Multi-model estimation of understorey shrub, herb and moss cover in temperate forest stands by laser scanner data

In temperate forests, the highest plant richness is regularly found in the understorey, i.e. shrub, tree regeneration, herbal and moss covers, which provides important food and shelter for other plant and animal species. Here, Light Detection And Ranging (LiDAR) remote sensing was investigated as a surrogate to laborious field surveys to improve understanding of the causal and predictive attributes of understorey. We designed a study in which we used a high-density LiDAR point cloud and applied a thinning algorithm to simulate two lower density point clouds including first and last returns and half of the remaining points (half-thinned data) and only first and last returns (F/L-thinned data). From each dataset, several over-and understorey-related statistical metrics were derived. Each of the three sets of LiDAR metrics was then combined with the forest habitat information to estimate the recorded proportions of shrub, herb and moss coverages. We used three different model procedures including zero-and-one-inflated beta regression (ZOINBR), ordinary least squares with logit-transformed response variables (logistic model) and a machine learning random forest (RF) method. The logistic and ZOINBR model results showed highly significant relationships between LiDAR metrics and habitat types in explaining understorey coverage. The highest coefficients of determination included r(2) = 0.80 for shrub cover (estimated by F/L-thinned data and ZOINBR model), r(2) = 0.53 for herb cover (estimated by half-thinned data and logistic model) and r(2) = 0.48 for moss cover (estimated by half-thinned data and logistic model). RF models returned the best predictive performances (i.e. the lowest root mean square errors). Despite slight differences, no substantial difference was observed amongst the performances achieved by the original, halfthinned and F/L-thinned point clouds. Moreover, the ZOINBR models did not improve predictive performances compared with the logistic model, which suggests that the latter should be preferred due to its greater simplicity and parsimony. Despite the differences between our simulated data and the real-world LiDAR point clouds of different point densities, the results of this study are thought to mostly reflect how LiDAR and forest habitat data can be combined for deriving ecologically relevant information on temperate forest understorey vegetation layers. This, in turn, increases the applicability of prediction results for overarching aims such as forest and wildlife management.