Agriculture is the human activity that uses and transforms most of the Earth's surface. Agricultural intensification and simplification have created numerous ecological and environmental problems. To better manage the future development of agricultural landscapes, it is important to understand the past and current dynamics of agricultural landscapes at regional scales compatible with the scales where environmental and ecological services manifest themselves. Yet, most studies on agricultural dynamics at regional scales do not distinguish between the dynamics related to a steady-state farming activity and the dynamics related to changes in the mechanisms of farming activity. Meanwhile, studies reported in the literature that make this distinction have the disadvantage of being difficult to duplicate. The purpose of this thesis is to develop a generic method for modelling the past and current dynamics of Landscape Organization of Farming Activity (LOFA). We developed a stochastic modelling method based on Hidden Markov Models that allows data mining within a corpus of spatio-temporal land use data to segment the corpus and reveal hidden agricultural dynamics. We applied this method to land use corpora from various sources (field surveys, remote sensing) belonging to two agricultural landscapes of regional dimension : the study site of Chizé (430 km², Poitou-Charentes, France) and the catchment area of Yar (60 km², Brittany, France). This method provides three contributions to the modeling of LOFA : (i) LOFA description following a temporo-spatial approach that first identifies temporal regularities and then localizes them by segmenting the agricultural landscape into compact areas having similar temporal regularities ; (ii) data mining of the neighborhood of land use successions and their dynamics ; (iii) combining of the regularities revealed by our data mining approach at the regional level with rules identified by agronomy and ecology experts at more local scales to explain the regularities and validate the experts' hypotheses. We tested the generic nature of the first contribution on the two study sites. The last two LOFA modelling contributions were developed and tested on the study site of Chizé. Our results validate the hypothesis according to which LOFA fits well a Markov field of land-use successions. This thesis opens the door to a new LOFA modelling approach that investigates the combining of regularities and rules and that further exploits artificial intelligence tools. This work could serve as the beginning of what could become a numerical landscape agronomy.
