CN105678419A  Fine gritbased forest fire hazard probability forecasting system  Google Patents
Fine gritbased forest fire hazard probability forecasting system Download PDFInfo
 Publication number
 CN105678419A CN105678419A CN201610010997.8A CN201610010997A CN105678419A CN 105678419 A CN105678419 A CN 105678419A CN 201610010997 A CN201610010997 A CN 201610010997A CN 105678419 A CN105678419 A CN 105678419A
 Authority
 CN
 China
 Prior art keywords
 forest fires
 probability
 data
 forest
 fire
 Prior art date
 Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
 Pending
Links
 238000004364 calculation method Methods 0.000 claims description 18
 238000007619 statistical method Methods 0.000 claims description 16
 230000000694 effects Effects 0.000 claims description 14
 239000000203 mixture Substances 0.000 claims description 13
 230000000875 corresponding Effects 0.000 claims description 10
 238000000034 method Methods 0.000 claims description 9
 XLYOFNOQVPJJNPUHFFFAOYSAN water Substances O XLYOFNOQVPJJNPUHFFFAOYSAN 0.000 claims description 6
 238000001556 precipitation Methods 0.000 claims description 5
 230000002123 temporal effect Effects 0.000 claims description 5
 241001269238 Data Species 0.000 claims description 4
 238000009472 formulation Methods 0.000 claims description 4
 238000007476 Maximum Likelihood Methods 0.000 claims description 3
 230000005540 biological transmission Effects 0.000 claims description 3
 238000010192 crystallographic characterization Methods 0.000 claims description 3
 230000001419 dependent Effects 0.000 claims description 3
 239000000284 extract Substances 0.000 claims description 2
 238000005094 computer simulation Methods 0.000 abstract 1
 238000010276 construction Methods 0.000 abstract 1
 238000004393 prognosis Methods 0.000 description 10
 230000002354 daily Effects 0.000 description 2
 230000018109 developmental process Effects 0.000 description 2
 238000010586 diagram Methods 0.000 description 2
 238000005516 engineering process Methods 0.000 description 2
 230000003203 everyday Effects 0.000 description 2
 238000011068 load Methods 0.000 description 2
 238000010606 normalization Methods 0.000 description 2
 230000002265 prevention Effects 0.000 description 2
 230000000007 visual effect Effects 0.000 description 2
 230000032683 aging Effects 0.000 description 1
 230000015572 biosynthetic process Effects 0.000 description 1
 230000001364 causal effect Effects 0.000 description 1
 238000002485 combustion reaction Methods 0.000 description 1
 230000000295 complement Effects 0.000 description 1
 238000011156 evaluation Methods 0.000 description 1
 238000005755 formation reaction Methods 0.000 description 1
 239000000446 fuel Substances 0.000 description 1
 230000002452 interceptive Effects 0.000 description 1
 238000009533 lab test Methods 0.000 description 1
 239000011159 matrix material Substances 0.000 description 1
 238000007670 refining Methods 0.000 description 1
 238000005070 sampling Methods 0.000 description 1
Classifications

 G—PHYSICS
 G06—COMPUTING; CALCULATING; COUNTING
 G06Q—DATA PROCESSING SYSTEMS OR METHODS, SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL, SUPERVISORY OR FORECASTING PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL, SUPERVISORY OR FORECASTING PURPOSES, NOT OTHERWISE PROVIDED FOR
 G06Q10/00—Administration; Management
 G06Q10/04—Forecasting or optimisation, e.g. linear programming, "travelling salesman problem" or "cutting stock problem"
Abstract
The invention relates to the filed of forest fire prediction and aims to solve the problem in the prior art that a conventional forest fire forecasting system is not high in prediction precision, coarse in spatial granularity and the like, which provides a novel forest fire probability prediction system. According to the technical scheme of the invention, a fine gritbased forest fire hazard probability forecasting system comprises a relational database module. The relational database module further comprises a weather observation database, a weather forecast database, a model input factor database, a model database and a forest fire probability database. The weather observation database is used for storing weather observation data on that day and updating data in the model input factor database. The weather forecast database is used for storing weather forecast data applied to the bayesian network inference. The model input factor database is used for storing historical data and spatial data adapted to the construction of a bayesian network model and a probability distribution model. The model database is used for storing the model files of the system and providing a computational model for calculating the forest fire probability in each region. The system is mainly applied to the forest fire forecasting field.
Description
Technical field
The present invention relates to a kind of forest fire prognosis and prediction system, particularly relate to a kind of forest fire prognosis and prediction system based on datadriven method modeling.
Background technology
Forest is one of important resource, and forest fire is the major casualty affecting forest development, not only directly affects forest ecology balance, causes economy and ecological resources to run off, and injures the safety [1] of people's lives and properties. Design simple and practical forest fire prognosis and prediction system, be the important measures reducing forest fire harm.
The external research work to forest fire prognosis and prediction is carried out relatively early. The U.S. develops country's fire size class system (NFDRS) [2], using many factors such as meteorology, landform, combustibles as input, develops based on the theory of combustion and laboratory test and sets up physical model, calculate risk of forest fire index. But it is extremely complex that NFDRS forms structure, more difficult enforcement [3] in actual applications; And the method that its forecasting model is based on physics, it is necessary to carry out substantial amounts of experiment.
Canada develops Canadian Forest weather conditions conducive to wildfires index (FWI) system [4], and its structure is as shown in Figure 1. This system is from Fuel loads balance theory, using meteorological measuring 4 kinds basic as input, sets up computation model by a large amount of fire trials, through a series of derivations, calculating, finally carries out fire danger prediction with weather conditions conducive to wildfires index. But it only considers the impact that forest fires are occurred by meteorological factor, have ignored the spatial variations of the factor such as combustible and landform [3].
Korea S, for the feature of this country's forest fire, develops National Forest fire risk rating forecast system (KFFDRI) [5] suitable in Korea S area, as shown in Figure 2. KFFDRI forest fire historical data between Korea S 19972001 and the same period meteorological data and 126 scene of fire survey datas based on, set up half mechanism half statistical model based on meteorological factor, the combustible factor and terrain factor and carry out forest fire size class forecast. But KFFDRI is only applicable to Korea S area, not easily promote [6].
Australia constructs the forest fire danger class system (McArthur) [3] of this country in the later stage fifties 19th century.Four submodels are studied and formulated to this system, based on the igniting experiments of field: arid factor model, combustible Humidity Model, rate of propagation model and difficulty model of putting out a fire to save life and property, for the evaluation of fire size class. The input factor of this system includes longterm drought index and recent precipitation, temperature, humidity and wind speed, formulates forest fire danger class according to the degree of difficulty of putting out a fire to save life and property of fire.
The research of forest fire forecast is started late by China, and mainly on the performance basis of the country such as the U.S., Canada, carry out in conjunction with the practical situation of China, such as " China Meteorological industry standard (QX/T772007) forest fire danger class " issued by China Meteorological Administration and implement on October 1st, 2007. The input of the method includes 5 kinds of current meteorological factors and 7 kinds of early stage meteorological factors, each meteorological factor contribution degree to risk of forest fire is calculated first with computing formula, then calculate current factor index and preceding factor index, finally utilize the result of previous step to calculate Forest Fire Danger Weather Index. The system that current China is applied to Forest Fire Monitoring is more, but the formation system for forest fire forecast is less. Although China has developed and developed tens of kinds of fire danger prediction methods, but up to the present but without setting up a set of national forest fire prognosis and prediction system [7] considering the influence factors such as meteorology, vegetation, geography and mankind's activity.
Cause that the factor that forest fires occur is numerous, mechanism is complicated, and in nonlinear relation between cause and effect, traditional forest fire prognosis and prediction system is intended to carry out forest fire forecast [8] by multiple meteorological factor is set up linear model by certain combination, but the landform residing for each forest farm is different, each factor is different to the influence degree of forest fires, and the generation of forest fires is also not only by the impact of meteorological factor, thus causing prediction effect unsatisfactory [9]; And the spatial granularity of tradition forest fire prognosis and prediction system is comparatively rough, ageing poor, it is impossible to meet the actual demand of forest fire protection at different levels work.
Summary of the invention
For overcoming the deficiencies in the prior art, for the tradition problems such as forest fire prognosis and prediction system prediction precision is not high, spatial granularity is thicker, design and Implement the forest fires probability forecast system of practicality to reach following target:
(1) consider the combined influence relation that forest fires are occurred by many factors, improve the precision of prediction of system;
(2) build Bayesian network model, the correlation theory of Bayesian network is applied to forest fire prognosis and prediction field, play Bayesian network in the advantage processing this type of uncertain problem, improve the precision of prediction of system;
(3) build probability assignments model, largescale forest fires probability is allocated, the spatial granularity of refining system;
(4) formulate forest fires risk class and by visualization, intuitively characterize forest fires occur complexity.
The technical solution used in the present invention is, finegrained forest fire probability forecast system, including:
Relational data library module, relational data library module farther includes meteorological measuring storehouse, weather forecast data base, mode input factor data storehouse, model database and forest fires probability database; The meteorological measuring on meteorological measuring library storage same day, inputs the data in factor data storehouse for Renewal model; Weather forecast database purchase is for the weather forecast data of Bayesian Network Inference; Mode input factor data library storage is for building historical data and the spatial data of Bayesian network model and probability assignments model, including meteorological historical data, forest fires historical data, geodata, ground mulching data, population distribution data; The result of calculation of forest fires probability data library storage system, including the forest fires probability at county level obtained by Bayesian Network Inference, by probability assignments model calculated 300m × 300m granularity forest fires probability, and corresponding forest fires risk class;The model file of model data library storage system, the forest fires probability for each area provides computation model;
Database management module, it is provided that to the process of Back ground Information in data base, it is achieved the importing of data, Data Format Transform, data edition, amendment, inquiry;
Data acquisition module completes system is inputted the collecting work of data, and wherein meteorological historical data needs to write web crawler and is acquired; Raster data Image Via Gis instrument including ground mulching, height above sea level extracts;
FWI Index for Calculation module completes forest fire weather Index for Calculation function, first by the period of the day from 11 a.m. to 1 p.m temperature, the period of the day from 11 a.m. to 1 p.m relative humidity, the period of the day from 11 a.m. to 1 p.m wind speed, 24 hours precipitation calculates and obtains 5 intermediate indexs, the FWI index on the same day is calculated again by 5 intermediate indexs, the regular module of data completes the discretization function to continuous data, to generate the sample data for Bayesian network model training and prediction;
Bayesian Network Inference module completes BN modeling and inference function, and first according to expertise, subjectivity determines node and the network structure of Bayesian network; Then, carry out network parameter study based on sample data, set up the Bayesian network model for forest fires probabilistic forecasting; Finally, application Junction tree makes inferences, it was predicted that go out forest fires probability at county level;
Probability assignments module completes forest fires probability assignments at county level to 300m × 300m granularity function, first forest fires historical record is carried out statistical analysis, obtain the value weight of each forest fires factor, it is then based on spatial data and the forest fires probability of forest fires probability calculation 300m × 300m granularity at county level, and is shown by display module.
BN modeling and inference function and the concrete steps doping forest fires probability at county level:
Node in the Bayesian network of forest fires probabilistic forecasting is divided into two classes: a class is forest fires generation nodes, characterization result; Another kind of is forest fires inducement nodes, characterizes reason; Internodal directed connection arc represents the cause effect relation between forest fires and inducement thereof, and the parameter value of each node characterizes the probability dependency between forest fires and inducement thereof;
BN modeling is carried out: the space reference of Bayesian network model is county or district based on temporal data and spatial data; The time reference of model is one day; Each node has several value states discrete, mutual exclusion, it means that the continuous number observed includes height above sea level, the continuous number of population density must be beforehand with slidingmodel control, and data are carried out discretization by application equifrequent discretization method [17];
The conditional probability table of height above sea level node, ground mulching node and population density node is determined by priori, it may be assumed that the mean sea level in each county, ground mulching and be constant in the population density short time;
The conditional probability table of all the other nodes except forest fires generation node, application maximum Likelihood learns, maximal possibility estimation, based on traditional statistical analysis thought, passes judgment on the fitting degree of sample and model, if data D is by sample (D according to the likelihood degree of sample with parameter_{1},D_{2},…,D_{m}) composition, then nodes X_{i}Parameter θ_{i}The general type of loglikelihood function be: L (θ_{i} D) and=logP (D  θ_{i}), θ_{i}Maximal possibility estimation, it is simply that make L (θ_{i} D) reach that maximum valueNamely
If nodes X_{i}Total r_{i}Individual value, its father node π (X_{i}) the total q of value_{i}Individual combination, assumes the architectural feature with Bayesian network according to independent same distribution, has:
Wherein D_{l}It is an observation sample, m_{ijk}For D meets X_{i}=k and π (X_{i}The quantity of the sample of)=j, noteThen θ_{i}Maximal possibility estimation be:
For forest fires generation node, his father's node set comprises ground mulching node, owing to the space reference of model is county or district, in real data, the value of this node does not determine that, but a probability distribution: when county's name is known, the probability of each ground mulching type is the ratio of the type area coverage in this county and this county area, therefore forest fires node needs to carry out maximal possibility estimation [18] based on " broken power sample ", it is assumed that X_{lc}For ground mulching node, X_{fire}For forest fires node, due to X_{lc}There are 13 kinds of value states, therefore by D_{l}It is split as 13 broken power samples:
Wherein x_{i}For X_{lc}Certain value,For sample (D_{l},X_{lc}=x_{i}) weight, meet constraints:Being known in real data, its value is sample D_{l}Ground mulching type x in affiliated county_{i}The ratio of area coverage and this county gross area, in forest fires nodes X_{fire}Parameter learning process in, each observation sample D_{l}All substituted by 13 complete broken power samples, remember D^{t}For whole broken power partial datas, then X_{fire}Method for parameter estimation be:
Wherein m_{fire,jk}For D^{t}In all meet X_{fire}=k and π (X_{fire}The weight sum of the sample of)=j, m_{fire,j}For D^{t}In all meet π (X_{fire}The weight sum of the sample of)=j;
Build after Bayesian network model, it is necessary to carry out Bayesian Network Inference and calculate forest fires probability at county level and Posterior probability distribution, it may be assumed that name node in county's in known network, FWI index node and month node value, month the value of node be evidence variable, be designated as X_{E}, calculate forest fires generation querying node variable X_{fire}Posterior probability distribution, carry out the Forecast reasoning from reason to result, the Junction tree (JunctionTreeAlgorithm, JTA) in application Accurate Reasoning method calculates forest fires probability at county level.
Application Junction tree makes inferences, dope forest fires probability at county level to comprise the concrete steps that, first bayesian network structure figure G is converted to the data structure of an auxiliary, namely combines tree T, then pass through the initialization definitions on T and information transmission carries out posterior probability calculating:
Wherein
Probability assignments model and the concrete steps calculating 300m × 300m granularity forest fires probability:
Probability assignments carries out in units of county, according to the forest fires probability at county level that Bayesian network forecasts, the region of 300m × 300m each in this county is carried out probability assignments, main on forest fires Probability Basis at county level, considering landform and the ground mulching disturbance degree to forest fires, mode input is the average forest fires probability in certain county, mean sea level H, gradient S, position, slope T in 300m × 300m region of required calculating, ground mulching L, totally 4 kinds of factors of influence; It is output as the forest fires probability of this fritter;
First the disturbance degree of abovementioned 4 elite stand cause of fire is calculated. The disturbance degree of the note height above sea level factor is P_{h}, the disturbance degree of slope location factor is P_{t}, the disturbance degree of slope factor is P_{s}, the disturbance degree of the ground mulching factor is P_{l}. According to the forest fires Effects of Factors Du He county forest fires probability in 300m × 300m regions all in whole county, computation rule coefficient, 4 kinds of Effects of Factors degree of each fritter are multiplied, then are multiplied by regularization coefficient, obtain the forest fires probability of this fritter. Computing formula is as follows:
Wherein, NA represents some special value, such as water body, glaciers etc. of ground mulching, and these local forest fires probability are 0, and forest fires probability is that the disturbance degree of every elite stand cause of fire is multiplied by regularization coefficient elsewhere;
Factor of influence disturbance degree adopts statistical method to be calculated: occur carrying out statistical analysis at forest fires historical record to every kind of factor of influence, using the value of every kind of factor of influence as independent variable, its to the disturbance degree of forest fires as dependent variable, the two is fitted by application normal distribution, and corresponding coefficient value is finely tuned, add up the sufficient statistic in forest fires historical record of the every kind of factor respectively;
Simulating the relation between height above sea level and forest fires disturbance degree according to normal distribution, in forest fires historical record, forest fires point mean sea level is 2257, and sample variance is 624, obtains height above sea level according to normal distribution the relation between forest fires disturbance degree and height above sea level is as follows:
Statistical nature and Korea National forest fire danger class forecast system, P according to existing forest fires historical record_{t}Value compares as follows with position, required fritter actual slope:
The foot of the hill: P_{t}=0.5,
Mountain valley: P_{t}=0.25,
Middle descending: P_{t}=0.1,
Middle upward slope: P_{t}=0.05.
Abovementioned rule obtains according to position, the slope statistics of forest fires scene in forest fires historical record, wherein forest fires have half to occur at the foot of the hill, have 1/4th generations in mountain valley, have 1/10th generations at middle descending, residue forest fires scene is middle upward slope, and the forest fires number of times of middle upward slope is less;
Equally, between slope factor disturbance degree and the gradient, relation is obtained by normal distribution matching, and computing formula is as follows:
Wherein, S represents the mean inclination in fritter;
Ground mulching is divided into 32 kinds, according to China Meteorological industry standard, each ground mulching type is arranged disturbance degree, wherein, glacier, farmland, water body, city etc. are NA beyond the ground mulching value of forest scope, and remaining every kind cover type assignment is the real number between 0 to 1;
After calculating each factor contributions degree, also need normalized coefficient: computational methods are as follows:
Wherein n is the fritter number of 300m × 300m in county, P_{i}Forest fires probability for each fritter.
Formulate the concrete steps of forest fires risk class:
Forest fires output ratio corresponding to forest fires probability carries out the formulation of forest fires risk class, makes P_{fire}Represent certain forest fires probability determined, N_{fire}Represent and predict the outcome middle forest fires probability more than P_{fire}Forest fires sample number, N_{total}Represent and predict the outcome middle forest fires probability more than P_{fire}Total sample number, then P_{fire}Corresponding forest fires output ratio computational methods are:
First the forest fires probability of all observation samples is calculated, several threshold values selected, the codomain of forest fires probability is divided into several intervals, for each threshold value, calculates the output ratio of its correspondence, according to the relative size of output ratio, formulate forest fires risk class.
The feature of the present invention and providing the benefit that:
Owing to present invention employs Bayesian network model and probability assignments model, thus the forecast result of the forest fires probability forecast system of the present invention is more accurate, and forecast granularity is thinner, is effective, practical system. This system can provide reliable reference for forest fire prevention, deployment of forces, equipment configuration etc.
Accompanying drawing illustrates:
Fig. 1: Canadian Forest weather conditions conducive to wildfires index system (FWI) structure chart.
Fig. 2: Korean forest fire risk rating forecast system (KFFDRI) structure chart.
Fig. 3: forest fires probability forecast system architecture diagram.
Fig. 4: forest fires probability forecast system function module figure.
Fig. 5: forest fires probability forecast system flow chart.
Fig. 6: the bayesian network structure of forest fires probability forecast.
Fig. 7: Junction tree flow chart.
Fig. 8: probability assignments schematic flow sheet.
Fig. 9: forest fires risk class formulates schematic diagram.
ROC curve on Figure 10: the second group of experiment test collection.
Figure 11: 2015 year daily forecast in February 26 result figure.
Forest fires probability forecast on February 26th, Figure 12: 2015 system forecast result partial enlarged drawing.
Figure 13: 2015 year daily forecast in March 22 result figure.
Forest fires probability forecast on March 22nd, Figure 14: 2015 system forecast result partial enlarged drawing.
Detailed description of the invention
Forest fires probability forecast system is using Java as development language, and MySQL, as database management tools, accesses system database by JDBC, develops under Eclipse programmed environment. System adopts standard threelayer architecture, respectively by representing that layer, application layer and data Layer form, as shown in Figure 3. Data Layer provides data support for whole system, stores the calculation result data of system simultaneously; Application layer provides system concrete function; Represent that layer provides the interactive interface of system and user.
System adopts relevant database MySQL to store data, in conjunction with JDBC fulfillment database basic operation. System database includes meteorological measuring storehouse, weather forecast data base, mode input factor data storehouse, model database and forest fires probability database. The meteorological measuring on meteorological measuring library storage same day, inputs the data in factor data storehouse for Renewal model; Weather forecast database purchase is for the weather forecast data of Bayesian Network Inference; Mode input factor data library storage is for building historical data and the spatial data of Bayesian network model and probability assignments model, including meteorological historical data, forest fires historical data, geodata, ground mulching data, population distribution data etc.; The result of calculation of forest fires probability data library storage system, including the forest fires probability at county level obtained by Bayesian Network Inference, by probability assignments model calculated 300m × 300m granularity forest fires probability, and corresponding forest fires risk class;The model file of model data library storage system, the forest fires probability for each area provides computation model.
System need to possess database management function, data acquisition function, FWI Index for Calculation function, the regular function of data, Bayesian Network Inference function and probability assignments function, as shown in Figure 4.
Database management module comprises the subfunctions such as temporal data management, spatial data management, it is provided that to the process of Back ground Information in data base, it is achieved functions such as the importing of data, Data Format Transform, data edition, amendment, inquiries. Data acquisition module completes system is inputted the collecting work of data, and wherein meteorological historical data needs to write web crawler and is acquired; The raster data such as ground mulching, height above sea level needs Image Via Gis instrument to extract. FWI Index for Calculation module completes Canadian Forest weather conditions conducive to wildfires Index for Calculation function, first by the period of the day from 11 a.m. to 1 p.m temperature, the period of the day from 11 a.m. to 1 p.m relative humidity, the period of the day from 11 a.m. to 1 p.m wind speed, 24 hours precipitation calculates and obtains 5 intermediate indexs, then calculated the FWI index on the same day by 5 intermediate indexs. The regular module of data completes the discretization function to continuous data, to generate the sample data for Bayesian network model training and prediction. Bayesian Network Inference module completes BN modeling and inference function, and first according to expertise, subjectivity determines node and the network structure of Bayesian network; Then, carry out network parameter study based on sample data, set up the Bayesian network model for forest fires probabilistic forecasting; Finally, application Junction tree makes inferences, it was predicted that go out forest fires probability at county level. Probability assignments module completes forest fires probability assignments at county level to 300m × 300m granularity function, first forest fires historical record is carried out statistical analysis, obtain the value weight of each forest fires factor, be then based on spatial data and the forest fires probability of forest fires probability calculation 300m × 300m granularity at county level.
System flow is as shown in Figure 5. First, using data such as meteorology, vegetation, geography, population distribution as input, comprehensive forest fires historical data sets up Bayesian network model, and applies Junction tree and carry out probability inference, it was predicted that go out forest fires probability at county level. Then, historical data is carried out statistical analysis and builds probability assignments model, forest fires probability at county level is adjusted to 300m × 300m granularity. Finally, according to forest fires probability formulate forest fires risk class, and by visualization.
The Bayesian network of forest fires probability forecast
Bayesian network (BayesianNetworks, BN) [11], also known as belief network, it is a kind of Statistical Inference grown up on the basis of Bayesian decision method in multivariate statistical analysis technology, was proposed [12] in 1988 first by Pearl. Bayesian network with unique uncertainty knowledge expressionform, abundant probability ability to express, comprehensive priori incremental learning characteristic become the focus of theoretical research in recent years, it is widely used in numerous areas, as, tsunami risk assessment [13], hydrologic(al) prognosis [14], injures assessment [15] etc.
Bayesian network based on theory of probability, complementary statistical regularity between multiple factors in research objective things, with the form of graph theory visual in image express the causal correlation between stochastic variable. It comes the dependence between qualitative representation variable and independence not only by directed acyclic graph (network structure), and can pass through conditional probability distribution (network parameter) and quantitatively portray the variable dependence to its father node.
Defining 1 Bayesian network is two tuple B=<G, θ>. G=<X, A>for directed acyclic graph, each of which nodes X_{i}∈ X represents domain variables, every directed connection arc A_{ij}∈ A represents respective nodes X_{i}And X_{j}Between incidence relation X_{i}→X_{j}; θ={ θ_{i}Represent network conditional probability parameter sets, θ_{i}=P (X_{i}π(X_{i})) represent nodes X_{i}At given its father node π (X_{i}) conditional probability distribution under state.
According to Bayes theorem, chain rule and conditional independence, Bayesian network provides a kind of method that joint probability distribution is decomposed:
Thus reducing the complexity of probabilistic model, provide a great convenience for probability inference.
Node in the Bayesian network of forest fires probabilistic forecasting is divided into two classes: a class is forest fires generation nodes, characterization result; Another kind of is forest fires inducement nodes, characterizes reason. Internodal directed connection arc represents the cause effect relation between forest fires and inducement thereof, and the parameter value of each node characterizes the probability dependency between forest fires and inducement thereof.
BN modeling is carried out based on temporal data and spatial data. It is (meteorological data and population distribution data can only correspond at county level) at county level owing to collecting the thickest spatial granularity of data, therefore the reference of the space of Bayesian network model is county (district); FWI index calculates once every day, therefore the time reference of model is one day. So, what Bayesian network model calculated is the forest fires probability in each county every day. The Bayesian network model structure for forest fires probability forecast is set up as shown in Figure 6 according to expertise [5] [16].
Each node has several value states discrete, mutual exclusion, it means that the continuous number (such as height above sea level, population density etc.) observed must be beforehand with slidingmodel control. Data are carried out discretization by present invention application equifrequent discretization method [17].
The conditional probability table of height above sea level node, ground mulching node and population density node is determined by priori, it may be assumed that the mean sea level in each county, ground mulching and be constant in the population density short time.
The conditional probability table of all the other nodes except forest fires generation node, application maximum Likelihood learns. Maximal possibility estimation, based on traditional statistical analysis thought, passes judgment on the fitting degree of sample and model according to the likelihood degree of sample with parameter. If data D is by sample (D_{1},D_{2},…,D_{m}) composition, then nodes X_{i}Parameter θ_{i}The general type of loglikelihood function be: L (θ_{i} D) and=logP (D  θ_{i})。θ_{i}Maximal possibility estimation, it is simply that make L (θ_{i} D) reach that maximum valueNamely
If nodes X_{i}Total r_{i}Individual value, its father node π (X_{i}) the total q of value_{i}Individual combination, assumes the architectural feature with Bayesian network according to independent same distribution, has:
Wherein m_{ijk}For D meets X_{i}=k and π (X_{i}The quantity of the sample of)=j. NoteThen θ_{i}Maximal possibility estimation be:
For forest fires generation node, his father's node set comprises ground mulching node. Owing to the space reference of model is county, therefore the value of this node does not determine that in real data, but a probability distribution: when county's name is known, the probability of each ground mulching type is the ratio of the type area coverage in this county and this county area. Therefore forest fires node needs to carry out maximal possibility estimation [18] based on " broken power sample ". Assume D_{l}It is an observation sample, X_{lc}For ground mulching node, X_{fire}For forest fires node. Due to X_{lc}There are 13 kinds of value states, therefore by D_{l}It is split as 13 broken power samples:
Wherein x_{i}For X_{lc}Certain value,For sample (D_{l},X_{lc}=x_{i}) weight, meet constraints:Being known in real data, its value is sample D_{l}Ground mulching type x in affiliated county_{i}The ratio of area coverage and this county gross area.In forest fires nodes X_{fire}Parameter learning process in, each observation sample D_{l}All substituted by 13 complete broken power samples. Note D^{t}For whole broken power partial datas, then X_{fire}Method for parameter estimation be:
Wherein m_{fire,jk}For D^{t}In all meet X_{fire}=k and π (X_{fire}The weight sum of the sample of)=j, m_{fire,j}For D^{t}In all meet π (X_{fire}The weight sum of the sample of)=j.
After building Bayesian network model, it is necessary to carry out Bayesian Network Inference and calculate forest fires probability at county level. Posterior probability, maximum a posteriori hypothesis and maximum possible interpretation problems are by the three types of Bayesian Network Inference. Forest fires probability forecast problem mainly calculates Posterior probability distribution, it may be assumed that name node in county's in known network, FWI index node and month node (evidence variable, is designated as X_{E}) value, calculate forest fires generation nodes X_{fire}The Posterior probability distribution of (query interface), carries out the Forecast reasoning from reason to result. Bayesian Network Inference has Accurate Reasoning and approximate resoning two class method, and the Junction tree (JunctionTreeAlgorithm, JTA) [19] in present invention application Accurate Reasoning method calculates forest fires probability at county level.
First the G of bayesian network structure figure shown in Fig. 6 is converted to the data structure of an auxiliary, namely combines tree T, then pass through the initialization definitions on T and information transmission carries out posterior probability calculating:
WhereinJunction tree flow process is as shown in Figure 7.
Probability assignments model
Probability assignments carries out in units of county. According to the forest fires probability at county level that Bayesian network forecasts, the region of 300m × 300m each in this county is carried out probability assignments. The method is mainly on forest fires Probability Basis at county level, it is considered to landform and the ground mulching disturbance degree to forest fires. Mode input is the average forest fires probability in certain county, mean sea level H, gradient S, position, slope T in 300m × 300m region of required calculating, ground mulching L, totally 4 kinds of factors of influence; It is output as the forest fires probability of this fritter, as shown in Figure 8.
First the disturbance degree of abovementioned 4 elite stand cause of fire is calculated. The disturbance degree of the note height above sea level factor is P_{h}, the disturbance degree of slope location factor is P_{t}, the disturbance degree of slope factor is P_{s}, the disturbance degree of the ground mulching factor is P_{l}. According to the forest fires Effects of Factors Du He county forest fires probability in 300m × 300m regions all in whole county, computation rule coefficient, 4 kinds of Effects of Factors degree of each fritter are multiplied, then are multiplied by regularization coefficient, obtain the forest fires probability of this fritter. Computing formula is as follows:
Wherein, NA represents some special value, such as water body, glaciers etc. of ground mulching, and these local forest fires probability are 0, and forest fires probability is that the disturbance degree of every elite stand cause of fire is multiplied by regularization coefficient elsewhere.
Factor of influence disturbance degree adopts statistical method to be calculated. Occurring carrying out statistical analysis at forest fires historical record to every kind of factor of influence, using the value of every kind of factor of influence as independent variable, it is to the disturbance degree of forest fires as dependent variable, and the two is fitted by application normal distribution, and corresponding coefficient value is finely tuned. Add up the sufficient statistic in forest fires historical record of the every kind of factor respectively.
The relation between height above sea level and forest fires disturbance degree is simulated according to normal distribution. In forest fires historical record, forest fires point mean sea level is 2257, and sample variance is 624. Height above sea level is obtained the relation between forest fires disturbance degree and height above sea level is as follows according to normal distribution:
Statistical nature and Korea National forest fire danger class forecast system, P according to existing forest fires historical record_{t}Value compares as follows with position, required fritter actual slope:
The foot of the hill: P_{t}=0.5,
Mountain valley: P_{t}=0.25,
Middle descending: P_{t}=0.1,
Middle upward slope: P_{t}=0.05.
Abovementioned rule obtains according to position, the slope statistics of forest fires scene in forest fires historical record, wherein forest fires have half to occur at the foot of the hill, have 1/4th generations in mountain valley, have 1/10th generations at middle descending, residue forest fires scene is middle upward slope, and the forest fires number of times of middle upward slope is less.
Equally, between disturbance degree and the gradient of slope factor, relation is obtained by normal distribution matching. Computing formula is as follows:
Wherein, S represents the mean inclination in fritter.
Ground mulching is divided into 32 kinds, according to China Meteorological industry standard, each ground mulching type is arranged disturbance degree, wherein, glacier, farmland, water body, city etc. are NA beyond the ground mulching value of forest scope, and remaining every kind cover type assignment is the real number between 0 to 1.
After calculating each factor contributions degree, also need normalized coefficient. Introduce normalisation coefft in order that make the forest fires probability of fritter keep consistent with this county forest fires probability, on average probability basis, county, be multiplied by the disturbance degree that every elite stand cause of fire is sub, and divided by constant so that fritter forest fires probability is maintained between 0 to 1. The computational methods that normalisation coefft is are as follows:
Wherein n is the fritter number of 300m × 300m in county, P_{i}Forest fires probability for each fritter.
The formulation of forest fires risk class
Forest fires output ratio corresponding to forest fires probability carries out the formulation of forest fires risk class, as shown in Figure 9. Make P_{fire}Represent certain forest fires probability determined, N_{fire}Represent and predict the outcome middle forest fires probability more than P_{fire}Forest fires sample number, N_{total}Represent and predict the outcome middle forest fires probability more than P_{fire}Total sample number, then P_{fire}Corresponding forest fires output ratio computational methods are:
First the forest fires probability of all observation samples is calculated. Several threshold values selected, are divided into several interval by the codomain of forest fires probability. For each threshold value, calculate the output ratio of its correspondence, according to the relative size of output ratio, formulate forest fires risk class.
Beneficial effect
Using Yunnan Province's forest fires probability forecast as case, beneficial effects of the present invention is described. The observation sample collected totally 452455, the historical data in corresponding Yunnan Province in 20052015 years every days of each county.
In testing at first group, randomly selecting the forest fires sample (totally 935 example) of 80% and the nonforest fires sample (totally 361028 example) of 80% forms training set, all the other samples of 20% are used as test. Keeping on the basis of stochastical sampling, experiment repeats 5 times. Forest fires probability forecast system and FWI system test the average of the accuracy rate on divided test set, hit rate and false alarm rate and variance in Table 1 in this group. Forest fires probability forecast system shoots straight in FWI system in this group is tested, and relatively FWI system is low for false alarm rate simultaneously, and this illustrates that this system is more more effective than FWI system. Forest fires probability forecast system Average Accuracy on test set is 81.55%, and this illustrates that system energy Accurate Prediction goes out most observation sample. System is more suitable for the prediction that forest fires are occurred, it may be assumed that its mean hit rate is up to 90.08%.
Table 1 forest fires probability forecast system and FWI system predicting the outcome on first group of experiment test collection
Table 2 forest fires probability forecast system and FWI system predicting the outcome on second group of experiment test collection
In order to verify the actual prediction effectiveness of forest fires probability forecast system, in testing at second group, choose Yunnan Province in 2006 December in2013 years on the 1st January all observation samples of 31 days composition training set, the observation sample composition test set on May 14 ,2015 years on the 1st January in 2014. Forest fires probability forecast system and FWI system ROC curve on this group experiment test collection are shown in Fig. 9, it was predicted that the confusion matrix of result is in Table 2. For 175 example forest fires samples in test set, forest fires probability forecast system energy Accurate Prediction goes out 156 examples, and hit rate is 89.14%, and this illustrates that the generation of most forest fires all can by Accurate Prediction.For the 62144 nonforest fires samples of example in test set, 12521 examples therein are predicted as forest fires by system by mistake, and false alarm rate is 20.15%, lower than the 24.08% of FWI system. Meanwhile, the ROC curve comparing result on test set shows, forest fires probability forecast system is better than Canada's weather conditions conducive to wildfires index (FWI) system on precision of prediction, and this system is effective, feasible.
The fine granularity forecast result of forest fires probability forecast system is shown in visual mode. Choose whole nation Forest Fire Danger Weather alarm figure as a comparison. The forecast result on February 26th, 2015 is shown in Figure 10. Left figure is the control methods forecast result in this day, and right figure is the forest fires risk class scattergram of forest fires probability forecast system forecast. Black circles on figure represents actual fire point on same day position, and partial enlarged drawing is shown in Figure 11. The forecast result on March 22nd, 2015 is shown in Figure 12. Left figure is the control methods forecast result in this day, and right figure is the forest fires risk class scattergram of forest fires probability forecast system forecast. Black circles on figure represents actual fire point on same day position, and partial enlarged drawing is shown in Figure 13.
It can be seen that the forecast result of forest fires probability forecast system is more accurate from the case of Yunnan Province's forest fires probability forecast, forecast granularity is thinner, is effective, practical system. This system can provide reliable reference for forest fire prevention, deployment of forces, equipment configuration etc.
System need to possess database management function, data acquisition function, FWI Index for Calculation function, the regular function of data, Bayesian Network Inference function and probability assignments function, as shown in Figure 4.
Database management module comprises the subfunctions such as temporal data management, spatial data management, it is provided that to the process of Back ground Information in data base, it is achieved functions such as the importing of data, Data Format Transform, data edition, amendment, inquiries. Data acquisition module completes system is inputted the collecting work of data, and wherein meteorological historical data needs to write web crawler and is acquired; The raster data such as ground mulching, height above sea level needs Image Via Gis instrument to extract. FWI Index for Calculation module completes Canadian Forest weather conditions conducive to wildfires Index for Calculation function, first by the period of the day from 11 a.m. to 1 p.m temperature, the period of the day from 11 a.m. to 1 p.m relative humidity, the period of the day from 11 a.m. to 1 p.m wind speed, 24 hours precipitation calculates and obtains 5 intermediate indexs, then calculated the FWI index on the same day by 5 intermediate indexs. The regular module of data completes the discretization function to continuous data, to generate the sample data for Bayesian network model training and prediction. Bayesian Network Inference module completes BN modeling and inference function, and first according to expertise, subjectivity determines node and the network structure of Bayesian network; Then, carry out network parameter study based on sample data, set up the Bayesian network model for forest fires probabilistic forecasting; Finally, application Junction tree makes inferences, it was predicted that go out forest fires probability at county level. Probability assignments module completes forest fires probability assignments at county level to 300m × 300m granularity function, first forest fires historical record is carried out statistical analysis, obtain the value weight of each forest fires factor, be then based on spatial data and the forest fires probability of forest fires probability calculation 300m × 300m granularity at county level.
System flow is as shown in Figure 5. First, using data such as meteorology, vegetation, geography, population distribution as input, comprehensive forest fires historical data sets up Bayesian network model, and applies Junction tree and carry out probability inference, it was predicted that go out forest fires probability at county level. Then, historical data is carried out statistical analysis and builds probability assignments model, forest fires probability at county level is adjusted to 300m × 300m granularity.Finally, according to forest fires probability formulate forest fires risk class, and by visualization.
For the operating process understanding the present invention making user become apparent from, now set forth the using method of the present invention. The present invention is all applicable to windows/linux operating system. The programming language of system is Java, and data base is MySQL, runs executable file FPS.jar and can start native system. It is embodied as step as follows:
1) under windows, JRE installs. Its download link is:http://www.oracle.com/technetwork/java/javase/downloads. Directly install after download.
2) under windows, MySQL installs. Its download link is:http://www.mysql.com/downloads/. Decompressing after download, and revise configuration file my.ini, arranging character encoding format is utf8.
3) import system runs desired data, keys in sourceforestfire.sql in order line.
4) loading system runs required third party's Jar bag. Load xxx.jar bag: this bag is put into the lib file under the catalogue of FPS.jar place.
List of references
[1] Shu Lifu, Tian Xiaorui, Kou Xiaojun. fire research summary (I) study hotspot and progress [J]. World Forestry is studied, and 2003,16 (3): 3740.
[2] DeemingJE, BurganRE, CohenJD.TheNationalFireDangerRatingSystem 1978 [R] .U.S.DepartmentofAgriculture.ForestServiceGeneralTechnic alReport:INT39,1977.
[3] Tian Xiaorui, DouglasJM, Zhang Youhui. Assessment of Forest Fire Danger Rating Systems [J]. World Forestry is studied, and 2006,19 (2): 3946.
[4]VanWagnerCE.DevelopmentandstructureoftheCanadianForestFireWeatherIndexSystem[R].CanadianForestService.ForestTechnicalReport:35,1987.
[5]WonMS,LeeSY,LeeMB,etal.DevelopmentandapplicationofaforestfiredangerratingsysteminsouthKorea[J].JournaloftheFacultyofAgriculture,2010,55(2):221229.
[6] Yang Guang, Shu Lifu, Di Xueying, etc. Korea National forest fire danger class forecast system general introduction [J]. World Forestry is studied, and 2013,26 (6): 6468.
[7] Bai Shangbin, Xiaoli Zhang. forest fire prognosis and prediction Review Study [J]. agriculture network information, 2008,2008 (6): 2225.
[8] Zeng Wei, Liu Xiangfeng, Yang Longguang. forest fire danger forecasting progress [J]. China's Forest byproduct and speciality, 2013,2013 (1): 8284.
[9] Yi Haoruo, Ji Ping, Qin Xianlin. whole nation forest fire danger forecasting systematic research and operation [J]. forestscience, 2004,40 (3): 203207.
[10] Zhu Yunhai, Tan Jing, Wu Hongzhi, etc. Jinan City's urban operating mechanism project management and Designing Decisionmaking System and realization [J]. computer engineering and application, 2012,48 (2): 241244.
[11]HeckermanD.AtutorialonLearningwithBayesianNetworks[R].MicrosoftResearch.MicrosoftTechnicalReport:MSRTR9506,1995.
[12]PearlJ.Probabilisticreasoninginintelligentsystems:networksofplausibleinference[C].MorganKaufmann,SanMateo,California,1988.
[13]BlaserL,OhrnbergerM,RiggelsenC,etal.Bayesianbeliefnetworkfortsunamiwarningdecisionsupport[C].Proceedingsofthe10thEuropeanConferenceonSymbolicandQuantitativeApproachestoReasoningwithUncertainty.Verona:Springer,2009:757768.
[14] Li Weiqian, solution is opened a position, Zhang Yongjin, etc. dynamic bayesian network application [J] in hydrologic(al) prognosis. computer engineering and application, 2010,46 (6): 231234.
[15] Wang Fengshan, Zhang Hongjun. the military engineering dam age assessment model based on Bayesian network studies [J]. computer engineering and application, 2011,47 (12): 242245.
[16]DlaminiWM.ABayesianbeliefnetworkanalysisoffactorsinfluencingwildfireoccurrenceinSwaziland[J].EnvironmentalModelling&Software,2010,25(2):199208.
[17]DoughertyJ,KohaviR,SahamiM.Supervisedandunsuperviseddiscretizationofcontinuousfeatures[C].Proceedingsofthe12thInternationalConferenceonMachineLearning.LakeTahoe:MorganKaufmann,1995:194202.
[18] Zhang Lianwen, Guo Haipeng. Bayesian network draws opinion [M]. Beijing: Science Press, 2006.
[19]ShaferGR,ShenoyPP.Probabilitypropagation[J].AnnalsofMathematicsandArtificialIntelligence,1990,2(14):327351。
Claims (5)
1. a finegrained forest fire probability forecast system, is characterized in that, including:
Relational data library module, relational data library module farther includes meteorological measuring storehouse, weather forecast data base, mode input factor data storehouse, model database and forest fires probability database; The meteorological measuring on meteorological measuring library storage same day, inputs the data in factor data storehouse for Renewal model; Weather forecast database purchase is for the weather forecast data of Bayesian Network Inference; Mode input factor data library storage is for building historical data and the spatial data of Bayesian network model and probability assignments model, including meteorological historical data, forest fires historical data, geodata, ground mulching data, population distribution data;The result of calculation of forest fires probability data library storage system, including the forest fires probability at county level obtained by Bayesian Network Inference, by probability assignments model calculated 300m × 300m granularity forest fires probability, and corresponding forest fires risk class; The model file of model data library storage system, the forest fires probability for each area provides computation model;
Database management module, it is provided that to the process of Back ground Information in data base, it is achieved the importing of data, Data Format Transform, data edition, amendment, inquiry;
Data acquisition module completes system is inputted the collecting work of data, and wherein meteorological historical data needs to write web crawler and is acquired; Raster data Image Via Gis instrument including ground mulching, height above sea level extracts;
FWI Index for Calculation module completes forest fire weather Index for Calculation function, first by the period of the day from 11 a.m. to 1 p.m temperature, the period of the day from 11 a.m. to 1 p.m relative humidity, the period of the day from 11 a.m. to 1 p.m wind speed, 24 hours precipitation calculates and obtains 5 intermediate indexs, the FWI index on the same day is calculated again by 5 intermediate indexs, the regular module of data completes the discretization function to continuous data, to generate the sample data for Bayesian network model training and prediction;
Bayesian Network Inference module completes BN modeling and inference function, and first according to expertise, subjectivity determines node and the network structure of Bayesian network; Then, carry out network parameter study based on sample data, set up the Bayesian network model for forest fires probabilistic forecasting; Finally, application Junction tree makes inferences, it was predicted that go out forest fires probability at county level;
Probability assignments module completes forest fires probability assignments at county level to 300m × 300m granularity function, first forest fires historical record is carried out statistical analysis, obtain the value weight of each forest fires factor, it is then based on spatial data and the forest fires probability of forest fires probability calculation 300m × 300m granularity at county level, and is shown by display module.
2. finegrained forest fire probability forecast system as claimed in claim 1, is characterized in that, BN modeling and inference function and the concrete steps doping forest fires probability at county level:
Node in the Bayesian network of forest fires probabilistic forecasting is divided into two classes: a class is forest fires generation nodes, characterization result; Another kind of is forest fires inducement nodes, characterizes reason; Internodal directed connection arc represents the cause effect relation between forest fires and inducement thereof, and the parameter value of each node characterizes the probability dependency between forest fires and inducement thereof;
BN modeling is carried out: the space reference of Bayesian network model is county or district based on temporal data and spatial data; The time reference of model is one day; Each node has several value states discrete, mutual exclusion, it means that the continuous number observed includes height above sea level, the continuous number of population density must be beforehand with slidingmodel control, and data are carried out discretization by application equifrequent discretization method;
The conditional probability table of height above sea level node, ground mulching node and population density node is determined by priori, it may be assumed that the mean sea level in each county, ground mulching and be constant in the population density short time;
The conditional probability table of all the other nodes except forest fires generation node, application maximum Likelihood learns, maximal possibility estimation, based on traditional statistical analysis thought, passes judgment on the fitting degree of sample and model, if data D is by sample (D according to the likelihood degree of sample with parameter_{1},D_{2},…,D_{m}) composition, then nodes X_{i}Parameter θ_{i}The general type of loglikelihood function be: L (θ_{i} D) and=logP (D  θ_{i}), θ_{i}Maximal possibility estimation, it is simply that make L (θ_{i} D) reach that maximum valueNamely
If nodes X_{i}Total r_{i}Individual value, its father node π (X_{i}) the total q of value_{i}Individual combination, assumes the architectural feature with Bayesian network according to independent same distribution, has:
Wherein D_{l}It is an observation sample, m_{ijk}For D meets X_{i}=k and π (X_{i}The quantity of the sample of)=j, noteThen θ_{i}Maximal possibility estimation be:
For forest fires generation node, his father's node set comprises ground mulching node, owing to the space reference of model is county or district, in real data, the value of this node does not determine that, but a probability distribution: when county's name is known, the probability of each ground mulching type is the ratio of the type area coverage in this county and this county area, therefore forest fires node needs to carry out maximal possibility estimation based on " broken power sample ", it is assumed that X_{lc}For ground mulching node, X_{fire}For forest fires node, due to X_{lc}There are 13 kinds of value states, therefore by D_{l}It is split as 13 broken power samples:
Wherein x_{i}For X_{lc}Certain value,For sample (D_{l},X_{lc}=x_{i}) weight, meet constraints: Being known in real data, its value is sample D_{l}Ground mulching type x in affiliated county_{i}The ratio of area coverage and this county gross area, in forest fires nodes X_{fire}Parameter learning process in, each observation sample D_{l}All substituted by 13 complete broken power samples, remember D^{t}For whole broken power partial datas, then X_{fire}Method for parameter estimation be:
Wherein m_{fire,jk}For D^{t}In all meet X_{fire}=k and π (X_{fire}The weight sum of the sample of)=j, m_{fire,j}For D^{t}In all meet π (X_{fire}The weight sum of the sample of)=j;
Build after Bayesian network model, it is necessary to carry out Bayesian Network Inference and calculate forest fires probability at county level and Posterior probability distribution, it may be assumed that name node in county's in known network, FWI index node and month node value, month the value of node be evidence variable, be designated as X_{E}, calculate forest fires generation querying node variable X_{fire}Posterior probability distribution, carry out the Forecast reasoning from reason to result, the Junction tree (JunctionTreeAlgorithm, JTA) in application Accurate Reasoning method calculates forest fires probability at county level.
3. finegrained forest fire probability forecast system as claimed in claim 2, it is characterized in that, application Junction tree makes inferences, dope forest fires probability at county level to comprise the concrete steps that, first bayesian network structure figure G is converted to the data structure of an auxiliary, namely combine tree T, then pass through the initialization definitions on T and information transmission carries out posterior probability calculating:
Wherein $P({X}_{E}={x}_{e})=\underset{{X}_{fire}}{\Σ}P({X}_{fire},{X}_{E}={x}_{e}).$
4. finegrained forest fire probability forecast system as claimed in claim 1, is characterized in that, the concrete steps of probability assignments model and calculating 300m × 300m granularity forest fires probability:
Probability assignments carries out in units of county, according to the forest fires probability at county level that Bayesian network forecasts, the region of 300m × 300m each in this county is carried out probability assignments, main on forest fires Probability Basis at county level, considering landform and the ground mulching disturbance degree to forest fires, mode input is the average forest fires probability in certain countyMean sea level H, gradient S, position, slope T in the required 300m × 300m region calculated, ground mulching L, totally 4 kinds of factors of influence; It is output as the forest fires probability of this fritter;
First the disturbance degree of abovementioned 4 elite stand cause of fire is calculated. The disturbance degree of the note height above sea level factor is P_{h}, the disturbance degree of slope location factor is P_{t}, the disturbance degree of slope factor is P_{s}, the disturbance degree of the ground mulching factor is P_{l}. According to the forest fires Effects of Factors Du He county forest fires probability in 300m × 300m regions all in whole county, computation rule coefficient, 4 kinds of Effects of Factors degree of each fritter are multiplied, then are multiplied by regularization coefficient, obtain the forest fires probability of this fritter. Computing formula is as follows:
Wherein, NA represents some special value, such as water body, glaciers etc. of ground mulching, and these local forest fires probability are 0, and forest fires probability is that the disturbance degree of every elite stand cause of fire is multiplied by regularization coefficient elsewhere;
Factor of influence disturbance degree adopts statistical method to be calculated: occur carrying out statistical analysis at forest fires historical record to every kind of factor of influence, using the value of every kind of factor of influence as independent variable, its to the disturbance degree of forest fires as dependent variable, the two is fitted by application normal distribution, and corresponding coefficient value is finely tuned, add up the sufficient statistic in forest fires historical record of the every kind of factor respectively;
Simulating the relation between height above sea level and forest fires disturbance degree according to normal distribution, in forest fires historical record, forest fires point mean sea level is 2257, and sample variance is 624, obtains height above sea level according to normal distribution the relation between forest fires disturbance degree and height above sea level is as follows:
Statistical nature and Korea National forest fire danger class forecast system, P according to existing forest fires historical record_{t}Value compares as follows with position, required fritter actual slope:
The foot of the hill: P_{t}=0.5,
Mountain valley: P_{t}=0.25,
Middle descending: P_{t}=0.1,
Middle upward slope: P_{t}=0.05.
Abovementioned rule obtains according to position, the slope statistics of forest fires scene in forest fires historical record, wherein forest fires have half to occur at the foot of the hill, have 1/4th generations in mountain valley, have 1/10th generations at middle descending, residue forest fires scene is middle upward slope, and the forest fires number of times of middle upward slope is less;
Equally, between slope factor disturbance degree and the gradient, relation is obtained by normal distribution matching, and computing formula is as follows:
Wherein, S represents the mean inclination in fritter;
Ground mulching is divided into 32 kinds, according to China Meteorological industry standard, each ground mulching type is arranged disturbance degree, wherein, glacier, farmland, water body, city etc. are NA beyond the ground mulching value of forest scope, and remaining every kind cover type assignment is the real number between 0 to 1;
After calculating each factor contributions degree, also need normalized coefficient: computational methods are as follows:
Wherein n is the fritter number of 300m × 300m in county, P_{i}Forest fires probability for each fritter.
5. finegrained forest fire probability forecast system as claimed in claim 1, is characterized in that, formulates the concrete steps of forest fires risk class:
Forest fires output ratio corresponding to forest fires probability carries out the formulation of forest fires risk class, makes P_{fire}Represent certain forest fires probability determined, N_{fire}Represent and predict the outcome middle forest fires probability more than P_{fire}Forest fires sample number, N_{total}Represent and predict the outcome middle forest fires probability more than P_{fire}Total sample number, then P_{fire}Corresponding forest fires output ratio computational methods are:
First the forest fires probability of all observation samples is calculated, several threshold values selected, the codomain of forest fires probability is divided into several intervals, for each threshold value, calculates the output ratio of its correspondence, according to the relative size of output ratio, formulate forest fires risk class.
Priority Applications (1)
Application Number  Priority Date  Filing Date  Title 

CN201610010997.8A CN105678419A (en)  20160105  20160105  Fine gritbased forest fire hazard probability forecasting system 
Applications Claiming Priority (1)
Application Number  Priority Date  Filing Date  Title 

CN201610010997.8A CN105678419A (en)  20160105  20160105  Fine gritbased forest fire hazard probability forecasting system 
Publications (1)
Publication Number  Publication Date 

CN105678419A true CN105678419A (en)  20160615 
Family
ID=56299483
Family Applications (1)
Application Number  Title  Priority Date  Filing Date 

CN201610010997.8A Pending CN105678419A (en)  20160105  20160105  Fine gritbased forest fire hazard probability forecasting system 
Country Status (1)
Country  Link 

CN (1)  CN105678419A (en) 
Cited By (17)
Publication number  Priority date  Publication date  Assignee  Title 

CN106776730A (en) *  20161117  20170531  中国银行股份有限公司  A kind of Chinese compression decompression method and device based on JAVA language 
CN107748735A (en) *  20171010  20180302  广西壮族自治区林业科学研究院  The sample number of forest fuel average value investigation determines method 
CN108399469A (en) *  20180131  20180814  中国人民解放军空军研究院战场环境研究所  A kind of weather phenomenon forecasting procedure based on deep learning and numerical weather forecast 
CN108765836A (en) *  20180522  20181106  深圳源广安智能科技有限公司  A kind of forest fire earlywarning system based on wireless sensor network 
CN108875806A (en) *  20180531  20181123  中南林业科技大学  False forest fires hot spot method for digging based on spacetime data 
CN108898159A (en) *  20180531  20181127  中南林业科技大学  False forest fires hot spot filter method based on DBSCAN algorithm 
CN109118001A (en) *  20180809  20190101  成都天地量子科技有限公司  A kind of mountain fire monitoring method and system based on satellite remote sensing date 
CN109146187A (en) *  20180830  20190104  昆明理工大学  A kind of precipitation predicting method average based on adaptive Bayesian model 
CN109214716A (en) *  20181017  20190115  四川佳联众合企业管理咨询有限公司  Mountain fire risk profile modeling method based on stacking algorithm 
CN109447331A (en) *  20181017  20190308  四川佳联众合企业管理咨询有限公司  Mountain fire Risk Forecast Method based on stacking algorithm 
CN109472396A (en) *  20181017  20190315  四川佳联众合企业管理咨询有限公司  Mountain fire prediction technique based on depth elearning 
CN110276477A (en) *  20190529  20190924  河海大学  A kind of Flood Forecasting Method based on layering Bayesian network and incremental learning 
CN110309961A (en) *  20190620  20191008  京东城市(北京)数字科技有限公司  Fire alarm method and apparatus 
CN110543540A (en) *  20190830  20191206  国网湖南省电力有限公司  method and system for identifying icecoated microterrain based on cyclic correlation 
CN111583610A (en) *  20200430  20200825  深圳市前海用电物联网科技有限公司  Firefighting linkage control method and system of causal model 
CN111784994A (en) *  20200714  20201016  中国民航大学  Fire detection method and device 
CN113298311A (en) *  20210608  20210824  平安国际智慧城市科技股份有限公司  Forest fire danger level prediction method, device, equipment and readable storage medium 

2016
 20160105 CN CN201610010997.8A patent/CN105678419A/en active Pending
Cited By (24)
Publication number  Priority date  Publication date  Assignee  Title 

CN106776730B (en) *  20161117  20200117  中国银行股份有限公司  Chinese compression and decompression method and device based on JAVA language 
CN106776730A (en) *  20161117  20170531  中国银行股份有限公司  A kind of Chinese compression decompression method and device based on JAVA language 
CN107748735A (en) *  20171010  20180302  广西壮族自治区林业科学研究院  The sample number of forest fuel average value investigation determines method 
CN107748735B (en) *  20171010  20210226  广西壮族自治区林业科学研究院  Sample number determination method for forest combustible average value survey 
CN108399469A (en) *  20180131  20180814  中国人民解放军空军研究院战场环境研究所  A kind of weather phenomenon forecasting procedure based on deep learning and numerical weather forecast 
CN108765836A (en) *  20180522  20181106  深圳源广安智能科技有限公司  A kind of forest fire earlywarning system based on wireless sensor network 
CN108875806A (en) *  20180531  20181123  中南林业科技大学  False forest fires hot spot method for digging based on spacetime data 
CN108898159A (en) *  20180531  20181127  中南林业科技大学  False forest fires hot spot filter method based on DBSCAN algorithm 
CN108898159B (en) *  20180531  20220527  中南林业科技大学  False forest fire hot spot filtering method based on DBSCAN algorithm 
CN109118001A (en) *  20180809  20190101  成都天地量子科技有限公司  A kind of mountain fire monitoring method and system based on satellite remote sensing date 
CN109146187A (en) *  20180830  20190104  昆明理工大学  A kind of precipitation predicting method average based on adaptive Bayesian model 
CN109214716B (en) *  20181017  20201027  四川佳联众合企业管理咨询有限公司  Mountain fire risk prediction modeling method based on stacking algorithm 
CN109447331B (en) *  20181017  20201027  四川佳联众合企业管理咨询有限公司  Mountain fire risk prediction method based on stacking algorithm 
CN109472396A (en) *  20181017  20190315  四川佳联众合企业管理咨询有限公司  Mountain fire prediction technique based on depth elearning 
CN109447331A (en) *  20181017  20190308  四川佳联众合企业管理咨询有限公司  Mountain fire Risk Forecast Method based on stacking algorithm 
CN109214716A (en) *  20181017  20190115  四川佳联众合企业管理咨询有限公司  Mountain fire risk profile modeling method based on stacking algorithm 
CN110276477A (en) *  20190529  20190924  河海大学  A kind of Flood Forecasting Method based on layering Bayesian network and incremental learning 
CN110276477B (en) *  20190529  20220816  河海大学  Flood forecasting method based on hierarchical Bayesian network and incremental learning 
CN110309961A (en) *  20190620  20191008  京东城市(北京)数字科技有限公司  Fire alarm method and apparatus 
CN110543540A (en) *  20190830  20191206  国网湖南省电力有限公司  method and system for identifying icecoated microterrain based on cyclic correlation 
CN111583610A (en) *  20200430  20200825  深圳市前海用电物联网科技有限公司  Firefighting linkage control method and system of causal model 
CN111583610B (en) *  20200430  20210716  深圳市前海用电物联网科技有限公司  Firefighting linkage control method and system of causal model 
CN111784994A (en) *  20200714  20201016  中国民航大学  Fire detection method and device 
CN113298311A (en) *  20210608  20210824  平安国际智慧城市科技股份有限公司  Forest fire danger level prediction method, device, equipment and readable storage medium 
Similar Documents
Publication  Publication Date  Title 

CN105678419A (en)  Fine gritbased forest fire hazard probability forecasting system  
Liu et al.  Random forest regression evaluation model of regional flood disaster resilience based on the whale optimization algorithm  
Liang et al.  Assessment of debris flow hazards using a Bayesian Network  
CN106651211A (en)  Differentscale regional flood damage risk evaluation method  
Yue et al.  Projection of wildfire activity in southern California in the midtwentyfirst century  
Dlamini  Analysis of deforestation patterns and drivers in Swaziland using efficient Bayesian multivariate classifiers  
CN106022596A (en)  Urban gas pipeline system danger forecast and evaluation method  
Morales et al.  A stochastic fire spread model for north Patagonia based on fire occurrence maps  
Feng et al.  Analysis on fuzzy risk of landfall typhoon in Zhejiang province of China  
Zhang et al.  Spatial distribution of grassland fires at the regional scale based on the MODIS active fire products  
Hagen et al.  Development and evaluation of flood forecasting models for forecastbased financing using a novel model suitability matrix  
Zhou et al.  An interpretable model for the susceptibility of rainfallinduced shallow landslides based on SHAP and XGBoost  
Monrat et al.  A belief rule based flood risk assessment expert system using real time sensor data streaming  
Jin et al.  Comprehensive Evaluation of Water Ecological Civilization Based on Random Forest Regression Algorithm  
Wang et al.  Assessment of landslide susceptibility and risk factors in China  
Bordbar et al.  Multihazard spatial modeling via ensembles of machine learning and metaheuristic techniques  
Anezakis et al.  Hybrid intelligent modeling of wild fires risk  
Rosti et al.  Comparison of PSH results with historical macroseismic observations at different scales. Part 2: application to SouthEast France  
Giuntoli et al.  Going Beyond the Ensemble Mean: Assessment of Future Floods From Global Multi‐Models  
Zhou et al.  Applying Bayesian belief networks to assess alpine grassland degradation risks: A case study in northwest Sichuan, China  
Rohan et al.  Landslide susceptibility analysis based on citizen reports  
Mangeas et al.  A spatially explicit integrative model for estimating the risk of wildfire impacts in NewCaledonia  
Lai et al.  A method for the hazard assessment of regional geological disasters: a case study of the Panxi area, China  
Luo et al.  Drought grade assessment method based on grey cloud incidence clustering model  
Liu et al.  A similaritybased quantitative model for assessing regional debrisflow hazard 
Legal Events
Date  Code  Title  Description 

C06  Publication  
PB01  Publication  
C10  Entry into substantive examination  
SE01  Entry into force of request for substantive examination  
WD01  Invention patent application deemed withdrawn after publication 
Application publication date: 20160615 

WD01  Invention patent application deemed withdrawn after publication 