Apply for this PhD here https://applyto.newcastle.ac.uk/ using application studentship code FLOOD246. Please contact Caspar Hewett (caspar.hewett@newcastle.ac.uk) if you have any questions about the application process.
The IPCC’s 6th Assessment Report shows that complex interactions between multiple flood hazards, exposures, and (human) vulnerabilities are poorly understood. Inequity, conflict, poverty, weak governance, access to insurance, limited access to basic services, and other factors increases vulnerability to flood hazards in different ways, limiting different people’s ability to adapt to current and future flood risk.
Flood risk models are increasingly adept at modelling how exposure and hazard change over time. There is an urgent need to understand how vulnerability of population and the built environment could evolve. This is crucial to determine not just how risk will change but also the most appropriate adaptation action to manage flood risk in a changing world. Currently, exposure models are usually based on quite simplistic spatial growth/weighting approaches to predice future population and building development, and do not measure vulnerability (e.g. Wang et al., 2022). Moreover, existing models of vulnerability are typically static snapshots or operate at very coarse granularity (Bang and Burton, 2021).
This project will use a generative AI approach to develop, validate and demonstrate a high resolution, scalable, model of flood vulnerability.
Generative AI (best known for ChatGPT) has shown huge potential to produce realistic images and patterns for a number of applications. Generative Adversarial Network (GAN) methods can reproduce urban morphology through image training (Zhang et al., 2022). This project will extend this with (i) a conditional variational auto-encoder (CVAE) to create a synthetic population model with flood vulnerability characteristics, (ii) the CVAE population model into the GAN, and (iii) a policy model (e.g. restricted floodplain development). This will improve performance and enable scenario testing by constraining the GAN to account for aspects of vulnerability and development policies and output a spatial model with population vulnerability attributes.
Data will be split into three: training, validation, and application. The vulnerability model will be validated using standard spatial statistics approaches in urban modelling and against socio-economic statistics to ensure the correct diversity of the population vulnerability characteristics.
The vulnerability model will be embedded into the open-access OpenCLIM framework, maintained on DAFNI by STFC. This demonstration will test the model’s ability to identify hotspots of flood risk, not captured by exposure models alone, in Great Britain, helping decision-makers target adaptation investment.
Wang X, Meng X, Long Y. (2022) Projecting 1 km-grid population distributions from 2020 to 2100 globally under shared socioeconomic pathways. Scientific Data, 9(1):563.
Bang HN, Burton NC. (2021) Contemporary flood risk perceptions in England: Implications for flood risk management foresight. Climate Risk Management. 1;32:100317.
Zhang, W., Ma, Y., Zhu, D., Dong, L. and Liu, Y. (2022) Metrogan: Simulating urban morphology with generative adversarial network. In Proc. 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 2482-2492).
