A Deep Learning-based Physics-aware Model for Detecting Land-Sea Breeze over Tropical Islands.

Land-sea breeze circulation is one of the most prominent and well-observed meteorological phenomena in coastal zones. While the land-sea breeze and its influence are well-documented and observable using mesoscale-resolving operational NWP models over larger islands, it is typically much weaker and more challenging to detect over smaller islands due to their limited land mass, which cannot generate strong pressure gradients between land and the ocean. We use high-resolution observations from a Department of Energy - Atmospheric Radiation Measurement (DOE-ARM) site on Manus island (1639 square km) in the tropical western Pacific to detect signatures of land-sea breeze circulations. These signatures demonstrate a nonlinear relationship between the land-sea breeze circulations, the topography of the island, and surrounding environmental conditions. We leverage this nonlinear relationship to develop a spatially-aware deep learning model to detect land-sea breeze circulations. The model is trained over different land structures to detect circulation in the atmosphere, irrespective of the size of the landmass. The work is further extended to another DOE-ARM site located on Nauru island (21 square km), which is even smaller than Manus, to advance our understanding of the land-sea breeze dynamics.