Developing drought index-based forecasts for tropical climates using wavelet neural network: an application in Fiji

Hydrological extremes are complex phenomena that will intensify in the future as a result of the changing climate. Seasonal forecasts are one of a multitude of tools that can be used to support an early warning system to monitor hydrological extremes, such as droughts. In this study, different drought indices were calculated for regions around Fiji and a correlational study was performed using two satellite vegetation products: the Normalized Difference Vegetation Index and Enhanced Vegetation Index. The Effective Drought Index (EDI) had the highest direct correlation with the satellite vegetation products. The vegetation response time to the EDI is short, and during rain deficit periods, changes in vegetation can be observed within a month. Thereafter, the study employed a wavelet-transformed artificial neural network (WANN) model to generate short-term (1 month lead time) EDI forecasts using the Southern Oscillation Index (SOI) and Nino 3.4. The model performance was measured by mean square error (MSE) and coefficient of determination (R-2). The forecast accuracy was further verified using a confusion matrix based on predicted EDI classes. The model showed promising results in the testing stage, with R-2 values ranging from 0.93 to 0.97. The MSE values were between 0.05 and 0.09. Under-estimation and over-estimation of some of the predicted EDI categories were also noted. Marginal improvements were noted in the model performance using Nino 3.4 and SOI as exogenous variable. The model showed good skill in predicting EDI, thus can be used to develop an operational drought early warning system for Fiji.

Automated summary

This study calculated different drought indices and conducted a correlational study using satellite vegetation products, finding that the Effective Drought Index (EDI) had the highest correlation with vegetation and that vegetation responds quickly to changes in the EDI. By using a wavelet-transformed artificial neural network (WANN) model to generate short-term drought index forecasts, the model showed high accuracy with R-2 values ranging from 0.93 to 0.97 during testing.