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

Abstract

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