Case Study: WeatherBench [PeterDueben]

WeatherBench provides a new benchmark to test data‐driven approachesto weather forecasting. Traditional weather models are based on the discretizedequations governing the atmosphere. They perform very well for many tasks butare still found lacking for some others. Data‐driven approaches, such as deeplearning, directly learn from the best available observations and couldpotentially produce better forecasts. In this paper, we define a benchmarktask—predicting pressure and temperature across the globe 3 and 5 daysahead—which will hopefully lead to progress in data‐driven weather predictionand foster collaboration across disciplines.

Additional Case Study:MetNorway

MetNorway use crowd-sourced in-situ weather observations in order tofill in data-gaps. This is used to improve the accuracy of their weatherprediction model - and therefore allows them to provide more accurateforecasts.

In Norway, these in-situ observations are taken from sensors that arewidely owned by the general population - but in the future, these observationsmay be able to be taken on a much smaller scale using the internet of things -for example using mobile phones.

‍