Google updates WeatherNext model to use raw satellite data

In short: Google has released WeatherNext 3, an AI weather forecasting model that now uses more raw satellite data and can update forecasts every hour.

What happened

Google released version 3 of its WeatherNext weather model, described in a new research paper. The biggest change is that it now “ingests” some satellite observations, meaning it takes in fresh measurements from weather satellites instead of relying only on a pre-packaged global dataset.

Many AI weather models mainly use something called a reanalysis (a stitched-together global snapshot made by blending lots of weather readings into one consistent picture). Reanalyses are typically produced every six hours, which can create a time lag between what is happening outside and when the model can start a new forecast.

WeatherNext 3 adds satellite data so it can run forecasts hourly. Google also increased the model’s detail level, made the model larger, and added an extra system focused on rainfall using satellite-based precipitation estimates. The model also uses a small amount of basic location information, like whether a point is land or ocean and its elevation, to improve local temperature and dew point predictions.

Why it matters

More frequent updates can be useful when weather changes quickly, like during storms. It is similar to checking traffic with more recent road reports instead of waiting for a summary that only updates a few times a day.

Google says WeatherNext 3 now powers weather forecasts across Google services, including Search, Gemini, and Maps. The paper reports accuracy gains in some areas, but it also notes some odd behavior, including weaker performance in the first few hours for certain measures and visible grid-shaped artifacts in some maps.

Source: Arstechnica