Detecting hydrological droughts in ungauged areas from remotely sensed hydro-meteorological variables using rule-based models

Abstract

As a method of detecting hydrological droughts in ungauged areas, we propose rule-based models using percentiles from remotely sensed key hydro-meteorological variables. Four rule-based models of the Decision Trees, Adaptive Boosting of Decision Trees (Adaboost), Random Forest, and Extremely Randomized Trees are used for their capabilities of modeling nonlinear relationships, and their results are compared to the multiple linear regression. The temporal information of month and the percentiles of key variables of water and energy balance including precipitation, actual evapotranspiration, Normalized Difference Vegetation Index (NDVI), land surface temperature, and soil moisture are used as input variables. Drought severity values are calculated from streamflow percentiles for 3-, 6-, 9-, and 12-month time scales as an indicator for hydrological droughts. Data from six basins of the case study area are used for tuning model parameters and training, and the remaining two basins are used for final evaluation. Models with an ensemble of trees successfully detect hydrological droughts despite the limited input variables (for Adaboost, correlation coefficients ≥ 0.85, mean absolute error ≤ 0.12, root-mean-square error–observations standard deviation ratio ≤ 0.53, and larger Nash–Sutcliffe efficiency of drought severity ≥ 0.72 for the test data set). The most important variable is precipitation, followed by soil moisture (3-month time scale) or NDVI (longer time scales). Hydrological droughts in various time scales are detected in ungauged areas of the case study area. Serious droughts in early 2002, from late 2006 to mid-2007, from early 2008 to 2009, and from mid-2013 to 2017 are detected.