Focal species: Buffel grass (C. ciliaris L. P. ciliare)
Buffel grass is a perennial, summer-growing (C4) African bunch grass (Sharif-Zadeh and Murdoch 2001). It reproduces via seed and rhizomes and, as a result, can be seen in the landscape as both lone tussocks and dense monocultures. It does not drop its leaves; they accumulate at the base of the plant, often forming a ring of dry foliage around the tussock. The grass is spread by wind, water and traffic. In arid environments of Australia, where this study is based, it is typically found at highest density in riparian environments, depressions, and wherever soils are disturbed, including roadsides, construction sites and fire beds (Marshall et al. 2012). The plant responds rapidly to rain and often emerges before native grasses. It is also quick to dry-off and burn. The window for image capture of growing plants is brief; in Australia, we consider it is usually restricted to about a month after the first summer rains.
Located in the remote far north-west corner of South Australia, the study site occupies 15 × 12 km of the aboriginal owned Anangu Pitjantjatjara Yankunytjatjara (APY) lands. The site encompasses two indigenous communities—Kalka (26° 7′11.50″S, 129° 8′59.04″E) and Pipalyatjara (26° 9′37.45″S, 129°10′20.64″E) (Fig. 1)—with a combined population of less than 350. Climate is arid, with hot summers, mild winters and annual rainfall below 300 mm. Elevation ranges from 650 to 900 m. Plains comprise alluvial and fluvial sediments, vegetated by Aristida grasslands, sparsely distributed low shrubs and Hakea trees. These grasslands are increasingly dominated by Buffel grass. The Tomkinson ranges (Fig. 1) comprise mafic rock dominated by Spinifex hummock grasses; ranges in the north-west of the study site (Fig. 1) comprise felsic rock dominated by Enneapogon sp. grasses.
Buffel grass was introduced by direct seeding around Kalka in October 1987, along with Cenchrus setigerus and native drought-tolerant shrubs, Atriplex nummularia, Acacia kempeana and Acacia ligulata to combat dust storms on the alluvial flats; dust became a problem after an uncontrolled wildfire burnt a substantial area near the settlements, drought followed, and vegetation never regenerated. As a result of the direct seeding in 1987, this region is now largely dominated by Buffel grass.
Colour digital photography and four-band (visible to NIR) multispectral images were obtained over the study area. Image specifications are given in Table 1. The imagery was acquired on 14 February 2012 between 1134 and 1430 hours. Multispectral imagery was flown after the aerial photography from 1352 hours in the afternoon; consequently, shadow effects vary between the images. Conditions at the time of image capture were slightly hazy with less than 1 % high cirrus cloud cover. Buffel grass was approximately 50 % dried off on the day of image capture.
The aerial photography was acquired for a grid of 3 × 3 transects across the study site (north–south transects approx. 17 km; east–west transects approx. 12 km; spaced 5 km apart) (Fig. 1). Transects were positioned to capture the diversity of vegetation and geological settings while avoiding high elevations that are potentially dangerous for aerial navigation (Fig. 1). Photography was received as 930 un-georeferenced frames, in TIFF format. To save time georeferencing these frames individually, three-five frames were stitched together in the automated image matching program, Microsoft image composite editor (ICE). Image frames were exported from ICE as JPEG files, georeferenced in ArcGIS and saved as raster data using the minimum cell size for the image. These raster files were used for all subsequent analyses.
The four-band imagery, collected using the Spec Terra multispectral sensor, was acquired for three smaller areas, in highly diverse local environments, and overlapping the aerial photography flight paths (Fig. 2). The multispectral data was delivered corrected for radiometric and geometric artefacts, as orthorectified and georegistered mosaics in TIFF format. All image analysis was carried out in the 1994 Geocentric Datum of Australia, projected to UTM zone 52.
Ground validation sites
Field work was conducted from 7 to 12 February 2012. Selection of sites for ground validation was governed by in situ interpretation of environmental units, such as vegetation structure, soil colour and land use, aided by a 2007 ALOS colour mosaic of the region (2.5 m GSD). The goal was to represent the diversity of landscapes in which Buffel grass was present or absent and at varying densities. In total, 95 field sites were documented. Within these circular sites (10 m in diameter), projected cover (the vertical projection of plant foliage onto a horizontal surface) was estimated for Buffel grass and land cover units categorised as “herbs and forbs”, “other grasses”, “woody”, “leaf litter” and “soil”. The cover for each cover type was recorded as discrete classes: absent, 0 %; low, 0–25 %; moderate–low, 25–55 %; moderate–high, 55–85 %; and high, 85–100 % (Fig. 3). The centre point of each ground validation site was recorded using a Garmin eTrex High Sensitivity hand-held global positioning system receiver, which achieved a spatial accuracy of approximately 2–5 m.
For remote sensing analysis, the ground validation sites were co-registered to the aerial photography and separately to the multispectral imagery using the GPS coordinates recorded in the field and personal knowledge on the site. Of the 95 sites, 18 lay outside of the imagery coverage. A further 41 were not used because of image quality (which diminished over hilly terrain), obstruction from trees or insufficient geographical information to accurately place the site. Of the remaining sites, 43 lay within the coverage of the multispectral image scenes. Ultimately, a total of 53 and 43 sites were used for interpretation and classification of the aerial colour photography and four-band multispectral imagery, respectively.
Aerial photography image classification
We evaluated three commonly used image classification techniques for discriminating and quantifying Buffel grass in the aerial photography including visual cover estimates, manual digitisation, and a pixel-based unsupervised classification. Classifications were run separately for each field site (53 sites × 3 approaches).
Visual cover estimates for each ground validation site were scored using the same cover classes employed in the field survey. For consistency, sites were viewed at a scale of 1:125 to make the estimates. Visual standards (Fig. 4) also aided in making observations consistent.
For the manual digitisation method, individual Buffel grass plants or clumps within the 10-m diameter circular plots were digitised from the imagery, at a display scale of 1:125 m. The digitiser did not alter the viewing scale in order to more precisely circle plants. The total digitised area of Buffel grass for each site was then tabulated.
For the pixel-based assessment, an unsupervised classification was performed on the imagery at each site. A circular area of a diameter of 30 m, centred on the ground validation site, was used to run the classification. This accounted for the possibility to have a “Buffel grass” class even if Buffel grass was not present within the more tightly prescribed sample site. The classification was performed using the Iso Cluster Unsupervised Classification tool in ArcGIS 10 Spatial Analyst. The number of classes was set to 20; classes most representative of Buffel grass were then manually aggregated on the basis of visual examination. The aggregation process is producer-directed, allowing for some flexibility in the number of classes selected as representative of Buffel grass. The total area classified as Buffel grass for each site was then tabulated.
Four-band imagery classification
To exploit the additional spectral information in the multispectral imagery, the NDVI was applied. In the desert environment in which our study is situated, Buffel grass is often the “greenest” vegetation in the understorey during the summer months (December–February). Hence, the NDVI is a suitable index to identify cover type. The NDVI output was visually compared with the higher-resolution aerial photography to identify an NDVI threshold that best represented Buffel grass cover. The total area classified as Buffel grass was then calculated for each site.
To explore differences between each of the cover estimation approaches (visual cover estimates, manual digitisation, unsupervised classification, NDVI thresholds), classification results for selected sites were viewed concurrently, and disparities were described. The effectiveness of each approach in quantifying Buffel grass cover was then examined using regression analyses. The four image classifiers were compared not only with the field-based estimates, but with each other—to compare like with like. This is important, because whilst Buffel grass presence–absence is best interpreted from field results, field cover estimates are also subjective, and not necessarily more correct than the image-based estimates. The strength of each relationship was interpreted using Pearson's r-squared.