CyprusGPT GeoAI: Sentinel-2 Coastal Indicators, Short-Horizon Prediction and the Evidential Limits of Conversational Decision Support in Cyprus

Coastal geospatial artificial intelligence (GeoAI) can make environmental information accessible, but predictive skill alone does not validate beach-specific or safety-related advice. This study evaluates CyprusGPT at three bathing sites in Cyprus using Sentinel-2 imagery, ERA5 reanalysis and an audit of the rules that turn environmental inputs into user-facing ratings. A normalized difference turbidity index (NDTI) from 465 Sentinel-2 overpasses (2021–2026) separates the sites, which share one ERA5 wave cell, but is unrelated to the application's wave-derived clarity score (Spearman ρ from −0.04 to 0.03). Random forest (RF) and XGBoost forecasts of daily wave height and wind are compared with persistence and seasonal climatology. Next-day RF wave forecasts reach a coefficient of determination of 0.537 and reduce mean squared error by 35.8% relative to persistence (95% interval 26.4–44.4%); an independently specified reconstruction reaches 0.465. No three-day model improves on climatology, and high waves are underpredicted. The code audit also shows that missing inputs can yield favourable star ratings. No calibrated water-quality retrieval, field validation or conversational evaluation is claimed. The findings support limited short-horizon predictive value and define the spatial, temporal and evidential controls needed before predictions become coastal decision support.