Input selection and data-driven model performance optimization to predict the Standardized Precipitation and Evaporation Index in a drought-prone region

Accurate predictions of drought events to plan and manage the adverse effects of drought on agriculture and the environment requires tools that precisely predict standardized drought metrics. Improving on the World Meteorological Organization approved Standardized Precipitation Index (SPI), the multi-scalar Standardized Precipitation and Evapotranspiration Index (SPED, a variant of the SPI, is a relatively recent drought index, which takes into account the impacts of temperature change on overall dryness, along with precipitation and evapotranspiration effects. In this paper, an extreme learning machine (ELM) model was applied to predict SPEI in a drought-prone region in eastern Australia, and the quality of the model's performance was compared to that of a multiple linear regression (MLR), an artificial neural network (ANN), and a least support vector regression (LSSVR) model. The SPEI data were derived from climatic variables recorded at six weather stations between January 1915 and December 2012. Model performance was evaluated by means of the normalized root mean square error (NRMSE), normalized mean absolute error (NMAE), coefficients of determination (r(2)), and the Nash-Sutcliffe efficiency coefficient (NASH) in the testing period. Results showed that the ELM and ANN models outperformed the MLR and LSSVR models, and all four models revealed a greater predictive accuracy for the 12 month compared to the 3-month SPEI predictions. For the 12-month SPEI predictions, optimal models had r(2) that ranged from 0.668 for the LSSVR model (Station 6) to 0.894 for the ANN model (Station 4). The good agreement between observed and predicted SPEI at different locations within the study region indicated the potential of the developed models to contribute to a more thorough understanding of potential future drought-risks in eastern Australia, and their applicability to drought assessments over multiple timescales. The models and findings have useful implications for water resources assessment in drought-prone regions.