2.2. Automated Detection of Surface Structures
The automated detection of surface structures using the RAMP AMM-1 SAR image mosaic is based on an edge detection method (sigma-on-mu) described by Septhon et al
] and Young et al
]. The floating ice masses (shelves and tongues) were extracted automatically using the MOA grounding line [23
] and the coastline of the RAMP AMM-1 project ( http://bprc.osu.edu/rsl/radarsat/data
). We identified 183 ice shelves and glacier tongues with areas between 20 and about 490,000 km2
for the Ross Ice Shelf. Ice shelves and ice tongues smaller than 20 km2
were neglected in this study (comprising 0.16% of total ice-shelf area), because a reliable feature detection on these small areas was not possible. A detailed look at the radar image mosaic shows sharp drops of the radar intensity at the boundary between adjacent SAR scenes at several sites, where the mean backscattering coefficients differ by up to 2 dB. These transitions might be caused by changing environmental conditions during the acquisition period, as a small amount of liquid water in the upper layer increases the absorption of the radar signal significantly, which in turn decreases the backscattering coefficient [24
]. However, European Centre for Medium-Range Weather Forecasts (ECMWF) daily temperature data ( http://data-portal.ecmwf.int/data/d/interim_daily/
, 2-m above surface) for September and October 1997 do not indicate melting conditions on any ice shelf. Another parameter directly influencing the backscatter intensity is the incidence angle of the radar signal. Wesche and Dierking [25
] showed that an incidence angle varying in the range between 20 to 40° has an effect on the backscattering intensity of icebergs with possible drops of 3–6 dB (see [25
]). In the data collection used for this study, radar incidence angles on the ice shelves vary between 22° and 31°. In the absence of any temperature effects we therefore assume that the sharp intensity transitions are caused by the different incidence angles.
Surface structures in the SAR images are recognized as rapid changes of the backscattering coefficient within a short distance (neighboring image pixels). Thus, the detectability of surface structures is influenced by the pixel size (in our case 100 × 100 m) and by the presence of speckle. In general, only structures which are at least as wide as the pixel size or which reveal a large intensity contrast to the adjacent undisturbed surface areas can be recognized. To reduce the influence of speckle on the detection result, we applied an enhanced Lee filter [26
] with a kernel of 3 × 3 pixels. After filtering the speckle is reduced, while the texture of the images is preserved [26
]. We employed mean (μ
) and standard deviation (σ
) of the backscattering intensity to separate homogeneous from inhomogeneous areas in the image. Both parameters were determined within a moving 3 × 3 pixel window. Using the resulting images of the mean and standard deviation, the sigma-on-mu
values were calculated pixel-by-pixel. Large
-ratios indicate areas of rapidly changing backscattering coefficients, which are then interpreted as edges.
To fix the thresholds of the
-ratios for automatically separating ice structures from undisturbed surface areas, we generated a reference data set by manually placing regions of interests (ROIs) on the boundaries of selected visible surface structures. These structures appear in most cases as linear patterns. The thresholds are obtained by comparing the
-ratios of the structures (obtained from the ROIs) and of the undisturbed surface areas in between. To consider the effect of the radar incidence angle,
-thresholds were determined separately for different incidence angle ranges.
By investigating the radar intensity variations along the coastlines in detail, we found that the backscattering contrast between surface features and undisturbed ice varied significantly, independent of the incidence angle. Hence we defined three thresholds using the median (second quantile), the third quantile and the ninths percentile of the cumulative frequency distributions of the
-values obtained from the ROIs. The surface structure detection results obtained with all three thresholds were then visually inspected. The median threshold showed optimal performance in areas with high contrasts, the ninth percentile in areas with low contrasts and the third quantile was used in areas of moderate intensity contrasts. On 42 small ice shelves, with areas between 20 and 327 km2
, the contrasts were too low to identify any structures. The detection results for the other ice shelves were converted to polygon style features indicating the location of a surface structure such as crevasse, rift or surface depression. The overall performance of the detection was tested by using a reference data set generated by manual identification of all surface structures on the Shackleton Ice Shelf as test site. As a quality criterion for the automated structure identification we compared the total length of the automatically detected linear structures (2,797 km) to the total length of the visually delineated structures (4,372 km). The result is a mean detection performance of 64 % for the test site (Figure 1
The result of the detection performance is very similar to the ones presented by [27
] and [22
]. Both presented iceberg detection algorithms using thresholding and the sigma-on-mu, respectively, with variable detection performances [22
]. Our detection performance can be traced back to the application of the enhanced Lee filter, which causes the radar intensity variations in the SAR scene to become smoother and the contrast between homogeneous and heterogeneous areas to become lower. Some surface structures appeared blurred after filtering and were therefore invisible for the our method.
To minimize false detections (Figure 1b
), we removed all identified structures that consisted of less than 10 pixels or 0.1 km2
. Some automatically detected structures around islands and close to the grounding line, where closely spaced structures were merged to one large patch in the radar image, could not be identified by visual inspection of the Radarsat-1 mosaic. High resolution Landsat-7 imagery of a part of the Shackleton Ice Shelf (not shown) resolves the dense pattern of surface structures. This means that not all features that do not match the visually determined reference data set are false detections.
Considering the problems with the performance of the automated detection, we used the result as starting point for a thorough visual identification of surface structures along the entire ice shelf margins within a 15 km wide strip inland. Although the visual detection was rather time-consuming, it was nevertheless important for generating a reliable data base for further analyses. Taking into account that the size of calving icebergs is limited by the crevasses spacing at the surface close to the front of the parent ice shelf [18
], we fixed the width of the strip for inspection to 15 km from the calving front as a compromise between the width of the largest iceberg observed by remote sensing techniques (B15: 37 km [28
]) and the extent of smaller ice shelves.