基于 GEE 云平台的农作物种植结构以及覆膜农田的遥感提取
Alternative TitleCropping Structure Extracting and Plastic-mulched Farmland Mapping with Multi-Source Remote Sensing Data Based on the Google Earth Engine Cloud Platform
Thesis Advisor张清凌
Degree Grantor中国科学院大学
Place of Conferral北京
Degree Discipline理学硕士
Keyword农业种植结构提取 覆膜农田提取 随机森林分类 决策树分类 绿洲农业 Crop Structure Extraction, Plastic-mulched Farmland Random Forest Classification Decision tree Classification Oasis Agriculture
Abstract水资源匮乏是干旱区实现可持续发展的最大障碍。干旱区农业灌溉耗费大量的水资源,而不同农作物在生长期所需的灌溉水量存在较大的差异,因此快速、准确的了解干旱区的农作物种植结构可以为干旱区的水资源优化提供重要依据。另一方面, 由于农业地膜具有增加农作物产量以及增温保墒的功能, 其已被广泛地应用于干旱区的农作物栽培。在我国,农业地膜的使用量以平均每年 6%的速度增长,而新疆则是我国农业地膜使用范围最广的地区。随着农业地膜的长期使用以及回收不力等因素的影响,新疆已成为全国农业地膜碎片残留最为严重的地区, 地膜残留污染已经给该地区的生态环境等方面带来了一系列的负面影响。因此, 为实现新疆地区的可持续发展, 对该地区的农业地膜使用情况的动态监测迫在眉睫。 在此基础上, 本文以新疆天山北坡经济带为研究区,以谷歌地球引擎(Google Earth Engine, GEE)云平台为支撑,以 Sentinel-2 以及 Landsat 7-8 等遥感数据为数据源,采取以下步骤进行研究区的农作物种植结构以及覆膜农田的遥感提取:(1)农作物种植结构的遥感提取:第一,为了简化农作物种植结构提取过程,本文利用一年最大 NDVI 值以及坡度信息构建了耕地掩膜图层;第二,根据研究区内主要农作物的物候历,获取 2018 年 3 月 1 日至 2018 年 11 月 30 日期间每月最大 NDVI 值的时间序列以及农作物在一年中出现 NDVI 最大值的日期数据,并在此基础上构建一个 10 波段的特征波段影像;第三,结合野外实地考察获得的有效样本点以及经耕地掩膜图层掩膜后的 10 波段的特征波段影像,利用随机森林分类器进行研究区的农作物种植结构提取;第四, 在 GEE 云平台上编写农作物种植结构提取算法;最后,对农作物种植结构提取结果进行众数滤波(4×4 pixels)处理,同时利用相关的有效验证样本点进行精度验证。 分类结果表明: 2018 年研究区内棉花、玉米、小麦等的总体分类精度为 92.19%, Kappa 系数为 0.883;为了进一步与统计数据进行对比,本文同时提取了研究区 2017 年的农作物种植结构,其分类结果表明 2017 年研究区内棉花、玉米、小麦的种植面积分别为 5270km2、2000km2、2340km2。与统计数据相比,其相对精度分别为 86.53%、77.54%、86.19%。(2)覆膜农田的遥感提取:第一,我们在 USGS 光谱数据库中找到了研究区内主要地物(植被、地膜、裸土、不透水面层、水体等)的光谱特征曲线,同时结合野外实地考察获得的有效样本点的光谱等信息,在此基础上,本文选择了光谱特征、时相特征以及辅助特征作为覆膜农田提取的分类特征;第二,根据研究区内覆膜农作物的物候历,本文将每年 4 月至 5 月以及 7 月至 8 月这个两个时间段作为覆膜农田识别的关键期,并将 Sentinel-2 和 Landsat 7-8 的遥感数据作为数据源,生成研究区无云的 5 波段特征波段影像;第三,我们将获得的有效训练样本点利用回归决策树进行决策树分类,进而获取覆膜农田识别规则;第四,在 GEE 云平台上编写覆膜农田识别算法;最后,我们对覆膜农田提取结果进行众数滤波(4×4 pixels)处理,同时利用相关的有效验证样本点进行精度验证。分类结果表明:2016 年研究区内覆膜农田提取的总体精度为 88.94%,F-score 为 0.925。为了进一步与统计数据进行对比,本文同时进行研究区内 2015 年的覆膜农田提取,分类结果表明,2016年以及 2015年研究区内地膜覆盖的面积分别为 8256km2、7791km2。 与统计数据相比,其覆膜农田提取的相对精度分别为 86.11%、 91.7%。
Other AbstractLimited water resource is the major factor affecting sustainable development inarid areas, and most of it are used for agricultural. At the same time, the amount ofirrigation water required for different crops in the growing season vary significantly.Therefore, rapid and accurate mapping of cropping structure in arid areas can providean important basis for optimizing water use in agriculture. At the same time, accordingto reports, plastic mulch area has been expanded at an average rate of 6% per year inChina over the last decade, because plastic mulch can significantly increase the yieldof crops and increase temperature and conserve soil moisture. Xinjiang is now theregion with the largest area of plastic mulch coverage in China. Due to years of residualplastic film accumulation in soils and insufficient recycling, Xinjiang is aslo the regionwith the largest amount of plastic mulch residue in China, which has been leading to aseries of negative impacts on regional climate and eco-environment. Therefore, thelong-term monitoring of the use of the plastic mulch in Xinjiang is conducive to thesustainable development of this region.This paper chose the Northern Tianshan Economic Belt as the study area, proposinga method to map the cropping structure and the spatial distribution of plastic mulch inthis region with multi-source remote sensing data based on the Google Earth Engine(GEE) cloud platform. The main contents are as the following:(1) Crop planting structure extraction with remote sensing data: First, in order tosimplify the cropping structure extraction process and minizing impacts from non-cropvegetation, a cropland mask is generated with slope information calculated from DEMdata and the maximum NDVI value calcualted from time series satellite data throughoutone entire year in the study area. Second, according to the phenology calendars of majorcrops in the study area, time-series of the monthly maximum NDVI value during theperiod of March 1, 2018 to November 30, 2018 and the corresponding date when theyearly maximum NDVI reached were calculated with remote sensing data, which form10 feature bands for further analysis. Third, we masked the above 10 feature bands with the cropland mask. Based on these data together with field samples, we trained arandom forest classifier for cropping structure extraction. Fourth, the crop plantingstructure extraction algorithm was implemented on the GEE cloud platform. Finally,we performed the majority filtering (4×4 pixels) on the crop planting structureextraction results, and then accuracy assessment. The accuracy assessment resultsshowed the overall accuracy of the classification results was 92.19%, and the Kappacoefficient was 0.883. In order to further verify the accuracy of the classificationalgorithm, the cropping structure in the study area for 2017 also was extracted, and wasthen compared with the results of statistical yearbook data in 2017. The classificationresults showed that the planted area of cotton, corn and wheat in the study area was5270km2, 2000km2, and 2340km2 respectively in 2017. The relative accuracy of cotton,corn and wheat planted area were 86.53%, 77.54% and 86.19%, respectively.(2) Plastic-mulched farmland extraction with remote sensing: First, the spectralcharacteristics of main objects (vegetation cover, plastic mulch, bare soil, impervioussurface, water body, etc.) in the study area were extracted in the USGS spectral database,and then were combined with the spectral, temporal and auxiliary features as theclassification features. Second, according to the phenology of plastic mulch crops inthe study area, the two periods from April to May and July to August were identified asthe critical periods for the identification of plastic-mulched farmland, and then a cloudfree 5-band image of the study area was generated. Third, the samples extracted fromhigh-resolution imagery were fed to a previously trained decision tree. Fourth, theplastic-mulched farmland extraction algorithm was implemented on the GEE cloudplatform. Finally, we performed the majority filtering (4×4 pixels) on the plasticmulched farmland extraction results, and then accuracy assessment. The accuracyassessment results showed the overall accuracy of the classification results was 88.94%,and the F-score was 0.883. In order to further verify the accuracy of the classificationalgorithm, the plastic-mulched farmland in the study area for 2015 also was extracted,and was then compared with the results of statistical yearbook data in 2015 and 2016.The classification results showed that the plastic-mulched farmland in the study areawas 8256km2, 7791km2 respectively in 2016 and 2015. The relative accuracy of 2016 and 2015 were 86.11% and 91.7%, respectively.
Subject Area地图学与地理信息系统
Document Type学位论文
First Author Affilication中国科学院新疆生态与地理研究所
Recommended Citation
GB/T 7714
熊元康. 基于 GEE 云平台的农作物种植结构以及覆膜农田的遥感提取[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
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[熊元康]'s Articles
Baidu academic
Similar articles in Baidu academic
[熊元康]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[熊元康]'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.