Semi-Supervised Temporal Domain Adaptation for Artificial Intelligence Cyclone-Intensity Estimation Under Reanalysis Label Scarcity: Error Bounds and Evacuation-Trigger Reliability

Tropical-cyclone intensity estimates derived from reanalysis products are scarce, temporally heterogeneous and systematically biased relative to the best-track record, which limits the reliability of artificial-intelligence models used to trigger evacuation advisories. The framework combines a temporal consistency regulariser, a domain-adversarial feature extractor and a pseudo-label refinement step, and it is analysed through a Rademacher-complexity argument that yields a generalisation bound decomposing target risk into source risk, a temporal discrepancy term and a label-scarcity term. The bound is then propagated through a decision-theoretic evacuation-trigger model to obtain a reliability condition relating the intensity error to the probability of both missed and false alarms. The analysis is methodological and theoretical: no model was trained and no observational verification was performed. The contribution is a set of explicit error bounds and a trigger-reliability criterion that together specify when a semi-supervised intensity estimator is admissible for advisory use, and when label scarcity in the reanalysis record makes it inadmissible regardless of architecture.