Land cover change prediction with a new theory of plausible and paradoxical reasoning
by
Samuel Corgne, Laurence Hubert-Moy, Jean Dezert, Grégoire Mercier
CNRS Costel Lab, Rennes, France; Onera, France; ENST Bretagne, France

In intensive agricultural regions, land cover during winter has an important impact on the water quality, and the identification and monitoring of vegetation covering dynamics at high spatial scales constitute a prior approach for the restoration of water resources. The spatial prediction modeling of land cover at the field scale in winter that appears useful for land management and helping local decision making, is specially complex because of the high variability of the factors that motivate the land cover changes between each winter. Thus, uncertainty in the data and the results has to be integrated in the modelling process for better decision making. Dempster's fusion rule has been used in a preceding study to spatially predict the location of bare winter fields for the next season on a watershed located in an intensive agricultural region. The data integrated in the model come from different sources : remote sensing, expert knowledge and alphanumeric data. They statistically express the driven factors that motivate land cover changes : the past-observed bare soils, field size, distance from farm buildings and agro-environmental actions. Identified factors which support the hypothesis "bare soils" and "covered soils" are transformed through fuzzy membership functions into mass function maps before being fused using the Dempster's rule. The model well predicts the presence of bare soils on 4/5 of the total area, but clearly presents some limits in generating errors in land cover assignment when the level of conflict, between the sources of evidence that support the hypothesis, becomes important.To solve this problem, we applied the Dezert-Smarandache Theory (DSmT), which can be considered as a generalization of the Dempster-Shafer Theory (DST). In this new theory, the rule of combination takes into account both uncertain and paradoxical information. This method offers a specific framework because unlike the DST, the frame of discernment is exhaustive but not necessarily exclusive. Thus, any source of information that can be rational, uncertain or paradoxical can be combined. The aim of this study is to evaluate the reliability of DSmT in managing conflicts between sources which support the hypothesis defined here to predict land cover vegetation presence in the fields. The framework is set up with the dealing of paradoxical information for the four sources of evidence, through an hyper-power set created with union and intersection operators. The fusion process applied is justified from the maximum entropy principle and decision making relies on a pignistic probability function. Results are presented and compared with the results drawn from the classical Dempster-Shafer theory.We point out that higher levels of well predicted fields are achieved for both assignment classes "bare soils" and "covered soils". Furthermore, the conflict management between the sources of evidence allows to spatially represent fields where the conflict is the strongest and to contribute to a better understanding of the factors that motivate land cover changes. Finally, the fusion process lead to relevant results to make a decision for the issue of bare soils reduction in agricultural intensive regions.Through this first application of the Dezert-Smarandache Theory, we show an example of this new approach ability to solve practical problems where the Dempster-Shafer usually fails.

Date received: February 19, 2003