Enhancing Tropical Cyclone Intensity Estimation via Multi-Scale Spatial Feature Correlation

Tropical cyclones pose significant threats to coastal regions, necessitating accurate intensity predictions for effective disaster management. Traditional methods, relying on manual feature extraction, suffer from subjectivity and scalability issues. This study introduces a novel model leveraging multi-scale and spatial feature correlation for precise tropical cyclone intensity estimation. By integrating deformable convolution, residual networks, and strip pooling, our model captures complex cloud patterns and long-range dependencies. Experimental results on the TCIR dataset demonstrate superior performance, achieving a Root Mean Square Error (RMSE) of 8.46kt, representing a 10.7% improvement over recent deep learning approaches and a 33.3% improvement over traditional methods. The source code is available at our GitHub repository, and the dataset can be accessed from the official TCIR website.