Using UAV imagery to map invasive Phragmites australis on the Crow Island State Game Area, Michigan, USA

Abstract

Wetland managers in North America spend a great deal of time and money trying to control invasive Phragmites australis. Accurate mapping with remote sensing imagery is key to these efforts, which are increasingly employing uncrewed aerial vehicle (UAV) imagery. We mapped P. australis on the Crow Island State Game Area using UAV-derived single-date and multi-date RGB imagery combined with a Digital Surface Model (DSM). In addition to a traditional maximum likelihood classification (MLC), we used two machine-learning (ML) classification algorithms: support vector machine (SVM) and neural network (NN). We assessed accuracy based on both the traditional global model (overall accuracy [OA], omission [OE] and commission [CE] errors for the Phragmites class, and Kappa statistic) and local, per-patch accuracy broken down across 5 density classes and 3 size classes. Our global accuracy assessment for single-date imagery found that SVM (72% OA, 10% OE, 16% CE) performed similar to MLC (70% OA, 17% OE, 8% CE), while NN (33% OA, 7% OE, 41% CE) performed worse. The use of multi-date imagery had little effect on accuracy (MLC 64% OA, 21% OE, 12% CE; SVM 71% OA, 11% OE, 17% CE) except with NN, where the additional bands led to much higher accuracy (67% OA, 7% OE, 22% CE). These results were largely mirrored in the per-patch accuracy assessment, where SVM performed slightly better than MLC and NN performed poorly due to high commission errors. Regarding patch size and density, both larger and medium sized patches, as well as denser patches, were identified relatively accurately, but smaller patches tended to be overestimated and lower-density patches exhibited high omission errors. These results show that wetland managers can achieve very acceptable mapping accuracies with simple methods that require little in the way of resources and expertise.