Predicting Tropical Cyclone Genesis Location Using STAG-Net: A Spatio-Temporal Attention-Gated Network

Kalim Sattar¹, Malik Muhammad Saad Missen², Syeda Zoupash Zahra^1,3, Najia Saher⁴, Rab Nawaz Bashir^3,5,6,*, Oumaima Saidani⁷, Shahid Kamal⁵, Muhammad I. Khan⁶
1 Department of Information Technology, The Islamia University of Bahawalpur, Bahawalpur, Pakistan
2 Department of Software Engineering, The Islamia University of Bahawalpur, Bahawalpur, Pakistan
3 Department of Computer Science, COMSATS University Islamabad, Vehari Campus, Vehari, Pakistan
4 Department of Artificial Intelligence, The Islamia University of Bahawalpur, Bahawalpur, Pakistan
5 Center for Advanced Analytics, CoE for Artificial Intelligence, Faculty of Computing and Informatics, Multimedia University, Cyberjaya, Selangor, Malaysia
6 Artificial Intelligence and Data Analytics Laboratory (AIDA), College of Computer and Information Sciences (CCIS), Prince Sultan University, Riyadh, Saudi Arabia
7 Department of Information Systems, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University, P.O. Box 84428, Riyadh, 11671, Saudi Arabia
* Corresponding Author: Rab Nawaz Bashir. Email: email
(This article belongs to the Special Issue: Applied Machine Learning for FAIR and Responsible Modelling)

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2026.078569

Received 04 January 2026; Accepted 23 March 2026; Published online 06 May 2026

Download PDF

Abstract

Tropical Cyclone (TC) genesis forecasting is an important aspect of early warning systems, as it allows the adoption of early warnings and mitigation plans. However, existing methods often rely on binary classification or fail to capture the complex spatio-temporal dependencies that govern TC formation. To address this limitation, this study introduces STAG-Net, a novel Spatio-Temporal Attention-Gated Network designed to directly predict the geographical coordinates of TC genesis. The model uses multivariate variables of meteorological factors such as u-wind, v-wind, relative humidity, temperature, and large-scale dynamic features using a Convolutional Neural Network (CNN), Gated Recurrent Units (GRUs), and a channel-wise attention mechanism in identifying both spatial and temporal characteristics. The methodology takes the initial tropical disturbance data as an input and obtains spatial features in the ERA5 reanalysis dataset that covers 37 isobaric pressure levels. The study also investigates the effect of grid resolution on prediction performance, as four grid sizes were compared, namely 10×10, 20×20, 30×30, and 40×40. The experimental results demonstrate that STAG-Net significantly outperforms existing baselines such as the Dynamic Spatio-temporal model (DST), Spatial Attention Fusing Network (Saf-Net), and a temporal-only model. Notably, the model achieves an average MAE of 2.67∘, MSE of 13.24, RMSE of 3.45, and R2 of 0.87045, corresponding to performance improvements of 9.75%, 26.25%, 12.92%, and 4.27%, respectively, over the baseline model. The results also indicate that the 30×30 grid configuration was found to be the most effective. The results highlight the significance of the proposed approach for the TC genesis location prediction task.