地球信息科学学报 ›› 2017, Vol. 19 ›› Issue (1): 125-133.doi: 10.3724/SP.J.1047.2017.00125

• 遥感科学与应用技术 • 上一篇    下一篇

基于随机森林的遥感干旱监测模型的构建

沈润平(), 郭佳, 张婧娴, 李洛晞   

  1. 南京信息工程大学地理与遥感学院,南京 210044
  • 收稿日期:2015-11-11 修回日期:2016-01-08 出版日期:2017-01-20 发布日期:2017-01-13
  • 作者简介:

    作者简介:沈润平(1963-),男,教授,研究方向为遥感建模与分析。E-mail: rpshen@nuist. edu. cn

  • 基金资助:
    国家自然科学重点基金项目“青藏高原陆面再分析关键技术及数据集”(91437220)

Construction of a Drought Monitoring Model Using the Random Forest Based Remote Sensing

SHEN Runping*(), GUO Jia, ZHANG Jingxian, LI Luoxi   

  1. School of Geography and Remote Sensing, Nangjing University of Information Science and Technology, Nangjing 210044, China
  • Received:2015-11-11 Revised:2016-01-08 Online:2017-01-20 Published:2017-01-13
  • Contact: SHEN Runping E-mail:rpshen@nuist.edu.cn

摘要:

利用遥感数据进行大面积旱情监测是现有干旱监测的重要方法之一,然而传统的遥感干旱监测方法主要侧重于对土壤湿度或植被状况等单一干旱响应因子进行监测,对综合多因子的干旱监测研究较为有限。随机森林是一种机器学习方法,具有学习过程快速、运算速度快、稳定性好、预测精度高的优点,近年来被应用于生态环境等多个领域。本文利用2001-2010年4-9月的MODIS数据提取的植被状态指数(VCI)、温度状态指数(TCI)和土地覆盖类型(LC),TRMM降水资料计算的TRMM-Z指数及SRTM-DEM、土壤有效含水量(AWC)等多个遥感及土壤资料提取的干旱因子为自变量,以气象站点的综合气象干旱指数(CI)为因变量,利用随机森林模型构建遥感干旱监测模型,并以河南省为研究区进行了评价和分析。该模型在2009-2010年的监测值和实测CI值的具有显著的相关性,并且二者干旱等级的一致率为81%。在2001-2010年4-9月间,模型监测值与气象站点的标准降水蒸散发指数(SPEI)总体干旱等级一致率为74.9%,较为一致,其中9月的模型结果与SPEI的干旱等级一致率最高,达到82.4%,空评估率和漏评估率最低;与10 cm土壤相对湿度的相关系数在0.475-0.639之间,达到极显著水平。河南省2011年4-6月干旱事件同样验证了本文构建的模型旱情监测结果,说明本模型能较好地就应用于监测区域旱情监测。

关键词: 干旱, 遥感, 随机森林

Abstract:

The drought detection of a large area by using the remote sensing data has been an important method in drought monitoring. However, the conventional remote sensing methods mainly focus on some single drought response factors, such as the soil moisture or vegetation status, and the drought monitoring study that integrated with multiple factors is relatively limited. In order to explore the relationships among multiple drought factors, a random forest algorithm was applied. Random forest is a machine learning method, which has many advantages such as being accurate, handy, fast and stable, and it has been used in many fields in recent years. In this paper, a remote sensing drought model was developed using the random forest algorithm and the multi-source remote sensing data, including MODIS, TRMM and SRTM-DEM. Vegetation Condition Index (VCI), Temperature Condition Index (TCI), Land Cover types (LC), TRMM-Z, DEM and Available Water Capacity (AWC), which were extracted from the remote sensing data and other soil data, were used as the independent variables, and the comprehensive meteorological drought index (CI) was used as the dependent variable. The training and testing experiments were carried out in Henan Province from April to September annually between 2001 and 2010. The results showed that the model and CI had highly significant correlation and their concordance rate reached 81% with respect to the drought classes from 2009 to 2010. In the study case′s period, the overall concordance rate was 74.9% between the model results and the Standardized Precipitation Evapotranspiration Index (SPEI) of the meteorological stations, from which the concordance was found to be the highest and the vacancy and miss rate was the lowest in September. The correlation between the model results and the soil moisture within 10 cm depth was highly significant, and their correlation coefficient varied between 0.475 and 0.639, which indicated that this model could effectively detect the agriculture drought. In addition, the drought event of Henan Province from April to June in 2011 was simulated by the proposed model, and the results could reflect the actual drought situation and its spatial variation. Therefore, this method could be well applied to monitor regional drought events.

Key words: drought, remote sensing, random forest