Tropical Cyclone Track Prediction with an Encoding-to-Forecasting Deep Learning Model

Abstract

Abstract Recently, with the accumulation of remote sensing data, the traditional tropical cyclone (TC) track prediction methods (e.g., dynamic methods and statistical methods) have limitations in prediction efficiency and accuracy when dealing with a large amount of data. However, deep learning methods begin to show their advantages to capture the complex spatiotemporal features in high-dimensional data. The task of TC track prediction based on remote sensing images can be formulated as a spatiotemporal sequence-to-sequence problem. Therefore, a novel encoding-to-forecasting model with convolutional long short-term memory (ConvLSTM) and spatial attention network (SAN-EFSModel) was proposed to predict TC tracks in this paper, which can fully extract the long-term spatial and temporal features. The proposed model was evaluated on the real remote sensing images in the western North Pacific Ocean to forecast 24-h TC tracks. Compared with ECMWF-HRES model and NCEP-GFS model, the proposed method has a better prediction accuracy in the testing set with an average position error about 30 km less. Compared with the deep learning methods, the proposed method also has the best performance. Significance Statement Tropical cyclones have a great impact on human life and natural environment due to their high frequency of occurrence, heavy degree of harm, wide impact range, and long disaster chain. In this article, we propose an encoding-to-forecasting model with convolutional long short-term memory (ConvLSTM) and spatial attention network (SAN-EFSModel) to predict tropical cyclone tracks, in which the SAN module and the convolution module of ConvLSTM have the ability to extract spatial features, and the long short-term memory (LSTM) module of ConvLSTM can extract temporal features from remote sensing images and tropical cyclone tracks. The results show that the average position error (APE) of our proposed method has about 30% improvement in a 24-h forecast compared with prevailing numerical weather prediction models.