Ice sea the future: statistical machine learning to predict Antarctic sea ice with quantified uncertainty
Conference
Format: IPS Abstract
Keywords: machine learning,, sea, statistics
Session: Invited Session 10B - Modern applied and theoretical approaches to environmental statistics
Thursday 5 December 1:30 p.m. - 3 p.m. (Australia/Adelaide)
Abstract
Antarctic sea ice has reached record lows in recent years. These changes are unprecedented in the observational record. It is important to model these short-term, unusual changes due to the role sea ice plays in regulating the global climate system and potential sea level rise.
I will describe the Stochastic Spatial Random Forest (SS-RF)[1] method and how we are using it to accurately predict Antarctic sea ice extent a year in advance. This is a peer reviewed, open source machine learning method which is accurate and effective for predicting land cover change based on satellite data. It is fast, non-parametric and generalisable to many types of big data. Importantly, it explicitly quantifies uncertainty by producing posterior probabilities of land cover. This is highly valuable when modelling sea ice extent change which is inherently uncertain, and increasingly so as the climate is changing.
[1] J. Holloway-Brown, K. J. Helmstedt, and K. L. Mengersen, “Interpolating missing land cover data using stochastic spatial random forests for improved change detection,” Remote Sens. Ecol. Conserv., p. rse2.221, Jun. 2021, doi: 10.1002/RSE2.221.