A method for calculating the probability of high-intensity forest fires by combining BiLSTM and kernel density estimation

Abstract

This invention discloses a method for calculating the probability of high-intensity forest fires by combining BiLSTM and kernel density estimation, belonging to the field of remote sensing inversion technology. The method for assessing the probability of high-intensity forest fires described in this invention is simple to operate, requiring only input of information on combustible materials, meteorology, topography, and human activities. It can quickly assess the probability of high-intensity forest fires based on BiLSTM quantile regression and kernel density estimation. This method fully considers key risk factors for high-intensity forest fires, incorporating the temporal variation characteristics of combustible materials and meteorological factors. It is well-suited for estimating the probability of high-intensity forest fires in complex forest areas and is of great significance for high-intensity forest fire risk monitoring and early warning.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing inversion technology, specifically relating to a method for assessing the probability of high-intensity forest fires by combining BiLSTM and kernel density estimation. Background Technology

[0002] Wildfires severely damage the ecological environment, impact global climate, and cause enormous losses to people's lives and property. Typical low- to medium-intensity forest fires spread at speeds on the order of 100 m / h, while high-intensity forest fires spread at speeds on the order of 1 km / h or have a fire line intensity of 10... 3 High-intensity forest fires, occurring in the kW / m range, account for only 3% of all forest fires, but cause up to 95% of global forest fire losses. Therefore, strengthening high-intensity forest fire risk monitoring and early warning, curbing major forest fires, and minimizing enormous losses are equally urgent and important!

[0003] Combustible materials, topography, and meteorology are closely related to the occurrence and development of forest fires, and these factors are all incorporated into forest fire behavior models to simulate forest fire spread. However, forest fire behavior models have significant uncertainties when applied to complex forest areas, making it difficult to achieve high-precision, large-scale, high-intensity forest fire probability assessments. Furthermore, the forest fire spread process is influenced by various uncontrollable factors, and prediction methods based on forest fire behavior models cannot measure the uncertainty of simulated forest fire behavior characteristics. Satellite remote sensing technology has the advantage of large-scale, high-frequency Earth observation, providing an effective way to extract forest fire spread characteristics such as spread rate and fire radiation intensity. Currently, publicly available global fire datasets, such as Global FireAtlas and MOD14A1 / MYD14A1, have been used to extract spread rate and fire radiation intensity. Using these datasets and combining them with machine learning methods, nonlinear relationships between combustible materials, topography, meteorological factors, and historical forest fire behavior characteristics can be established, providing a new approach for predicting potential forest fire behavior characteristics and assessing the probability of high-intensity forest fires.

[0004] However, there is currently very little research on the probability assessment of high-intensity forest fires. The key to achieving wildfire risk early warning is to utilize multi-category combustible materials, meteorological and topographical information, and historical forest fire behavior characteristic datasets to construct a predictive model of potential forest fire behavior characteristics and conduct a probability assessment of high-intensity forest fires. This will help improve the accuracy of forest fire prevention and control. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a highly scalable and computationally simple method for assessing the probability of high-intensity forest fires based on BiLSTM and kernel density estimation.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a method for calculating the probability of high-intensity forest fires by combining BiLSTM and kernel density estimation, comprising the following steps:

[0007] Step 1: Determine the high-intensity forest fire risk factors for the preset area. The high-intensity forest fire risk factors include: combustible material factors, meteorological factors, and human activity factors. The combustible material factors include: canopy and surface combustible material moisture content, canopy combustible material load, forest canopy height, and tree coverage. The meteorological factors include: wind speed, temperature, rainfall, and relative humidity. The topographic factors include: elevation, slope, aspect, topographic humidity index, topographic location index, and plane curvature. The human activity factors include: straight-line distance from roads and straight-line distance from residential areas.

[0008] Step 2: Using the Global Fire Atlas dataset, extract the location, date, and spread rate of ignition points for each forest fire from 2003 to 2016 that have a burning area of ​​≥2 pixels and a burning period of ≥2 days. Also, extract at least 3 features from the risk factors of high-intensity forest fires within a specific time period before the fire occurs to construct a forest fire behavior feature case library.

[0009] Step 3: Construct a potential spread rate prediction model based on a bidirectional long short-term memory network (BiLSTM) model and quantile regression. Use the constructed wildfire behavior feature case library for training and learning, and use the non-dominated sorting genetic algorithm NSGA-II to optimize the hyperparameters of the neural network model to obtain the optimal potential spread rate prediction model.

[0010] Specifically, it includes three parts:

[0011] Step 3.1: Build a bidirectional long short-term memory network (BiLSTM) model. The two outputs of the BiLSTM model correspond to the lower and upper bounds of the predicted variables, respectively. Use formula (1) as the model loss function.

[0012] (1);

[0013] Where α is the confidence interval. and q1 and q2 are the loss functions for the upper and lower bounds of the predicted variables, respectively, where q1 and q2 are quantiles, and w1 and w2 are weight coefficients.

[0014] (2);

[0015] Where N is the total number of samples, y i For the true value of the sample, This represents the predicted value of the sample when the quantile is q.

[0016] Step 3.2: Normalize the forest fire behavior feature dataset and divide it into training, validation and test sets in a ratio of 6:3:1. Use average interval coverage (PICP) and average interval width (PINAW) as the accuracy evaluation targets. Use Epoch and select Adam as the model optimizer. Export the model after training.

[0017] (3);

[0018] (4);

[0019] in, and These are the lower and upper bounds for the prediction of the i-th sample, respectively;

[0020] Step 3.3: Set confidence intervals to 90%, 80% and 70%, use the non-dominated sorting genetic algorithm NSGA-II, with formulas (3) and (4) as optimization objectives, use the comprehensive index CWC to determine the optimal BiLSTM model, and obtain the number of hidden units (n_units), dropout rate (d_rate) and batch size (b_size) of the BiLSTM layer.

[0021] (5);

[0022] in, As a penalty factor;

[0023] Step 4: Based on the optimal prediction model, output the predicted value of the potential spread rate corresponding to each quantile q, estimate the probability density distribution of the potential spread rate using the kernel density estimation method, and evaluate the probability of high-intensity forest fires using numerical integration.

[0024] Specifically, it includes four parts:

[0025] Step 4.1: Based on the optimal prediction model, using formula (2) as the cost function, output the model prediction values ​​for the points q=0.01, 0.02, ..., 0.99 respectively, thus obtaining 99 potential spread rate prediction values. When the quantile is q0, the goodness of fit R between the true value and the predicted value is... 2 Highest;

[0026] Step 4.2: Using the kernel density estimation method, a Gaussian function is selected as the kernel function. The predicted spread velocity values ​​corresponding to different quantiles are input to calculate the probability density of the potential diurnal spread velocity. ;

[0027] (6);

[0028] Where n is the total number of quantiles, i.e., n=99; h is the bandwidth; K represents the predicted potential spread rate corresponding to the i-th quantile. u It is a Gaussian function, that is:

[0029] (7);

[0030] in, Represents the independent variable that has a Gaussian distribution;

[0031] Step 4.3: Calculate the diurnal spread rate R d Threshold R' d This makes the likelihood of high-intensity forest fires the greatest, assuming a potential forest fire hourly spread rate R. h With combustion time t h The fire exhibits a sinusoidal distribution; if the fire lasts for d days, the daily spread rate R... d satisfy:

[0032] (8);

[0033] Among them, t all The hours of forest fire spread, and meet the following conditions A represents the maximum hourly spread rate. If there is a possibility of a high-intensity forest fire, then A ≥ 1 km / h, i.e.:

[0034] (9);

[0035] When t all =24d, When t all When =24(d-1)+1, ;

[0036] Step 4.4: Using numerical integration methods and based on the estimated potential diurnal spread velocity probability density, calculate the potential spread velocity greater than R'. d The probability of a forest fire occurring at a speed of km / d is the probability of a high-intensity forest fire occurring.

[0037] (10);

[0038] Where h is the step size, y is the measured R d M is the number of numerical integration subintervals.

[0039] Furthermore, in step 2, features from high-intensity forest fire risk factors in the 16 days prior to the fire are extracted.

[0040] Furthermore, when the confidence interval is α, the values ​​of q1 and q2 in the formula (1) are 0.5-α / 2 and 0.5+α / 2, respectively; the weight coefficients w1 and w2 in the formula (1) are both 0.5.

[0041] Furthermore, the optimal confidence interval in formula (1) is 80%; the optimal hyperparameters are n_units=112, d_rate=0.15, and b_size=26.

[0042] Furthermore, the optimal quantile q0 is 0.56.

[0043] Furthermore, the optimal value of h in formula (6) is ,in The potential spread rate is predicted as ( The standard deviation of ).

[0044] Furthermore, R' in formula (9) d The value is km / d.

[0045] Furthermore, in the formula (10), h should be less than 0.1 km / d, and M should be 1000.

[0046] The beneficial effects of this invention are as follows: The method for assessing the probability of high-intensity forest fires described in this invention is simple to operate. It only requires input of information on combustible materials, meteorology, topography, and human activities, and can quickly assess the probability of high-intensity forest fires based on BiLSTM quantile regression and kernel density estimation methods. This method fully considers the key risk factors of high-intensity forest fires and incorporates the temporal variation characteristics of combustible materials and meteorological factors, making it more suitable for estimating the probability of high-intensity forest fires in complex forest areas. It is of great significance for high-intensity forest fire risk monitoring and early warning. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the overall method flow of the present invention.

[0048] Figure 2 This is a schematic diagram of the network structure of a potential wildfire spread rate prediction model based on BiLSTM.

[0049] Figure 3 This diagram illustrates the change in the loss function during model training in our specific implementation plan.

[0050] Figure 4 This is an evaluation of the accuracy of the prediction of the potential spread rate in the specific implementation plan of this company. Detailed Implementation

[0051] The following description, in conjunction with specific embodiments and accompanying drawings, further illustrates the high-intensity forest fire probability estimation method based on BiLSTM and kernel density estimation provided by the present invention:

[0052] A method for estimating the probability of high-intensity forest fires that combines BiLSTM and kernel density estimation, such as... Figure 1 As shown, it includes the following steps:

[0053] Step 1: Preparation of Risk Factors for High-Intensity Forest Fires

[0054] High-intensity forest fire risk factors include combustible factors (canopy and surface combustible moisture content, canopy combustible load, forest canopy height, and tree cover), meteorological factors (wind speed, temperature, rainfall, and relative humidity), topographic factors (elevation, slope, aspect, topographic humidity index, topographic location index, and plane curvature), and human activity factors (straight-line distance from roads and straight-line distance from residential areas). Canopy combustible moisture content and load were obtained from the MODIS MCD43A4 product using a lookup table algorithm. The forest canopy height factor was extracted from the global forest canopy height product published by Peter Potapov. The tree cover factor was extracted from the GFCC product. Surface dead combustible moisture content data came from the Canadian Forest Fire Hazard Rating System, meteorological data used the ERA5-Land dataset, topographic data used the GMTED2010 dataset, and road and residential area information came from the National Geographic Information Resource Catalog Service System. The temporal resolution for combustible material moisture content and load, as well as meteorological factors, is 1 day, and the spatial resolution is 500 meters. The spatial resolution for other factors is 500 meters.

[0055] Step 2: Construction of a Forest Fire Behavior Characteristic Case Database

[0056] Using the Global Fire Atlas dataset, we extracted the location, date, and spread rate of each forest fire (fire area ≥ 2 pixels, burning period ≥ 2 days) from 2003 to 2016. We also extracted the daily combustibles and meteorological factors, as well as topographic and human activity factors for the 16 days before the fire, to construct a case library of forest fire behavior characteristics.

[0057] Step 3: Construct a potential spread rate prediction model based on a bidirectional long short-term memory network (BiLSTM) and quantile regression (see...) Figure 2 The model was trained using a constructed forest fire behavior case library and optimized using the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) to obtain the optimal potential spread rate prediction model.

[0058] Specifically, it includes three parts:

[0059] Step 3.1: Build a two-layer BiLSTM neural network model. The two outputs of the model correspond to the lower bound L and the upper bound U of the predicted variables, respectively. Use formula (1) as the model loss function.

[0060] (1);

[0061] Where α is the confidence interval. and q1 and q2 are the loss functions for the upper and lower bounds of the predictor variables, respectively; q1 and q2 are the quantiles, with values ​​of 0.5-α / 2 and 0.5+α / 2, respectively; w1 and w2 are the weighting coefficients, both of which are 0.5.

[0062] (2);

[0063] Where N is the total number of samples, y i For the true value of the sample, This represents the predicted value of the sample when the quantile is q.

[0064] Step 3.2: Normalize the forest fire behavior feature dataset and divide it into training, validation, and test sets in a 6:3:1 ratio. Use average interval coverage (PICP) (Formula 3) and average interval width (PINAW) (Formula 4) as accuracy evaluation targets. Perform 200 epochs, selecting Adam as the model optimizer with an initial learning rate of 0.001. Stop training and export the model when the validation set accuracy no longer decreases. The loss function changes during training as follows: Figure 3 As shown;

[0065] (3);

[0066] (4);

[0067] in, and These are the lower and upper bounds for the prediction of the i-th sample, respectively;

[0068] Step 3.3: Set confidence intervals to 90%, 80%, and 70%, and use the Non-Dominated Sort Genetic Algorithm II (NSGA-II) with formulas (3) and (4) as optimization objectives. Use the comprehensive index CWC (formula 5) to determine the optimal BiLSTM model. When the confidence interval is 80%, the CWC is the lowest, that is, the model performance is the best. The hyperparameters of the BiLSTM layer are determined, namely n_units=112, d_rate=0.15, b_size=26.

[0069] (5);

[0070] in, As a penalty factor, it can be taken as 0.5;

[0071] Step 4: Based on the optimal prediction model, output the predicted value of the potential spread rate corresponding to each quantile q, estimate the probability density distribution of the potential spread rate using the kernel density estimation method, and evaluate the probability of high-intensity forest fires using numerical integration.

[0072] Specifically, it includes three parts:

[0073] Step 4.1: Based on the optimal prediction model, using formula (2) as the cost function, for each forest pixel, when the division points are successively set to 0.01, 0.02, ..., 0.09, 99 potential spread rate prediction values ​​can be obtained ( When the quantile is 0.56, the goodness of fit (R²) between the true and predicted values ​​is... 2 R is at its highest at this time. 2 =0.83, such as Figure 4 As shown; when the quantile is less than 0.18, the model predicts... All are the lowest observed values;

[0074] Step 4.2: Using the kernel density estimation method (Equation 6), select the Gaussian function as the kernel function, input the predicted spread velocity values ​​corresponding to different quantiles, and calculate the probability density of the potential diurnal spread velocity. ;

[0075] (6);

[0076] Where n is the total number of quantiles, i.e., n=99; h is the bandwidth, the optimal value of which is ( (where the variance of the n predicted values ​​is n). K represents the predicted potential spread rate corresponding to the i-th quantile. u It is a Gaussian function, i.e. ;

[0077] Step 4.3: For each forest pixel, using the numerical integration method (Equation 7) and based on the estimated potential diurnal spread velocity probability density, calculate the potential spread velocity greater than... The probability of a forest fire occurring at a rate of km / d is considered the probability of a high-intensity forest fire.

[0078] (7);

[0079] Where h is the step size, y is the measured R d h should be less than 0.1 km / d; M is the number of numerical integration subintervals, with a value of 1000;

[0080] Step 5: Mapping the probability of high-intensity forest fires.

Claims

1. A method for calculating the probability of high-intensity forest fires combining BiLSTM and kernel density estimation, comprising the following steps: Step 1: Determine the high-intensity forest fire risk factors for the preset area. The high-intensity forest fire risk factors include: combustible material factors, meteorological factors, and human activity factors. The combustible material factors include: canopy and surface combustible material moisture content, canopy combustible material load, forest canopy height, and tree coverage. The meteorological factors include: wind speed, temperature, rainfall, and relative humidity. The topographic factors include: elevation, slope, aspect, topographic humidity index, topographic location index, and plane curvature. The human activity factors include: straight-line distance from roads and straight-line distance from residential areas. Step 2: Using the Global Fire Atlas dataset, extract the location, date, and spread rate of ignition points for each forest fire from 2003 to 2016 that have a burning area of ​​≥2 pixels and a burning period of ≥2 days. Also, extract at least 3 features from the risk factors of high-intensity forest fires within a specific time period before the fire occurs to construct a forest fire behavior feature case library. Step 3: Construct a potential spread rate prediction model based on a bidirectional long short-term memory network (BiLSTM) model and quantile regression. Use the constructed wildfire behavior feature case library for training and learning, and use the non-dominated sorting genetic algorithm NSGA-II to optimize the hyperparameters of the neural network model to obtain the optimal potential spread rate prediction model. Specifically, it includes three parts: Step 3.1: Build a bidirectional long short-term memory network (BiLSTM) model. The two outputs of the BiLSTM model correspond to the lower and upper bounds of the predicted variables, respectively. Use formula (1) as the model loss function. （1）； Where α is the confidence interval. and q1 and q2 are the loss functions for the upper and lower bounds of the predictor variables, respectively, where q1 and q2 are quantiles and w1 and w2 are weight coefficients. （2）； Where N is the total number of samples, y i For the true value of the sample, This represents the predicted value of the sample when the quantile is q. Step 3.2: Normalize the forest fire behavior feature dataset and divide it into training, validation and test sets in a ratio of 6:3:

1. Use average interval coverage (PICP) and average interval width (PINAW) as the accuracy evaluation targets. Use Epoch and select Adam as the model optimizer. Export the model after training. （3）； （4）； in, and These are the lower and upper bounds for the prediction of the i-th sample, respectively; Step 3.3: Set confidence intervals to 90%, 80% and 70%, use the non-dominated sorting genetic algorithm NSGA-II, with formulas (3) and (4) as optimization objectives, use the comprehensive index CWC to determine the optimal BiLSTM model, and obtain the number of hidden neurons in the BiLSTM layer, random inactivation rate and batch size; （5）； in, As a penalty factor; Step 4: Based on the optimal prediction model, output the predicted value of the potential spread rate corresponding to each quantile q, estimate the probability density distribution of the potential spread rate using the kernel density estimation method, and evaluate the probability of high-intensity forest fires using numerical integration. Specifically, it includes four parts: Step 4.1: Based on the optimal prediction model, using formula (2) as the cost function, output the model prediction values ​​for the points q=0.01, 0.02, ..., 0.99 respectively, thus obtaining 99 potential spread rate prediction values. When the quantile is q0, the goodness of fit R between the true value and the predicted value is... 2 Highest; Step 4.2: Using the kernel density estimation method, a Gaussian function is selected as the kernel function. The predicted spread velocity values ​​corresponding to different quantiles are input to calculate the probability density of the potential diurnal spread velocity. ; （6）； Where n is the total number of quantiles, i.e., n=99; h is the bandwidth; K represents the predicted potential spread rate corresponding to the i-th quantile. u It is a Gaussian function, that is: （7）； in, Represents the independent variable that has a Gaussian distribution; Step 4.3: Calculate the diurnal spread rate R d Threshold R' d This makes the likelihood of high-intensity forest fires the greatest, assuming a potential forest fire hourly spread rate R. h With combustion time t h The fire exhibits a sinusoidal distribution; if the fire lasts for d days, the daily spread rate R... d satisfy: （8）； Among them, t all The hours of forest fire spread, and meet the following conditions A represents the maximum hourly spread rate. If there is a possibility of a high-intensity forest fire, then A ≥ 1 km / h, i.e.: （9）； When t all =24d, When t all When =24(d-1)+1, ; Step 4.4: Using numerical integration methods and based on the estimated potential diurnal spread velocity probability density, calculate the potential spread velocity greater than R'. d The probability of a forest fire occurring at a speed of km / d is the probability of a high-intensity forest fire occurring. （10）； Where h is the step size, y is the measured R d M is the number of numerical integration subintervals.

2. The method for calculating the probability of high-intensity forest fires combining BiLSTM and kernel density estimation as described in claim 1, characterized in that, Step 2 extracts features from the high-intensity forest fire risk factors in the 16 days prior to the fire.

3. The method for calculating the probability of high-intensity forest fires combining BiLSTM and kernel density estimation as described in claim 1, characterized in that, When the confidence interval is α, the values ​​of q1 and q2 in the formula (1) are 0.5-α / 2 and 0.5+α / 2, respectively; the weight coefficients w1 and w2 in the formula (1) are both 0.

5.

4. The method for calculating the probability of high-intensity forest fires combining BiLSTM and kernel density estimation as described in claim 1, characterized in that, The optimal quantile q0 is 0.

56.

5. The method for calculating the probability of high-intensity forest fires combining BiLSTM and kernel density estimation as described in claim 1, characterized in that, The optimal value of h in formula (6) is ,in The predicted potential spread rate ( The standard deviation of ).

6. The method for calculating the probability of high-intensity forest fires combining BiLSTM and kernel density estimation as described in claim 1, characterized in that, R' in formula (9) d The value is km / d.

7. The method for calculating the probability of high-intensity forest fires combining BiLSTM and kernel density estimation as described in claim 1, characterized in that, In the formula (10), h should be less than 0.1 km / d, and M should be 1000.

Images