Tropical Cyclone Intensity Prediction Using Spatio-Temporal Data Fusion

Tropical cyclone is a sea storm that causes important life and economic losses in the coastal regions in the tropical zone around the equator of the earth. Tropical cyclone intensity is an important characteristic used to estimate the strength of the tropical cyclone. This study aims to improve the tropical cyclone intensity prediction by concatenating the spatial and temporal features of tropical cyclones. The proposed methodology utilized the deep learning based approach for handling 3D and 2D features for 24h early intensity prediction. In the first phase, a dynamic grid-based approach is utilized to extract the spatial features in a ( $3\times 3$ ) grid format from the eye of the TC. These spatial features are extracted for four different components (u,v,t,r) and 37 different isobaric planes. In the second step, multiple convolutional layers are used to process each spatial component separately, and a fusion method is used to combine the spatial and temporal features. The proposed method achieved state-of-the-art results by reducing the MAE up to 3.31% overall and 8.5%,14.78%, 5.67% for u,v, and (u,v) add fusion components, respectively. The proposed methodology outperformed the state-of-the-art Saf-net model by 8.5 %,14.78%,5.67% for u,v, and (u,v) add fusion, respectively. A performance comparison on four real-time tropical cyclones (Bavi 2015, AERE 2016, NANMADOL 2017, HECTOR 2018) is also performed. The proposed model achieved MAE 2.92, 2.99, 2.46, 3.95 that are 10.08%, 34.35%, 23.65%, and 3.2% lower than state-of-the-art spatio-temporal models, respectively.