Multi-scale deformable detection transformer with multi-modal data fusion for cyclone detection and intensity estimation at global scale

Cyclone remote sensing provides critical data for climate research. However, effectively extracting accurate information from large datasets remains a challenge. Given that the cyclones are characterized as small objects at global scale, this study developed a multi-scale deformable detection transformer (MS-DDETR) to detect the cyclones and estimated their intensities. The wind field obtained from Advanced Scatterometer (ASCAT) and the precipitation data obtained from Global Precipitation Mission (GPM) were integrated as a multi-modal input dataset for cyclone detection and identification. The results demonstrate that the proposed MS-DDETR model achieves comprehensive improvements in cyclone detection by combining multi-scale feature extraction with a deformable self-attention mechanism. This mechanism enables the model to focus on key points and extract multi-scale information from them. As a result, errors in intensity estimation are also reduced because the model makes more physically consistent judgements based on richer multi-scale contextual information. The large-scale detection and identification method developed in this study is expected to advance the research and applications on global cyclone monitoring through multi-modal data fusion.