MT-GN: Multi-Task-Learning-Based Graph Residual Network for Tropical Cyclone Intensity Estimation

Abstract

A tropical cyclone (TC) is a type of severe weather system that damages human property. Understanding TC mechanics is crucial for disaster management. In this study, we propose a multi-task learning framework named Multi-Task Graph Residual Network (MT-GN) to classify and estimate the intensity of TCs from FY-4A geostationary meteorological satellite images. And we construct a new benchmark dataset collected from the FY-4A satellite for both TC classification and intensity estimation tasks. Four different methodologies to classify TCs and estimate the intensity of TCs are fairly compared in our dataset. We discover that accurate classification and estimation of TCs, which are usually achieved separately, requires co-related knowledge from each process. Thus, we train a convolution feature extractor in a multi-task way. Furthermore, we build a task-dependency embedding module using a Graph Convolution Network (GCN) that further drives our model to reach better performance. Finally, to overcome the influence of the unbalanced distribution of TC category samples, we introduce class-balanced loss to our model. Experimental results on the dataset show that the classification and estimation performance are improved. With an overall root mean square error (RMSE) of 9.50 knots and F1-score of 0.64, our MT-GN model achieves satisfactory performance. The results demonstrate the potential of applying multi-task learning for the study of TCs.