ICEBERG OBSERVATIONS IN THE BARENTS SEA BY RADAR

AND OPTICAL SATELLITE IMAGES

Stein Sandven(1), Mohamed Babiker (1) and Kjell Kloster(1),

(1)Nansen Environmental and Remote Sensing Center (NERSC), Thormøhlensgte 47, N-5006 Bergen, Norway. E-mail: stein.sandven@nersc.no

ABSTRACT

Studies of iceberg detection in the Franz Josef land region have been performed by use of optical images in combination with and Synthetic Aperture Radar images during April 2006. Alternating Polarization (AP) images (HH- and VV-pol) from ASAR and RADARSAT ScanSAR Narrow images were tested against Landsat images for identification of 15 icebergs of size between 50 and 400 m. The three types of SAR images had fairly similar capability to detect these icebergs, with the HH-pol from ASAR showing the most reliable results. The ScanSAR Narrow image failed to observe the smallest icebergs of size around 50 m. The icebergs were located in fastice with relatively homogeneous backscatter which was favourable for iceberg detection. The studies showed that combined use of optical and SAR images for iceberg detection gives better results compared to previous studies where SAR and optical images were used separately. 1. INTRODUCTION

In recent years the icebergs in the Arctic have become increasingly important because of changes in the cryosphere and the risk for offshore operations and sea transportation. The majority of the icebergs are small, typically 20 m (bergy bits and growlers), while tabular icebergs, glacier bergs and other types at scales of 100 m or more represent a minority (Abramov, 1996; Zubakin et al., 2004). An example of a tabular iceberg is shown in Fig. 1. Satellites can observe the largest icebergs in the Barents Sea, typically 100 m or more in horizontal extent, and under specific wind and sea ice conditions. Several studies have been conducted on detection of small-scale Arctic icebergs from satellite SAR data in eastern Canada (Power, et al., 2001), off the southern coasts of Greenland (e.g. Gill, 2001) and in the northern Barents Sea (Kloster and Spring, 1993; Knapskog, 1996; Sandven et al., 1999). These studies present various techniques and algorithms to detect objects in SAR images that are most likely icebergs drifting in partly ice-covered and partly open water conditions. Iceberg observations in high-resolution optical images are not hampered by the speckle noise that is characteristic for the SAR images. Observations in optical images are therefore more reliable for icebergs of size of 100 m or less. The limitation of optical images is cloud cover and darkness, allowing good quality observations only occasionally.

Figure 1. Photograph of a tabular iceberg of about 100 m in horizontal extent, embedded in sea ice. Note that

the icebergs and sea ice have different drift speed, producing open leads in the wake of the iceberg.

Photograph courtesy: A. Glazovsky. Iceberg signatures in SAR images are characterized by bright signature compared to a darker background. The possibility to detect icebergs therefore depends on the iceberg properties (size, height, shape) as well as by the background (sea ice or open water). SAR backscatter from icebergs are caused by two main mechanisms: (1) backscatter form the top surface(s), and (2) backscatter from ocean surface and the sidewalls of an iceberg in case there is a well-defined wall facing the radar beam (denoted “double-bouncing”). Icebergs are found in three different situations, each requires a specific strategy for observation: (1) Icebergs located in open water will appear as bright spots against dark background in both optical and SAR images. During high wind speed the contrast between open water and icebergs is reduced in the SAR images. (2) Icebergs in drifting ice create open leads due to the differential drift speed (Fig. 1). Icebergs can create long tracks in the drifting ice if the icebergs are grounded. It can be difficult to distinguish icebergs from the background both for optical and SAR images. (3) Icebergs in fastice near calving areas can be readily observed in high-resolution optical as well as in SAR images if image resolution is sufficient. Icebergs can be stationary in the fastice for a long time, making it possible to identify the same icebergs over several months. The present study will address the detection of icebergs in a fastice situation.

_____________________________________________________

Proc. ‘Envisat Symposium 2007’, Montreux, Switzerland 23–27 April 2007 (ESA SP-636, July 2007)

2. METHOD

In this study we have investigated iceberg detection capabilites using ENVISAT ASAR alternating polarisation images (AP), RADARSAT ScanSAR Narrow (SN) mode image and optical images from Landsat and Terra ASTER. The specification of the images is given in Table 1. The study took place in April 2006 in the southern part of the Franz Josef land archipelago. Table 1. Specification of images used for iceberg detection Satellite sensor Image mode Pixel

size Image size/

Swath width Landsat ETM+ Panchromatic 12.5 m 180 by 180 km Multichannel 25 m 180 by 180 km Terra ASTER Multichannel 15 m 60 by 60 km ENVISAT ASAR Wideswath 75 m 400 km Alternating

polarisation (HH- and VV polarisation)

12.5 m 100 km

RADARSAT ScanSAR Narrow mode

25 m 300 km

The study demonstrates how synergetic use of SAR and optical images can improve the observation of icebergs in the Barents Sea region. A selection of 15 icebergs of size from 50 to 400 m were identified in a subset of the Landsat panchromatic image (pixel size 15 m) covering the southern part of the Franz Josef Land archipelago (Fig. 2a). The subimage covered about 30 by 30 km and overlapped with the SAR AP image (Fig. 2b). It should be mentioned that several hundred icebergs could be identified in the Landsat image, but only a small subset (15 icebergs) was selected for further analysis. The iceberg objects in the Landsat image were then compared to similar objects in the SAR images. In optical images icebergs are identified as bright objects against a less bright background combined with a dark shadow due to the low sun angle and height of the iceberg. The brightness signature of the icebergs and the surrounding sea ice can be similar, so the key criterion is the dark shadow. In SAR images icebergs are also identified as bright objects, but detection capability depends on the backscatter of the surrounding sea ice or open water. In this example the background is level fast-ice with relative low backscatter, allowing detection of objects with bright signature. Shadows can in principle be generated in a SAR image, but images need higher resolution to show this effect. For each of the selected icebergs the mean backscatter of the icebergs was calculated and

a

b Figure 2. (a) Subset of a Landsat image from 14 April 2006 covering the area around Salm Island (the island in the middle of the image covered by a glacier) in the

southern Franz Josef Land. There are 15 icebergs (A to R) which are compared with the SAR HH- and VV- image (b) obtained over the same area on 12 April.

compared to the mean backscatter of the surrounding ice area. The backscatter for both HH- and VV-polarisation ASAR images and RADARSAT HH-polarisation image was quantified and compared for 15 icebergs selected from the Landsat image (Fig. 2a).

3. DATA ANALYSIS AND RESULTS

The size of the 15 icebergs was determined from the number of pixels representing the iceberg objects in the Landsat image. The length varied between 50 and 400m, while the width varied from 30 to 230 m. 30 m is the smallest length scale that could be determined from images with pixel size of 15 m (Fig. 3). The mean backscatter for each iceberg is compared to the mean background signal for the fastice surrounding the icebergs, as shown in Fig. 4. The fastice in the study area is generally smooth, undeformed ice with low backscatter from both HH- and VV-polarisation. The smoothness of the fastice around the icebergs is documented in the Landsat image where ridges and other surface features appear as structures in the image. An example of this is seen between point H and J in Fig. 2 a.

Figure 3. Horizontal size of the 15 icebergs selected form the Landsat image. Length: blue bars, width: red

bars. The HH image showed significant backscatter difference between iceberg and background for all the 15 cases, while VV-images failed to identify icebergs in two of the cases. The HH image had better signal-to-noise ratio for all 15 icebergs compared to the VV image (Fig. 4). When comparing VV and HH it should be noted that corresponding pixels in the two images are not exactly in the same position, due to the alternating HH and VV pulses. It was also noted that the maximum backscatter representing iceberg signature in several cases came from different pixels in HH and VV. One should therefore be cautious in comparing objects of only a few pixels in size in alternating polarisation images. The main result is that the mean backscatter of the icebergs compared to the background is generally higher for HH-polarisation compared to VV-polarisation. The HH backscatter is more stable, while the VV backscatter is more variable. For VV two icebergs could not be identified (M, N), while two others had a weak signal similar to the background (D, P).

a

b

Figure 4. Average backscatter for the icebergs (blue bars) compared to the surrounding background ice (red bars). The digital values are taken form the AP

images provided by ESA on CD-ROM. (a) shows backscatter for HH-polarization, while (b) shows

backscatter for the VV-polarisation. A test of a RADARSAT ScanSAR Narrow image with pixel size of 25 m and HH polarisation was also done for detection of the 15 selected icebergs. The backscatter from the 15 icebergs was compared with the background ice. To compare digital numbers from two different SAR systems, we calculated sigma-0 for both data sets, and presented sigma-0 for icebergs as well as for the background fastice. The results are shown in Fig. 5, where sigma-0 is presented in dB. Note that sigma-0 level is different between the two satellite images because the incidence angle is different. The lower backscatter level of the icebergs in the RADARSAT image (about -12 dB) compared to the ENVISAT image (about – 5 dB) is mainly due to the higher incidence angle of the RADARSAT data. The same situation is found for the background fastice.

a

b Figure 5. (a) SAR backscatter (sigma-0) for the iceberg

objects (dark blue dots) and the surrounding fastice (pink dots) for RADARSAT ScanSAR HH-polarisation and (b) for ENVISAT HH-polarisation. The sigma-0 values are estimated for higher incidence angle for

RADARSAT compared to ENVISAT. Also the pixel size is different: 25 m for RADARSAT and 12.5 m for

ENVISAT APP images. Both SAR systems provide good contrast between iceberg signatures and the background, which is a key criterion for detection. An exception is iceberg M, N, O and P which are detected only in the ENVISAT and not in the RADARSAT image. The pixel size of the RADARSAT image is twice that of ENVISAT (25 m versus 12.5 m), but size is not the main criteria for detection in SAR images. Icebergs M, N, O and P have size of about 50 m, but could not be detected in the RADARSAT image. Other icebergs with size of 50 m, such as D and I, were well detected in the RADARSAT image. Detection capability in SAR images depends on geometry and orientation of reflecting planes of the target. The impact of iceberg shape and geometry on SAR backscatter could not be directly investigated because we did not have data on shape and size of the 15 icebergs. A first and simple approach to investigate the effect of iceberg geometry was to classify the icebergs into “simple and “complex” geometry based on

information in the Landsat image. Icebergs were classified as “simple” if they appeared as single objects with bright signature and a corresponding object with dark signature representing the shadow. Examples of simple icebergs are F and P in Fig. 6a. Icebergs were classified as complex if the object had several backscatter maxima or complex shadow pattern. Complex icebergs can represent clusters of icebergs grouped together or single icebergs with complex topography producing irregular shadows, as shown in A and B of Fig. 6a. The backscatter of the simple and complex icebergs were compared by taking the difference between the mean backscatter for HH and VV polarisation. The results show that for the simple icebergs HH had higher backscatter than VV, while for the complex icebergs VV was higher than HH (Fig. 6b). Only one exception was found: iceberg I was classified as simple but VV was higher than HH. This is an interesting result, but it raises new questions and we cannot draw any conclusions at this stage. It is clear that more quantitative data on shape and geometry of icebergs is needed in order to study the effect on different SAR polarisations. Such study is needed to prepare for use of SAR in iceberg monitoring systems. New SAR satellites (RADARSAT-2, TerraSAR-X, Sentinel-1, etc.) will all have possibilities to select different polarisations (HH, VV, HV, VH), and for iceberg monitoring we need to know which polarisation is best suited for the task. Both HH- and VV-image showed many bright spots that were not identified as any object in the Landsat image. An example of this is shown in Fig. 7. The signatures of icebergs M, N, O and P were compared between the Landsat and the ENVISAT HH-polarisation image. M represents two very similar objects within the circle in both images. P and O represent well-defined icebergs in the centre of the circles in the Landsat image which can also be identified as bright spots in the SAR image. There are also other bright spots in the neighbourhood which cannot be recognised as icebergs. N was only observed in the ENVISAT image and not in the Landsat image. N is therefore not included in the analysis of the 15 icebergs. N is assumed to be speckle noise producing many bright spots which are similar to the iceberg signature. The two subsets from Landsat and ENVISAT HH-polarisation image presented in Fig. 7 demonstrate very well that many bright spots in the SAR are not associated with icebergs. This has very severe impact on the use of SAR data for iceberg detection. Even if SAR images with pixels size of about 10 m can detect icebergs of size 50 m and more, the problem is the number of false detections.

a

b

Figure 6. (a) Landsat sub-images showing examples of complex (A, B) and simple (F , P) iceberg geometry;

(b) difference between HH- and VV-polarisation (digital values from ESA) for each iceberg and

annotation of simple (s) and complex (c) geometry of the icebergs.

Figure 7. Subsets of Landsat image (left) and ASAR APP H-polarisation image (right) in the area where

icebergs M, O and P were identified. Note that N was not observed in the Landsat image, but analysis of a nearby feature was done for the SAR images. These

icebergs were not identified in the RADARSAT image. There are many more icebergs that can be identified in the Landsat image as well as in the SAR image.

However, many of the bright points in the SAR image are not icebergs but speckle noise.

This study has only compared SAR and optical data for iceberg detection in fastice. Detection of icebergs in open water and in drifting ice has not been investigated as far as SAR data is concerned. More specific studies are needed to find solutions for more reliable iceberg detection in SAR images. 4. CONCLUSIONS

To improve iceberg detection in SAR images, it is necessary to filter out as much as possible of the high-frequency speckle noise as part of the image analysis, but this means that iceberg objects are also filtered out. A solution is to use SAR images with higher resolution initially, allowing speckle filtering to be done without loosing too much iceberg information. It is clear that iceberg geometry needs to be better quantified in order to understand the detection capability for the two polarisations. This is ongoing work where iceberg data from ship and aircraft are used to compare with satellite data. Since future SAR systems will have different polarisation options, it is important to determine which SAR modes should be used for iceberg detection. An optimal solution for iceberg detection is to use a synergy of optical and SAR images, both with resolution of 10 m or better. A daily acquisition scheme should be used to ensure that data are captured under favorable conditions, which means no-clouds for optical images and low-wind for SAR images. 5. ACKNOWLEDGEMENT The studies have been supported by the ESA AO project no. 1260 and Hydro Oil and Energy. 6. REFERENCES Abramov, V. Atlas of Arctic Icebergs. Backbone

Publishing Company, 1996, 70 pp. Gill, R. S. Operational detection of sea ice edges and

icebergs using SAR (2001). Canadian Jour. Rem. Sens. Vol. 27 No. 5, pp. 411 – 232.

Kloster, K. and W. Spring, Iceberg and glacier mapping using satellite optical imagery during the sea ice data acquisition programme (IDAP). Proceedings of POAC 1993, pp. 413 – 424.

Knapskog, A. O., (1996) Assessment of RADARSAT for detection and classification of icebergs. MSc Thesis (in Norwegian), Norwegian Technical University, Trondheim, Norway, 72 pp.

Power, D., J. Youden , K. Lane, C. Randell and D. Flett (2001). Iceberg detection capabilities of RADARSAT Synthetic Aperture Radar.

Canadian Jour. Rem. Sens. Vol. 27, No. 5, pp. 476-486.

Sandven, S, O. M. Johannessen, Martin Miles, Lasse H. Pettersson and K. Kloster (1999). Barents Sea seasonal ice zone features and processes from ERS-1 SAR. J. Geophys. Res. Vol. 104, No. C7, pp. 15843 – 15857.

Zubakin, G. K., A. K. Naumov and I. V. Buzin. Estimates of ice and icebergs spreading in the Barents Sea. Paper no. 2004-JSC-381, 8 pp, 2004.