Peter Uhe
Peter is Principal Scientific Developer at Fathom, a company specialising in flood risk. His role involves solving a mix of scientific challenges to improve the performance of Fathom’s models, accounting for climate change, and using machine learning to develop new datasets. Peter has a PhD from the University of Bristol and is based in Melbourne.
Innovation Title:
Computer vision: from measuring terrain to downscaling climate data
Abstract:
Flooding is a hazard which is highly susceptible to changes in climate and increases in extreme rainfall. Estimating these changes is a huge challenge, and approaches are turning towards using machine learning and AI to help solve these problems. Computer vision is a field which is readily relatable to processing of maps and geographic data, so we present two applications of its use in developing improved datasets.
Firstly, a key determinant of flood model skill is the quality of terrain information, and we have recently developed an improved global terrain map using a novel vision transformer model to improve on existing data. This terrain map allows us to model global flood risk at 30m resolution globally. Secondly, my team is developing a diffusion model to downscale coarser resolution climate models to higher resolution. This will help to include a more granular representation of the changes in rainfall in future climates and make use of the statistics available from large ensembles of global climate models.
Previous post:
Coming up next: