Modelling tropical storms using survival analysis and AI: understand, describe, predict
Background
A characteristic feature of survival data is that some observations may be incomplete (censored), e.g. due to loss-to-follow-up or end of study. Statistical methods to analyse survival data are of critical importance in medical research, actuarial work, and many more, including environmental studies. For example, the lifetime of a convective storm may be defined as time from onset to dissipation, or until it reaches a certain location. In such applications, survival analysis is particularly efficient in assessing the impact of atmospheric and geographical covariates on the survival probability. This is important for assessing environmental risks at various timescales, from forecasting and early warnings to longer-term climate change prediction.
PhD opportunity
Convective storms play a fundamental role in the Earth climate system, influencing tropical rainfall and circulation patterns. Their lifecycles and attributes are determined by local conditions. The thrust of this project is to explore the utility of survival analysis in modelling environmental data such as the statistics of thunderstorms and hurricanes. The overarching aim is to improve understanding of lifecycles and predictability of the tracks of extreme meteorological events on account of precursor atmospheric and land-surface conditions. Project outcomes could assist in understanding future trends in the behaviour of storms and thus aid the evaluation of long-term climate models and projections. Established statistical models such as Cox proportional hazards, together with modern AI tools including Large Language Models (LLM), will underpin visual and textual summaries from data analysis to assist real-time consistent predictions at different timescales. In addition to climate science, the outcomes of this research can be used in the insurance industry and prospective management and planning.
A related objective is to apply suitable deep learning tools to translate large scale climate variables into local meteorological factors, such as rainfall, winds, temperature, surface pressure, humidity, etc. Tropical cyclones and other storms may serve as an excellent test bed for such research, leveraging vast data (both weather and climate) sourced from ground-based observations, satellite images and measurements, and also generated by physical computational models. Currently used tools include U-NET (convolutional neural network for image segmentation) and Generative Diffusion Model, but other deep learning architectures could be tested for the purpose of downscaling, such as Long Short-Term Memory (LSTM) or Reinforcement Learning. Importantly, the project will address the ability of such models to reproduce consistency between predicted characteristics.
This research will leverage the developing partnership with the Lloyds Banking Group via General Insurance Weather & Catastrophe Modelling Team. The student will be encouraged to play a role in joint quarterly science meetings, thus expanding potential applications.
Applicant Profile
A successful candidate will demonstrate the ability and interest in analysing complex real data. General experience of using R and/or Python (numpy, pandas useful) in data analysis will be expected. The required routines from the library ‘survival’ can be learned along the way. As an alternative, use of Python for survival analysis is also appropriate (via ‘scikit-survival’). Python will be required for loading and manipulating storm data, but experience with the appropriate packages (xarray) is not expected. Previous familiarity with survival analysis would be helpful but not compulsory. Previous meteorological/atmospheric science experience not necessary.
Other information
· Collett D. (2015). Modelling Survival Data in Medical Research, 3rd ed. Chapman & Hall.
· Hougaard, P. (1999). Fundamentals of survival data. Biometrics, 55, 13–22. (https://doi.org/10.1111/j.0006-341X.1999.00013.x)
· Klein C, Jackson LS, Parker DJ, Marsham JH, Taylor CM, Rowell DP, Guichard F, Vischel T, Famien AM, Diedhiou A. (2021). Combining CMIP data with a regional convection-permitting model and observations to project extreme rainfall under climate change. Environmental Research Letters, 16, 104023. (https://iopscience.iop.org/article/10.1088/1748-9326/ac26f1)
· Taylor C M, Belušić D, Guichard F, Parker DJ, Vischel T, Bock O, Harris PP, Janicot S, Klein C & Panthou G. (2017). Frequency of extreme Sahelian storms tripled since 1982 in satellite observations. Nature, 544, 475–478. (https://www.nature.com/articles/nature22069)
· Kendon EJ, Stratton RA, Tucker S, Marsham JH, Berthou S, Rowell DP & Senior CA. (2019). Enhanced future changes in wet and dry extremes over Africa at convection-permitting scale. Nature Communications, 10, 1794. (https://www.nature.com/articles/s41467-019-09776-9)
· Taylor CM, Klein C, Dione C, Parker DJ, Marsham J, Cheikh AD, Fletcher J, Chaibour AAS, Nafissa DB, Semeena VS. (2022). Nowcasting tracks of severe convective storms in West Africa from observations of land surface state. Environmental Research Letters, 17, 034016. (https://doi.org/10.1088/1748-9326/ac536d)
· Wilhelm J, Wapler K, Blahak U, Potthast R & Kunz M. (2023). Statistical relevance of meteorological ambient conditions and cell attributes for nowcasting the life cycle of convective storms. Quarterly Journal of the Royal Meteorological Society, 149, 2252–2280. (https://doi.org/10.1002/qj.4505)
