Classification of tropical cyclone containing images using a convolutional neural network: performance and sensitivity to the learning dataset

Abstract. Tropical cyclones (TCs) are one of the most devastating natural disasters, which justifies monitoring and prediction on short long timescales in context a changing climate. In this study, we have adapted tested convolutional neural network (CNN) for classification reanalysis outputs according to presence or absence TCs. This study compares performance sensitivity CNN learning dataset. For purpose, chose two meteorological reanalysis, ERA5 MERRA-2, used number variables from them form TC-containing background images. The TCs is labeled HURDAT2 Special attention was paid design image set make sure it samples similar locations times We assessed using accuracy but also more objective AUC AUPRC metrics. Many failed classifications can be explained by context, such as situation with cyclonic activity not yet classified HURDAT2. impact spatial interpolation “mixing matching” training test sets CNN. showed that applying an ERA5-trained MERRA-2 images works better than MERRA-2-trained