Session Information
| Session | Poster Session | | Date | Monday (2008-04-07) | | Time | 5:00 PM - 7:00 PM | | Room | Grand Terrace |
Presentation Information
| Presenter | Todd Hawbaker | | Title | Hierarchical stratified sampling optimizes forest structure maps derived from LiDAR data | | Affiliation | University of Wisconsin | | Authors | Todd Hawbaker, Adrian Lesak, Volker Radeloff | | Keywords | Forest structure, LiDAR, Remote sensing, Sampling design, Statistics | | Presentation Type | Poster | Abstract:
The vertical structure of forests is important for forest, wildlife, and ecosystem processes. However, structure can be highly variable, even within seemingly homogenous forest stands, and is notoriously hard to measure for large areas. Light detection and ranging (LiDAR) data measures vegetation height and, when used with ground-truth data, can assess vertical structure at landscape scales. The problem is that selection of ground-truth data collection locations can affect the accuracy of models, and model accuracy is poor unless the entire range of variability is sampled adequately. We used LiDAR-derived vegetation heights as prior data to select sampling locations across a 73,500 acre landscape in northern Wisconsin. We summarized the mean and variance of LiDAR heights within 30 meter grid cells, divided mean heights into 10 equally-spaced strata and, within each height strata, grouped variance into 3 strata, producing 30 strata total. In each stratum, one field plot was randomly selected. At each field plot, we measured tree diameters and heights. We compared models of tree diameters and heights developed using data from our sampling design with a completely random design. Both sets of models captured the same general patterns, but models developed with the hierarchical systematic design had better fit and lower error. The best LiDAR-based models of tree height had r-squared values of 60% and mean diameter had r-squared values of 55%. Plot locations selected using our hierarchical stratified design captured a greater range of variability in the mean and variance of vegetation heights. Plot locations selected using the completely random design included the most common values of height, but failed to sample the full range of variability. Random sample designs are statistically rigorous and simple to implement, especially when no prior knowledge is available. However, when prior information is available, as is the case in many LiDAR based mapping effort, the use of prior data to select field sampling locations ensures that the entire range of variability is sampled and this will greatly improve model results. |
|