Modelling and forecasting of drought index using machine learning techniques

Abstract

A reliable and computationally efficient drought model is a useful tool for water resources management since a prior knowledge of drought-risk depends on our ability to accurately forecast its future occurrence. The Inter-governmental Panel on Climate Change Report on extreme events has recognized drought as an extreme climatic event that needs to be mitigated to reduce its negative effects. Therefore, fast, accurate and reliable drought forecasting models that provide lead-time information on the future drought-risk is a useful tool for drought management. Droughts pose risks to economic areas (e.g., agriculture), and so must be defined by the deficiency of water resources rather than a rainfall deficiency. Therefore, forecasting of the DIs based on the concept of water resources (e.g., the effective DI, EDI; Byun and Wilhite 1999) are useful for decision-making in the field of drought hydrology and water resource management. To extract information on drought-risk, the downscaling of rainfall from global climate models (GCMs) is performed using statistical or dynamical methods. Dynamical methods utilize a limited-area high-resolution model (e.g., regional climate model) using boundary conditions from GCMs to derive small-scale drought information while statistical models discover links between future climate and large-scale predictors. In statistical models using machine learning (ML) algorithms, the simulations are empirically calibrated from the observed data. ML is recognised as an alternative for local-scale drought forecasting. ML is less complex as it assimilates datasets to ‘learn’ from climatic trends. Other advantages include the forecasting of climate without the need for complex physical equations, easy model development, low computational cost, fast training and testing, the possibility of local applications (e.g., farms) using point-based forecasting data and their competitive performance relative to fully-dynamic models. In this study an attempt has been made to develop a reliable machine learning model for the forecasting of effective drought index