Identifying indicators of the spatial variation of agricultural practices by a tree partitioning method: The case of weed control practices in a vine growing catchment

Environmental impact assessments of agricultural practices on a regional scale may be computed by running spatially distributed biophysical models using mapped input data on agricultural practices. In cases of hydrological impact assessments, such as herbicide pollution through runoff, methods for generating these data over the entire water resource catchment and at the plot resolution are needed. In this study, we aimed to identify indicators for simulating the spatial distribution of weed control practices (WCP) in a French vine growing catchment. On the basis of interviews of 63 winegrowers, a spatially explicit database was developed that included 1007 vine plots and information regarding practices and potential explanatory variables. Four practices were differentiated according to the methods used (chemical weed control, shallow tillage, grass cover or a combination) that determine the intensity of herbicide use and potential surface runoff. Three groups of explanatory variables corresponding to three assumed levels of spatial organisation of WCP (the plot, the farm and the local government area (LGA)) were tested and compared. In the first step, selection of explanatory variables within each group was performed using a tree-partitioning method that combined the advantages of the CART algorithm (building an interpretable and controlled model) and the Random Forest algorithm (limiting overfitting) algorithm. In the second step, the performance of the selected variables for reproducing the observed repartition of practices was evaluated by a stochastic use of the tree, leading to a set of equiprobable spatial distributions of practices at the plot resolution. The results indicate that plot characteristics related to alley width play an important role in the weed control choices; however, to take into account the total diversity of the WCP, it appears to be necessary to focus on the farm holding variables and, in particular, on the variable LGA. However, the interpretation of these results is still difficult. Specifically, the great relevance of the variable LGA to discriminate the practices may be related to various factors, one of which is the distribution of soil properties within the Peyne catchment that still requires more precise characterization. The results also indicate that the combination of the three groups of variables leads to the highest-performing simulations of the spatial distribution of WCP. Nevertheless, the farm holding variables provided little additional spatial information, which supports the idea that they may be omitted without significantly impacting the final results.