An Objective Technique for Typhoon Monitoring with Satellite Infrared Imagery

Abstract

In this paper, an objective technique was developed for monitoring typhoons over the Northwestern Pacific Ocean with Himawari-8 geostationary satellite infrared imagery. Two convolutional neural networks (CNNs) were designed to locate a typhoon and estimate its intensity, respectively. The mean error of the typhoon center location (CNN-Location) model is 5.4 pixels (54 km), and the top-1 accuracy and root mean square error (RMSE) of the intensity estimation (CNN-Intensity) model are 79.6% and 11.66 kt, respectively. By changing the loss function from categorical_crossentropy to focal_loss in the CNN-Intensity model, higher top-1 accuracy of 82.9% and lower RMSE of 10.84 kt are obtained. The results demonstrate that CNN has great potential in the application of automatic typhoon monitoring.