Research On Wild Fire Risk Prediction Based On Deep Learning

Forest is a valuable resource given by nature to mankind.They can not only regulate the climate,but also maintain a stable ecosystem.However,in recent years,a large number of wild fires have caused forest fires,especially in Yunnan Province.Yunnan Province has a large proportion of mountain types and high forest coverage.Affected by human factors and weather factors,forest fires occur frequently.Forest fires have caused considerable losses to the ecology,society and economy of local people.In order to reduce these losses and reduce forest fires,modeling and predicting the occurrence of forest fires can support forest fire prevention and management losses,and has very important practical significance.At present,there are few researches on wild fire prediction in Yunnan Province Based on deep learning.Most researchers use statistical analysis methods and traditional machine learning methods to predict wild fire risk.Based on the temporal and spatial regularity of wild fire in Yunnan Province,this thesis uses deep learning to model wild fire prediction,and proposed two deep learning models to effectively predict wild fire risk in Yunnan Province.（1）Summarize and collect the wild fire impact factors in Yunnan Province,eliminate the wild fire impact factors with high correlation,and analyze the importance of wild fire impact factors.This thesis collects the data sources of wild fires and wild fire impact factors in Yunnan Province from 2018 to 2019,including the distance from roads（Road）,distance from river（River）,distance from villages（Villages）,altitude（DEM）,slope（Slope）,slope direction（Aspect）,land surface temperature（LST）,normalized difference infrared index7（NDII7）,normalized difference vegetation index（NDVI）,rainfall（Rain）,average relative humidity（Hum）,maximum temperature（Maxtem）,minimum temperature（Mintem）,average temperature（Avetem）,maximum gust wind speed（Wind）,maximum gust wind direction（Windasp）,and landuse（Landuse）.The minimum temperature and average temperature with high correlation coefficient are eliminated by Pearson coefficient analysis,and there is no collinearity between the eliminated factors by multiple collinearity test.Secondly,after eliminating the redundant factors,the importance of the remaining 15 wild fire impact factors are analyzed by information gain rate.Among them,the rainfall has the greatest impact on wild fire in Yunnan Province and the land type with the least impact.（2）According to the law that different wild fire influence factors have different weights for causing wild fire,an Attconvlstm（Attention Convolutional Long Short-Term Memory）model combined with channel attention is proposed.Wild fire has regularity in time and space,therefore this thesis uses Convlstm,which can extract spatio-temporal characteristics,as the backbone network for predicting wild fire to train wild fire data set.The verification set is verified by False Alarm Rate,Specificity,RMSE（Root Mean Square Error）,Kappa,Recall,F1,Precision,AUC（Area Under Curve）.The experimental results show that the proposed Att Convlstm model can effectively improve the accuracy of wild fire prediction and is more suitable for wild fire prediction.On the other hand,the wild fire risk map is generated by Att Convlstm with the data in the test set,and the wild fire risk map is verified with the actual fire points.The experimental results show that the proposed model can make more fire points fall into the area with high risk level,and the channel attention can effectively predict the wild fire.（3）According to the space-time law of wild fire,a space-time cascade wild fire prediction network is proposed.3D Convolution Neural Network（3DCNN）is inferior to Convlstm in extracting temporal features,while Convlstm is inferior to 3DCNN in extracting spatial features.A spatiotemporal cascade wild fire prediction network of 3DCNN and Convlstm is proposed.The network can fully extract the temporal and spatial law of wild fire at the same time.3DCNN,Convlstm and 3DCNN-Convlstm are compared.The experimental results show that the proposed model is optimal in False Alarm Rate,Specificity,RMSE,Kappa,Precision and AUC.On the other hand,through the proposed model,the wild fire risk map is generated with the data in the test set,and verified with the actual fire points.The experimental results show that the proposed model can make more fire points fall into the area with high risk level,and can effectively predict the wild fire.