This study developed a drought forecast model for Fiji based on high-resolution dynamic downscaling and machine learning of long-range climate forecast data and remote sensing data. The downscaled dataset with 8 km resolution adequately represented the detailed temporal and spatial features of rainfall beside the add value. Bias correction further increased the reliability of the downscaled rainfall dataset. The downscaled dataset did not show clear dry trends during the past 30 years. The leeward side of Viti Levu at an elevation between 200 m and 300 m was very sensitive to SPCZ variability. The downscaled dataset effectively responded to the anomalous large-scale patterns, which provided a solid ground to develop the drought prediction model.

Drought index SPI6 was used as the target variable to indicate drought conditions. Machine learning models were trained using the drought index, which was calculated using in-situ data. There are few weather stations with available precipitation data; this provided insufficient datasets for machine learning models. Limited distribution of weather stations may also hamper examination of spatial characteristics of drought occurrences.

Bias-corrected WRF model precipitation outputs with 8 km spatial resolution were used to calculate SPI6_OBS instead of the index using in-situ data. The use of these WRF model outputs increased drought accuracy and decreased MAE for drought categories in SPI6 forecasting with shorter lead times, whether either the bias-corrected climate model outputs or machine-learning models were used.

1. INTRODUCTION

1.1 Regional Historical Droughts and Their Impacts in Fiji

1.2 Existing Monitoring and Prediction Systems Related to Drought

1.3 Research Questions and Objectives of the Study

2. STUDY AREA

3. LITERATURE REVIEW

3.1 High-Resolution Dynamic Downscaling of Climate Data

3.2 Drought Forecasting

4. MATERIAL

4.1 In-Situ Data

4.2 SPI

4.3 Reanalysis Data

4.4 Seasonal Forecast Data

4.5 Remote Sensing Data

4.5.1 CMORPH Data

4.5.2 PERSIANN-CDR

4.5.3 TRMM

4.5.4 GPM

4.5.5 MODIS Land Surface Temperature

4.5.6 MODIS Vegetation Indices

4.5.7 Elevation Data

4.6 SPCZ

4.7 ENSO

4.8 PDO

5. METHODOLOGY

5.1 High-Resolution Dynamic Downscaling of Climate Data

5.2 Drought Forecasting

5.2.1 Machine Learning Models

5.2.2 Preprocessing of Input Variables

5.2.3 Performance Measures

6. RESULTS

6.1 Impact of Large-Scale Modes on Wet and Dry Conditions over Fiji

6.1.1 Observed Rainfall Features and SPI

6.1.2 Composite of Observed Wet and Dry Events

6.2 Production of 8km WRF Model Outputs

6.2.1 Regional Simulation of Climatology over Fiji

6.2.2 Temporal Evolution of Rainfall

6.2.3 Spatial and Temporal Distribution of SPI

6.2.4 Composite Analysis of Simulated Wet and Dry Events

6.3 Bias Correction

6.4 Development of a Drought Forecast Model

6.4.1 SPI6 Forecast Based on Bias-Corrected Climate Models (FCST_ONLY)

6.4.2 SPI6 Forecast Based on Machine Learning Models

6.5 Validation of the Drought Forecast Model

7. CONCLUSION

Appendix A. Guideline for the Application of the Drought Forecast Model

A.1 Select a Drought Index

A.2 Select Performance Measures

A.3 Select a Method and Climate Model

A.4 Create Spatially Distributed SPI6 Forecast Maps

A.5 Create Zonal Average Maps Acknowledgements

REFERENCES

TABLES

FIGURES