Latest posts by Kevin Synnott (see all)
- Why Climate Change needs its Playboy Bunny. - May 4, 2017
- Lick your teeth and recycle: The power of Habit and climate change. - April 13, 2017
- What’s the Job?Climate change in the media. - February 23, 2017
Firstly, what is Machine Learning, other than a “sexy” buzzword used by the science community to make statistics sound cool? Well, per Andrew Ng, chief scientist at Baidu Inc. and an associate professor at Stanford it is “The science of getting computers to act without being programmed”. That statement leaves a lot to the imagination. One might start off by thinking of movies like I-Robot, Her and Ex Machina. For this post let’s stay grounded and say that Machine Learning explores the study and construction of algorithms that can learn from and make predictions from data. So, what can such an algorithm do for Climate Science? Turns out a lot.
Dr Claire Monteleoni, Assistance Professor at George Washington University, uses Machine Learning to track climate models. She along with many others are building a new field of Climate Informatics, a term she coined, with the aim of encouraging collaborations between climate scientists and Machine Learning researchers in order the bridge the gap between data and understanding. In her paper  Tracking Climate Models, she demonstrates the advantage of a Machine Learning approach for combining the predictions of multiple climate models. The Intergovernmental Panel on Climate Change currently use about 20 climate models to make informed decisions and predictions on climate change. This paper introduces the use of an online learning algorithm called Learn-α to make predictions that match or surpass that of the best climate model.
Machine Learning is also useful in application where there is little data or the data is sparse. For example, global maps of the partial pressure of CO2 (pCO2). Observations of sea surface pCO2 are taken mostly by commercial ships and consequently are sparse in both time and space, especially before the 1990s. Knowledge of this pCO2 is essential to investigate the variation of the ocean CO2 sink, from which data comparisons to the Global Carbon Budget can be made. To “fill in the gaps” Dr Peter Landschützer, from ETH Zürich, employed a Forward-Feed Neural Network. Neural Networks are based on the way the biological brain solves problems, using a large cluster of neurons connected by axons.
Finally, looking a bit closer to home, the Informatics Lab at the Met Office are applying Machine Learning techniques to traffic cameras. They are currently undertaking a project which will use data taken from traffic cameras to train a machine to recognize the weather. This is especially useful when considering snow. Snow is the hardest weather to forecast as it depends on small differences in pressure, temperature and heights of clouds. To know if it’s snowy it’s much easier to look at the ground. From the images on the traffic cameras the amount of white could tell you and furthermore characterise the snowy weather. What’s particularly cool about this project is that all the code and data is freely available on the Informatics Lab website http://www.informaticslab.co.uk
To conclude, we can see many applications of Machine Learning from combining the strengths of different Climate models to just wondering if it’s snowing. The field of Climate Informatics is without doubt exciting and becoming more and more important. We might be a long way off a computer feeling cold but until then let’s use it to tell us more about our complicated climate.
 Monteleoni, C., Saroha, S., Schmidt, G., and Asplund, E.: Tracking Climate Models, Journal of Statistical Analysis and Data Mining, 4, 372–392, 2011.