fields/climate-weather/model families
Random forests in weather, climate and oceans
A look at how random forests, ensembles of simple branching decision rules, are used in weather, climate and ocean science, and why researchers often prefer them to other kinds of model.
How it is used
A decision tree is a chain of yes-or-no questions about the inputs, each one splitting the data until a prediction is reached at the end of a branch. A random forest grows many such trees, each on a different random sample of the data and the input variables, then averages their answers. In this field the inputs are usually the sort of thing an instrument or a weather model already provides: temperature, pressure, humidity, wind, satellite channels, the time of day and the day of the year. The output is a quantity that is harder to measure directly, such as rainfall at the ground, a pollutant concentration, or a chemical property of a sample.
Two uses recur. The first is prediction: learn the relationship between cheap, widely available measurements and a sparse reference set, then apply it everywhere the cheap measurements exist. The second is sorting out which inputs matter. Because a forest records how much each variable improves its splits, it produces a ranking of inputs, often used to trim a long candidate list down to a workable set before a different model is fitted. Forests also serve as a way of removing the weather's own influence from a record, by predicting what a measurement would have been given the meteorology alone.
Why researchers use it
The relationships here are rarely straight lines and rarely independent. Rainfall responds to cloud temperature only beyond a threshold; a pollutant builds up when wind is light and the air is stable, not in response to either alone. A forest picks up thresholds and interactions without being told to look for them, and it tolerates inputs on very different scales, skewed distributions and the odd outlier. It needs little tuning and gives a usable answer from a few hundred samples, which matters when the reference data are gauge readings, laboratory assays or ship measurements rather than model output.
In the record
Limits
A forest averages trees, so its predictions cannot go beyond the range of values it was trained on: it will understate record heat or extreme rainfall, which is often exactly what is wanted. It carries no physics, so it can break conservation of mass or energy and will not transfer to a climate unlike its training period. Importance rankings mislead when inputs are correlated, as weather variables usually are, and a high ranking is not evidence of cause. Reference data are often clustered in space and time, which makes test scores look better than real skill.
Written by AIxSci: the background from general knowledge, the examples from checked records.
Related
Every article
aix-00286
AI fills gaps in river temperature records to track overlapping heatwaves
aix-00288
Machine learning turns satellite cloud images into hourly rainfall maps for central Vietnam
aix-00289
Machine learning points to filter loading as main driver of levoglucosan desorption temperature
aix-00281
Bacteria-derived aerosol matter made to produce oxidants, then modelled from its optical fingerprints
aix-00290
Machine learning separates weather from policy in China's winter heating air pollution
aix-00293
Machine learning used to predict fine particle emissions at a Greek regional airport
aix-00294