Typhoon intensity forecasts stop reading cloud images alone: TCFNet steers visual encoding with pressure trends, cutting 24-hour pressure error to 9.52 hPa

Tropical cyclone intensity forecasting gains a route that does not rely on generic feature fusion: the visual representation of satellite cloud imagery is guided by the intensity variables themselves, rather than letting the model decide on its own which cloud features matter. Earlier multimodal methods concatenated or summed geostationary satellite images with best-track time series without explicitly modeling consistency between physical variables and visual features, while similar visual structures in cloud imagery can correspond to intensification, steady state and weakening. Across the Western North Pacific and South Pacific basins, forecasting 24 hours ahead from the past 12 hours of observations, TCFNet achieves the lowest error among the compared methods at most lead times, with 24-hour pressure RMSE of 9.52 hPa in the Western North Pacific and 7.00 hPa in the South Pacific.

The model treats tropical cyclone intensity as a continuously evolving physical process, so it can predict under irregular sampling and at arbitrary forecast times. Discrete recurrent models are constrained by fixed time steps, and existing methods require complete, regular historical observations, making them hard to adapt to missing data and flexible lead times. Training uses regular grid-aligned sequences, but timestamps enter through time encoding into time-aware gated recurrent units, and state evolution is integrated by neural ordinary differential equations with an adaptive-step solver; on the Western North Pacific the model still runs when 25%, 50% and 75% of input observations are randomly removed, with wind-speed performance dropping about 18.1% and 31.8% at 50% and 75% missing rates.

The contribution of cloud imagery to longer-lead forecasts can be quantified separately: removing the visual stream raises error, and more so as lead time grows. Relative to the full model, the variant without the visual stream raises pressure and wind-speed RMSE at the 24-hour lead by 21.71% and 13.07% respectively. This ablation is repeated over three random seeds on the Western North Pacific, where the full model's 24-hour pressure error is 9.49±0.04 hPa and the variant without the visual stream reaches 11.55±0.05 hPa.

Evaluation is confined to the Western North Pacific and South Pacific basins in DigitalTyphoon, with the past 12 hours of 3-hourly observations as input and 24-hour pressure and wind-speed forecasts as output; training, validation and test sets are split at the tropical cyclone event level. The authors name improved estimation or assimilation of physical variables as a direction for future work.

Model effectiveness depends on the availability and quality of physical variables; accurate measurements are hard to obtain under extreme weather, and in-situ observations over ocean regions are sparse, so these uncertainties may affect robustness. In the missing-observation experiments performance declines gradually as the missing rate rises, and the authors call missing inputs a challenging factor. The continuous and discrete versions reach comparable accuracy at the evaluated lead times, while the continuous version takes 0.23 s of inference per sample versus 0.0093 s for the discrete version.