Cloud Classification Using Artificial Intelligence

Stratocumulus clouds cover more of the planet than any other cloud type, making them extremely important to the planet’s radiation budget. This study will look at one of the largest stratocumulus cloud decks in the world which is found over the Atlantic Ocean, off the west coast of Africa. Cloud classification schemes have been made to identify these clouds without the use of spatial and structural data, although inclusion of such information may improve classification accuracy. Using machine learning techniques with a capability of producing higher spatial and temporal resolution cloud fields than what is currently available would aid in the process of assessing the relationships between stratocumulus clouds and their radiative effects. To promote this, a deep learning model was created using data from the geostationary Spinning Enhanced Visible and Infrared Imager (SEVIRI) satellite. Data was used from a combination of both visible and IR SEVIRI’s channels focusing on the area over the Atlantic Ocean off the west coast of Africa. Unsupervised K-means clustering was first used to explore the data. The k-means clustering performed moderately well at identifying cloud types at 4 or 5 clusters. The next step included the development of a sequential convolutional neural network (CNN) model that was used in a supervised manner to identify open and closed cell stratocumulus clouds. Three bands of SEVIRI satellite channels (visible, water vapor, and infrared) are used as input during the training step of the model, while the ground truth dataset contained four classes (open, closed, clear, unknown). We will present the model’s training and test accuracies for the various architecture and parameters tested, and will discuss its pros and limitations for the task in hand.