【作者】 亓兴兰；

【导师】 刘健；

【作者基本信息】 福建农林大学 ， 森林经理学， 2011， 博士

【摘要】 马尾松毛虫(Dendrolimus Punctatus Walker)是马尾松的主要害虫,严重制约了马尾松的生长,其发生频繁,常猖獗成灾,被称为“无烟的森林火灾”。在我国南方地区,马尾松毛虫常年发生,一般3至5年就成灾一次,为害面积高达2.0x106-3.3×106hm2,减少木材量约3.0x106m3,经济损失达数十亿元,是发生面积最广、为害最为严重的森林害虫,危害很大。实现马尾松毛虫害的有效监测,是防治松毛虫危害的关键,但传统地面抽样调查方法费时费力,已经不能满足林业生产实际需要,而遥感技术具有大面积覆盖且实时探测的优点,具有技术上的可行性与经济上的节约性。为此,本文以沙县作为研究区,以研究区SPOT-5遥感影像多光谱数据与全色数据为基础数据,结合外业调查数据、研究区森林资源数据库、研究区DEM等辅助数据,分析虫害影像的光谱变化与纹理变化,通过光谱指标的选取、纹理特征的分析等,从多方面对影像虫害信息进行提取,从而建立遥感监测森林病虫害的技术体系,构建马尾松毛虫害的遥感监测指标体系,并建立一新的针对马尾松毛虫害的基于片层—面向类的影像分类新方法,为实际生产监测马尾松毛虫害提供技术支持与理论基础,并为高分辨率影像用于森林病虫害的监测提供借鉴。主要研究结果如下：(1)参考国土资源部的土地利用一级分类系统,结合研究区实际情况,针对研究需要,确定研究区沙县的土地利用分类系统为：林地与非林地,林地又分为马尾松、杉木和阔叶树。基于影像的光谱特征,基于分层分类的思想,结合阈值设定与决策树分类QUEST算法,提取出林地影像,在此基础上,实现马尾松林专题信息的提取,其提取精度达到92.89%。(2)对SPOT-5多光谱影像进行光谱特征分析,发现受害马尾松在2、3、4波段变化明显,基于此,首次构建了马尾松毛虫害的遥感监测指标体系。(3)基于构建的光谱指标,采用岭迹分析方法进行监测因子的优选,建立了表征虫害程度的虫情级数的岭回归估测模型,其模型估测精度达到81.69%。进而实现研究区虫情级数的反演,并进行马尾松毛虫害的专题信息提取,其信息提取总精度=70.75%,Kappa=0.6759,分类精度不高,且椒盐现象比较严重,健康林分与轻度受害林分混淆较严重。(4)提取图像的纹理特征并分析,分别采用基于像元的最大似然法与面向对象分类方法,进行马尾松毛虫害信息提取。前者分类总精度=72.75%,Kappa=0.6913;后者分类总精度=74.75%,Kappa=0.7283,后者精度高于前者。两者相比于基于光谱信息分类来说,其分类精度有了一定的提高,分别为：基于像元统计提高了2%,面向对象方法提高了4%。说明纹理特征对于图像的分类起关键作用。(5)研究进行光谱信息与纹理信息的融合,引入纹理信息参与虫害图像的分类,采用融合多尺度纹理与光谱信息的SVM分类方法进行虫害信息提取,并与单尺度纹理的分类方法进行比较,前者分类总精度=82.50%,Kappa=0.8059,后者分类总精度=80.75%,Kappa=0.7824,前者分类精度高于后者。同时两者相比于单纯基于光谱信息与纹理信息都有了提高,分别为：单尺度纹理分别提高了10%、6%,多尺度纹理提高了11.75%、7.75%。光谱信息与纹理信息的融合显著地增强了马尾松毛虫害的光谱响应能力,其信息量明显增多,同时多源信息的引入有利于提高影像的分类精度。(6)融合影像的光谱信息与纹理信息,并考虑影响昆虫发生发展及变化的生态及林木自身因子,首次构建了马尾松毛虫害遥感监测指标体系。基于建立的监测指标体系,通过岭回归建模,实现了马尾松毛虫害专题信息的提取,其分类总精度=85.75%,Kappa=0.8328。相比于没有考虑地形及林木因子的三种方法,其精度分别提高了15%、11%、3%。(7)基于虫害影像特征,提出基于生态因子分片—面向类的新算法,该法基于海拔、坡度与坡向,对马尾松进行空间分区划片,然后针对每片影像的特点,综合运用决策树、面向对象等分类方法,进行虫害信息的提取,其分类总精度=87.50%,Kappa=0.8559,其精度高于岭回归建模的方法,其分类总体精度提高了1.75%。该法综合考虑了各方面的因素,消除了地形的影响,有效避免了椒盐现象的发生。(8)对于各种程度受害的马尾松来说,从各种方法提取的精度来分析,健康马尾松与重度受害马尾松提取精度最高,而轻度受害马尾松与中度受害马尾松精度稍低。各种提取虫害信息的方法,以生态因子分区划片-面向类的方法精度最高,虫害信息分布图质量也最高,是最佳的信息提取方法。(9)森林病虫害的遥感监测不仅仅是一个图像识别的问题,病虫害的发生及发展与其生存的环境有密切的关系,同时遭受病虫害的植被在遥感图像上的响应,还受林木自身的生长规律影响,所以在进行森林病虫害监测时,综合考虑病虫寄主的生态条件及其自身林分状况等因子的影响,才能正确识别病虫害区域,提高监测精度。更多还原

【Abstract】 Dendrolimus Punctatus Walker (the main insect pests of pinus massoniana) is seriously restrain the growth of pinus massoniana, it occurs frequently, often rampant hazard, known as "the forest fire of smokeless". In southern China region, dendrolimus punctatus perennial happen, cause disaster generally three to five years, as once flood-gates pestcide area 2.0×106～3.3×106hm2, reduce wood quantity about 3.0×106 m3, economic loss of one billion yuan, is the most pest in the case of area of forest pest and serious harm. Realize the effective monitoring of Dendrolimus punctatus Damage, is the key to reduce harm. Traditional ground sampling survey method is time-consuming,and cannot satisfied with forestry production actual need.Remote sensing technology has the advantages of covering and real-time detection of large area, has the technical feasibility and economic save. Therefore, ShaXian was chosen as study area in this paper, based on multi-spectral and panchromatic images of SPOT-5, combined with the field survey data, research area forest resources database, research area DEM auxiliary data, analyze spectral changes and textures change of pests images, the information of Dendrolimus punctatus Damage was extracted with the spectrum index selection and the analysis of the texture characteristics etc. Thus the remote sensing monitoring forest pest technology system was established, and the remote sensing monitoring Dendrolimus punctatus Damage index system was constructed, and a new method based on pests in pinus massoniana slice layer of image classification -- facing classes was established, provide technical support and theoretical basis for production monitoring, and as reference for high resolution images used for forest pest monitoring. The main research results are as follows:(1) Referenced to land use classification system of ministry of land resources, combined with actual situation of study area, according to research needs, the land use classification system of Shaxian county was determined:woodland and unwoodland, woodland were divided into Pinus assoniana, fir,and hardwood Based on the spectral features of images and sub-region and hierarchical theory, combining the threshold set and decision tree algorithm of QUEST classification, the woodland image was extracted, on this basis, the pinus massoniana project information was extracted, the extraction accuracy reached 92.89%.(2) The victims of pinus massoniana of images changed greatly in 2,3,4 band with SPOT-5 spectral analysis. Based on this, the remote sensing monitoring index system of Dendrolimus punctatus Damage was firstly constructed. (3) Based on the spectral index, the insect scouting series estimation model that Characterizing pest degree was established, its model estimation accuracy reached 81.69%. So as to realize the inversion of insect scouting series and extract the project information of pinus massoniana damage, the information extraction total precision was 70.75%, Kappa=0.6759, the classification accuracy was not high, Salt-pepper phenomenon was serious, and Health stand with mild victims stand confused seriously.(4) The image texture feature were extracted and analysised, and the information of pinus massoniana wool pests were extracted with the maximum likelihood method and object-oriented classification method. The former classification total precision was 72.75%, Kappa=0.6913,The latter classification total precision was 74.75%, Kappa=0.7283, the latter is higher than that of the former. compared to the classification that based on spectral information, both classification accuracy have improved, as followed:the classification method that based on pixel statistics raised 2%, object-oriented method increased 4%.It showed that the texture characteristics for image classification played a key role.(5) Study on fusion algorithms of spectral information and texture information, introducing texture information to the image classification of Dendrolimus punctatus damage, extracting pests information with SVM classification method based on multi-scale texture and spectrum fusion, compared to the single yardstick texture method, the former classification total precision was 82.50%, Kappa=0.8059, the latter classification total precision was 80.75%, Kappa= 0.7824, the latter classification accuracy was higher than the former. Meanwhile compared to only based on spectral information and texture information, both classification precision had increased, as followed:monoscale texture respectively increased by 10%、6%, multi-scale texture improved 11.75%,7.75%. Spectrum information and texture information fusion significantly enhanced the pests spectral response, and the introduction of multi-source information were helpful to improve the image classification accuracy.(6) Fusing spectrum information and texture information of image, considering the ecological and forest itself factor that impacting occurrence and development and change of insects, the remote sensing monitoring index system of Dendrolimus punctatus damage were firstly constructed. Based on the monitoring index system, the project information of Dendrolimus punctatus damage were extracted by ridge regression model its classification total precision was 85.75%, Kappa=0.8328. Compared to the method without terrain and forest factor, its accuracy increased 15%,11%,3%.(7) Based on the characteristics of image of pests, the new method-based on the ecological factors patches for class was put forward.Based on the altitude, slope and aspect, the pinus massoniana were partited in space, then according to the characteristics of each piece, the decision-making tree, object-oriented classification method were used comprehensively to extract information of pests, the classification total precision was 87.50%, Kappa=0.8559, its precision was higher than ridge regression modeling method, and the classification total accuracy improved 1.75%. This method considered comprehensively many factors, eliminated the influence of terrain, effectively avoid salt-pepper happening.(8) Analysising various methods to extracting information of Pinus massoniana wool pest, the infonnation extraction accuracy of health stand and severe suffer stand were highest, the information extraction accuracy of mild suffer stand and moderate suffer stand were slightly lower. Among various method of extracting pest information, the classification precision of new method -based on the ecological factors patches for class was highest, This new method was the best information extraction method.(9) The remote sensing monitoring of forest diseases and pests is not just a question of image recognition, the plant diseases and pests occurrence and development have a close relationship with its survival environment, and the spectrum responses of pests were influenced by the forest itself. So remote sensing monitoring the forest diseases and pests, the ecological conditions and its own stand condition factor of host should be considered, thus the pest areas are correctly identified and the monitoring accuracy are improved.更多还原