Data-driven models for the study of extremes.
The topic deals with a data-driven study of extreme events and is situated in the field of extreme value statistics while exploring links with computational and machine learning methods. Assessing the probability of extreme events is of great importance in various life science applications given their potential for catastrophic impacts, e.g. in nature tsunamis, floods or heat waves can cause significant economic and human losses. While statistical models for univariate extremes are well-understood, classical models for multivariate extremes often lead to complex analyses that don't scale well when the number of variables increases and are hard to interpret or visualise. In this PhD track, we aim to build further on recent work in the field that allows more flexibility towards modelling the dependency structure in the tail of a distribution. A starting point of the research is the study of sparse dependency structures that can be based on e.g. latent variable approaches or graphical models. Applications are being explored in the broad context of climate change issues among others.
Doctorandus: Bastiaan Aelbrecht
Optimum survey design for the acquisition of spatial(-temporal) data.
In various life science applications, the acquisition of spatial data can be labor intensive or costly such that in practice one is
interested in an efficient design for data collection, e.g. for the acquisition of species observations to model species distribution
or for the acquisition of weather observations for the study of climate extremes. Although a rich literature is available on spatial survey designs,
less is known on how to develop optimum designs for integrating field surveys with modern acquisition techniques such as citizen science where data
is collected by volunteers with different levels of expertise and in an unstandardized manner or remote sensing techniques where a parameter of
interest is inferred from high-dimensional data. In this PhD project, we will test and adapt principles from the theory of optimum design to
integrate optimum survey design of in-situ data with other types of sensed data. Computational methods (e.g. resampling methods, Monte Carlo theory
among others) will play an important role in this research next to Bayesian statistical modelling with e.g. point processes, kriging and
Doctorandus: Thierry Rondagh
Citizen Science Data for Biodiversity Conservation Policy in Agricultural Landscapes
The aim of the project is to investigate how opportunistic citizen science data can support biodiversity conservation
policy in agriculture in an era of growing environmental concerns and increasing land use intensity. Species distribution models will be used to reduce the simultaneous effects of several biases that occur in opportunistic citizen science data. A combined approach will be proposed to minimize bias and we will use landscape metrics obtained through remote sensing to improve model predictions of species occurrence. Our findings will provide insight in the drivers of biodiversity in an agricultural landscape and will allow to simulate the future state of biodiversity by implementing different land use scenarios.
To illustrate the application potential of our results, we address the most prominent question in today’s ‘green’ agricultural policy: do the current policy efforts, meant to safeguard and improve biodiversity, pay off and how can their effectiveness be increased? With this project, we aim to take a step forward in the use of citizen science data for policy purposes. We wish to support the uptake of environmentally beneficial practices in farms, while aspiring a more sustainable land use future.
Doctorandus: Camille Van Eupen
Open set recognition using extreme value statistics applied on (big) biological data
Classification and anomaly detection tasks are frequently encountered in today's data driven era. A large part of classification methods are based on an optimization process (a so-called training phase) that learns the boundaries between classes from available examples. These algorithms assume that all classes are known and available during training. Open set recognition is the process of detecting new classes that were not seen by the model during training. In this project solutions for the open set recognition problem will be developped that are based on the use of extreme value statistics, a field in statistics that relies on extrapolation from the data of observed classes. This work builds further on our previous obtained
results in novelty detection [10,12] . A potential application of these methods is the detection of invasive species using image data. Automatic recognition of species based on image data could ease and support the labour intensive job of experts to go through all images.
Doctorandus: Matthys Lucas Steyn