Meta-modeling light interception in crop : weed canopies

EAGESTADAGROSUPShort paper + poster; Introduction Weeds are a major pest of crop production but are important for biological diversity. In order to design cropping systems that reconcile crop production and biodiversity we need tools that allow us to test multiple cropping systems, such as FLORSYS (Colbach et al., 2014). This mechanistic, multi-annual and multi-specific model ("virtual field") represents the crop:weed canopy in 3D, with an individual-based light interception sub-model (Munier-Jolain et al., 2013), which is time-consuming and thus limits the number of tested cropping systems. The 3D canopy is discretized with voxels (3D pixels), and each plant usually consists of several voxels. Photosynthetically Active Radiation interception (PARi) and absorption (PARa) are calculated for each voxel, and then summed for each plant. Here, our objective was to accelerate the sub-model by developing a metamodel predicting light interception and absorption directly at the scale of the plant. Materials and Methods The light interception of a target plant depends on: (1) plant parameters (e.g. height, extinction coefficient); (2) physical environment (latitude, day), (3) model precision (voxel, field sample area), and (4) biological environment (density, location and parameters of neighbouring plants). We used a new global sensitivity method based on a truncated Legendre polynomial chaos expansion (PCE) whose coefficients are estimated by PLS regression (Gauchi et al., 2016) to: 1, rank inputs as a function of their polynomial and total effects on outputs via the so-called PCE-PLS, and; 2, to provide a meta-model predicting PARi and PARa at the plant level as a function of the four input types. The number of computer-based iteratuions required to build the metamodels were based on a space-filling design following a Sobol quasi-random sequence. For the meta-model, the canopy of target plants were aggregated into synthetic input variables (such as mean neighbour plant height, as opposed to actual input variables crop sowing pattern and weed density. Results and Discussion In isolated plants, target-plant characteristics are more important than physical environment or model precision for determining PARa (Fig. 1). Among the former, the variables determining the volume occupied by the target plant (height, width) dominate all others. The PARa meta-model consisted of a partial Legendre polynomial of 14th ESA Congress 5–9th September 2016 Edinburgh, Scotland 44 4000 mononomials of degree 7; presenting a high predictive quality (Fig. 2), with an index Q2cum of 0.95 (a specific PLS fitting-prediction criterion which should be close to 1). The results for target plants inside canopies will be available soon. Figure 1. Sensitivity index sensu PCE-PLS, in white polynomial effect (i.e. with no interactions), in black total effect of the inputs (plant parameters, physical environment and model precision). Figure 2. Scatter plot of the meta-model predictions and the ligth interception sub-model outputs. Conclusions The sensitivity analysis allowed us to discriminate those plant variables that must be predicted precisely (e.g. plant height) from those that can be estimated with less precision (e.g. leaf area distribution parameters). The meta-model reduces computation time without impairing prediction quality, which allows us to simulate numerous cropping systems and inform best-management practices and advice for farmers. Acknowledgements The present work was financed by INRA (EA and MIA divisions), the French project CoSAC (ANR -14 -CE18CE18CE18 -0007) and the Burgundy Region.