~/aixsci
weekly updates

fields/climate-weather/model families

Convolutional neural networks in weather, climate and oceans

How convolutional neural networks, which learn patterns from grids of numbers, are used to forecast weather, study climate and read the ocean from satellite and radar data.

How it is used

A convolutional neural network is a model that slides small filters across a grid of numbers, learning to spot local patterns such as edges, swirls or sharp changes, then combining them into larger ones. Weather, climate and ocean data arrive naturally in that form: maps of temperature, wind and rainfall laid out on a grid, stacked through time and depth. In the examples the filters read sequences of radar images to produce the next images, and read satellite pictures of sea surface temperature to infer the currents beneath. Elsewhere convolution is used only as a first step, turning raw fields into compact summaries that a different kind of model then works on.

Two patterns recur across the examples. One is the encoder-decoder, where convolutions compress a map into a condensed description and then build a new map back out of it, with direct connections between matching layers so fine detail is not lost. The other is convolution as a front end: a stack of filters reads the input, and attention layers or a recurrent layer handle how things change over time. The examples also show convolution applied to single time series rather than maps, with the filters reading along time instead of space to classify how a system is behaving.

Why researchers use it

Physical simulation of the atmosphere and ocean is expensive, and some things it cannot do at all: radar echoes minutes ahead, or currents at the scale of a few kilometres seen from orbit. Convolutional networks offer a cheap function fitted to observations instead. They suit the data because weather is local: what happens at a point depends mostly on its neighbourhood, which is exactly the assumption built into a sliding filter. The same filter is reused everywhere on the map, so the model has far fewer things to learn than one which treats every grid point separately, and it can be trained on patches and applied to larger regions.

Researchers also reach for convolution when the input is an image in all but name. Satellite and radar archives are long, regular and already gridded, which makes them unusually good training material in a field where labelled data is often scarce. In the examples training data is sometimes generated rather than observed: output from an ocean circulation model, or time series simulated from simple dynamical equations, used to teach a network a relationship that can then be tried on real measurements. Convolution stands in for a hand-written rule about how neighbouring values relate, learned from data instead of derived.

In the record

Limits

A sliding filter assumes the answer depends on a local neighbourhood, so these models struggle with influences that travel far across the globe, and must be extended with attention or recurrence to handle long spans of time. They learn the statistics of their training data: a network taught on simulated oceans or idealised equations may not transfer to real measurements, and one taught on past decades may mislead as the climate shifts. They tend to blur fine detail and under-represent rare extremes, which are often what matters most. Missing observations are awkward, since the grid must be complete.

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

Related

Every article