TY - JOUR
T1 - Performance of one-class classifiers for invasive species mapping using airborne imaging spectroscopy
AU - Skowronek, Sandra
AU - Asner, Gregory P.
AU - Feilhauer, Hannes
N1 - Funding Information:
The authors would like to thank N. Chiariello and T. Herbert for their support at JRBP. Many thanks to J. Barbosa and B. Mack for their advice on biased SVM classification and to D. Chadwick for proofreading the manuscript. S. Skowronek is funded by the ERA-Net BiodivERsA project DIARS (Detection of Invasive plant species and Assessment of their impact on ecosystem properties through Remote Sensing) through German Research Foundation (DFG) research grant FE 1331/3-1. This study was funded by the David and Lucille Packard Foundation and was supported by a scholarship for PhD students provided by the German Academic Exchange Service (DAAD). We thank N. Vaughn, C. Anderson, D. Knapp, R. Martin and the CAO team for airborne data collection and processing. The Carnegie Airborne Observatory has been made possible by grants and donations to G.P. Asner from the Avatar Alliance Foundation, Margaret A. Cargill Foundation, David and Lucile Packard Foundation, Gordon and Betty Moore Foundation, Grantham Foundation for the Protection of the Environment, W. M. Keck Foundation, John D. and Catherine T. MacArthur Foundation, Andrew Mellon Foundation, Mary Anne Nyburg Baker and G. Leonard Baker Jr., and William R. Hearst III.
Publisher Copyright:
© 2016 Elsevier B.V.
PY - 2017/1/1
Y1 - 2017/1/1
N2 - Most remote sensing approaches for mapping invasive plant species focus on species in a prominent phenological stage, such as during flowering, and do not systematically evaluate the performance for mapping lower cover fractions. In this study, we used airborne imaging spectroscopy (also known as hyperspectral imaging) to detect the invasive grass Phalaris aquatica and the invasive herb Centaurea solstitialis in a pre-flowering stage in the Jasper Ridge Biological Preserve, California, and compared the performance of three different one-class classifiers: Maxent, biased support vector machines and boosted regression trees. We collected presence data for C. solstitialis and P. aquatica to calibrate each approach and additional presence-absence data to validate model performance on 3 m × 3 m plots. The imaging spectroscopy data were acquired using the Carnegie Airborne Observatory Visible-to-Shortwave Infrared (VSWIR) imaging spectrometer (400–2500 nm range) with a pixel size of 1 m × 1 m. The resulting overall accuracies were 72–74% for C. solstitialis, and 83–88% for P. aquatica. For both species, the overall performance was slightly better for Maxent and BRT than for biased SVM. The detection rates for low cover plots were considerably higher for C. solstitialis than for P. aquatica. The models relied on different areas of the reflectance spectrum, but still produced the same general pattern of predicted species occurrences. We conclude that the different one-class classifiers allow for the detection and monitoring of target species with similar success rates.
AB - Most remote sensing approaches for mapping invasive plant species focus on species in a prominent phenological stage, such as during flowering, and do not systematically evaluate the performance for mapping lower cover fractions. In this study, we used airborne imaging spectroscopy (also known as hyperspectral imaging) to detect the invasive grass Phalaris aquatica and the invasive herb Centaurea solstitialis in a pre-flowering stage in the Jasper Ridge Biological Preserve, California, and compared the performance of three different one-class classifiers: Maxent, biased support vector machines and boosted regression trees. We collected presence data for C. solstitialis and P. aquatica to calibrate each approach and additional presence-absence data to validate model performance on 3 m × 3 m plots. The imaging spectroscopy data were acquired using the Carnegie Airborne Observatory Visible-to-Shortwave Infrared (VSWIR) imaging spectrometer (400–2500 nm range) with a pixel size of 1 m × 1 m. The resulting overall accuracies were 72–74% for C. solstitialis, and 83–88% for P. aquatica. For both species, the overall performance was slightly better for Maxent and BRT than for biased SVM. The detection rates for low cover plots were considerably higher for C. solstitialis than for P. aquatica. The models relied on different areas of the reflectance spectrum, but still produced the same general pattern of predicted species occurrences. We conclude that the different one-class classifiers allow for the detection and monitoring of target species with similar success rates.
KW - Biased support vector machine
KW - Boosted regression trees
KW - Carnegie airborne observatory
KW - Centaurea solstitialis
KW - Hyperspectral
KW - Imaging spectroscopy
KW - Maxent
KW - Phalaris aquatica
KW - Species distribution model
KW - Vegetation
KW - Yellow starthistle
UR - http://www.scopus.com/inward/record.url?scp=85007178913&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85007178913&partnerID=8YFLogxK
U2 - 10.1016/j.ecoinf.2016.11.005
DO - 10.1016/j.ecoinf.2016.11.005
M3 - Article
AN - SCOPUS:85007178913
SN - 1574-9541
VL - 37
SP - 66
EP - 76
JO - Ecological Informatics
JF - Ecological Informatics
ER -