the high-dimensional third feature into the output predicted value, and output the fire spread state prediction map; dynamically update the cell state and environmental data, and use steps 4.1 to 4.4 to generate real-time fire prediction results.

8. The method as described in claim 7, characterized in that, This method further includes training a spatiotemporal convolutional neural network: Define a cross-entropy loss function to measure the difference between the predicted and true values ​​of the output of a 1×1 convolutional layer, and update the network parameters of the spatiotemporal convolutional neural network using the Adam optimizer.

9. The method as described in claim 1, characterized in that, In step 4, the fire spread spatiotemporal matrix and environmental data are aligned according to time steps and spatial locations, and fused to form a joint feature matrix, which includes static feature coding and dynamic feature coding. Static features describe the inherent properties of a space, are bound to each spatial location, and retain fixed values ​​after encoding, including terrain data and vegetation data; Dynamic features describe parameters that evolve over time, are updated step-by-step and aligned with spatial location, including meteorological data and the spatiotemporal matrix of fire spread; Static and dynamic features are aligned according to time steps and spatial locations to construct a multidimensional joint feature matrix.

Images