GENERALIZED SPATIAL MODELS OF FOREST STRUCTURE USING AIRBORNE MULTISPECTRAL AND LASER SCANNER DATA

Forest structure was modelled to explore the nonparametric approaches for inventory of small forests. Models were built using information from airborne optical and LiDAR data across two adjacent forest sites using the remote sensing data collected by a similar instrument. Off-site samples from the inventoried reference stands were combined with drawn samples from the target area for modelling responses. The applied combinations included two imputation methods, three sampling designs, and two predictor subset sizes, where a Genetic Algorithm (GA) was used to prune the predictors. Diagnostic tools including Root Mean Square Error, bias and Standard Error of Imputation were employed to evaluate the results. Results showed that Random Forests produced more accurate results than Most Similar Neighbour inference, yet was slightly more biased than MSN. Appending systematically stratified samples from the target dataset yielded more accurate results, yet it was most influential up to a low-to-medium intensity. The use of pruned predictors resulted in reduced bias. By bootstrapping the RMSE, the majority of the simulations lied within the confidence intervals. The applied methods show positive potential towards producing spatial models of forest structure.