ML has the potential to apply to any field. Google today shared work on “Radar Images Machine Learning for Precipitation Nowcasting” which hopes to answer how forecasts of localized rain storms and other short-term weather events remain “particularly challenging.”
Today, numerical methods that simulate directly atmospheric dynamics, ocean effects, thermal radiation, and other processes / effects are limited by limits of computational resources. The National Oceanic and Atmospheric Administration (NOAA), for example, collects nearly 100 terabytes of data a day.
The company wants to address short-term predictions— known as “nowcasting” — that are not being well supported by current methods. This is useful for “immediate decisions” such as traffic control, logistics and even planning for evacuations.
Google’s solution uses radar data, and treats weather prediction as an issue with computer vision. The neural network must learn about “atmospheric physics from the examples of training alone, not by a priori knowledge of how the atmosphere actually works.”
In the short term, Google’s ML-powered rain forecasting “outperforms all three of these models” compared with three widely used forecasting models.
Google is looking forward to combining its system with a current method known as High Resolution Rapid Refresh (HRRR), which is better for long-term forecasts by using a physical 3D model.