地球信息科学学报 ›› 2019, Vol. 21 ›› Issue (5): 752-766.doi: 10.12082/dqxxkx.2019.180420

• 遥感科学与应用技术 • 上一篇    下一篇

Google Earth Engine支持下的江苏省夏收作物遥感提取

何昭欣1,2(), 张淼1, 吴炳方1,2,*(), 邢强1   

  1. 1. 中国科学院遥感与数字地球研究所 遥感科学国家重点实验室,北京 100101
    2. 中国科学院大学,北京 100049
  • 收稿日期:2018-08-29 修回日期:2019-01-11 出版日期:2019-05-25 发布日期:2019-05-25
  • 通讯作者: 吴炳方 E-mail:hezx@radi.ac.cn;wubf@radi.ac.cn
  • 作者简介:

    作者简介:何昭欣(1991-),男,山东莒县人,硕士,主要从事农作物与果园遥感分类研究。E-mail: hezx@radi.ac.cn

  • 基金资助:
    国家重点研发计划项目(2016YFD0300608);中国科学院科技服务网络计划(STS计划)项目(KFJ-STS-ZDTP-009);国家自然科学基金项目(41561144013、41861144019、41701496)

Extraction of Summer Crop in Jiangsu based on Google Earth Engine

Zhaoxin HE1,2(), Miao ZHANG1, Bingfang WU1,2,*(), Qiang XING1   

  1. 1. State Key Laboratory of Remote Sensing Science, Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Beijing 100101, China
    2. University of Chinese Academy of Sciences, Beijing 100049, China
  • Received:2018-08-29 Revised:2019-01-11 Online:2019-05-25 Published:2019-05-25
  • Contact: Bingfang WU E-mail:hezx@radi.ac.cn;wubf@radi.ac.cn
  • Supported by:
    National Key R&D Program of China, No.2016YFD0300608;Science and Technology Service Network Initiative (STS), No.KFJ-STS-ZDTP-009;National Natural Science Foundation of China, No. 41561144013, 41861144019, 41701496.

摘要:

江苏省是农作物种植大省,国家统计局统计数据显示,江苏省近10年冬小麦、冬油菜的总播种面积分列全国第五、第七,快速准确地获取冬小麦和冬油菜的空间分布对于该省的农业发展具有重意义。基于单机的传统遥感分类能够准确获取农作物的空间分布信息,但是耗时较长。随着地理大数据与云平台、云计算的发展,Google Earth Engine(GEE)作为一个基于云平台的全球尺度地理空间分析平台,为快速遥感分类带来了新的机遇。本文基于GEE,使用Sentinel-2数据快速提取了江苏省2017年冬小麦与冬油菜的空间分布。首先,利用GEE获得覆盖江苏省119景无云质优的Sentinel-2影像;其次,在此基础上分别计算了遥感指数、纹理特征、地形特征,并完成原始特征的构建与优化;最后,分别试验了朴素贝叶斯、支持向量机、分类回归树和随机森林4种分类器,比较了各分类器的分类精度,并提取了冬小麦与冬油菜的空间分布信息。得出以下结论:①GEE能够快速完成覆盖江苏省影像数据的去云、镶嵌、裁剪及特征构建等预处理,较本地处理具有明显优势;②J-M距离值位于前两位且大于1将特征数量从28个压缩到11个,有效压缩了原始特征空间;③光谱+纹理+地形特征组合训练,朴素贝叶斯、支持向量机、分类回归树、随机森林的平均验证精度分别为61%、87%、89%、92%。

关键词: GEE云平台, Sentinel-2, 特征优选, 作物识别, 江苏省

Abstract:

Jiangsu province, with 13 municipalities and located in the east of China, is an important part of the Yangtze river delta economy belt. The temperature is appropriate and the rainfall is moderate. Jiangsu province enjoys a moderate climate, which is suitable for the agricultural development. Winter wheat is distributed throughout the whole province, whereas the planting structure of winter rapeseed is complex and mainly scattered in Southern Jiangsu. As reported by the State Statistics Bureau, the total planting area of winter wheat and winter rapeseed in Jiangsu ranked the fifth and seventh in China, respectively, during the last 10 years. Fast obtaining the precise planting area of these two crops in Jiangsu is crucial for the agricultural development. Remote sensing classification based on local host can obtain spatial distribution of crops with high accuracy, but is time-consuming. With the development of geographical big data, cloud platform, and cloud computation, the Google Earth Engine (GEE), a global scale geospatial analysis platform based on the cloud platform, has brought new opportunities for rapid remote sensing classification. Based on the GEE cloud platform, a time-saving method of obtaining the spatial distribution of winter wheat and winter rapeseed by use of sentinel-2 data in Jiangsu was proposed. First, 119 sentinel-2 images without cloud were obtained using the GEE in Jiangsu. The time interval was set from March 1 to June 1, 2017, and the space area was Jiangsu province. Based on the spatio-temporal information, the 119 remote sensing images were mosaicked and clipped. Secondly, remote sensing indices, texture, and terrain features were calculated respectively, and the original features were extracted. The original feature space was optimized by an algorithm named Separability and Thresholds (SEaTH algorithm). Finally, four classifiers including naive Bayes, support vector machine, classification regression tree, and random forest were tested and evaluated by the average assessment accuracy. The spatial distribution information of winter wheat and winter rapeseed were obtained quickly. The following conclusions are drawn: (1) the GEE can quickly complete pre-processing of cloud-masking, image-mosaicking, image-clipping, and feature extraction, which is superior to the local processing. (2) The distance values of J-M that are higher than 1 and rank top two highest can reduce the number of features from 28 to 11 and effectively compress the original feature space. (3) With the combined training of spectral, texture and terrain features, the average assessment accuracy of naive Bayes, support vector machine, classification regression tree, and random forest was 61%, 87%, 89% and 92%, respectively.

Key words: GEE cloud platform, Sentinel-2, feature optimization, crop identification, Jiangsu Province