Résumé
This study investigates a new methodology to estimate canopy plant area index (PAI) and the distribution of PAI in 3D from point clouds acquired using a system of 3 LiDAR (2 looking at nadir, and a third one at 45°) mounted in the Phenomobile ground robot designed high-throughput phenotyping of field crops. The method relies in the computation of the gap fraction at the voxel level using the trajectories of the laser beams to invert numerically the Beer-Lambert law (BL) and estimate simultaneously PAI and ALA (average leaf angle, in°). More specifically, this study addresses two research questions: (i) can this methodology provide a reliable estimation of both voxel- and canopy-level PAI and ALA of wheat using the LiDAR system of the Phenomobile? and (ii) how the reliability of PAI and ALA estimations may be affected by the number of LiDAR used, the assumptions of BL and the uncertainties in the point cloud?The reliability of this methodology was evaluated both in silico (using the AdelWheat 3D model coupled with a point cloud simulator) and in actual phenotyping trials of a panel of 10 bread wheat genotypes grown in two study sites under different treatments of plant density, water stress and sowing dates. The in silico analysis indicated that the numeric inversion of BL with the LiDAR system of the Phenomobile permits to estimate PAI and ALA at the voxel and at the canopy level with relative error of 9.8% (RMSE = 0.32, r2 = 0.98) and 8.4% (RMSE =5.4°, r2 = 0.95) respectively. The use of 3 LiDAR provided the best results. Possible uncertainties in the LiDAR point cloud impact negatively the voxel-level PAI retrieval but not the total canopy PAI. The validation in actual phenotyping experiments against destructive measurements produced satisfactory results for canopy PAI (RMSE = 1.43, r2=0.82). Furthermore, the methodology permitted to characterize the variability of 3D architecture across actual genotypes through the voxel-level estimates of PAI and ALA.