KMS XINJIANG INSTITUTE OF ECOLOGY AND GEOGRAPHY，CAS
| 时间序列卫星数据的农作物分类研究及应用——以天山北坡为例 | |
| 黄双燕 | |
| Subtype | 硕士 |
| Thesis Advisor | 陈曦 |
| 2018-06-01 | |
| Degree Grantor | 中国科学院大学 |
| Place of Conferral | 新疆乌鲁木齐 |
| Degree Discipline | 工程硕士 |
| Keyword | 机器学习 随机森林 农作物分类 地块基元 多源数据 Machine learning Random forest Crop classification Parcel data set Multi-sources data |
| Abstract | 当前，基于机器学习方法开展农作物分类研究，对于确保干旱区粮食安全和生态安全有着极为重要的现实意义。 近年来，遥感技术在农作物分类和种植结构提取方面已取得长足的发展，但仍面临诸多困难与挑战：一是农作物种类自动识别方法有待进一步提升，二是时间序列多源数据融合技术有待进一步研究。本研究选择时间序列 Sentinel 2A与 Landsat 8卫星数据，结合野外调查数据，采用随机森林机器学习方法，探索基于地块基元和红边特征的多源遥感数据农作物自动分类方法，实现了天山北坡地区主要农作物种植结构的自动提取。主要内容包括： 引入地块基元和红边特征， 探讨不同分类特征组合对机器学习分类精度的影响； 探讨分别在数据级和特征级联合利用中分辨率时间序列多源卫星数据对农作物分类精度有何影响，确定时间序列 Sentinel 2A 数据和 Landsat 8 数据在干旱区分类应用过程中的最佳融合层次。 结果表明：（1） 基于时间序列 Sentinel 2A， 分别利用 8 种时间序列光谱、植被指数分类特征集对沙湾县进行农作物分类识别， 总体精度均在 89 % 以上。分类组最高精度达 94.02 %，表明随机森林机器学习分类器可有效集成多维向量的优势， 是一种行之有效的干旱区典型农作物分类方法。引入地块基元后， 有红边光谱组及有红边指数组的分类精度分别提高 3.13 % 和 4.07 %，表明利用地块基元中心点作为同质地块的代表像元， 提取的分类特征可供机器学习分类器使用，能较大幅度提高分类精度及效率，平滑地块内部同种作物间的“椒盐噪声”。红边特征的引入使光谱组及指数组的分类精度分别提高 2.39 % 和 1.63 %，对随机森林分类模型的精度提升十分有效。（2） 探讨基于数据级和特征级的多源数据联合利用方式对农作物分类精度的影响， 结果表明 Landsat 8 与 Sentinel 2A 的融合层次在植被指数特征级精度最高。 在此基础上对温泉县进行农作物遥感分类识别的精度达 92.92%。 特征级分类特征的增加会促使随机森林分类器的总体分类精度及各作物精度提升。（3） 引入地块基元与红边特征，基于 Landsat 8 与 Sentinel 2A 的时间序列植被指数特征级融合数据， 采用上述基于随机森林的机器学习方法，对天山北坡主要农作物进行分类识别，其总体分类精度达到了 81.51 %，符合大区域遥感监测应用需求。 2016 年天山北坡研究区（含兵团） 内，小麦种植面积最大的区域是奇台县，占研究区小麦种植面积的 18.66 %。春玉米种植面积最大的区域是呼图壁，占研究区春玉米种植面积的 20.59 %。棉花种植面积最大的区域是沙湾县，占研究区棉花种植面积的 31.16 %。提取的天山北坡 2016 年种植结构表明， 棉花是天山北坡经济带种植面积最大的作物， 主要分布在天山北坡的中部区域的冲积平原区及洪积冲积扇区。春玉米主要分布在天山北坡的山前丘陵区。小麦在整个天山北坡经济带的均有分布，但种植面积较少且地块较为零星破碎。研究结果具有科学意义和实用价值， 为机器学习方法及时间序列多源遥感数据在干旱区农业遥感的应用提供了参考。部分研究结果在甘家湖棉花受灾案例中得到了应用。 |
| Other Abstract | Currently, research on crop classification based on machine learning methods is ofgreat practical significance for ensuring food security and ecological security in aridregions. In recent years, remote sensing technology has made great progress in theclassification of crops and extraction of planting structures, but it still faces manydifficulties and challenges. First, the method of automatic identification of crop typesneeds to be further improved. Second, the time series multi-source data fusiontechnology needs further study.The project choose the random forest machine learning method based on theselection of time series Sentinel 2A and Landsat 8 satellite data, and combined withfield survey data, so that we can explore a multi-source remote sensing data cropclassification method for automatic extraction of planting structure of major cropsbased on the characteristics of the ground block and red edge in the northern slope ofthe Tianshan Mountains. The main contents include several poins listed in the following:the introduction of land-based primitives and red edge features, and explores the effectsof different classification feature combinations on the accuracy of machine learning;and the use of multiple-source satellite data pairs at the data level and feature levelrespectively. What effect does the crop classification accuracy have on thedetermination of the best integration level of time series Sentinel 2A data and Landsat8 data in the arid zone? The results are listed in the following.Firstly, based on the time series Sentinel 2A, eight kinds of time series spectra andvegetation index classification feature sets were used to classify and identify crops inShawan County, and the overall accuracy was over 89%. The highest accuracy of theclassification group is 94.02 %, which indicates that the random forest machine learningclassifier can effectively integrate the advantages of multidimensional vectors, and isan effective method for classifying typical crops in arid regions. After the introductionof the block-based primitives, the classification accuracy of the red-edge spectral groupand the red-edge index group increased by 3.13% and 4.07% respectively. It indicatsthat the use of the center point of the block-based block as the representative pixel ofthe homogeneous block, the extracted classification The feature can be used by themachine learning classifier to greatly improve the classification accuracy and efficiency,and to smooth the "pepper and salt noise" among the same kinds of crops within the block. The introduction of the red edge feature improved the classification accuracy ofthe spectral group and the exponent group by 2.39 % and 1.63 %, respectively, whichwas very effective in improving the accuracy of the random forest classification model.Secondly, explore the impact of the combined use of multi-source data based onthe data level and the feature level on the accuracy of crop classification. The resultsshow that the integration level of Landsat 8 and Sentinel 2A has the highest accuracyin the vegetation index. On this basis, the accuracy of remote sensing cropsclassification reached 92.92% in Wenquan county. The increase of feature-levelclassification features will promote the overall classification accuracy of random forestclassifiers and improve the accuracy of various crops.Finally, incorporating the characteristics of land parcels and red edges, based onthe characteristics of Landsat 8 and Sentinel 2A time series vegetation indices andfusion data, using the above-mentioned machine learning method based on randomforests to classify and identify the main crops on the northern slope of the TianshanMountains. The overall classification accuracy reached 81.51%, which is in line withthe needs of large-scale remote sensing monitoring applications. In the 2016 TianshanNorth Slope Research Area (including the Bingtuan), the largest wheat planting areawas in Qitai County, accounting for 18.66% of the wheat planted area in the study area.The area with the largest spring maize planting area is Hutubi, which accounts for 20.59%of the planting area of spring maize in the study area. The area with the largest cottonplanting area is Shawan County, accounting for 31.16% of the cotton planted area inthe study area. The extracted planting structure of the northern slope of the TianshanMountains in 2016 shows that cotton is the largest crop planted in the economic belt onthe northern slope of the Tianshan Mountains. It is mainly distributed in the alluvialplains and alluvial sectors in the central area of the northern slope of the TianshanMountains. Spring corn is mainly distributed in the hilly area on the northern slope ofthe Tianshan Mountains. Wheat is distributed throughout the northern slope of theTianshan Mountains, but the planting area is small and the land is broken up.The research results have scientific significance and practical value. It provides areference for machine learning methods and time series multi-source remote sensingdata for agricultural remote sensing applications in arid regions. Some research resultshave been applied in the case of cotton disaster in Ganjia Lake. |
| Subject Area | 测绘工程 |
| Language | 中文 |
| Document Type | 学位论文 |
| Identifier | http://ir.xjlas.org/handle/365004/14969 |
| Collection | 研究系统_荒漠环境研究室 |
| 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 | 黄双燕. 时间序列卫星数据的农作物分类研究及应用——以天山北坡为例[D]. 新疆乌鲁木齐. 中国科学院大学,2018. |
| Files in This Item: | There are no files associated with this item. | |||||
Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.
Edit Comment