| |
Drought forecasting can help set out mitigation strategies and minimize losses. Because a drought index is an indicator of drought, the latter may be forecast by forecasting the former. We investigated the potential of using Agricultural Reference Index for Drought (ARID) and various climate indices (CIs) for drought forecasting; the applicability of linear regression (LR), artificial neural network (ANN), adaptive neuron-fuzzy inference system (ANFIS), and autoregression moving average (ARMA) models to forecasting ARID; and the performance of these methods relative to the El Niño-Southern Oscillation (ENSO) approach. Using historical weather data of 56 years of five locations in the southeast USA, monthly values of ARID were computed. Also, monthly values of six CIs that have significant connections with the weather phenomena in this region – AMO, JMA, NAO, NIÑO 3.4, PDO, and PNA – were collected. For ENSO, values of ARID were separated into three ENSO categories and averaged by phases. For ARMA, monthly time series of ARID were fitted to ARMA models. For the other methods, CIs were used as predictors. To avoid possible correlations among CIs, their first principal component (PC1) was used as an input variable. Using LR, ANN, and ANFIS, values of ARID were predicted from the past values of PC1. The performance of these methods was assessed using cross-validation, RMSE, d-index, and modeling efficiency. The ANN models showed the highest performance followed by the ANFIS models. The performance of ENSO, LR, and ARMA models varied depending on months and locations. While ENSO performed better during the winter, LR and ARMA did better during the summer. Although no method could make perfect forecasts, forecasts of ANN and ANFIS models were significantly better than those of ENSO. Results indicated that using CIs and/or artificial intelligence techniques might improve forecasting and that ARID can be used to forecast agricultural drought. |