Analysis of Drought Hazard in Indonesia Based on Decade-Scale Global Climate Predictions

Drought is a slow onset disaster that affects large areas. Extreme drought events occur in a decadal time scale with nationwide impacts, which require a decade-scale climate prediction to anticipate. Currently, WMO has established the Decadal Climate Prediction Project (DCPP), which delivers decade-scale climate forecast. However, the disseminated DCPP products do not include information regarding the probability of drought events (especially extreme droughts) in the next 5 years. A method to perform drought hazard analysis from DCPP rainfall predictions has been developed in this study. The method involves computation of 12-monthly SPI (SPI-12) from the raw ensemble outputs of three (out of nine) selected DCPP models. Model selection is based on performance evaluation against GPCC rainfall data for the period of 1962 to 2017. Both model evaluation and SPI-12 calculations are carried out for spatial aggregate defined by island and zone of season (ZOM), as well as temporal aggregate of five years. The skill of probabilistic forecasts of SPI-12 for moderate and extreme drought categories is evaluated using ROC method. The results show that there are three different selected models for each of the aggregated regions, but MOHC model consistently appears in 9 out of 11 regions. It is also found that, despite no calibration being applied, the SPI values calculated from models show comparable distributions with those from observations. Comparisons between probabilities computed from the ensemble model and observations for SPI-12 of different categories indicate a trend in model bias, with a tendency to underestimate at the beginning of the study year and overestimate at the end of the 2000s. The ROC scores that are evaluated per decade also show fluctuating values with the best period of 1991 to 2000. This suggests that observed rainfall in the regions of aggregation has non-stationary property, which seems to be unrepresented by the models. It might be important to consider such non-stationarity in future analysis.