The goal of this AI challenge is to establish public baselines and to advance state-of-the-art in the field of weather nowcasting.   Below we provide answers to some basic questions regarding nowcasts and the challenge.

What are Weather Nowcasts?

Nowcasts (https://en.wikipedia.org/wiki/Nowcasting_(meteorology)) are short-term forecast of weather variables typically measured by weather radar or satellite.   Nowcasts are different from traditional weather forecasts (like those you see on the news, or on your phones weather app) in that they are based (mostly) on statistical extrapolations of recent data, rather than full physics-based numerical weather prediction (NWP) models.   The advantages of using statistical extrapolation techniques is that they generally run much faster than NWP models (seconds compared to hours), and because of this, nowcasts are able to  leverage the most recently observed data and use it for forecasting.  Because of this extremely low latency, nowcast out-perform traditional weather models in terms of accuracy and precision for short look-aheads (typically 1-2 hours at most).  

How are nowcast computed

Nowcast are computed in a variety of ways, but one of the most common approaches is to apply optical flow techniques to a sequence of radar images.   These techniques track the motion of storm objects, which is then used to extrapolate the location of storms into the future.   See e.g. https://github.com/hydrogo/rainymotion. Recently there has been a great deal of interest in applying deep learning to this problem.   The goal of this challenge is to assess the skill of many different nowcasting techniques and decide which method for creating nowcast perform the best.

How are weather nowcasts used today?

Nowcasts are used in a variety of application where short range and very precise forecasts are required.   These applications include aviation weather forecasting for terminal and enroute, HADR applications like monitoring of forecast fires, maritime applications, and for tracking severe weather for public safety.  Performance matters, and hopefully this challenge will produce the best performing nowcast for a variety of applications.  

What data is used in this challenge?

The nowcasting challenge will use the radar imagery in the SEVIR dataset.  See here for more information on this dataset, including how to download it.

What exactly is the challenge task?

The challenge is posed as a video prediction task similar to that described in (ref neurons paper). In this task, each input in the training and test datasets represents an hour long sequence of radar images.  The associated target for each of these sequence is the radar sequence observed in the following hour.   The goal is to develop an algorithm that uses the input to estimate the output sequence.

How do I submit my solution

If you have a solution that improves upon the benchmarks provided,   you may submit an updated results table as a pull request, along with a reference to a write-up of your method.

Where do I start?

Data and code for some benchmark nowcasting models can be found here.