It looks like you're new here. If you want to get involved, click one of these buttons!
Yesterday I went to a tutorial on Climate change: challenges for machine learning by Arindam Banerjee and Claire Monteleoni.
I took rather sketchy notes because they said their slides would be made public. Here are my notes:
We don't know how climate change will affect the tails of the temperature distribution - the probability of extreme events.
World Climate Research Programme 2013 grand challenge: understanding and improving predictions of extreme weather events.
Climate models are a very interesting playground for latent variables - in a climate model you can measure variables that can't be seen otherwise.
There's a lot of low-hanging fruit for machine learning techniques.
There's a workshop called Climate Informatics, which first met in 2011. The next meeting will be on September 25-26 in Boulder, Colorado.
Li, Nychka and Ammann, JASA, 2010 - Bayesian hierarchical model used to study "hockey stick".
Paleoclimate data can be reconstructed using "sparse matrix completion techniques" - if we can discover latent structure.
Also "data fusion" - combining data from different sources - is important.
Chatterjee et al, SDM 2012 studied the influence of ocean temperatures on on land temperature and precipitation. They used the "sparse group lasso" for high-dimensional regression. The idea: only a few ocean locations are relevant. This reminds me of Daniel Mahler's attempt to locate the most significant regions for El Niño prediction.
Climate downscaling: LatticeKrig is a method for "spatial downsizing" to generate good local climate predictions from global ones. Benestad et al NCC 2012.
Why use model ensembles?
"Ensembles of opportunity" - when there just happen to be lots of modeling groups making models.
"Initial condition ensembles" - change initial conditions to increase robustness of forecasts
"Perturbed physics ensembles" - change parameters in the model to increas robustness.
The Coupled Model Intercomparison Project or CMIP tried to improve the results of the IPCC ensemble. Average prediction over all models is better than any one. But what's the best way to combine model predictions? It's not simply taking the average.
Tracking Climate Models is a method of finding the best way to combine forecasts - a method which changes with time as conditions change.
Jia DelSole and Tippett, J. Climate 2013 - discovered a low intrinsic dimensionality of world climate models, due to El Niño and a few other features! This looks interesting for our El Niño project.
Kawale et al, SDM 2011, Steinback et al, KDD 2003 did automated discovery of pressure dipoles. This also looks very interesting, combining machine learning ideas with Paul Pukite's fondness for dipoles!
Ebert-Uphoff et al, A new type of climate network based on probabilistic graphical models: results of boreal winter versus summer J. Clim. 2012 used Bayesian network ideas to infer causal relationships between the 4 biggest teleconnections!
Deng et al, GRL 2014 - in climate models, information flow in the weather diminishes as we move forwards in time in global warming scenarios: the biggest northern-latitude "hubs" in climate networks disappear; remaining hubs move poleward.