Challenges in Modelling and Forecasting Rainfall
Stochastic models of precipitation are generally constructed to help planning and decision making in agriculture, hydrology, ecology and various other earth sciences. Here they are useful in modelling processes which depend on rainfall, such as flooding, runoff and crop growth. In addition, risk assessment and planning in these areas often requires specialised models which focus on the extremal behaviour of rainfall.
There are a variety of ways of modelling rainfall, ranging from empirical statistical models, which seek only to provide a description of aggregated, e.g. daily, totals to more physically motivated models which attempt to describe the underlying rainfall generating mechanism. Each presents its own challenges. Modelling extremes of rainfall presents different problems, compounded by the sparsity of data on extreme events.
This joint meeting will address recent progress made on addressing the challenges discussed above. The format of the afternoon will include three talks, as well as a coffee break for networking and sharing ideas.
Malcolm Farrow, School of Mathematics & Statistics, Newcastle University
TITLE: Joint modelling of occurrence and amount: mixed zero-positive distributions with applications to rainfall data
ABSTRACT: Rainfall measured over short time intervals, such as days, has a mixed distribution with a nonzero probability of zero and the remaining probability spread continuously over positive values. In this talk, approaches to modelling such distributions are reviewed. In particular, models for dependence on covariates and dependence among collections of such mixed variables, for example in spatio-temporal rainfall models, are considered. Bayesian inference for such models is discussed.
Jo Kaczmarska, Risk Management Solutions Inc.
TITLE: Local estimation for point process-based rainfall models: Allowing for a nonstationary climate
ABSTRACT: Simulations of rainfall at sub-daily time scales are required as input into various models, for example the rainfall-runoff models used for hydrological design. Point process-based rainfall models, fitted to rain-gauge data, have been used for generating such simulations for many decades. An appealing property of these models is their mechanistic structure, which reflects important aspects of the physical process: the clustering behaviour exhibited by rainfall and the fact that rainfall totals over short intervals are often zero. However, the models suffer from a major limitation: the only non-stationary feature that they can incorporate is seasonality, which is achieved by fitting separate models for each calendar month or season. We address this limitation by extending the existing fitting method. The new approach allows the discrete covariate, calendar month, to be replaced or supplemented with continuous covariates that are more directly related to the incidence and nature of rainfall, thereby allowing the models to be used for climate impact studies. The covariate-dependent model parameters are estimated for each time interval using a kernel-based nonparametric approach within a Generalised Method of Moments framework ("local GMM"). An empirical study demonstrates the new methodology at a single site, using a time series of five-minute rainfall data.
Francesco Serinaldi, School of Civil Engineering & Geosciences, Newcastle University
TITLE: Stochastic rainfall synthesis: a matter of space, time, tails ... and statistical oversights
ABSTRACT: Rainfall is the main source of surface and underground water resources, but also the driver of natural hazards such as floods and landslides; it has short and long term effects, impacting small and/or large areas. Based on their spatio-temporal scales of evolution, these effects can be associated to the average fluctuations of the rainfall process and/or to extreme realizations. Since rainfall is a complex phenomenon driven by multiple physical mechanisms acting at multiple spatio-temporal scales, a stochastic representation has been widely used in the last decades, leading to a large number of different models devised to reproduce different aspects of the precipitation process. This talk provides an overview of such approaches focusing on their rationale, properties, and shortcomings. Since sometimes the haste of modeling leads to forget what we are modeling, I discuss the importance of using reliable and unbiased diagnostics and long rainfall records to distinguish actual properties and possible statistical artifacts.
For more information and to register please visit www.ncl.ac.uk/maths/rain
Organiser Name Sarah Heaps
Organising Group(s) RSS Environmental Statistics Section and RSS North East Local Group