Short-Term Rolling Prediction of Tropical Cyclone Intensity Based on Multi-Task Learning with Fusion of Deviation-Angle Variance and Satellite Imagery

Tropical cyclones (TCs) are one of the most serious types of natural disasters, and accurate TC activity predictions are key to disaster prevention and mitigation. Recently, TC track predictions have made significant progress, but the ability to predict their intensity is obviously lagging behind. At present, research on TC intensity prediction takes atmospheric reanalysis data as the research object and mines the relationship between TC-related environmental factors and intensity through deep learning. However, reanalysis data are non-real-time in nature, which does not meet the requirements for operational forecasting applications. Therefore, a TC intensity prediction model named TC-Rolling is proposed, which can simultaneously extract the degree of symmetry for strong TC convective cloud and convection intensity, and fuse the deviation-angle variance with satellite images to construct the correlation between TC convection structure and intensity. For TCs’ complex dynamic processes, a convolutional neural network (CNN) is used to learn their temporal and spatial features. For real-time intensity estimation, multi-task learning acts as an implicit time-series enhancement. The model is designed with a rolling strategy that aims to moderate the long-term dependent decay problem and improve accuracy for short-term intensity predictions. Since multiple tasks are correlated, the loss function of 12 h and 24 h are corrected. After testing on a sample of TCs in the Northwest Pacific, with a 4.48 kt root-mean-square error (RMSE) of 6 h intensity prediction, 5.78 kt for 12 h, and 13.94 kt for 24 h, TC records from official agencies are used to assess the validity of TC-Rolling.