Tradeoffs between lidar pulse density and forest measurement accuracy

Discrete airborne lidar is increasingly used to analyze forest structure. Technological improvements in lidar sensors have led to the acquisition of increasingly high pulse densities, possibly reflecting the assumption that higher densities will yield better results. In this study, we systematically investigated the relationship between pulse density and the ability to predict several commonly used forest measures and metrics at the plot scale. The accuracy of predicted metrics was largely invariant to changes in pulse density at moderate to high densities. In particular, correlations between metrics such as tree height, diameter at breast height, shrub height and total basal area were relatively unaffected until pulse densities dropped below 1 pulse/m2. Metrics pertaining to coverage, such as canopy cover, tree density and shrub cover were more sensitive to changes in pulse density, although in some cases high prediction accuracy was still possible at lower densities. Our findings did not depend on the type of predictive algorithm used, although we found that support vector regression (SVR) and Gaussian processes (GP) consistently outperformed multiple regression across a range of pulse densities. Further, we found that SVR yielded higher accuracies at low densities (< 0.3 pl/m2), while GP was better at high densities (> 1 pl/m2). Our results suggest that low-density lidar data may be capable of estimating typical forest structure metrics reliably in some situations. These results provide practical guidance to forest ecologists and land managers who are faced with tradeoff in price, quality and coverage, when planning new lidar data acquisition.

► Plot-scale forest metrics are derived from airborne lidar at 17 pulse densities. ► Most forest metric accuracies were invariant to pulse density until very low levels. ► Tree density, tree and shrub cover-related metrics required higher pulse density. ► SVM and Gaussian processes outperformed multiple regression.