Seasonal Phragmites australis classification in Long Point National Wildlife Area wetlands using a remotely piloted aircraft system and random forest machine learning

Abstract

This study produced a high-accuracy remotely piloted aircraft system (RPAS) imagery classification method for identifying the invasive reed Phragmites australis ( Cav.) Trin. Ex Steud subsp. australis using random forest (RF) machine learning. RPAS imagery was collected in the spring and fall of 2019 using a fixed-wing RPAS equipped with a visible spectrum camera (eBee X, S.O.D.A. 3D; senseFly) in Long Point, Ontario, Canada. Imagery was used to produce separate early and late season classifications and a bi-temporal classification which used imagery from both dates. The overall accuracy achieved for each was 97%, 96%, and 91%, respectively. Digital surface models (DSMs) were the most important variable for identifying Phragmites in all classifications due to their greater height when compared to surrounding herbaceous vegetation. The bi-temporal classification, which utilized change in DSM value during the growing season, resulted in an estimated 47.8% new growth of Phragmites and appeared to capture sparse growth better than traditional classification differencing alone. This study highlights the promising use of high-resolution DSMs produced from RPAS imagery to classify invasive Phragmites and monitor within-year patch expansions.