~/aixsci
weekly updates

models/DGMR (deep generative model of radar)

DGMR (deep generative model of radar)

DGMR is a rainfall forecasting model that learns from past radar pictures and then generates whole plausible sequences of how the rain might move and change over the next few hours.

What it is

Radar is a way of seeing rain from the ground: a transmitter sends out radio pulses, raindrops bounce some energy back, and the strength of the echo becomes a map of how hard it is raining. Repeat the scan every few minutes and you get a film of weather moving across a region. DGMR, short for deep generative model of radar, is a computer model trained on many such films. It takes the most recent frames and produces what comes next. The word generative matters: rather than printing one average answer, it draws samples, so each run gives a different but internally consistent future. Run it many times and the spread of samples describes the uncertainty.

Inside, the model has two halves. One half reads the recent radar frames and compresses them into a summary of the current state of the rain. The other half injects randomness, drawn from a simple bell-shaped distribution of numbers, which is what makes each sample different. A recurrent part then steps forward in time, adding one future frame after another, each conditioned on the summary and on the frames it has already made. Training pits the model against critics, other networks whose job is to tell real radar films from invented ones. The model improves by making films the critics cannot separate from the real thing.

Why scientists use it

Forecasts for the next couple of hours, called nowcasts, are a hard case. The big physics-based weather models that work well for tomorrow need time to settle and run at too coarse a grid to track an individual shower. The usual quick alternative is to assume the rain pattern simply drifts along at its current speed, which works briefly but cannot invent new showers or let old ones die. A model learned directly from radar can pick up those habits of growth and decay from data instead of from equations.

The other reason is the look of the output. Models trained to minimise average error tend to hedge: they smear rain into a blur that is right on average and useless in detail, because a blur is the safest single guess. A generative model is pushed instead towards outputs that resemble real radar, with sharp edges, realistic texture and heavy cores in the right places. That matters for the people who act on forecasts, who need to know whether a specific place will be hit hard, not whether a vague wash of drizzle is likely somewhere.

In the record

Limits

A model like this learns what rain has looked like, not why it falls, so it is only trusted for the next hour or two and fades towards the average beyond that. Rare extremes are scarce in the training record and may be underdone. Realistic texture can read as confidence: a sharp, plausible picture of a storm in the wrong place still looks convincing, and judging it needs many samples rather than one. It also inherits whatever the radar network misses or mismeasures, and may not transfer to other regions.

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

Model family

Every article

1articles
0structural-biology
0materials-chemistry
0astronomy
1climate-weather