Microsoft separates the organized structure of precipitation, which is predictable for a long time, from the local noise that must be modeled generatively. Tested on the United States, Europe and China.
Generative models applied to radar data can produce hyperlocal rainfall forecasts, but their accuracy on intense precipitation collapses after the first few hours — which limits the warning during flash floods. MW-Nowcast's (Microsoft Weather Nowcast) bet is based on a simple observation: the structure of a storm system remains predictable longer than each of the cells that make it up.
The model thus learns two things together: a deterministic predictor that captures the organized structure shared by all members of the whole, and a generator that produces various local residuals around this common prediction—growth, dissipation, reorganization, and thunderstorm initiation. The evaluation is based on independent test sets in the United States, Europe, and China.
Two hours saved, and a real decision value at 4-6 hours
Over the entire six-hour horizon, MW-Nowcast has a better detection capacity than the reference methods for heavy and extreme rainfall. For the most intense rainfall, it doubles the warning time available in the three regions: a 6-hour forecast reaches the quality that the best generative reference only held up to 3 hours. A cost-loss analysis shows that the model retains a substantial decision-making value over the 4-6 hour range, where other approaches provide almost nothing — the interval in which the population is effectively alerted.
Read the full article on the original source
Go to the source →