Bayesian deep learning for large scale environmental data modelling

Abstract

Deep learning – machine learning using deep neural networks – is an efficient way to discover patterns in data that may be more complex than we could manually hypothesise. Here we learn spatio-temporal models that harness information from gridded auxiliary datasets, such as digital terrain models and satellite imagery, by learning task-relevant derivatives of these with no requirement for manual feature engineering. By operating within the Bayesian probabilistic framework, we can learn well-calibrated deep models that quantify epistemic and aleatoric uncertainties and avoid overfitting despite the capacity of deep models to do so.