A GIS Evaluation of Cobb's Snowfall Algorithm as a Prediction Tool for Emergency Managers

Snow is one of the more challenging meteorological events to forecast. When trying to predict snow-ratio and total snowfall accumulation at specific locations and times the forecast becomes even more difficult. Accurate forecasts are particularly important for emergency managers who would like to make plans two to three days in advance of a major snowfall event. Two tools have recently been developed that may help improve snow forecasting by facilitating comparisons between snowfall predictions and actual snowfall amounts. The snowfall algorithm developed by Daniel K. Cobb Jr. from the National Weather Service (NWS) in Caribou, Maine uses a top-down approach to calculate snow-ratio at the surface. The calculated surface snow-ratio can then be multiplied by the quantitative precipitation forecast (QPF) to determine predicted snowfall total. Christopher Karsten from Iowa State University has developed ‘The Bufkit Warehouse', which runs different forecast models through Cobb's algorithm and outputs the results in a text format for individual stations. The purpose of this research project is to use geographic information system (GIS) to interpolate snowfall totals using Cobb's algorithm and compare the results with actual snowfall measurements. Forecasted snowfall totals were collected from the North American Model (NAM), for the 00Z and 12Z model runs one, two and three runs (one to two days) prior to the start of each studied snowfall event using ‘The Bufkit Warehouse'. Snowfall data was then interpolated over the study area in Northeast Kansas and compared to interpolated actual snowfall totals from the National Weather Service and storm spotter reports. A total of six events were used between the years of 2011 and 2013. In addition to the snowfall, QPF was also examined for each event. The events were looked at as individual snowfall measurements and as snowfall event totals. When all of the events were treated as one dataset, a positive, statistically significant correlation between the forecast and actual event totals was found. Not surprisingly the forecast model run closest in time to the start of the event had the strongest correlation, whereas the forecast model farthest from the start of the event had the weakest correlation. All model runs tended to over predict snowfall totals. It was concluded that the models do a good job of predicting that snowfall will occur but since there was an average percent difference of 44.36, 43.61 and 74.07 percent between actual snowfall and predicted snowfall for model runs one, two and three respectively they do not accurately predict actual snowfall totals. Results varied for each storm when treated individually, however the two model runs closest to the start of the snowfall event typically performed best.