天堂视频

A great challenge in flood risk management to better model extreme events, particularly as flood frequency and severity are likely to increase in the UK as climate changes. A critical limitation for high-resolution flood modelling is inadequate information about the shape and roughness of the bed of our river channels.

This PhD project is ambitious and exploratory, yet based on a simple breakthrough idea: Water surface motion visibly reflects the key factors shaping flow in a river, almost by definition, so information for enhanced flood prediction should be observable from short videos of rivers (e.g. on a mobile phone). Its vision is to create a handheld, easily and widely deployable ‘tool’ for improving real-time flood risk assessment. It will have one critical conceptual advantage over other approaches (e.g. green lidar, structure from motion); it does not require light or sophisticated, expensive, equipment to penetrate to the river’s bed. It potentially reveals how riverbed shape changes during high flow events (i.e. when the water is entirely opaque), and Storm Desmond in 2015 demonstrated how important riverbed shape changes during high flows (i.e. River Greta) are in influencing flooding.

Methodology:

The PhD will use short video-clips of the water surface of a river to get a surface velocity field (established techniques) and spatially classify flow regimes (e.g. pool, riffle). Then, these data will be inverted via flow modelling & machine learning (physics-informed neural network - PINN) to get a detailed & spatially varying bathymetry and/or riverbed roughness. As such the methodology has four elements
1. Flume tank: In controlled conditions, data will be collected relating surface observables from videos, with underlying riverbed characteristics designed and known at 1-10 m scale.
2. Hydraulic modelling: A large number of simulations will be run, for known riverbeds, yielding predictions of surface observables.
3. PINN construction: Will be designed, using adversarial training and semi-supervised learning strategies to blend these data types, enabling prediction from this (what is for machine learning) sparse data.
4. Fieldwork: Building on pilot data collected of 10 days in 2024 between on the river Greta near Keswick in the Lake District, videos and field surveys will serve to make this work applicable to the real rivers (10-100m scale). Pragmatically, a variety of video capture technology will be investigated (i.e. mobile phone, fixed GoPro camera, aerial drone).

References:
Environment Agency (2021) Understanding river channel sensitivity to geomorphological changes, Report number - FRS17183/R1, ISBN: 978-1-84911-479-0.

Muste et al. (2009) Large-scale particle image velocimetry for measurements in riverine environments. Water Resources Research, 44, W00D19. doi:10.1029/2008WR006950.

Xia et al. (2019) A full-scale fluvial flood modelling framework based on a high-performance integrated hydrodynamic modelling system (HiPIMS). Advances in Water Resources 132, 103392.