We develop a new technique to integrate realistic uncertainties in probabilistic models of past sea-level change. The new framework performs better than past methods (in precision, accuracy, bias, and model fit) because it enables the incorporation of previously unused data and exploits correlations in the data. This method has the potential to assess the validity of past estimates of extreme sea-level rise and highstands providing better context in which to place current sea-level change.

Cape Town is situated in a small but ecologically and climatically highly diverse and vulnerable pocket of South Africa uniquely receiving its rain mostly in winter. We show complex structures in the spatial patterns of rainfall seasonality and year-to-year changes in rainfall within this domain, tied to spatial differences in the rain-bearing winds. This allows us to develop a new spatial subdivision of the region to help future studies distinguish spatially distinct climate change responses.

The Day Zero water crisis affecting Cape Town after the severe 2015–2017 drought motivated renewed research interest into causes and projections of rainfall variability and change in this water-stressed region. Unusually few wet months and very wet days characterised the Day Zero Drought. Its extent expanded as it shifted gradually north-eastward, concurrent with changes in the weather systems driving drought. Our results emphasise the need to consider the interplay between drought drivers.

In this study, we introduce a deep learning algorithm to establish the relationship between wind and waves in order to predict the latter. The performance of the proposed method has been evaluated both on the output of numerical wave models and on in situ data and compared to other statistical methods developed by our research team. The results obtained confirm the interest of deep learning methods for forecasting ocean data.

A common problem in climate studies is to decide whether a climate model is realistic. Such decisions are not straightforward because the time series are serially correlated and multivariate. Part II derived a test for deciding wether a simulation is statistically distinguishable from observations. However, the test itself provides no information about the nature of those differences. This paper develops a systematic and optimal approach to diagnosing differences between stochastic processes.

We use machine learning to assess how different geophysical uncertainties relate to the severity of future sea-level rise. We show how the contributions to coastal hazard from different sea-level processes evolve over time and find that near-term sea-level hazards are driven by thermal expansion and the melting of glaciers and ice caps, while long-term hazards are driven by ice loss from the major ice sheets.

We developed a framework to produce global real-time estimates of how human-caused climate change affects the likelihood of daily weather events. A multi-method approach provides ensemble attribution estimates accompanied by confidence intervals, creating new opportunities for climate change communication. Methodological efficiency permits daily analysis using forecasts or observations. Applications with daily maximum temperature highlight the framework's capacity on daily and global scales.