A method for constructing a forest fire early warning model based on time-decay precipitation algorithm

Abstract

The present invention discloses a method for constructing a forest fire early warning model based on a time decay precipitation algorithm, comprising: obtaining precipitation data and calculating a comprehensive precipitation value based on a time decay algorithm; obtaining historical fire data and converting forest fire points into forest fire point density based on a two-dimensional Gaussian convolution function; obtaining meteorological data, remote sensing data and elevation data and preprocessing the data to obtain preprocessed data; using the comprehensive precipitation value, forest fire point density and preprocessed data as training sets, training an SVM regression model to obtain a trained SVM regression model; predicting forest fire density data based on the trained SVM regression model; dividing fire levels according to forest fire density data, and combining the trained SVM regression model to obtain a forest fire early warning model. By using the present invention, the accuracy of forest fire early warning can be improved. The present invention can be widely used in the field of fire risk prediction.

Description

Technical Field

[0001] The present invention relates to the field of fire risk prediction, and in particular to a method for constructing a forest fire early warning model based on a time-attenuated precipitation algorithm. Background Art

[0002] In most of the forest fire warning models constructed by previous researchers, precipitation was selected as an important characteristic factor and ranked among the top of many characteristic factors, indicating that precipitation values ​​play an important role in forest fire warning. However, in most of the previous models, precipitation is calculated as an equal-weighted average of monthly average precipitation or regional average precipitation. Using average precipitation as the precipitation value does not take into account the attenuation effect of precipitation at different time nodes, but only takes into account the average weight. Using this method to calculate precipitation values ​​cannot well express the impact of precipitation values ​​on forest fires. Summary of the invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide a method for constructing a forest fire warning model based on a time-decaying precipitation algorithm, which uses a time-decaying algorithm to calculate the comprehensive precipitation value to better express the impact of precipitation values ​​on forest fires and improve the accuracy of forest fire warnings.

[0004] The first technical solution adopted by the present invention is: a method for constructing a forest fire early warning model based on a time-attenuated precipitation algorithm, comprising the following steps:

[0005] S1, obtaining precipitation data and calculating the comprehensive precipitation value based on the time decay algorithm;

[0006] S2, obtaining historical fire data and converting forest fire points into forest fire point density based on a two-dimensional Gaussian convolution function;

[0007] S3, obtaining meteorological data, remote sensing data and elevation data and preprocessing the data to obtain preprocessed data;

[0008] S4, using the comprehensive precipitation value, forest fire point density and pre-processed data as a training set, training the SVM regression model to obtain a trained SVM regression model;

[0009] S5. Predict forest fire density data based on the trained SVM regression model;

[0010] S6. Divide the fire levels according to the forest fire density data, and obtain the forest fire early warning model by combining the trained SVM regression model.

[0011] Furthermore, the calculation formula of the comprehensive precipitation value is as follows:

[0012]

[0013] In the above formula, D Precipitation represents the comprehensive precipitation value, A represents the undetermined constant, n represents the previous n days, t represents the time, and D t Indicates the precipitation value t days ago.

[0014] Further, the formula of the two-dimensional Gaussian convolution function is expressed as follows:

[0015]

[0016] In the above formula, σ represents the convolution range, x represents the horizontal coordinate difference from the center point, and y represents the vertical coordinate difference from the center point.

[0017] Furthermore, the preprocessed data includes relative humidity of the day, maximum wind speed of the day, maximum temperature of the day, normalized difference vegetation index, vegetation water supply index and altitude.

[0018] Furthermore, the step of obtaining meteorological data, remote sensing data and elevation data and preprocessing the data to obtain preprocessed data specifically includes:

[0019] S31. Obtain meteorological data, remote sensing data and elevation data of forest areas;

[0020] S32, interpolating the meteorological data based on the Kriging interpolation method to obtain uniform grid meteorological data in the forest area;

[0021] S33, calculating the normalized difference vegetation index and the vegetation water supply index based on remote sensing data;

[0022] S34. Obtaining altitude data according to the elevation data.

[0023] Furthermore, the step of using the comprehensive precipitation value, forest fire point density and pre-processed data as a training set to train the SVM regression model to obtain a trained SVM regression model specifically includes:

[0024] S41, taking the comprehensive precipitation value, uniform grid meteorological data in the forest area, normalized vegetation index, vegetation water supply index and altitude data as independent variables, and forest fire point density as dependent variable, and training the SVM regression model in combination with the objective function to obtain a trained SVM regression model;

[0025] S42. The SVM regression model normalizes the independent variables and uses a Gaussian kernel.

[0026] Further, the formula of the objective function is expressed as follows:

[0027]

[0028]

[0029] ξ 1i ≥0,ξ 2i ≥0(i=1,2,…m)

[0030] In the above formula, ω represents the weight vector of the output space, represents the nonlinear mapping from input space to output space, ξ 1i ,ξ 2i represents the slack variable, b represents the offset value, ε represents the insensitive range of the loss value, m represents the dimension, and T represents the transpose of the matrix.

[0031] Furthermore, the step of dividing the fire levels according to the forest fire density data and obtaining the forest fire early warning model by combining the trained SVM regression model specifically includes:

[0032] S61. Based on the natural breakpoint method, the forest fire density data is divided into five different levels of forest fire risk areas;

[0033] S62, determining that the forest fire risk level in the area is higher than a preset level, and issuing an early warning;

[0034] S63. A forest fire warning model is obtained by combining the trained SVM regression model.

[0035] The beneficial effect of the method of the present invention is as follows: the present invention calculates the comprehensive precipitation index through the time decay algorithm, and then combines other characteristic factors as the independent variable of the forest fire warning model, and at the same time convolves the discrete forest fire points into forest density as the dependent variable of the model. After data preprocessing, combined with the SVM regression model, a high-accuracy forest fire warning model is constructed. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The present invention is a flowchart of the steps of a method for constructing a forest fire warning model based on a time-attenuated precipitation algorithm. DETAILED DESCRIPTION

[0037] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only provided for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0038] Reference Figure 1 The present invention provides a method for constructing a forest fire warning model based on a time-attenuated precipitation algorithm, the method comprising the following steps:

[0039] S1, obtaining precipitation data and calculating the comprehensive precipitation value based on the time decay algorithm;

[0040] Specifically, in the time decay model, the change amplitude of the value is proportional to its time, which is consistent with the impact of precipitation on different days on the occurrence of forest fires. In the time decay model, the above mapping relationship can be converted into a mathematical formula and expressed as:

[0041]

[0042] In the above formula, N represents a value, t represents time, and γ represents a decay constant.

[0043] Applied to this patent, the solution of the time decay equation can be expressed as:

[0044] N(t)=D t e -γt (2)

[0045] In the above formula, N(t) is a quantity related to time t, which represents the impact value of precipitation t days ago, and D t Indicates the precipitation value t days ago.

[0046] Next, the precipitation impact values ​​of the previous n days are added together to obtain the comprehensive precipitation value. The calculation formula for the comprehensive precipitation value is:

[0047]

[0048] In the above formula, D Prcipitation represents the comprehensive precipitation value, A represents the undetermined constant, t represents the time, D t represents the precipitation value t days ago. Now we need to determine the value of the unknown constant A and select the precipitation value of the previous n days. Mutual information can represent the amount of information contained in a random variable by another random variable. It can be used to measure the correlation between them and is not limited to linear relationships. The calculation formula for mutual information is:

[0049]

[0050] Where p(x, y) is the joint probability distribution function of X and Y, and p(x) and p(y) are the marginal probability distribution functions of X and Y, respectively. According to the comprehensive precipitation data obtained by different combinations of the undetermined constant term A value and the precipitation value of the previous n days, the constant term A value and the precipitation value of the previous n days with the largest mutual information value between the comprehensive precipitation data and the forest fire point density are selected. After iterative calculation, when A is 0.1657 and n is 6, the calculated comprehensive precipitation value can best represent the influence of precipitation value on fire prediction in the present invention.

[0051] S2, obtaining historical fire data and converting forest fire points into forest fire point density based on a two-dimensional Gaussian convolution function;

[0052] Specifically, the historical fire data used in the present invention comes from the global forest fire historical data website provided by NASA. The data provided by the website is low-latency fire activity data recorded from satellite images provided by NASA and the Forest Service of the United States Department of Agriculture. The data form of forest fire points is a scatter plot within the forest area. In the present invention, discrete forest fire points are converted into continuous forest fire point density through two-dimensional Gaussian convolution, and the classification problem is converted into a regression problem. Each discrete forest fire point represents the situation of forest fires in the surrounding area. After two-dimensional Gaussian convolution, the higher the forest fire point density value, the more serious the surrounding forest fires.

[0053] The expression of two-dimensional Gaussian convolution is as follows:

[0054]

[0055] In the above formula, σ represents the convolution range. In the present invention, ten kilometers is used as the convolution radius.

[0056] S3, obtaining meteorological data, remote sensing data and elevation data and preprocessing the data to obtain preprocessed data;

[0057] S3.1. Obtain meteorological data, remote sensing data and elevation data of forest areas;

[0058] Specifically, the present invention also selects the relative humidity of the forest area on the day, the maximum wind speed on the day, the maximum temperature on the day, the normalized difference vegetation index (NDVI), the vegetation water supply index (VSWI), and the altitude as characteristic factors. The relative humidity of the day, the maximum wind speed on the day, and the maximum temperature on the day are obtained through data provided by the local meteorological bureau.

[0059] S3.2. Interpolate the meteorological data based on the Kriging interpolation method to obtain continuous meteorological data in the forest area.

[0060] Specifically, the above-mentioned meteorological data are discrete and discontinuous in the forest area; and the present invention needs to obtain continuous meteorological data in the forest area, so it is necessary to perform linear interpolation on the data so that each data point in the forest area has its meteorological data. Therefore, the present invention uses the Kriging interpolation algorithm to interpolate the meteorological data.

[0061] S3.3. Calculate the normalized difference vegetation index and vegetation water supply index based on remote sensing data;

[0062] The vegetation coverage of the forest is represented by the NDVI model data. NDVI is the normalized vegetation index, which reflects the growth status and vegetation coverage of the vegetation. NDVI is calculated using the reflectance values ​​of the near-infrared band and the red light band. In this paper, NDVI is calculated using the reflectance data of band 1 and band 2 in the MOD021KM remote sensing data. The formula is:

[0063]

[0064] The NDVI value is between -1,1. When it is a negative value, it means that the ground is covered by clouds, water, snow, etc., and has high reflectivity to visible light. When it is a positive value, it means that there is vegetation coverage, and it increases with the coverage.

[0065] The VSWI index is also called the vegetation water supply index, which can reflect the growth status of crops and the moisture content of vegetation to a certain extent. Its formula is:

[0066]

[0067] NDVI is the Normalized Difference Vegetation Index, Ts represents the surface temperature, and Ts data is obtained from MOD11A2 remote sensing data, which records the comprehensive surface temperature values ​​for eight days. The larger the VSWI value, the higher the moisture content of the vegetation.

[0068] S3.4. Obtain altitude data based on elevation data.

[0069] The altitude data is obtained from the SPTMDEM 90M elevation data.

[0070] S4, using the comprehensive precipitation value, forest fire point density and pre-processed meteorological data as a training set, training the SVM regression model to obtain a trained SVM regression model;

[0071] Specifically, the comprehensive precipitation value and the preprocessed meteorological data are used as independent variables, and the forest fire point density is used as the dependent variable. Then the SVM kernel function selects the Gaussian kernel to normalize the data in the independent variables. In addition, the Gram matrix is ​​used in the SVM model training. The meaning of the Gram matrix is ​​that when the distance between two samples is close, the value is 1, and when the distance is far, the value is 0. The kernel scale is a scaling parameter used to evaluate the scaling size of the appropriate Gram matrix. After screening and calculation, the kernel scale is specified to be 3, and the training set data is used for training in the SVM regression model. Then the trained SVM regression model is used in the test set to obtain the forest fire density data in the prediction set. The objective function expression of the SVM regression model is as follows:

[0072]

[0073] S5. Predict forest fire density data based on the trained SVM regression model;

[0074] S6. Divide the fire levels according to the forest fire density data, and obtain the forest fire early warning model by combining the trained SVM regression model.

[0075] S6.1. Divide forest fire density data into multiple levels of forest fire risk areas based on the natural breakpoint method;

[0076] After obtaining the forest fire density data in the prediction set, the natural breakpoint method is used to divide the forest area into five fire warning levels. The higher the level, the greater the risk of fire. The natural breakpoint method calculates the variance of each category, and then calculates the sum of these variances. The size of the sum of variances is used to compare the quality of classification. The smallest value is the best classification result. The calculation method is as follows:

[0077] GVF=(SDAM-SCDM) / SDAM (9)

[0078] Where GVF is the goodness of fit of the variance and SDAM is the sum of squared deviations from the mean, which are calculated as follows:

[0079]

[0080] in is the mean, x i is the forest fire density. SCDM is the sum of squared deviations of the category mean. It is calculated as follows:

[0081]

[0082] Using the natural breakpoint method, the present invention divides the forest fire risk areas into five different levels according to the following numerical ranges, and divides the forest fire point density of (-∞, 0) into level one, [0, 0.08) into level two, [0.08, 0.27) into level three, [0.27, 0.45) into level four, and [0.45, ∞) into level five.

[0083] S6.2. If the forest fire risk level in the area is judged to be higher than the preset level, an early warning is issued, where the preset level is level 3;

[0084] S6.3. Combine the trained SVM regression model to obtain a forest fire warning model.

[0085] At this point, the forest fire warning model has been built. The forest areas with fire risk level three or above are used to predict the occurrence of forest fires. The prediction accuracy rates of the forest areas in the test set are 90.13%, 93.04% and 87.5% respectively. If the average precipitation value is used to represent the impact of precipitation on forest fires, and other factors remain unchanged, the prediction accuracy rates of the forest areas in the test set are 85.17%, 89.50% and 84.90% respectively. By comparison, it can be seen that the comprehensive precipitation value calculated by the time decay algorithm as a characteristic factor can improve the accuracy of the forest fire warning model by about 5%. This shows that the model results can better predict the risk of forest fires.

[0086] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for constructing a forest fire warning model based on a time-decay precipitation algorithm, characterized in that: The following steps are involved: S1, obtaining precipitation data and calculating the comprehensive precipitation value based on the time decay algorithm; S2, obtaining historical fire data and converting forest fire points into forest fire point density based on a two-dimensional Gaussian convolution function; S3, obtaining meteorological data, remote sensing data and elevation data and preprocessing the data to obtain preprocessed data; S4, using the comprehensive precipitation value, forest fire point density and pre-processed data as a training set, training the SVM regression model to obtain a trained SVM regression model; S5. Predict forest fire density data based on the trained SVM regression model; S6. Divide the fire levels according to the forest fire density data, and obtain the forest fire early warning model by combining the trained SVM regression model; The calculation formula of the comprehensive precipitation value is as follows: In the above formula, D Precipitation represents the comprehensive precipitation value, A represents the undetermined constant, n represents the previous n days, t represents the time, and D t Indicates the precipitation value t days ago.

2. The method for constructing a forest fire early warning model based on a time-attenuated precipitation algorithm according to claim 1, characterized in that: The formula of the two-dimensional Gaussian convolution function is as follows: In the above formula, σ represents the convolution range, x represents the horizontal coordinate difference from the center point, and y represents the vertical coordinate difference from the center point.

3. The method for constructing a forest fire early warning model based on a time-attenuated precipitation algorithm according to claim 2 is characterized in that: The pre-processed data include the relative humidity of the day, the maximum wind speed of the day, the maximum temperature of the day, the normalized difference vegetation index, the vegetation water supply index and the altitude.

4. The method for constructing a forest fire early warning model based on a time-attenuated precipitation algorithm according to claim 3 is characterized in that: The step of obtaining meteorological data, remote sensing data and elevation data and preprocessing the data to obtain preprocessed data specifically includes: S31. Obtain meteorological data, remote sensing data and elevation data of forest areas; S32, interpolating the meteorological data based on the Kriging interpolation method to obtain uniform grid meteorological data in the forest area; S33, calculating the normalized difference vegetation index and the vegetation water supply index based on remote sensing data; S34. Obtaining altitude data according to the elevation data.

5. The method for constructing a forest fire early warning model based on a time-attenuated precipitation algorithm according to claim 4 is characterized in that: The step of using the comprehensive precipitation value, forest fire point density and pre-processed data as a training set to train the SVM regression model to obtain a trained SVM regression model specifically includes: S41, taking the comprehensive precipitation value, uniform grid meteorological data in the forest area, normalized vegetation index, vegetation water supply index and altitude data as independent variables, and forest fire point density as dependent variable, and training the SVM regression model in combination with the objective function to obtain a trained SVM regression model; S42. The SVM regression model normalizes the independent variables and uses a Gaussian kernel.

6. The method for constructing a forest fire early warning model based on a time-attenuated precipitation algorithm according to claim 5, characterized in that: The objective function of the SVM regression model is expressed as follows: x 1i ≥0.ξ 2i ≥0(i=1,2,…m) In the above formula, ω represents the weight vector of the output space, represents the nonlinear mapping from input space to output space, ξ 1i ,ξ 2i represents the slack variable, b represents the offset value, ε represents the insensitive range of the loss value, m represents the dimension, and T represents the transpose of the matrix.

7. The method for constructing a forest fire warning model based on a time-attenuated precipitation algorithm according to claim 6, characterized in that: The step of dividing the fire levels according to the forest fire density data and obtaining the forest fire early warning model by combining the trained SVM regression model specifically includes: S61. Based on the natural breakpoint method, the forest fire density data is divided into five different levels of forest fire risk areas; S62, determining that the forest fire risk level in the area is higher than a preset level, and issuing an early warning; S63. A forest fire warning model is obtained by combining the trained SVM regression model.

Images