Description
Exotic plant invasion is a major environmental issue and is particularly critical for riparian ecosystems of California. Early detection of invasive plants is crucial for effective weed management. In this study, high spatial resolution multispectral imagery was captured and utilized for detecting the invasive plant Arundo donax. This invasive plant was detected and delineated using imagery captured with a digital color infrared imaging sensor mounted on an unmanned aerial system (UAS) in August of 2011 and a light manned aircraft along the San Diego River in May 2009. The multispectral image data are characterized by 12 cm and 81cm spatial resolutions respectively with spectral coverage in the visible (V) and near infrared (NIR) portions of the light spectrum. Classification techniques were tested to exploit different attributes of the very high spatial resolution imagery. Three classification approaches were evaluated including: (1) a per-pixel maximum likelihood approach using ERDAS Imagine software, (2) a neighborhood based or pseudo-object oriented approach using Feature Analyst software, and (3) a true object-based approach using Imagine Objective software. To test the reliability and accuracy of the three different classification routines, the accuracy assessment included a stratified random sampling design scheme utilizing a 16 m point-to-point distance rule. Results indicate A. donax detection varied substantially throughout both study sites, between image datasets, and across a range of image classification algorithms. A. donax exhibits a wide range of biophysical characteristics within the riparian environment which make it difficult to determine a single optimal detection routine. Results derived from the statistical accuracy assessment suggest that: (1) A. donax patches as small as 3.5 m_ can be detected by off-the- shelf multispectral imaging sensors aboard low cost, flexible imaging systems based on unmanned aircraft systems (UAS) and light-sport aircraft (LSA) platforms, (2) a maximum likelihood classifier can accurately detect expansive A. donax colonies using 81 cm multispectral imagery, (3) the spatial contextual classifier of Feature Analyst was more efficient at detecting small and mixed A. donax patches, (4) the object-based Imagine Objective classifier utilizing image segmentation consistently yielded less accurate detection results, (5) ultra-high spatial resolution imagery obtained by a UAS can achieve similar A. donax detection results compared to coarser spatial resolution imagery obtained by a manned LSA in small and sparse communities, and (6) detection accuracy of all classification routines decreases when A. donax distributions are sparse, relatively small and mixed within other riparian vegetation types. Based on these findings an optimal routine to map A. donax would utilize a maximum likelihood classifier to accurately detect existing, large colonies combined with the spatial contextual classifier of Feature Analyst to more efficiently map small and sparse distributions
