EGI OpenIR
基于多源数据的耕地提取与新疆主产作物识别的对比研究
Alternative TitleComparative Study on Cultivated Land Extraction and Identification of Main Crops in Xinjiang Base on Muilt-Source Data
刘松江
Subtype硕士
Thesis Advisor陈曦
2019-06-30
Degree Grantor中国科学院大学
Place of Conferral北京
Degree Discipline工程硕士
Keyword多源数据 耕地提取 农作物识别 对比研究 光学和雷达 Muilt-Source Data Cultivated land extraction Crop Identification Comparative Study Optics and radar
Abstract目前,农情遥感监测的主要手段还是光学数据,利用光学数据进行耕地提取和农作物分类技术相对来说都比较成熟,但是光学影像易受云雨天气的影响,在部分场景下往往无法获取可用的影像数据。 在这种情况下, 合成孔径雷达数据以其特有的主动性、敏感性、穿透性优势, 逐渐成为了一种全新的农情监测手段。本研究选取新疆典型农业发展带玛纳斯河流域作为研究区, 分别选择光学和合成孔径雷达数据进行了研究区的耕地提取和主要农作物识别两部分对比实验。(1) 首先, 选择单时相的 Landast8 数据和 Sentinel-1 数据,在 DEM 坡度数据的基础上采用支持向量机和最大似然法提取玛纳斯流域耕地,综合对比雷达数据、光学数据、雷达光学融合数据以及雷达光学特征组合数据的耕地提取精度,探究不同数据融合方法、不同类型数据、不同特征组合及不同分类方法对耕地提取的影响。通过耕地提取的对比实验,得到了如下结论: a)通过雷达光学融合影像耕地提取对比实验证明: 小波变换的影像融合效果比正交化变换要好; b)同等条件下,单时相的雷达与光学融合数据耕地提取精度高于单时相光学数据的提取精度, 单时相光学数据耕地提取精度高于单时相雷达数据; c)纹理特征有助于耕地信息的提取,雷达数据的纹理信息比光学数据的纹理信息更为丰富; 雷达后向散射特征比雷达纹理特征对提取耕地的正向影响更大; 不同特征组合的耕地提取精度由高到低为:光学光谱+雷达后向散射>光学光谱+雷达纹理>光学光谱+光学纹理>光学光谱数据>雷达后向散射+雷达纹理; d) 在本文的耕地提取方案下, 支持向量机的耕地提取结果优于最大似然法。(2) 其次, 在耕地信息提取的基础上,对不同的时序组合数据(未处理的多时相 Landsat-8 数据、 主成分分析后的多时相 Landsat-8 数据、全作物生育期的多时相 Sentinel-A 数据、作物关键物候期的多时相 Sentinel-1A 数据、 多时相Landsat-8 和单时相 Sentinel-1 时相组合数据、单时相 Landsat-8 和多时相Sentinel-1 组合数据、多时相 Landsat-8 和多时相 Sentinel-1 时相组合数据)采用随机森林和最大似然法进行研究区主产作物识别(棉花、玉米、小麦、其他)的对比实验,探究不同数据处理方式、 不同长度时相、不同数据源、 不同数据组合和不同分类方法对农作物识别精度的影响。通过农作物分类对比实验,得到了如下结论: a)基于多时相波段组合数据的农作物分类方法可行, 基于不同作物之间光学数据的时序光谱差异和雷达数据的时序后向散射系数差异,能够实现农作物的分类; b)将主成分分析引入多时相光学波段组合的农作物分类,不仅能提高分类精度,而且能大大提高提高分类速度;c)基于多时相雷达波段组合数据进行作物分类,全生育期的数据比关键物候期数据可以获得更高的分类精度,增加时相数可以提高分类精度;相比全生育期数据,正确选择关键物候期进行农作物分类,能够大大减轻前期数据处理任务,提高农作物分类效率; d)引入光学数据可以有效提高多时相 SAR 数据农作物总体分类精度, 比单独使用多时相 SAR 数据总体精度提高 1%-6%, 其中 6 月 24 日光学数据的加入获得了最大的分类精度; e)在多时相光学数据中引入雷达数据,农作物分类精度有所提高。但引入不同时相的 SAR 数据对农作物分类总体精度提高基本一致,时相间分类精度差异较小; f)多时相光学与多时相雷达组合数据比单一多时相数据的农作物分类精度更高, 3 者分类精度由高到低排序为:多时相光学与多时相雷达组合数据>多时相光学数据>全生育期的多时相雷达数据>关键物候期的多时相雷达数据; g )本文的农作物分类方法中, 随机森林的分类精度高于最大似然分类精度。
Other AbstractAt present, the main means of agricultural remote sensing monitoring is opticaldata. The use of optical data for arable land extraction and crop classificationtechnology is relatively mature. But optical images are susceptible to cloud and rainweather. In some scenarios, it is often impossible to obtain available images. In thiscase, Synthetic Aperture Radar (SAR) has gradually become a new means ofmonitoring agricultural conditions with its unique initiative, sensitivity andpenetrating advantages. In this study, the Manas River Watershed was selected as thestudy area, which is the typical agricultural development belt in XinJiang. The opticaland synthetic aperture radar data were selected to study the two aspects of cultivatedland extraction and main crop identification.(1) Firstly, we select the single-phase Landast-8 data and Sentinel-1 data, and usethe support vector machine and the maximum likelihood method to extract the ManasRiver Watershed cultivated land based on the DEM slope data. We comprehensivelycompare the accuracy of cultivated land extraction with radar data, optical data, radaroptical fusion data, and radar optical feature combined data and explore the effects ofdifferent data fusion methods, different types of data, different feature combinationsand different classification methods on cultivated land extraction.Through the comparative experiments of cultivated land extraction, the followingconclusions are obtained: a) Contrast experiments with radar optical fusion imagecultivated land prove that the image fusion effect of wavelet transform is better thanGram-Schmidt transform; b) The single-phase radar and The accuracy of opticalfusion data for arable land extraction is higher than that for single-phase optical data.The accuracy of single-temporal radar and optical fusion data for arable landextraction is higher than that for single-phase optical data.The accuracy ofsingle-phase optical data for arable land extraction is higher than that for single-phaseradar data; c) The texture feature is helpful for the extraction of cultivated landinformation and the texture information of the radar data is richer than the textureinformation of the optical data;the radar backscattering feature has more positiveinfluence on the extracted cultivated land than the radar texture feature; d) theaccuracy of cultivated land extraction with different feature combinations is from highto low: Optical Spectrum+Radar Backscatter > Optical Spectrum+Radar Texture >Optical Spectrum+Optical Texture > Optical Spectral Data>Radar Backscatter+ Radar Texture; e) Under the cultivated land extraction scheme of this paper, the results ofcultivated land extraction by the support vector machine are better than the maximumlikelihood method.(2) After the extraction of cultivated land information, we do comparativeexperiments on the identification of main crops (cotton, corn, wheat, others) in thestudy area using random forest and maximum likelihood method based on thedifferent time series combinations data (unprocessed multi-temporal phase Landsat-8data, multi-temporal phase Landsat-8 data with principal component analysis,multi-temporal phase Sentinel-A data of whole crops growth period, multi-temporalphase Sentinel-1A data for crop critical phenology, the combination data ofmulti-temporal phase Landsat-8 and single-phase Sentinel-1A , the combination dataof single-phase Landsat-8 and multi-temporal phase Sentinel-1A, the combinationdata of multi-temporal phase Landsat-8 and multi-temporal phase Sentinel-1).Weexplore the impact of different data processing methods, different length phases,different data sources, different data combinations and different classification methodson crop identification accuracy.Through the crop classification's comparable experiments, the followingconclusions are obtained: a) The crop classification method based on the combinationdata of multi-temporal bands is feasible.We can achieve the classification of cropsbased on the time series spectral difference of different crops optical data and the timeseries backscattering coefficient difference of different crops radar data; b) Weintroduce the Principal Component Analysis into crop classification of themulti-temporal optical band combination, which can not only improve classificationaccuracy, but also can greatly improve the classification speed; c) Crop classificationbased on multi-temporal radar band combination data, the data of whole growthperiod can obtain higher classification accuracy than the data of the keyphenology.We can improve classification accuracy by increasing the number ofphases. Compared with data covering the entire growth period, correctly selecting thedata of the essential phenological period can greatly reduce the previous dataprocessing tasks and improve efficiency of crop classification; d) The introduction ofoptical data can effectively improve the overall crops classification accuracy ofmulti-temporal SAR data, which is 1%-6% higher than the overall accuracy ofmulti-temporal SAR data alone. The addition of optical data on June 24 has obtained the largest classification accuracy; e) After the introduction of radar data, the cropsclassification accuracy of multi-temporal optical data has improved. However, theintroduction of SAR data with different time phases is basically consistent with theoverall accuracy of crop classification. f) Combined data from multi-temporal opticsand multi-temporal radar can achieve higher crop classification accuracy than singlemulti-temporal data. The classification accuracy of the three types of data is sortedfrom high to low: the combination data of multi-temporal phase optics andmulti-temporal phase radar data> multi-temporal phase optical data> multi-temporalphase radar data of the whole crops growth period> multi-temporal phase radar dataof critical phenology; g) Between the crop classification methods of this paper, thecrops classification accuracy of random forest is higher than the maximumlikelihood's classification accuracy.
Subject Area测绘工程
Language中文
Document Type学位论文
Identifierhttp://ir.xjlas.org/handle/365004/15341
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
刘松江. 基于多源数据的耕地提取与新疆主产作物识别的对比研究[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
[刘松江]'s Articles
Baidu academic
Similar articles in Baidu academic
[刘松江]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[刘松江]'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.