Comparing Modeling Methods for Predicting Forest Attributes Using LiDAR Metrics and Ground Measurements

Selected modeling methods are compared for predicting 5 forest attributes, basal area (BA), stem volume (VOL), Lorey's height (LOR), quadratic mean diameter (QMD), and tree density (DEN), from airborne LiDAR metrics in southwestern Oregon, in the United States. The selected methods included most similar neighbor (MSN) imputation, gradient nearest neighbor (GNN) imputation, Random Forest (RF)-based imputation, BestNN imputation, ordinary least square (OLS) regression, spatial linear model (SLM), and geographically weighted regression (GWR). Several performances of each method were assessed by 500 simulations with different numbers of training data. No single modeling method was always superior to the others in prediction of the forest attributes. The best method varied according to response variable, prediction type, and performance measures, even though there was a leading group (SLM, OLS, BestNN, and GWR) that always outperformed the other methods in root mean squared prediction error (RMSPE). The model's performance was quite affected when the number of training data used in the modeling procedure was small. The optimal sizes of training data were 100-150 for point prediction and 200-250 for total prediction. SLM showed its applicability to wider conditions in that it produced better performance in most cases. RF imputation produced poorer performances than the other methods, particularly with lower prediction interval coverage. This might be because RF imputation had some bias and a smaller prediction standard error; the poor performance by RF did not stem from the smaller number of predictor variables.