Typhoon Intensity Prediction Based on an Improved Swin Transformer

Typhoons pose a huge threat to human life and property. The typhoon intensity prediction is very important for disaster prevention. A novel hybrid deep learning model for typhoon intensity prediction is proposed, which combines multiple data preprocessing techniques, transfer learning, and Swin Transformer model. Firstly, the multiple data preprocessing techniques such as image preprocessing, image enhancement and data enhancement are utilized to reduce typhoon image noise, enhance image contrast and address data category imbalance. Secondly, based on Swin Transformer, an improved typhoon remote sensing image classification is proposed. In the improved model, a CNN layer is added before the patch block of the original Swin Transformer. This design can better utilize the hierarchical nature of CNN networks to extract local and global features of images. We also utilize the transfer learning pretrained using the ImageNet-1K dataset to accelerate the model training time and further improve the prediction performance. Finally, real typhoon data are used for validating the model performance. The comprehensive experiment results show that the proposed model outperforms most baseline models in terms of performance.