Estimation of forest height, biomass and volume using support vector regression and segmentation from lidar transects and Quickbird imagery

Lidar (light detection and ranging) remote sensing can accurately characterize forest vertical structure, such as canopy height, above-ground biomass (AGB) and timber volume; however, data acquisition is expensive. To reduce costs, one potential method is to integrate (small area) lidar transects and (large extent) optical imagery to estimate forest characteristics. Typically, multiple regression is used to link variables extracted from lidar transect data and optical imagery. Height information is then generalized from the area covered by lidar transects to other areas without lidar coverage. However, multiple regression models may not fully capture the complex relationship between variables. Fortunately, Support vector regression (SVR) provides a solution to deal with such complex nonlinear problems. Using a case study in Vancouver Island, Canada, SVR was applied to generalize canopy height from lidar transect(s) to the entire study area (2601 ha) based on a segmented Quickbird image. Results show that: (i) compared to typical multiple regression models, the SVR models provided better results for estimating canopy height; (ii) by using only one lidar transect (i.e., 8.8% cover), the SVR model generates an average canopy height estimation error of 6.2 m - which is less than a British Columbia forest inventory height class (9.0 m); and (iii) the final model estimates have relatively high correlations with field data for forest canopy height (R²: 0.81), AGB (R²: 0.76) and volume (R²: 0.64), while representing dramatically reduced acquisition costs.