Our goal was that every mobile user could participate and share observations on present weather and its impacts. Therefore, we decided to keep it simple, rather than exhaustive and detailed, which challenged our culture of weather experts...
In other words, we favored quantity over quality of the weather reports. And it is a success! The amount of real-time human observations that we have today is the largest we have ever had. This new source of information is already used every day by our forecasters in operation. It represents a complement to other sources of data or even helps forecaster to track weather phenomena that other observation systems can not detect.
But to be really useful, these large amounts of crowd-sourced data must be filtered and classified automatically.
Weather observations from mobile users are filtered in order to eliminate irrelevant observations. Machine Learning methods have also been tested (anomaly detection using the multivariate gaussian distribution) and shew that the ratio of fakes was rather low.
Deep Learning image classification methods are used to filter the pictures taken by mobile users before publication. These methods are also tested to identify the type of weather on the images, in order to provide information on weather phenomena or impacts that are very difficult to observe, for instance, snow cover on roads.
![](images/flag.png)
![](images/blue_ribbon.png)