Time Series Analyses across Weather and Climate Scales

March 11, 2021

Maria Molina from the National Center for Atmospheric Research

Hosted by Russ Schumacher

Abstract

A time series is data ordered sequentially, which can be used to describe the internal structure of an atmospheric or climate variable as it varies over time. Signal processing and machine learning techniques can be used to describe features within the data and predict what the time series may look like at a future time. Here we leveraged time series analysis methods for two projects of differing temporal scales. The first project involved the application of a random convolutional kernel transform method, which randomly extracts a large number of features from time series that are subsequently input into a classifier. The classifier outputs a probability of a thunderstorm being severe (tornado, large hail) or non-severe using a time series of observations from the Geostationary Lightning Mapper instrument aboard the NOAA GOES-16 satellite. The trained machine learning model performs well in predicting tornado and hail activity with 15-minute lead time, as evaluated using the critical success index, and also reduces the number of false alarms of previously warned convection. The second project involved the application of signal processing techniques (spectrum and wavelet analyses) to five multi-centennial simulations created with the Community Earth System Model, in order to explore the influence of the Atlantic Meridional Overturning Circulation (AMOC) and Pacific Meridional Overturning Circulation (PMOC) on global sea surface temperatures and ENSO. It was found that the amplitude of annual cycle sea surface temperatures across the tropical Pacific decreases and ENSO amplitude increases as a result of an AMOC shutdown, irrespective of the development of PMOC. These results suggest that if climate simulations projecting a weakening of AMOC are realized, compounding climate impacts could occur given the far-reaching ENSO teleconnections to extreme weather and climate events. The underlying physical reasons for results from both studies will also be discussed.