| Přispěvatelé: |
UCL - SST/ELI/ELIE - Environmental Sciences, UCL - Ingénierie biologique, agronomique et environnementale, Allard, Denis, Hristopulos, Dionssios, Defourny, Pierre, Bragard, Claude, Bogaert, Patrick |
| Popis: |
Categorical variables have always played an important role in a wide variety of statistical applications in several scientific fields, including environmental sciences. In a spatial environmental context, such data naturally arise in geography, geology or remote sensing classification, although they might be used with quite different goals in mind. Due to sampling limitations, these data often are not spatially exhaustive (i.e., values are not known everywhere in the spatial domain of interest); therefore, modeling and spatial prediction steps are required at some stage of the study. Although dealing with spatial continuous data based on a statistical framework has generated considerable literature for a long time and has led to well-established methods, modeling and predicting spatial categorical data have proved to be much more complex. Elegant approaches have been advocated, such as the Bayesian maximum Entropy (BME) methodology that originates from the concepts of entropy maximization and posterior conditioning. Among its advantages, this methodology is distribution-free, allows us to integrate both hard and soft data, does not rely on restrictive assumptions and provides a complete posterior distribution while classical methods are often limited in providing only few statistical moments. Although the BME approach for categorical data is very general and is a real outsider compared to more traditional approaches, this method also suffers from some drawbacks. This work aims at generalizing further the theoretical results of the BME framework in order to (i) account for qualitative data, such as experts’ opinions or frugal information, through a minimum divergence approach, and (ii) ease the integration multiple information sources through a Bayesian data fusion methodology suitable for categorical data. The first part of this work sets the theoretical background while the second part is dedicated to applications that illustrate the benefits of the suggested methodology in the field of land cover mapping, where accounting for qualitative data (e.g., crowdsourced information) and integrating multiple data sources (e.g., combining several land cover products) are of prime interest. (AGRO - Sciences agronomiques et ingénierie biologique) -- UCL, 2018 |
