Abstract:Tropical cyclone (TC) is an extreme tropical weather system and its trajectory can be described by a variety of spatio-temporal data. Effective mining of these data is the key to accurate TCs track forecasting. However, existing methods face the problem that the model complexity is too high or it is difficult to efficiently extract features from multi-modal data. In this paper, we propose the Dual-Branched spatio-temporal Fusion Network (DBF-Net) -- a novel multi-horizon tropical cyclone track forecasting model which fuses the multi-modal features efficiently. DBF-Net contains a TC features branch that extracts temporal features from 1D inherent features of TCs and a pressure field branch that extracts spatio-temporal features from reanalysis 2D pressure field. Through the encoder-decoder-based architecture and efficient feature fusion, DBF-Net can fully mine the information of the two types of data, and achieve good TCs track prediction results. Extensive experiments on historical TCs track data in the Northwest Pacific show that our DBF-Net achieves significant improvement compared with existing statistical and deep learning TCs track forecast methods.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to solve several key problems in tropical cyclone (TC) trajectory prediction: 1. **Model complexity and feature extraction efficiency**: - Existing methods are either too complex in model or difficult to extract features from multi - modal data efficiently. For example, traditional methods such as statistical regression and numerical weather prediction (NWP) have high precision but huge computational costs and require supercomputer support. 2. **Multi - modal data fusion**: - The trajectory of tropical cyclones is affected by various atmospheric physical quantities, such as pressure fields, wind fields, etc. How to efficiently fuse these multi - modal spatio - temporal data to improve prediction accuracy is an important challenge. 3. **Full utilization of time - series features**: - Time - series models (such as LSTM, BiGRU, etc.) that solely rely on the inherent features of TC usually have lower prediction accuracy due to the lack of consideration of factors affecting TC trajectories. 4. **Multi - horizon forecasting**: - Existing methods perform poorly in multi - step forecasting, especially when predicting the trajectories at multiple future time points, and the errors will gradually accumulate. To solve the above problems, the paper proposes a new dual - branched spatio - temporal fusion network (DBF - Net) for multi - step tropical cyclone trajectory prediction. DBF - Net solves these problems in the following ways: - **TC feature branch**: Use an LSTM - based encoder - decoder network to extract time features from the one - dimensional inherent features of TC. - **Pressure field branch**: Use a 2D - CNN - based encoder - decoder network to extract spatio - temporal features from the re - analyzed two - dimensional pressure field. - **Efficient feature fusion**: Through the fully - connected layer, fuse the high - level spatio - temporal features extracted by the pressure field branch into the LSTM decoder of the TC feature branch, thus making full use of the information of both types of data. Through this method, DBF - Net can achieve prediction results on the historical TC trajectory data in the Northwest Pacific that are significantly better than existing statistical and deep - learning methods.

Dual-Branched Spatio-temporal Fusion Network for Multi-horizon Tropical Cyclone Track Forecast

Tropical Cyclone Intensity Prediction by Inter- and Intra-Pattern Fusion Based on Multi-Source Data

A Neural Network Framework for Fine-Grained Tropical Cyclone Intensity Prediction

Enhanced Rainfall Nowcasting of Tropical Cyclone by an Interpretable Deep Learning Model and Its Application in Real-Time Flood Forecasting

Short-Term Intensity Prediction of Tropical Cyclones Based on Multi-Source Data Fusion with Adaptive Weight Learning

Enhanced Tropical Cyclone Inflow Flood Forecasts by Using Deep Learning and Spatial&#8209;temporal Information

A Spatio-Temporal Graph-Guided Convolutional LSTM for Tropical Cyclones Precipitation Nowcasting.

Spatiotemporal fusion convolutional neural network: tropical cyclone intensity estimation from multisource remote sensing images

TFSF: Topological and Feature Space Fusion with Spatio-Temporal Modeling for Crop Yield Prediction.

Tropical Cyclone Track Forecasting Using Fused Deep Learning From Aligned Reanalysis Data

A Novel Data-Driven Tropical Cyclone Track Prediction Model Based on CNN and GRU with Multi-Dimensional Feature Selection

Deep Learning Approaches in Predicting Tropical Cyclone Tracks: an Analysis Focused on the Northwest Pacific Region

A Novel Tropical Cyclone Track Forecast Model Based on Attention Mechanism

Predicting the Intensity of Tropical Cyclones over the Western North Pacific Using Dual-Branch Spatio-Temporal Attention Convolutional Network

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

A Data-Driven Approach for Regional-Scale Fine-Resolution Disaster Impact Prediction under Tropical Cyclones

TCP-Diffusion: A Multi-modal Diffusion Model for Global Tropical Cyclone Precipitation Forecasting with Change Awareness

A Multiscale and Multilayer Feature Extraction Network With Dual Attention for Tropical Cyclone Intensity Estimation

A multimodal deep learning approach for hurricane tack forecast based on encoder-decoder framework

TCIP-Net: Quantifying Radial Structure Evolution for Tropical Cyclone Intensity Prediction

A Novel Deep Learning Approach for Tropical Cyclone Track Prediction Based on Auto-Encoder and Gated Recurrent Unit Networks