Mapping Invasive Plant Species with a Combination of Field and Remote Sensing Data

Friday, 18 December 2015
Poster Hall (Moscone South)
Sandra Skowronek1, Hannes Feilhauer1, Ruben Van De Kerchove2, Michael Ewald3, Raf Aerts4, Ben Somers4, Jens Warrie4, Pieter Kempeneers5, Jonathan Lenoir6, Olivier Honnay4, Gregory Paul Asner7, Sebastian Schmidtlein3, Tarek Hattab6 and Ducchio Rocchini8, (1)University of Erlangen-Nuremberg, Erlangen, Germany, (2)VITO, Mol, Belgium, (3)Karlsruhe Institute of Technology, Karlsruhe, Germany, (4)Katholieke Universiteit Leuven, Leuven, Belgium, (5)VITO, VITO, Mol, Belgium, (6)Jules Verne University of Picardie, Amiens, France, (7)Carnegie Institution for Science Washington, Washington, DC, United States, (8)Fondazione Edmund Mach and Innovation Center, San Michele all'Adige, Italy
Advanced hyperspectral and LIDAR data offer a great potential to map and monitor invasive plant species and their impact on ecosystems. These species are often difficult to detect over large areas with traditional mapping approaches. One challenge is the combination of the remote sensing data with the field data for calibration and validation. Therefore, our goals were to (1) develop an approach that allows to efficiently map species invasions based on presence-only data of the target species and remote sensing data; and (2) use this approach to create distribution maps for invasive plant species in two study areas in western Europe, which offer the basis for further analysis of the impact of invasions and to infer possible management options.

For this purpose, on the island of Sylt in Northern Germany, we collected vegetation data on 120 plots with a size of 3 m x 3 m with different cover fractions of two invasive plant species; the moss Campylopus introflexus and the shrub Rosa rugosa. In the forest of Compiègne in Northern France, we sampled a total of 50 plots with a size of 25 x 25 m, targeting the invasive tree Prunus serotina. In both study areas, independent validation datasets containing presence and absence points of the target species were collected. Airborne hyperspectral data (APEX), which were simultaneously acquired for both study areas in summer 2014, provided 285 spectral bands covering the visible, near infrared and short-wave infrared region with a pixel size of 1.8 and 3 m. First results showed that mapping using one-class classifiers is possible: For C. introflexus, AUC value was 0.89 and OAC 0.72, for R. rugosa., AUC was 0.93 and OAC 0.92. However, for both species, a few areas were mapped incorrectly. Possible explanations are the different appearances of the target species in different biotope types underrepresented in the calibration data, and a high cover of species with similar reflectance properties.