Atmospheric Fronts in Climate Models: Inter-comparison Across Historical and Future Scenarios
Abstract
Global warming is projected to increase the future intensity and frequency of extreme precipitation events, even in regions where overall precipitation may decrease. In the United States extreme precipitation events most often occur in proximity to atmospheric fronts associated with Extra Tropical Cyclones (ETCs). In order to better understand the way in which the spatial and temporal distribution of extreme precipitation events will change in response to further warming, we need long-term records of atmospheric front locations and time evolution from both historical observations and from climate model simulations. Due to the large number of model simulations required to assess future changes an associated uncertainties, such records can only be generated through reliable automated detection of atmospheric fronts. There have been a number of efforts, particularly over the last 20 years, to develop automated methods for detecting fronts using a number of different approaches. In this presentation we detail the results of using a Deep Learning (DL) Convolutional Neural Network (CNN) to detect atmospheric fronts using the same inputs that meteorologists are trained to use - namely, maps of pressure, temperature, humidity, and wind speed and direction. Over the past few years CNNs have been used with tremendous success in a variety of disciplines for feature and object recognition tasks. We designed a CNN that we trained using input data from the NASA MERRA-2 reanalysis dataset and truth data from the National Weather Service Coded Surface Bulletin dataset, which is a multi-year collection of digitized atmospheric front coordinates acquired during the 3-hourly North America surface analyses done by meteorologists at the NWS Weather Prediction Center. Once trained with three years of data, this CNN was used to produce a 15-year record of 3-hourly atmospheric front locations using the MERRA-2 dataset. We then produced records of front locations using data from the CAM5, CM3, ESM2G, and ESM2M models over a number of historical and future scenario runs. On analyzing these results, we find the primary trend in the climate model future scenarios is a reduction in the number of summer atmospheric fronts over much of the United States.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2018
- Bibcode:
- 2018AGUFMIN11A..03B
- Keywords:
-
- 1906 Computational models;
- algorithms;
- INFORMATICSDE: 1914 Data mining;
- INFORMATICSDE: 1942 Machine learning;
- INFORMATICSDE: 1978 Software re-use;
- INFORMATICS