Using Deep Learning to Identify Tropical Easterly Waves, the Intertropical Convergence Zone, and the Monsoon Trough

Abstract Current tropical easterly wave (TEW) tracking algorithms struggle over the Caribbean Sea, and no algorithm exists to automatically identify and distinguish between the intertropical convergence zone (ITCZ) and the monsoon trough (MT). ERA5 reanalysis data and human-analyzed labels of TEWs and the ITCZ/MT are used to train convolutional neural networks (CNNs) to identify these features in the eastern North Pacific and Atlantic Oceans. The CNNs successfully capture TEW activity over the Caribbean and are able to distinguish between the ITCZ and the MT. The CNNs are used to generate a climatology of TEW and ITCZ/MT objects from 1981 to 2023. TEW vorticity and moisture signatures are significantly weaker in the Caribbean than in the open Atlantic, but these weaker waves are still readily identified by a characteristic local moisture maximum and curvature in wind fields. The MT is a deeper vorticity feature than the ITCZ, and the MT is also generally moister than the ITCZ when comparing at similar longitudes and latitudes. The ITCZ/MT boundary in the eastern North Pacific moves eastward during the strongest summer El Niño events. The MT has expanded westward in the Atlantic over the past few decades from 15 July to 30 September time frame. The dataset of TEW and ITCZ/MT object locations generated in this study can be used for further investigation into rainfall, circulation, and tropical cyclogenesis patterns associated with these features. The CNNs used to identify these features can also help operational analysts to identify these features in real time. Significance Statement Tropical waves and the intertropical convergence zone/monsoon trough directly impact people through wind, rain, and hurricane formation in the Pacific and Atlantic Oceans. Existing methods for identifying these important weather patterns do not produce sufficiently accurate real-time coverage. We create a machine learning methodology for identifying these weather features, and we use this methodology to create a record of these weather events going back to 1981. This new record can be used by researchers to learn more about these weather features, and the machine learning methodology can help human analysts to identify these features in real-time data.