基于卷积神经网络的林火烟雾检测  

Forest fire smoke detection based on convolutional neural network

作　　者：徐海文 张贵[2] 谭三清[2] 肖化顺[2] 杨志高 文东新[2] 吴鑫[2] XU Haiwen;ZHANG Gui;TAN Sanqing;XIAO Huashun;YANG Zhigao;WEN Dongxin;WU Xin(The Center of Monitoring,Dispatching and Evaluation of Forest and Grassland Fire Prevention in Hunan Province,Changsha 410004,Hunan,China;College of Forestry,Central South University of Forestry&Technology,Changsha 410004,Hunan,China)

机构地区：[1]湖南省森林草原防火监测调度评估中心,湖南长沙410004 [2]中南林业科技大学林学院,湖南长沙410004

出　　处：《中南林业科技大学学报》2023年第7期23-31,64,共10页Journal of Central South University of Forestry & Technology

基　　金：国家自然科学基金项目(32271879);湖南省科技创新平台与人才计划项目(2017TP1022)。

摘　　要：【目的】随着卫星遥感技术的蓬勃发展,卫星遥感已成为林火监测的重要手段。林火发生初期,由于燃烧温度不高致使卫星红外波段接收不到足以成像的能量辐射。林火发生时会首先产生烟雾,采用深度学习方法利用气象卫星影像进行林火烟雾检测,相较于利用卫星红外通道监测林火而言可更早地发现林火。【方法】以高时间分辨率国产静止气象卫星FY-4A数据为基础,采集研究区内1500张林火烟雾图像和1500张云图像作为数据集,以4︰1的比例划分训练集与验证集并进行数据预处理,采用卷积神经网络AlexNet、MobileNet、ResNet及Inception-ResNet(IRNet)结构对数据集进行实验分析,采用准确率、精确率、召回率和Kappa系数评价模型的总体效果,选取最优结果建立基于卷积神经网络的林火烟雾检测模型。【结果】利用准确率、精确率、召回率及Kappa系数定量评价各模型的总体效果,得出AlexNet模型的准确率达89.3%,精确率达100%,召回率达78.7%,Kappa系数为78.7%;MobileNet模型的准确率达98.2%,精确率达99.7%,召回率达96.7%,Kappa系数为96.3%;ResNet模型的准确率达98.0%,精确率达100%,召回率达96.0%,Kappa系数为96.0%;IRNet模型的准确率达99.8%,精确率达100%,召回率达99.7%,Kappa系数为99.7%。IRNet模型的总体效果高于AlexNet模型、MobileNet模型与ResNet模型,选取IRNet为林火烟雾检测的最优模型。【结论】利用高时效性FY-4A静止气象卫星遥感数据,采用IRNet模型进行林火烟雾检测的总体效果最好,能有效地减少卫星监测时的林火漏判和迟判现象,提高对森林火灾的早期监测预警能力。【Objective】With the vigorous development of satellite remote sensing technology,satellite remote sensing has become an important means of forest fire monitoring.In the initial stage of a forest fire,the infrared band of satellites cannot receive enough energy radiation for imaging due to the low combustion temperature.Smoke is first generated when a forest fire occurs.Using deep learning methods and meteorological satellite images to detect forest fire smoke can detect forest fires earlier than using satellite infrared channels.【Method】In this study,data from high temporal resolution domestic geostationary meteorological satellite FY-4A was used to collect 1500 images of forest fire smoke and 1500 cloud images in the study area,which were used as the dataset.The dataset was divided into the training set and validation set in a 4∶1 ratio and preprocessed.Convolutional neural network(CNN)models including AlexNet,MobileNet,ResNet and Inception-ResNet(IRNet)were used to conduct experiments on the dataset.The overall performance of the models was evaluated using Accuracy,Precision,Recall and Kappa coefficient,and the best model was selected to establish a forest fire smoke detection model based on CNN.【Result】The overall effectiveness of the models was evaluated quantitatively using Accuracy,Precision,Recall and Kappa coefficients.The Accuracy of the AlexNet model was 89.3%,the Precision was 100%,the Recall was 78.7%,and the Kappa coefficient was 78.7%.The MobileNet model achieved 98.2%Accuracy,99.7%Precision,96.7%Recall and 96.3%Kappa coefficient.The ResNet model achieved 98.0%Accuracy,100%Precision,96.0%Recall and 96.3%Kappa coefficient,and the IRNet model achieves 99.8%Accuracy,100%Precision,99.7%Recall and 99.7%Kappa coefficient.The results showed that the overall performance of the IRNet model was superior to that of the AlexNet,MobileNet,and ResNet models,which was selected as the optimal model for forest fire smoke detection.【Conclusion】Using high-temporal FY-4A geostationary meteorological sa

关 键 词：卷积神经网络 林火烟雾 图像识别 IRNet模型 林火监测 FY-4A 

分 类 号：S762.3[农业科学—森林保护学]