Our Research
Our research focuses on spatio-temporal modelling and environmental applications. Below we list some of our current research projects.
AI-enhanced spatio-temporal modelling and forecasting
Complex environmental systems—such as weather patterns, ocean dynamics, and atmospheric phenomena—are inherently high-dimensional and nonlinear. We use deep learning to model these systems in ways that would be intractable with classical methods, enabling both more accurate representations of complex dynamics and dramatically faster computations. Crucially, our approaches are probabilistic: we produce full predictive distributions rather than point forecasts, rigorously quantifying uncertainty in our predictions. This uncertainty quantification is essential for risk assessment and decision-making under environmental uncertainty.
Flexible modelling for climate extremes
Risk assessment of natural hazards—such as heatwaves, floods, and wildfires—requires statistical analysis of extreme events, often beyond observed levels. We develop flexible models for multivariate and spatial extremes with rich tail dependence in both the upper and lower tails, enabling realistic characterisation of joint and compound risk when standard dependence assumptions fail. Alongside model development, we design efficient, scalable inference methods that deliver principled uncertainty quantification for extrapolation and decision support.
→ Quantifying joint heatwave risk via flexible tail dependence
→ Graph-based discovery of river discharge risk network
→ Joint spatial prediction of extreme wildfire frequency and spread
Neural networks for accelerated inference
Estimating parameters in a statistical model is often the computational bottleneck of a spatio-temporal data analysis. Further, often computations often need to be done several times (e.g., millions of times) daily. For this reason, we have been investigating deep learning methods for fast approximate inference in large-scale spatio-temporal models, to enable real-time environmental monitoring and forecasting in classically intractable settings. These tools are lightning fast, do not require knowledge of the likelihood function, and can yield accurate inferences.
→ Neural networks for Bayes estimation
Modern spectral analysis
The three main families of nonparametric estimators for the power spectral density—lag-window, multitaper, and Welch—are statistically consistent, yet each can suffer from nonignorable bias when estimating from finite samples. Welch’s method, which averages periodograms over overlapping segments, faces a bias–variance trade-off: more segments reduce variance but increase bias. We have developed novel debiasing methods that maintain computational complexity and asymptotic consistency while substantially improving finite-sample performance. This theory extends to the broader family of quadratic estimators, enabling bias correction for lag-window and multitaper estimators and their combinations, and works in conjunction with any tapers and lag-sequences designed for bias reduction.
Survival analysis for time-to-event data
Many environmental processes are naturally described by the time until a critical event occurs, including mortality, habitat loss, ecological collapse, pathogen exposure, disease progression or system failure. Such data are inherently incomplete because events may be right-censored, interval-censored, or subject to left truncation, and observation windows often vary across individuals, sites or regions. We develop semi-parametric survival models for time-to-event data that explicitly account for censoring and truncation while retaining the flexibility of hazard-based regression without imposing restrictive parametric assumptions. Our work advances survival analysis by combining flexible hazard modelling, penalized likelihood estimation and time-varying or nonlinear effects, enabling interpretable and reliable risk assessment in complex environmental systems.
Statistical methodology for Antarctic science
We collaborate with the program Securing Antarctica’s Environmental Future (SAEF) to develop statistical methods for Antarctic research, including techniques for predicting sea-level rise contribution, calibrating ice sheet models, and modelling biodiversity in polar ecosystems. This work requires fusion from multiple data sources—such as satellites, numerical models, and in-situ observations—as well as statistical downscaling and high-performance computing to enable inference at scale.
