Atmosphere Ocean Science Colloquium
Dimensionality reduction and network inference for climate data using 𝛿-MAPS
Speaker: Fabrizio Falasco, Courant, NYU
Date: Wednesday, October 20, 2021, 3:30 p.m.
The exponential growth of climate data combined with advances in machine learning offers new opportunities to understand the climate system and its response to external forcings. In this talk, I will present 𝛿-MAPS, a recently developed complex network analysis method for climate fields.
Given a spatiotemporal climate field embedded on a grid, the framework first identifies patterns (domains) defined as spatially contiguous, homogeneous sets of time series. At a second step, it infers a weighted and direct network between domains. The resulting climate network allows to condense large spatiotemporal fields in few domains plus their connectivity patterns, therefore providing a simple new dimensionality reduction scheme for climate data. Applications ranging from climate models evaluation to the investigation of regime shifts at paleoclimate (~103 years) scales are discussed.