Fire Point Detection Based On Deep Learning

Forest fires pose a serious threat to the global ecological environment and human security.They even cause irreversible damage to the diversity of biological systems.In addition,when lots of materials are burned,the smoke containing chemical substances disturb the atmospheric balance and cause serious air pollution.With the rapid development of global satellite remote sensing in the past decades,more and more researchers can detect and track the fire scene through remote sensing image processing.Their researches provide timely fire information and understand the situation of small fire points such as straw burning.Therefore,fire point detection based on remote sensing images is very meaningful.The traditional fire detection methods are mostly based on threshold judgment using remote sensing images,which has poor adaptability to the environment.The selection of threshold can only meet existing images.And the methods cannot consider the rate of missed detection and false detection at the same time.Considering the disadvantages of traditional methods,this thesis designs a classifier combining threshold processing and deep learning for fire point detection.The work of this thesis is as follows:Preliminary screening of fire based on the threshold and the brightness temperature relation model.The main purpose of this part is to determine the candidate points of fire points and quickly remove most non-fire point pixels.Firstly,the high temperature center of remote sensing image is detected by short wave infrared and near-infrared bands.Secondly,the common high-temperature points are searched for the neighborhood around the high-temperature points.And the pixels determined as high-temperature points are clustered to record the coordinates and areas of the high-temperature center points.Finally,the model based on the relationship between area and brightness temperature is established to complete the preliminary fire detection using the appropriate temperature threshold for different fire area.The block images including different bands combination are outputted from the positions of candidate fire points and part of them used as the training data for accurate discrimination classifier.Fire-image classifier based on deep learning.In order to accurately discriminate the preliminary screening results,the block images of candidate fire points is used as the input of the training network which has been improved based on VGG16 network.And we finally got the classification model corresponding to the combined images of each band.A band combination classifier based on VGG16 network model is proposed.To make full use of the advantages of each band,the network models trained from different band combination images are combined by weighted voting to obtain an accurate discriminating classifier.The block image obtained by the preliminary screening is used as the input of classification model and discriminated comprehensively accurately using the model.This thesis proposes a method combing the traditional method and deep learning for fire detection.The preliminary screening of this method belongs to the under-sampling technology in unbalanced data processing.Most of the non-fire data in the whole image are eliminated,which solves the problem of large gap between positive and negative samples in accurate discrimination.In the test of the whole image,the preliminary screening can delete most invalid non-fire pixels and keep some fire points.Using the accurate discrimination model,the average accuracy rate can reach 0.96,the accuracy rate can reach 0.83,and the missing detection rate can reach 0.07.Compared with the traditional learning algorithm,the algorithm in this paper improves the accuracy and robustness of the algorithm.