Key points are not available for this paper at this time.
Publicly available land cover maps do not accurately represent shrubs and saplings, an uncommon but ecologically relevant cover type represented by woody vegetation 6000 km2) mountainous region. We hypothesized that the combination of spectral (imagery) and structural (LIDAR) data would allow for discrimination of shrub/sapling cover from other cover types. Specifically, we created training data using segmented four-band imagery from the National Agricultural Imagery Program (NAIP). In addition to spectral information from imagery, we used topographic information (elevation, slope, and aspect) and a LIDAR-derived canopy height model to classify land cover within a pixel-based random forests framework. To assess model accuracy, we used image interpretation and an independent sample of validation points. Due to the fine resolution of predictor rasters across such a large geographic region, we classified five subregions (counties) separately. We also compared the landscape metrics calculated for our custom classification at fine (1 m) and coarse resolution (resampled to 30 m) to metrics calculated with National Land Cover Data (NLCD). We achieved an overall accuracy of 89% and >80% accuracy for each land cover class. The LIDAR-derived canopy height model was consistently ranked as the most important predictor of vegetative land cover classes. Compared with our custom classification, NLCD underrepresented pasture/grassland by up to 10% and overrepresented forest up to 30%. There was no correlation between percent shrub/sapling cover in our custom classification and NLCD, suggesting that NLCD is not reliable for applications concerned with this ecologically relevant cover type.
Bulluck et al. (Fri,) studied this question.