Deep Spatiotemporal Learning for Climate Science

Climate systems are highly complex and modelling their dynamics is hard. In recent times, there has been an explosion of climate and geophysical data from satellites and sensory instruments. With increased data, machine learning methods have become viable climate modelling techniques and have been adopted in many forecasting tasks. Gaussian processes have been used for a long time in geostatistics for the probabilistic modelling of spatiotemporal processes. However, Gaussian processes are fundamentally shallow models and their expressiveness is limited by the choice of kernel function. Recently, deep learning algorithms have proven very successful for learning spatiotemporal processes. Convolutional Neural Networks (CNNs), traditionally developed for image recognition, are able to uncover complex hierarchical correlations in spatial data that Gaussian processes cannot. Traditional deterministic CNNs and deep learning algorithms however, lose the benefits of probabilistic modelling, such as uncertainty quantification, and are often difficult to interpret. With increasing research into deep Gaussian processes and deep convolutional Gaussian processes, this project will seek to develop deep probabilistic spatiotemporal models to bring together the benefits of both deep learning and Gaussian process, for use in climate science forecasting.