Water surface elevation time series estimated from video segmentation for real-time tsunami monitoring using Segment Anything Model 2

The Segment Anything Model 2 (SAM 2), a general-purpose segmentation model based on deep learning that does not require prior training by individual users, was applied to analyze video images that captured the tsunami caused by the 2024 Noto Peninsula earthquake. Subsequently, a time series of water-surface elevations was estimated from the segmentation results obtained with SAM 2 for the analyzed case. A comparative analysis was conducted between the results obtained using SAM 2 and those from existing Canny edge detection methods. The values measured by humans were used as the reference standard. For the examined tsunami event, SAM 2 exhibited accuracy and detection performance comparable to or higher than those obtained using Canny edge detection with optimized parameters, while maintaining nearly identical processing speed. Furthermore, whereas edge detection necessitates a process of trial and error to optimize the parameters, SAM 2 does not require such adjustments. The results demonstrate that SAM 2 can robustly extract time series of water-surface elevations during the 2024 Noto Peninsula earthquake tsunami and may suggest its potential for application to real-time tsunami monitoring.