Evaluation Of Airborne LiDAR Data To Predict Presence / Absence

This study evaluates the capabilities of the NASA Experimental Advanced Airborne Research Lidar (EAARL) system in delineating vegetation assemblages in Jean Lafitte National Park, Louisiana. Five-meter-resolution grids of bare earth (BE), canopy height (CH), canopy-reflection ratio (CRR), and height of median energy (HOME) were derived from EAARL data acquired in September 2006. Ground-truth data were collected along transects to assess species composition, canopy cover, and ground cover. Comparisons of the capabilities of general linear models (GLM) and generalized additive models (GAM) were conducted using conventional evaluation methods (sensitivity, specificity, kappa statistics, area under the curve) and two new indexes, net reclassification improvement (NRI) and integrated discrimination improvement (IDI). GAMs were superior to GLMs in modeling the vegetation training data, but no statistically significant differences between the two models were achieved for predicting vegetation validation data using conventional evaluation methods, although statistically significant improvements in net reclassifications were observed. The goodness-of-fit and prediction accuracy for both models are influenced by data prevalence and occurrence, although GAM models perform much better in the case of training data for all vegetation communities. Vegetation communities with less than 60% prevalence (e.g., coarse woody debris, herbs, shrubs, floating aquatics, and palms) have less than 40% maximum deviance explained, with the exception of bare ground for which the deviance explained by the GAM model is 64%. Midstory and canopy trees that have over 80% prevalence and over 10% occurrence have 99% deviance explained by the GAM model. For the validation dataset with vegetation community prevalence above 35%, GAM models show improvements in NRI only for vegetation categories with occurrences above 10%, and improvements in IDI for vegetation categories with occurrences above 20%. Different vegetation categories may be defined by the same structural characteristics, since LiDAR metrics are more closely related to vegetation structure rather than to species alone.