Modélisation prédictive de l'occupation des sols en contexte agricole intensif Application à la couverture hivernale des sols en Bretagne

In intensive agricultural regions , winter vegetation covering is a key indicator of water transfer processes. Its prediction can help local decision making for restoring water quality. Though, spatial prediction modelling of winter land cover is complex and thus it appears necessary to introduce uncertainty in the modelling process, especially as high spatial and temporal variability are encountered. The objective of this work is to develop a reproducible method to predict the land cover distribution for the following winter season for the two hypotheses "bare soils" and "covered soils". The selected modelling approach is based on an expert model using the Dempster-Shafer rule, because it considers both uncertainty and imprecision both in the modelling process and in the results, as confidence levels are associated with the predictions. The model has been applied on an experimental watershed located in Brittany. After having assessed the spatial and temporal dynamics of winter coverage from a multitemporal series of remote sensing images, and identified the driving factors of these changes, sources of information representing these factors are fused with the Dempster's rule. Results are ambivalent according to the studied hypotesis. The prediction scores are good at the watershed scale but present limits for allocating the land class at a field scale, especially for the "bare soils" class, because the level of conflict between sources of information is high. The Dezert-Smarandache theory, that allows to introduce paradoxical information in the modelling process, is then applied and increases the prediction scores for the "bare soils" class. The reproductibility of the modeling approach is then evaluated in applying the model on another study site, wider and less documented than the previous one. The "bare soils" class remains well evaluated with both fusion rules, but the spatial allocation of the land class at a field scale is still not well managed, that indicates the need to integrate new sources of ...