DIAF-Net: difference-driven multispectral satellite image sequence fusion for tropical cyclone intensity forecasting

Tropical cyclones (TCs) are among the most destructive natural hazards in coastal regions, making timely intensity monitoring and forecasting important for disaster risk reduction. Satellite remote sensing provides continuous observations of TC cloud-top temperature, convective organization, and upper-level moisture distribution. However, extending satellite-based TC intensity prediction beyond the commonly studied 24 h range remains challenging, because single-channel image textures alone cannot fully characterize both inner-core structural evolution and moisture-related environmental signals. To address this issue, this study proposes DIAF-Net, a difference-driven multispectral satellite image sequence fusion network for TC intensity forecasting at 24-, 36-, and 48-h lead times. The model uses infrared and water vapour satellite imagery as the main inputs. A difference-driven adaptive fusion (DDAF) module is designed to combine the two channels and enhance their complementary information. It preserves cloud-top structure from infrared imagery while incorporating water vapour information related to upper-level moisture and circulation. Radial profile and deviation angle variance features are introduced to describe radial organization and cloud symmetry. Historical intensity and latitude–longitude encoding are further used to support multi-lead-time prediction. A non-negative residual compensation module is applied to reduce the underestimation of severe and super typhoons. Experiments show that DIAF-Net achieves mean absolute errors of 8.91, 12.51, and 16.84 kt for 24, 36, and 48 h forecasts of TC maximum sustained wind speed, respectively. These results demonstrate the potential of difference-driven multispectral satellite image sequence fusion for extracting TC structural information and supporting short- to medium-range intensity prediction.