The role of ground reference data collection in the prediction of stem volume with LiDAR data in mountain areas

Ground reference data collection represents an important element in the prediction of stem volume with LiDAR-derived variables, and at present it is the most expensive part of such analyses. In this paper two aspects of ground reference data collection were analyzed: (1) the positioning error of the ground plots; and (2) the optimal number of training plots. A system for the prediction of stem volume at area-based level was adopted. LiDAR data were preprocessed and 13 variables describing both height and coverage were extracted. Models were defined using a stepwise ordinary least square (OLS) regression. Three experiments were conducted: (i) the role of the plots positioning error on prediction accuracy; (ii) the influence of random downsampling of plot numbers on prediction accuracy; and (iii) the influence of a stratified downsampling of plot numbers on prediction accuracy based on LiDAR-derived variables. A dataset comprising 799 ground plots was used. They were distributed throughout a mountainous area in the Southern Alps, where the presence of a complex landscape increases the uncertainty of the Global Positioning System (GPS) accuracy, and where a large variety of tree forest species and climatic environments make it necessary to have a large number of sample plots for accurate characterization of the study area. All the experiments provided important indications for LiDAR based forest inventories: the GPS error did not significantly influence the prediction accuracy and it was possible to reduce the number of training samples without compromising the generalization ability of the prediction model. Leading on from these findings, a new ground sampling protocol based on genetic algorithms was proposed. The new protocol allowed us to obtain promising results for the considered dataset: using only 53 training plots, instead of 534 in the original dataset, we obtained the same results for the validation set. These results, obtained in a complex mountainous area, are representative of Alpine environments and allow us to infer that similar (or better) results could also be obtained within non mountainous areas.