BAMOS Vol 31 No.2 June 2018 | Page 19

BAMOS Jun 2018 Figure 2. Typical data coverage for a run of the European Centre for medium-range Weather Forecasting (ECMWF) operational global model: a) radiosonde coverage; and b) coverage from the Advanced Microwave Sounding Unit (AMSU) on polar-orbiting satellites with over 600,000 observations covering the Northern and Southern Hemispheres equally. (Source: ECMWF web site) This approximation (or discretisation) introduces resolved and unresolved scales: the resolved scales are those larger than a few multiples of the spacing of the discrete points, while the unresolved scales are those scales smaller than that. The physical processes on scales close to or smaller than the grid spacing cannot be properly resolved by such a discretisation, and must be represented in terms of the resolved quantities. In other words, the effects of physical processes too small to be resolved explicitly must be incorporated through empirical laws or parameterised. For example, in most models, convective clouds are part of the unresolved scales and their effects, which are critical to the success of the forecast, must be added in the computer model through rules that reflect the underlying physics and the state of the atmosphere on the resolved scales. Much of the predictive skill of the computer models hinges on how well the unresolved physics is represented by the parameterisations. Numerical weather prediction is not the only way forecasts are made. Forecasts for a few minutes to a few hours, called nowcasts, are usually based on the statistical extrapolation of observations from radar, satellite, lightning networks and conventional instruments. Nowcasts are made most often for rain and thunderstorms. Increasingly, observation-based nowcasts are blended with short-range computer model forecasts to generate short-term forecasts. Forecasters interpret the model predictions and nowcasts, and communicate this information as weather forecasts and warnings to the general public and specific users. At times, forecasters use their local knowledge to adjust the objective forecasts or to make specialised forecasts for a particular place. An example of the former is fog prediction for the aviation industry, while an example of the latter is fire weather forecasts for firefighters. National weather agencies around the world share their model predictions in real-time, and in addition to the output from the suite of models operating on the Bureau’s supercomputer, forecasters in Australia have access to the predictions from at least four international computer models. 3. The Importance of Observations and the Initial State Making a weather forecast requires a detailed knowledge of the current state of the atmosphere, land surface and ocean. (From a technical mathematical perspective, weather prediction is an initial-value problem.) The current state of the atmosphere is known as the analysis. The main sources of input data for numerical weather forecasts are radiosondes (weather balloons) and satellites, as well as commercial aircraft, floating buoys, ships, winds derived from the measured motion of clouds (called cloud drift winds), radars and surface-based measurements from automatically recording manual and automated weather stations. Figure 2 shows the typical data coverage used in making a forecast with the ECMWF (European Centre for Medium-range Weather Forecasting) operational global numerical weather prediction model. The chart for radiosonde soundings (TEMP) shows the poor coverage over ocean areas and the Southern Hemisphere (Figure 2a). In contrast, the chart for observations from the Advanced Microwave Sounding Unit (AMSU), now carried on several polar orbiting satellites, shows over 600,000 observations available for the 1200 UTC analysis, and the coverage in the Southern Hemisphere just the same as the Northern Hemisphere (Figure 2b). The analysis blends the observations with short-term computer model forecasts for the current time, taking into account estimates of the respective errors in the observations and computer model forecasts. This process of estimating the current state is known as data assimilation, and it is one of the most important reasons for the great improvement in the predictive skill of numerical weather forecasts. Weather forecasts can be highly sensitive to small changes in the analysis. The observations used to construct the analysis are affected by instrumental and sampling errors. In most circumstances these initial errors result in forecast errors which can amplify in time. Moreover, the computer models themselves have errors as they are based on approximate forms of the governing laws of physics. In the 1960s, recognition of this sensitivity of the forecast to small changes in the initial conditions led the American 19