Individual tree species classification at the Acadia Research Forest using airborne lidar

Thumbnail Image
Journal Title
Journal ISSN
Volume Title
University of New Brunswick
Species knowledge is an essential component of any accurate forest inventory. In this study, 14 Acadian tree species were remotely classified with the Random Forests classifier using two LiDAR datasets (1 pulse per m2 and 6 pulses per m2). The 6 pulses per m2 dataset was found to perform better for individual tree species identification with 11.1 % higher overall accuracy (1 pulse per m2 accuracy = 32.8 %, 6 pulses per m2 = 43.9 %). Additional modeling strategies were explored to further boost accuracy, including aggregating species by product market class (+ 16.9 % increase in accuracy) and adding contextual features such as elevation, depth to water and site information (+ 6.9 %). Further improvements were possible by using a multi-step hierarchical species-group modeling approach (+ 3.4 % for softwood groups to 71.1 % overall accuracy and +16.1 % for hardwoods to 83.8 % overall accuracy). To further boost accuracy to an operational level, higher pulse density and/or additional variables including multi-spectral and multitemporal imagery should be explored.