Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Accueil du site → Doctorat → Royaume-Uni → 2001 → Reasoning with uncertainty in remote sensing

University of Surrey (2001)

Reasoning with uncertainty in remote sensing

Ahmadzadeh, M R

Titre : Reasoning with uncertainty in remote sensing.

Auteur : Ahmadzadeh, M R

Université de soutenance : University of Surrey

Grade : Doctor of Philosophy (Ph.D.) : 2001

Résumé
This thesis is on information fusion in remote sensing. Several fusion approaches are inves- tigated and some of them are successfully implemented. Assessing the risk of desertification of a forest after a fire, which is the main motivation of this work, depends on many factors. Aggregation of these factors, which are derived from different sources, gives a basis for the evaluation of the risk of desertification. Different ways of considering the uncertainty and imprecision due to data, concepts, mea- suring instruments etc in decision making systems, lead scientists to develop different in- formation fusion approaches. In this work I concentrate on uncertainty in the data due to errors in interpolation. The slope and aspect of the terrain are among those factors which influence the risk of desertification. The slope and aspect of the terrain are derived by a Geographic Information System (GIS) from the Digital Elevation Model (DEM). The prob- lein is that although the sources of these data are usually of diverse resolution, all of them are re-sampled to refer to the same resolution. Re-sampling, which is done by interpolation, introduces errors in the data. Most commercial GISs, in spite of these errors, deal with the data during decision making as if they were precise. Modelling the errors of slope and aspect when computed from interpolated data is the first objective of this thesis. The proposed error models may be used subsequently in the decision making process. Studying different fusion of information approaches to combat the problem at hand is the second contribution of this thesis. Especially I focus on the Dempster-Shafer evidence theory and its application in combining multi source data. First I use Dempster’s rule of combination as a tool for combining two classifiers : a Bayesian network and a fuzzy logic classifier. These two classifiers have been proposed in the past to assess the risk of desertification of burnt forests. The problem is that one of the classifiers has 3 classes (Bayesian network classifier) and the other one 5 classes (fuzzy logic classifier). To combine these two classifiers a superset of classes is defined, with the help of which the classes of each classifier can be defined by the union of few superset classes. The novelty of the proposed methodology is that not only the two classifiers are of different types (a probabilistic classifier and a Fuzzy logic-based classifier), but also the number of output classes are different. Finally I examine three other combining approaches : neural network approach, fuzzy neural network approach and application of Dempster-Shafer evidence theory in propagating the belief functions through a network in an expert system. In each experiment the results are compared with the expert results which are derived by inspecting the field data.

Mots clés : Information fusion ; Desertification ; GIS ; Neural Pattern recognition systems Pattern perception Image processing Geography Artificial intelligence Pattern recognition systems Geography Artificial intelligence

Annonce

Version intégrale (10 Mb)

Page publiée le 4 juin 2009, mise à jour le 22 octobre 2018