An algorithm to automate the filtering and classifying of 2D LiDAR data for site-specific estimations of canopy height and width in vineyards

The 3D characterisation of individual vine canopies with a LiDAR sensor requires point cloud classification. A Bayesian point cloud classification algorithm (BPCC) is proposed that combines an automatic filtering method (AFM) and a classification method based on clustering to process LiDAR data. Data were collected on several grape varieties with two different modes of training. To evaluate the quality of the BPCC algorithm and its influence on the estimation of canopy parameters (height and width), it was compared to an expert manual method and to an established semi-automatic research method requiring interactive pre-treatment (PROTOLIDAR). The results showed that the AFM filtering was similar to the expert manual method and retained on average 9% more points than the PROTOLIDAR method over the whole growing season. Estimates of vegetation height and width that were obtained from classification of the AFM-filtered LiDAR data were strongly correlated with estimates made by the PROTOLIDAR method (R2 = 0.94 and 0.89, respectively). The classification algorithm was most effective if its parameters were permitted to be variable through the season. Optimal values for classification parameters were established for both height and width at different phenological stages. On the whole, the results demonstrated that although the BPCC algorithm operates at a higher level of automation than PROTOLIDAR, the estimates of canopy dimensions in the vineyards were equivalent. BPCC enables the possibility to adjust the spray rate according to local vegetative characteristics in an automated way.