Method for predicting occurrence probability of lightning fire by fusing extreme climate factors

Abstract

The invention provides a lightning fire occurrence probability prediction method fusing extreme climate factors, and relates to the technical field of forest fire prediction and risk assessment. The method comprises the following steps: collecting multi-source data such as thunder and lightning, extreme climate, weather, vegetation, soil and terrain, and constructing a balanced sample set; training an ignition probability model by adopting a random forest, XGBoost, a convolutional neural network and other models, and evaluating the performance difference before and after the extreme climate index is included; explaining variable contribution by using an SHAP method, and identifying a key driving factor; a risk grading grid map is generated according to the prediction probability, and the fire danger grade is automatically improved under the composite condition of heat waves and drought. The method has high precision and strong interpretability, and is suitable for fine prediction and prevention and control decision support of the occurrence probability of the lightning fire under the extreme climate background.

Description

Technical Field

[0001] This invention relates to the field of forest fire prediction and risk assessment technology, specifically a method for predicting the probability of lightning-induced fires by incorporating extreme climate factors. Background Technology

[0002] Lightning-induced fires are a type of natural fire caused by cloud-to-ground lightning directly igniting combustible materials, and are widely distributed in mid- to high-latitude forest ecosystems. Due to global warming, the frequency and intensity of extreme weather events such as extreme heat, drought, and heavy rainfall are continuously increasing. These extreme events prolong the fire season, reduce fuel moisture content, and significantly increase the incidence of lightning-induced fires.

[0003] Current forest fire risk predictions largely rely on conventional meteorological variables such as temperature, humidity, and wind speed, as well as fire risk indices. These models perform well under average weather conditions but struggle to capture the sudden impact of extreme events, potentially underestimating the risk of future lightning-induced fires. Although existing research has explored the impact of extreme weather on lightning-induced fires, current models often focus on specific regions or single models, lacking comprehensive data collection processes, sample balancing strategies, and interpretable analysis.

[0004] Therefore, there is an urgent need to build a systematic technical solution that integrates multi-source data acquisition, sample construction, model training, performance evaluation, and interpretive analysis to provide a comprehensive solution for lightning-fire prediction. Summary of the Invention

[0005] To address the technical problems of existing methods for predicting the probability of lightning-induced fires, such as their inability to fully utilize extreme climate information, limited accuracy, and lack of interpretability, this invention proposes a method for predicting the probability of lightning-induced fires that integrates extreme climate factors. This method collects multi-source data, including lightning strike events, extreme climate indices, meteorological data, vegetation data, soil data, and topographical data, constructs a balanced training sample, uses various machine learning models to predict the ignition probability of lightning-induced fires, and utilizes the SHAP method to explain the contribution of variables. Finally, it generates a lightning-induced fire risk classification map, providing a basis for fire prevention early warning and resource allocation.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Step 1: Collect data such as lightning strike points, fire point dates and locations, extreme climate indices, meteorological factors, vegetation indices, soil moisture and topographic factors by using lightning location data, global fire point data, ERA5-Land reanalysis data, remote sensing images and digital elevation models, etc. Divide the study area into grid units with uniform resolution, extract the environmental features of each grid, and resample and normalize the data from different sources.

[0008] Step 2: Based on the spatiotemporal matching of lightning strike events and fire points, the daily grid is marked as either "ignition" sample (value 1) or "non-ignition" sample (value 0) to form a binary response variable. Since the number of ignition samples is much smaller than the number of non-ignition samples, undersampling is used to randomly select non-ignition samples to make the number of positive and negative samples equal, thereby alleviating the class imbalance problem.

[0009] Step 3: Divide the grid features and response variables into training and test sets (preferably in a 7:3 ratio). Establish lightning strike fire probability models using logistic regression, random forest, XGBoost, and convolutional neural networks, respectively. Logistic regression is used for benchmark comparison; random forest uses random sampling and voting mechanisms to handle non-linear relationships; XGBoost improves accuracy by sequentializing decision trees and adding regularization terms; convolutional neural networks use convolutional and pooling layers to capture spatial correlations. The model output is the ignition probability of each grid cell on a given date.

[0010] Step 4: Calculate the accuracy, area under the ROC curve (AUC), and other metrics for each model, and compare the performance of models with and without the extreme climate index.

[0011] Step 5: Use the SHAP method to interpret the model output and evaluate the contribution of each variable to the occurrence of lightning-induced fires;

[0012] Step 6: Generate a lightning fire risk grid map based on the predicted ignition probability, and divide it into extremely high, high, medium, low and extremely low risk zones according to preset thresholds. When the duration of the heat wave and the daily temperature difference exceed the threshold and the precipitation index is low, the fire risk level will be automatically increased. This information will be used by forest fire prevention departments to formulate patrol plans and allocate resources.

[0013] Preferably, the extreme climate index includes 16 extreme temperature indicators and 10 extreme precipitation indicators. The extreme temperature indicators include: number of summer days, number of frost days, number of ice days, duration of cold weather, duration of warm weather, number of hot nights, length of growing season, maximum daily temperature, minimum daily temperature, maximum daily temperature, minimum daily temperature, percentage of warm days, percentage of cold days, percentage of warm nights, percentage of cold nights, and daily temperature difference. The extreme precipitation indicators include: number of days with heavy precipitation, number of days with extreme heavy precipitation, number of days with daily precipitation ≥10mm, number of days with daily precipitation ≥20mm, number of consecutive drought days, number of consecutive wet days, total precipitation, daily precipitation intensity index, maximum daily precipitation, and maximum consecutive 5-day precipitation. The lightning strike data can be obtained from the Global Lightning Location Network, and the fire point data can be obtained from the Global Fire Database; the vegetation index adopts the MODIS Normalized Difference Vegetation Index product; soil moisture and high and low vegetation cover are obtained from ERA5; the terrain factors are extracted from the digital elevation model; all data are resampled to 1 km resolution in ArcGIS software and standardized to form the input feature matrix.

[0014] Preferably, the positive and negative sample screening strategy means that when matching lightning strike events and fire point locations, if the distance between the lightning strike and the center of the fire point is within 10km and the ignition time difference is no more than 7 days, it is considered a lightning strike ignition sample; an equal number of non-ignition samples are randomly selected from areas with active lightning strikes but no fire point occurrence as controls to form a balanced sample set.

[0015] Preferably, the model parameter settings are as follows: the random forest model has 500 trees and a minimum number of samples per node; the XGBoost model has a learning rate of 0.05, a maximum depth of 6, regularization parameters L1 and L2 set to 0.3 and 0.5 respectively, and an early stopping strategy is used to prevent overfitting; the convolutional neural network takes a 3×3 grid neighborhood feature tensor as input, and the network contains two convolutional layers (32 and 64 convolutional kernels), one max pooling layer and two fully connected layers, with a Dropout ratio of 0.3; the logistic regression model uses L2 regularization.

[0016] Preferably, the model evaluation not only calculates accuracy and area under the ROC curve (AUC), but also precision, recall and F1 score, and analyzes the model's classification performance on ignition and non-ignition samples through confusion matrix analysis; k-fold cross-validation is used during evaluation to reduce partitioning bias, the differences between models with and without extreme climate indices are compared on each index, and statistical tests are used to confirm the significance of performance improvement.

[0017] Preferably, the SHAP method interprets the model output by grouping features according to extreme climate factors, meteorological factors, vegetation factors, soil factors, and topographic factors, calculating the contribution of each group and each individual indicator, and presenting the direction and magnitude of the variable's influence through SHAP bee scatter plots and bar charts. To ensure stability, the SHAP value can be calculated 100 times, and the average contribution is taken as the final ranking.

[0018] Preferably, the lightning fire risk grid map is divided into extremely high, high, medium, low and extremely low risk zones based on the ignition probability output by the model, and the risk map is output in GeoTIFF format; when the predicted number of days of heat wave and daily temperature difference exceed the historical high, and the extreme precipitation index is at a low value, the system automatically increases the fire risk level of the current area and triggers the advance deployment of patrol personnel and equipment.

[0019] This invention provides a method for predicting the probability of lightning-induced fires by incorporating extreme climate factors, which has the following beneficial technical effects:

[0020] 1. In this invention, the extreme climate index is incorporated into the lightning fire prediction framework, which significantly improves the model's ability to identify ignition probability, and at the same time, it can quantitatively assess the contribution of extreme temperature and precipitation indices to the occurrence of lightning fires.

[0021] 2. In this invention, the lightning fire prediction model takes into account both performance and interpretability. It reveals the contribution of each environmental factor to the prediction results through the SHAP method, identifies the positive driving role of key indicators such as the minimum daily maximum temperature, daily temperature range, and duration of warm weather, clarifies the inhibitory effect of heavy precipitation indicators such as maximum daily precipitation and maximum consecutive 5-day precipitation, and analyzes the combined effects of heat waves and drought.

[0022] 3. In this invention, the lightning fire prediction model has wide applicability and is easy to operate. It proposes a complete process for data collection, sample construction, model training and prediction. The input data sources are wide and the model parameters are easy to adjust. The risk classification map can be directly used in forest fire prevention and management systems and has good prospects for promotion and application. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the process provided by the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] like Figure 1 As shown, this embodiment of the invention provides a method for predicting the probability of lightning-induced fires by incorporating extreme climate factors, including:

[0026] This invention employs multi-source data acquisition and gridded processing. The study area is divided into grid cells of uniform resolution, with each grid cell serving as a sample unit. The time, location, and duration of lightning strikes are obtained through the Global Lightning Location Network, and fire point vector boundary information is acquired from the Global Fire Database. Combined with ERA5-Land reanalysis data, conventional meteorological factors such as temperature, relative humidity, precipitation, and wind speed, as well as 16 extreme temperature indices and 10 extreme precipitation indices, are extracted. Simultaneously, vegetation indices are calculated from MODIS NDVI products, and high and low vegetation cover and soil moisture indices are extracted from ERA5. Topographic factors such as slope, aspect, and altitude are extracted from the Digital Elevation Model. All data undergoes projection transformation, resampling, and normalization before being associated with each grid cell to form a feature vector.

[0027] For sample construction and balancing, this invention is based on the spatiotemporal matching relationship between lightning strike events and fire point data. If a lightning strike occurs in a certain grid on a certain day and a fire point appears within a specified range, the grid is marked as 1; otherwise, it is marked as 0. Since there are far more non-ignition samples than ignition samples, an undersampling method is used to randomly remove some non-ignition samples to keep the number of positive and negative samples consistent. The samples are randomly divided into training set and test set, with the preferred ratio being 7:3, to ensure the balance between model training and validation.

[0028] Model training and prediction: Logistic regression, random forest, XGBoost, and convolutional neural network models were trained on the constructed training set respectively. The logistic regression model was used for benchmark comparison, assuming a linear relationship between the response variable and the independent variable. The random forest model handled non-linear features by integrating multiple decision trees. The XGBoost model used a gradient boosting strategy to iteratively train the tree model and introduced a regularization term to control complexity. The convolutional neural network model treated the multidimensional features of the grid neighborhood as input tensors, captured spatial correlations through convolutional and pooling layers, and then output the ignition probability through a fully connected layer. Cross-validation and early stopping strategies were used during training to prevent overfitting.

[0029] For model evaluation and comparison, this invention calculates the accuracy, AUC, precision, recall, and other indicators of each model on the test set, and compares the performance differences between models with and without the extreme climate index.

[0030] This invention explores variable contribution analysis and interaction effects by applying the SHAP method to the best-performing model, drawing SHAP value distribution plots and bar charts, and analyzing the positive and negative impacts of each variable on the prediction results.

[0031] Risk grading and visualization output: This invention divides the risk level into extremely high, high, medium, low and extremely low risk levels based on the predicted ignition probability by setting a threshold, and generates a risk raster map; land use and road network information are superimposed on the risk map to identify key protection areas; when the predicted heat wave duration index and daily temperature difference exceed the threshold and the precipitation index is low, the system automatically upgrades the fire risk level.

[0032] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0033] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A lightning-caused fire occurrence probability prediction method that fuses extreme climate factors, characterized by, Comprise: Step 1: Collect lightning strike points, fire point dates and locations, extreme climate indices, meteorological factors, vegetation indices, soil moisture, and terrain factors, etc. by lightning location data, global fire point data, ERA5-Land reanalysis data, remote sensing images, and digital elevation models, etc. Divide the study area into grid cells of uniform resolution, extract the environmental characteristics of each grid, and resample and normalize the data from different sources; Step 2: According to the spatio-temporal matching of lightning strike events and fire points, mark each daily grid as "ignition" sample (value 1) or "non-ignition" sample (value 0) to form a binary response variable. Since the number of ignition samples is much smaller than that of non-ignition samples, undersampling is used to randomly select non-ignition samples to make the number of positive and negative samples equal, thereby alleviating the class imbalance problem; Step 3: Divide the grid features and response variables into training set and test set (ratio preferably 7:3); Use logistic regression, random forest, XGBoost, and convolutional neural network models to establish lightning-caused fire occurrence probability models, where logistic regression is used for benchmark comparison; Random forest uses random sampling and voting mechanism to handle nonlinear relationships; XGBoost improves accuracy by serializing decision trees and adding regularization terms; Convolutional neural network uses convolutional and pooling layers to capture spatial correlation, and the model output is the ignition probability of each grid on a given date; Step 4: Calculate the accuracy, ROC curve area (AUC), and other indicators of each model, and compare the model performance with and without extreme climate indices; Step 5: Use SHAP method to explain the model output and evaluate the contribution of each variable to lightning-caused fire occurrence; Step 6: Generate lightning-caused fire risk grid map according to the predicted ignition probability, divide it into extremely high, high, medium, low, and extremely low risk areas according to the preset threshold, and automatically upgrade the fire risk level when the duration of heat wave and daily temperature difference exceed the threshold and the precipitation index is low, which can be used by forest fire prevention departments to develop patrol plans and resource allocation.

2. The method of claim 1, wherein: The extreme climate index includes 16 extreme temperature indicators and 10 extreme precipitation indicators, wherein the extreme temperature indicators include summer day number, frost day number, ice day number, cold duration, warm heat duration, hot night day number, growing season length, maximum value of daily maximum temperature, minimum value of daily maximum temperature, maximum value of daily minimum temperature, minimum value of daily minimum temperature, warm day percentage, cold day percentage, warm night percentage, cold night percentage and daily temperature difference; the extreme precipitation indicators include strong precipitation day number, extreme strong precipitation day number, day number of daily precipitation greater than or equal to 10 mm, day number of daily precipitation greater than or equal to 20 mm, continuous drought day number, continuous wet day number, total precipitation, daily precipitation intensity index, maximum one-day precipitation and maximum continuous 5-day precipitation; the lightning data can come from the global lightning location network, the fire point data come from the global fire database; the vegetation index adopts the MODIS normalized vegetation index product; the soil moisture and high and low vegetation coverage come from ERA5; the terrain factor is extracted from the digital elevation model; all data are resampled to 1 km resolution in the ArcGIS software, and the standardized processing is used to form an input feature matrix.

3. The method of claim 1, wherein: The positive and negative sample screening strategy refers to, when matching the lightning event and the fire point position, if the lightning distance from the fire point center is within 10 km and the ignition time difference is not more than 7 days, the lightning ignition sample is considered; the same amount of non-ignition samples are randomly extracted in the lightning active area without fire points as controls to form a balanced sample set.

4. The method of claim 1, wherein: The model parameter setting refers to that the number of random forest model trees is preferably 500, and the minimum sample number of nodes is 2; the learning rate of the XGBoost model is 0.05, the maximum depth is 6, the regularization parameter L1 is set to 0.3, L2 is set to 0.5, and the early stopping strategy is used to prevent overfitting; the convolutional neural network input is a feature tensor of a 3*3 grid neighborhood, the network includes two convolutional layers (32 and 64 convolutional kernels), one maximum pooling layer and two fully connected layers, and the Dropout ratio is 0.3; the logistic regression model adopts L2 regularization.

5. The method of claim 1, wherein: The lightning fire risk grid map is divided into extremely high, high, medium, low and extremely low risk areas according to the ignition probability output by the model, and the risk map is output in the GeoTIFF format; when the predicted heat wave duration and daily temperature difference exceed the historical high value, and the extreme precipitation index is at a low value, the system automatically raises the fire danger grade of the current area, triggering the advance deployment of patrol personnel and equipment.

Images