A regional fire risk prediction method based on multi-index data

Abstract

The application discloses a kind of regional fire risk prediction methods based on multi-index data, it is related to fire prediction method.The application specifically includes the following steps: step 1. Vector grid is generated according to the set longitude and latitude interval to the predicted area, as basic space unit;Step 2. Set up key index system;Step 3. Key index analytic hierarchy process weight calculation and verification are carried out to each basic space unit;Step 4: data standardization and forest grassland fire danger index calculation;Step 5: neural network regression model construction and training;Step 6: regional danger index prediction.The application integrates multi-dimensional real forest grassland fire risk data, uses basic grid unit as model training data, adopts neural network to fit complex nonlinear relationship to improve prediction accuracy, and has good engineering application prospect and social benefit.

Description

Technical Field

[0001] This invention belongs to the field of disaster prevention and prediction technology, and relates to fire prediction methods, specifically a regional fire risk prediction method based on multi-index data. Background Technology

[0002] Forest and grassland fires are among the most far-reaching types of natural disasters affecting ecosystem security and the lives and property of residents in forest and pastoral areas. In recent years, influenced by factors such as global warming, exacerbated regional droughts, and frequent extreme heat events, forest and grassland fires have shown significant characteristics, including a year-on-year increase in their frequency, more extreme fire behaviors, and a significantly faster spread.

[0003] Currently, the forest and grassland fire risk level forecasting system still faces the following key technical bottlenecks in the face of new situations and the need for higher accuracy:

[0004] (1) Over-reliance on meteorological drivers without adequate consideration of the spatial heterogeneity of combustibles and ignition sources.

[0005] Current operational forecasts primarily rely on meteorological factors (such as temperature, precipitation, wind speed, and humidity), while paying insufficient attention to core combustion parameters such as forest and grassland combustible load (especially fine combustibles), flammability, and moisture content. Furthermore, key influencing factors such as the intensity of human activity in forest and pastoral areas, the density of open fire sources, and the localized wind field enhancement caused by the complex terrain of high mountain and canyon regions have not been effectively integrated, making it difficult to comprehensively depict the fire hazard formation mechanism.

[0006] (2) Lack of ability to express fine-grained differences at the regional scale

[0007] Currently, meteorological fire risk levels are mostly based on a spatial resolution of 1km or more, which makes it difficult to reflect the significant fine-grained differences between counties (cities, districts), towns (townships), and even different vegetation types. This results in the actual fire risk pattern within a region being overly smoothed, which is not conducive to precise prevention and control and localized decision-making.

[0008] (3) It is difficult to integrate multi-source heterogeneous data.

[0009] The forest and grassland fire risk survey (a special project on comprehensive natural disaster risk survey) has for the first time obtained high-precision feature data from multiple sources, including combustibles, wildfire sources, meteorology, and topography. However, the current operational system lacks a complete multi-source data fusion model and technical methodology, resulting in a low operational conversion rate of the survey results and making it difficult to fully support refined forest and grassland fire risk level forecasting.

[0010] (4) Traditional statistical models are difficult to characterize the complex nonlinear features of forest and grassland fire risk.

[0011] Forest and grassland fire risk is driven by multiple factors, and these factors exhibit significant nonlinear and coupled relationships. Examples include the nonlinear response of combustible material moisture content to meteorological conditions, the interaction between field fire density and vegetation cover, and the accelerating effect of topography on local wind fields. Existing linear statistical models struggle to effectively characterize the complex relationships of forest and grassland fire risk, thus limiting their ability to provide high-precision forecasts. Summary of the Invention

[0012] To overcome the technical deficiencies of existing technologies, this invention discloses a regional fire risk prediction method based on multi-index data.

[0013] The regional fire risk prediction method based on multi-index data described in this invention includes the following steps:

[0014] Step 1. Generate a vector grid for the predicted area according to the set latitude and longitude intervals, which will serve as the basic spatial unit;

[0015] Step 2. Establish a key indicator system, including n key indicators related to the probability of fire occurrence.

[0016] x b1 ,x b2 ,...x bn ;

[0017] Generate an n-dimensional key indicator feature vector Xb.

[0018] Xb=(x b1 ,x b2 ,...x bn );

[0019] Step 3. Perform weight calculation and verification of key indicators using the Analytic Hierarchy Process (AHP) for each basic spatial unit;

[0020] Step 31. Based on the key indicators in Step 2, construct a pairwise judgment matrix A for each indicator using the Analytic Hierarchy Process (AHP):

[0021] A=(a ij ), where matrix element a ij The subscripts i and j represent the row and column of the matrix, and each matrix element a ij This indicates the importance of the i-th key indicator to the j-th key indicator;

[0022] Step 32. Calculate the weight w of the i-th key indicator according to the weight calculation formula. i :

[0023] ;

[0024] The weight vectors of each key indicator are obtained by solving the problem.

[0025] W = (w1, w2, ... w) n );

[0026] Step 4: Data standardization and calculation of forest and grassland fire hazard index;

[0027] The key indicator data of all basic spatial units collected in step 2 are standardized. The standardized key indicator data...

[0028] ;

[0029] μ i σ is the average of the i-th key indicator of all basic spatial units. i Let x be the standard deviation of the i-th key indicator, where x i This is the raw data for the i-th key indicator.

[0030] Based on standardized key indicator data Calculate the forest and grassland fire hazard index H for each basic spatial unit. b :

[0031] ;

[0032] Where w i The weight of the i-th key indicator, The standardized value of the i-th key indicator of the grid cell is used to calculate the forest and grassland fire hazard index H. b Label data used for training neural networks;

[0033] Step 5: Building and training the neural network regression model;

[0034] A neural network regression model is constructed, using the key indicator feature vector Xb obtained in step 2 and the label data obtained in step 4 as the neural network dataset. The neural network dataset is divided into a training set and a test set.

[0035] The neural network regression model is trained using N samples, which is data from N basic spatial units.

[0036] The coefficient of determination R² and the root mean square error RMSE are used as indicators to evaluate model accuracy, respectively:

[0037] ;

[0038] ;

[0039] Where N represents the sample size, H b This represents the forest-grassland fire hazard index of the b-th basic spatial unit calculated according to step 4. Represents N H b The average value, This represents the predicted value of the forest and grassland fire hazard index for the b-th basic spatial unit.

[0040] And set the optimal value of the coefficient of determination RT and the optimal value of the root mean square error RMT. The model parameters with the coefficient of determination R²≥RT and the root mean square error RMSE≤RMT are determined as the optimal model parameters, and the corresponding deep learning training model is determined as the final model ME.

[0041] Step 6: Predicting the regional risk index;

[0042] For the basic spatial unit that needs to be predicted, after collecting the key indicator data of the basic spatial unit, the forest and grassland fire hazard index H is calculated according to steps 2 to 4. b The key indicator feature vector Xb is input into the final model ME obtained in step 5, and the model is trained to output the predicted value of the forest and grassland fire hazard index for this basic spatial unit. .

[0043] Preferably, step 2 specifically includes:

[0044] Step 21. Establish an initial key indicator system and collect data;

[0045] Step 22. Preprocess the collected data to obtain a standardized indicator dataset;

[0046] Step 23 calculates the mutual information value between each indicator and the fire hazard level label using a spatiotemporal weighted mutual information algorithm, with the following formula:

[0047] ;

[0048] Where I w (i,Y) represents the spatiotemporal weighted mutual information value between indicator i and fire risk level Y, T represents the total number of time windows, S represents the total number of spatial zones of the forest area divided by a certain area, and x i,t,s Let y be the value of the i-th key indicator in the t-th time window and the s-th spatial partition. t,s For the fire hazard level label of the corresponding spatiotemporal region, p() is the probability distribution within that spatiotemporal unit; ω t For time weights, ω s For spatial weights.

[0049] Preferably, the preprocessing specifically includes:

[0050] Step 221. Perform outlier identification and removal, mark values ​​exceeding the threshold as outliers, and use box plot method to identify outliers without clear environmental threshold indicators, and remove all outliers;

[0051] Step 222: performing missing value filling on different types of data;

[0052] Step 223: performing normalization processing, adopting the extreme value mapping method exclusive to forest areas, mapping the value of each index to the interval [0, 1], and finally obtaining a standardized index data set.

[0053] Preferably, step 3 further comprises a data screening step, specifically:

[0054] Step 33: setting a consistency ratio threshold CRT, and calculating the consistency ratio CR of the pairwise judgment matrix A;

[0055] CR=CI / RI,

[0056] wherein RI is an average random consistency index, and CI is a consistency index;

[0057] CI=(λ max -n) / (n-1)λ max represents the maximum eigenvalue of the pairwise judgment matrix A, and represents the maximum value among the eigenvalues of the 13 weight indicators

[0058] that constitute the pairwise judgment matrix A;

[0059] when the consistency ratio CR<CRT, it is considered reasonable and enters the subsequent steps; otherwise, it is considered unreasonable, and the unreasonable basic spatial unit does not enter the subsequent steps.

[0060] Preferably, in step 31, for matrix element a in the analytic hierarchy process ij , the assignment method is defined as follows:

[0061] M gears are set for assignment, M is an odd number, which are arranged in ascending order as D(1), D(2)...D(M), wherein the middle item, namely D((1+M) / 2), is set to 1, n is the number of key indicators, which is also the number of rows and columns of the pairwise analysis matrix, and E is any positive number;

[0062] for values of other gears, for those whose serial numbers are greater than (1+M) / 2, the values are taken as E 1 , E 2 ... in ascending order; for those whose serial numbers are less than (1+M) / 2, the values are taken as E -1 , E -2 ... in descending order;

[0063] A table is constructed, the table stores 2M+1 numerical values, which respectively represent the value obtained by taking the p-th power of E and then extracting the n-th root, that is, storing E p / n , p=-M,-(M-1)...1,2...M,

[0064] In step 32, when performing a multiplication operation on the weight calculation formula, the multiplication of the values ​​expressed in exponential form is simplified to the addition of the exponents; and the calculation result is obtained by looking up a table.

[0065] Preferably, in step 5, the neural network regression model employs a dual-branch CNN-LSTM hybrid neural network, which includes a CNN branch that outputs spatial feature vectors, an LSTM branch that outputs temporal feature vectors, and a residual connection layer that connects to the CNN branch and the LSTM branch.

[0066] Preferably, the CNN branch has two convolutional layers, with ReLU activation function; a max pooling layer is connected after each convolutional layer; and a flattening layer is connected after each pooling layer.

[0067] The LSTM branch includes two LSTM layers, with a DROP-OUT layer between the two LSTM layers. The output of the second LSTM layer is used as the temporal feature vector of the dynamic index and passed to the subsequent residual connection module.

[0068] The residual connection module includes a feature dimension matching layer and a residual block. The feature dimension matching layer expands the dimensions of the spatial feature vector output by the CNN branch and the temporal feature vector output by the LSTM branch using zero-padding, so that the dimensions of the two are consistent and then added together. The residual block contains two fully connected layers. The first fully connected layer has 128 neurons and uses the ReLU function for activation. The second fully connected layer has the same number of neurons as the feature vector dimension and uses the Sigmoid function for activation. The feature dimension matching layer adds the dimension-matched spatial feature vector and the temporal feature vector to obtain the initial fused feature. The initial fused feature is then input into the residual block to obtain the final fused feature vector. The final fused feature vector is then passed to the fully connected layer.

[0069] Preferably, in step 5, a multilayer perceptron model is used for neural network learning and prediction, such as... Figure 2 As shown, the multilayer perceptron model includes:

[0070] (1) Input layer, with n input nodes, the same number as the key indicators;

[0071] (2) First hidden layer: It has 20 output nodes and a ReLU activation function;

[0072] (3) Second hidden layer: It has 20 output nodes and a ReLU activation function;

[0073] (4) Dropout layer: Set the dropout rate to 0.5;

[0074] (5) Output layer: 1 node, outputting the predicted value of forest and grassland fire hazard index of the b-th basic spatial unit. .

[0075] Preferably, the loss function L is set to mean squared error during training:

[0076] ;

[0077] The model training is complete when the loss function L is less than the set learning target.

[0078] Preferably, step 6 is included after step 5, which calculates the regional fire hazard index of the region. Specifically, the formula for calculating the regional fire hazard index HA of region A, which contains m basic spatial units, is as follows:

[0079] ;

[0080] in Sb is the predicted value of the forest-grassland fire hazard index for the b-th basic spatial unit, and Sb is the area of ​​the b-th basic spatial unit.

[0081] This invention integrates detailed, multi-dimensional real-world forest and grassland fire risk data obtained from forest and grassland fire risk surveys, including data on combustible materials, fire weather, wildfire sources, and topography. It utilizes basic grid cells as model training data and employs a multilayer perceptron neural network to fit complex nonlinear relationships, thereby improving prediction accuracy. Compared to existing forest and grassland fire risk level forecasts that are primarily driven by weather conditions, this invention has promising engineering applications and social benefits. Attached Figure Description

[0082] Figure 1 This is a schematic diagram of a specific implementation of the regional fire risk prediction method based on multi-index data according to the present invention.

[0083] Figure 2 This is a schematic diagram of a specific implementation of the multilayer perceptron model described in this invention;

[0084] Figure 3 This is a schematic diagram of a specific implementation of the dual-branch CNN-LSTM hybrid neural network described in this invention;

[0085] Figure 4 In a specific embodiment of the present invention, Figure 2 The curves show the mean square error, coefficient of determination, and root mean square error as the loss function when the model is trained and tested.

[0086] Figure 5This is a schematic diagram comparing the predicted value of the forest and grassland fire risk index and the calculated forest and grassland fire risk index obtained in a specific embodiment of the present invention.

[0087] Figure 6 This is a spatial distribution map of different fire risk levels of basic spatial units in a certain area obtained using the method described in this invention;

[0088] Figure 7 To obtain using the method described in this invention Figure 5 The area shown is a spatial distribution map of different fire risk levels under township-level administrative divisions. Detailed Implementation

[0089] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0090] The regional fire risk prediction method based on multi-index data described in this invention, such as... Figure 1 As shown, it includes the following steps:

[0091] Step 1: Constructing vector grid cells;

[0092] A vector grid is generated according to the set latitude and longitude intervals in a coordinate system established by the World Geodetic Coordinate System, such as WGS-84, as the basic spatial unit;

[0093] For example, taking Sichuan Province as the study area, a 30-arc-second vector grid is generated within the administrative area of ​​Sichuan Province using a latitude and longitude interval of 0.008333° × 0.008333° as the basic spatial unit. The set of all basic spatial units is used as the set G of basic spatial units for the next step of analysis and calculation.

[0094] Through spatial calculations, a total of 658,948 grids were generated within the Sichuan province, and various forest and grassland fire risk survey data for the region were spatially mapped according to the grid units.

[0095] Step 2: Construction of Key Indicator System

[0096] Step 21. Establish an initial key indicator system. In this embodiment, the key indicators constituting the risk of forest and grassland fires are divided into 4 groups and 14 items:

[0097] 1. Based on the forest and grassland fire risk survey, four indicators of forest and grassland combustibles were obtained: total combustible load, fine combustible load, combustible combustibility, and combustible moisture content.

[0098] 2. Three fire weather indicators obtained from the forest and grassland fire risk survey, and the average temperature, average precipitation, and average wind speed during the fire prevention period calculated based on meteorological data from the past 20 years.

[0099] 3. Based on the forest and grassland fire risk survey, four indicators of wildfire sources were obtained: road density, frequency of fire use, number of key populations, and frequency of forest ranger patrols.

[0100] 4. Based on the topographic indicators obtained from the forest and grassland fire risk survey, three indicators were obtained: slope, aspect, and altitude.

[0101] Step 22. After the above 14 data items are collected, preprocessing can be performed. Preprocessing specifically includes:

[0102] Step 221. Perform outlier identification and removal. Combine the actual environmental threshold of the forest area, mark the values ​​that exceed the threshold as outliers. At the same time, use the box plot method to identify outliers without clear environmental threshold indicators and remove all outliers.

[0103] Step 222. Fill missing values. For fire weather indicators, use the neighborhood time window interpolation method and fill with the average of the three time windows before and after the indicator. For topographic and geomorphological indicators and forest and grassland combustibles indicators, use the neighborhood spatial partition interpolation method and fill with the average of the four spatial partitions around the indicator. For human activity indicators such as wildfire sources, fill with the average of the same period in history.

[0104] Step 223. Perform normalization processing. Use the forest area-specific extreme value mapping method to map the value of each indicator to the interval [0, 1] to obtain a standardized indicator dataset, thereby ensuring the validity and consistency of the indicator data.

[0105] Step 23. For the preprocessed standardized index dataset, calculate the mutual information value between each index and the fire risk level label, as well as the mutual information value between the indices. By setting a mutual information threshold, redundant indices are eliminated to obtain a subset of selected indices.

[0106] Step 23 calculates the mutual information value between each indicator and the fire hazard level label using a spatiotemporal weighted mutual information algorithm, with the following formula:

[0107] ;

[0108] Where I w (i,Y) represents the spatiotemporal weighted mutual information value between indicator i and fire risk level Y, T represents the total number of time windows divided into 7-day periods, S represents the total number of spatial zones of the forest area divided into certain areas, and x represents the total number of spatial zones of the forest area divided into certain areas. i,t,s Let y be the value of indicator i in the t-th time window and the s-th spatial partition. t,s For the fire hazard level label of the corresponding spatiotemporal region, P() is the probability distribution within that spatiotemporal unit; ω t The time weight is determined by the proportion of valid data for indicator i in the t-th time window. Valid data proportion = (Number of non-missing data in that window) / (Total number of data in the window), ω sThe spatial weight is determined by the historical frequency of forest fire risk in the s-th spatial partition, where the frequency is calculated as: historical fire risk number in the partition / total forest fire risk number.

[0109] The specific operational steps are as follows: First, divide the target forest area into time windows and spatial partitions; then calculate the probability distribution of each spatiotemporal unit; and substitute the results into the formula to obtain I. w (i, Y), finally set I w Indicators with (i,Y) less than the mutual information threshold of 0.2 are considered redundant and are removed. This is to avoid misjudging of redundant indicators caused by the general mutual information ignoring spatiotemporal differences when collecting data in step 21.

[0110] In one specific embodiment, the time window is set to 7 days, and the collection period is 140 days, so the total number of time windows T=20; the spatial partitions are divided into 0.1km×0.1km grids, resulting in 500 spatial partitions, i.e., S=500. Based on the correlation between historical fire risk and indicators in this forest area, a mutual information threshold of 0.2 is set, and indicators with mutual information values ​​lower than 0.2 with the fire risk level label are removed. Finally, the forest ranger patrol frequency indicator, which is highly correlated with road density, is removed.

[0111] After removal, the remaining 13 key indicators of each basic spatial unit are used to generate a 13-dimensional key indicator feature vector Xb through spatial mapping, i.e., n=13, which serves as the feature data for subsequent neural network training.

[0112] Xb=(x b1 ,x b2 ,...x b13 ), where Xb represents the eigenvector of the b-th basic spatial unit, x b1 ,x b2 ,...x b13 This represents the first to thirteenth key indicators of the b-th basic spatial unit, which constitute the first to thirteenth vector elements of the key indicator feature vector Xb;

[0113] Step 3: Calculate and verify the weights of key indicators using the Analytic Hierarchy Process (AHP) for each basic spatial unit;

[0114] Step 31. Based on the 13 key indicators in Step 2, construct a pairwise judgment matrix A for each indicator using the Analytic Hierarchy Process (AHP):

[0115] A=(a ij ), where matrix element a ij The subscripts i and j represent the row and column of the matrix, and each matrix element a ij This indicates the importance of the i-th key indicator to the j-th key indicator. For example, for a certain basic space unit, if the importance of the third indicator, the flammability of combustibles, to the fifth indicator, temperature, is 3.0, then a35 =3.0.

[0116] Step 32. Calculate the weight w of the i-th key indicator according to the weight calculation formula. i :

[0117] ;

[0118] When n=13, Formula 1 becomes

[0119] ;

[0120] The weight vectors of the 13 key indicators were obtained by solving the problem.

[0121] W = (w1, w2, ... w) 13 );

[0122] In the existing analytic hierarchy process (AHP), the assignment of values ​​to each element usually adopts a 1-9 level assignment method, typically 1 / 9, 1 / 7, 1 / 5, 1 / 3, 1, 3, 5, 7, 9. The pairwise judgment matrix A in the AHP is essentially a proportional scale relationship rather than an additive scale relationship. The use of multiplication and square root extraction in Formula 1 can maintain the proportional meaning of the feature data.

[0123] However, the high-order multiplication and square root calculations in the weight calculation formula also bring the disadvantage of large computational cost, especially when updating large areas of large grid data. Under the condition of limited hardware and software resources, it will reduce the model training and prediction time.

[0124] In this invention, to simplify the calculation of the weight calculation formula, the matrix element a in the analytic hierarchy process is... ij The assignment method is defined as follows:

[0125] The assignment is set to M, where M is an odd number, and the values ​​are arranged in ascending order as D(1), D(2)...D(M), with the middle term D((1+M) / 2) set to 1. n is the number of key indicators, which is also the number of rows and columns of the pairwise analysis matrix, and E is any positive number.

[0126] For other values, those with serial numbers greater than (1+M) / 2 are respectively taken as E from smallest to largest. 1 E 2 ...For serial numbers less than (1+M) / 2, take E from largest to smallest. -1 E -2 ...

[0127] For example, assuming E=2, n=4, M=9, the middle item is the fifth item, D(5)=1, and the serial numbers D(6) to D(9) which are greater than D(5) are 2, 4, 8, and 16 respectively; the serial numbers D(1) to D(4) which are less than D(5) are 1 / 16, 1 / 8, 1 / 4, and 1 / 2 respectively.

[0128] The above assignment method maintains the multi-level, reciprocal assignment method of the existing hierarchical analysis method, while widening the gap between each level. The example of E=2 is merely for illustration. To control the difference between the values ​​of each level, E can also be an integer greater than 1 and less than 2. For example, when E=1.5, the first five powers are approximate to 1.5, 2.25, 3.38, 5.63, and 7.60, respectively, which are close to the intervals of 1, 3, 5, 7, and 9 in the commonly used 9-level setting method in the existing paired judgment matrix A.

[0129] When performing multiplication using the above assignment method, multiplying the assignments expressed in exponential form simplifies to adding the exponents.

[0130] For example, suppose E=2, n=4, M=9, and the values ​​assigned to the third column of the pairwise judgment matrix A from top to bottom are 2. 1 2 3 2 1 2 -1 2 1 2 1 2 -4 2 1 2 -2 Multiplication is converted into the addition of exponents. The sum of the exponents is 1, and taking the fourth root gives 2. 1 / 4 .

[0131] The sum of the exponents is an integer. For example, if the sum of the exponents is 'a', then the product of the exponents is E. a .

[0132] For E a We can use a lookup table instead of performing specific calculations. Analyzing the possible forms of the sum of exponents, we can conclude that the value 'a' after adding the exponents must be distributed within the integer range of [-M*n, M*n], meaning there can only be 2M*n+1 values. Since integer multiples of n can be directly raised to the nth power, the actual number of values ​​that need to be calculated is at most 2M+1. When 2M>n, this number may be even smaller.

[0133] Therefore, a table can be constructed that stores 2M+1 values, representing the p-th power of the n-th root of E, i.e., storing E p / n, p=-M, -(M-1)...1, 2...M, calculation can be performed by table lookup, and there is no need for the tedious calculation of continuous multiplication followed by square root extraction, thereby increasing the operation speed and reducing the occupation of hardware resources.

[0134] In a preferred embodiment of step 3, data screening is required, specifically

[0135] Step 33. Set a consistency ratio threshold CRT, calculate the consistency ratio CR of the pairwise judgment matrix A. When CR<CRT, it is considered reasonable, otherwise it is considered unreasonable;

[0136] For example, when the consistency ratio threshold CRT=0.1

[0137] ;

[0138] where RI (Random Index) is the average random consistency index. In the analytic hierarchy process, RI is determined by the number n of key indicators. When n=13, RI=1.56 can be obtained by table lookup.

[0139] CI is the consistency index,

[0140] CI=(λ max -n) / (n-1); λ max represents the maximum eigenvalue of the pairwise judgment matrix A, which constitutes the

[0141] maximum value of the eigenvalues of 13 weight indicators of the pairwise judgment matrix A;

[0142] When the consistency ratio CR<CRT, it is considered reasonable and the process proceeds to the subsequent steps; otherwise, it is considered unreasonable. For unreasonable basic spatial units, return to step 2 to check whether the collected data is accurate, re-perform the calculation process of this step after updating the data, and the consistency ratio threshold CRT can also be adjusted.

[0143] Step 4: Data standardization and calculation of forest and grassland fire risk index;

[0144] Since various key indicators have different measurement units, direct calculation and statistical analysis cannot be performed. It is necessary to perform standardization processing on the data, unify the dimensions, and then calculate the forest and grassland fire risk index.

[0145] Perform Z-score standardization on the key indicator data of all basic spatial units collected in step 2, and the standardized key indicator data

[0146] ;

[0147] μ i is the average value of the i-th key indicator of all basic spatial units, σ iLet x be the standard deviation of the i-th key indicator, where x i This is the raw data for the i-th key indicator.

[0148] The standardized key indicator data no longer have dimensions. Based on the standardized key indicator data, the forest and grassland fire hazard index H of each basic spatial unit is calculated. b :

[0149] ;

[0150] Where w i The weight of the i-th key indicator, The standardized value of the i-th key indicator of the grid cell is used to calculate the forest and grassland fire hazard index H. b Label data used for training neural networks.

[0151] Step 5: Building and training the neural network regression model;

[0152] The key indicator feature vector Xb obtained in step 2 and the label data obtained in step 4 are used as the neural network dataset. The neural network dataset is divided into training set and test set in a ratio of 8:2.

[0153] In step 5, a multilayer perceptron (MLP) model can be used for neural network learning and prediction, such as... Figure 2 As shown, the MLP model includes:

[0154] (1) Input layer, with 13 input nodes, the same number as the key indicators;

[0155] (2) First hidden layer: It has 20 output nodes and a ReLU activation function;

[0156] (3) Second hidden layer: It has 20 output nodes and a ReLU activation function;

[0157] (4) Dropout layer: Set the dropout rate to 0.5;

[0158] (5) Output layer: 1 node, outputting the predicted value of forest and grassland fire hazard index of the b-th basic spatial unit. .

[0159] The model is trained using N samples, representing data from N basic spatial units, and the loss function L is the mean squared error (MSE).

[0160] ;

[0161] H bThis represents the forest and grassland fire hazard index of the b-th basic spatial unit calculated according to step 4; the model training is complete when the loss function L is less than the set learning target.

[0162] The Adam optimizer (Adaptive Moment Estimation optimizer) can be used for parameter optimization, with a learning rate of 0.001 and 1000 training epochs.

[0163] During training, the coefficient of determination (R²) and the root mean square error (RMSE) are used as metrics to evaluate model accuracy, respectively:

[0164] ;

[0165] ;

[0166] Where N represents the sample size, H b This represents the forest-grassland fire hazard index of the b-th basic spatial unit calculated according to step 4. This represents the average value of the 1st, 2nd, ..., Nth basic spatial units. This represents the predicted value of the forest and grassland fire hazard index for the b-th basic spatial unit.

[0167] like Figure 4 As shown, the present invention is presented in the form of Figure 2 The curves shown represent the mean squared error, coefficient of determination, and root mean squared error as loss functions during model training and testing. The horizontal axis represents the number of training iterations, and the vertical axis represents the specific values. After 400 training iterations, the various indicators tend to stabilize.

[0168] like Figure 5 The image shows the predicted and calculated values ​​of the forest and grassland fire hazard index after the model has been trained and input into the test set. Figure 4 The horizontal axis represents different basic spatial units, and the difference between the vertical axes at the top and bottom of each vertical line represents the predicted value and the calculated value. It can be seen that the predicted value and the calculated value of the model after training are very close.

[0169] And set the optimal values ​​for the coefficient of determination R² and the root mean square error RMT. The model parameters with R² ≥ RT and RMT ≤ RMT are determined as the optimal model parameters, such as... Figure 4 , Figure 5As shown, RT=0.7 and RMT=0.2 were set. After multiple rounds of training and verification, the model parameters with a coefficient of determination R²≥0.7 and root mean square error RMSE≤0.2 were determined as the optimal model parameters. For models that meet the requirements, the deep learning training model with the largest R² / RMSE ratio was selected as the final model ME, and this model was used to carry out the risk index prediction calculation for each basic spatial unit.

[0170] In a preferred embodiment, step 5 employs a dual-branch CNN-LSTM hybrid neural network for neural network learning and prediction, such as... Figure 3 The dual-branch CNN-LSTM hybrid neural network shown includes a CNN branch, an LSTM branch, and a residual connection layer:

[0171] In the CNN branch, the convolutional kernel parameters are set according to the spatial resolution of the target forest area terrain. The kernel size is set to 3×3, and each kernel corresponds to the indicator features of 3 adjacent spatial partitions. The CNN branch has 2 convolutional layers. The first convolutional layer has 16 kernels and uses the ReLU activation function to adapt to the non-linear features of vegetation and terrain indicators. The second convolutional layer has 32 kernels and also uses the ReLU activation function. After each convolutional layer, a max pooling layer with a 2×2 kernel size is connected to extract the key features of each spatial region. After the pooling layer, a flattening layer is connected to flatten the features output by the pooling layer to obtain the spatial feature vector of the static indicators, thereby achieving accurate extraction of the static features of the forest area.

[0172] In the LSTM branch of the dual-branch CNN-LSTM hybrid neural network, the time window length is set to 14, taking into account the 14-day average cycle from potential forest fire risk to occurrence. Two LSTM layers are constructed, with 64 hidden units in the first LSTM layer and 32 hidden units in the second LSTM layer. A dropout mechanism with a dropout rate of 0.2 is used between layers to avoid overfitting. In the fourth step, the output of the second LSTM layer is used as the time-series feature vector of the dynamic index and passed to the subsequent residual connection module to achieve the capture of the time-series evolution of the dynamic features of the forest area.

[0173] The residual connection module includes a feature dimension matching layer, which performs feature dimension matching by expanding the dimensions of the spatial feature vector output from the CNN branch and the temporal feature vector output from the LSTM branch using zero-padding, ensuring they are of the same dimension before adding them. The residual block contains two fully connected layers: the first fully connected layer has 128 neurons and uses the ReLU activation function, while the second fully connected layer has the same number of neurons as the feature vector dimension and uses the Sigmoid activation function. The feature dimension matching layer adds the dimension-matched spatial feature vector to the temporal feature vector to obtain the initial fused feature, which is then input into the residual block to obtain the final fused feature vector. This final fused feature vector is then passed to the fully connected layer to improve the fusion effect of the two-branch features and the training stability of the model. Finally, the fully connected layer outputs the predicted value of the forest-grassland fire hazard index for the b-th basic spatial unit. .

[0174] The dual-branch CNN-LSTM hybrid neural network combines the spatial feature extraction capability of the CNN branch with the temporal feature capture capability of the LSTM branch, effectively fusing the spatial features of static indicators and the temporal features of dynamic indicators. Furthermore, the residual connection module enhances the feature fusion effect and the training stability of the model.

[0175] Step 6: Predicting the regional risk index;

[0176] For the basic spatial unit that needs to be predicted, after collecting the key indicator data of the basic spatial unit, the key indicator feature vector Xb is calculated according to step 2, and input into the final model ME obtained in step 5 to train and output the predicted value of the forest and grassland fire hazard index of the basic spatial unit. .

[0177] For a given area containing multiple basic spatial units, the predicted forest-grassland fire hazard index value of the grid unit is weighted according to the area of ​​all basic spatial units contained in the area to obtain the regional fire hazard index of the area.

[0178] This method can be applied to existing administrative blocks such as towns (townships) and counties (cities, districts). Each administrative division contains multiple contiguous basic spatial units. For basic spatial units at the intersection of multiple administrative blocks, it is possible to manually assign which administrative block the basic spatial unit belongs to.

[0179] Specifically, the fire hazard index H of region A, which contains m basic spatial units. A The calculation formula is:

[0180] ;

[0181] in S is the predicted value of the forest-grassland fire hazard index for the b-th basic spatial unit. b Let be the area of ​​the b-th basic spatial unit. It should be noted that although the areas of each basic spatial unit divided in step 1 are theoretically the same, due to the curvature of the Earth's sphere, the areas of each basic spatial unit vary with latitude and longitude. At the same time, in actual operation, in order to reduce the amount of calculation, the areas of each basic spatial unit will be excluded from areas such as water bodies and urban areas where forest and grassland fires are unlikely to occur. Therefore, the areas of each basic spatial unit are not completely consistent.

[0182] Using the Jenks method (natural discontinuity method), based on the predicted forest and grassland fire hazard index, each basic spatial unit and the various regions derived from the township-level administrative divisions are divided into four different fire risk levels: high risk, medium-high risk, medium-low risk, and low risk. Figure 6 The image shows a spatial distribution map of different fire risk levels for various basic spatial units within a certain area of ​​Sichuan Province, obtained using the method described in this invention. Figure 7 The figure shows the result obtained using the method described in this invention. Figure 6 A spatial distribution map of different fire risk levels in various townships within the same region, based on administrative divisions.

[0183] This invention integrates detailed, multi-dimensional real-world forest and grassland fire risk data obtained from forest and grassland fire risk surveys, including data on combustible materials, fire weather, wildfire sources, and topography. It utilizes basic grid cells as model training data and employs a multilayer perceptron neural network to fit complex nonlinear relationships, thereby improving prediction accuracy. Compared to existing forest and grassland fire risk level forecasts that are primarily driven by weather conditions, this invention has promising engineering applications and social benefits.

[0184] The foregoing descriptions are preferred embodiments of the present invention. Unless there is a clear contradiction between the preferred embodiments or a prerequisite for a particular preferred embodiment, the preferred embodiments can be arbitrarily combined and used. The embodiments and specific parameters described are only for clearly illustrating the inventor's invention verification process and are not intended to limit the scope of patent protection of the present invention. The scope of patent protection of the present invention shall still be determined by its claims. Similarly, any equivalent structural changes made based on the description and drawings of the present invention shall also be included within the scope of protection of the present invention.

Claims

1. A method for predicting regional fire risk based on multi-indicator data, characterized in that, Comprising the following steps: Step 1. Generating a vector grid for the prediction area according to a set longitude and latitude interval, which is used as a basic spatial unit; Step 2. Establish a key indicator system, including n key indicators x related to the probability of fire occurrence. b1 ,x b2 ,...x bn ; Generate an n-dimensional key indicator feature vector Xb, Xb=(x b1 ,x b2 ,...x bn ); Step 3. Performing analytic hierarchy process weight calculation and verification on key indicators for each basic spatial unit; Step 31. Constructing a pairwise judgment matrix A for each key indicator based on the analytic hierarchy process according to the key indicators in step 2: A=(a ij ), where matrix element a ij The subscripts i and j represent the row and column of the matrix, and each matrix element a ij This indicates the importance of the i-th key indicator to the j-th key indicator; Step 32. Calculate the weight w of the i-th key indicator according to the weight calculation formula. i : ； Solving to obtain the weight vector of each key indicator: W=(w1,w2,...w n ); Step 4: Data standardization and calculation of forest and grassland fire risk index; The key indicator data of all basic spatial units collected in step 2 are standardized. The standardized key indicator data are as follows: : ； μ i σ is the average of the i-th key indicator of all basic spatial units. i Let x be the standard deviation of the i-th key indicator. i This is the raw data for the i-th key indicator; Based on standardized key indicator data Calculate the forest and grassland fire hazard index H for each basic spatial unit. b : ； Where w i The weight of the i-th key indicator, The standardized value of the i-th key indicator of the grid cell is used to calculate the forest and grassland fire hazard index H. b Label data used for training neural networks; Step 5: Construction and training of neural network regression model; Constructing a neural network regression model, taking the key indicator feature vector Xb obtained in step 2 and the label data obtained in step 4 as a neural network data set, and dividing the neural network data set into a training set and a test set; Training the neural network regression model, where the number of training samples is N, that is, data of N basic spatial units; Using the coefficient of determination R² and the root mean square error RMSE as model accuracy evaluation indicators, which are respectively: ； ； Where N represents the sample size, H b This represents the forest-grassland fire hazard index of the b-th basic spatial unit calculated according to step 4. Represents N H b The average value, This represents the predicted value of the forest and grassland fire hazard index for the b-th basic spatial unit. Setting an optimal coefficient of determination value RT and an optimal root mean square error value RMT, determining model parameters with the coefficient of determination R²≥RT and the root mean square error RMSE≤RMT as optimal model parameters, and determining the corresponding obtained deep learning training model as a final model ME; Step 6: Prediction of regional risk index; For the basic spatial unit that needs to be predicted, after collecting the key indicator data of the basic spatial unit, the forest and grassland fire hazard index H is calculated according to steps 2 to 4. b The key indicator feature vector Xb is input into the final model ME obtained in step 5, and the model is trained to output the predicted value of the forest and grassland fire hazard index for this basic spatial unit. .

2. The regional fire risk prediction method based on multi-index data as described in claim 1, characterized in that, The step 2 specifically comprises: Step 21. Establishing an initial key indicator system and collecting data; Step 22. Preprocessing the collected data to obtain a standardized indicator data set; Step 23. For the preprocessed standardized indicator data set, calculating mutual information values between each indicator and fire risk level labels and mutual information values between indicators, eliminating redundant indicators by setting a mutual information threshold, and obtaining a filtered indicator subset; Calculate mutual information value The formula is: ； Where I w (i,Y) represents the spatiotemporal weighted mutual information value between indicator i and fire risk level Y, T represents the total number of time windows, S represents the total number of spatial zones of the forest area divided by a certain area, and x i,t,s Let y be the value of the i-th key indicator in the t-th time window and the s-th spatial partition. t,s For the fire hazard level label of the corresponding spatiotemporal region, p() is the probability distribution within that spatiotemporal unit; ω t For time weights, ω s For spatial weights.

3. The regional fire risk prediction method based on multi-index data as described in claim 2, characterized in that, The preprocessing specifically comprises: Step 221. Identifying and eliminating outliers, marking values exceeding the threshold as outliers, simultaneously identifying outliers of indicators without clear environmental thresholds by a box plot method, and eliminating all outliers; Step 222. Performing missing value filling for different types of data; Step 223. Performing normalization processing, adopting a forest region-specific extreme value mapping method to map the value of each indicator to the interval [0, 1], and finally obtaining the standardized indicator data set.

4. The regional fire risk prediction method based on multi-index data as described in claim 1, characterized in that, The step 3 further comprises a data screening step, specifically: Step 33. Setting a consistency ratio threshold CRT, and calculating the consistency ratio CR of the pairwise judgment matrix A; CR=CI / RI, wherein RI is an average random consistency index, and CI is a consistency index; CI=(λ max -n) / (n-1); λ max represents the largest eigenvalue of the pairwise judgment matrix A, and represents the maximum value of the eigenvalues ​​of the 13 weight indices that constitute the pairwise judgment matrix A; When the consistency ratio CR<CRT, it is considered reasonable and the process proceeds to the subsequent steps; otherwise it is considered unreasonable, and the unreasonable basic spatial units do not enter the subsequent steps.

5. The regional fire risk prediction method based on multi-index data as described in claim 1, characterized in that, In step 31, the matrix element a in the analytic hierarchy process... ij The assignment method is defined as follows: Setting M gears for value assignment, M is an odd number, which are arranged from small to large as D(1), D(2)...D(M), wherein the middle term, namely D((1+M) / 2), is set to 1, n is the number of key indicators, and E is any positive number; For other values, those with serial numbers greater than (1+M) / 2 are respectively taken as E from smallest to largest. 1 E 2 ...For serial numbers less than (1+M) / 2, take E from largest to smallest. -1 E -2 ...; In step 32, a table is constructed, which stores 2M+1 values, representing the p-th power of the n-th root of E, i.e., storing E. p / n p = -M, -(M-1)...1, 2...M; When the weight calculation formula performs continuous multiplication, the multiplication of assignments expressed in various exponential forms is simplified to the addition of exponents, and the calculation result is obtained by looking up a table.

6. The regional fire risk prediction method based on multi-index data as described in claim 1, characterized in that, In step 5, the neural network regression model adopts a dual-branch CNN-LSTM hybrid neural network, which includes a CNN branch that outputs spatial feature vectors, an LSTM branch that outputs temporal feature vectors, and a residual connection layer that connects to the CNN branch and the LSTM branch.

7. The regional fire risk prediction method based on multi-index data as described in claim 6, characterized in that, The CNN branch has two convolutional layers, with ReLU activation function; a max pooling layer is connected after each convolutional layer; and a flattening layer is connected after each pooling layer. The LSTM branch includes two LSTM layers, with a DROP-OUT layer between the two LSTM layers. The output of the second LSTM layer is used as the temporal feature vector of the dynamic index and passed to the subsequent residual connection module. The residual connection module includes a feature dimension matching layer and a residual block. The feature dimension matching layer expands the dimensions of the spatial feature vector output by the CNN branch and the temporal feature vector output by the LSTM branch using zero-padding, making their dimensions consistent and then adding them together. The residual block contains two fully connected layers. The first fully connected layer has 128 neurons and uses the ReLU activation function. The second fully connected layer has the same number of neurons as the feature vector dimension and uses the Sigmoid activation function. The feature dimension matching layer adds the dimension-matched spatial feature vector and the temporal feature vector to obtain the initial fused feature, and then inputs the initial fused feature into the residual block to obtain the final fused feature vector. The final fused feature vector is then passed to the fully connected layer.

8. The regional fire risk prediction method based on multi-index data as described in claim 1, characterized in that, In step 5, a multilayer perceptron model is used for neural network learning and prediction. The multilayer perceptron model includes: (1) Input layer, with n input nodes, the same number as the key indicators; (2) First hidden layer: It has 20 output nodes and a ReLU activation function; (3) Second hidden layer: It has 20 output nodes and a ReLU activation function; (4) Dropout layer: Set the dropout rate to 0.5; (5) Output layer: 1 node, outputting the predicted value of forest and grassland fire hazard index of the b-th basic spatial unit. .

9. The regional fire risk prediction method based on multi-index data as described in claim 1, characterized in that, During training, the loss function L is set to mean squared error: ； The model training is complete when the loss function L is less than the set learning target.

10. The regional fire risk prediction method based on multi-index data as described in claim 1, characterized in that, Step 5 is followed by step 6, which calculates the regional fire hazard index for the region. Specifically, the regional fire hazard index H of region A, which contains m basic spatial units, is calculated. A The calculation formula is: ； in Sb is the predicted value of the forest-grassland fire hazard index for the b-th basic spatial unit, and Sb is the area of ​​the b-th basic spatial unit.

Images