4.6 Using Deep Learning to Create a Long-Term Climatology of Warm and Cold Fronts

Tuesday, 14 January 2020: 11:45 AM
156A (Boston Convention and Exhibition Center)
Ryan A. Lagerquist, CIMMS, Norman, OK; and J. T. Allen and A. McGovern

Several studies have investigated the climatology of fronts (e.g., Serreze et al., 2001; Berry et al., 2011; Catto and Pfahl, 2013). In these studies fronts have typically been detected by numerical frontal analysis (NFA; Hewson, 1998), a type of expert system applied to gridded data. NFA has many deficiencies, including sensitivity to noise (algorithms often depend on second and third derivatives of the underlying field), dependence on rules defined a priori (rather than learned from the data), and inability to handle diverse situations (e.g., methods that cannot detect warm fronts or shear-induced fronts, discussed in Schemm et al., 2015). These deficiencies can be overcome by machine learning (ML). Specifically, we use convolutional neural networks (CNN) to detect warm and cold fronts at every grid cell in a model field. This talk will focus on developments made since Lagerquist et al. (2019).
Predictors are temperature, specific-humidity, u-wind, and v-wind fields from the surface and 850 mb in the ERA-5 reanalysis. Labels (correct answers) are human-drawn fronts from Weather Prediction Center surface bulletins. The CNN outputs three probabilities – no front, warm front, and cold front – for every grid cell. A novel procedure is used to turn this probability grid into polygons (warm-frontal and cold-frontal zones). Evaluating the CNN against human fronts on testing data with a 200-km matching distance, the CNN achieves a probability of detection (POD) of 0.75, false-alarm rate (FAR) of 0.49, critical success index (CSI) of 0.44, and frequency bias of 1.46. A 100-km matching distance leads to POD of 0.70, FAR of 0.67, CSI of 0.29, and frequency bias of 2.10. However, the human fronts are not perfect, and we will show that the CNN is more temporally and spatially consistent.
We apply the trained CNN to an area covering most of North America, every three hours from 1979-2018. The resulting labels are used to study the climatology and variability of warm and cold fronts over the domain. This talk will focus on the tuning and application of the CNN and discuss basic findings from the climatology, while a complimentary talk by John Allen will discuss details of the climatology and variability. This talk will also show common sources of disagreement between the CNN and humans, along with interpretation maps to understand why the CNN makes a different decision.
Berry, G., M. Reeder, and C. Jakob, 2011: "A global climatology of atmospheric fronts." Geophysical Research Letters, 38 (4).
Catto, J., and S. Pfahl, 2013: "The importance of fronts for extreme precipitation." Journal of Geophysical Research: Atmospheres, 118 (19), 10791–10801.
Hewson, T., 1998: "Objective fronts." Meteorological Applications, 5 (1), 37–65.
Lagerquist, R., A. McGovern, and D. Gagne II, 2019: "Deep learning for spatially explicit prediction of synoptic-scale fronts." Weather and Forecasting, early online release, URL https://journals.ametsoc.org/doi/abs/10.1175/WAF-D-18-0183.1.
Schemm, S., I. Rudeva, and I. Simmonds, 2015: "Extratropical fronts in the lower troposphere – global perspectives obtained from two automated methods." Quarterly Journal of the Royal Meteorological Society, 141 (690), 1686–1698.
Serreze, M., A. Lynch, and M. Clark, 2001: "The Arctic frontal zone as seen in the NCEP-NCAR reanalysis." Journal of Climate, 14 (7), 1550–1567.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner