A deep learning-based algorithm for spatiotemporal monitoring of debris in dam-break waves

Entrained debris in extreme flow conditions, such as tsunami and flash floods, pose severe risks to human safety, structural stability, and critical infrastructure. Numerous objects can be mobilized, entrained, accelerated, and dispersed by the momentum flux of such events, leading to complex patterns of debris transport, impact, and obstruction. Automated detection and tracking of multiple debris objects remain challenging, as conventional methods, such as color-thresholding, point cloud analysis, or Lagrangian sensor-based methods, often become unreliable or impractical when debris counts exceed more than one or two dozen, or flow velocities surpass 1.0 m s -1 . This study presents a novel hybrid deep-learning algorithm for debris identification and tracking, DeepRIS (Deep-learning Regression-based Identification and tracking of debriS), integrating the YOLO (You Only Look Once) object detector with the DeepSORT association framework. DeepRIS was validated conducting and using 91 medium-scale (1:20) dam-break experiments. Up to N = 240 debris objects in categories of varying density, initial debris orientations, and hydrodynamic conditions were investigated. The detection module achieved a mean confidence score of 88 %, while objects retained a 65 % probability of being tracked successfully. A novel performance score Q and a phenomenological cluster rate C are introduced to allow qualification of optical monitoring methods in coastal engineering. Overall, DeepRIS demonstrated robustness across diverse test configurations, exceeding the state-of-the-art maximum number of tracked debris by a factor of ten and reducing analysis duration by factors of 14 to 270, while maintaining trajectory update rates up to 71.4 Hz and operating under flow velocities up to 3.13 m s -1 . These findings highlight the method’s potential for full-scale debris monitoring to support forensic analysis, real-time hazard assessment, and predictive modeling, thereby enhancing the resilience of infrastructure exposed to extreme flow conditions.