~/aixsci
weekly updates

fields/climate-weather/what the AI was for

Property prediction in weather, climate and oceans

How researchers in weather, climate and ocean science use machine learning to estimate a measurable quantity of something that already exists, from data that only partly reveals it.

How it is used

Weather, climate and ocean science measures the atmosphere, the land surface and the water flowing between them. Many of the quantities researchers want are hard to observe directly, so they are inferred from things that can be measured. Property prediction here means training a model on cases where both the measurements and the target quantity are known, then applying it to cases where only the measurements exist. In the examples, models estimate water temperature and dissolved oxygen in river basins from daily weather and flow, fine particle pollution from aircraft activity and weather averages, storm intensity from satellite images, and the chemical reactivity of airborne particle samples from their optical and chemical signatures.

The models are mostly ordinary supervised ones: regressions, decision-tree ensembles and neural networks. Where the target varies through time, the examples use recurrent networks, which carry information forward from earlier days to later ones. Where the input is an image, the examples treat it as an image-to-number problem and predict a single intensity value. Much of the described work is preparation rather than modelling: lining up measurements taken at different intervals onto a common timetable, discarding observations with missing channels rather than filling them in, and splitting data by basin or by season so that the test cases are genuinely held back.

Why researchers use it

Direct measurement in this field is expensive and sparse. Instruments sit on towers, buoys and ships, and they cover a tiny fraction of the land and ocean. Satellites see much more but see it indirectly, as brightness rather than as the quantity of interest. Property prediction stands in for a measuring instrument that cannot be placed everywhere, extending a few well-instrumented sites to many unmonitored ones, or turning a continuous image record into a continuous record of a physical property.

It also stands in for a physical calculation that is either too slow or not fully known. Some processes, such as how organic matter in the air produces reactive chemicals, or how a compound comes off a heated filter, have no tractable equation linking the easily measured inputs to the outcome. A model fitted to laboratory results supplies the link empirically. In one example the fitted model is used to reconstruct decades of daily values across many basins, a record that no measurement campaign had produced.

In the record

Limits

These models learn the conditions they were shown. A model fitted to instrumented basins or recent storm seasons may drift when applied to a different climate, a different instrument or an unusual event, which is exactly the case researchers most want. Discarding incomplete observations keeps the data clean but can quietly remove the hardest conditions. The models give a number without the physical reasoning behind it, so an estimate can look plausible and still be wrong, and errors in a long reconstructed record are hard to detect when there is nothing independent to compare against.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Related

Every article