As the accuracy and sensitivity of remote-sensing
satellites improve, there is an increasing demand for more accurate and
updated base datasets for surveying and monitoring. However,
differentiating rock outcrop from snow and ice is a particular problem in
Antarctica, where extensive cloud cover and widespread shaded regions lead to
classification errors. The existing rock outcrop dataset has significant
georeferencing issues as well as overestimation and generalisation of rock
exposure areas. The most commonly used method for automated rock and snow
differentiation, the normalised difference snow index (NDSI), has difficulty
differentiating rock and snow in Antarctica due to misclassification of
shaded pixels and is not able to differentiate illuminated rock from clouds.
This study presents a new method for identifying rock exposures using
Landsat 8 data. This is the first automated methodology for snow and rock
differentiation that excludes areas of snow (both illuminated and shaded),
clouds and liquid water whilst identifying both sunlit and shaded rock,
achieving higher and more consistent accuracies than alternative data and
methods such as the NDSI. The new methodology has been applied to the whole
Antarctic continent (north of 82
Differentiating areas of snow and exposed rock in Antarctica is important in
a variety of contexts, including mapping; navigation; glaciological,
geological and geomorphological research; and monitoring changes in the ice
sheet and its response to climate change. The only existing continent-wide
geospatial dataset for exposed rock in Antarctica is available from the
Scientific Committee on Antarctic Research (SCAR) Antarctic Digital Database
(ADD) website,
Example of the issues with the existing ADD rock outcrop dataset showing the problems with the georeferencing, overestimation and generalisation of areas of rock outcrop. The example uses a false-colour image using the band combination red: SWIR2; green: blue; and blue: blue. This combination accentuates rock, snow and cloud distinctions, with red/pink pixels representing rock or clouds and turquoise pixels representing snow. The Landsat 8 scene used is LC82081132013343LGN00.
In temperate regions methods have been formulated to automatically identify
exposed rock outcrop from satellite imagery (e.g. Racoviteanu et al., 2010;
Dozier, 1989; Hall et al., 1995; Paul et al., 2002, 2009; Bolch et al., 2010;
Zhu and Woodcock, 2012; Zhu et al., 2015), but the methods have never been
applied to Antarctica. The most commonly used existing method for delineating
snow cover and rock outcrop is the normalised difference snow index (NDSI;
Hall et al., 1995; Dozier, 1989). The NDSI was developed following other
indices, such as the normalised difference vegetation index (NDVI; Tucker
1986, 1979), initially for application to MODIS and Landsat satellite
imagery. The NDSI is calculated according to Eq. (1) (modified for Landsat 8
data) where Landsat 8 OLI (the Landsat 8 Operational Land Imager sensor) band
3 equates to spectral wavelengths of 0.53 to 0.59
Equation (1) works on the basis that snow reflects visible wavelengths stronger than middle-infrared wavelengths whilst rock displays a slightly higher reflectance for middle-infrared wavelengths than visible wavelengths (Fig. 2) and so a threshold value can be determined for the NDSI of an image differentiating pixels of snow and rock (typically in the range 0.25 to 0.45; Hall et al., 1995). One problem for application of the thresholded NDSI technique to automated snow and rock differentiation is that the optimal threshold value must be determined for each individual image being analysed or even varied within the same image due to changes in illumination or fresh snow cover across the image's area (Burns and Nolin, 2014). It is often the case that the optimal threshold is manually determined on each scene by comparison to reference data; however this becomes a problem when large numbers of images need to be analysed or reference data are not available.
Although the application of the NDSI has been successful at lower latitudes (e.g. Burns and Nolin, 2014) where vertically illuminated imagery is available, high solar elevation angles in Antarctica lead to exclusion of shaded rock. This issue of shaded rock is greater in Antarctica, where unavoidably low solar elevation angles result in large percentages of the outcrop being in the shade. The problem has been addressed for glacier mapping at lower latitudes by thresholding the Landsat blue band (in addition to an NDSI or alternative band ratio threshold) due to the higher reflectance of shaded snow than shaded rock in blue wavelengths (Arendt et al., 2012; Bishop et al., 2004; Paul et al., 2007; Paul and Kääb, 2005).
Spectral reflectance data for snow and rock (granite, basalt and
sandstone) from the ASTER Spectral Library v1.2 (Baldridge et al., 2009).
Designations of spectral regions as defined by the Landsat 8 bands: blue –
band 2, 0.45–0.51
Illustration of the misclassification of cloud cover as rock pixels when using the NDSI technique. As in Fig. 1, the example uses a false-colour image using the band combination red: SWIR2; green: blue; and blue: blue. An NDSI threshold of 0.6 is used here to identify the rock outcrops, but at this threshold much of the cloud cover is also included. The Landsat 8 scene used is LC82161092014338LGN00.
Comparison of debris cover for glaciers at low latitudes:
Unavoidable cloud cover in some Antarctic images, especially on the Antarctic Peninsula, leads to the classification of clouds as rock exposure by the NDSI technique (Fig. 3) as the two are indiscernible using this methodology. Any effective dataset of rock outcrop in Antarctica would have to ensure that clouds are not misrepresented.
A further problem for automated rock identification at lower latitudes is debris cover on glaciers which is indiscernible in multispectral imagery from exposed rock (Paul et al., 2004). This is accentuated by the melting and ablation of low-latitude glaciers (Stokes et al., 2007) and is intensified by the large amount of debris from frost shattering and freeze–thaw activity (Fig. 4a and b). However, Antarctic glaciers are rarely debris-covered due the prevailing climatic conditions where constant sub-freezing conditions result in a lack of ablation (Fig. 4c and d). The limited number of positive-degree days and the lack of a day–night cycle at polar latitudes reduces freeze–thaw activity, meaning that less frost shattering takes place. Most Antarctic glaciers and ice streams are marine-terminating, and relatively few have active ablation zones (with the exception of a small percentage on the northern and eastern Antarctic Peninsula). The result is that most Antarctic glaciers are largely debris-free, removing this limitation from our study.
Here we present a new technique for automated rock outcrop identification using freely available Landsat 8 satellite data. The method is a composite technique combining separate algorithms that divide the image into cloud, liquid water, shaded snow and sunlit snow, and shaded and sunlit rock exposures. We test the method against manually digitised polygons, the existing ADD rock outcrop dataset and the NDSI to validate and compare its accuracy.
We apply the new methodology to the entire landmass of Antarctica
(
To produce a rock outcrop map for the entire Antarctic continent requires a
freely available georeferenced multiband dataset. The dataset must cover high
latitudes; be recently acquired; be of a high enough resolution to identify
individual outcrops and geomorphological features; and be divided into
sufficiently large scenes to allow for manual selection of suitable tiles for
the entire continent. On this basis, the Landsat 8 multispectral satellite
data were chosen for analysis. Landsat 8 is the latest and continuing
satellite mission for multispectral global data acquisition launched by NASA
and the United States Geological Survey (Roy et al., 2014). The satellite's
Operational Land Imager (OLI) sensor records eight electromagnetic bands
(0.43–2.29
For the production of an Antarctica-wide rock outcrop map, tiles were selected
that display strong illumination and minimal cloud cover. To ensure strong
illumination, we only used images taken during the day in the austral summer
between September and March, with all but 17 images having solar elevation
angles
In addition to the raw data, pre-processed tiles (170 km north–south by
183 km east–west) corrected for top-of-atmosphere (TOA) reflectance, surface
reflectance and brightness temperature are freely available for download
(
The new methodology identifies areas of sunlit and shaded rock through two
separate workflows and then merges both outputs to produce the final dataset.
Within both procedures a series of masks are produced to identify areas of
exposed outcrop and to exclude areas of snow, cloud and liquid water. At each
stage band ratios were used in preference to threshold values for individual
bands to allow application of a single set of threshold values to a large
dataset. These two procedures are detailed below, and a flowchart for
executing this process is shown in Fig. 5. The complete methodology was
automated within ArcPy (Zandbergen, 2013). The script is available from
GitHub (
Flowchart for the automated identification of rock outcrops in Antarctica. Threshold values are given without the 16 bit scaling used in the corrected Landsat 8 raster images.
Threshold values used in the methodology were determined by manually classifying 8741 pixels from three different Landsat 8 images from the Antarctic Peninsula of different latitudes, geology, illumination and cloud cover (images LC82081132013343LGN00, LC82191052013340LGN00 and LC82201072015017LGN00). Pixels were classified as representing “clouds”, “sea”, “sunlit rock”, “shaded rock”, “sunlit snow” or “shaded snow”. Pixel values were extracted for the spectral bands of interest to determine the spectral properties of these six land cover classes (Fig. 6), with thresholds being set that best distinguished them.
Although the NDSI is unable to identify shaded rock and often misclassifies
clouds as rock outcrop, it remains the best method for identifying regions of
exposed sunlit rock. Consequently, it is the primary input for this
methodology, with a threshold value of
One of the main problems of rock outcrop identification in Antarctica is that
sunlit rock and clouds are indiscernible using the NDSI alone (Fig. 6a).
Consequently we have derived a mask for sunlit snow and clouds using the
thermal infrared band (Landsat 8 TIRS1, 10.60 to 11.19
Box plots of extracted pixel values from three Landsat 8 tiles
illustrating the different spectral properties of clouds (number of extracted
pixels,
The most widely applied approach for the identification of liquid water in
multispectral imagery is the normalised difference water index (NDWI;
McFeeters, 1996). Modified for Landsat 8 data with the Landsat 8 OLI band 3
equating to spectral wavelengths of 0.53–0.59
Even in the shade, snow is more reflective at blue wavelengths than shaded
rock. By comparing the blue reflectance values of pixels representing rock
and snow, a threshold reflectance value of
Although a blue wavelength threshold successfully differentiates shaded snow
and rock, liquid water is also misclassified as rock. Thus, the NDWI and
coastline mask applied to the sunlit rock data are also applied to the
shaded rock data (again using the NDWI threshold value of
Pixels that were identified as rock by the NDSI mask and not identified as
cloud or water represent sunlit rock outcrops. Similarly, pixels with blue
band intensities below the threshold for shaded rock that are not
subsequently identified as liquid water by the NDWI threshold represent
shaded rock exposures. Merging these two outputs produced the rock outcrop
map for each tile. Tiles not already projected with the WGS 1984
Stereographic South Pole spatial reference, EPSG 3031 (i.e. those at scenes
with a centre latitude greater than or equal to
As most areas were covered by multiple overlapping Landsat tiles, any pixels identified as rock exposure by any of the overlying tiles were included as exposed rock in the final dataset. This was achieved by mosaicking the binary raster files produced by the workflow and taking the maximum pixel value. If a pixel was classified as snow, it was designated “0” by the script, or “1” if it represents rock. Consequently this mosaicking process stores rock outcrop pixels (“1”) in the raster mosaic in preference to snow (“0”). By analysing multiple overlapping tiles, the methodology becomes more sensitive to identifying rock outcrops; allows detection of rock outcrops even when they are obscured by clouds in one tile of the input data; and makes the methodology less sensitive to seasonal or short-term variation in snow cover.
Finally, the extent of the mosaicked raster dataset was converted into a new polygon shapefile and merged with the existing ADD rock outcrop dataset for areas not covered by the Landsat 8 imagery (Fig. 7).
Rock exposure map of Antarctica showing the data sources for the new
dataset. Outcrops shown in red were derived using the new remote-sensing
methodology, and outcrops in blue were derived from the existing ADD rock
outcrop dataset to supplement areas not covered by the Landsat 8 imagery
(areas south of 82
To quantify the accuracy of the new methodology and its limitations, the
extent of rock exposure was manually delineated using ten
Images used for the quality assessment overlain by the three alternative methodologies and datasets: pixels extracted using optimum NDSI thresholds for each image (NDSI threshold values shown in brackets); pixels extracted using the new methodology presented here; and the extents of the current ADD rock outcrop map. Enlargements of these images can be downloaded from the Supplement. Scene locations are indicated in Fig. 9.
Locations of the 249 Landsat 8 tiles (blue squares) used to identify
rock outcrop in Antarctica and the locations
Summary of mean accuracy assessment vales for the 10 images evaluated.
The manually derived land cover was compared with the existing ADD rock outcrop dataset, the new automated method and the optimum NDSI-determined output for each image. Optimum NDSI threshold values (maximum values for pixels identified as rock) were taken as those with the lowest total quantity disagreement (abundance accuracy) and allocation disagreement (location accuracy) (Pontius Jr. and Millones, 2011). As shown by Fig. 10, optimum NDSI threshold values are highly variable. For well-illuminated images without any cloud cover (Fig. 8a–e), NDSI threshold values of 0.6 or 0.7 are optimal. Images of extensive shade achieve more accurate results at higher NDSI threshold values (0.8, Fig. 8f), allowing identification of shaded rock. In contrast, images with extensive thick cloud require lower values (0.3 to 0.5, Fig. 8g and h) so as not to include the cloud as misidentified rock outcrop pixels. Thinner, low clouds (Fig. 8i) are not so problematic and high values (0.7) remain optimal. For mixed images (Fig. 8j) with shaded and illuminated rock with minor cloud cover, 0.7 remained the optimal threshold value.
Well-illuminated, cloud-free images produce similar classification accuracies
(CA; Eq. 3) for the optimal NDSI technique and the new method (Fig. 8a–e) with
low commission or omission disagreements (Fig. 11a). However, the required
determination of an optimal NDSI threshold value renders this alternative
methodology more involved than that used for our new dataset. In addition,
even when using the optimal threshold value, the NDSI technique omits areas of
rock in shaded images as well as both shaded and sunlit rock in cloudy
images, leading to high and variable omission disagreements (Fig. 11b).
Total quantity and allocation disagreement values (Pontius Jr. and Millones, 2011) for pixels extracted from the images in Fig. 8 using the NDSI threshold technique.
The new methodology performed poorest in images with limited areas of rock outcrop (e.g. Fig. 8h, 0.1 % rock), although shade, clouds and mixed pixels of snow and rock in Fig. 8h make even manual pixel identification difficult. There are omission disagreements in shaded images (Fig. 8f and j), although these are much lower than for the alternative techniques (a mean of 15 % for all images compared to 38 % for the NDSI technique and 30 % for the ADD rock outcrop dataset, Fig. 11b). Clouds were successfully masked and do not contribute to the commission disagreement (Fig. 8h–j).
Mean statistics for the quality assessment are recorded in Table 1. The
quality assessment shows higher accuracies for the new method (a mean of
This is the first automated methodology for the differentiation of snow and
rock in Antarctica, from which a new outcrop map of the entire Antarctic
continent has been produced at higher and more consistent accuracies than
existing data and techniques (Fig. 11). The new dataset is available online
via the SCAR Antarctic Digital Database (
Despite the poorer accuracy of the ADD rock outcrop dataset (39 %
classification accuracy compared to 74 % for the new methodology,
Fig. 11a), due to the methodology by which it was derived, certain features
are better represented. This includes South Georgia and the South Orkney
Islands, where a lack of cloud-free imagery in the late austral summer (when
the outcrops are not covered by snow) prevents automated outcrop
identification. Consequently, rock outcrop extents in these areas are derived
from the existing ADD dataset rather than remote-sensing imagery, in addition
to outcrops south of 82
It is important when using the new Landsat 8 rock outcrop map to consider seasonal variability in snow cover and that most outcrops were derived from multiple tiles from different years and different months of the austral summer. As a result the map may not be representative of current conditions and may not consistently represent maximum outcrop extent across the continent.
Using the new methodology, we have produced a revised map of rock outcrops in
Antarctica. Landsat 8 does not provide coverage south of 82
Examples of the new methodology's limitations, comparing the new
dataset with false-colour Landsat 8 images. The band combination (red: SWIR2;
green: blue; blue: blue) is chosen to accentuate rock, snow and seawater
distinctions, with red pixels representing rock, turquoise pixels for snow and
dark green to black pixels for seawater.
Because an overlap exists between the NDWI values of shaded rock and
liquid water (Fig. 6d) and because of inaccuracies in the existing coastal
vector dataset, some pixels of coastal seawater not masked by the ADD
coastline have been misidentified as exposed rock in all coastal scenes
containing seawater pixels. This is particularly problematic for pixels
adjacent to seawater rich in calved ice and glacial debris (Fig. 12a). These
pixels are spectrally identical to shaded rock and thus cannot be excluded
automatically from the data. Consequently these pixels were manually removed
from the final dataset, with the distinction of shaded rock and liquid water
being made by eye. It should be noted that some of these misidentified
pixels may still be present. However, as no manual editing was done on land,
the repeatability of this methodology should not be affected. Even though spectral properties have been chosen that distinguish rock
pixels from those of snow, clouds or sea, some overlap exists where pixels
remain ambiguous (Fig. 6). Consequently, to allow automated analysis over
such a large area, mildly conservative threshold values were chosen. For
example, the NDSI threshold for sunlit rock was set at the 95th percentile
rather than the complete range exhibited by sunlit outcrops as this excludes
any overlap with the range of NDSI values for sunlit snow (Fig. 6a). This
results in the exclusion of some pixels of exposed rock that are spectrally
similar to clouds and snow (e.g. Fig. 12b). Due to the 100 m spatial resolution of the TIRS band, small outcrops
around the continent (especially those less than 60 m or 2 pixels across)
are often excluded by the new technique and may be better represented in the
ADD rock outcrop dataset. Whilst Antarctic glaciers rarely show any debris cover (Fig. 4), there
are local occurrences where extensive debris cover does occur (most notably
in the vicinity of the Dry Valleys and the NW coast of the Ross Ice Shelf)
which are mapped as outcrop in the new dataset. However, it should be noted
that these occurrences are isolated on the continental scale. As this
project aims to provide a consistent and automated approach that can be
reproduced in the future (for example to monitor change in ice cover over
time or season), our methodology attempts to be as free as possible from
manual changes. We accept that in some areas localised occurrences of
debris-covered glaciers may need to be manually altered if detailed topographic
maps of rock outcrop are required.
We calculate (using the South Pole Lambert azimuthal equal-area projection)
that the existing ADD rock outcrop dataset has a 44 900 km
The new Landsat 8 rock outcrop map will provide a revised and accurate base dataset for future topographical, glaciological, geological and geomorphological mapping. A number of satellite programmes collecting new high-resolution colour images have recently been launched or are planned for launch in the near future, including the DigitalGlobe WorldView-3 satellite (launched 2014), NASA's HyspIRI satellite (proposed but not yet under development), European Space Agency's Sentinel program (three satellites already launched with more under development) and the continuing Landsat data acquisition (continuing acquisition from Landsat 7 and 8, with Landsat 9 planned for launch in 2023). These new datasets will allow further application of this technique at higher resolutions and consequently higher accuracies, allowing future improvement of the datasets' broader applications. Application of the new technique to these alternative datasets would however require modification of the threshold values for each mask in the procedure.
Once the available imagery has improved, the Antarctic rock outcrop dataset
will again be updated to exploit the new data and increase coverage of the
continent (especially south of 82
A new map of exposed rock outcrop has been developed for the Antarctic
continent. The new map was achieved via an automated methodology employing
Landsat 8 multispectral imagery. The new methodology uses the NDSI technique
to identify sunlit rock exposure and low blue intensities for shaded rock,
and then applies separate masks to remove incorrectly classified pixels of
cloud, snow and liquid water. This is the first automated methodology for
rock outcrop identification in Antarctica and achieves higher and more
consistent accuracies than the existing dataset or what can be achieved using
the alternative automated technique (the NDSI). Assessing the accuracy of
these alternative techniques and datasets across a range of images gives a
mean value for correct pixel identification of
The new map, supplemented by existing data for latitudes south of
82
The new Landsat 8 derived rock outcrop dataset is available in the Supplement
to this article; from the SCAR Antarctic Digital Database,
We would like to thank Allen Pope and our anonymous second reviewer for their positive, helpful and thorough reviews. This study is part of the British Antarctic Survey Polar Science for Planet Earth programme, funded by the Natural Environment Research Council (NERC). Martin Black was funded by a NERC research studentship (NE/K50094X/1). Edited by: K. Matsuoka Reviewed by: A. Pope and one anonymous referee