How Google Is Using Machine Learning to Predict the Weather
Google is using machine learning to predict the weather. Named precipitation nowcasting, it focuses on forecasts for right now and up to the next 6 hours, with a total latency of 5–10 minutes. It is a play on the word forecast. Fore meaning the front of something, or situated in the front. So it is forecasting vs nowcasting.
The goal of nowcasting is to take current conditions and answer questions such as:
“Is it going to start raining when I step out to run my errands in the next 30 minutes?”
“When is this heavy rain going to let up? (How long am I going to have to stand under this awning?)”
Using radar images, Google treats this as a computer vision problem. They use a “data-driven physics-free approach,” which means they are not using atmospheric conditions and physics to predict the weather. Instead, they treat weather prediction as an image-to-image translation problem. One where image analysis of radar images and the use of convolutional neural networks (CNNs) can be utilized to predict the weather.
