引用本文:毛克彪,杨军,韩秀珍,唐世浩,袁紫晋,高春雨.基于深度动态学习神经网络和辐射传输模型地表温度反演算法研究[J].中国农业信息,2018,30(5):53-63
【打印本页】   【HTML】   【下载PDF全文】   查看/发表评论  【EndNote】   【RefMan】   【BibTex】
←前一篇|后一篇→ 过刊浏览    高级检索
本文已被:浏览 238次   下载 179 本文二维码信息
码上扫一扫!
分享到: 微信 更多
基于深度动态学习神经网络和辐射传输模型地表温度反演算法研究
毛克彪1,2,杨军3,韩秀珍3,唐世浩3,袁紫晋1,高春雨1
1.中国农业科学院农业资源与农业区划研究所,北京100081;2.湖南农业大学资源环境学院,长沙410128;3.国家气象卫星中心,北京100081
摘要:
目的 地表温度反演是一个典型的病态反演问题,深度动态学习神经网络的出现提供了一条新的地表温度反演途径。文章以MODIS中红外和热红外波段作为参照模拟研究对象,利用深度动态学习神经网络和辐射传输模型(MODTRAN)进行地表温度反演研究,选择最适合于MODIS地表温度反演的波段组合,从而为国产卫星风云系列和高分数据红外波段反演地表温度提供参考算法。方法 根据中红外波段受太阳的影响以及水汽波段的特征,将反演组合波段分成3组。第1组适合白天和晚上同时反演地表温度的组合(MODIS波段29、31、32和33);第2组适合白天的热红外波段和水汽波段组合(MODIS波段29、31、32、33和水汽波段);第3组是只适合晚上的中外波段(MODIS 20、22、23)与热红外波段(MODIS 29、31、32和33)的组合。结果 利用辐射传输模型(MODTRAN)和深度动态神经网络(NN)反演分析表明,深度动态学习神经网络能够被用来精确地从单景MODIS数据中反演地表温度,克服了传统MODIS白天/黑夜产品算法的缺陷。3种类型的组合地表温度的平均反演误差都在1 K以下,最高精度为热红外波段与水汽波段的组合,平均最高精度为0.251 K,标准差是0.255 K,相关系数是1。结论 利用深度动态学习神经网络和辐射传输模型彻底解决了地表温度和发射率病态反演难题,为风云系列卫星和高分数据地表温度反演算法提供参考算法模式,深度动态学习神经网络与辐射传输模型相结合反演地表温度和发射率在地表温度反演史上具有里程牌意义。
关键词:  深度学习神经网络;辐射传输模型;地表温度;发射率
DOI:10.12105/j.issn.1672-0423.20180506
分类号:
基金项目:国家重点研发计划课题“高时空分辨率多源卫星遥感气象灾害产品融合技术”(2018YFC1506502),国家自然科学基金项目“基于遥感研究气候变化背景下农业旱灾时空变化对粮食生产影响41571427国家重点研发计划课题“高时空分辨率多源卫星遥感气象灾害产品融合技术”(2018YFC1506502),国家自然科学基金项目“基于遥感研究气候变化背景下农业旱灾时空变化对粮食生产影响(41571427)
Retrieving land surface temperature based on deep dynamic learning NN algorithm and radiation transmission model
Mao Kebiao1,2,Yang Jun3,Han Xiuzhen3,Tang Shihao3,Yuan Zijin1,Gao Chunyu1
1.Institute of Agricultural Resources and Regional Planning,Chinese Academy of Agricultural Sciences, Beijing 100081,China;2.College of Resources & Environment,Hunan Agricultural University, Changsha 410128,China;3.National Satellite Meteorological Center,Beijing 100081,China
Abstract:
Purpose Land surface temperature retrieval is a typical ill-posed problem. The deep dynamic learning neural network provides us with a new method to retrieve land surface temperature. In this paper,the mid-infrared and thermal infrared bands of MODIS are used as reference simulation objects. The deep dynamic learning neural network and radiation transmission model (MODTRAN) are used to study the surface temperature retrieval,and select which band combination is most suitable for MODIS surface temperature retrieval.Methods According to the influence of the sun for the mid-infrared band and the characteristics of the water vapor band,the retrieval combined band is divided into three types. The first group is suitable for simultaneous retrieval of surface temperature during the day and night (MODIS bands 29,31,32 and 33);the second group is thermal infrared and water-wave band combinations (MODIS bands 29,31,32,33 and the water vapor band) which is suitable for daytime;the third group is a combination of the mid-infrared band (MODIS 20,22,23) and the thermal infrared band (MODIS 29,31,32 and 33) for the evening.Results The retrieval analysis using the radiation transfer model (MODTRAN) and deep dynamic neural network (NN) shows that the deep dynamic learning neural network can be used to accurately retrieve the surface temperature from the MODIS data. The average retrieval error of the three types of combined surface temperatures is below 1K. The highest precision is the combination of the thermal infrared band and the water vapor band. The average highest precision is 0.251 K,the standard deviation is 0.255K,and the correlation coefficient is 1.Conclusion It has the significance of mileage in the history of surface temperature retrieval when the deep dynamic learning neural network combined with radiation transmission model is used to retrieve surface temperature and emissivity.
Key words:  deep dynamic learning neural network;radiation transmission model;land surface temperature;emissivity