fields/climate-weather/what the AI was for
Classification in weather, climate and oceans
How researchers in weather, climate and ocean science sort measurements into set categories: is it raining, is a freak wave coming, is the system about to shift.
How it is used
Classification means sorting things into categories that are already decided in advance. The simplest version is a yes-or-no question asked of each case: rain or no rain, dangerous wave or ordinary sea. The categories can also be graded, such as light, moderate and heavy rainfall. In this field the thing being sorted is usually a measurement or a short stretch of recording: a satellite view of a patch of sky, a gauge reading, a run of sea surface heights from a floating buoy. The model is shown many labelled examples, works out which patterns in the input go with which label, and then applies that to cases it has not seen.
A notable trick is to turn a forecast into a classification. Instead of predicting what the sea will do, researchers cut out the recording that came just before each known event, label those stretches as leading to the event or not, and ask the model to tell the two apart. The same framing is used for abrupt shifts in a system's behaviour, where simulated series are labelled by the kind of change that followed. Models range from decision-tree ensembles, which build many simple rule-based classifiers and pool their votes, to neural networks that scan a series for shapes and track it over time.
Why researchers use it
Much of what this field needs to know is already a question about categories. Warnings, thresholds and operational decisions are categorical: a port either closes or it does not. A model that answers in categories answers in the form the decision takes, and it can be judged plainly by how often it is right and what kind of mistake it makes. Classification also copes with inputs that resist a tidy equation. The physics of a freak wave or of the clouds that hold rain is understood in outline but hard to compute from a handful of local measurements, so a model learned from records stands in for a calculation that cannot be carried out.
It also stands in for human reading of data. Sorting records by eye is possible but slow, and these fields produce streams of measurements continuously, from satellites, gauges and instruments at sea. A trained classifier applies the same criteria to every case at the same speed. And where real labelled events are scarce, because the events themselves are rare, models can be trained on series produced by simple simulated systems and then applied to measurements from the real world, on the hope that the patterns carry across.
In the record
Limits
A classifier only knows the categories it was given, and the boundaries between them are chosen by people, not by nature: three competing definitions of a rogue wave give three different label sets. Rare events make for lopsided training data, and the fixes, copying scarce examples or weighting them, can flatter the scores without improving the warnings. Labels inherited from gauges or simulations carry those sources' errors. Models trained on simple simulated systems may meet real measurements that are noisier, shorter or simply unlike anything in training, and will still return a confident answer.
Written by AIxSci: the background from general knowledge, the examples from checked records.