EGI OpenIR
基于集成分类器的多时相高分一号/二号影像农林间作系统分类方法研究
Alternative TitleIntercropping Classification from Multi-temporal GF-1 and GF-2 Satellite Imagery Using Ensemble classifier
刘萍
Subtype博士
Thesis Advisor陈曦
2019-06-30
Degree Grantor中国科学院大学
Place of Conferral北京
Degree Discipline理学博士
Keyword集成分类器 农林间作分类 特征提取 GF-2 intercropping classification classifier ensemble rotation forest GF-2
Abstract农林间作是广泛存在的立体耕地利用模式,这种种植模式可实现对耕地系统的光、热、水、土壤等资源最大化的综合利用,已有大量学者从农业管理、生态效应的角度对农林间作系统展开了讨论和研究,但少有遥感监测方面的应用。高分一号和高分二号卫星具备高空间分辨率,高时间分辨率宽覆盖,多波段等的特征,具备广阔的应用前景。本研究选择了位于塔里木盆地墨玉县南部的核桃-玉米间作区域,通过对多时相高分一号和高分二号卫星影像特征的挖掘,从其高空间分辨率,多时相的特征出发,研究基于集成分类器的不同特征组合情况下的间作地区作物分类。 本文的主要研究内容如下:1 )在对各分类算法进行分析的基础上, 对传统采用决策树作为旋转森林基分类器的算法进行改进,将能够很好支持小样本,多维度特征的支持向量机算法引入作为基分类器,改进旋转森林集成支持向量机(RoF-SVM)的方法,该方法既具有 SVM 算法的优点,又能通过旋转森林增强各个基分类器之间的差异性从而提高了分类的精度。并将该分类结果与常见的最大似然法, 支持向量机以及旋转森林算法进行实验对比。实验证明 RoF-SVM 在分类精度方面实现了最优,并且有效降低了影像分类图上的椒盐噪声。2 )GF-1 卫星的宽覆盖(Wide Field View, WFV )数据空间分辨率 16m,重访周期为 2 天, 可以为农作物生长全周期的监测过程免费提供持续可靠的数据来源。收集本研究区域作物生长时间段内时间序列高分一号卫星的 WFV 影像,计算出 NDVI 时间序列数据并进行拟合,通过与当地物候历结合,研究区作物发育情况,为选取高空间分辨率影像的时相进行作物精细提取提供依据。3 )由于空间分布的多样性和复杂性,为实现对不同类型植被的精细识别, 需要对多特征信息进行提取。基于研究区融合后的 1m 空间分辨率 GF-2 卫星影像,采用灰度共生矩阵提取了 4 个波段的 32 个纹理信息;结合配准后的多时相 GF-1卫星影像,提取不同时相的归一化植被指数、以及增强型植被指数信息。通过 a多光谱波段, b 单一时相 GF-2 多特征数据, c 多时相多特征数据,三种组合之间的对比实验,证明了结合多时相的光谱特征及 GF-2 纹理特征,可增大不同类型作物之间的差异性,获得较好的分类精度。本研究针对 GF-1, GF-2 卫星影像特征,通过对农林间作样区的分类研究,形成了比较完善的特征提取和间作地区作物分类方案。实验表明, GF-1 和 GF-2卫星多光谱影像在结合纹理和多时相植被指数特征后可以在光谱信息复杂,植被信息多样化的间作区域内提供足够的植被覆盖细节,实现农林信息的精细提取(总体精度达到 86.87%, kappa 系数为 0.78 )。该研究为充分发掘国产高分卫星在农作物和林业方面的应用潜力提供了理论依据。
Other AbstractIntercropping serves as a widely used means to increase the food productivity,thereby enhancing farm incomes, improving soil and water quality, and reducinggreenhouse gas emissions. Surveying and mapping the plant types, quantity structures,and spatial distribution characteristics are important in improving intercroppingsystems and in estimating potential yields and tree crop system adjustments. Butremote sensing is rarely used for intercropping area monitoring.Chinese GF-1 and GF-2 satellites with high spatial resolution, high timeresolution, wide coverage, and four spectral bands have broad application potentials.Walnut and maize intercropping areas in Moyu County south of the Tarim Basin wereselected as the study area. The ensemble classification methods was studied in theintercropping area based on the combination of different features of the GF-1 andGF-2 images. The main research contents of this paper are as follows:1) The rotation forest method was then adopted based on a Support VectorMachine (RoF-SVM), which offers the advantage of using an SVM algorithm and thatboosts the diversity of individual base classifiers by a rotation forest. The results werecompared with those of the maximum likelihood classifier, support vector machineand random forest method. Different types of trees and crops in intercropping areascan be effectively distinguished by RoF-SVM (the OA was measured at 86.87 and thekappa coefficient was measured at 0.78). Moreover, classification results effectivelyeliminated salt and pepper noise.2) The Wide Field View (WFV) sensors of GF-1 sattellite with data spatialresolution of 16m and a revisit period of 2 days, providing a free and reliable datasource for crop growth and change detectation information. For the study area, NDVItime series data were calculated using WFV images of time series GF-1 satellitesduring crop growth period, and Gaussian fitting was performed to obtain NDVI timeseries data. Combined with the local phenology, the crop growth characteristics of the study area provide a basis for the fine extraction of crops for the selection of keyphases of high spatial resolution images.3) To evaluate optimal feature selection, three group feature combinationsderived from GF-1 and GF-2 images were combined, and the RoF-SVM ensemblelearning method was adopted to classify the three groups. The group that combinedspectral-textural-multitemporal features achieved the best classification results. Then,the classification results were compared with those of the MLC and SVM via optimalfeature selection.Due to the diversity and complexity of spatial distribution, in order to achieveaccurate identification of different types of vegetation, multi-feature informationneeds to be extracted.。 Based on the 1m spatial resolution GF-2 satellite image afterthe fusion of the study area, 32 texture information of 4 bands were extracted by aGray level co-occurrence matrix (GLCM); combined with the multi-temporal GF-1satellite images after pre-processing, NDVI and EVI of different temporal wereextracted. Through the classification and comparison experiments between the threefeature combinations, it is proved that combining the spectral characteristics ofmulti-temporal phase and GF-2 texture features can increase the difference betweendifferent types of crops and obtain better classification accuracy. It is shown that theRoF-SVM algorithm for the combined spectral-textural-multitemporal features caneffectively classify an intercropping area (overall accuracy of 86.87% and kappa coefficientof 0.78), and the classification result effectively eliminated salt and pepper noise.This study used GF-1 and GF-2 imagery developed a comprehensive featureextraction and intercropping classification scheme. And proved by experiments thatthe GF-1 and GF-2 satellite images combined with spectral, textural, andmulti-temporal features can provide sufficient information on vegetation cover locatedin an extremely complex and diverse intercropping area. This study provides a basisfor exploring the application potential of Chinese Gaofen satellites in cropidentification and monitoring, and has certain application prospects.
Subject Area地图学与地理信息系统
Language中文
Document Type学位论文
Identifierhttp://ir.xjlas.org/handle/365004/15340
Collection中国科学院新疆生态与地理研究所
研究系统
Affiliation中国科学院新疆生态与地理研究所
First Author Affilication中国科学院新疆生态与地理研究所
Recommended Citation
GB/T 7714
刘萍. 基于集成分类器的多时相高分一号/二号影像农林间作系统分类方法研究[D]. 北京. 中国科学院大学,2019.
Files in This Item:
There are no files associated with this item.
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[刘萍]'s Articles
Baidu academic
Similar articles in Baidu academic
[刘萍]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[刘萍]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.