Detecting Cloud Cover in Webcam Images using Neural Networks: A Nowcasting Application

Cloud cover is an important parameter in public weather forecasting. When numerical weather prediction (NWP) models fail to represent the current weather accurately, observations are often used in post-processing to correct these model errors. However, traditional ground-based measurements of cloud cover are often sparse. Webcams on the other hand, are becoming more prevalent and many are directed at the sky.

We show how a convolutional neural network can be trained to detect cloud cover from webcam images. The network is trained using over 80,000 manually labeled images from around 100 high quality cameras in Norway. A team of meteorologists labeled the images with oktas, ranging in value from 0 (no clouds) to 8 (overcast).

The neural network then estimates cloud cover from new images received in near real-time. These cloud cover estimates are used to correct output from an NWP model by using a statistical interpolation technique. The end result is a gridded nowcast of cloud cover that better reflects the current conditions.