Forest fire risk prediction method and system based on machine learning

Abstract

The invention discloses a forest fire risk prediction method and system based on machine learning, and the method comprises the steps: taking data of a monitoring data source of a preset forest in a specified time period as to-be-analyzed data; extracting space-time distribution characteristics of the to-be-analyzed data and the reference data source, performing area division on the space-time distribution characteristics according to similarity, taking an area where the to-be-analyzed data with the similarity higher than a similarity threshold value is located as a candidate area, and otherwise, taking the area as a fuzzy area; performing autocorrelation analysis on the fuzzy region based on the undetermined data source, adding the fuzzy region of which the autocorrelation is higher than an autocorrelation threshold to the candidate region, and performing time-varying influence analysis on the environmental data and the human activity data to obtain a fire driving coefficient; and constructing a forest fire risk prediction model according to the fire driving coefficient and the candidate region, optimizing the forest fire risk prediction model according to a fire spread rule, and outputting a prediction result.

Description

Technical Field

[0001] The present invention relates to the field of risk prediction, and in particular to a forest fire risk prediction method and system based on machine learning. Background Art

[0002] As a global natural disaster, forest fires are characterized by suddenness, destructiveness, and rapid spread, posing a serious threat to the ecological environment, social economy, and human security. In recent years, the frequency and severity of forest fires have increased significantly due to factors such as climate change, vegetation drying, and intensified human activities. There is an urgent need to develop efficient and accurate risk prediction technologies to support early warning and prevention and control decisions.

[0003] Traditional forest fire risk prediction methods mostly rely on a single data source and use statistical models or empirical indices for evaluation. However, such methods have significant limitations: first, they fail to fully integrate the spatiotemporal correlation characteristics of multi-dimensional data, resulting in an incomplete characterization of fire driving factors; second, they ignore the spatiotemporal heterogeneity of fire risk in a large area, making it difficult to distinguish the dynamic differences between high-risk and low-risk sub-areas; third, the time-varying impact of human activities and the physical laws of fire spread have not been effectively modeled, limiting the adaptability and accuracy of the prediction model.

[0004] With the development of remote sensing technology, Internet of Things monitoring and machine learning, multi-source data fusion and spatiotemporal feature mining have provided new ideas for fire prediction. However, existing research still faces challenges: the heterogeneity of multi-source data makes feature alignment difficult; the reliability differences of different monitoring data sources may introduce noise; at the same time, the nonlinear relationship between fire risk and driving factors is difficult to analyze through traditional models under complex terrain and vegetation conditions. In addition, the existing methods deal with "fuzzy areas" in a rough manner, and do not fully utilize the spatial autocorrelation between data for risk correction, which can easily cause prediction blind spots.

[0005] Therefore, it is necessary to propose a new forest fire risk prediction method based on machine learning. By integrating multi-source monitoring data, combining spatiotemporal distribution feature analysis, autocorrelation correction and time-varying driven modeling, a dynamic risk prediction model can be constructed to improve the spatial rationality and temporal sensitivity of the prediction results, providing technical support for the refined risk management of forest fires. Summary of the invention

[0006] The purpose of the present invention is to provide a forest fire risk prediction method based on machine learning.

[0007] To achieve the above object, the present invention is implemented according to the following technical solutions: The present invention comprises the following steps: The data of the monitoring data source of the preset forest within the specified time period is used as the data to be analyzed; the monitoring data source includes a reference data source and a pending data source; the fire risk law of the reference data source is higher than that of the pending data source; the data includes environmental data, image data, historical fire data, human activity data, and remote sensing data; the environmental data includes meteorological data, vegetation data, and terrain data; Extracting the spatiotemporal distribution features of the data to be analyzed and the reference data source, dividing the spatiotemporal distribution features into regions according to similarity, and defining the region where the data to be analyzed is located with a similarity higher than a similarity threshold as a candidate region, and vice versa as a fuzzy region; Performing autocorrelation analysis on the fuzzy area based on the pending data source, adding the fuzzy area with autocorrelation higher than the autocorrelation threshold to the candidate area, and performing time-varying impact analysis on the environmental data and the human activity data to obtain a fire driving coefficient; A forest fire risk prediction model is constructed according to the fire driving coefficient and the candidate area, the forest fire risk prediction model is optimized according to the fire spread law, the data to be predicted is input into the forest fire risk prediction model, and the prediction result is output.

[0008] Furthermore, the method for dividing the spatiotemporal distribution features into regions according to similarity includes: Calculate the similarity between the spatiotemporal distribution features in the data to be analyzed and the spatiotemporal distribution features in the reference data source, and mark the spatiotemporal distribution features in the data to be analyzed with a similarity higher than 0.473 in the latitude and longitude map of the preset area; The spatiotemporal distribution characteristics are divided into primary regions according to the clustering density of the pre-annotated latitude and longitude map of the preset area to obtain the primary region; the fire probability of the spatiotemporal distribution characteristics is calculated: The ignition probability of the cth spatiotemporal distribution feature is , the wind intensity in the ath direction is , the fire temperature of the neighbor in the ath direction is , the ignition point of the cth spatiotemporal distribution feature is , the wind direction vector in the ath direction is , the neighbor coordinate vector in the ath direction is , the number of directions is ; The density clustering algorithm is used to perform primary regional adjustment on the spatiotemporal distribution features with a fire probability higher than 0.316 to obtain the adjusted area. The adjusted area containing the spatiotemporal distribution features with a fire probability higher than 0.316 is taken as the first area, otherwise it is taken as the fuzzy area.

[0009] Furthermore, the method of performing autocorrelation analysis on the fuzzy area based on the pending data source includes: Extract spatial distribution features of the pending data source to obtain a pending spatial feature set, take the spatial distribution features of the reference data source as reference features; take the spatial distribution features of the fuzzy area as fuzzy features; The different spatial distribution characteristics of the to-be-determined spatial feature set and the reference features are taken as the to-be-determined spatial features, and the autocorrelation between the to-be-determined spatial features and the fuzzy features is calculated.

[0010] Furthermore, the method of performing time-varying impact analysis on the environmental data and the human activity data to obtain a fire driving coefficient includes: The long short-term memory network is used to analyze the meteorological data, vegetation data, and terrain data in the environmental data to obtain the meteorological coefficient, combustible coefficient, and terrain coefficient. The elevation, slope, and slope direction are processed by binning to obtain the terrain coefficient. The human footprint index method is used to analyze the human activity data to obtain the human factor coefficient. The meteorological coefficient includes the influence of wind speed, temperature, and precipitation; the terrain coefficient includes the influence of elevation, slope, and slope direction. The meteorological data, vegetation data, and terrain data are input into the coefficient fusion model, and the meteorological coefficient, combustible coefficient, terrain coefficient, and human factor coefficient are standardized. The wind speed, temperature, and precipitation in the meteorological coefficient are standardized by standard scores and weighted. The dynamic coefficients of wind speed, temperature, and precipitation are obtained by reinforcement learning parameter adjustment method. The expression of the meteorological coefficient after coefficient standardization is: The meteorological coefficient after coefficient standardization is , wind speed is , the temperature is T, the rainfall is P, and the dynamic coefficient of wind speed is , the dynamic coefficient of temperature is , the dynamic coefficient of rainfall is ; The fire driving coefficient is obtained based on the standardized meteorological coefficient, combustible coefficient, terrain coefficient and human factor coefficient. The expression is: The fire driving coefficient of the xth area is , the meteorological data after the x-th regional coefficient is standardized: , the combustible coefficient after the x-th region coefficient is standardized , the terrain coefficient after normalization of the x-th regional coefficient is , the human factor coefficient after standardization of the x-th regional coefficient is , the trainable parameters are , ; The sliding window mechanism is used to update the trainable parameters, and the expression is: The learning rate is , the loss function is Loss, and the updated trainable parameters for , the updated trainable parameters for ; Output the adjusted fire driving coefficient.

[0011] Furthermore, the method for constructing a forest fire risk prediction model according to the fire driving coefficient and the candidate area includes: The candidate areas are marked by the fire driving coefficient, and the candidate areas with a fire driving coefficient greater than 0.615 are taken as key monitoring areas. The fire driving coefficient is taken as the risk probability of the key monitoring area. The objective function is constructed based on the actual fire situation and the key monitoring area. The expression is: The objective function is , the fire classification loss function is , the fire intensity regression loss is , the consistency loss of fire spread direction is , the first parameter is , the second parameter is , the third parameter is ; The fire classification loss function is the consistency between the risk probability and whether there is a fire. When there is a fire, the fire classification loss function dimension is 0, otherwise it is 1; The forest fire risk prediction model includes autoencoders, comparison algorithms, and machine learning algorithms; The autoencoder compresses the input data into a low-dimensional feature representation through the encoder, reconstructs the input data through the decoder, and is trained based on minimizing the reconstruction error to extract a low-dimensional feature vector that effectively represents the core features of the input data; The comparison algorithm analyzes the difference between the low-dimensional feature vector and the preset standard low-dimensional feature vector, and uses the low-dimensional feature vector with a difference less than 0.371 as the key feature; The machine learning algorithm optimizes the objective function, uses historical data to train the model based on key features to learn the fire occurrence pattern, and adjusts the parameters through the validation set to predict the risk of forest fires.

[0012] Furthermore, the method for optimizing the forest fire risk prediction model according to the fire spread law includes: The spatiotemporal distribution features are taken as nodes, and the spatiotemporal distribution features are screened according to the fire probability. The nodes with a value greater than the fire threshold are taken as wave source nodes, and the spreading radius and energy value of the wave source nodes are calculated as follows: The fire spread speed at time t is , the energy attenuation coefficient is , the spreading radius of the wave source node at time t is , the spreading radius of the wave source node at time t-1 is , the energy value of the wave source node at time t is , the energy value of the initial wave source node is , the circumference constant is , the energy of the bth wave source node at the tth time is ; If the distance from the wave source node to the jth node is greater than , then the initial spread fails to spread to the jth node and needs to arrive at t=t+2 or later; if the distance from the source node to the jth node is less than or equal to ,and If the value is greater than or equal to the connection threshold of the jth node, the initial spread spreads to the jth node and the jth node is activated, and a directed edge is established from the source node to the jth node; Calculate the fire spread probability of the activated node: The probability attenuation coefficient is , the connection threshold of the jth activated node is , the energy value of the j-th wave source node at the t-th time is , the fire spread probability of the jth activated node is ; Calculate the initial energy value of the activated node: The response coefficient of the jth activated node is , the initial energy value of the jth activated node is ; Calculate the updated node spreading radius and energy value: The spreading radius of the bth activated node at the tth moment after the update is , the spreading radius of the bth activated node at the t-1th moment after the update is , the energy value of the bth activated node at the tth moment after the update is , the initial energy value of the bth activated node is ; If the distance from the activated node to the i-th neighboring node is greater than , the fire spread of the bth activated node fails to spread to the i-th neighboring node; the distance from the activated node to the i-th neighboring node is less than or equal to ,and If the connection threshold of the bth node is greater than or equal to that of the bth node, the fire of the bth activated node spreads to the i-th neighboring node, and the energy value generated by the activation of the i-th neighboring node is calculated: The energy value of the i-th neighboring node is , the response coefficient of the i-th neighboring node is , the energy value of the i-th neighboring node at the t-th time after the update is , the fire spread probability of the i-th activated node is ; Continue to iterate until all nodes are traversed and the fire spread situation is given. The fire driving coefficient of the forest fire risk prediction model is adjusted according to the variance between the fire spread situation and the actual spread situation.

[0013] In the second aspect, a forest fire risk prediction system based on machine learning includes: Data acquisition module: used to use the data of the monitoring data source of the preset forest within the specified time period as the data to be analyzed; the monitoring data source includes a reference data source and a pending data source; the fire risk law of the reference data source is higher than that of the pending data source; the data includes environmental data, image data, historical fire data, human activity data, and remote sensing data; the environmental data includes meteorological data, vegetation data, and terrain data; Comparison and division module: used to extract the spatiotemporal distribution features of the data to be analyzed and the reference data source, and divide the spatiotemporal distribution features into regions according to the similarity, and take the region where the data to be analyzed with a similarity higher than a similarity threshold as a candidate region, and vice versa as a fuzzy region; Adding an impact analysis module: used to perform autocorrelation analysis on the fuzzy area based on the pending data source, add the fuzzy area with autocorrelation higher than the autocorrelation threshold to the candidate area, and perform time-varying impact analysis on the environmental data and the human activity data to obtain a fire driving coefficient; Model building and optimization module: used to build a forest fire risk prediction model according to the fire driving coefficient and the candidate area, optimize the forest fire risk prediction model according to the fire spread law, input the data to be predicted into the forest fire risk prediction model, and output the prediction result.

[0014] The beneficial effects of the present invention are: The present invention is a forest fire risk prediction method and system based on machine learning. Compared with the prior art, the present invention has the following technical effects: The present invention can improve the accuracy of forest fire risk prediction by extracting spatiotemporal distribution characteristics, regional division, autocorrelation analysis, time-varying impact analysis, acquiring model construction and optimizing model steps, thereby improving the accuracy of forest fire risk prediction, and optimizing forest fire risk prediction, which can greatly save resources and improve work efficiency. It can realize intelligent prediction of forest fire risk, and perform regional division and fire spread optimization for forest fire risk prediction in real time, which is of great significance to forest fire risk prediction, can adapt to forest fire risk prediction of different standards and different forest fire risk prediction needs, and has a certain universality. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flowchart of the steps of a forest fire risk prediction method based on machine learning in the present invention. DETAILED DESCRIPTION

[0016] The present invention is further described below by means of specific embodiments. The illustrative embodiments and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0017] The present invention provides a forest fire risk prediction method and system based on machine learning, comprising the following steps: like Figure 1 As shown, in this embodiment, the following steps are included: The data of the monitoring data source of the preset forest within the specified time period is used as the data to be analyzed; the monitoring data source includes a reference data source and a pending data source; the fire risk law of the reference data source is higher than that of the pending data source; the data includes environmental data, image data, historical fire data, human activity data, and remote sensing data; the environmental data includes meteorological data, vegetation data, and terrain data; In the actual assessment, a certain forest was taken as the research object, and data from October 2019 to October 2024 were collected as historical data to study the forest fire risk from 17:34 to 19:00 on December 3, 2024. The meteorological data were temperature 25℃, humidity 40%, wind speed 5m / s, and southeast wind direction; vegetation data were vegetation coverage 70%, vegetation type mainly coniferous forest, and vegetation dryness 0.6; terrain data were 513 meters above sea level, slope 17°, and slope to the south; historical fire data showed that there had been three small fires in the area in the past five years, all of which occurred in the hot and dry summer period; human activity data showed that there was a small road around the forest, with about 100 vehicles passing by every day, mainly local residents' vehicles and a small number of tourist vehicles; remote sensing data was the overall distribution of the forest and vegetation health information obtained through satellite remote sensing, showing that there were no obvious abnormal hot spots in the current forest; Extracting the spatiotemporal distribution features of the data to be analyzed and the reference data source, dividing the spatiotemporal distribution features into regions according to similarity, and defining the region where the data to be analyzed is located with a similarity higher than a similarity threshold as a candidate region, and vice versa as a fuzzy region; The spatiotemporal distribution characteristics of the data to be analyzed are vegetation distribution, terrain distribution, human activity distribution, meteorological conditions, and time patterns of human activities; vegetation distribution includes type and cover, dryness, and health status; terrain distribution includes altitude and slope, and the impact of terrain on vegetation; human activity distribution includes roadsides and camping areas; meteorological conditions include seasonal changes and short-term meteorological changes; and time patterns of human activities include camping activities and vehicle traffic. The similarity threshold is 0.748; the candidate area is D, and the fuzzy areas are A, B, and C; Performing autocorrelation analysis on the fuzzy area based on the pending data source, adding the fuzzy area with autocorrelation higher than the autocorrelation threshold to the candidate area, and performing time-varying impact analysis on the environmental data and the human activity data to obtain a fire driving coefficient; In the actual evaluation, the autocorrelation threshold is 0.263, and region A is added to the candidate regions; the fire driving coefficients of candidate regions A and D are 0.532 and 0.624, respectively; Constructing a forest fire risk prediction model according to the fire driving coefficient and the candidate area, optimizing the forest fire risk prediction model according to the fire spread law, inputting the data to be predicted into the forest fire risk prediction model, and outputting the prediction result; In the actual evaluation, the risk probabilities of candidate regions A and D are 0 and 0.624, respectively.

[0018] In this embodiment, the method of dividing the spatiotemporal distribution features into regions according to similarity includes: Calculate the similarity between the spatiotemporal distribution features in the data to be analyzed and the spatiotemporal distribution features in the reference data source, and mark the spatiotemporal distribution features in the data to be analyzed with a similarity higher than 0.473 in the latitude and longitude map of the preset area; The spatiotemporal distribution characteristics are divided into primary regions according to the clustering density of the pre-annotated latitude and longitude map of the preset area to obtain the primary region; the fire probability of the spatiotemporal distribution characteristics is calculated: The ignition probability of the cth spatiotemporal distribution feature is , the wind intensity in the ath direction is , the fire temperature of the neighbor in the ath direction is , the ignition point of the cth spatiotemporal distribution feature is , the wind direction vector in the ath direction is , the neighbor coordinate vector in the ath direction is , the number of directions is ; The density clustering algorithm is used to perform primary regional adjustment on the spatiotemporal distribution features with a fire probability higher than 0.316 to obtain the adjusted area. The adjusted area containing the spatiotemporal distribution features with a fire probability higher than 0.316 is taken as the first area, otherwise it is taken as the fuzzy area.

[0019] In this embodiment, the method of performing autocorrelation analysis on the fuzzy area based on the pending data source includes: Extract spatial distribution features of the pending data source to obtain a pending spatial feature set, take the spatial distribution features of the reference data source as reference features; take the spatial distribution features of the fuzzy area as fuzzy features; The different spatial distribution characteristics of the to-be-determined spatial feature set and the reference features are taken as the to-be-determined spatial features, and the autocorrelation between the to-be-determined spatial features and the fuzzy features is calculated.

[0020] In this embodiment, the method for performing time-varying impact analysis on the environmental data and the human activity data to obtain a fire driving coefficient includes: The long short-term memory network is used to analyze the meteorological data, vegetation data, and terrain data in the environmental data to obtain the meteorological coefficient, combustible coefficient, and terrain coefficient. The elevation, slope, and slope direction are processed by binning to obtain the terrain coefficient. The human footprint index method is used to analyze the human activity data to obtain the human factor coefficient. The meteorological coefficient includes the influence of wind speed, temperature, and precipitation; the terrain coefficient includes the influence of elevation, slope, and slope direction. The meteorological data, vegetation data, and terrain data are input into the coefficient fusion model, and the meteorological coefficient, combustible coefficient, terrain coefficient, and human factor coefficient are standardized. The wind speed, temperature, and precipitation in the meteorological coefficient are standardized by standard scores and weighted. The dynamic coefficients of wind speed, temperature, and precipitation are obtained by reinforcement learning parameter adjustment method. The expression of the meteorological coefficient after coefficient standardization is: The meteorological coefficient after coefficient standardization is , wind speed is , the temperature is T, the rainfall is P, and the dynamic coefficient of wind speed is , the dynamic coefficient of temperature is , the dynamic coefficient of rainfall is ; The fire driving coefficient is obtained based on the standardized meteorological coefficient, combustible coefficient, terrain coefficient and human factor coefficient. The expression is: The fire driving coefficient of the xth area is , the meteorological data after the x-th regional coefficient is standardized: , the combustible coefficient after the x-th region coefficient is standardized , the terrain coefficient after normalization of the x-th regional coefficient is , the human factor coefficient after standardization of the x-th regional coefficient is , the trainable parameters are , ; The sliding window mechanism is used to update the trainable parameters, and the expression is: The learning rate is , the loss function is Loss, and the updated trainable parameters for , the updated trainable parameters for ; Output the adjusted fire driving coefficient.

[0021] In this embodiment, the method for constructing a forest fire risk prediction model according to the fire driving coefficient and the candidate area includes: The candidate areas are marked by the fire driving coefficient, and the candidate areas with a fire driving coefficient greater than 0.615 are taken as key monitoring areas. The fire driving coefficient is taken as the risk probability of the key monitoring area. The objective function is constructed based on the actual fire situation and the key monitoring area. The expression is: The objective function is , the fire classification loss function is , the fire intensity regression loss is , the consistency loss of fire spread direction is , the first parameter is , the second parameter is , the third parameter is ; The fire classification loss function is the consistency between the risk probability and whether there is a fire. When there is a fire, the fire classification loss function dimension is 0, otherwise it is 1; The forest fire risk prediction model includes autoencoders, comparison algorithms, and machine learning algorithms; The autoencoder compresses the input data into a low-dimensional feature representation through the encoder, reconstructs the input data through the decoder, and is trained based on minimizing the reconstruction error to extract a low-dimensional feature vector that effectively represents the core features of the input data; The comparison algorithm analyzes the difference between the low-dimensional feature vector and the preset standard low-dimensional feature vector, and uses the low-dimensional feature vector with a difference less than 0.371 as the key feature; The machine learning algorithm optimizes the objective function, uses historical data to train the model based on key features to learn the fire occurrence pattern, and adjusts the parameters through the validation set to predict the risk of forest fires.

[0022] In this embodiment, the method for optimizing the forest fire risk prediction model according to the fire spread law includes: The spatiotemporal distribution features are taken as nodes, and the spatiotemporal distribution features are screened according to the probability of ignition. The nodes with a value greater than the ignition threshold are taken as wave source nodes, and the spreading radius and energy value of the wave source nodes are calculated as follows: The fire spread speed at time t is , the energy attenuation coefficient is , the spreading radius of the wave source node at time t is , the spreading radius of the wave source node at time t-1 is , the energy value of the wave source node at time t is , the energy value of the initial wave source node is , the circumference constant is , the energy of the bth wave source node at the tth time is ; If the distance from the wave source node to the jth node is greater than , then the initial spread fails to spread to the jth node and needs to arrive at t=t+2 or later; if the distance from the source node to the jth node is less than or equal to ,and If the value is greater than or equal to the connection threshold of the jth node, the initial spread spreads to the jth node and the jth node is activated, and a directed edge is established from the source node to the jth node; Calculate the fire spread probability of the activated node: The probability attenuation coefficient is , the connection threshold of the jth activated node is , the energy value of the j-th wave source node at the t-th time is , the fire spread probability of the jth activated node is ; Calculate the initial energy value of the activated node: The response coefficient of the jth activated node is , the initial energy value of the jth activated node is ; Calculate the updated node spreading radius and energy value: The spreading radius of the bth activated node at the tth moment after the update is , the spreading radius of the bth activated node at the t-1th moment after the update is , the energy value of the bth activated node at the tth moment after the update is , the initial energy value of the bth activated node is ; If the distance from the activated node to the i-th neighboring node is greater than , the fire spread of the bth activated node fails to spread to the i-th neighboring node; the distance from the activated node to the i-th neighboring node is less than or equal to ,and If the connection threshold of the bth node is greater than or equal to that of the bth node, the fire of the bth activated node spreads to the i-th neighboring node, and the energy value generated by the activation of the i-th neighboring node is calculated: The energy value of the i-th neighboring node is , the response coefficient of the i-th neighboring node is , the energy value of the i-th neighboring node at the t-th time after the update is , the fire spread probability of the i-th activated node is ; Continue to iterate until all nodes are traversed and the fire spread situation is given. The fire driving coefficient of the forest fire risk prediction model is adjusted according to the variance between the fire spread situation and the actual spread situation.

[0023] In the second aspect, a forest fire risk prediction system based on machine learning includes: Data acquisition module: used to use the data of the monitoring data source of the preset forest within the specified time period as the data to be analyzed; the monitoring data source includes a reference data source and a pending data source; the fire risk law of the reference data source is higher than that of the pending data source; the data includes environmental data, image data, historical fire data, human activity data, and remote sensing data; the environmental data includes meteorological data, vegetation data, and terrain data; Comparison and division module: used to extract the spatiotemporal distribution features of the data to be analyzed and the reference data source, divide the spatiotemporal distribution features into regions according to the similarity, and take the region where the data to be analyzed with a similarity higher than a similarity threshold as a candidate region, otherwise as a fuzzy region; Adding an impact analysis module: used to perform autocorrelation analysis on the fuzzy area based on the pending data source, add the fuzzy area with autocorrelation higher than the autocorrelation threshold to the candidate area, and perform time-varying impact analysis on the environmental data and the human activity data to obtain a fire driving coefficient; Model building and optimization module: used to build a forest fire risk prediction model according to the fire driving coefficient and the candidate area, optimize the forest fire risk prediction model according to the fire spread law, input the data to be predicted into the forest fire risk prediction model, and output the prediction result.

[0024] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A forest fire risk prediction method based on machine learning, characterized in that: The following steps are involved: The data of the monitoring data source of the preset forest within the specified time period is used as the data to be analyzed; the monitoring data source includes a reference data source and a pending data source; The fire risk pattern of the reference data source is higher than that of the pending data source; the data includes environmental data, image data, historical fire data, human activity data, and remote sensing data; the environmental data includes meteorological data, vegetation data, and terrain data; Extracting the spatiotemporal distribution features of the data to be analyzed and the reference data source, dividing the spatiotemporal distribution features into regions according to similarity, and defining the region where the data to be analyzed is located with a similarity higher than a similarity threshold as a candidate region, and vice versa as a fuzzy region; Performing autocorrelation analysis on the fuzzy area based on the pending data source, adding the fuzzy area with autocorrelation higher than the autocorrelation threshold to the candidate area, and performing time-varying impact analysis on the environmental data and the human activity data to obtain a fire driving coefficient; A forest fire risk prediction model is constructed according to the fire driving coefficient and the candidate area, the forest fire risk prediction model is optimized according to the fire spread law, the data to be predicted is input into the forest fire risk prediction model, and the prediction result is output.

2. A forest fire risk prediction method based on machine learning according to claim 1, characterized in that: The method for dividing the spatiotemporal distribution features into regions according to similarity includes: Calculate the similarity between the spatiotemporal distribution features in the data to be analyzed and the spatiotemporal distribution features in the reference data source, and mark the spatiotemporal distribution features in the data to be analyzed with a similarity higher than 0.473 in the latitude and longitude map of the preset area; The spatiotemporal distribution characteristics are divided into primary regions according to the clustering density of the pre-annotated latitude and longitude map of the preset area to obtain the primary region; the fire probability of the spatiotemporal distribution characteristics is calculated: The ignition probability of the cth spatiotemporal distribution feature is , the wind intensity in the ath direction is , the ignition temperature of the neighbor in the ath direction is , the ignition point of the cth spatiotemporal distribution feature is , the wind direction vector in the ath direction is , the neighbor coordinate vector in the ath direction is , the number of directions is ; The density clustering algorithm is used to perform primary regional adjustment on the spatiotemporal distribution features with a fire probability higher than 0.316 to obtain the adjusted area. The adjusted area containing the spatiotemporal distribution features with a fire probability higher than 0.316 is taken as the first area, otherwise it is taken as the fuzzy area.

3. The method for predicting forest fire risk based on machine learning according to claim 1, characterized in that: The method for performing autocorrelation analysis on the fuzzy area based on the pending data source includes: Extract spatial distribution features of the pending data source to obtain a pending spatial feature set, take the spatial distribution features of the reference data source as reference features; take the spatial distribution features of the fuzzy area as fuzzy features; The different spatial distribution characteristics of the to-be-determined spatial feature set and the reference features are taken as the to-be-determined spatial features, and the autocorrelation between the to-be-determined spatial features and the fuzzy features is calculated.

4. The method for predicting forest fire risk based on machine learning according to claim 1, characterized in that: The method for performing time-varying impact analysis on the environmental data and the human activity data to obtain a fire driving coefficient includes: The long short-term memory network is used to analyze the meteorological data, vegetation data, and terrain data in the environmental data to obtain the meteorological coefficient, combustible coefficient, and terrain coefficient. The elevation, slope, and slope direction are processed by binning to obtain the terrain coefficient. The human footprint index method is used to analyze the human activity data to obtain the human factor coefficient. The meteorological coefficient includes the influence of wind speed, temperature, and precipitation; the terrain coefficient includes the influence of elevation, slope, and slope direction. The meteorological data, vegetation data, and terrain data are input into the coefficient fusion model, and the meteorological coefficient, combustible coefficient, terrain coefficient, and human factor coefficient are standardized. The wind speed, temperature, and precipitation in the meteorological coefficient are standardized by standard scores and weighted. The dynamic coefficients of wind speed, temperature, and precipitation are obtained by reinforcement learning parameter adjustment method. The expression of the meteorological coefficient after coefficient standardization is: The meteorological coefficient after coefficient standardization is , wind speed is , the temperature is T, the rainfall is P, and the dynamic coefficient of wind speed is , the dynamic coefficient of temperature is , the dynamic coefficient of rainfall is ; The fire driving coefficient is obtained based on the standardized meteorological coefficient, combustible coefficient, terrain coefficient and human factor coefficient. The expression is: The fire driving coefficient of the xth area is , the meteorological data after the x-th regional coefficient is standardized: , the combustible coefficient after the x-th region coefficient is standardized , the terrain coefficient after normalization of the x-th regional coefficient is , the human factor coefficient after standardization of the x-th regional coefficient is , the trainable parameters are , ; The sliding window mechanism is used to update the trainable parameters, and the expression is: The learning rate is , the loss function is Loss, and the updated trainable parameters for , the updated trainable parameters for ; Output the adjusted fire driving coefficient.

5. The method for predicting forest fire risk based on machine learning according to claim 1, characterized in that: The method for constructing a forest fire risk prediction model according to the fire driving coefficient and the candidate area comprises: The candidate areas are marked by the fire driving coefficient, and the candidate areas with a fire driving coefficient greater than 0.615 are taken as key monitoring areas. The fire driving coefficient is taken as the risk probability of the key monitoring area. The objective function is constructed based on the actual fire situation and the key monitoring area. The expression is: The objective function is , the fire classification loss function is , the fire intensity regression loss is , the consistency loss of fire spread direction is , the first parameter is , the second parameter is , the third parameter is ; The fire classification loss function is the consistency between the risk probability and whether there is a fire. When there is a fire, the fire classification loss function dimension is 0, otherwise it is 1; The forest fire risk prediction model includes autoencoders, comparison algorithms, and machine learning algorithms; The autoencoder compresses the input data into a low-dimensional feature representation through the encoder, reconstructs the input data through the decoder, and is trained based on minimizing the reconstruction error to extract a low-dimensional feature vector that effectively represents the core features of the input data; The comparison algorithm analyzes the difference between the low-dimensional feature vector and the preset standard low-dimensional feature vector, and uses the low-dimensional feature vector with a difference less than 0.371 as the key feature; The machine learning algorithm optimizes the objective function, uses historical data to train the model based on key features to learn the fire occurrence pattern, and adjusts the parameters through the validation set to predict the risk of forest fires.

6. The method for predicting forest fire risk based on machine learning according to claim 1, characterized in that: The method for optimizing the forest fire risk prediction model according to the fire spread law includes: The spatiotemporal distribution features are taken as nodes, and the spatiotemporal distribution features are screened according to the probability of ignition. The nodes with a value greater than the ignition threshold are taken as wave source nodes, and the spreading radius and energy value of the wave source nodes are calculated as follows: The fire spread speed at time t is , the energy attenuation coefficient is , the spreading radius of the wave source node at time t is , the spreading radius of the wave source node at time t-1 is , the energy value of the wave source node at time t is , the energy value of the initial wave source node is , the circumference constant is , the energy of the bth wave source node at the tth time is ; If the distance from the wave source node to the jth node is greater than , then the initial spread fails to spread to the jth node and needs to arrive at t=t+2 or later; if the distance from the source node to the jth node is less than or equal to ,and If the value is greater than or equal to the connection threshold of the jth node, the initial spread spreads to the jth node and the jth node is activated, and a directed edge is established from the source node to the jth node; Calculate the fire spread probability of the activated node: The probability attenuation coefficient is , the connection threshold of the jth activated node is , the energy value of the j-th wave source node at the t-th time is , the fire spread probability of the jth activated node is ; Calculate the initial energy value of the activated node: The response coefficient of the jth activated node is , the initial energy value of the jth activated node is ; Calculate the updated node spreading radius and energy value: The spreading radius of the bth activated node at the tth moment after the update is , the spreading radius of the bth activated node at the t-1th moment after the update is , the energy value of the bth activated node at the tth moment after the update is , the initial energy value of the bth activated node is ; If the distance from the activated node to the i-th neighboring node is greater than , the fire spread of the bth activated node fails to spread to the i-th neighboring node; the distance from the activated node to the i-th neighboring node is less than or equal to ,and If the connection threshold of the bth node is greater than or equal to that of the bth node, the fire of the bth activated node spreads to the i-th neighboring node, and the energy value generated by the activation of the i-th neighboring node is calculated: The energy value of the i-th neighboring node is , the response coefficient of the i-th neighboring node is , the energy value of the i-th neighboring node at the t-th time after the update is , the fire spread probability of the i-th activated node is ; Continue to iterate until all nodes are traversed and the fire spread situation is given. The fire driving coefficient of the forest fire risk prediction model is adjusted according to the variance between the fire spread situation and the actual spread situation.

7. A forest fire risk prediction system based on machine learning, used to execute the method according to any one of claims 1 to 6, characterized in that: include: Data collection module: used to collect data from the monitoring data source of the preset forest within a specified time period as data to be analyzed; The monitoring data source includes a reference data source and a pending data source; The fire risk pattern of the reference data source is higher than that of the pending data source; the data includes environmental data, image data, historical fire data, human activity data, and remote sensing data; the environmental data includes meteorological data, vegetation data, and terrain data; Comparison and division module: used to extract the spatiotemporal distribution features of the data to be analyzed and the reference data source, divide the spatiotemporal distribution features into regions according to the similarity, and take the region where the data to be analyzed with a similarity higher than a similarity threshold as a candidate region, otherwise as a fuzzy region; Adding an impact analysis module: used to perform autocorrelation analysis on the fuzzy area based on the pending data source, add the fuzzy area with autocorrelation higher than the autocorrelation threshold to the candidate area, and perform time-varying impact analysis on the environmental data and the human activity data to obtain a fire driving coefficient; Model building and optimization module: used to build a forest fire risk prediction model according to the fire driving coefficient and the candidate area, optimize the forest fire risk prediction model according to the fire spread law, input the data to be predicted into the forest fire risk prediction model, and output the prediction result.

Images