Research On Forest Fire Detection Using Unmanned Aerial Vehicles Based On Convolutional Neural Network

With the common sense that forest is a precious natural resource while maintaining ecological balance,a forest fire can bring about enormous ecological and property loss.If forest fire being identified at the beginning stage,the loss can be furthest reduced.At present,the researches on forest fire detection based on computer vision and unmanned aerial vehicles(UAVs)have already been developed worldwide in recent years,but there are still some defects,such as a high false positive rate and long detection time.In the research of forest fire identification technology,we can choose to use visual cameras with UAVs,which are flexible in operation,cheap and efficient.This paper takes forest fire detection based on UAVs as the research object.Relevant research work has been carried out as follows:(1)Smoke detection of forest fires was performed by LBP feature extraction combined with S VM classifier.Observing the image data of forest fires that have occurred,it is found that in the forests with lush vegetation,the proportion of flames in the fire images captured is extremely small,so the study of smoke detection is first considered.In this regard,the local binary pattern(LBP)feature extraction and support vector machine(SVM)classifier are used for smoke detection,so as to make preliminary discrimination of forest fire.(2)Two sets of models were built to implement CNN-based smoke and flame detection.For the purpose of identifying it in the early stage of the fire exactly,according to the convolutional neural network(CNN),it has the characteristics of local receptive domain,weight sharing and pooling,this thesis proposes a method for detecting forest fires in UAV based on CNN.In order to simultaneously detect smoke and flame and improve accuracy,CNN-9 and CNN-17 models were established.(3)The effect of image preprocessing on convolutional neural network detection performance was considered.Before the image is input into the CNN network,in order to enhance the detectability of the related information and to simplify the data to the greatest extent,the image preprocessing related operations such as histogram equalization,smooth low pass filtering,and the like are performed.The effectiveness of the proposed method is verified by the experiments on real forest fire images,i.e.histogram matching and neighbourhood averaging method can improve the accuracy of forest fire detection.(4)With SqueezeNet as the backbone network,the flame region was segmented after the model was optimized.In the testing experiment,the deep learning model was trained by using the actual forest fire data set,and carried out some single-image forest fire detection experiments.The experimental results show that the proposed network realizes the accurate location of the flame region in the image,thus effectively detecting early forest fires.