Enhancing forecasting flood inundation mapping through data assimilation

Timely flood inundation forecasts allow pro-active flood management, reducing loss of life and damage to infrastructure. This project aims to investigate new methods using observations of floods to verify and improve flood inundation forecasts.

The state-of-the art in operational flood inundation forecasting at national/trans-national scales uses a simulation library, where rainfall data from numerical weather prediction (NWP) or observations drives a chain of process models and probabilistic analysis, and the final flood forecasts are produced from a pre-computed library of flood maps. This approach saves computation time, allowing near-real-time updating for large areas, which otherwise presents a significant challenge. However, it poses challenges for data assimilation, where observations are combined with model forecasts in a dynamical feedback loop, to keep forecasts on track. Nevertheless, the data assimilation framework still provides consistent techniques for comparison of heterogeneous observations with model data, forecast verification and calibration.

The PhD project will address the following questions:

1. What are the lengthscales of variability in flood inundation forecasts made using a simulation library approach? New generic methods are needed to understand the scales of ensemble variability, which may be determined by the driving NWP and hydrological models rather than the resolutions of the pre-computed flood hazard maps. This will be valuable for the qualitative interpretation of forecasts for users, as well as setting quantitative scale parameters for appropriate comparisons of observations with models. The approach will use the ideas of the meteorological Fractions Skill Score in a new context.

2. Can we use low resolution satellite data to give useful urban flood observations? Satellite-based Synthetic Aperture Radar (SAR) instruments measure backscatter, which in rural areas can be converted to flood extent using image processing. Recent work by the supervisors has identified flood extent in urban areas using expensive high resolution (2-3m) SAR data together with high resolution (2m) lidar digital surface models (DSMs) and modelled flood return period maps. Can these techniques be usefully extended to use lower resolution Sentinel 1 open data or for locations where high resolution lidar is not available?

3. Which observation operators are most appropriate? Different observation-model comparison approaches may provide better or worse discrimination between forecasts, depending on the lengthscales of forecast variability and the features that can be resolved by the observations. This lends itself to pixel-by-pixel flooded/not flooded binary comparisons or related probabilistic approaches, flooded area estimations and water level estimation by intersection of flood extents with a digital terrain model.

4. Is data assimilation cycling beneficial for a simulation library system? Implementing an ensemble Kalman filter or resampling particle filter may provide forecasts that are closer to the observed reality than the standard simulation library approach. The student will build a system using the observation operators developed in Year 2 and existing data assimilation software libraries (such as PDAF), and evaluate the impact on forecasts.

The student is expected to develop generic methods and explore these questions through case studies for different locations (e.g., Myanmar, India, Bangladesh, Ireland, UK) using both ensemble and single deterministic forecast approaches, with data from satellites, gauges and river / road cameras. During the project, the student will have opportunities to gain further experience through work placements with JBA.