sar-based estimation of glacial extent and velocity ...

SAR-BASED ESTIMATION OF GLACIAL EXTENT AND VELOCITY FIELDS ON ISANOTSKI VOLCANO, ALEUTIAN ISLANDS, ALASKA Daniel J. Sousa1,2,*, Audrey M. Lee1,3, Owen P. Parker1,4, Yamina Pressler1,5, Sung-Yee Guo1,6, Cindy Schmidt1,7, Batuhan Osmanoglu8 1

NASA Ames DEVELOP Program, NASA Ames Research Center, 2Columbia University, 3University of California, Los Angeles, 4San Francisco State University, 5California Polytechnic State University, San Luis Obispo, 6Los Altos High School, 7Bay Area Environmental Research Institute, 8University of Alaska, Fairbanks Geophysical Institute DEVELOP NASA Ames Research Center M.S. 239-20 Moffett Field, California 94035 * [email protected]

ABSTRACT Observations worldwide demonstrate Earth’s glaciers are ablating at increasing rates. Direct monitoring of glacial environments is often precluded by obstacles including restricted field access and paucity of cloud-free optical imagery. Synthetic aperture radar (SAR) sensors overcome these challenges by using microwave-frequency energy to observe Earth’s surface regardless of cloud cover or solar illumination. This study evaluated the applicability of two SAR monitoring methods to small (< 7 km2) glaciers in rugged topography. The glaciers of interest are on the flanks of Isanotski Volcano on Unimak Island in the Aleutian Archipelago. Two techniques were used to estimate glacier velocity fields and extents: (1) feature tracking and (2) coherence-based mapping. Data were analyzed from NASA Uninhabited Aerial Vehicle SAR (UAVSAR) and JAXA Advanced Land Observing Satellite Phased Array type L-band SAR (ALOS PALSAR) sensors over the period 2008-2011. UAVSAR power images were coregistered to sub-pixel accuracy and processed with the Coregistration of Optically Sensed Images and Correlation (COSICorr) feature tracking module to obtain estimates of glacial velocity fields. Maximum glacier velocities ranged from 28.9 m/yr to 58.3 m/yr. Glacial boundaries were mapped from interferometric coherence of ALOS PALSAR data then refined with masking operations based on terrain slope and segment size. Accuracy was assessed by comparison with manually digitized outlines, yielding 83.0% producer’s accuracy (errors of omission) and 86.1% user’s accuracy (errors of commission). Results represent refinement of a decades-old entry in the USGS National Hydrography Dataset (NHD). This procedure could be replicated in similar environments worldwide to provide a consistent methodology for ongoing evaluation of the global cryosphere. KEYWORDS: Satellite remote sensing, Synthetic aperture radar interferometry, SAR, Aleutian Islands, Unimak, Isanotski, Alaska, Glacier, Feature tracking, Boundary mapping, InSAR Coherence

INTRODUCTION Glaciers are integral to the wellbeing of communities worldwide, serving as sources of drinking water, hydropower, ecological diversity, and natural capital. Studies have shown globally-averaged trends of decreasing glacial length (Oerlemans, 2005) and mass (Cogley et al., 2005) over the 20th century. Observations have indicated mass loss of Alaskan glaciers in particular to be more rapid than any region on Earth other than Patagonia (Dyurgerov and Meier, 2005). These trends threaten the reliability of melt-fed water resources and the stability of nearby ecosystems. Additionally, Alaskan glacial melt contributes to sea level rise on the order of 0.12 mm/year (Berthier et al., 2010). Such concerns highlight the pressing need for development of precise, accurate and repeatable methods for ongoing glacial monitoring. The challenges of accurately monitoring glacial change are exacerbated by the intricacy of the systems involved. For instance, simple volume disequilibrium calculations based on changes in surface area would not account for possible changes in depth. In addition, surface observations generally fail to capture basal processes and different types of ice may exhibit variations in density and shear strength (König et al., 2001; Bardel et al., 2002). Any model seeking to capture this complexity requires extensive data over large spatial and temporal scales. Ancillary datasets are available for use, but are not exhaustive in critical information needed for glacial monitoring. For example, meteorological stations often provide long time series of climatic data that can be ASPRS 2013 Annual Conference Baltimore, Maryland ♦ March 24-28, 2013

synthesized with satellite observations for interpolation over most portions of the globe (e.g. NASA Global Precipitation Climatology Project; NOAA Unified Gauge-Based Analysis of Global Daily Precipitation). Moderate resolution digital elevation models are available nearly worldwide (e.g. NASA Shuttle Radar Topography Mission; NASA Advanced Spaceborne Thermal Emission and Reflection Radiometer global DEM). However, data specific to glacial variables are sparse. Field campaigns are costly and present a sampling bias skewed towards easily accessible regions (Atwood et al., 2010; Fallourd et al., 2011). The Global Land Ice Measurements from Space Project (GLIMS) is an attempt to capitalize on the coverage of global Earth observations to monitor the world’s glaciers. Originally launched by the NASA Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) science team, it is a multi-institutional effort “to establish a global inventory of land ice … to measure changes in extent of glaciers and, where possible, their surface velocities” (GLIMS, 1998). Two GLIMS products of interest are glacial outlines and velocity field maps. Glacier outlines are used as inputs for global climate models and aid in the analysis of glacial change over time by informing estimates of glacial mass. Velocity field maps are critical for accuracy assessment of glacial models. Both outlines and velocity fields can also serve as boundary conditions for such models. The Randolph Glacier Inventory (RGI 2.0) is a global dataset of glacier outlines intended for use in the upcoming Fifth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC AR5). As an amalgamation from many sources, some RGI 2.0 glacier outlines are based on recent analysis while others have proven difficult to update due to the limitations of optical satellite imagery and ground-based observations. These outlines persist as remnants of decades-old studies with little knowledge of data accuracy (Boyack, 2012). For glacial velocity fields, no comprehensive database exists. Recent work has utilized synthetic aperture radar (SAR) to improve estimates of these glacial parameters. Automated boundary delineation from SAR interferometry (InSAR) has been used to improve the RGI 2.0 effort in challenging locations (Atwood et al., 2010; Miller et al., 2011). SAR feature tracking techniques have been used to accurately track glacial motion (Giles et al., 2009; Ahn and Howat, 2011). These novel approaches circumvent several of the limitations of optical imagery and allow for a single consistent method of glacial outline and velocity field determination in diverse environments worldwide. Preliminary results from both methods are promising; however, studies to date have focused on large inland glaciers and expansive areas of landfast sea ice. This study will attempt to address the following question: Can the methods of (1) feature-tracking-based velocity field estimation and (2) coherence-based boundary delineation be applied to small mountain glaciers? If successful, these results may contribute to the improvement of glacial models and an enhanced understanding of the relationship between these glaciers and the global climate system.

BACKGROUND Glacial change is traditionally monitored by repeat survey of field markers. Despite improvements gained from GPS technology in recent decades, direct monitoring of glacial systems is often impractical due to inaccessibility and cost (Bardel et al., 2002). Remote sensing is a feasible alternative, providing the potential for uniform, routine and cost-effective monitoring of glacier status over large spatial and temporal scales. However, conventional remote sensing techniques are often invalidated over many glacial environments due to cloud cover and seasonal darkness (Atwood et al., 2010). Optical glacial monitoring is further limited by the presence of overlying debris and poor spectral differentiation between snow and ice. InSAR (Goldstein et al., 1993; Joughin et al., 1996; Bardel et al., 2002; Joughin et al., 2002) and SAR feature tracking (Lucchinta et al., 1995; Jackson et al., 2005; Giles et al., 2009; Ahn and Howat, 2011) have been proven as viable alternatives. This study builds on emerging trends in interferometry and feature tracking customization to explore new analysis solutions. SAR Techniques Delineation of glacier extent is a necessary first step in mass balance calculations (Gao and Liu, 2001; Atwood et al., 2010). Although differences between areas of seasonal snow and glaciated ice are obscured at the surface, radar reflectance responds to changes in subsurface density (König et al., 2001), which has traditionally been leveraged for boundary detection (Bardel et al., 2002). More recently, Atwood et al. (2010) simplified boundary mapping with InSAR by capitalizing on decorrelation of glacial surfaces over long time-series. Acquisitions for this purpose were chosen from summer months in order to both maximize glacial movement and minimize non-glacial snow cover. Meyer et al. (2011) expanded on these methods with improved techniques for automation and more rigorous assessment measures. The processing workflow in both studies included a morphological filter regime to overcome the persistent speckle patterns that are characteristic of radar imagery. Rough boundary delineations were then refined by image segmentation to reduce the number of spurious ASPRS 2013 Annual Conference Baltimore, Maryland ♦ March 24-28, 2013

identifications. Both methods are reliant on consistent detection of stable background features. L-band radar provides the necessary correlation at seasonal time scales not previously possible with shorter-wavelength C-band systems (e.g. Joughin, 2002). Although boundary mapping research has shown promise for global applicability, previous study (Atwood et al., 2010; Meyer et al., 2011) has focused on extensive inland glacial complexes and landfast sea ice. There is need to test similar methods over smaller, discontinuous glaciers. Velocity monitoring is another fundamental tool in the study of glacial change (Bindschadler, 1998). Although InSAR has been used successfully to derive glacial flow fields (Joughin et al., 1996; Joughin et al., 2002), recent research has focused on feature-tracking methods to reduce difficulties related to poor correlation. RepeatImage Feature Tracking (RIFT) methods have been commonly used for coregistration of SAR imagery in software packages such as IMCORR and Coregistration of Optically Sensed Images and Correlation (COSI-Corr; ENVI 4.8). RIFT processing algorithms utilize a moving search window of correlation coefficients to identify potential matches between images. Feature tracking for movement detection represents an innovative use of the procedure; rather than attempting to align stable image features, displacement of target objects is measured against the background of coregistered image pairs (Giles et al., 2009; Ahn and Howat, 2011). Edge detectors such as Sobel or Prewitt filters have also been used to emphasize subtle characteristics of the glacial surface before applying the RIFT algorithm (Ahn and Howat, 2011; Meyer et al., 2011). Sources of Error SAR data collection and signal processing have inherent errors due to side-looking imaging geometry; particularly shadow, foreshortening, and layover. Conventional SAR signal processing operates on the simplified assumption that all radar echoes are returning to the sensor from a planar surface (Richards, 2007). Errors thus become more severe over rough topography: shadow occurs when a surface is obscured by terrain and cannot present a return, foreshortening is a compression of scatterers due to sloping ground surface, and layover is a more severe case when the surface is sufficiently steep to reverse the scatterer order. Given that the target glaciers occur in mountainous terrain, all datasets in this study experienced some form of shadow, foreshortening, and/or layover. Another possible error source for feature tracking is the change of glacier reflectivity over time due to melting, snowfall, or deformation. Different surface conditions can hinder the performance of the feature matching algorithm, increasing the errors in estimated velocities. Effects of such errors can be reduced by using pairs with shorter temporal baselines, or removing the poorly matched (low signal-to-noise ratio) features from the analysis. Accuracy and coverage differ between airborne and spaceborne SAR sensors. Airborne systems typically yield higher spatial resolution SAR images (Rosen et al., 2006), whereas satellite systems provide relatively inexpensive global coverage (Richards, 2007). Both modalities experience errors due to design and mechanics of data acquisition. Airborne radar sensors are susceptible to attitude adjustments of roll, pitch, and yaw as well as instability due to airplane vibration, wind, and atmospheric turbulence (Rosen et al., 2006). Spaceborne radar sensors feature more stable orbital tracks but still have discrepancies which must be precisely corrected (Richards, 2007).

STUDY AREA The Aleutian Archipelago is located on the Pacific Ring of Fire between the Alaska Peninsula and the Kamchatka Peninsula. It formed volcanically from subduction of the Pacific plate under the North American plate. This plate boundary is an area of active orogeny, featuring more than 90 documented volcanoes as well as earthquakes up to moment magnitude 9.2. Estimates of plate motion indicate the Pacific plate to be sliding upward at a rate of approximately 6-7 cm per year (DeMets et al., 1994). Unimak is the largest island in the Aleutian chain and is composed of several major volcanic complexes which each range from 10-20 km in basal diameter (Figure 1). This study examines glaciers located on the flanks of Isanotski Volcano (2,470 m; Figure 2), an inactive dissected stratovolcano located on Unimak Island in the eastern Aleutians (Miller et al., 1998). Unimak Island is surrounded by the Pacific Ocean to the south and the Bering Sea to the north, resulting in a cool, wet, and windy maritime climate. While there is no distinct dry season, the wettest months occur between October and December (Rodionov et al., 2005). The islands are subject to the dominant influence of the Aleutian Low pressure cell as well as the Pacific Decadal Oscillation (PDO), resulting in persistently strong winds and

ASPRS 2013 Annual Conference Baltimore, Maryland ♦ March 24-28, 2013

Figure 1. Unimak Island, showing Isanotski Volcano and adjacent volcanic peaks. unrelenting cloud cover. The eastern Aleutian Islands have experienced a shift towards warmer temperatures after the PDO phase reversal of 1976/1977 (Rodionov et al., 2005) and temperatures are expected to continue to rise. According to modeling by Stahl et al. (2008), glacial equilibrium is not expected for at least 100 years. The eastern Aleutian Islands are a region subject to high risk of environmental impact from changes in climate (Berthier et al., 2010). Consistent with observations in similar locations worldwide, glacial watersheds on Unimak Island are experiencing a shift from high summer flows to more frequent low flows as temperatures gradually increase (Stahl et al., 2008). With surrounding communities dependent on these fresh water resources for economic vitality, water supply, and hydropower (Baker and Bolling, 2010), glacial monitoring plays a crucial role in adapting to the effects of glacier variability. Currently, the RGI 2.0 delineation of glacier extent in the area of interest for this study is derived from the USGS National Hydrography Dataset (NHD), a compilation of spatial information for all water bodies in the United States. The NHD outline was likely drawn from a military snowline and assessed circa 1950 (Boyack, 2012), encompassing a large contiguous area across multiple peaks. The GLIMS database has no ground or remotely sensed information on the expected velocities of the glaciers on Isanotski Volcano.

METHODS Data were acquired from the NASA Jet Propulsion Laboratory (JPL) Uninhabited Aerial Vehicle SAR (UAVSAR) multi-look complex (MLC) imagery archive. UAVSAR produces L-band imagery at high resolution and is preprocessed by JPL to correct for airplane attitude and flight path. Although the processing software used in this study was unable to perform interferometry on UAVSAR data, the multi-looked intensity images were appropriate for feature tracking applications. Alaska Satellite Facility (ASF) MapReady software was used to geocode and resample the radar source files to 7.2 m cell resolution GeoTIFFs. In addition, JAXA Advanced Land Observation Satellite (ALOS) Phased Array type L-band Synthetic Aperture Radar (PALSAR) Level 1.1 single look complex (SLC)

Figure 2. Peak of Isanotski Volcano, also known as “Ragged Jack.”


images were used for interferometric coherence-based boundary delineation. Acquisition dates and sensor information are listed in Table 1. 3 arc second (~60 m) NASA Shuttle Radar Topography Mission (SRTM) digital elevation model (DEM) was used for topographic modeling. Velocity field mapping was tested using IMCORR and COSI-Corr software. Image pairs provided by JPL corresponded with sub-pixel precision, making UAVSAR flight paths sufficiently stable to obviate co-registration preprocessing. For each image Figure 3. Conceptual schematic of RIFT algorithm. pair, change scenes were resampled to match reference scenes in terms of cell center and extent. RIFT algorithms within IMCORR and COSI-Corr were then used to compute the magnitude and direction of pixel displacements between the images. Although RIFT is intended to estimate scene-wide shifts for image transformations, it is also capable of tracking slow-moving environmental features such as glaciers. Settings were standardized as follows (see Figure 3): initial search window of 32 pixels; final of 16 pixels; step of 1 pixel. Best settings for the signal-to-noise confidence (i.e. ‘mask threshold’) varied by scene (Table 1). Once fields of vector displacements were derived, the results for each glacier were rectified against a standardized orthographic scene produced by JPL. Manual orthorectification was performed by individual glacier subsets in order to minimize topographic impacts and allow low-order polynomial transformations, thus reducing image warping. In addition, polynomial coefficients were applied against the derived displacement values to transform them into rectified ground geometry. Post-processing refinement was necessary to remove outlier vectors present in the initial RIFT output. First, the average background movement over stable features was attributed to sub-pixel image misregistration and detrended globally (Ahn and Howat, 2011). A multi-stage algorithm was then applied to remove spurious vectors within glacier boundaries. Assuming consistent flow direction (e.g., Ahn and Howat, 2011), vectors that deviated more than 110 degrees to either side of each glacier’s dominant downslope movement were masked as erroneous. Vectors with magnitudes greater than 2 standard deviations from the mean were also rejected. Finally, following Liu et al. (2012), a normalized median absolute deviation (MAD) statistic was used to identify directional displacements that varied more than 2 standard deviations from a local 11x11 moving window. After outlier vectors had been masked, remaining null value regions were replaced with inverse distance weighting (IDW) interpolation (Liu et al., 2012). Finally, a low pass 5x5 moving window filter was used to smooth local values (Ahn and Howat, 2011) and displacements were rescaled to represent one year time steps (Figure 4). The dataset used for boundary delineation (ALOS PALSAR scenes acquired in August of subsequent years) was chosen primarily because it operates in the L-band, which has been shown by other studies to preserve correlation over stationary targets covered with vegetation or loose material more consistently than shorterwavelength X- or C-band sensors (Lu et al., 2005; Atwood et al., 2010). Because the two images used were acquired with different sensing modalities, the higher resolution product was resampled to the coarser (i.e. ‘Fine Beam Single-polarization’ and ‘Fine Beam Dual-polarization,’ respectively), resulting in pixel sizes of 5 m azimuth by 19 m range. Another strength of the PALSAR data is global coverage, allowing similar analysis to be possible on small mountain glaciers worldwide. InSAR processing for boundary delineation was tested with both ADORE-DORIS and SARscape (ENVI 4.8). An interferogram was derived between multi-temporal image pairs. A reference phase image was then calculated from the DEM and subtracted from the raw interferogram to correct for topographic effects. The result was then filtered by the Goldstein-Werner method (Gabriel et al., 1989; Goldstein and Werner, 1998) to reduce noise within the phase fringes. Finally, an output of phase coherence was produced, which was then terrain corrected and geocoded in SARscape. To assist in model refinement, desired glacial boundaries were manually mapped from orthorectified UAVSAR MLC amplitude images. Key parameters for the boundary-delineation method of Atwood et al. (2010) were chosen by trial-and-error comparison (Figure 5). Distinct boundaries between glaciated ice and surrounding terrain were apparent in unwrapped coherence images, with a correlation limit of 0.4 chosen to extract a preliminary spatial boundary. A mask to remove slopes with gradient above 24% was applied. Artifacts of image noise were then eliminated with a morphological erosion filter, followed by a size filter applied to image segments. At this stage, target glacial objects were readily apparent and the appropriate size minimum (2.25 km2) was determined by cursory analysis. A morphological opening was then performed to grow the boundaries and correct for the earlier erosion filter. ASPRS 2013 Annual Conference Baltimore, Maryland ♦ March 24-28, 2013

RESULTS Feature Tracking For the five initial target glaciers on Isanotski Volcano (Figure 4), feature tracking was effective at deriving flow velocity fields in all but one case. Acquisition dates were hand-selected for each glacier to optimize signal-tonoise ratio for the RIFT algorithm in the face of topographic disturbances in the data. Because the flight tracks passed on different sides of the island, viewing geometry was often significantly enhanced on one scene versus another, even when captured on nearly identical dates. For glacier ‘B,’ poor signal return and abundance of radar shadow resulted in insufficient displacement detection over the glacial surface as well as difficulty in orthorectification. Although IMCORR may be more robust than COSI-Corr for such applications (Liu et al., 2012), Table 1. Feature Tracking Results by Glacier. Glacier A C D E

UAVSAR S c e n e P r o p e r t i e s Scene ID # Date 24003_10063_006_100804_L090 08/04/2010 24003_11054_007_110802_L090 08/02/2011 24003_09050_100_090722_L090 07/22/2009 24003_10063_006_100804_L090 08/04/2010 06004_09051_004_090723_L090 07/23/2009 06004_10063_011_100804_L090 08/04/2010 06004_10063_011_100804_L090 08/04/2010 06004_11054_006_110802_L090 08/02/2011

Analysis Parameters RMSE (Pixels) Mask Threshold

F l o w V e l o c i t y (m/yr) Maximum Mean Median

0.698

0.9

28.85

9.53

8.59

1.248

0.9

40.39

10.44

8.17

1.130

0.8

55.99

18.54

16.36

1.758

0.9

58.33

11.78

8.28

Figure 4. Velocity fields of annual flow rates by glacier. Hatched zones indicate areas of snow accumulation for which lack of coherence precluded stable displacement measurement. Note that for glacier ‘B’ excessive errors of radar shadow and layover prevented analysis.


command-line implementation of IMCORR was found to be operationally impractical and COSI-Corr was used for displacement detection. The final product for each image pair was a field of variable flow over the glacial surface, with summary statistics provided in Table 1. In Figure 4, cross-hatched symbology represents areas of erroneous RIFT output. These all occur in the accumulation zones, likely due to the rapid, discontinuous motion and heavy snowfall which is characteristic of the upper portion of most glaciers. Maximum flow velocities by glacier ranged from 28.9 meters/year to 58.3 meters/year, and were generally found to be located approximately halfway from accumulation zone to toe along the major axis of the glacier. Slowest glacial velocities were typically observed at glacial toes over areas of low topographic slope. Boundary Delineation Relative phase coherence was derived from a single suitable pair of L-band ALOS PALSAR summer captures (ALPSRP135121090, Aug. 2008, and ALPSRP188801090, Aug. 2009). As with flow field analysis, command-line ADORE-DORIS implementation of interferometry was tested and found to be impractical under time constraints; ENVI SARscape was used for all final results. The results of boundary delineation are summarized in Table 2. Four of the five glacier boundaries were mapped using the semi-automated procedure. Confident determination of boundary in the case of glacier ‘B’ was confounded by shadow. Accuracy was assessed by comparison with manually digitized outlines from the UAVSAR power images used for feature tracking, yielding 83.0% producer’s accuracy (errors of omission) and 86.1% user’s accuracy (errors of commission) across all glaciers. Total detected surface area by glacier ranged from 89.3% to 100.3%. While glacier boundaries (Figure 5) appeared generally correct, there were morphological variations inconsistent with the hand-digitized outlines. These were most apparent in accumulation zones and were likely a direct result of the deep incisions and rugged topography found in the upper regions of the dissected stratovolcano. Furthermore, inconsistencies at the toes of several glaciers were observed. In the case of glacier ‘E’ – the most prominent example – this was caused by the erosion filter severing a previously contiguous segment of incoherence from the body of the glacier. The glacier ‘E’ result also had noticeable discrepancies in the area where it rapidly changes direction, likely due to radar shadow. While the sample size was insufficient to determine whether these are persistent limitations of the method, they are noteworthy for future work with this technique.

Figure 5. Final glacial outlines in blue (i), with manually digitized outlines shown in black. Boundary delineation processing shown in black after (ii) coherence mask < 0.4, (iii) slope mask < 24%, (iv) erosion filter.


Table 2. Boundary Delineation Results by Glacier Glacier A B C D E

A r e a (km2) Manual Derived 3.358 3.173 4.467 N/A 4.198 4.380 3.061 3.246 6.675 6.845

Producer’s 83.7 N/A 87.4 86.3 78.4

A c c u r a c y (%) User’s 93.8 N/A 88.7 86.1 80.9

By Area 89.3 N/A 98.5 100.3 96.9

DISCUSSION Velocity Field Estimation Maximum glacier velocity results are from 28 to 59 meters/year, falling within the typical distribution of small mountain glaciers (Scherler et al., 2008). It must be noted, however, that without objective field validation these results remain tentative. Installation of corner reflectors on the glacial surface prior to SAR flight tracks would have allowed for rigorous quantitative analysis. The image cross-correlation software used for velocity field estimation relied on several assumptions. First, features not on the glacial surface were assumed to be stable over the 1-year time intervals between UAVSAR acquisitions. Second, despite motion, features on the glacier surface were assumed to maintain stable localized texture from year to year (i.e. as represented in the radar power image). The strengths and limitations of UAVSAR for this application both stem primarily from its airborne nature. Relatively low sensor altitude results in higher spatial resolution and ability to commission image acquisitions on demand. However, the tropospheric flight path adds significant atmospheric turbulence and subjects the sensor to challenges of airplane attitude and flight track instability. High spatial resolution also results in large file sizes and subsequent computational hurdles. Because continuous global coverage is not possible, analysis was limited to the yearly repeat time of the Aleutian UAVSAR campaign. Viewing geometry is another key factor for this study. UAVSAR has typically flown twice every summer over the Aleutians: one pass from NE to SW, south of Unimak Island, and one pass SW to NE, north of Unimak Island. Isanotski presents steep topography and the targeted glaciers flank it on all sides, resulting in preferable acquisition geometry for each glacier as indicated in Table 1. Furthermore, edge effects introduce a large source of error. RIFT compares localized correlation values to estimate displacement, such that search windows containing mixed glacier pixels and stationary background pixels will necessarily confound the algorithm. This was noticeable at glacier boundaries in every COSI-Corr result and generally led to an underestimate of the spatial extent of the glacier velocity fields. This effect is expected to be more pronounced at coarser image resolutions (e.g. from orbital sensors) or with smaller glaciers. Boundary Delineation The total glacier boundaries derived from this study are more than an order of magnitude smaller (21.8 km2 vs. 229.6 km2) than the previously existing outline in the GLIMS database (Figure 6). The GLIMS outline is over 60 years old and there is no record as to how the NHD data was originally composed or if it was in fact originally intended as a glacier outline (Boyack, 2012). While the refinement of the particular glacial outlines generated by this study is unlikely to have far-reaching impacts, the use of this method to correct inaccuracies in boundaries of other cloud- and darkness-obscured glaciers such as these could impact the quality of global Earth system models. Several assumptions underlie this processing workflow. First, glacial surfaces were assumed to behave like incoherent radar scatterers over a one year temporal baseline. This has been shown empirically in literature (Atwood et al., 2010) and is likely valid over the majority of the glacial surfaces. Second, glaciers were assumed not to exist below 180 m elevation, at slopes steeper than 24%, or at sizes smaller than 2.25 km2. These are user-derived values which are dependent Figure 6. Manually derived glacier boundaries (blue) on the area of interest. For this study, these values over archive NHD glacial dataset (purple). effectively removed spurious patches of incoherence and ASPRS 2013 Annual Conference Baltimore, Maryland ♦ March 24-28, 2013

may be useful guidelines for applications in similar locales elsewhere. The glacial outline map in Figure 5 shows the hand-digitized glacier outlines as well as those generated by the semi-automated procedure. While the general position and morphology appear to be similar, errors of omission and commission are apparent (Table 2). Perhaps the most noticeable feature is the large “false positive” region in the west of the image mistakenly identified by the algorithm as a glacier. This area falls in a topographic saddle between Mount Shishaldin (west of image) and Isanotski. It was not excised by the elevation mask as it lies at an elevation comparable to the glacial toes, nor was it removed by the size filter as it is a large contiguous area. This may indicate that the method of Atwood et al. (2010) is poorly suited to snow fields over extensive and evenly sloped accumulation zones such as those existing on youthful stratovolcanos. Several of the parameters used by the boundary delineation algorithm (e.g. slope threshold, coherence threshold, elevation threshold) require a priori knowledge by the user. While this method was useful in streamlining glacier boundary delineation, it represents a supervised algorithm that must be calibrated by an experienced operator. This method should therefore be used with caution in regions where the user has little or no prior experience. The largest sources of error for this method were due to the steep topography at the study site. Regions of foreshortening, layover, and shadow have adverse impacts on the quality of SAR-based results. Furthermore, a temporal baseline of one year often contains sufficient disturbance to the landscape to fully decorrelate interferograms. Fortunately, this study focused on a remote area where there is minor human disturbance on the coherence over a long time scale. Each of the applied post-interferometry filters may have introduced further errors. For instance, glacier edge morphology was noticeably modified by the iterative erosion and dilation steps used to remove spurious patches of decorrelation. The most extreme example of this occurred at the toe of glacier ‘E,’ where a patch of glacier became disconnected from the body and was removed after the area-based filter was applied (Figure 5). A key limitation of analysis was the lack of available scenes. This study was limited to only one pair of summer scenes, precluding the possibility of time series analysis. While the computations involved with interferometry and subsequent spatial analysis were complex, they did not pose an insurmountable hurdle once proper parameters were established. Future work may build on these methods to allow semi-automated determinations of change in glacier extent over time.

CONCLUSIONS The purpose of this study was to determine the viability of two SAR-based monitoring methods – coherence mapping and feature tracking – for small glaciers over rugged terrain. The coherence-based boundary mapping technique was found to be useful, with reasonable errors of both omission and commission when compared to a hand-digitized outline. The feature-tracking-based velocity field estimation technique produced results with good agreement spatially over the glaciers of interest, but was not possible to validate on the ground due to resource and time constraints. The main strength of both methods used in this study is their global applicability, regardless of cloud cover or solar illumination. Orbital SAR data is available worldwide and the techniques used in this study have been demonstrated to achieve reasonable accuracy despite rugged topography, perennial cloud cover, and small glacial extents. In the location of this study, the findings are a significant refinement over the existing outline in the GLIMS RGI 2.0 database. When viewed globally, this could present a substantial change in global estimates of cryospheric parameters which are integral to global Earth system models. A key benefit of the dual-method approach used in this study is that the strengths of one sensor complement the limitations of the other. ALOS PALSAR data was too coarse for successful velocity field estimation for glaciers of this size; the UAVSAR flight path is too complex to be used for interferometry by a typical end-user. Furthermore, the spatial extent of the velocity field provides confirmation of the boundary provided by coherencebased mapping. Similarly, the loss of coherence seen in the interferometric result is confirmation that there is considerable motion over the area of the glacier. Taken together, the interplay between two independent methods improves confidence in the findings of each. With no publicly available SAR satellite, analysis was limited to the few compatible scenes to which access was granted. All UAVSAR multi-looked scenes are freely available to any user. However, as noted above the spatial coverage of these data is limited to the flight tracks commissioned by JPL. Free ERS-1/2 data was made available after submitting an application, but presented issues which could not be overcome given time and resource constraints. This study was limited to one year-long time step for each method, but more powerful time-series analyses should be possible given better data availability. ASPRS 2013 Annual Conference Baltimore, Maryland ♦ March 24-28, 2013

Another limitation to the methods used is computational complexity. While freeware software programs exist (e.g. ADORE-DORIS, ROI-PAC), they generally function only from shell in Linux-based environments, requiring advanced knowledge of both computer science and InSAR. More user-friendly software is often limited in computational capability (e.g. NEST, MapReady). The new WInSAR ISCE package, with a pending release date, will likely be an improvement from the current state. The final interferometric processing in this study was performed using commercial software (ENVI SARscape). The implications of improved glacial monitoring are far-reaching, from community (water resources) to global (sea level rise) scales. These results can be used by other studies as boundary conditions for models of glacial dynamics or serve as an independent source of model output validation. If produced in time series, similar outputs could be correlated against other geophysical observations such as air temperature, precipitation, and geothermal activity to provide further insight into glacial change.

ACKNOWLEDGEMENTS The authors would like to thank Dave Hulslander at ENVI SARscape for a complimentary software trial license and troubleshooting assistance, without which the level of analysis would have been difficult to achieve, and Sheryl Boyack at USGS for information about the existing NHD outline. They would also like to thank Don Atwood and Jeff Freymueller at the University of Alaska, Fairbanks as well as Tom Lee and Kim Richardson at the United States Naval Research Laboratory in Monterey, California for their support throughout the research process.

REFERENCES Ahn, Y., and I.M. Howat, 2011. Efficient automated glacier surface velocity measurement from repeat images using multi-image/multichip and null exclusion feature tracking, IEEE Transactions on Geoscience and Remote Sensing 49(8):2838-2846. Atwood, D.K., F. Meyer, and A. Arendt, 2010. Using L-band SAR coherence to delineate glacier extent, Canadian Journal of Remote Sensing 36(1):S186-S195. Baker, A., and L. Bolling, 2010. Renewable Energy Resource Assessment for the Communities of Cold Bay, False Pass, and Nelson Lagoon. Aleutians East Borough, Alaska. 54 p. Bardel, P., A.G. Fountain, D.K. Hall, and R. Kwok, 2002. Synthetic aperture radar detection of the snowline on Commonwealth and Howard Glaciers, Taylor Valley, Antarctica, Annals of Glaciology 34(1):177-183. Berthier, E., E. Schiefer, G.K.C. Clarke, B. Menounos, and F. Rémy, 2010. Contribution of Alaskan glaciers to sealevel rise derived from satellite imagery, Nature Geoscience 3(2):92-95. Bindschadler, R., 1998. Monitoring ice sheet behavior from space, Reviews of Geophysics 36(1):79. Boyack, S., 2012. Personal communication, June 26, USGS Utah Water Science Center, Salt Lake City, UT. Cogley, J.G., M.A. Ecclestone, and D. T. Andersen, 2001. Melting on glaciers: environmental controls examined with orbiting radar, Hydrological Processes 15(18):3541-3558. DeMets, C., R.G. Gordon, D.F. Argus, and S. Stein, 1994. Effects of recent revisions to the geomagnetic reversal time scale on estimates of current plate motions, Geophysical Research Letters 21(20):2191-2194. Dyurgerov, M., and M.F. Meier, 2005. Glaciers and the Changing Earth System: A 2004 Snapshot. Occasional Paper 58. Boulder, CO: Institute of Arctic and Alpine Research, University of Colorado, Boulder. 118 pp. Fallourd, R., O. Harant, E. Trouve, J.-M. Nicolas, M. Gay, A. Walpersdorf, J.-L. Mugnier, J. Serafini, D. Rosu, L. Bombrun, G. Vasile, N. Cotte, F. Vernier, F. Tupin, L. Moreau, and P. Bolon, 2011. Monitoring temperate glacier displacement by multi-temporal TerraSAR-X images and continuous GPS measurements, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 4(2): 372-386. Gabriel, A.K., R.M. Goldstein, and H.A. Zebker, 1989. Mapping small elevation changes over large areas: Differential radar interferometry, Journal of Geophysical Research 94(B7):9183-9191. Gao, J., and Y. Liu, 2001. Applications of remote sensing, GIS and GPS in glaciology: a review, Progress in Physical Geography 25(4):520-540. Giles, A.B., R.A. Massom, and R.C. Warner. 2009. A method for sub-pixel scale feature-tracking using Radarsat images applied to the Mertz Glacier Tongue, East Antarctica, Remote Sensing of Environment 113(8):16911699. GLIMS, 1998. An overview of the GLIMS operational plan. http://www.glims.org/About/whitepaper.htm Goldstein, R.M., H. Engelhardt, B. Kamb, and R.M. Frolich, 1993. Satellite radar interferometry for monitoring ice sheet motion: application to an antarctic ice stream, Science 262(5139):1525-30.


Goldstein, R.M., and C.L. Werner, 1998. Radar interferogram filtering for geophysical applications, Geophysical Research Letters 25(21):4035-4038. Jackson, M., I.A. Brown, and H. Elvehøy, 2005. Velocity measurements on Engabreen, Norway, Annals of Glaciology 42(1):29-34. Joughin, I., 2002. Ice-sheet velocity mapping: a combined interferometric and speckle-tracking approach, Annals of Glaciology 34(1):195-201. Joughin, I., R. Kwok, and M. Fahnestock, 1996. Estimation of ice-sheet motion using satellite radar interferometry: Method and error analysis with application to Humboldt Glacier, Greenland, Journal of Glaciology 42(142):564-575. König, M., J.-G. Winther, and E. Isaksson, 2001. Measuring snow and glacier ice properties from satellite, Reviews of Geophysics 39(1):1. Liu, H., L. Wang, S.-J. Tang, and K.C. Jezek, 2012. Robust multi-scale image matching for deriving ice surface velocity field from sequential satellite images, International Journal of Remote Sensing 33(6): 1799-1822. Lu, Z., C. Wicks, O. Kwoun, J.A. Power, D. Dzurisin, 2005. Surface deformation associated with the March 1996 earthquake swarm at Akutan Island, Alaska, revealed by C-band ERS and L-band JERS radar interferometry, Canadian Journal of Remote Sensing 31(1):7-20. Lucchitta, B. K., C. E. Rosanova, and K. F. Mullins, 1995. Velocities of Pine Island Glacier, West Antarctica, from ERS-1 SAR images, Annals of Glaciology 21:277-283. Meyer, F.J., A.R. Mahoney, H. Eicken, C.L. Denny, H.C. Druckenmiller, and S. Hendricks, 2011. Mapping arctic landfast ice extent using L-band synthetic aperture radar interferometry, Remote Sensing of Environment 115(12):3029-3043. Miller, T.P., R.G. McGimsey, J.R. Richter, J.R. Riehle, C.J. Nye, M.E. Yount, and J.A. Dumoulin, 1998. Catalog of the Historically Active Volcanoes of Alaska, Open-File Report 98-582, U.S. Geological Survey, 104 p. Oerlemans, J., 1988. Simulation of historic glacier variations with a simple climate-glacier model, Journal of Glaciology 34(118):333-341. Richards, M.A., 2007. A beginner’s guide to interferometric SAR concepts and signal processing, IEEE Aerospace and Electronic Systems Magazine 22(9):5-29. Rodionov, S.N., J.E. Overland, and N.A. Bond, 2005. Spatial and temporal variability of the Aleutian climate, Fisheries Oceanography 14(s1):3-21. Rosen, P.A., S. Hensley, K. Wheeler, G. Sadowy, T. Miller, S. Shaffer, R. Muellerschoen, C. Jones, H. Zebker, and S. Madsen, 2006. UAVSAR: a new NASA airborne SAR system for science and technology research, Proceedings of the IEEE Radar Conference, April 24-27, 2006, Pasadena, CA: Jet Propulsion Laboratory, NASA. Scherler, D., S. Leprince, and M. Strecker, 2008. Glacier-surface velocities in alpine terrain from optical satellite imagery—Accuracy improvement and quality assessment, Remote Sensing of Environment 112(10):38063819. Stahl, K., R.D. Moore, J.M. Shea, D. Hutchinson, and A.J. Cannon, 2008. Coupled modelling of glacier and streamflow response to future climate scenarios, Water Resources Research 44(2):1-13.
