Analysis of Classification Methods for Identifying Stands for Commercial Thinning Using LiDAR

Abstract

Commercial thinning (CT) is an important tool that meets a diverse set of forest management objectives, including the generation of intermediate revenue, promotion of regeneration tree growth, and the modification of vertical and horizontal fuel structure for wildfire mitigation. Using a set of 653 fixed radius plots and a coincident LiDAR acquisition, we compared three different classification methods to predict CT eligibility for Douglas-fir (Pseudotsuga menziesii) stands in southwestern Oregon. We assessed logistic regression (LOG), random forests (RF), XGBoost (XGB) to classify areas eligible for CT operations based on three structural attributes, volume (VOL), basal area (BA) and Curtis’ Relative Density index (CRD). We also assessed their predictive performance and reliability via cross-validation at different sample sizes. We used the area under the receiver operating characteristic curve (AUC) as our primary performance measure. Estimated AUCs were 0.86, 0.77 and 0.68 for XGB, RF and LOG, respectively. We observed that classifier performance stabilized between sample sizes of 200 and 300 plots, which suggests that the development of a CT eligibility classifier is appropriate for operational applications of the method with similar sample sizes and large area attributes.