(b) Downscaling Methods

We can think of downscaling in two stages, though in practice, the two stages are often combined:

i) Meteorological Step - demonstration of the ability to predict climate features that are closer to many aspects of applications problems. Examples include seasonal rainfall total for a particular small region or even point location, the number of storms, the number of dry spells etc. In making this demonstration, we are confined to questions about the predictability of the atmosphere.

ii) Consequences Step - demonstration of the ability to predict features most directly connected to climate applications. Examples include streamflow, crop yield, mosquito distributions. Ultimately, the downscaled prediction information can be created in a format that integrates smoothly with decision support.

It is often valuable to start by thinking about the first (meteorological) step - especially given that establishing the predictability of many of these features is still in a research phase. However, each problem needs to be considered on its own merits. If a watershed covers a region of known high seasonal predictability, then high predictability of the important atmospheric features for streamflow is, in this case, highly likely. Extensive testing of the meteorological step (the downscaled atmospheric features that impact stream flow) is less critical to the development of a reliable operational stream-flow forecasting scheme. In contrast, predictability of the features that impact a crop at the farm scale is still uncertain, and it would be dangerous to try to operationally predict crop yield at the farm level and use this information in decision support, without first evaluating the downscaled atmospheric predictability.

Purely Statistical Approach to Downscaling

The simplest approach to downscaling is to form a predictand of the feature of interest and relate the predictand to known predictors that have a physical relationship with the climate of the region. For example, the predictand may be stream flow or crop yield. The predictors may be SST, such as the Nino3 SST index. For this to be effective, there is a need for:

Requirement 1. Reliable predictors to have been established for the large-scale climate through modelling and diagnostic studies - this issue was discussed in Parts 2 and 3 (see especially discussion of statistical methods in Part 3)

Requirement 2. Confidence that the large scale atmospheric predictability cascades into predictability of the feature of interest.

Streamflow is influenced by rainfall averaged across a catchment, which if the catchment is large enough, implies a scale that is similar to that on which seasonal predictability may already have been established by previous seasonal prediction research. In regions like Northeastern Brazil where the seasonal rainfall total is strongly tied to known SST indices, the result that seasonal mean streamflows have good predictability from these SST indices has a sound physical basis. For example, in Fig. 4.1, streamflow is predicted using two indices of sea-surface temperature measured in ocean locations (central tropical Pacific and tropical Atlantic) that have been shown to strongly impact atmospheric circulation and rainfall in the region of Northeastern Brazil.