Comparison of Two Physical- and Statistical-Based Postprocessing Methods for High-Resolution NWP Visibility Forecasts over 13 North European Airports

Short-term visibility forecasting for Terminal Aerodrome Forecasts (TAFs) is one of the remaining strongholds where human forecasters are still able to use their expertise to outperform direct model output (DMO) or many of the solely model-based post-processing products. This applies especially in the extended nowcasting range, where information from both observations and model forecasts can be exploited to create a synthesis forecast. As the forecast length exceeds the nowcasting range, the model information becomes increasingly reliable as compared to persistence forecast. As TAF forecasts are issued for 24 hours, there is correspondingly a high demand also for calibrated model forecasts extending beyond the nowcasting range.

In this study, we apply two distinct post-processing methods for visibility forecasts over the Northern European domain covering 13 airports and compare their forecasts with the DMO of a high-resolution HARMONIE-AROME model. The first post-processing method is a physical-based diagnostic method, which has been developed and empirically refined by the duty forecasters at the Finnish Meteorological Institute over the course of several years. The method takes into account the worsening of the visibility from the default value, caused either by precipitation of various forms or by low stratus clouds. The second method is a statistical-based analogue method, where the prevailing weather conditions from the recent past 6-hour METAR observations are compared with the longer observational time series of several years in order to distinguish the most similar past situation to the current one. After identifying the most similar past situation, that situation is correspondingly used to forecast the present situation. The method gives each past situation a scaled similarity index, based on fuzzy logic. The method applies varying weighting for data points, depending on the used variable and the time difference as compared to present time.

The physical-based postprocessing scheme can be applied to DMO of any model that provides the output of the required parameters (precipitation form, intensity, humidity, cloud coverage) for the algorithm and this is also being done at FMI for several models. The analogue method does not need any model forecasts as it is a purely observation-based product, but the observational time series preferably needs to be as long as possible.

Our verification of visibility forecasts is mostly based on the ICAO visibility classification defined in Annex 3 (2016). The verification period of the physical-based postprocessing method covers the year 2017 whereas analogue forecasts are verified from Sep2017 to Dec2017. Our results show that the post-processing has mixed effects for the forecasts: Excluding the visibility class 0...150m, the performance of the diagnostic post-processing is better for the low visibility classes as compared with DMO, but the differences might not be statistically significant. Unfortunately, our analogue forecasts do not perform at all as were hoped. The performance is obviously lower as compared to DMO, in terms of both error variance and bias. In addition, the analogue forecasts do not show a similar degradation as a function of forecast length and the additional benefit of having access to observational information is not anyhow evident from the forecasts: Even the 3-hour forecasts perform very poorly as compared with DMO.

Analogue method could be extensively refined though using different weighting schemes, but unfortunately this kind of optimization was computationally too demanding due to long time series used in the calculation of analogies. Physical-based postprocessing method clearly shows more promise to it, which has also been observed in operational setting. There remains plenty of room for the forecaster to make subjective evaluation of the available products.