Wildfire Risk Assessment Based On Ensemble Learning

In recent years,the increasingly severe wildfires have posed a great threat to the earth’s ecology and transmission lines.Wildfire risk assessment and early warning have become an important research topic in the power sector and forestry sector.The occurrence of mountain fire disasters is affected by a variety of environmental factors,including human factors,meteorological factors,vegetation factors,etc.,and the amount of data is huge.In order to reflect the advantages of ensemble learning and make fire point prediction more efficient and accurate,this thesis uses ensemble learning to evaluate wildfire risk.The main research contents and results are:（1）Through MODIS,NPP,and Landsat8 satellite data,the fire points in Yunnan Province from 2016 to 2020 were collected.At the same time,non-fire point samples near the fire point were extracted based on the arcgis ring buffer.In the research process,in order to quantify the human factors,the distance from the residential area and the distance from the road are introduced,and then they are correlated with meteorological data,vegetation data,and terrain data to construct the dataset used in this thesis.（2）The thesis analyzes in detail the temporal and spatial distribution of wildfires and the distribution of land cover in Yunnan Province.The temporal distribution mainly includes the analysis on the annual scale and the monthly scale,and the spatial analysis of the fire point nuclei density.The new surface coverage distribution in 2020 was introduced,with various types.On this basis,the distribution characteristics of fire points were analyzed,and then the relationship between wildfire disasters and various environmental factors was analyzed.（3）KNN filling and random forest filling are compared in the filling of missing values in the data set.Random forest filling is more suitable for multiple types of feature variables in the wildfire risk assessment data set.The random forest algorithm is improved for feature selection,making it more suitable for imbalanced wildfire risk assessment datasets.（4）A wildfire risk assessment model based on LightGBM ensemble learning is proposed.Using the idea of Stacking algorithm,Logistic Regression（LR）,Random Forest（RF）,KNearest Neighbors（KNN）,Support Vector Machine（SVM）and Naive Bay The Yes（NBM）model is used as the first-layer learner,and the LightGBM model is used as the second-layer learner,which overcomes the defects of low accuracy and slow training speed of a single model,and improves the classification effect and stability of the wildfire risk assessment model.The results show that,compared with the single LightGBM model with the best prediction result,the wildfire risk assessment model based on LightGBM ensemble learning improves the accuracy by 9.52%,the precision by 8.15%,and the recall by 10.35%.%,the F1 value has increased by 9.49%,and the AUC value has increased by 8.27%,and the running time of the wildfire risk assessment model based on LightGBM ensemble learning is only 10.68s,which is greatly reduced.（5）In order to realize short-term early warning of wildfire risk,a new time series prediction method is proposed to fill in the missing values of remote sensing data.Combining the populated remote sensing data and GFS meteorological data,short-term early warning of wildfire risk is realized.The final case analysis shows that the model can be effectively applied to wildfire risk assessment in the study area.