Hydrological drought class early warning using support vector machines and rough sets

Abstract

Prediction of drought severity class/state can provide useful insight into preparedness actions. In this study, hydrological drought class, as determined by the standardized hydrological drought index (SHDI), was predicted. SHDI1 and SHDI3 drought classes were determined in seven and nine drought class systems. Support vector regression (SVR), support vector classification (SVC), and rough set theory (RST) were tested and compared as forecast modelling tools. Different combination of historic monthly streamflow and precipitation time series, or values/classes of SHDI and SPI, were considered as predictors to tackle the following question; is it possible to achieve the same or better accuracy by directly predicting the drought class, instead of predicting the streamflow (or drought index values) first and subsequently estimating the qualitative severity (class) of the drought? The results were more accurate in case of considering drought classes as inputs/output. However, the number of drought/wet classes may affect prediction accuracy. Prediction in case of fewer drought/wet classes leads to more accurate results. In addition, the SHDI3 prediction was more accurate than the SHDI1 prediction. RST showed slightly better accuracy than SVC and SVR. In case of nine-class forecast, the overall accuracy of RST model in correctly predicting the exact class, or with maximum one class shift, was 90.6% for SHDI1 and 96.5% for SHDI3. These values were 92.9% for SHDI1 and 98.8% for SHDI3 in seven-class system. Overall, direct classification methods (SVC and RST), in addition to simplicity, were more successful than the indirect method (SVR) in determination of severe dry/wet classes.