A. Science issues - Zawadzki and Wilson

1. The crossover point where NWP improves over extrapolation depends more on the radar coverage than model resolution. This is one way to say that the extent and resolution of the observations and the model science are more important than model resolution.
   Observations - For example the resolution of observations of water vapor or the resolution of boundary layer divergence just may be more important than model resolution in pin pointing the future location and intensity of convective storms.
   Science - For example is the model microphysics good enough to determine the timing and characteristics of thunderstorm outflows. These items are critical for determining storm evolution and secondary initiation.


2. We need to quantify the uncertainties in Nowcasting so we can define in a probabilistic manner the nowcast.
   For example there may be a developing storm for which we want to define its likelihood to become severe. There are conceptual models based on years of research to do just this. These models include meteorological variables like CAPE, vertical wind shear and radar characteristics. What is needed is quantification of how variations in the magnitude of these parameters effect the probability of the storm becoming severe.
   We note that ensembles and fuzzy logic are well suited for deriving nowcast probabilities.


3. Does the human have any role in real-time Nowcasting? During the week we have heard arguments on both side of this question.


4. The science leads for this panel are making a plea to the attendees to read the literature and not "reinvent the wheel". There are plenty of new problems to resolve; try and make a leap in understanding rather than resolve old problems.

B. Nowcasting Systems - Golding

1. Choosing the appropriate balance between NWP, Extrapolation and Conceptual Models.
   NWP is certainly gaining rapidly in capability at the storm scale, and projections indicate that it will be possible to make very short range NWP forecasts available within one hour of data time. The critical area needing further development is in provision and assimilation of fine scale observations. Both extrapolation and conceptual models were originally manual techniques, which have been reproduced automatically with varying degrees of success. Both can be expected to retain a role in nowcasting for lead times shorter than the processing time for NWP (i.e. ~1 hour). They also provide a cheap alternative to NWP for specific problems. While automation of extrapolation has been generally successful, there are few examples of the successful automation of conceptual models. This may be related to the difficulty of pattern recognition, problems arising from the artificial limit to patterns that fit, or to the inherent nonlinearity of interaction with the environment. In general, it would appear that nonlinear modelling is a better approach to solving these problems.


2. Role of the human
   Presentations at the symposium provide a wide spectrum of views on the role of the human in nowcasting. There is a clear issue about the scope of responsibility, with the greatest benefits found where manual responsibility is restricted to a rather small area ? perhaps the range of one radar. A related issue is that, whereas an experienced forecaster can routinely improve on automated products, inexperienced forecasters can equally routinely degrade them, so that there is a substantial additional cost in lengthy training if all operational input is to be from experienced staff. In other words, a human role appears to be justified only if the cost of the human resource is small and/or the potential benefits large. Where the staff cost is high or the weather impacts are modest, automation appears to be the preferred route. However, even in this situation, there is strong evidence of a continuing need for human quality control of the final warnings, and of human interpretation of the forecasts for sensitive users.


3. Monitoring
   Automated systems need to have a much higher degree of reliability than those that merely form an input to the forecaster. In particular, automated monitoring and quality control of automated end products will be needed.


4. Non-precipitation nowcasting
   While the symposium was dominated by convective precipitation nowcasting, a substantial requirement for very short range prediction of other variables was demonstrated. As well as looking at opportunities for better integration of approaches to precipitation nowcasting across different countries, there is also a need to look at integrating nowcasting approaches for different variables. If the future of nowcasting lies in convective scale NWP, this should happen naturally, since NWP models deal with the full structure of the atmosphere, though it will be important to ensure that appropriate observations, data assimilation, and physical representations are included to meet these additional needs.

C. Issues on NWP and data assimilation- Koizumi

1. NWP and data assimilation are required to provide reliable atmospheric states in a timely manner. In order to meet the requirements, aside from improving NWP model and data assimilation system themselves, we are to exploit all available observational data (not only radar reflectivity).


2. Currently many people seem to think that time-consuming data assimilation method like 4D-Var is not practical to be used in nowcasting/very-short-range-forecasting context.
   Currently many people seem to think that time-consuming data assimilation method like 4D-Var is not practical to be used in nowcasting/very-short-range-forecasting context.


3. Estimating and providing "Uncertainty" of model forecasts is getting more and more important. However, there is no standard method to estimate it yet. Since model errors for mesoscale events can have quite ill-shaped (e.g. not Gaussian) probability distribution, it is not guaranteed that "ensemble method" provides appropriate estimation of model errors. Error estimation methods and their limitations are to be investigated.