Spatial Simulation of Agricultural Practices using a Robust Extension of Randomized Classification Tree Algorithms

In this paper, extensions of the classification tree algorithm and analysis for spatial data are proposed. These extensions focus on: (1) a robust manner to prune a classification tree to smooth sampling (e.g., spatial sampling effects), (2) an assessment of tree spatial prediction performances with respect to its ability to satisfactorily represent the actual spatial distribution of the variable of interest, and (3) a unified framework to aid in the interpretation of the classification tree results due to variable correlations. These methodological developments are studied on an agricultural practices classification problem at an agricultural plot scale, specifically, the weed control practices on vine plots over a 75 km2 catchment in the South of France. The results show that, with these methodological developments, we obtain an explicit view of the uncertainty associated with the classification process through the simulation of the spatial distribution of agricultural practices. Such an approach may further facilitate the assessment of model sensitivities to categorical variable map uncertainties when using these maps as input data in environmental impact assessment modelling.