Mikhail Kanevski : On Spatial Data Science; Wednesday 9th June at 10am
Daniela Castro Camillo : Practical strategies for fitting Extreme-Value statistical models with a view towards environmental and ecological applications; Wednesday 9th June at 2pm
Marc Barthelemy : Understanding the spatial structure of cities: some results and challenges; Thursday 10th June at 10am
Devis Tuia : Interactive deep learning for animal conservation from above; Friday 11th June at 10am
Faculty of Geosciences and Environment, University of Lausanne (Switzerland)
On Spatial Data Science
Abstract: Data science is an emerging scientific discipline concerned with the development and application of theoretical and computational methods to work with and extract knowledge from Data. (Geo)statistics, geoinformatics and machine learning are basic and complementary methodological approaches contributing to the spatial data science (SDS). In the current presentation the main attention is paid to the analysis, modelling, prediction and visualisation of complex spatial (spatio-temporal) environmental data. A problem-oriented approach, which starts with the objectives of the study and quality and quantity of data, is adapted. It follows a generic data driven methodology: from data collection via intelligent exploratory data analysis and modelling with careful validation and testing to the interpretability/explainability of the results. SDS is considered as an experimental science, therefore experimentation with data by applying different methods, algorithms and tools is considered as very important. Such point of view helps in better understanding of data and phenomena, obtaining reliable and robust results and making intelligent decisions. The presentation is accompanied by simulated and real data case studies. In conclusion some general remarks and future perspectives are discussed.
BIO: Mikhail Kanevski is Professor on Environmental data mining & Geostatistics. His current scientific interests cover a wide range of topics: geographical information science, environmental modelling, spatial statistics, time series forecasting, machine learning and environmental data mining. The major applications deal with natural hazards, environmental pollution and renewable energy analyses and assessments. He is. co-author of several books on spatial data modelling, introducing geostatistics and machine learning algorithms in the environmental sciences, namely: “Analysis and modelling of spatial environmental data” (2004), along with the Geostat Office software; “Advanced Mapping of Spatial Data. Geostatistics, Machine learning, Maximum Bayesian Entropy” (2008); “Machine Learning for Spatial Environmental Data. Theory, Applications and Software” (2009). The models developed and adapted by the group of Prof. Kanevski were successfully applied to geo-, environmental and socio-economic spatio-temporal data analyses. The fundamental scientific research was supported by several SNSF grants. At present Prof. Kanevski is a co-PI (collaboration with EPFL) of the PNR75 “Big Data” project “Hybrid renewable energy potential for the built environment using big data: forecasting and uncertainty estimation”.
Lecturer in Statistics, University of Glasgow, UK
Practical strategies for fitting Extreme-Value statistical models with a view towards environmental and ecological applications
Abstract: Over the last years, Extreme-Value Statistics (EVS) has gained considerable attention in environmental science, as extreme observations have increased in size and frequency. Mathematically, EVS has a well-developed asymptotic framework that allows us to study extreme events of single or multiple processes observed in one or many locations over space and time. Moreover, it enables us to make statements regarding future events that can be even more extremes than those observed. The mathematical elegance of these methods faces a couple of challenges in the applied arena. For instance, most asymptotically justified EVS models are computationally expensive, and their application to spatial data is limited to few locations. Moreover, some of these models cannot account for well-known features in environmental data, such as decaying dependence strength as events become more extreme. Other problems are related to constraints imposed by the limiting models that do not naturally exist in the observed processes.
In this talk, I will present three different approaches to tackle the previous issues. The first approach is a computationally appealing method to model multiple extreme events over spatially rich regions that successfully captures weakening extremal dependence. The second and third approaches leverage the integrated nested Laplace approximation (INLA) framework, which allows fast and accurate inference in complex models applied to data with different levels of spatial coverage. We will see how to apply these methodologies using precipitation, wind speed, fishery and pollution data. I will conclude with some reflections on how EVS can be incorporated into widely used classical statistical models.
BIO: Daniela Castro-Camilo is a Lecturer in Statistics at the University of Glasgow. Her research focuses on the theory and applications of multivariate and spatial extremes, with a particular interest in environmental, geological, and ecological applications. During the last few years, her work has gravitated around the integrated nested Laplace approximation (INLA) method for Bayesian inference. Specifically, she has developed methods promoting the need to adequately capturing extremes observations within the usual statistical analysis centred around mean values. She has worked closely with INLA developers to implement and improve extreme value models to help to bridge the gap between statistical theory and practice. She co-authored the book “Advanced spatial modelling with stochastic partial differential equations using R and INLA” (CRC Press, 2018).
Institut de Physique Théorique, CEA, CNRS-URA, France
Understanding the spatial structure of cities: some results and challenges
Abstract: The recent availability of large amounts of data about cities allowed us to better understand the spatial organization of cities and how they evolve in time. In this talk I will present a small selection of results and also discuss some challenges. I will first present some tools for the characterization of infrastructure networks (such as roads and subways) and their temporal evolution. I will then discuss mobility patterns obtained from mobile phone data and the polycentric structure of cities. If time allows, I will end this talk by discussing theoretical and empirical challenges about urban sprawl.
BIO: Marc Barthelemy is a former student of the Ecole Normale Superieure of Paris. In 1992, he graduated at the University of Paris VI with a thesis in theoretical physics titled “Random walks in random media”. Since 1992, he has held a permanent position at the CEA (Paris) and since 2009 is a research director at the Institute of Theoretical Physics (IPhT) in Saclay and a member of the Center of Social Analysis and Mathematics (CAMS) at the Ecole des Hautes Etudes en Sciences Sociales (EHESS). He has worked on applications of statistical physics to complex networks, epidemiology, and more recently on spatial networks. Focusing on both data analysis and modeling, he is currently working on urban networks and various aspects of the emerging science of cities. Marc Barthelemy co-authored the book “Dynamical Processes on Complex Networks” (Cambridge Univ. Press., 2008), and published recently the books “The Structure and Dynamics of Cities” (Cambridge Univ. Press, 2016) and “Morphogenesis of spatial networks” (Springer 2018).
Associate Professor, Environmental Computational Science and Earth Observation Laboratory, Ecole Polytechnique Fédérale de Lausanne (EPFL)
Interactive deep learning for animal conservation from above
Abstract: Monitoring wildlife populations is a complex business, since it involves monitoring over large areas, with complex terrains and counting living animals that move (and can also be dangerous at close range). For all these reasons, as well as to increase frequency and reduce costs, Unmanned Aerial Vehicles (UAVs) are more and more used. UAVs indeed acquire large amounts of data, but then also raise the problem of detecting and counting the animals, in order to provide accurate counts, in an automatic way. In this talk, I will talk about how deep learning can help, especially when helped by enthusiastic nature lovers willing to screen images for protecting wildlife.
BIO: Devis Tuia received a Ph.D. in Environmental Sciences at University of Lausanne in 2009. He was then a postdoc researcher at the University of València, Spain, the University of Colorado, Boulder, CO, USA and EPFL Lausanne. In 2014-2017, he was an Assistant Professor at the University of Zurich. He is now full professor at Wageningen University, the Netherlands. Since 2020, I joined EPFL Valais, to start the ECEO lab, working at the interface between Earth observation, machine learning and environmental sciences. His research focuses on geospatial computer vision, a field at the interface between GIscience, remote sensing and machine learning. He develops digital solutions to address problems of land planning and the environment. He led most of his efforts in urban recognition, land-use modeling and analysis, but he also have experience in wildlife tracking, environmental risk reduction and forest management through scientific collaborations.
Current works: (i) Making remote sensing accessible to everyone! Developing algorithms for human machine interaction; (ii) Open the black box: interpretable deep learning and uncertainties in environmental modeling; (iii) Digital wildlife conservation: using imaging to automatize censuses and conservation efforts.