Drought modeling: a comparative study between time series and neuro-fuzzy approaches

Meteorological drought is one of the inseparable climatic phenomena in sub-tropical countries such as Iran. In these areas, which encompass the vastest deserts of the world, the effects of precipitation scarcity on water resources manifest themselves promptly. This study employed the standardized precipitation index (SPI), as a meteorological drought assessment tool, over 3- and 12-month time scales during the years 1970 to 2014. We compared the accuracy of the neuro-fuzzy model (as a non-linear model) with time-series models for modeling of drought. Time-series analysis was conducted according to the Box–Jenkins method. ARIMA (3, 0, 4) and ARIMA (2, 0, 1) were selected as the best-fitting time-series models for modeling SPI at time scales of 3 and 12 months, respectively. The results indicated that the neuro-fuzzy model significantly outperforms the time-series models. The Nash–Sutcliffe efficiency (NSE) coefficients are equal to 0.12 and 0.60 respectively for SPI3 and SPI12 estimated by ARIMA model, while NSE coefficients for neuro-fuzzy model are equal to 0.52 and 0.80 respectively for SPI3 and SPI12 in validation period. Also, the violin plots demonstrated that the neuro-fuzzy model (unlike the ARIMA model) is well-suited to estimate the volatility of SPI values for wet and dry periods, which is a very important prerequisite for efficient water resources’ management.