Study On Temporal And Spatial Variation Of Forest Fire And Fire Risk Prediction In Large-scale Areas

Forest can improve the ecological environment,but also has economic and other aspects of the value.However,in recent years,the frequent occurrence of forest fires has continuously destroyed the ecological environment,threatened human life and safety,and suffered economic losses.Therefore,how to accurately and efficiently forecast and prevent forest fire has become a key concern and an urgent problem to be solved.But at present,most of the research on this problem is focused on specific or small areas.According to this situation,this research mainly focuses on large scale,that is,national scale research.Based on the historical forest fire data of China from 2003 to 2016(satellite monitoring fire point data and forest fire area data),this study analyzed the change rule of forest fire during the 14 years from two aspects of time and space.In this study,five types of data including meteorology,topography,vegetation,infrastructure and social economy were obtained,and 20 influencing factors related to the occurrence of forest fires were finally obtained.The main driving factors of forest fires were screened out through feature selection and analyzed.In this research,four machine learning algorithms are selected,namely radial basis neural network,support vector machine,artificial neural network and random forest.These four algorithms are used to construct a large-scale forest fire prediction model,and the optimal model is selected through model performance evaluation and comparison.The optimal model is used to predict the probability of forest fires,and finally a spatial zoning map of China’s forest fire risk rating and a zoning map of China’s seasonal fire risk rating are made.A large-scale forest fire area prediction model was constructed by using feedforward neural network and support vector machine regression algorithm.The prediction ability of the prediction model was evaluated by the precision evaluation index,and the optimal forest fire area prediction model was selected.This study can provide some data basis and scientific reference for the study of forest fire risk prediction in China,so as to improve the management level of forest fire prevention more effectively.The main research contents are as follows:(1)The occurrence regularity of forest fire in China.From 2003 to 2016,the number of fires gradually decreased with the passage of time,and the area burned by fire decreased year by year.Spring and winter are more frequent than summer and autumn.March is the month with high frequency of forest fires,and the months with large areas of forest disaster mainly concentrated in the first five months of the year.Spatially:The areas where forest fires are prone to occur are mainly concentrated in the central south region and southwest regions of China,and the Northeast region is the most severely affected region.Provincially,Yunnan province is the most fire-prone city,while Heilongjiang province has been hit the hardest.From the perspective of topography,the number of forest fires and the size of the affected area increased with decreasing elevation.From the perspective of vegetation type,mixed coniferous and broad forests are most likely to burn.Severe forest fires occur more in the south and north and less in the center.(2)Development of forest fire prediction model.In this study,13 major forest fire drivers were selected,including longitude,latitude,average surface temperature,daily maximum surface temperature,cumulative precipitation,average relative humidity,sunshine duration,average temperature,daily maximum temperature,altitude,population,GDP and NDVI.A large-scale regional forest fire prediction model based on four machine learning algorithms(artificial neural network,random forest,radial basis function neural network,and support vector machine)was established.The accuracy of the four models is more than 75%,and the area under the curve is more than 0.80.Therefore,they are all suitable for large-scale forest fire prediction.The prediction accuracy of the model is 89.2%for random forest,84.3%for support vector machine,83.0%for artificial neural network,and 75.8%for radial basis neural network.According to the comprehensive analysis,the random forest model is the best among the four models.By predicting the probability of forest fire occurrence,a spatial zoning map of forest fire risk levels was produced.The results showed that the forest fire levelⅠfire risk zones were mainly distributed in Heilongjiang province,northern Inner Mongolia and the southern region of China,etc.,the seasonal forest fire risk grade was in the order of spring,winter,summer and autumn.Therefore,these are the key fire prevention objects,need to strengthen forest fire monitoring and prevention.(3)In this study,the influencing factors of forest fire overfire area were determined:average relative humidity,average wind speed of the day,average temperature of the day,elevation,sunshine hours,and slope,and two prediction models of overfire area,BP neural network and support vector regression,were constructed.The accuracy verification results show that the two models are both applicable to the prediction of the actual large-scale forest fire area,and the fitting R~2 of the models is between 0.40 and0.70.RMSE of BPNN model was 10.29,RMSE of SVR model was 11.78.Therefore,the BP neural network model has better ability to predict the area of forest fire and is more suitable for the prediction of the area of forest fire in large-scale.