Classifying Tree Species in a Semiarid Woody Encroached Landscape Using ML and DL Algorithms from Remotely Sensed UAV Imagery

The use of UAV optical sensors for image classification provide significant advantages overtraditional satellite and aerial sensors. Increases in spatial and temporal resolution allow forcontinuous monitoring of high level information classes. Furthermore, 3D point clouds can becreated through photogrammetry, adding physical information, thereby improving classificationresults. In this study, UAV flights were conducted in December 2018 over a semiarid junipershrub land in west central Texas to classify tree species using pixel and object-basedclassification, as well as a suite of traditional, machine learning, and deep learning algorithms. Atotal of 999 images were collected over 345.8 acres, with a spatial resolution of .0245m. A RGBorthomosaic was created, alongside ancillary spectral layers and a canopy height model. Adecision tree classifier was used to mask out a majority of ground and herbaceous pixels thatwould have decreased classification accuracy. The mosaic was segmented twice using differingsegmentation parameters to create two distinct segmented mosaics. The classification schemeincluded 5 classes: shadow, ground, Prosopis glandulosa (Honey Mesquite), Quercus virginiana(Live Oak), and Juniperus ashei (Juniper class). Using visual interpretation, 500 trainingpixels/20 training segments and 500 testing pixels were identified per class. Pixel and object-based maximum likelihood, random forest, and support vector machine classifications were runon the 3 mosaics and accuracy assessments were produced. The VGG-19 CNN developed byOxford University (Simonyan & Zisserman 2014) was used and trained using 5000 fully labeledimages and tested on 1000 labeled images using the same classification scheme. The model wasthen applied to the 3 masked mosaics and accuracy assessments were produced. The importanceof producing accurate classification maps for monitoring and environmental management willthen be discussed.