中国农业气象 ›› 2021, Vol. 42 ›› Issue (03): 230-242.doi: 10.3969/j.issn.1000-6362.2021.03.007

• 农业气象信息技术 栏目 • 上一篇    下一篇

基于机器学习融合多源遥感数据模拟SPEI监测山东干旱

杨晋云,张莎,白雲,黄安齐,张佳华   

  1. 1.青岛大学计算机科学技术学院,青岛 266071;2.中国科学院遥感与数字地球研究所,北京 100094
  • 收稿日期:2020-09-11 出版日期:2021-03-20 发布日期:2021-03-20
  • 通讯作者: 张佳华,研究员,主要研究卫星遥感及其在气象、灾害、环境与生态中的应用、全球变化的区域响应、城市生态遥感等,E-mail:zhangjh@radi.ac.cn E-mail:zhangjh@radi.ac.cn
  • 作者简介:杨晋云,E-mail:yjy5144@163.com
  • 基金资助:
    中国科学院战略先导项目(XDA19030402);国家自然科学基金(41871253);山东省基础研究计划(2018GNC110025)

SPEI Simulation for Monitoring Drought Based Machine Learning Integrating Multi-Source Remote Sensing Data in Shandong

YANG Jin-yun, ZHANG Sha, BAI Yun, HUANG An-qi, ZHANG Jia-hua   

  1. 1.College of Computer Science and Technology, Qingdao University, Qingdao 266071,China; 2.Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Beijing 100094
  • Received:2020-09-11 Online:2021-03-20 Published:2021-03-20

摘要: 以山东省为研究区,选择偏差校正随机森林BRF(Bias-corrected random forest),支持向量回归SVR(Support vector regression)和Cubist模型三种机器学习方法融合多影响因子模拟3个月时间尺度的标准化降水蒸散指数SPEI-3,以期为精确监测山东地区干旱提供一种方法。将2001−2017年23个站点的SPEI-3值作为因变量,多源遥感数据包括降水量、地表温度、蒸散发、潜在蒸散发、归一化植被指数以及土壤湿度六类7个影响因子作为自变量,自变量和因变量构成数据集的80%作为训练集,20%作为测试集。根据BRF模型得到研究区各个站点的模拟值以及各影响因子的相对重要性,绘制SPEI-3的空间分布图,并进行验证。结果表明,综合因子比单一因子模拟效果好,BRF模型测试集中的模拟值和观测值的决定系数R2达到了0.856,均方根误差RMSE为0.359,BRF模型能较好模拟站点SPEI-3值。大部分站点模拟值与观测值反映的干旱趋势一致,反映站点不同程度旱情的月份个数基本相同。此外,BRF模型模拟的SPEI-3的空间分布与站点SPEI-3观测值表现的干旱程度基本一致,且SPEI-3空间分布站点之外栅格数据也可以较准确地反映旱情,说明根据BRF模型可在站点和空间尺度上较精确地监测山东地区干旱情况。

关键词: 机器学习, 多源遥感, 干旱, 标准化降水蒸散指数, 山东省

Abstract: In this study, Shandong province was taken as the research area, and three machine learning methods, namely Bias corrected Random Forest(BRF), Support Vector Regression(SVR) and Cubist model were selected to integrate multiple impact factors to simulate the standardized precipitation evapotranspiration index on a three-month time scale(SPEI-3), so as to provide a method for accurate monitoring of drought in Shandong province. SPEI-3 values of 23 stations from 2001 to 2017 were taken as dependent variables. Multi-source remote sensing data including precipitation, surface temperature, evapotranspiration, potential evapotranspiration, normalized difference vegetation index and soil moisture were taken as independent variables. Independent variables and dependent variables constituted 80% of the data set as training set and 20% as test set. According to the BRF model, the simulated values of each site in the study area and the relative importance of each impact factor were obtained. The spatial distribution diagram of SPEI-3 was drawn and verify them. The results showed that the multiple factor simulation was more effective than the single factor. The R2 of the simulated value and observed value in the BRF model’s test set reached 0.856, and the RMSE of the root mean square error was 0.359. The BRF model can well simulate the SPEI-3’s values of the sites. Simulated and observed values for most sites are consistent with the drought trend, and reflect the number of months of different drought conditions is basically the same. The spatial distribution of SPEI-3 simulated by the BRF model is basically consistent with the drought degree of observed SPEI-3 at the site, and SPEI-3 spatial distribution of raster data outside the sites can also reflect the drought situation more accurately. According to the BRF model, the drought situation in Shandong province can be monitored more accurately at the site and spatial scale.

Key words: Machine learning, Multi-source remote sensing, Drought, Standardized precipitation evapotranspiration index, Shandong province