Random forests to detect subsoiling and subsurface drainage effects on corn plant height and water table depth

Soil compaction is an agricultural and environmental concern. Understanding the parameters affecting crop productivity is essential for an efficient agriculture. We compared two statistical methods: random forest (RF), and multiple linear regression (MLR), to evaluate the effects of various parameters, including subsoiling, on corn (Zea mays L.) plant height and on seasonal mean water table depth. We performed subsoiling treatments in a clay field of low permeability, provided with subsurface drains and remodeled with rounded beds. These experiments allowed comparing a control without subsoiling to four subsoiling treatments: a tractor and a bulldozer operated parallel and perpendicular to the subsurface drains. Each treatment was randomized and repeated three times. In the spring of 2016, we drilled 198 60-cm deep water table depth observation wells. The corn height was estimated by photogrammetry. The results showed that both RF and MLR allowed determining the main factors affecting plant height and seasonal mean water table depth, for which subsoiling treatments played a negligible role. Coefficients of determination were much higher and prediction intervals much smaller for RF (R² = 0.94) than for MLR (R² : 0.28-0.69). RF allowed visualizing nonlinear relationships between mean water table depth, plant height and the predictor variables, such as XY coordinates, horizontal distance to the field open drain and depth to subsurface drains which are all related to the position on the beds. To maximize corn plant height in this type of soil, RF also showed that the horizontal distance to drain must be less than 3–4?m, and therefore optimal drain spacing is 7?m, horizontal distance to field open drains is more than 8?m, and optimal mean seasonal water table depth is greater than 0.25?m.