23AI Spatial-temporal Convolution Neural Networks for Tropical Cyclone Detection from Geostationary Satellite Images

In this study, a deep learning approach is investigated to detect and classify tropical cyclones (TCs) according to its stage in an end-to-end manner. A two-stream Convolution Neural Network (CNN) which captures both spatial and temporal features of time-series of geostationary satellite cloud images (SCIs) is proposed. A dataset of Himawari-8 SCIs of 104 TC over Western North Pacific and Bien Dong Sea from 2014 to 2019 are collected. Additional inputs, namely atmospheric motion vector (AMV), related to tropical cyclone circulation features for the learning processes are derived by tracking clouds through consecutive SCIs. Results show that our proposed model significantly improves the performance of TC detection in comparison to other state-of-the-art based CNN models. Moreover, our case study for the detection of TC Doksuri in 2017 yielded promising results as our proposed model is capable of detecting TC formation 24 hours prior to the first time reported by the best-track data