Analysis of Atmospheric Circulation Associated with Extreme Floods in Brazil Using Autoencoder Networks

Carlos Lima, Upmanu Lall

Frequency studies of flood hydrology are commonly based on statistical models and thereof rely on the classical assumptions of independence, homogeneity and stationarity of the flood data. In this ongoing work we aim to advance traditional flood frequency studies by investigating extreme streamflow events under the flood hydroclimatology framework, where a formal consideration of the physical mechanisms responsible for the generation of extreme floods is contemplated through the analysis of the synoptic atmospheric and oceanic fields in the days that preceded the events. Large scale fields of wind vector, sea surface temperature and moisture divergence as well as storm track data built on the integrated moisture flux in the atmosphere are evaluated in a reduced dimensional space obtained by autoencoder networks, which consist of multilayer neural networks whose goal is to reduce the dimensionality of high-dimensional input data. Extreme hydrological events in two flood-prone regions in Brazil are used as case studies for the current work and a hypothesis of the causal chain of extreme floods in such regions is offered and investigated using the proposed methodology and the machine learning tools.

Link to Recording: http://video.ucar.edu/mms/image/CI2015_carlos_lima.mp4