Using Multi-Task Learning to Isolate the Response of Western US Precipitation During ENSO Events
Abstract
This study is the first to use a multi-task learning (MTL) approach to isolate the non-linear response of El Nino-Southern Oscillation (ENSO) on precipitation over the Western United States, namely southern California. We construct a multi-input multi-output autoencoder (MIMO-AE) trained on SST anomalies over the tropical Pacific (20S - 20N, 150E - 80W) and southern California precipitation to capture the non-linear shared latent space between them. By extending this general autoencoder architecture to include multiple inputs and outputs, we capture the variability that is shared by SST in the tropical Pacific and precipitation over southern California. We use this approach to identify a new index for categorizing ENSO events based on their influence on western US precipitation. We found that this new index is similar to other non-linear indices used for ENSO event identification, namely the ENSO Longitude Index (ELI). The MIMO-AE is able to successfully capture a stronger relationship between southern California precipitation and ENSO over traditional methods like the Nino 3.4 index. For instance, it categorizes the 2015-2016 ENSO event, where western US precipitation was below normal, differently than the 1982-1983 and 1997-1998 events that showed above-average precipitation anomalies. This study is the first to use a MTL approach to isolate known ENSO teleconnections in model and observational data and project them onto a temporal index.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.A14C..03P