BlackBox: generalizable reconstruction of extremal values from incomplete spatio-temporal data (Q2028575)

In the frame of the 11th international conference on Extreme-Value Analysis (EVA 2019) a data competition was organized, for which several teams modeled and predicted Red Sea surface temperature extremes over space and time. The present article was a submission to the EVA 2019 and refers to the problem of predicting extremes of the Red Sea surface temperature anomaly within spatio-temporal regions of missing data. Daily temperatures at fixed points, covering a special geographical area of Red Sea and for over 31 years were provided to the teams. A certain percentage of data has been deliberately removed from the dataset. The aim was to predict the distribution of extremes of the sea surface temperature anomaly on a number of specified space-time regions. The quality of the predicted extreme values was then evaluated by a so called ``threshold-weighted continuous ranked probability score'', averaged over all predicted regions. In the present work, the authors develop a method to reconstruct missing data based on neural networks. The main steps are: one removes first more data from the dataset with the goal to train an autoencoder to recover missing data based on remaining information and then, trained models are evaluated on original data and provide stochastic plausible reconstructions of temperature anomaly within regions of missing data. Next, a direct calculation of the extreme values for the regions of interest follows. The implementation of the autoencoder model is based on convolutional deep neural networks. The entire methodology and the construction of the model, the convolutional autoencoder architecture, the training protocol and implementation are very detailed presented in the third section of the article. Implementation results and performance are discussed in the fourth section. A very interesting research paper. Actually I would like to point out the importance of competitions as EVA for developing more efficient methodologies to solve difficult practical problems.