Machine Learning Approaches for Polar Lows Detection in Numerical Atmospheric Model Data

Polar mesocyclones (PMCs) are small but intensive atmospheric vortexes. They usually occur above oceans in high latitudes and are characterized by a short lifetime, relatively small diameter and high surface wind speed. They are dangerous for marine infrastructure and affect deep ocean convection. Conventional method of identifying PMCs is tracking them on satellite images. Recently, high-resolution atmospheric models are also used for identifying PMCs, but this method is not highly accurate because of the absence of objective criterias for identifying PMCs. A novel method of estimating efficiency of criteria to identify polar lows in numerical atmospheric models in high-resolution atmospheric models is introduced in this paper. Four classical machine learning models were trained on all available criteria in the scientific literature. After feature selection and comparing feature importance of selected criteria in different models we conclude that the most important criteria are mcao700, difference between temperature on 700 hPa and sea surface temperature, absolute vorticity on 850 hPa and sea level pressure. These criteria are selected by most ML models and they are well explained by physical mechanism of PMCs. Machine learning methods and atmospheric model data analysis is implemented in python programming language. RAS NAAD reanalysis and calendars of objective PMCs are used as data sources.