The Digital Typhoon Dataset V3, described in a new arXiv preprint, extends the longest collection of typhoon satellite images to nearly five decades. The authors position it as a benchmark for machine learning models working on long-term spatio-temporal data, a setting that tests both memory and generalization across changing conditions.

The new version adds a multi-spectral extension, broadening the input channels available for model training and evaluation. Beyond the dataset itself, the paper reports an analysis of systematic bias in intensity regression, suggesting the resource is meant not only for benchmarking but also for diagnosing recurring errors in prediction tasks.

Because this is based on a single abstract, the details are limited. The paper does not yet describe the full methodology or results, but the stated goals point to a resource aimed at improving both model robustness and the understanding of bias in typhoon intensity estimation.