Effect of ground surface interpolation methods on the accuracy of forest attribute modelling using unmanned aerial systems-based digital aerial photogrammetry

Abstract

ABSTRACT Unmanned aerial systems digital aerial photogrammetry (UAS-DAP) is an emerging technology that has the capacity to generate dense three-dimensional point clouds similar to airborne laser scanning (ALS). Over forested areas, these point clouds can be used to model forest attributes using the area-based approach (ABA). However, with DAP point clouds, canopy occlusion contributes to larger gaps in terrain registration from UAS-DAP compared to ALS point-clouds. Few studies have investigated the terrain modelling and forest inventory capacity of UAS-DAP over complex coniferous forests. In this study, we applied common terrain surface-interpolation routines using an established set of optimal UAS-DAP ground points and analysed how these routines influenced the prediction accuracy of forest stand attributes. Interpolation routines included inverse-distance weighted (IDW), natural neighbour (NATN), triangulated irregular network (TIN), and spline with tension (SPLT). The forest attributes of interest included mean tree height (H mean), Lorey’s height (H Lorey) and stem volume per hectare (V stem). Models were developed using metrics calculated from the vertical distribution of the UAS-DAP point cloud normalized by the different UAS-DAP terrain surfaces in addition to a reference surface generated from commercially provided ALS ground points. Results showed no significant difference between predictions derived from different terrain surfaces for all three dependent variables; however, the IDW method produced a distribution of wall-to-wall predictions most similar to those from the ALS-DEM. The best performing forest attribute models for H mean, H Lorey and V stem yielded mean RMSE values of 1.19 m (7.29%), 0.92 m (5.04%) and 54.55 m3 ha−1 (26.66%) respectively across the four UAS-DAP terrain surfaces generated. Model performance was higher yet comparable when using the ALS-DEM for point cloud height normalization with RMSE values of 0.73 m (4.43%), 0.59 m (3.24%) and 37.31 m3 ha−1 (18.24%).