Tropical Surface Analysis Using Deep Learning

Tropical waves and the intertropical convergence zone (ITCZ) / monsoon trough impact human interests directly through wind and rainfall and indirectly through their relationship to tropical cyclogenesis. The Tropical Analysis and Forecast Branch (TAFB) of the National Hurricane Center releases an analysis of current weather features in the tropical North Atlantic and Northeast Pacific every six hours. The text discussions of these analyses are available from 2003 to the present and include location descriptions for features like tropical waves and the ITCZ / monsoon trough. The archive of TAFB discussions is thus a trove of information on significant tropical weather, though its text-based format and limited temporal record are obstacles to easily using this information in comprehensive research.

We use tropical wave and ITCZ / monsoon trough locations extracted from TAFB’s discussions and ERA5 reanalysis data to train two neural networks based on the U-Net 3+ architecture. One network is trained to identify tropical wave axes, while the other is trained to identify and distinguish between the ITCZ and monsoon trough. Both of our networks demonstrate significant skill at identifying the locations of their respective target objects when verifying against TAFB’s feature labels. We also compare the performance of our wave-classifying network against existing objective wave tracking algorithms and identify conditions under which our classifier’s performance decreases. We present a climatology of wave and ITCZ / monsoon trough locations across the North Atlantic and East Pacific based off of our network outputs. The effectiveness of both of our networks suggests that they will serve as useful tools for studying different regimes of tropical cyclogenesis in the reanalysis era and that they may be a helpful tool for operational tropical surface analysis.