基于GF-2卫星影像的川东丘陵地区耕地信息提取

发布时间：2018-05-08 14:48

本文选题：GF-2 + 影像　；参考：《成都理工大学》2017年硕士论文

【摘要】：耕地资源是人类生存与发展不可或缺的物质基础,是区域内粮食自给能力的根本所在,是国家粮食安全的可靠保障,我国必须守住18亿亩耕地红线。在城市不断发展过程中,土地资源的消耗越来越大,耕地资源在这一过程中被逐步侵占的问题十分严重。耕地面积锐减是这一时期土地利用/覆被(LUCC)变化的最显著特征,在一定时空程度上而言,建设用地与耕地资源之间的关系是此消彼长。因此,如何快速、准确地获取耕地信息对耕地动态监测、耕地保护以及土地资源利用程度分析,协调好人地关系具有重要意义,也可以为党和国家制定相关政策提供依据。而遥感影像中普遍存在的同物异谱、异物同谱现象使得地物分离精度大大降低。并且由于耕地与相邻建设用地、林地等地物交织混杂,形成不易区分的混合体,且分布破碎无规律,加之土壤特性、作物以及耕作方式的不同,地类间光谱特征的差异,这使得基于中低分辨率卫星影像的耕地信息提取效果不如人意。随着遥感卫星技术的迅猛发展,尤其是高空间分辨率卫星的发展,使得高精度地物分类成为可能。目前,有关利用高分辨率遥感影像提取耕地信息已渐成研究热点。这为进行大区域、高精度的遥感信息尤其是耕地信息的快速有效提取提供了新思路。本文以地处川东丘陵地区的隆昌市龙市镇为研究区进行耕地信息提取,龙市镇是隆昌市基本农田保护示范乡镇,镇内耕地面积大,具有代表性。研究采用基于像元的最小距离法、最大似然法、K均值法、ISODATA法与面向对象的隶属度函数分类法、CART决策树法、支持向量机法对研究区GF-2卫星影像数据进行耕地信息提取。主要取得以下成果:(1)研究区GF-2影像数据,分别采用Gram-Schmidt、PCA变换与NNDiffuse四种方法进行影像融合,通过对比亮度、对比度、纹理等特征得出该区域影像的最佳融合方法为NNDiffuse融合。(2)基于影像的光谱特征对研究区融合影像进行非监督分类实验和监督分类实验。在易康(eCognition)软件平台上进行多尺度分割,确定适合研究区影像的分割尺度与相关参数,建立分类体系,通过特征优化选择优化特征集,结合研究区数字高程模型(DEM)进行耕地信息提取,结果表明基于优化特征的分类效果优于单纯基于光谱特征的分类效果。(3)结合野外实地验证,评价不同方法提取耕地信息的精度,结果表明,面向对象的分类方法能更快速、准确地提取丘陵地区耕地信息。(4)分析研究区耕地分布与区内地形、坡度之间的关系,并对耕地保护与基本农田保护区划定给出相关建议。
[Abstract]:The cultivated land resource is the indispensable material foundation for the human survival and development, is the basic of the grain self-sufficiency ability in the region, is the reliable guarantee of the national food security, our country must keep the 1.8 billion mu cultivated land red line. In the process of continuous development of cities, the consumption of land resources is increasing, and the problem of cultivated land resources being gradually encroached on in this process is very serious. The sharp decrease of cultivated land area is the most significant feature of LUCCs change in this period. In a certain time and space, the relationship between construction land and cultivated land resources is the same. Therefore, how to obtain cultivated land information quickly and accurately is of great significance to the dynamic monitoring of cultivated land, the analysis of cultivated land protection and the degree of land resource utilization, and the coordination of human-land relations, which can also provide the basis for the Party and the state to formulate relevant policies. In remote sensing images, the common isospectral and foreign body isospectral phenomena reduce the precision of ground object separation greatly. And because cultivated land, adjacent construction land, woodland and other ground objects are intertwined and mixed, the mixture is not easy to distinguish, and the distribution is irregular, in addition, the characteristics of soil, the difference of crop and tillage methods, the difference of spectral characteristics among land types. This makes the effect of farmland information extraction based on low-resolution satellite image unsatisfactory. With the rapid development of remote sensing satellite technology, especially the development of high spatial resolution satellite, it is possible to classify ground objects with high precision. At present, the use of high-resolution remote sensing images to extract cultivated land information has become a research hotspot. This provides a new idea for fast and effective extraction of remote sensing information, especially cultivated land information. This paper takes Longchang City, located in the hilly region of East Sichuan, as the research area to extract cultivated land information. Longcheng Town is the demonstration town of basic farmland protection in Longchang City. The cultivated land area in the town is large and representative. In this paper, the method of minimum distance based on pixel, the method of maximum likelihood and the method of ISODATA and the method of object oriented membership function classification are used to extract cultivated land information from GF-2 satellite image data in the study area by support vector machine (SVM). The main achievements are as follows: 1) the GF-2 image data in the research area are fused by Gram-SchmidtPCA transform and NNDiffuse respectively, and the brightness and contrast are compared. Texture and other features show that the best fusion method for the region is NNDiffuse fusion. (2) based on the spectral features of the image, the unsupervised classification experiment and supervised classification experiment are carried out for the fusion image in the study area. Multi-scale segmentation is carried out on the software platform of Accom Cognition to determine the segmentation scale and relevant parameters suitable for the image in the study area, to establish a classification system, and to select the optimized feature set through feature optimization. The result shows that the classification effect based on the optimized feature is better than the classification effect based solely on the spectral feature. Combined with field verification, the accuracy of different methods for extracting cultivated land information is evaluated. The results show that the object-oriented classification method can extract the cultivated land information of hilly area more quickly and accurately. Some suggestions on farmland protection and basic farmland protection are also given.
【学位授予单位】：成都理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S127

【参考文献】