Tropical Cyclone Monitoring Based on Geostationary Satellite Imagery

Abstract

AbstractDue to the devastating wind, heavy floods, and coastal inundation from storm surges, tropical cyclones (TC) can cause significant damage to human lives and property. Forecasters and emergency responders both rely on accurate TC location and intensity estimates. In this chapter, two deep convolutional neural networks (CNNs) were designed for locating TC centers (CNN-L) and estimating their intensities (CNN-I) from the brightness temperature data observed by the Himawari-8 geostationary satellite. To train the CNNs, we used 97 TC incidents from the Northwest Pacific from 2015 to 2018. When compared to the Best Track dataset of TC centers and intensities as a reference, the mean location error of CNN-L model is 30 km for TCs in categories H1-H5. The best multi-category CNN classification model provided a pretty good accuracy (84.0%) and low root mean square error (RMSE, 10.19 kt) in TC intensity estimate using a combination of four channels of data as input. We added a focal_loss function to the CNN model to mitigate the negative effect of the severely imbalanced distribution of TC category samples. The accuracy increases to 88.9% if we convert the multi-classification problem to a binary classification problem, and the RMSE is dramatically lowered to 8.99 kt. Our CNN models are robust in assessing TC intensity from geostationary satellite pictures, according to the results.