Tropical Cyclone Intensity Estimation From Geostationary Satellite Imagery Using Deep Convolutional Neural Networks

In this study, a set of deep convolutional neural networks (CNNs) was designed for estimating the intensity of tropical cyclones (TCs) over the Northwest Pacific Ocean from the brightness temperature data observed by the Advanced Himawari Imager onboard the Himawari-8 geostationary satellite. We used 97 TC cases from 2015 to 2018 to train the CNN models. Several models with different inputs and parameters are designed. A comparative study showed that the selection of different infrared (IR) channels has a significant impact on the performance of the TC intensity estimate from the CNN models. Compared with the ground truth Best Track data of the maximum sustained wind speed, with a combination of four channels of data as input, the best multicategory CNN classification model has generated a fairly good accuracy (84.8%) and low root mean square error (RMSE, 5.24 m/s) and mean bias (-2.15 m/s) in TC intensity estimation. Adding attention layers after the input layer in the CNN helps to improve the model accuracy. The model is quite stable even with the influence of image noise. To reduce the side-effect of the very unbalanced distribution of TC category samples, we introduced a focal_loss function into the CNN model. After we transformed the multiclassification problem into a binary classification problem, the accuracy increased to 88.9%, and the RMSE and the mean bias are significantly reduced to 4.62 and -0.76 m/s, respectively. The results show that our CNN models are robust in estimating TC intensity from geostationary satellite images.