BlackBox: Generalizable Reconstruction of Extremal Values from Incomplete Spatio-Temporal Data

We describe our submission to the Extreme Value Analysis 2019 Data Challenge\nin which teams were asked to predict extremes of sea surface temperature\nanomaly within spatio-temporal regions of missing data. We present a\ncomputational framework which reconstructs missing data using convolutional\ndeep neural networks. Conditioned on incomplete data, we employ\nautoencoder-like models as multivariate conditional distributions from which\npossible reconstructions of the complete dataset are sampled using imputed\nnoise. In order to mitigate bias introduced by any one particular model, a\nprediction ensemble is constructed to create the final distribution of extremal\nvalues. Our method does not rely on expert knowledge in order to accurately\nreproduce dynamic features of a complex oceanographic system with minimal\nassumptions. The obtained results promise reusability and generalization to\nother domains.\n