Using LiDAR and Random Forest to improve deer habitat models in a managed forest landscape

Abstract Conservation strategies are hindered by a lack of accurate maps of important habitat for many wildlife species, but especially for species inhabiting managed forest landscapes. Prioritizing restoration efforts on Alaska’s Tongass National Forest from past extensive clearcut logging is extremely challenging given the difficulty in accurately mapping its remote, rugged temperate rainforest landscapes. We tested the application of airborne light detection and ranging (LiDAR) technology to build a winter habitat model for Sitka black-tailed deer (Odocoileus hemionus sitkensis), the primary herbivore in the coastal temperate rainforest. We analyzed the importance of geomorphometric and forest structure characteristics as predictors of deer winter habitat selection using Random Forest applied to a 3-year GPS relocation dataset collected from 40 adult female deer. The LiDAR-based habitat model had a predictive performance of 94% (Out-of-bag error = 6%), a 10% lower model error compared to air-photo interpreted polygons and modeled plot data. Random Forest also outperformed analogous resource selection function models based on a comprehensive k-fold cross-validation. Deer habitat selection patterns in the LiDAR-based model were nonlinear across geomorphometric and forest structure predictive variables, and generally supported existing studies of deer habitat selection. Besides improving deer conservation and management on the Tongass National Forest, our approach could greatly enhance the accuracy and resolution of habitat maps used for conservation and restoration planning across large managed forest landscapes.