Paysages virtuels et analyse de scénarios pour évaluer les impacts ...

1996; Wu and Webster, 1998; Verburg et al., 2004), and explicitly rule-based models (Thornton and Jones,
1998; Largouët and Cordier, 2001; Gaucherel et al., 2006a) require arbitrary decisions and strong expertise
(Li and Yeh, 2002; Houet and Hubert-Moy, 2006). Generalized linear models (Aspinall et al., 2004; Huang et
al., 2007) rely on parametric hypotheses and are obscure to interpret, which complicates the design of new
landcover change scenarios. On the contrary, decision trees (DT) are non-parametric methods known to be
efficient without parameter tuning, accessible and useful for non-experts (Quinlan, 1981; Breiman et al.,
1984; Murthy, 1998; Perlich et al., 2003). In the field of landcover modeling, DT are mainly used as
classification tools based on remotely sensed data, pedological and geomorphological factors, or spatial
descriptors of the farm territory (Lynn et al., 1995; DeFries and Chan, 2000; Friedl et al., 2002; Lawrence et
al., 2004; Thenail and Baudry, 2004). Although DT have been used to simulate iterative temporal processes
in other research fields (Gladwin, 1989; Heidenberger, 1996; Jordan et al., 1997; Hazen, 2002; Tarim et al.,
2006), such applications of DT have not yet been done in landcover change modeling, to our knowledge.
Our objectives were to bridge the gap between exploratory approaches and crop transition modeling by
inferring stochastic decision trees trained by a data-mining analysis of a crop transition learning dataset, and
to use them to simulate crop transitions over several decades with the incorporated agronomic driving
factors. These objectives involved the following steps: (1) to identify relevant driving factors of crop
transition on an agricultural catchment, (2) to build a crop transition model based on stochastic decision trees
(named Stochastree hereafter) and simulate summer crop transition while accounting for agronomic driving
factors, (3) to evaluate the abilities of Stochastree and a reference temporal first-order Markov chain model
(named Rotomatrix hereafter) to comply with agronomic constraints. Finally, we discuss the structure of the
models and the mutual benefits of a compared evaluation of Stochastree and Rotomatrix simulations.
III. Stochastree, un modèle de successions de cultures basé sur des arbres de décision stochastique – p. 73