Forecasting E0

Forecasts of E0 at timescales ranging from daily to seasonal are increasingly desired by stakeholders and managers in a number of sectors, including agriculture, water-resource management, and wildland-fire management, largely driven by recent developments highlighting the value of E0 for drought monitoring. However, few such E0 forecast tools currently exist.

Weather-scale forecasts of E0 (i.e., for lead times of up to 2 weeks) can be valuable to agricultural producers to assist in irrigation scheduling. Dynamical weather forecast models, such as those used by the US National Weather Service (NWS), output all the necessary variables to compute a physically based E0 as a post-processing step; ET estimates can then be derived using crop coefficients. Users need to be aware that raw dynamical model output will often contain biases, and the spatial resolution of the models (particularly global models) can be quite low. Some recent studies have examined bias-correction and downscaling methods to improve raw E0 forecasts (e.g., Ishak et al. 2010; Silva et al. 2010). Other potential improvements include the use of retrospective forecast analogs (Tian and Martinez 2012a, 2012b) and ensemble forecasting to improve skill over single deterministic forecast runs (Tian and Martinez 2014). The only operational E0 weather forecast product is the Forecast Reference ET (FRET) developed by the NWS. FRET produces E0 forecasts of values and anomalies for Days 1-7 over CONUS. While more research is needed, the 7-day accumulated E0 anomalies could be useful to provide early warning of developing flash drought conditions in the growing season.

E0 forecasts at subseasonal (3 weeks to 3 months) and seasonal (3 to 9 months) scales can be used for long-term planning purposes as opposed to day-to-day operations. As yet, only a few studies have examined the skill and potential application of seasonal E0 forecasts. Tian et al. (2014) used the Climate

Forecast System Version 2 (CFSv2; Saha et al. 2014), a global dynamical seasonal forecast model, to evaluate bias-corrected and downscaled E0 quantities over the southeast US. Moderate skill was found during the cold season, but little forecast skill was found during the growing season due to the inability of the coarse global model to resolve convective processes. A broader analysis over CONUS examined the skill of using E0 anomalies derived from CFSv2 to forecast droughts (McEvoy et al. 2016b) and found E0 forecasts to be nearly universally more skillful than Prcp forecasts in predicting drought, with the greatest skill during the growing season in major agricultural regions of the central and northeast United States. Figure 11.6 demonstrates that, averaged over CONUS, E0 forecast skill is greater than that for Prcp during all seasons. The east north central region in Figure 11.6 shows large differences between E0 and Prcp forecast skill with E0 having much greater skill during the growing season, while the southeast region in Figure 11.6 shows one region where E0 forecast skill is quite weak and often similar to that of Prcp.

The question then arises: Why are E0 forecasts typically much more skillful than Prcp forecasts? In general, T is more predictable than Prcp, and several studies have linked a multidecadal warming trend to improved T predictions in seasonal forecast models (Jia et al. 2015; Peng et al. 2013). The ability of CFSv2 to change atmospheric CO2 over time leads to more consistent abovenormal T forecasts (Peng et al. 2013), which has been realized over the last

FIGURE 11.6

One-month lead time forecast skill (based on the correlation between forecasted and observed anomalies) for each seasonal period area-averaged over CONUS (a) and East North Central (b; Iowa, Minnesota, Wisconsin, and Michigan). (Continued)

(Continued)

FIGURE 11.6 (Continued)

One-month lead time forecast skill (based on the correlation between forecasted and observed anomalies) for each seasonal period area-averaged over and southeast (c; Florida, Alabama, Georgia, South Carolina, North Carolina, and Virginia) regions. One-month lead forecasts were initialized during the month prior to the target date (e.g., JFM forecasts were initialized in December). The straight reference line indicates an anomaly correlation of 0.3, which represents the start of moderate skill according to NOAA's Climate Prediction Center. Correlations were computed between CFSv2 reforecast and the University of Idaho's gridded meteorological data (Abatzoglou 2013) for the period 1982-2009. (Adapted from McEvoy et al. 2016b.)

several decades. However, T is not the only E0 driver (nor the most important in some regions), and other factors could be contributing to increased forecast skill over Prcp. SM memory is the primary land surface variable (with El Nino Southern Oscillation [ENSO] being the primary oceanic variable) contributing to seasonal predictability. Dirmeyer and Halder (2016) used CFSv2 to show that T, humidity, surface heat fluxes, and daytime boundary layer development all respond to variations in SM, while Prcp is unresponsive. The strong link to the land surface, as opposed to the free atmosphere, thus leads to improved seasonal E0-forecast skill.

A major hurdle in producing operational seasonal E0 forecasts is the availability of all E0 drivers from operational forecast producers. At the time of writing, the CFSv2 is the only model in the North American Multimodel Ensemble (NMME; a state-of-the-art seasonal forecast tool that uses an ensemble of nine models from the United States and Canadian institutions) to provide public access to forecasts of all E0 drivers in real time.

 
Source
< Prev   CONTENTS   Source   Next >