Mapping Phragmites cover using WorldView 2/3 and Sentinel 2 images at Lake Erie Wetlands, Canada

Abstract

Phragmites australis (Cav.) Trin. ex Steudel subspecies australis is an aggressive plant invader in North American wetlands. Remote sensing provides cost-effective methods to track its spread given its widespread distribution. We classified Phragmites in three Lake Erie wetlands (two in Long Point Wetland Complex (LP) and one in Rondeau Bay Marsh (RBM)), using commercial, high-resolution (WorldView2/3: WV2 for RBM, WV3 for LP) and free, moderate-resolution (Sentinel 2; S2) satellite images. For image classification, we used mixture-tuned match filtering (MTMF) and then either maximum likelihood (ML) or support vector machines (SVM) classification methods. Using WV2/3 images with ML classification, we obtained higher overall accuracy for both LP sites (93.1%) compared with the RBM site (86.4%); both Phragmites users’ and producers’ accuracies were also higher for LP (89.3% and 92.7%, respectively) compared with RBM (84.3% and 88.4%, respectively). S2 images with SVM classification provided similar overall accuracies for LP (74.7%) and for the RBM (74.3%); Phragmites users’ and producers’ accuracies for LP were 85.3% and 76.3%, and for the RBM, 69.1% and 79.2%, respectively. Using WV2/3, we could quantify small patches (percentage cover ≥ 20%; shoots ≥ 1 m tall; stem counts > 25) with accuracy > 80%, whereas parallel effort with S2 images only accurately quantified high density (> 60% cover), mature shoots (> 1 m tall; Stem counts > 100). By simultaneously mapping young or sparsely distributed Phragmites shoots and dense mature stands accurately, we show our approach can be used for routine mapping and regular updating purposes, especially for post-treatment effectiveness monitoring.