iCAST: A severe storm nowcasting prototype focused on optimization of the human-machine mix

Nowcasting of severe and/or high-impact weather is now the main focus for forecasters at the Meteorological Service of Canada (MSC). Nowcasting in Canada is particularly challenging due to the wide variety and frequency of severe weather that affects the country as well as the scarcity of observational data, especially in its many remote regions.

At each MSC Storm Prediction Centre, staffing is modest and the area of responsibility is immense - more than 1,000,000 squared km. MSC severe weather forecasters must monitor data from a large number of sources, including 10 or more radars, to maintain situational awareness.

To further complicate the problem, storm development in many regions is greatly influenced by mesoscale processes (e.g. thunderstorm initiation at drylines and lake breeze fronts, lake effect snow in winter). Thus, nowcasting in Canada requires intensive use of mesoscale data (observations and NWP), and mesoscale features are of critical importance.

A nowcasting system combining sophisticated representation of observational data, NWP, and algorithm output with an intuitive interaction interface is needed to help the forecaster cope with nowcasting in the Canadian context. It is recognized, however, that such a system needs to be carefully designed so that it does not erode forecaster expertise. Thus, the human-machine mix must be optimized to make the best use of both human and machine strengths.

What is the ‘optimum' human-machine mix? There are several areas where the ability of the human forecaster remains superior and will be so for some time, including pattern recognition, the use of conceptual models, judgment / decision-making, and adaptive strategies. Conversely, machine strengths currently include dealing with large volumes of data, integrating numerous datasets, rapid handling of complex calculations / complicated parameter interactions, and performing tedious tasks.

The ‘optimum' human-machine mix may be different for various applications. For some applications, an entirely automated forecast product may be sufficient, providing useful information 90% of the time. However, for the purposes of nowcasting severe weather, that extra 10% may represent the most significant weather events. In addition, the ability to respond to any type of situation, even those that are rapidly changing or unexpected, is essential. Thus, we suggest that the optimum human-machine mix for a severe weather nowcasting system will be achieved when both the human and machine strengths listed above are fully exploited in a complementary manner.

To this end, a prototype severe storm nowcasting system, tentatively named iCAST (interactive Convective Analysis and Storm Tracking), is being developed with an emphasis on optimizing the human-machine mix. iCAST is being prototyped on a platform developed at Environment Canada called Aurora. Aurora is an area-based, object-oriented system that allows forecaster interaction with graphical objects such as points, lines, areas, gridded surfaces, and time-linked tracks. Aurora ingests high-resolution satellite data, and radar data, surface station observations, and lightning network data from both Canada and the United States. It also ingests NWP (currently the Canadian GEM and US RUC models) and NWP-based statistical guidance.

iCAST employs a three-stage approach to nowcasting. First, past weather, current observations, NWP model / statistical guidance, and conceptual models are used to generate a mesoscale prognosis for the T0+6 time frame. The forecaster identifies, and creates objects for, both synoptic-scale features (such as fronts and jets) and mesoscale features (such as lake-breeze fronts and outflow boundaries). In addition, the forecaster identifies areas where thunderstorms and severe weather may be expected to occur in a 3-hour period beginning at that time.

Next, mesoscale analyses are performed hourly by the forecaster to verify the existence and locations of mesoscale features using mainly observations, but also NWP guidance where complementary. The forecaster must use pattern recognition and conceptual models to build a coherent depiction of the weather, despite any missing or conflicting data. Analyses are used to test the prognosis and revise it as necessary.

Lastly, once thunderstorms are present or imminent, storm-scale nowcasting is undertaken. We are currently experimenting with storm track filtering and editing as a way to achieve an optimal human-machine mix at the storm-scale. The forecaster may use conceptual models, analyzed / nowcast mesoscale feature information, current observations, and some NWP model output to modify cell tracks and intensity trends for the highest priority storms. A significant challenge here is dealing with the high frequency of radar data updates and keeping the workload manageable.

In the prototype phase, iCAST will help to assess the value of human input to the severe storm nowcast problem via an automated verification process. In addition, we are testing whether a semi-automated warning production interface leads to more efficient and effective warning generation.

iCAST has been evaluated in a real-time, operational setting during the past several summers via a Research Support Desk at the Ontario Storm Prediction Centre in Toronto. It is anticipated that successful components of the iCAST prototype will be proposed for transfer to Environment Canada's national forecaster workstation (NinJo).