UAV IMAGERY FOR TREE SPECIES CLASSIFICATION IN HAWAI'I: A COMPARISON OF MLC, RF, AND CNN SUPERVISED CLASSIFICATION

Abstract

Very-high resolution unmanned aerial vehicle (UAV) imagery coupled with emergent automated classification methods show great promise for fast and affordable remote sensing analysis. Tree species classification through remote sensing has traditionally been limited by spatial resolution of satellite imagery, or cost and logistics associated with aerial imagery collection. In this study, the use of red-green-blue (RGB) UAV imagery was assessed for supervised classification of multiple tree species within a tropical wet forest in Hawai‘i characterized by high species diversity and limited site accessibility. Three classifiers were tested: maximum likelihood classifier (MLC), random forest (RF), and convolutional neural network (CNN) U-Net. MLC and RF were additionally tested with the addition of texture statistics. U-Net achieved highest overall accuracy of 71.2%, compared to MLC with 48.1% and RF with 52.1%. MLC and RF both benefited from the addition of texture statistics. This study presents a novel comparison of three important classifier types and their capabilities with an emergent remote sensing data source. Findings from this study are consistent with those of recent studies and suggest that easily-acquirable RGB UAV imagery contains the necessary information for fine-grained classification at the species level, especially when utilizing a CNN.