In a blog post and accompanying paper, researchers at Google detail an AI system — MetNet — that can predict precipitation up to eight hours into the future. They say that it outperforms the current state-of-the-art physics model in use by the U.S. National Oceanic and Atmospheric Administration (NOAA) and that it makes a prediction over the entire U.S. in seconds as opposed to an hour.
It builds on previous work from Google, which created an AI system that ingested satellite images to produce forecasts with a roughly one-kilometer resolution and a latency of only 5-10 minutes. And while it’s early days, it could lay the runway for a forecasting tool that could help businesses, residents, and local governments better prepare for inclement weather.
MetNet takes a data-driven and physics-free approach to weather modeling, meaning it learns to approximate atmospheric physics from examples and not by incorporating prior knowledge. Specifically, it uses precipitation estimates derived from ground-based radar stations and measurements from NOAA’s Geostationary Operational Environmental Satellite that provide a top-down view of clouds in the atmosphere. Both sources cover the continental U.S., providing image-like inputs that can be processed by the model.
MetNet is executed for every 64-by-64-kilometer square covering the U.S. at a 1-kilometer resolution. As the paper’s authors explain, the physical coverage corresponding to each output region is much larger — a 1,024-by-1,024-kilometer square — since the model must take into account the possible motion of the clouds and precipitation fields over time. For example, to make a prediction 8 hours ahead, assuming that clouds move up to 60 kilometers per hour, MetNet needs 480 kilometers (60 x 8) of context.
Above: Performance evaluated in terms of F1-score at 1.0 millimeter per hour precipitation rate (higher is better). The neural weather model (MetNet) outperforms the physics-based model (HRRR) currently operational in the U.S. for timescales up to 8 hours ahead.
MetNet’s spatial downsampler component decreases the memory consumption while finding and retaining the relevant weather patterns, and its temporal encoder encodes snapshots from the previous 90 minutes of input data in 15-minute segments. The output is a discrete probability distribution estimating the probability of a given rate of precipitation for each square kilometer in the continental U.S.
One key advantage of MetNet is that it’s optimized for dense and parallel computation and well-suited for running on specialty hardware such as Google-designed tensor processing units (TPUs). This allows predictions to be made in parallel in a matter of seconds, whether for a specific location like New York City or for the entire U.S.
The researchers tested MetNet on a precipitation rate forecasting benchmark and compared the results with two baselines — the NOAA High Resolution Rapid Refresh (HRRR) system, which is the physical weather forecasting model currently operational in the U.S., and a baseline model that estimates the motion of the precipitation field, or optical field. They report that in terms of F1-score at a precipitation rate threshold of 1 millimeter per hour, which corresponds to light rain, MetNet outperformed both the flow-based model and HRRR system for timescales up to 8 hours ahead.
“We are actively researching how to improve global weather forecasting, especially in regions where the impacts of rapid climate change are most profound,” wrote Google research scientists Nal Kalchbrenner and Casper Sønderby. “While we demonstrate the present MetNet model for the continental U.S., it could be extended to cover any region for which adequate radar and optical satellite data are available. The work presented here is a small stepping stone in this effort that we hope leads to even greater improvements through future collaboration with the meteorological community.”
Author: Kyle Wiggers.
Source: Venturebeat
