Identifying Blizzards in Present and Future Climates 3 May 2017 Climate Prediction Applications Science Workshop Dr. Aaron Kennedy Brooke Hagenhoff University of North Dakota Supported by: NSF project IIA-1355466 1
Extreme Events of the Northern Plains Identify patterns responsible for hazardous or societally relevant weather (blizzards, severe convection, etc.) These types of events are some of the least understood with regards to climate / climate change. USGS 2
A Source of Heartburn Jan. Feb. 3
Northern Plains Blizzards Blizzard Frequency (1959-2014) Number of Blizzards Coleman and Schwartz (2016) 4
Blizzards in the Red River Valley Outlines represent forecast areas for National Weather Service Offices Kellenbenz/ Grand Forks NWS Note lull during mid-feb. 5
Causes of Blizzards Colorado Lows: Typically occur early/late in season. Falling snow + strong winds. Moisture from Gulf of Mexico = best potential for high snow totals. Alberta Clippers: Most common from Dec. > Feb. Less moisture = lower snowfall totals. Arctic Fronts: Responsible for ground blizzards. Winds loft snow fallen over previous days. Occur Jan-Feb Hybrids: Nature doesn t care about our simple taxonomies. Many systems have characteristics of multiple patterns. 6
Synoptic Pattern Classification Self Organizing Maps (SOMs) Kohonen (1995) Competitive neural network Unlike other techniques, classes are related to each other in a 2- dimensional matrix (feature map) If you remove the neighborhood function, the SOM is reduced to a k- means clustering technique (vectors compared using Euclidean distance) From http://www.lohninger.com Public domain software: SOM_PAK http://www.cis.hut.fi/research/som-research/nnrc-programs.shtml Routines in Matlab, Python Packages: PyMVPA, SOMpy, etc. 7
SOMs in the Arctic Changes in cloud fraction (Skific and Francis 2012) Extreme temp./ wind patterns for Barrow (Cassano et al. 2006) Connections between sea ice / atmosphere (Reusch and Alley 2007) 8
Classification of states begins with historical patterns Database of blizzards from the NWS, daily patterns at 12 UTC (for now) Rely on reanalyses (NARR) Regridded to GCM output (~1 ) Methodology Variables used: MSLP, 900 hpa winds and RH. 500 hpa height anomalies, winds, and RH. Warning: Multivariate data should be weighted so they contribute equally to the SOM. Also be aware of biases (seasonal, between datasets, etc.) 9
Example of Bias Problems 9x6 (54 class) SOM GCM has positive bias of 500 hpa RH Due to bias, patterns more likely to be classified to more humid patterns 10
4x2 Blizzard SOM Class 1 Class 4 Class 5 Class 8 500 hpa height anomalies MSLP, 900 hpa humidity 11
Classified Patterns 12
6x4 Blizzard SOM More complexity allowed. Drawback: fewer cases per class. Arctic Fronts 500 hpa height anomalies Colorado Lows 13
Historical Patterns Classify all historical patterns (daily from 1979-2014). Using error characteristics of each class, choose threshold error to count patterns as a blizzard. Cases must also be sub-freezing. Only a percentage of events are blizzards. Land surface conditions not yet considered (work in progress). 14
Historical Patterns Based on percentages, check frequency of classes by month. Tests ability of SOM to represent observations Class Count by Month Arctic Fronts Colorado Lows 15
Classifying GCM Patterns Threshold error and percentages are then used to type GCM Patterns Community Earth System Model (CESM) Mother of All Runs (MOAR) 6hr output 16
Next steps Include time dimension Final Thoughts Consideration of land surface properties Hierarchal SOMs? Take home point: SOMs, analogs, etc. are tools that can bridge communities. Weather Climate Research Operations Check out: Poster (same title) Kennedy, A., X. Dong, and B. Xi, 2016: Cloud Fraction at the ARM SGP Site. Reducing uncertainty with Self Organizing Maps. Theor. Appl. Climatol., 124:43-54. DOI:10.1007/ s00704-015-1384-3. 17