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

Stochastree, a crop transition model based on stochastic
decision trees, considering agronomic driving factors.
Luc Sorel a,b , Christian Walter a,b *, Patrick Durand a,b , Valérie Viaud a,b ;
a INRA, UMR1069, Sol Agro et hydrosystème Spatialisation, F-35000 Rennes
b Agrocampus Rennes, UMR1069, Sol Agro et hydrosystème Spatialisation, F-35000 Rennes
* Corresponding author: christian.walter@agrocampus-rennes.fr INRA-Agrocampus Rennes, UMR SAS, 65 rue de
Saint Brieuc, CS 84215, 35042 Rennes Cedex, France (phone: 0(+33) 2-23-48-54-39 / fax: 0(+33) 2-23-48-54-30)
Abbreviations: AWD, area-weighted distance; CTM, crop transition model; LCM, landcover change model;
MCM, Markov chain model; SDT, stochastic decision tree; SRP, self-replacement probability.
Keywords: crop transition, decision tree, landcover change modeling, transition probability matrices,
stochastic modeling.
Abstract
Evaluating the environmental impacts of agricultural practices requires the use of simulation approaches to
model hydrological fluxes at the landscape scale, whose response time can span over decades. The crop
allocation over the field pattern determines the spatial distribution of fertilization practices, which are
involved in soil and water quality issues. To facilitate the construction of agricultural scenarios, our objective
was to develop a crop transition model able to account for agronomic and spatial driving factors: crop
production objectives, spatial distribution of the crops around the farmsteads, and preferential allocation of
crops on soil waterlogging classes. We developed an innovative model based on stochastic decision trees,
Stochastree, to integrate farm-type and field characteristics (area, distance to farmstead, waterlogging,
current crop) in the crop transition simulation process without prior expert knowledge, but relying on a data-
mining approach. Simulation results of Stochastree were compared to field data and to a Markovian model
based on transition matrices, Rotomatrix, to test its abilities to follow presumed agronomic constraints.
Learned decision trees had a general structure similar to transition matrices. Stochastree and Rotomatrix
exhibited similar performances in predicting crop transitions and outputting expected crop productions.
Stochastree proved to be significantly superior in maintaining the spatial distribution of crops around the
farmsteads and slightly better in allocating crops to the proper soil waterlogging class. The ease to construct
decision trees suggests many potential couplings of Stochastree to various ecological models, like nutrient
diffuse transfer models or gene transmission models.
III. Stochastree, un modèle de successions de cultures basé sur des arbres de décision stochastique – p. 71