89th American Meteorological Society Annual Meeting

Monday, 12 January 2009: 4:15 PM
Using Data-Models to Monitor and Forecast Drought
Room 127BC (Phoenix Convention Center)
Ana P. Barros, Duke Univ., Durham, NC
A forecast data-model of severe drought was developed for the Murray-Darling Basin as the demonstration case-study. The model relies on spatial observations to describe the large and regional environmental conditions prior to drought onset. Data analyses were performed to derive suitable candidate predictors using near-global SSTA patterns over the Indian and Pacific Oceans and early indicators of ENSO onset, including: temporal and spatial zonal OLR gradients in the Pacific Ocean, and the far western Pacific windstress anomaly. OLR gradients proved to be dispensable predictors, whereas SPI-based predictors appear to control predictability when the SSTA in the region [87.5oN-87.5oS; 27.5oE-67.5oW] and eastward wind-stress anomalies in the region [4oN-4oS; 130oE-160oE] are small, respectively ±1o and ±0.01 dyne/ cm2. That is, when ENSO activity (large-scale conditions) is weak, regional processes (regional antecedent precipitation dominate). We also found evidence of association between changes in PDO regime, and changes in the low-frequency temporal variability in the SPI that are consistent with asymmetry of ENSO's influence on regional hydroclimatology. The areal averaged 12-month lead-time forecasts of SPI in the MDB explain up to 60% of the variance in the observations (r> 0.7). Based on a threshold SPI of (-0.5) for severe drought at the regional scale, and for a 12-month lead-time, the onset was estimated within 0-2 months of the actual threshold event. Spatial analysis suggests that forecast errors can be attributed in part to a mismatch between the spatial heterogeneity of rainfall and raingauge density of the observational network, and how that affects SPI estimates. Forecast uncertainty on the other hand appears associated with the phase of the climate regime. By examining the characteristics of multiple forecast models with different combinations of physically-meaningful predictors, we begin to show the potential for data-driven long-lead ensemble forecasts, as well as the value of this framework as a means of knowledge discovery and predictability attribution.

Supplementary URL: