干旱区研究 ›› 2021, Vol. 38 ›› Issue (6): 1637-1649.doi: 10.13866/j.azr.2021.06.15

• 天气与气候 • 上一篇    下一篇

基于Catboost的AMSR-2半经验地表温度降尺度

李永康1,2(),王新军1(),马燕飞3,胡贵锋1,桂海月1,张冠宏1   

  1. 1.新疆农业大学草业与环境科学学院,新疆土壤与植物生态过程实验室,新疆 乌鲁木齐 830052
    2.中国科学院空天信息创新研究院,数字地球重点实验室,北京 100094
    3.邯郸学院经济管理学院地理系,河北 邯郸 056005
  • 收稿日期:2020-12-10 修回日期:2021-05-05 出版日期:2021-11-15 发布日期:2021-11-29
  • 通讯作者: 王新军
  • 作者简介:李永康(1995-),男,硕士研究生,主要从事陆面数据同化研究. E-mail: yongkang_xau@163.com
  • 基金资助:
    国家自然科学基金项目(41761085);国家自然科学基金项目(41301205);自治区研究生科研创新项目(lXJ2020G152)

Downscaling land surface temperature through AMSR-2 passive microwave observations by Catboost semiempirical algorithms

LI Yongkang1,2(),WANG Xinjun1(),MA Yanfei3,HU Guifeng1,GUI Haiyue1,ZHANG Guanhong1   

  1. 1. College of Grassland and Environment Sciences of Xinjiang Agricultural University, Xinjiang Key Laboratory of Soil and Plant Ecological Processes, Urumqi 830052, Xinjiang, China
    2. Key Laboratory of Digital Earth Science,Aerospace Information Research Institute, Chinese Academy of Sciences,Beijing 100094, China
    3. Department of Geography of Handan University, Handan 056005, Hebei, China
  • Received:2020-12-10 Revised:2021-05-05 Online:2021-11-15 Published:2021-11-29
  • Contact: Xinjun WANG

摘要:

以古尔班通古特沙漠为研究对象,采用2019年AMSR-2的4通道被动微波亮温及MODIS植被指数,探索Catboost算法对被动微波地表温度空间降尺度的可行性,填补古尔班通古特沙漠MYD11A1的缺失像元,为获取全天候日、夜间多层土壤温度提供数据参考。结果表明:(1) 古尔班通古特沙漠日、夜间用于降尺度研究的特征向量(23.8 GHz V、36.5 GHz V、18.7 GHz H、89 GHz V、36.5-23.8 GHz V、36.5 V-18.7 GHz H、EVI、NDVI)与地表温度的相关性空间分异特征明显,显示沙漠相关性高,绿洲相关性低,且日间分异性更强;而盐矿覆盖会降低微波与地表温度相关性。(2) Catboost 4通道模型建立的被动微波亮温与地表温度间映射关系表现出鲁棒性。降尺度结果精度较高,日夜间R2分别为0.977、0.980,RMSE分别为3.69 K、2.38 K,MAE分别为2.71 K、1.70 K。(3) 单通道相关性统计结果与重要性分析结果存在差异,表明特征因子相关性统计结果不能直接作为Catboost被动微波反演地表温度特征选择依据。(4) 降尺度结果与站点6层土壤温度均呈极显著正相关,并随着深度的增加相关系数r总体呈下降趋势,RMSE呈增加趋势。

关键词: 地表温度, AMSR-2, Catboost, 降尺度, 古尔班通古特沙漠

Abstract:

This study aimed to fill in the missing pixels of MYD11A1 in the Gurbantunggut Desert and provide a theoretical basis for obtaining all-weather and multilayer soil temperatures during daytime and nighttime. We explored the feasibility of the CatBoost algorithm for the spatial downscaling of passive microwave surface temperature by using the four-channel passive microwave brightness temperature and MODIS vegetation index of the 2019 AMSR-2. Results show that (1) the spatial differentiation between feature vectors and surface temperature is evident during daytime and nighttime in the Gurbantunggut Desert. It indicates high desert correlation, low oasis correlation, and strong daytime differentiation. The salt mine coverage reduces the correlation between passive microwave brightness temperature and surface temperature. (2) The mapping relation between the passive microwave brightness temperature and surface temperature from the four-channel CatBoost model is robust. The accuracy of the downscaled results is high, with daytime-nighttime R2 of 0.977 and 0.980, RMSE of 3.69 and 2.38 K, and MAE of 2.71 K and 1.70 K, respectively. (3) Single-channel correlation and importance analysis results are different, suggesting that feature correlation results cannot be directly used as a basis for selecting surface temperature features with the CatBoost passive microwave inversion of LST. (4) Downscaling LST is significantly and positively correlated with the soil temperature in six layers from the Fukang site. The correlation coefficient decreases, and RMSE increases as depth increases.

Key words: land surface temperature, AMSR-2, Catboost, downscale, Gurbantunggut Desert