karen mckinnon, andrew poppick - neurips 2021 tutorial
- all systems are coupled (e.g. earth.nullschool.net)
- areas for ml <> climate science
- heterogenous things - understand climate locally
- causal inference
- pairing ml with physical models
- improve climate data via infilling, uncertainty estimation, etc.
- data
- in situ measurements - direct, set up measurement (e.g. temperature) at location and record over time
- land - global climate network temp. stations
- ocean - harder to measure, best measured along shipping roats
- argo program (~2005) - started recording using autonomous floats
- satellite data - challenge: can't see through clouds, sometimes doesn't properly quantify what we want
- ex. land-surface temperature
- ex. SMAP satellite - soil moisture + ocean salinity
- doesn't go far back (e.g. starts ~1979)
- gridded data
- noaa oisst v2 - satellite + in situe daily ocean
- berkely earth surface tempretaures - interpolated in situ measurements, monthly
- reanalyses - combination of many data sources
- e.g. ECMWF - physical model for weather which assimilates data
- proxy records / archives - ex. tree rings
- long time scales
- climate models / earth system models
- ex. carbonbrief.org
- divide earth into grid cells (can be too coarse)
- interact using known equations
- very computationally expensive
- easy to do experiments
- lots of noise, especially on small scales
- challenges
- inhomogeneities - observing platforms change over time
- in situ measurements - direct, set up measurement (e.g. temperature) at location and record over time