Forecasting Drought Indices using Machine learning and Spatial Modelling Approaches: An Application in Fiji

Compared with larger nations, small island developing states (SIDS) are disproportionately affected by natural disasters relative to size and frequency. Floods and droughts have drastic impacts on societies; however, few studies have been performed in the Pacific region as a means to reduce risks. Drought indices are useful to understand vulnerabilities and can be used to improve knowledge of spatial and temporal distribution of drought and floods, thereby improving preparedness. For this research, drought indices were tested for the Fiji region and verified using two plant productivity indices. It was concluded that the Effective Drought Index is the best-performing index and precipitation is a suitable parameter which can be used in monitoring hydrological extremes in island states such as Fiji. Further, a meta-analysis was carried out, and it was concluded that wavelet-transformed artificial neural networks (WANNs) have the highest accuracy to forecast drought indices. Therefore, we employed a WANN model to forecast the EDI using climate indices. The model showed promising results in predicting EDI; however, both the models showed average performance in predicting extreme events. Instances of over-prediction and under-prediction were noted from the categorical verification. We also used multivariate statistical techniques to carry out spatial drought modelling. The results were not better than those of the neural networks; however, multivariate techniques have added advantages such as identifying relationships between response and explanatory variables, which can be used with other modelling techniques. In conclusion, neural networks can be used to implement an operational hydrological extreme monitoring system in Fiji. Further testing and optimisation to better predict extreme events will be useful for informing the public; this area warrants more research in the near future