Quantitative validation of AI-derived tsunami water levels using co-located tide-gauge observations during the 2025 Kamchatka Earthquake

Camera-based extraction of tsunami water-level time series has recently become feasible through deep learning segmentation, but its quantitative validity has not been established using strictly co-located tide-gauge observations during a real tsunami event. In this study, we present a quantitative comparison between automatically extracted camera-derived water levels and tide-gauge measurements obtained simultaneously at the same site during the tsunami generated by the 30 July 2025 Kamchatka Earthquake. A tsunami monitoring camera installed within the grounds of the Japan Meteorological Agency (JMA) Mera Tide Gauge Station captured the tsunami inside Mera Port, Japan. Water-level time series were extracted from the camera images using an automated method based on Segment Anything Model 2 (SAM 2) and directly compared with tide-gauge records from the same site. The two waveforms showed strong agreement, with a root mean square error of approximately 0.07 m and a correlation coefficient exceeding 0.98 after low-pass filtering. To the best of our knowledge, this study provides the first fully automated quantitative validation of camera-derived tsunami water-level time series against co-located tide-gauge observations during a real tsunami event. These results demonstrate that, under co-located conditions, camera-derived tsunami time series can reproduce tide-gauge observations with quantitative accuracy, providing a practical framework for AI-based tsunami monitoring.