基于无人机遥感和深度学习的松材线虫病疫木自动提取方法研究  

Automatic Extraction Method of Pine woods Nematode Infected Wood based on Unmanned Aerial Vehicle Remote Sensing and Deep Learning

作　　者：时启龙 黄石明 张明霞 肖欣 喻俊 徐健[1,3] 邱琳 SHI Qi-long;HUANG Shi-ming;ZHANG Ming-xia;XIAO Xin;YU Jun;XU Jian;QIU Lin(Key Laboratory of Poyang Lake Wetland and Watershed Research,Ministry of Education,Jiangxi Normal University,Nanchang Jiangxi 330022,P.R.China;Jiangxi Forestry Resources Monitoring Eenter,Nanchang Jiangxi 330046;School of Geography and Environment,Jiangxi Normal University,Nanchang Jiangxi 330022,P.R.China)

机构地区：[1]江西师范大学鄱阳湖湿地与流域研究教育部重点实验室,江西南昌330022 [2]江西省林业资源监测中心,江西南昌330046 [3]江西师范大学地理与环境学院,江西南昌330022

出　　处：《西部林业科学》2022年第5期28-33,共6页Journal of West China Forestry Science

基　　金：鄱阳湖湿地与流域研究教育部重点实验室(江西师范大学)开放基金资助项目(PK2020005);江西省林业局林业科技创新专项项目(创新专项[2022]17号)。

摘　　要：松材线虫病是一种由松墨天牛携带传播松材线虫而引起的毁灭性病害,已经成为近40 a来我国发生最严重、最危险的重大林业病害。传统的监测方法主要依靠人工外业调查和卫星遥感目视判读的方法,不仅工作量大,而且效率低下。本研究以江西省赣州市为研究区域,利用无人机航空摄影获取松材线虫病疫区遥感影像,以被松材线虫感染的疫木为目标,通过面向对象提取和人工校核的方式对图像进行标注,并以256×256像素大小对图像和标签进行裁剪,以8︰2的比例将图像随机分成训练集和验证集,从而构建出包括图像和相应标签的深度学习样本数据库。然后基于TensorFlow开源框架搭建了U-Net和SegNet两种深度学习网络模型。结果显示,U-Net网络模型和SegNet网络模型训练精度均超过95%,U-Net网络模型验证精度要高于SegNet网络模型,SegNet网络模型在训练过程中出现了过拟合现象。最后本研究选取部分区域开展了预测应用,结果表明,U-Net网络模型精确率为81.36%,召回率为83.12%,调和平均值为82.23%,高于SegNet网络模型的75.21%、76.84%和76.01%,U-Net网络模型提取效果优于SegNet网络模型。Pine wood nematode disease is a devastating disease caused by pine wood nematode carried by Monochamus alternatus.It has become the most serious and dangerous major forestry disease in China in recent decades.Traditional monitoring methods mainly rely on manual field investigation and visual interpretation of satellite remote sensing,which is not only heavy workload,but also inefficient.In this study,Ganzhou City,Jiangxi Province was selected as the research area,and the remote sensing image of pine wood nematode epidemic area was obtained by unmanned aerial vehicle aerial photography.Taking the diseased wood infected by pine wood nematode as the target,the image was marked by object-oriented extraction and manual verification,and the image and label were cut with the size of 256×256 pixels,The images were randomly divided into training set and verification set at the ratio of 8︰2,so as to construct a deep learning sample database including images and corresponding labels.Then,two deep learning network models,u-net and segnet,were built based on tensorflow open source framework.The results showed that the training accuracy of u-net network model and segnet network model was more than 95%,and the verification accuracy of u-net network model was higher than segnet network model.There was an over fitting phenomenon in the training process of segnet network model.Finally,this study selected some regions for prediction application.The results indicated that the accuracy rate of u-net network model was 81.36%,the recall rate was 83.12%,and the harmonic average value was 82.23%,which was higher than 75.21%,76.84% and 76.01% of segnet network model.The extraction effect of u-net network model is better than segnet network model.

关 键 词：松材线虫 无人机遥感 深度学习 样本数据库 U-Net SegNet 网络模型 

分 类 号：S763[农业科学—森林保护学]