Digital Typhoon Dataset V3: Benchmarking for Multi-spectral Extension and Analyzing Systematic Bias in Intensity Regression
The Digital Typhoon Dataset V3, a 50-year satellite image collection, was used to test machine learning models for typhoon tracking and intensity prediction. Researchers found that adding more spectral channels did not significantly improve performance for pressure regression. They also identified a systematic bias where models underestimate the intensity of very strong typhoons, suggesting new weighting strategies are needed to better capture extreme events.