Deep Learning Methods for Cyclogenesis, Extratropical, and Tropical Cyclone Regions of Interest (ROI) from Satellite Observations

At present, there is a growing abundance of satellite data both in time and dimensionality providing greater details into cyclones that has not been observed before. We tested a method for cyclogenesis, extratropical and tropical cyclone satellite image recognition using deep learning (DL). In order to train the DL models, we designed a two-step combination of a heuristic-based model and a DL model. DL is a type of machine learning that initially runs with large quantities of well-labeled data to train itself and then can be used without the initial training process on unlabeled datasets to very quickly identify regions of interest (ROI). We used a heuristic model to derive the labeled data for training the DL models, which once trained were then run on unlabeled data.

Previously, we designed the heuristic-based model that was used to label extratropical and tropical cyclone ROI from Global Forecast System (GFS) analysis outputs. This model produced a dataset of cyclone ROI containing the date, centerpoint, and bounding latitudes and longitudes from 2011-2015. Our new work has evolved in two steps. First, we created additional datasets for cyclogenesis ROI that were derived from the original cyclone ROI. For cyclogenesis, we looked at leading time periods of either 12 or 24 hours before our existing cyclone ROI. Secondly, both the cyclone and cyclogenesis datasets were overlaid onto satellite data and used as labeled data inputs to train and validate different single and multi time-series dependent DL models. Once trained, the DL models were run on unlabeled satellite data to experiment with early detection of cyclone ROI. The models ran much quicker than the heuristic labeler and accurately identified regions within satellite data of meteorological interest.