Analysis And Research Of Wildfire Detection Model Based On Machine Learning

Forest fires destroy forest resources and cause serious losses to the ecological environment,life and property.In fire rescue,predicting the final burning area based on the early stages of a fire can effectively guide fire rescue and reduce losses caused by wildfires.Forest fires are affected by a variety of factors,including meteorological factors,topographic factors,vegetation factors,and human factors.Most of the existing studies that consider the impact of vegetation factors on forest fires use the normalized difference vegetation index,followed by most studies that use a single machine learning algorithm model to predict the area of forest fires.In order to avoid the shortcomings of NDVI such as sensitivity to soil reflectance and atmospheric disturbances and high noise,solve the problems of weak anti-interference ability and poor generalization ability of a single machine learning model,and more accurately predict the area of forest fires,this paper proposes using a new vegetation index to predict the area of forest fires,and constructs a model based on integrated learning algorithms to predict the area of forest fires.The main research content and results of this article are as follows:(1)Based on the MCD64A1 6th edition combustion zone data product collected by the MODIS sensor(500 resolution),historical fire spots from 2011 to 2016 in eastern Australia’s Queensland,New South Wales,and Victoria were collected.At the same time,various influencing factors such as remote sensing images,meteorological information,and vegetation index in the study area were manually retrieved.The impact of human factors is introduced and quantified as the distance from the road,and the dataset in this article is constructed together with meteorological data,terrain data,and vegetation data.(2)This paper analyzes in detail the spatiotemporal distribution characteristics of forest fires in Queensland,New South Wales,and Victoria from 2011 to 2016,as well as the spatiotemporal distribution characteristics of fire footprint area.It includes annual change analysis,monthly change analysis,quarterly change analysis,and spatial characteristic distribution of forest fire and fire footprint area.(3)Most previous studies have used the Normalized Difference Vegetation Index as an important indicator for predicting forest fire area,but NDVI is very sensitive to soil reflectance,with high data noise,and has interference in predicting forest fire area.Therefore,this paper uses a new forest fire impact factor,the Two-band Enhanced Vegetation Index,to predict the scale of wildfires.Compared to NDVI,this index can effectively avoid the impact of soil reflectance,canopy background interference,and atmospheric disturbances,thereby improving the accuracy of forest fire prediction models.The results showed that the R~2 of using EVI2 to predict forest fire scale increased by 6.05%compared to NDVI,and both MAE and MSE decreased,indicating that using EVI2 can predict forest fire scale more accurately,and the performance of the model was significantly improved.(4)Most of the existing research is based on a single machine learning prediction algorithm,which has low anti-interference ability and weak robustness.Therefore,a Stacking-XRSK model based on stack generalization ensemble learning is constructed to predict the area of forest fires.This model uses the idea of Stacking algorithm to improve the prediction performance of the model by integrating multiple single models.The Stacking XRSK model uses four models with significant structural differences,namely,extreme gradient lifting tree(XGBoost),random forest(RF),support vector regression(SVR),and K-nearest neighbor(KNN),as the first level learners and uses a5-fold cross validation method to train the model.The logical regression(LR)model is used as a second level learner to alleviate the overfitting phenomenon in the learning process of the Stacking model.Stacking-XRSK model can give full play to the single base model,improve the generalization ability of the model,and overcome the defect of low accuracy of the single model.The results show that,compared with the single model,the R~2 predicted by Stacking XRSK model is the highest,and MAE,MSE and AOC are the lowest.