Wildfire Danger Assessment And Early Warning From Multi-source Spatio-temporal Big Data

Global warming and frequent forest and grassland wildfires have caused great harm to regional and even global ecological security and the safety of people’s lives and property.Therefore,accurate assessment and early warning of wildfire danger,to nip wildfire in the bud,is particularly important and urgent.In recent years,the rise of big data mining model and multi-source spatio-temporal big data has brought new opportunities for wildfire danger assessment and early warning,and derived wildfire danger assessment and early warning research based on multiple mining models and multi-category spatio-temporal big data.These studies,however,are usually consider fuel information deficiencies and ignore key factor induced dynamic characteristic information,the data error and model error to unknown problems such as the influence of mining results,lead to the result wildfire danger assessment and early warning is still there is a big uncertainty,especially in the small area scale of uncertainty is stronger,Operational application of severe impact wildfire danger assessment and early warning.In view of this,thesis through the introduction of remote sensing technology for fuel information,consider wildfire induced dynamic factors of multidimensional feature information,build the mining model and the coupling model of intelligent optimization algorithm research,make up for the wildfire danger assessment and early warning of the theory and the actual defects,improve the credibility of wildfire danger early warning and assessment.The specific research contents are as follows:（1）Aimed at the present situation of wildfire danger assessment for fuel to consider the problem of insufficient information,thesis firstly according to fire triangle model constructed in 2002-2020 China regional wildfires historical space-time dataset,and according to the north and south climate specificity and cobain climate zoning study area can be divided into four areas: southern grassland area,southern forest area,northern grassland area and northern forest area.In addition to the basic information of each factor,the dataset also considered the time characteristic changes of Fuel Moisture Content（FMC）and meteorological variables within 16 days before the wildfire.Random Forests（RF）model was used to model and assess the wildfire danger in China,and the impact of fuel information on the accuracy and effectiveness of the assessment was explored.The results show that when the fuel information is considered in the model,The Area Under Curve（AUC）values of Receiver Operating Characteristic（ROC）curves of the four regions were 0.95,0.94,0.99 and 0.98,respectively.When fuel information was not considered in the model,the AUC value decreased to 0.91,0.87,0.94 and 0.89,which directly verified that fuel information played an irreplaceable role in the model.From the perspective of Fire Density（FD）index of the model,the effect of removing fuel information is worse and the important role of fuel information in wildfire danger assessment in China is clearly revealed.（2）In view of the lack of multi-dimensional characteristic information of dynamic factors in mainstream wildfire danger assessment,in thesis,Ts Fresh（Time Series Featu Re Extraction on Basis of Scalable Hypothesis Tests）,the current mainstream Time dimension feature enhancement tool,was firstly used to extracte the time-dimension characteristics of FMC,rainfall,wind speed,relative humidity and air temperature,and constructed wildfire dataset for the western Sichuan region.Combined with XGBoost model,the wildfire danger in the western Sichuan region was modeled and evaluated.Firstly,the spatial and temporal distribution of wildfire danger in western Sichuan showed that the model without meteorological and fuel factors had a certain proportion of false alarms in both fire season and non-fire season.Secondly,the spatial and temporal distribution map of wildfire danger in Lushan region of Xichang in 2020 and the time sequence diagram of average danger from January to March also show that the model with the time dimension features of meteorological and fuel factors is more accurate and more consistent with the change of real wildfire danger.Finally,from the comparison of AUC values of ROC curve and Precision Recall（PR）curve,it can be seen that the AUC of ROC curve and PR curve of the model with time dimension feature is higher than that of the model without time dimension feature,and other indicators are as follows: FD,commission errors and omission errors can be analyzed to show that the model with time dimension feature has better effect.This also proves that the lack of dynamic characteristic information of key triggers cannot achieve accurate wildfire danger assessment.（3）In view of the current vacancy of wildfire danger assessment on the impact of different non-fire sampling methods on the model,thesis first based on the current mainstream four non-fire sampling methods: Wildfire danger assessment and modeling in Yunnan were carried out by randomly selecting non-burning points from unburned areas,non-burning points with only one fire event in history,non-burning points from unvegetated areas or wasteland and non-burning points based on semi-variational function statistics,and the influences of four sampling methods on the model were explored.Firstly,the spatial and temporal distribution map of wildfire danger in Yunnan shows that the effect of selecting non-burning points in random unburned areas is better than other sampling methods.Secondly,the five evaluation indexes of the model: AUC of ROC curve,AUC of PR curve,MAE,Brier score（BS）and logarithmic score（LS）can be analyzed to find that the accuracy of randomly selecting non-fire points from nonvegetated areas or wastland is the highest.But from the perspective of FD for the analysis of the effect of four kind of sampling methods,random never burn area to select the fire effect of this approach is not the best,the reason is that this is the case: selected randomly from no vegetation area or wasteland not point the sampling way ascend in the sample data set differences in disguised forms,and false ascension model effect thereby.（4）According to the research results of the first three aspects in combustible information,time dimension characteristics and non fire point sampling methods,further combining Adaptive Neuro-Fuzzy Inference System（ANFIS）and Particle Swarm Optimization（PSO）,To study the wildfire danger early warning in western Sichuan.According to the model,the wildfire danger index products of the past five years were produced,and the wildfire danger warning model was constructed through the convolutional short and long time memory network to realize the wildfire danger warning in the western Sichuan region.The feasibility of the wildfire danger warning was proved by comparing the effect of the warning with the actual wildfire danger.