Voxel-based metrics from lidar improve predictions of wood volume in structurally diverse forests

Background/Question/Methods

Large scale wind events such as hurricanes, tornado outbreaks, and derechos are known to alter forest structure, species composition, and ecosystem function at landscape-scales, but assessment of changes at large spatial scales can be difficult. Estimates of forest attributes using lidar point clouds summarized over sample areas and modeling using field data are commonly utilized to quantify structural change over heterogenous landscapes; however, these models are often parameterized for a single forest type and are unlikely to be accurate (and thus nontransferable) in other forest types. A robust model that estimates forest structural attributes over a range of forest types could allow for the estimation of changes over heterogeneous landscapes affected by wind disturbances. Recent research has found that the use of volumetric pixels (voxels) to describe lidar point clouds can improve predictions of structural attributes. We compared random forest regression models using area-based versus voxel-based variables on 864 sample plots in nine forest types on the southeastern coastal plain. Our research questions were: 1) Do voxel-based models provide more accurate wood volume estimates than area-based models? and 2) Can a single model generate robust predictions of wood volume over a range of forest types?

Results/Conclusions

The voxel-based model that was agnostic to forest type had a lower symmetric median absolute percentage error (SMdAPE) than the corresponding area-based model (23.5 and 27.5, respectively). These results compare well to wood volume models developed in the southwestern United States. The agnostic voxel-based model compared favorably to voxel-based models by forest type, of which only two had lower SMdAPE. By individual forest type, the models using voxel-based variables had lower SMdAPE than the models using area-based metrics in six forest types dominated by pine species, nearly equal SMdAPE in two forest types dominated by hardwoods, and higher SMdAPE in cypress and gum forests. Voxel-based metrics appear to capture structural variability in pine dominated ecosystems that area-based metrics do not. It has been hypothesized that the voxel-based approach may account for horizontal variability better than the area-based approach, and much of the pine dominated forest in our study area is spatially heterogenous with an open-canopied structure, with heterogenous gaps. Our results show that assessments of forest change using voxel-based summaries of lidar data capture changes that affect growth and ecosystem functioning in structurally complex forests.