Same Storm, Two Satellites: Measuring Acquisition Sensitivity in Learned Tropical Cyclone Intensity Estimates

Historical tropical cyclone intensity records span several generations of geostationary satellite, so a learned estimator applied to them crosses acquisition regimes whose effect on its predictions is hard to measure. Overlapping coverage offers a direct route: a storm is routinely imaged by two spacecraft within one nominal time slot, giving two factual observations that share a best-track label and a closely matched atmospheric state while differing in instrument and viewing geometry. On such pairs from HURSAT-B1 a standard estimator disagrees with itself by roughly 10 kt, comparable in scale to the uncertainty discussed for operational intensity estimates and invisible to the homogenised operational record, which retains one estimate per storm-time. Using these pairs to audit invariance criteria, we find that common proxies need not track the measured target: a domain-adversarial objective does not reduce the gap and increases it on the near-coincident population, and a 23% reduction in the gap leaves the domain probe no weaker. A criterion defined on the paired coupling achieves that reduction, whereas matching or marginal criteria given the identical pairs recover about a fifth of it. We release the pair index.