资源科学 ›› 2020, Vol. 42 ›› Issue (10): 1998-2009.doi: 10.18402/resci.2020.10.16

• 实证与应用 • 上一篇    下一篇

基于机器学习的日尺度短波净辐射气候资源遥感反演研究

应王敏1(), 刘晓洁1, 房世峰1, 李秀娟1, 赖明2, 张旭振3, 吴骅1,4()   

  1. 1. 中国科学院地理科学与资源研究所,北京 100101
    2. 中国地质大学(武汉)资源学院,武汉 430000
    3. 中国地质调查局烟台海岸带地质调查中心,烟台 264004
    4. 中国科学院大学资源与环境学院,北京 100049
  • 收稿日期:2020-08-25 修回日期:2020-10-13 出版日期:2020-10-25 发布日期:2020-12-25
  • 通讯作者: 吴骅
  • 作者简介:应王敏,男,浙江台州人,硕士,主要研究方向为地表短波净辐射遥感反演、机器学习在地理数据上的应用。E-mail: qsy@zju.edu.cn
  • 基金资助:
    全国自然资源要素综合观测体系规划与部署项目(DD20208063);黄河流域自然资源要素综合观测试点项目(DD20208066);自然资源要素综合观测数据集成与应用服务项目(DD20208067)

Retrieval of daily net surface shortwave radiation climatic resources based on machine learning

YING Wangmin1(), LIU Xiaojie1, FANG Shifeng1, LI Xiujuan1, LAI Ming2, ZHANG Xuzhen3, WU Hua1,4()   

  1. 1. Institute of Geographic Sciences and Natural Resources Research, CAS, Beijing 100101, China
    2. School of Earth Resources, China University of Geosciences (Wuhan), Wuhan 430000, China
    3. Yantai Coastal Zone Geological Survey Center, China Geological Survey, Yantai 264004, China
    4. College of Resources and Environment, University of Chinese Academy of Sciences, Beijing 100049, China
  • Received:2020-08-25 Revised:2020-10-13 Online:2020-10-25 Published:2020-12-25
  • Contact: WU Hua

摘要:

日尺度地表短波净辐射(DNSSR)是大部分陆面过程模型、全球环流模型、陆-气交换过程模型和各种水文模型的重要输入参数,在自然资源调查、生态环境监测、能量平衡研究等领域具有重要的研究意义和实用价值。本文通过匹配MODIS双星遥感观测和FLUXNET日尺度地面观测数据,筛选出包含18个自变量总计15531对的有效样本,利用机器学习的随机森林方法构建了日尺度DNSSR遥感反演模型,并利用地面实测数据对模型结果进行了真实性检验。结果表明,构建的日尺度DNSSR遥感模型的偏差Bias为-0.1 W/m2,均方根误差RMSE为27.8 W/m2,决定系数R2为0.90,表现出良好的精度。基于此过程,得到MODIS双星反演的DNSSR全球分布结果,并与不同季节下再分析ERA5数据扩展得到的DNSSR数据进行了对比,发现两者全球分布特征基本一致,且均与太阳能量随季节变化分布特点密切相关。为进一步证实验证的结果,将ERA5和地面站点实测数据作了进一步的对比,结果从侧面证实了本文构建的MODIS的DNSSR产品精度远高于ERA5的DNSSR,而且其产品空间分辨率也有了极大提升。研究结果证明,本文提出的基于MODIS双星观测与机器学习的日尺度DNSSR反演模型具有反演精度高、空间分辨率高、具备时间连续性等优点,能够有效移植至其他气候资源的遥感反演。

关键词: 气候资源, 日尺度地表短波净辐射(DNSSR), 机器学习, 遥感, 反演

Abstract:

Daily net surface shortwave radiation (DNSSR) is one of the most important parameters in various global land process and hydrological models and is required in climate change, energy balance, ecological, and atmospheric circulation research. This study constructed a daily net surface shortwave radiation model using the random forest (RF) method and MODIS twin-satellite products. 15531 pairs of samples containing 18 independent variables were extracted by matching MODIS twin-satellite products and FLUXNET daily observations. The Bias, RMSE (root mean square error), and R2 for the proposed DNSSR model using the RF method are -0.1W/m2, 27.8 W/m2, and 0.90, respectively. Based on the process, MODIS-DNSSR global distribution in different seasons were presented. Verification with field observations shows that the results are similar to the ERA5 reanalysis data, which are closely related to the seasonal distribution of solar energy. To further verify the results, ERA5-DNSSR were compared with the FLUXNET-DNSSR. The result shows that the proposed DNSSR model has also better accuracy and higher resolution than the ERA5 data. The RF-based DNSSR model has a good retrieval accuracy, high spatial resolution, and good temporal continuity. It can be effectively transplanted to the retrieval of other climatic resources.

Key words: climatic resources, daily net surface shortwave radiation (DNSSR), machine learning, remote sensing, retrieval