Research On Global Wildfire Risk Spatiotemporal Mining And Prediction Methods

Wildfires are frequent in the context of global warming.Previous studies have pointed out that wildfires are strongly affected by three factors: fuel,weather,and topography.To improve the success rate of wildfire prevention and extinguishment plan,it has become a trend to spatiotemporally mine the changes of these factors and incorporate them into the regional wildfire prediction model.However,there are three main challenges in the research of wildfire prediction at large scale:(1)The inability to spatio-temporal continuously retrieve the Live Fuel Moisture Content(LFMC)at large scale and the lack of research on its relationship with wildfire,the database of wildfire induced factors is incomplete;(2)The lack of spatio-temporal mining methods for wildfire risk at large scale makes the accuracy of wildfire risk assessment need to be improved;(3)Wildfire risk prediction methods at pixel level need to be explored.Therefore,by solving the above problems,this study established a methodological system of spatiotemporal mining and prediction of wildfire risk.This study provided the new solutions and methods for the development of the wildfire prediction system.Moreover,it can be used to promote local fire management planning and resource allocation,and to contribute to wildfire risk warning,suppression,and response,as well as increasing people's awareness of life and property safety.Firstly,LFMC in wildfire-prone areas over southwestern China was spatio-temporal continuously retrieved based on multi-source remote sensing data and Radiative Transfer Model(RTM).The relationship between LFMC and wildfires was analyzed to verify LFMC's importance as a driver of wildfire occurrence;Secondly,a complete database of historical wildfire events and induced factors was constructed by integrating fuel-induced factors,weather-induced factors,topography-induced factors,and wildfire reference information factors.The historical wildfire risk assessment in southwestern China was completed based on a deep learning model;Finally,prediction and analysis of wildfire risk in forest fire-prone regions around the world are achieved based on the long time series wildfire risk mining at the pixel level.The main work and results can be summarized as:(1)Based on the RTM,Moderate-resolution Imaging Spectroradiometer(MODIS)reflectance product MCD43A4 and land cover product MCD12Q1,LFMC in wildfire-prone areas over southwest China was retrieved,and the wildfire events are extracted from MODIS burned area product MCD64A1.The relationship between LFMC and the occurrence of forest wildfires and grassland wildfires in two climate zone(Cwa and Cwb of Koppen climate classification)is analyzed.It pointed out that LFMC has a short-term threshold effect and a long-term control effect on wildfire occurrence.(2)By integrating fuel-induced factors(the Leaf Area Index(LAI)from MODIS MCD15A2 H product,fuel types based on MCD12Q1 and LFMC based on remote sensing data and RMT retrieval),weather-induced factors(relative humidity,precipitation,temperature,wind speed extracted and calculated by ERA-Interim meteorological reanalysis data),topography-induced factors(elevation,slope,and aspect extracted and calculated from GMTED2010 data)and wildfire reference information factors(burn year,burn date,latitude,and longitude extracted by MODIS MCD64A1)from 2000 to 2018 to build a complete database of historical wildfire events and wildfire induced factors.(3)The space-time buffer radius of the wildfire pixel was determined based on the semi-variogram and multi-temporal remote sensing information.Thus,the non-wildfire points and their relevant induced factors were extracted to construct the control database.70% of data is the training data,and the rest is the verification data.The AUC value and Accuracy are used as assessment indexes.A comparative study was also conducted using the Logistic Regression model as a traditional method.The results indicate that the constructed deep learning model has good assessment performance for wildfire risk and performs better than the Logistic Regression model.(4)Based on the global Fire Danger Index(FDI)product and the time series ARIMA model,the pixel-level simulation of forest wildfire risk over the global five forest wildfire-prone regions(Southwest China,Northern Australia,Southern Europe,Central Africa,and the US West Coast)in 2000-2018 was achieved.Moreover,the future trend prediction and analysis of forest wildfire risk in 2019-2020 over these areas were also achieved,and the validation of prediction was achieved using the actual forest wildfire data in 2019.