Tracking of atmospheric phenomena with artificial neural networks: a supervised approach.

Tracking of atmospheric phenomena, such as Tropical Cyclones (TC) and Polar Mesocyclones (PMC), — is a frequently performed task in the climate sciences. It is a crucial part of any climatic or short-period study since the occurrence and characteristics of a phenomenon at some moment carries significantly less information compared to its evolution. In some cases of synoptic-scale phenomena, this problem has established well-developed solutions (e.g., for extratropical cyclones, ETC). However, in the majority of other cases, there are no reliable tracking algorithms at the moment. In this study, we present the generic framework for the tracking of atmospheric phenomena of an arbitrary kind in the source data of nearly arbitrary origin. The core entity of our approach is the function of dissimilarity of representations of a phenomenon in two subsequent data frames which is learned under supervision of an artificial neural network. As an implementation of the proposed approach, we present the results of the tracking of Tropical Cyclones and Polar Mesocyclones. We demonstrate the success of our approach in terms of formal metric MOTA (Multiple Object Tracking Accuracy) and the empirical distributions of key lifecycle characteristics of TC and PMC. [ABSTRACT FROM AUTHOR]