Using Machine Learning to Compare Simulated and Observational Sea Ice Extent Data
Abstract
The extent of sea ice in the Arctic has been declining for decades. Simulation models forecast less sea ice decline than what we observe, leading to conservative predictions. It is important to identify why models are too conservative to effectively understand and plan for our future climate. To do this, we can train individual machine learning models on observed data and simulated data separately, and then compare the differences in the weights between the two differently trained models. In gaining this insight, we can shed light on which features the simulation deems important and compare to features the model trained on observational data deems important. We present results on the important features in data models trained on satellite sea ice concentration data, along with atmosphere and ocean reanalysis products, and compare with data models trained on historical simulation data from the Energy Exascale Earth System Model (E3SM). The data models predict sea ice concentration and use air temperatures, solar radiation, sea surface temperature, heat content anomalies, surface pressure, and wind speeds as input. Models are trained on data from 1979 to 2007, and tested on data from 2008 to 2017.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.C23D1599N
- Keywords:
-
- 3349 Polar meteorology;
- ATMOSPHERIC PROCESSES;
- 0456 Life in extreme environments;
- BIOGEOSCIENCES;
- 0750 Sea ice;
- CRYOSPHERE;
- 1620 Climate dynamics;
- GLOBAL CHANGE;
- 1621 Cryospheric change;
- GLOBAL CHANGE;
- 1622 Earth system modeling;
- GLOBAL CHANGE;
- 1694 Instruments and techniques;
- GLOBAL CHANGE;
- 4207 Arctic and Antarctic oceanography;
- OCEANOGRAPHY: GENERAL