Methods in machine learning for probabilistic modelling of environment, with applications in meteorology and geology

Abstract

Earth scientists increasingly deal with ‘big data’. Where once we may have struggled to obtain a handful of relevant measurements, we now often have data being collected from multiple sources, on the ground, in the air, and from space. These observations are accumulating at a rate that far outpaces our ability to make sense of them using traditional methods with limited scalability (e.g., mental modelling, or trial-and-error improvement of process based models). The revolution in machine learning offers a new paradigm for modelling the environment: rather than focusing on tweaking every aspect of models developed from the top down based largely on prior knowledge, we now have the capability to instead set up more abstract machine learning systems that can ‘do the tweaking for us’ in order to learn models from the bottom up that can be considered optimal in terms of how well they agree with our (rapidly increasing number of) observations of reality, while still being guided by our prior beliefs.

In this thesis, with the help of spatial, temporal, and spatio-temporal examples in meteorology and geology, I present methods for probabilistic modelling of environmental variables using machine learning, and explore the considerations involved in developing and adopting these technologies, as well as the potential benefits they stand to bring, which include improved knowledge-acquisition and decision-making. In each application, the common theme is that we would like to learn predictive distributions for the variables of interest that are well-calibrated and as sharp as possible (i.e., to provide answers that are as precise as possible while remaining honest about their uncertainty). Achieving this requires the adoption of statistical approaches, but the volume and complexity of data available mean that scalability is an important factor — we can only realise the value of available data if it can be successfully incorporated into our models.