干旱气象 ›› 2025, Vol. 43 ›› Issue (6): 997-1005.DOI: 10.11755/j.issn.1006-7639-2025-06-0997

• 技术报告 • 上一篇    

基于集成学习技术的青海高原地表温度反演研究

辛萍萍1,2(), 陈国茜1,3(), 周生蓉4, 文生祥1,5, 程俊清2, 田媛2   

  1. 1.青海省防灾减灾重点实验室青海 西宁 810001
    2.青海省大柴旦行政委员会气象局青海 海西 817000
    3.青海省气象科学研究所青海 西宁 810001
    4.青海省果洛州气象局青海 果洛 814000
    5.青海省诺木洪气象站青海 诺木洪 816102
  • 收稿日期:2024-08-29 修回日期:2024-11-28 出版日期:2025-12-31 发布日期:2026-01-19
  • 通讯作者: 陈国茜(1986—),女,广西北海人,硕士,正高级工程师,主要从事高寒生态遥感监测评估技术研发与业务应用。E-mail: 71153087@qq.com
  • 作者简介:辛萍萍(1993—),女,青海平安人,工程师,主要从事县级综合气象业务工作。E-mail: 2287485706@qq.com
  • 基金资助:
    青海省防灾减灾重点实验室开放基金项目(QFZ-2024-M20)

Research on land surface temperature retrieval over the Qinghai Plateau based on ensemble leaning technology

XIN Pingping1,2(), CHEN Guoqian1,3(), ZHOU Shengrong4, WEN Shengxiang1,5, CHENG Junqing2, TIAN Yuan2   

  1. 1. Key Laboratory of Disaster Prevention and Mitigation of Qinghai ProvinceXining 810001, China
    2. Dachaidan Administrative Committee Meteorological Bureau of Qinghai ProvinceHaixi 817000, Qinghai, China
    3. Qinghai Institute of Meteorological SciencesXining 810001, China
    4. Guoluo Tibetan Autonomous Prefecture Meteorological Bureau of Qinghai ProvinceGuoluo 814000, Qinghai, China
    5. Nuomuhong Meteorological Station of Qinghai ProvinceNuomuhong 816102, Qinghai, China
  • Received:2024-08-29 Revised:2024-11-28 Online:2025-12-31 Published:2026-01-19

摘要:

地表温度作为地表物理过程的重要参数,在有云情况下,微波遥感是获取其信息的主要途径之一。本文基于FY-3D/MWRI亮温数据,采用随机森林(Random Forest,RF)、XGBoost、LightGBM等集成学习算法构建青海高原地表温度估算模型。通过深入分析亮温数据、地形要素等特征在不同模型中的贡献度,发现RF模型中亮温数据贡献度较高,XGBoost模型中亮温数据和地形要素均表现出较高贡献度,LightGBM模型则地形要素贡献度较高。3个模型的训练和测试精度均超过0.8,其中LightGBM与XGBoost模型的训练、测试精度差异较小。3个模型均能较好地刻画青海高原2个低温区和2个高温区的空间分布特征,其中XGBoost模型在干旱和洪涝过程下表现尤为突出,其估算结果与地表温度实测值的偏差最小。综合表明,XGBoost模型为青海高原地表温度估算的最优模型,其反演结果为该区域开展干旱与洪涝遥感实时监测提供技术支持。

关键词: 地表温度, 被动微波遥感, 随机森林, XGBoost, LightGBM

Abstract:

Land surface temperature is an important parameter of surface physical processes. Under cloudy conditions, microwave remote sensing is one of the main methods to obtain its information. Based on the FY-3D/MWRI brightness temperature data, this paper builds land surface temperature retrieval models over the Qinghai Plateau using ensemble learning algorithms such as Random Forest (RF), XGBoost, and LightGBM. Through in-depth analysis of the contribution of brightness temperature data, terrain factors, and other features in different models, it is found that the contribution of brightness temperature data is relatively high in the RF model, both brightness temperature data and terrain factors show high contributions in the XGBoost model, and the contribution of terrain factors is relatively high in the LightGBM model. The training and testing accuracies of the three models all exceed 0.8, among which the training and testing accuracies of the LightGBM and XGBoost models are relatively small. All three models can well depict the spatial distribution characteristics of the two low-temperature and two high-temperature areas over the Qinghai Plateau. Among them, the XGBoost model performs more prominently in drought and flood processes, and its estimation results have the smallest deviation from the measured Land surface temperature values. In conclusion, the XGBoost model is the optimal model for land surface temperature retrieval over the Qinghai Plateau, and its inversion results provide technical support for real-time remote sensing monitoring of drought and flood in this region.

Key words: land surface temperature, passive microwave remote sensing, Random Forest, XGBoost, LightGBM

中图分类号: