A region-based convolutional fusion network for typhoon intensity estimation in satellite images.

Accurate typhoon intensity estimation is vital for timely risk warning and hazard assessment. The task is challenging since the subcategories of the typhoon have visually similar appearances. The local context extracted from the typhoon center region and the global context of cloud pattern is the most discriminative features for intensity estimation. However, the existing method does not utilize local and global contexts together. Furthermore, the interactions between local and global contexts remain unexplored. To make a better prediction, this paper proposes a local and global context attention mechanism-based fusion network that consists of a spatial attention fusion block and a channel attention fusion block. The channel attention fusion block infers an optimal channel fusion weight by exploiting the inter-channel relationship. The spatial fusion block learns the best inter-spatial relationship of features. Extensive experiments on DeepTi and TCIR benchmarks demonstrate the effectiveness of our method, and the proposed network achieves new state-of-the-art. The code is at https://github.com/chen-huanxin/RBCFN. • Local and global information fusion improve estimation accuracy. • A multi-level framework aggregates information from multiple areas. • Spatial and channel attention fusion improve feature representation ability. • Attention mechanism effectively fuses multi-modal information. [ABSTRACT FROM AUTHOR]