Satellite nowcasting of cloud coverage via machine learning

Satellites orbiting the earth make it possible to provide increasingly accurate weather forecasts. Through observations and numerical weather predictions, meteorologists can make near-future weather forecasts which have a major impact on everyday decisions in different societal sectors. However, unexpected cloud appearances often lead to inaccuracies in nowcasting weather or radiation. We propose the usage of machine learning via a Convolutional LSTM Neural Network. In particular, we validate the performance of learning spatial and temporal patterns within a sequence of satellite images using a Convolutional Long Short-Term Memory model architecture for cloud coverage prediction. The accuracy of the machine learning model was found to be 90 percent upon training on a dataset spanning a period of two years. However, additional efforts are deemed necessary to address and remove biases present within the model.