Vous êtes sur la page 1sur 20

CHAPTER 7

Quality in the GLIMS Glacier Database


Bruce H. Raup, Siri Jodha S. Khalsa, Richard L. Armstrong, William A. Sneed,
Gordon S. Hamilton, Frank Paul, Fiona Cawkwell, Matthew J. Beedle, Brian P. Menounos,
Roger D. Wheate, Helmut Rott, Liu Shiyin, Li Xin, Shangguan Donghui, Cheng Guodong,
Jeffrey S. Kargel, Chris F. Larsen, Bruce F. Molnia, Joni L. Kincaid, Andrew Klein,
and Vladimir Konovalov

ABSTRACT steps that must be passed before data are accepted


into the database. These steps ensure that ingested
Global Land Ice Measurements from Space data are well georeferenced and internally con-
(GLIMS) is an international initiative to map the sistent. The GLACE experiments and ingest time
world’s glaciers and to build a geospatial database quality control steps have led to improved quality
of glacier vector outlines that is usable via the and consistency of GLIMS data. This chapter pre-
World Wide Web. The GLIMS initiative includes sents the GLACE experiments and the quality con-
glaciologists at 82 institutions, organized into 27 trol steps incorporated in the data ingest process.
Regional Centers (RCs), who analyze satellite More recent similar studies are referenced.
imagery to map glaciers in their regions of expertise.
The results are collected at the U.S. National Snow
and Ice Data Center (NSIDC) and ingested into the 7.1 INTRODUCTION
GLIMS Glacier Database. A concern for users of
the database is data quality. The process of classify- GLIMS is the first attempt to build a globally
ing multispectral satellite data to extract vector out- complete, high-resolution map of glacier extents;
lines of glaciers has been automated to some degree, currently there are complete regional glacier in-
but there remain stages requiring human inter- ventories and incomplete global inventories. The
pretation. To quantify the repeatability and preci- GLIMS Glacier Database has begun to allow new
sion of data provided by different RCs, we designed scientific questions to be addressed, such as global
a method of comparative image analysis whereby statistics of glacier area and area elevation distribu-
analysts at the RCs and NSIDC could derive glacier tion, global trends in glacier area change and mass
outlines from the same set of images, chosen to change, and regional variability in rates of change.
contain a variety of glacier types. We carried out The GLIMS Glacier Database contains not just
four such experiments. The results were compiled, point locations for glaciers, as in the World Glacier
compared, and analyzed to quantify inter-RC anal- Inventory (WGI), but also glacier outlines as closed
ysis consistency. These comparisons have improved polygons, which record where the glacier bound-
RC ability to produce consistent data, and in aries were at a specific time. Also recorded for many
addition show that in the lower reaches of a glacier, glaciers are extents of supraglacial debris and lakes,
precision of glacier outlines is typically 3 to 4 pixels. proglacial lakes, snow lines, and approximate cen-
Variability in the accumulation area and over parts ter flow lines, as well as nonspatial data such as
of the glacier that are debris covered tends to be glacier name, source imagery and maps, and analyst
higher. The ingest process includes quality control details.
164 Quality in the GLIMS glacier database

As of early 2014, the number of glacier outlines in longitude/latitude or projected coordinates), but
the GLIMS Glacier Database was 122,414, repre- the choice for representation of the outcrop is less
senting approximately 70% of the estimated total of obvious. Geographic Information System (GIS)
Earth’s glaciers. The total area covered in the data- tools allow polygons to have ‘‘holes’’, and this
base is 520,000 km 2 , also about 70% of the esti- method is a frequent choice for representing
mated total area. nunataks. Holes are integral to the polygon, how-
For such a global database to be useful and ever, and must therefore share attributes with that
trustworthy to users, close attention must be paid polygon. Within GLIMS it was decided to allow for
to data quality and consistency. Glacier outlines the possibility that nunataks would have a separate
need to have good consistency between regions set of attributes, and so they are represented by
and over time in order for scientific questions to separate polygons instead of holes in the glacier
be addressed. The design of the database itself outline polygon.
imposes a consistent set of parameters and one data This one example illustrates the need for standard
model on GLIMS analysts, but despite large gains ways of representing glacier entities within GLIMS.
in the degree of automation of glacier classification The GLIMS Analysis Tutorial (http://glims.org/
in satellite imagery, automated algorithms must be MapsAndDocs/guides.html ) documents the GLIMS
tailored to the particular characteristics of glaciers approach to modeling glacier entities. Additionally,
from region to region, and human judgment and it is important to have standard formats for trans-
subjectivity remain necessary ingredients of the ferring glacier-mapping data from the analyst to
mapping process. The calculation of area changes the GLIMS Glacier Database. The GLIMS Core
introduces additional pitfalls. Factors that affect Technical Group defined a standard GLIMS data
the quality of glacier outlines derived from satellite transfer format, which is documented at http://
imagery include image georeferencing; variations of glims.org/MapsAndDocs/datatransfer/data_transfer
seasonal snow cover; debris cover on glaciers; work- _specification.html.
ing definition of ‘‘glacier’’ as an entity that may be A software tool called GLIMSView was created
connected to other ice bodies; and difficulties in in order to make it easier for GLIMS RCs to pro-
defining ice flow divides. Differing interpretations duce glacier-mapping data in the correct data
of snowfields in the accumulation area or of debris- model and to package these data in the GLIMS
covered ice in the ablation region can greatly affect data transfer format. It supports manual digitiza-
the calculated area for a glacier, possibly leading to tion of glacier boundaries from satellite imagery,
erroneous climatic interpretation. and exports the outlines and all attributes (e.g.,
This chapter presents the ways in which the name of analyst, Regional Center information,
GLIMS core developers and RCs have addressed physical parameters such as glacier area, etc.) in
methodological challenges encountered in space- the GLIMS data transfer format. It can also import
borne glacier mapping. These steps include the already existing glacier outlines, and therefore can
development of standard methods for mapping be used as a packaging tool for glacier outlines to
land ice from satellite imagery; the development prepare them for ingest into the GLIMS Glacier
of standard tools, such as GLIMSView, for glacier Database. It has been used for both purposes by
mapping and packaging of the resulting data; a number of RCs.
glacier analysis comparison experiments (GLACE), GLIMSView is free downloadable (open-source)
in which mapping results from multiple analysts are software that runs on Linux and Windows. Devel-
compared; the design of the GLIMS Glacier Data- opment ceased in 2009, and continued development
base; and the quality control steps in the data ingest is contingent on new funding for that purpose.
process. Similar functionality could be built in the form of
plug-ins for GIS software such as QGIS, GRASS,
or ArcGIS.
7.2 STANDARD METHODS AND TOOLS

Different people have different ideas about how to 7.3 ACCURACY AND PRECISION IN
represent glacier boundaries digitally. For example, GLACIER MAPPING
imagine a glacier with a rock outcrop in the middle
of it (a nunatak). The glacier outline is typically Given the distributed nature of glacier-mapping
represented by a polygon (sequence of vertices in efforts in GLIMS, it was recognized early on that
Accuracy and precision in glacier mapping 165

Figure 7.1. Five manual digitization trials described in Sneed (2007), performed separately from the GLACE
experiments. Five independent digitizations of a glacier boundary are plotted over the source image.

the differences in mapping results (from different was digitized five times independently, and the
algorithms and analysts) needed to be quantified. results were compared. They found that in the
Several experiments have been done, conducted case of a glacier of area 1.242 km 2 , variations in
either by individual Regional Centers or set up by digitization of the terminal boundary would result
the Core GLIMS Team, to compare results under in area uncertainty of approximately 1.7%. A part
controlled conditions. of the set of outlines is shown in Fig. 7.1. Paul
These experiments have focused on analytical (2007) tested the repeatability of manual digitiza-
variations and all sources of error arising from tion by one person, and also by two people, and
applying different image classification algorithms, found that relative error in resulting glacier area
manual image interpretation, and the complete exceeds 10% when the glacier area is 0.1 km 2 or
end-to-end effect of the mapping effort. To evaluate smaller. For larger glaciers, relative error was 5%
repeatability of manual digitization, the GLIMS or less.
participants in Sweden investigated the effects of Suites of automated methods used for the initial
human interpretation on manual digitization results mapping of glacier outlines have been analyzed in
by having nine operators outline distinct lake several previous studies (Albert 2002, Paul et al.
shorelines in a high-resolution aerial photograph, 2002, Paul and Kääb 2005, Racoviteanu et al.
and found that relative uncertainty in the resulting 2009) and generally show only marginal differences
outlines was 2.5 pixels, though this could be among the applied methods. Many common image
improved by applying binary-encoded transects classification algorithms perform well for clean
perpendicular to the lake boundaries (Sannel and glaciers (glaciers lacking rock-debris cover), and
Brown 2010). Similarly, Sneed (2007) describes a most of them perform poorly when glaciers are
test whereby the terminus of a glacier in Svalbard debris covered.
166 Quality in the GLIMS glacier database

Some additional characterizations and assess- experiments were not intended to assess the abso-
ments of error are given in many chapters in this lute accuracy of mapping results. Hence, mapping
book. We draw special attention to the treatments results were compared with each other, but not with
provided by Ramachandran et al. (Section 6.3 on any independent and validated source of glacier
sensor calibration and ASTER image geometric boundary information.
corrections and errors, and Section 6.4.2.3 on detec- In both GLACE 1 and GLACE 2, images were
tion versus full resolvability of features); Demuth et chosen to contain a variety of glacier types, and
al. (Online Supplement 16.3 on error of digitized various types of boundaries: ice–rock, ice–vegeta-
glacier boundaries); and Krumwiede et al. (Section tion, and ice–ice (Table 7.1). A digital elevation
22.4.6 on error of digitized glacier boundaries and model (DEM) was made available as ancillary data,
areas). Certainly as GLIMS and other glacier anal- to be used to aid interpretation of optical imagery.
ysis initiatives move toward change assessments However, because we knew that some participants
and other derivatives, the origination and propaga- had the facility to orthorectify and terrain-correct
tion of all significant errors must be tracked with imagery and others did not, we chose to prohibit
ever greater care. orthorectification for the purposes of these experi-
ments, so that the results would all be comparable.
The participants used a variety of methods, ranging
7.4 GLACIER ANALYSIS COMPARISON from manual digitization to fully automated tech-
EXPERIMENTS (GLACE) niques (Table 7.2). In GLACE 1, participants were
requested to digitize the boundary of one small
The GLIMS Core Team decided to implement a glacier manually.
series of glacier analysis comparison experiments GLACE 1 was conducted in 2004 and results
(GLACE, pronounced the same as ‘‘glass’’) to were reported at the August 2004 GLIMS Work-
quantify uncertainty in glacier mapping from satel- shop in Oslo, Norway, the Fall 2004 Meeting of the
lite imagery. American Geophysical Union (Raup et al. 2004),
Four GLACE experiments have been carried out and the December 2004 GLIMS Mini-workshop in
to date: GLACE 1, GLACE 2, GLACE 2A, and San Francisco. GLACE 2 was carried out in the
GLACE 3A. GLACE 1 and GLACE 2 focused on autumn of 2005, and results were reported at the
automated methods for glacier mapping from GLIMS Meeting in New Zealand in February 2006
imagery, and participants were allowed to use the and at the Arctic Workshop in Boulder, Colorado
software tools and algorithms of their choice. in March 2006. GLACE 2 included a change
GLACE 2A and GLACE 3A evaluated only detection component using multitemporal optical
manual digitization of the glacier boundaries. imagery. The analysis methods used in GLACE 1
and GLACE 2 are summarized in Table 7.2. Many
of the automated methods applied a threshold to
7.4.1 GLACE 1 and GLACE 2
the ratio of two sensor channels (Paul and Kääb
GLACE 1 and GLACE 2 allowed the participants 2005). The normalized difference snow index was
to use the tools and algorithms they plan to use also used, which for ASTER can be defined as
operationally in GLIMS. The goal was to assess ðB1  B4Þ=ðB1 þ B4Þ (where B1 ¼ Band 1 and
the precision and repeatability (variability) in the B4 ¼ Band 4) (Hall et al. 1995, Paul 2007). When
resulting data under realistic conditions. These B1 is saturated, B2 is sometimes used. Individual

Table 7.1. Satellite images used in the GLACE experiments.

Image ID Acquisition date Sensor GLACE No.

SC:AST_L1A.003:2004103566 September 6, 2001 ASTER 1

P050R24_5T910921 September 21, 1991 Landsat TM 2

SC:AST_L1A.003:2010881449 September 21, 2000 ASTER 2, 2A

SC:AST_L1A.003:2035265399 July 20, 2006 ASTER 3A


Glacier analysis comparison experiments (GLACE) 167

Table 7.2. Tools and techniques used in GLACE 1 and GLACE 2. The group (participant) numbers below have
been assigned randomly (separately for GLACE 1 and GLACE 2).

GLACE 1

Group Tools Techniques

1 Matlab, GLIMSView Band ratio 3/4, 3/6, 3/8 ! RGB; manual, maps
2 ERDAS Imagine, Arc/Info Band ratio 3/4, threshold 2.0, visual interpretation
3 Arc/Info, GLIMSView Unsupervised classification with manual editing
4 ENVI, PCI, Arc/Info Multistep ratio thresholding algorithm
5 Arc/Info Ratio 3/4, threshold 2.4; manual in shadows; >0.2 km 2
6 Matlab, GLIMSView, topo maps Ratio 3/4, threshold 2.5
7 PCI Works PCA on 1–4, NDSI

GLACE 2

1 PCI Band ratio enhancement; manual delineation of outlines


2 Matlab, ERDAS, GLIMSView Three different band ratios as RGB; manual interpretation
3 GLIMSView Manual delineation of outlines
4 PCI, ESRI Band ratio, threshold
5 ENVI 4.2, Google Earth Manual delineation of outlines
6 ENVI 4.1, ESRI Manual delineation of outlines
7 PCI, ESRI Bands 3–5 supervised classification for accumulation, ablation
8 GLIMSView, ESRI, ENVI, ERDAS Unsupervised classification based on NDSI and ASTER 2/5 ratio;
manual cleanup of automatically generated vectors

Abbreviations: RGB ¼ red, green, blue; PCA ¼ principal components analysis; NDSI ¼ normalized difference snow index. GLACE 2A
and GLACE 3A employed only manual digitization.

algorithm choices were based on participants’ pre- material in their normal GLIMS domains, the ice
vious experience applying them to glaciers in their masses in these images provided a region of clean
regions. Comparisons of automated glacier- ice that we predicted would work well with algo-
mapping algorithms are given by Albert (2002), rithms tuned for high-latitude types of glaciers (with
Paul et al. (2002), and Paul (2007). minimal debris cover).
While GLACE 1 revealed systematic problems
with image preprocessing and interpretation, the
goal of GLACE 2 was to derive a quantitative 7.4.2 GLACE 2A and GLACE 3A (manual
estimate of confidence in GLIMS analysis results, digitization)
with an additional focus on change detection. We GLACE 2A and GLACE 3A were performed as
selected two images covering the same area, the part of dedicated GLIMS workshops, and partici-
Klinaklini Glacier and surrounding glacier system pants interpreted the imagery and manually created
in the Coast Mountains of British Columbia, outlines while sitting together in a computer lab at
Canada (Fig. 7.2): an ASTER scene, acquired the workshop venues. The goal of these experiments
September 21, 2000 and a Landsat 5 TM scene, was to remove from consideration the differences
acquired September 21, 1991 (precisely nine years arising from the application of different algorithms
earlier; see Table 7.1). This allowed participants to and tools, and use only manual methods in order to
evaluate the ability to detect surface changes based evaluate variability in human interpretation of the
on images acquired from different instruments with imagery.
different characteristics, such as spatial and radio- GLACE 2A was conducted as part of a GLIMS
metric resolution. The region features a glacier sys- workshop held in Tucson, Arizona in September
tem containing many tributaries, a variety of sizes 2005. Approximately 10 participants used the
of mountain glaciers, clearly visible transient snow GLIMSView software package (http://glims.org/
lines, ice flow divides, various glacier boundary glimsview/) to manually digitize the boundaries of
types, and debris-covered as well as clean glaciers. a small glacier in British Columbia from the
While not all RCs have glaciers with morainal ASTER image used in GLACE 2 (Table 7.1).
168 Quality in the GLIMS glacier database

GLACE 1 GLACE 2
Figure 7.2. Images used in GLACE 1 and GLACE 2. (Left) False-color composite ASTER image acquired
September 6, 2001; (right) false-color composite Landsat TM image acquired September 21, 1991. Details on
the images used in all the GLACE experiments are listed in Table 7.1. Figure can also be viewed as Online
Supplement 7.1.

ASTER bands 1, 2, and 3 were displayed as blue, bridge, England. Participants manually digitized
green, and red, respectively, to create a visible near- the boundary of the terminus of the Klinaklini
infrared (VNIR) false-color composite image. Glacier, British Columbia, Canada.
After a short learning period to get familiar with Participants for all four experiments are listed in
GLIMSView, each participant visually interpreted Table 7.5 (p. 182).
the image and produced a vector outline of the
glacier extent by tracing its perimeter with the
mouse, basing their interpretation on their glacio-
7.5 GLACE RESULTS
logical expertise and previous experience viewing
satellite imagery of glaciers. They also produced
7.5.1 GLACE 1 and GLACE 2
vector lines to denote the location of snow lines
and center flow lines. The glacier’s boundaries The quality of the results in GLACE 1 was variable
included a flow boundary (ice–ice contact), as well and the experiments revealed problems such as
as ice–rock boundaries. (1) georeferencing errors (Fig. 7.3, left panel),
After producing a glacier outline using only the (2) interpretation errors in manual digitization,
ASTER image, the participants viewed the glacier (3) interpretation differences in manual digitization
using Google Earth, which at that time included a (Fig. 7.4), and (4) algorithmic deficiencies in auto-
moderate-resolution multispectral image (probably mated methods (Fig. 7.3, right panel). An example
from Landsat’s TM instrument) and a DEM. The of an interpretation error is the inclusion of non-
combination of the multispectral imagery and ele- glacier material, such as a rock slope or proglacial
vation data is viewable as a pseudo-3D scene from lake, within the glacier boundary. Interpretation
an arbitrary angle. The analysts used this new differences result from varying definitions of what
source of information with the ASTER image and to include as a ‘‘glacier’’ (e.g., should the laterally
created a new set of outlines. adjacent snow slope be part of the glacier? Where
The GLACE 3A experiment was similar to should the boundary between a debris-covered
GLACE 2A, and was held in conjunction with glacier and a partly ice-cored moraine that is sep-
the August 2006 GLIMS Workshop, held in Cam- arate from the glacier be drawn?). Algorithmic defi-
GLACE results 169

Figure 7.3. (Left) All GLACE 1 glacier boundaries overlaid on the ASTER image that was analyzed in the
experiment. Gross georeferencing errors, due to some initial difficulty in handling ASTER imagery, are apparent.
(Right) Some GLACE 1 glacier boundaries for Spencer Glacier overlaid on the ASTER image. Classification errors
include inclusion of the proglacial lake as part of the glacier (group 3, orange), and exclusion of lightly debris-
covered ice near the glacier terminus (group 6, yellow). Blue ¼ group 2; green ¼ group 1. Figure can also be viewed
as Online Supplements 7.2a and 7.2b.

Figure 7.4. (Left) GLACE 1 boundaries for Skookum Glacier overlaid on the ASTER image that was analyzed in
the experiment. A portion of the glacier is debris covered, making it dark in color. Some analysts mistakenly excluded
this from their glacier polygons. Analysts also differed in their interpretation of the snowfield on the glacier’s
northern side (north is up in image). (Right) Two GLACE 2 glacier outlines overlaid on the September 9, 2000
ASTER image from that experiment. Some analysts included the small tributary glacier (indicated by arrow) as part
of the Klinaklini Glacier, while others did not. Figure can also be viewed as Online Supplements 7.3a and 7.3b.
170 Quality in the GLIMS glacier database

Table 7.3. Quantitative comparison between different versions of manually digitized outlines for a specific glacier
(an unnamed glacier on the east side of Boggs Peak, 12 km east of Portage, Alaska; 60.835 N, 148.742 W, GLIMS
ID G211257E60835N, visible in Fig. 7.3 in upper central part of left image), produced by the participants in GLACE
1. Group numbers have been assigned randomly.

Group number 1 3 4 5 6 7 Mean Std. Dev.

Area (km2 ) 1.79 2.81 3.68 3.91 4.01 5.65 3.642 1.293

ciencies led to underestimation of glacier area in of the term glacier for use within the GLIMS
several cases. For example, parts of the tongues project. This resulted in a formal definition being
of some glaciers were lightly debris covered, leading included in the GLIMS Analysis Tutorial, as
some algorithms to misclassify those regions as rock discussed below.
(nonglacier). In order to quantify the differences between out-
GLACE 1 was the first of this kind of test, and lines produced from the same image, for a given
was therefore a learning experience at various pair of outlines (from two different analysts), we
levels. Notably, many in the GLIMS community calculated the straight line (shortest) Euclidean
were new to ASTER imagery, which poses unique distance between each vertex of one outline and
challenges (Abrams et al. 2002) for georeferencing the other outline. This was done by generating a
in some software. Additionally, at the time of this ‘‘distance grid’’ for each polygonal outline where
experiment, the GLIMS community had not yet the value at each grid cell is the normal distance
formulated a single definition of ‘‘glacier’’ for the from the cell center to that outline. Each grid was
purposes of GLIMS glacier delineation. These then sampled at the locations of the vertices of all
problems were starting to be addressed by the time the other outlines. These distances are similar to the
of the GLACE 2 experiment. Hausdorff distance (Alt et al. 1995) used in other
Given the different data models in which some of disciplines (polygonal feature matching in medical
the automatically generated data were delivered, imaging, for example). However, instead of retain-
meaningful quantitative comparisons among them ing the maximum of these distances (the Hausdorff
were impossible without modifying some of the distance), we examined the statistics for all of them.
data first. In light of this and the large qualitative The result is two sets of distances for every possible
differences, qualitative comparisons were deemed outline pair, each set consisting of distances
sufficient for most of the outlines submitted in this between each vertex of one outline and the other.
round of GLACE experiments. By contrast, the (There are two sets because calculation of the dis-
manually digitized glacier outlines of GLACE 1 tances from one set of vertices and the other outline
were all similar to each other. Table 7.3 shows is not a symmetric operation.) Each set of distances
the calculated areas and their summary statistics. represents a measure of the difference between two
Not all groups produced a manually digitized out- outlines, and these have been plotted as box-and-
line for this glacier. whisker plots in Fig. 7.6. The extent of the boxes is
All the outlines produced in the GLACE 2 the interquartile range, the whiskers extend from
experiment are shown in Fig. 7.5. The georeferenc- the 5th to 95th percentiles, and outliers are shown
ing problems encountered in GLACE 1 were largely as circles. The thick horizonal line is the median. In
mitigated in GLACE 2. However, interpretation terminus areas, the polygons generally had hun-
differences remained. Fig. 7.4 (right panel), for dreds of vertices. Distances are calculated between
example, shows that different analysts treated the vertex of one polygon and the interpolated
smaller tributary glaciers differently. In this case, straight line (within the UTM Zone 9 projection)
one analyst included the small tributary as part of connection to the other. Because vertex density is
the main glacier, while another excluded it. This high, there is no effect from varying numbers of
sort of problem led to an extensive discussion at vertices in the polygons. Fig. 7.7 shows the distribu-
the 2006 New Zealand GLIMS Meeting, and sub- tion of distances from the 581 vertices in polygon 1
sequently on the GLIMS electronic mailing list, to polygon 3, two of the better and more consistent
about how to specify a strict practical definition polygons. Standard deviation is 71 m, or approxi-
GLACE results 171

Figure 7.5. All outlines from GLACE 2, Landsat image. The outlines generally match well in the terminus area,
whereas there is high variability in the accumulation area. The analyst who produced the red outline applied a
different (non-GLIMS) data model, and digitized the contribution of each tributary to the terminus trunk separately.
The yellow outline excluded morainal material in the terminus area which should have been included in the glacier
outline. Figure can also be viewed as Online Supplement 7.4.

mately 4.7 pixels. Therefore, total positional uncer- and were excluded from the summary statistics in
tainty due to all sources for the best analysts was Table 7.4. Similarly, the change in area from group
about 4.7 pixels. 6 was omitted from the summary statistics of area
An additional feature of the GLACE 2 experi- change due to its obvious underestimate of area.
ment was analysis of two images, separated by nine Overall, the area of the Klinaklini Glacier does
years, of the same glacier system. Participating RCs not appear to have changed significantly during
produced a set of glacier outlines from each image the nine years between image acquisitions. Mass
and provided an estimate of area change for the loss can only be inferred, but there is evidence in
glacier. Some analyses showed a slight increase in the images, such as elevated vegetation trim lines, of
area, while others showed a slight decrease. On glacier thinning.
aggregate, the overall results showed area change
that was not statistically different from zero. How-
7.5.2 GLACE 2A and GLACE 3A
ever, the results from the most internally consistent
analysis indicated that the Klinaklini Glacier lost In the manual analysis of the small glacier near
approximately 1% of its area from 1992 to 2000 Klinaklini Glacier (GLACE 2A), the analysts each
(Table 7.4). Note that the standard deviation of produced either one or two outlines. Some pro-
the measured area changes is greater than the mean duced one, then after viewing the glacier in Google
(or median) change. The anomalously high area Earth, produced another using the additional infor-
from group 5 is due to inclusion of rock outcrops mation. Others had not produced an outline by the
internal to the glacier in the area computation. This, time they viewed the glacier in Google Earth, and
and the areas for group 2, were identified as outliers produced only the second outline. A few of the
172 Quality in the GLIMS glacier database

Figure 7.6. The distances between all the vertices of one glacier outline and the other outlines were determined.
This matrix of box plots summarizes interpolygon distances in GLACE 2. Distance between each polygon is shown
on the diagonal for comparison. One polygon (from group 6), was created from automated methods that led to a
narrower outline than the others; hence the distances between that outline and the others (bottom row, rightmost
column) are larger than the other pairs. The distances on the vertical axes are in meters. The width of the box plots
has no meaning. The calculation of distances from the vertices of one polygon to another polygon is not a symmetric
operation, though the values are generally similar.

participants were satisfied enough with their first Viewing the upper part of the glacier using only
outline that seeing Google Earth made no differ- the nadir image, analysts found it difficult to con-
ence, and they produced no second outline. sistently identify where the change in slope was
The results were highly variable, particularly in between the glacier of interest and its neighbor.
the interpretation of ice–ice flow boundaries (ice In the terminus region, a rocky or debris-covered
divides) in the upper snow-covered reaches of the area adjacent to the glacier was interpreted to be a
glacier, as well as the terminus region (Fig. 7.8). valley wall by some analysts, and a debris-covered
GLACE results 173

Figure 7.7. Histogram of the distances between outlines 1 and 3 in GLACE 2. These two outlines are visually
consistent with each other.

glacier by others. Fig. 7.8 shows the outlines super- ment with each other. This exercise emphasized
imposed over the imagery provided by Google the fact that topographic information is crucial
Earth, where it is clear from the topographic infor- for proper boundary delineation where there are
mation that the rocky area is a valley wall. The red ice flow divides and supraglacial debris.
lines were produced before viewing the glacier in The outlines produced in GLACE 3A are shown
Google Earth, and the blue lines were produced in Fig. 7.10. The lateral boundaries are well identi-
after. There is less variability in the blue outlines fied by all participants, but there are a few slight
compared with the red. differences in the terminus region.
Fig. 7.9 shows variability in the resulting areas
calculated from the outlines before viewing the
7.5.3 Discussion
three-dimensional data of Google Earth (left panel)
and after (right panel). In this case, the use of the As discussed above, errors can be categorized as
3D information led to less variability as well as a georeferencing errors, interpretation errors, inter-
smaller final outline for the glacier. In general, how- pretation differences, or algorithmic deficiencies.
ever, we do not expect the use of 3D information to The automated glacier-mapping methods used in
lead to smaller estimates of glacier size, but only to the GLACE tests were based only on multispectral
reduce variability of the estimates. The addition of data (not topography), and thus were best suited
topographic information enabled the analysts to for delineation of glaciers without optically thick
interpret the scene with higher confidence, and (opaque) and extensive debris cover or ice divides.
the resulting outlines were in much better agree- Some of the larger errors were due to debris cover
174 Quality in the GLIMS glacier database

Table 7.4. Changes in area of the Klinaklini Glacier as determined by the


different groups participating in GLACE 2. Group numbers have been
assigned randomly, and differently from GLACE 1. In both area and area
change measurements, data that were clearly outliers, marked in the table by
asterisks, were removed before calculating the summary statistics at the bottom
of the table. Acquisition dates for the ASTER and Landsat scenes were
September 21, 2000 and September 21, 1991, respectively.

Group number ASTER area TM area Area change Area change


(km 2 ) (km 2 ) (km 2 ) (%)

1 450.7 441.3 9.4 2.13

2 304.7  316.8  12.1 3.82

3 409.5 408.6 0.9 0.22

4 454.4 453.4 1 0.22

5 677.7  n/a n/a n/a

6 459.8 503.9 44.1  8.75 

7 402.1 413.7 11.6 2.8

8 474.4 479.9 5.5 1.15

Min 402.1 408.6 12.1 3.82

Max 474.4 503.9 9.4 2.13

Median 452.55 447.4 2.3 0.465

Mean 441.82 450.1 2.98 0.867

Std Dev 29.13 37.20 8.34 2.18


(6.6%) (8.3%)

or tributaries being excluded from the glacier area, ous. Automated algorithms remove human subjec-
varying interpretation of ice flow divides, and an tivity from the process and can map an entire
iceberg-filled lake being included in the glacier satellite scene hundreds of times faster than purely
area. The participants who edited the results from manual digitization (Bishop et al. 2001, Paul et al.
their automated algorithms to compensate for these 2004, Raup et al. 2007a, Racoviteanu et al. 2009).
effects achieved improved results. It is clear that The GLACE tests have helped the GLIMS com-
topographic information can be crucial for accurate munity converge on appropriate algorithms for
delineation of glacier boundaries, especially in accu- different glacier types (Paul et al. 2009).
mulation zones and where there is supraglacial For interpretation errors and differences, the
debris cover. While manual digitization is well GLACE experiments have been instrumental in
suited to final editing, automated algorithms are identifying and reducing errors and variability
recommended as the first step to produce a glacier among glacier outlines produced from satellite
map for an entire scene. Manual editing can then be imagery by different GLIMS collaborators. One
used to fix errors due to debris cover and cast of the chief outcomes is a standard definition of
shadow. Algorithms that use both multispectral ‘‘glacier’’ for the purposes of GLIMS. The
imagery and topography to map debris-covered definition was agreed upon through discussions at
glaciers automatically are being used increasingly, several conferences since the first GLACE experi-
and are valuable sources of a first map of glaciers ment, and has been included in a document called
in regions where debris-covered glaciers are numer- the GLIMS Analysis Tutorial (Raup and Khalsa
GLACE results 175

Figure 7.8. GLACE 2A outlines viewed over the default imagery in Google Earth. The red lines were produced
before viewing the glacier in Google Earth, and the blue lines were produced after.

2007). The definition is crafted to be specifically Another consequence of these experiments is the
applicable to satellite remote sensing of glaciers. continued development of a series of standard anal-
The definition reads ysis modules and algorithm descriptions that RCs
can use to produce uniform glacier data for
A glacier or perennial snow mass, identified by GLIMS. These standard methods can be implemen-
a single GLIMS glacier ID, consists of a body ted in GLIMSView, though this software is cur-
of ice and snow that is observed at the end of rently primarily used for manual digitization of
the melt season, or, in the case of tropical glacier outlines, surface facies, and glacier center
glaciers, after transient snow melts. This lines. GLIMSView provides a framework for con-
includes, at a minimum, all tributaries and con- sistent production and formatting of outline data,
nected feeders that contribute ice to the main and is extensible for future inclusion of additional
glacier, plus all debris-covered parts of it. processing algorithms (Raup et al. 2007b). Various
Excluded is all exposed ground, including nuna- processing protocols have been discussed in the
taks. An ice shelf shall be considered as a sep- GLIMS community, and have been documented
arate glacier. in some Regional Center work flow guides. All
these documents, and the GLIMS Analysis Tutorial,
The ramifications of this definition, such as how to are available at http://glims.org/
treat steep rock walls that are the source of snow As additional tools are implemented and guide-
that avalanches onto the glacier, are discussed in the lines are developed, we anticipate conducting
GLIMS Analysis Tutorial. The tutorial also docu- further comparative image analysis experiments to
ments recommended practices for the creation of validate the protocols and analysis modules as they
GLIMS data within the context of an appropriate evolve. Within the context of a recent European
data model. glacier-mapping project, the Glaciers Climate
176 Quality in the GLIMS glacier database

have their own preferred software tools, and have


tailored algorithms to the characteristics specific to
glaciers in their own regions. Several processing
protocols have therefore been recommended, each
tailored to a specific set of problems associated with
a particular type of glacier system. Development of
appropriate tools is an ongoing effort. GLIMS held
a workshop in Boulder, Colorado, U.S.A. in June
of 2008 to address these specific topics. The primary
outcome was a more detailed set of guidelines
tailored to different glacier types and software
packages. More information about this workshop
can be found in Racoviteanu et al. (2009).

7.6 GLIMS GLACIER DATABASE AND


THE DATA INGEST PROCESS

Glacier outlines, attributes, and related metadata


are stored at NSIDC in a relational database. The
database software, PostgreSQL with PostGIS
add-ons, is open source, and contains data types
and functions well suited to storing geospatial
and related nongeospatial data. Glacier outlines
are stored as polygons whose vertices are in the
longitude/latitude (geographic) coordinate system
on the WGS-84 datum.

7.6.1 Ingest quality control steps


When a Regional Center has produced a set of
Figure 7.9. Box plots showing the variability of the GLIMS glacier data, it submits the data to NSIDC
glacier area calculated from the glacier outlines pro- via a Web interface that captures metadata on the
duced before (left) viewing the glacier in Google processing steps used. These metadata include
Earth and after (right). The extent of the box in the information on tools used, how geocoding of the
interquartile range, the whiskers extend to the 5th source imagery was done, radiometric calibration,
and 95th percentiles, and outliers are shown as circles.
topographic correction, and the algorithms used for
The thick horizontal line is the median.
classification and interpretation of the imagery.
Several quality control (QC) steps are applied at
NSIDC before final ingest into the database. These
include automated checking for data completeness
Change Initiative (Glaciers_CCI), other compara- and integrity (e.g., existence of necessary IDs and
tive image analysis experiments have been carried other attributes, proper segment order and correct
out and have reached similar conclusions to the circulation direction (handedness) of polygons,
GLACE results (Paul et al. 2012). proper numeric range, polygon closure), and visual-
GLIMSView remains a useful tool for ensuring izing the data on a map and within Google Earth.
adherence to standardization protocols by guiding When problems are found, the RC is contacted with
the analyst through predetermined processing steps a request to fix the problems. Finally, after the data
in the protocol, or through its use as a ‘‘filter’’ have been inserted into the database and become
program, which ensures that certain processing publicly viewable via the web interface, the sub-
steps have been taken before exporting the data into mitter is notified and requested to view the dataset
the data transfer format. We recognize, however, via the GLIMS web map browser and do a final
that different researchers within the GLIMS project check for accuracy.
ASTER data for GLIMS: STARS, DARs, gain settings, and image seasons 177

Figure 7.10. Results of manual glacier delineation performed in GLACE 3A. Lateral boundaries are well identified
by all participants, but there was some disagreement about the details of the terminus, due primarily to differing
interpretations of broken ice.

The first QC steps are performed automatically do not get ingested into the database. Fig. 7.11
by software in the data submission system. Before shows the glacier inventory for British Columbia
any person sees the submission, uploaded data files displayed in Google Earth before ingest. This
are checked for proper formatting, presence and visualization method allows for easy identification
integrity of required attributes, proper coordinate of errors in the dataset. Fig. 7.12 shows an example
system for spatial data (latitude and longitude on where an error in one of the glacier IDs broke the
the WGS-84 datum), proper formatting and link between the outline and its metadata, and sub-
existence of GLIMS glacier identifiers, and proper sequently the outline for that glacier was dropped
closure and data model for glacier boundary by the ingest software. The ingest software issued a
polygons. warning about this, and the visual clue in Google
The quality control steps implemented in the Earth is unmistakable. Fig. 7.13 shows an example
ingest process, in the form of software and proce- from a different region where there was an offset
dures, have been effective in ensuring that bad data between submitted glacier outlines and imagery in
178 Quality in the GLIMS glacier database

Figure 7.11. The 17,585 GLIMS glacier outlines for British Columbia displayed in Google Earth for quality
checking before ingest into the GLIMS Glacier Database.

Figure 7.12. A missing outline for a glacier in British Columbia becomes obvious when displayed in Google Earth.
This error was corrected before ingest. Red lines represent glacier boundaries; green polygons surround rock
outcrops that are internal to the glacier.
GLIMS Glacier Database and the data ingest process 179

Figure 7.13. GLIMS glacier outlines showing a geographic offset in Google Earth. This prompted checking with
the Regional Center. Red lines represent glacier boundaries; green polygons surround rock outcrops that are internal
to the glacier.

Google Earth. In some cases it may be possible that


7.6.2 Representation of
the georeferencing of the imagery in Google Earth
measurement error
is incorrect, but in practice this is rare, and any
offset of the glacier outlines in Google Earth war- For each segment of each glacier outline, the
rants further investigation of the submitted data. GLIMS Glacier Database contains fields that store
The ingest software checks many other items for positional uncertainty. There are four different
basic data integrity, with the result that the data in fields for each polygon segment: ‘‘local’’ and
the GLIMS Glacier Database are consistent in ‘‘global’’ uncertainty in the x and y directions, both
metadata and ID links. Some metadata fields are expressed in meters. Local uncertainty is an esti-
optional, and population of these fields varies, but mate of the location precision of each vertex in
the QC steps in the ingest process strive to ensure the polygon, and is usually directly related to image
high-quality data in all mandatory fields and also resolution, though it can be affected by interpreta-
those optional ones that are populated. tion difficulty, such as at ice flow divides or debris-
As a result of these QC steps, the data that are covered ice at the terminus. Global uncertainty is an
ingested into the GLIMS Glacier Database typic- estimate of the accuracy of the entire segment’s
ally have only the types of errors and uncertainties position, generally related to georeferencing accu-
typified by the best of the outlines in the GLACE racy of the image. These fields are mandatory; they
experiments. Uncertainty is generally three or four cannot be left blank. At the time of ingest, these
pixels (1) in the terminus region, and can be con- numbers are compared with the positions of the
siderably higher (hundreds of meters) in the accu- polygons viewed over imagery as supplied by the
mulation areas at ice–ice boundaries. It is expected analyst and in Google Earth. Polygons are some-
that improvements in the georeferencing of source times also overlaid on Shuttle Radar Topography
imagery and mapping of ice divides will be achieved Mission (SRTM) DEM visualizations as an addi-
as more accurate DEMs become available. tional check.
180 Quality in the GLIMS glacier database

These uncertainty values are determined by the . " AND glacier_dynamic.analysis_id=$id"


. " AND glacier_dynamic.record_status=’okay’))/1000000"
analyst, taking into account image resolution, qual-
. " WHERE analysis_id = $id";
ity (e.g., extent of cloudiness), snow conditions, }
amount of debris cover on the glacier, and ease of else {
determining flow divides, if present, which in turn $update_statement = ’UPDATE glacier_dynamic SET

depends on the quality of ancillary data such as db_calculated_area =’


. ’ (select
DEMs, ground-based photographs, or field-based
sum(st_area(st_transform(glacier_polys,32767)))’
data such as velocity measurements. These consid- . ’ FROM glacier_polygons, glacier_dynamic’
erations are generally documented in the processing . " WHERE line_type=’glac_bound’"
description, also stored within the database. . " AND glacier_polygons.analysis_id=$id"
. " AND glacier_dynamic.analysis_id=$id"
. " AND glacier_dynamic.record_status=’okay’)/1000000"
. " WHERE analysis_id = $id";
7.6.3 Derived parameters in the database
}
As part of the ingest process, it is possible to derive
additional parameters from glacier outlines and The PostGIS function st_transform performs
store these as additional information in the data- the projection, and 32767 is an identifier for the
base. Currently, the primary derived parameter is cylindrical equal area projection.
glacier area. While many Regional Centers provide
the area of each glacier, at ingest time the area of
each glacier is calculated and stored within the
database. This ensures that (1) every glacier has 7.7 CONCLUSION
an associated area stored, and (2) all such areas
are calculated in a consistent manner. The GLIMS community has taken steps to ensure
For each glacier, PostGIS SQL functions are the high quality of data in the GLIMS Glacier
used to project the coordinates to cylindrical equal Database. GLIMS analysis comparison experi-
area (projected meters); calculate the area of the ments have revealed specific potential problems in
polygon(s) tagged with glac_bound—that is, deriving glacier outlines from satellite imagery that
the glacier boundary polygon(s); calculate the area can lead to inconsistent results when building a
of the polygon(s) tagged with intrnl_rock—that database of such outlines from multiple sources.
is, the internal rock (nunatak) boundary poly- The problems are more related to methodological
gon(s); subtract the internal rock area from the area questions during postprocessing than to technical
within the glacier boundary polygon(s) to get the issues of initial image classification. The central
final area for the glacier. question of what constitutes a ‘‘glacier’’ touches
The area calculations are done using the usual most of the problems encountered: treatment of
formula for calculating the area of planar polygons, tributaries and rock outcrops, location of ice
thus an equal area projection must be used. divides, interpretation of debris-covered glacier
This Perl code constructs the correct SQL query, parts and lakes with icebergs, and snowfields that
depending on whether there are internal rock may hide parts of the glacier perimeter or obscure a
polygons: small ice patch completely. The experiments have
led to the development, adoption, and documenta-
if (count_of_internal_rocks($id) > 0) {
tion of definitions, processing protocols, tools, and
$update_statement = ’UPDATE glacier_dynamic SET quality-control steps that have improved the con-
db_calculated_area =’ sistency and quality of glacier data going into the
. ’ ((select database. After the documents were distributed to
sum(st_area(st_transform(glacier_polys,32767)))’
the GLIMS community, the analysis quality was
. ’ FROM glacier_polygons, glacier_dynamic’
. " WHERE line_type=’glac_bound’"
observed at the data ingest stage to have improved.
. " AND glacier_polygons.analysis_id=$id" We estimate that glacier outline digitization repeat-
. " AND glacier_dynamic.analysis_id=$id" ability (1) is of the order of 3 to 4 pixels (45 to 60 m
. " AND glacier_dynamic.record_status=’okay’)" for ASTER) in regions where interpretation is
. ’ - (select
straightforward, but uncertainties can remain much
sum(st_area(st_transform(glacier_polys,32767)))’
. ’ FROM glacier_polygons, glacier_dynamic’
higher for individual glaciers where interpretation
. " WHERE line_type=’intrnl_rock’" is difficult (e.g., ice flow divides in regions without
. " AND glacier_polygons.analysis_id=$id" proper DEMs). Incorporation of topographic
References 181

information into the work flow is thus crucial to Paul, F., and Kääb, A. (2005) Perspectives on the produc-
reduce this uncertainty. As tools, protocols, and tion of a glacier inventory from multispectral satellite
data availability evolve, more GLACE tests will data in the Canadian Arctic: Cumberland Peninsula,
likely be carried out. Baffin Island. Annals of Glaciology, 42, 59–66.
Paul, F., Barry, R., Cogley, J., Frey, H., Haeberli, W.,
Ohmura, A., Ommanney, C., Raup, B., Rivera, A.,
and Zemp, M. (2009) Recommendations for the com-
7.8 ACKNOWLEDGMENTS
pilation of glacier inventory data from digital sources.
Annals of Glaciology, 50(53), 119–126.
The GLIMS initiative at the NSIDC was begun
Paul, F., Barrand, N., Berthier, E., Bolch, T., Casey, K.,
with the support of NASA awards NNG04GF51A
Frey, H., Joshi, S., Konovalov, V., Bris, R.L., Moelg,
and NNG04GM09G. We would like to thank the N. et al. (2012) On the accuracy of glacier outlines
late Mark Dyurgerov, Paul Geissler, Christian derived from remote sensing data. Annals of Glaciol-
Georges, Chris Helm, Ella Lee, and Claudia Riedl ogy, 54(63), 171–182.
for their involvement in the GLACE experiments. Racoviteanu, A.E., Paul, F., Raup, B., Khalsa, S.J.S.,
ASTER data courtesy of NASA/GSFC/METI/ and Armstrong, R. (2009) Challenges and recommen-
Japan Space Systems, the U.S./Japan ASTER dations in mapping of glacier parameters from space:
Science Team, and the GLIMS project. Results of the 2008 Global Land and Ice Measure-
ments from Space (GLIMS) workshop, Boulder,
Colorado, USA. Annals of Glaciology, 53, 53–69.
7.9 REFERENCES Raup, B., and Khalsa, S.J.S. (2007) GLIMS Analysis
Tutorial. National Snow and Ice Data Center, Boulder,
Abrams, M., Hook, S., Ramachandran, B. (2002) Aster CO. Available at http://glims.org/MapsAndDocs/
User Handbook, Version 2. NASA Jet Propulsion guides.html
Laboratory, Pasadena, CA. Raup, B., Khalsa, S., Armstrong, R., Cawkwell, F.,
Albert, T.H. (2002) Evaluation of remote sensing tech- Georges, C., Hamilton, G., Sneed, W., Jr., and
niques for ice-area classification applied to the tropical Wheate, R. (2004) Comparative image analysis to
Quelccaya Ice Cap, Peru. Polar Geography, 26(3), 210– ensure data quality in the global land ice measurements
226. from space (GLIMS) glacier database. EOS Trans.
Alt, H., Behrends, B., and Blömer, J. (1995) Approximate Am. Geophys. Union, 85(47), Supplement, abstract
matching of polygonal shapes. Annals of Mathematics H23D-1151.
and Artificial Intelligence, 13, 251–265. Raup, B., Kääb, A., Kargel, J.S., Bishop, M.P., Hamil-
Bishop, M.P., Bonk, R., Kamp, U., and Shroder, J. ton, G., Lee, E., Paul, F., Rau, F., Soltesz, D., Khalsa,
(2001) Terrain analysis and data modeling for alpine S.J.S. et al. (2007a) Remote sensing and GIS technol-
glacier mapping. Polar Geography, 25(3), 182–201. ogy in the Global Land Ice Measurements from Space
Hall, D.K., Riggs, G., and Salomonson, V. (1995) Devel-
(GLIMS) project. Computers and Geosciences, 33, 104–
opment of methods for mapping global snow cover
125, doi: 10.1016/j.cageo.2006.05.015.
using moderate resolution imaging spectroradiometer
Raup, B., Racoviteanu, A., Khalsa, S., Helm, C.,
data. Remote Sensing of Environment, 54(2), 127–140.
Armstrong, R., and Arnaud, Y. (2007b) The GLIMS
Paul, F. (2007) The New Swiss Glacier Inventory 2000:
Application of Remote Sensing and GIS (Schriftenreihe geospatial glacier database: A new tool for studying
Physische Geographie, Glaziologie und Geomorpho- glacier change. Global and Planetary Change, 56, 101–
dynamik No. 52). Universität Zürich, 210 pp. 110, doi: 10.1016/j.gloplacha.2006.07.018.
Paul, F., Kääb, A., Maisch, M., Kellenberger, T., Sannel, A.B.K., and Brown, I.A. (2010) High-resolution
Haeberli, W. (2002) The new remote-sensing-derived remote sensing identification of thermokarst lake
Swiss glacier inventory, I: Methods. Annals of Glaciol- dynamics in a subarctic peat plateau complex.
ogy, 34, 355–361. Canadian Journal of Remote Sensing, 36(Suppl. 1),
Paul, F., Huggel, C., and Kääb, A. (2004) Combining S26–S40.
satellite multispectral image data and a digital eleva- Sneed, W.A. (2007) Satellite remote sensing of Arctic
tion model for mapping debris-covered glaciers. glacier–climate interactions. Master’s thesis,
Remote Sensing of Environment, 89, 510–518. University of Maine.
182 Quality in the GLIMS glacier database

Table 7.5. Participants in GLACE 1, GLACE 2, GLACE 2A, and GLACE 3A, and their affiliations at the time of the
experiments.

Regional Center Number Institution, participants Experiments


(Steward nmber)

3 University of Alberta, Canada; Fiona CAWKWELL 1, 2, 3A

3 (536) University of Northern British Columbia, Canada; Roger WHEATE


and Brian MENOUNOS 1, 2

6 University of Maine, U.S.A.; Gordon HAMILTON and Bill SNEED 1, 2, 2A

—a Cambridge University; Narelle BAKER 3A

8 CAREERI, Lanzhou, China; Guodong CHENG, Shiyin LIU,


Xin LI, Donghui SHANGGUAN 1

5 Portland State University, U.S.A.; Matthew HOFFMAN 3A

11 University of Innsbruck, Austria; Claudia RIEDL, Helmut ROTT 1, 2

11 (507) University of Zurich, Switzerland; Frank PAUL 1, 2

13 University of Innsbruck, Austria; Christian GEORGES 1

602 University of Colorado, U.S.A.; Matthew BEEDLE 2

17 Russian Academy of Sciences, Russia; Vladimir KONOVALOV 2

—a Texas A&M University, U.S.A.; Andrew KLEIN, Joni KINCAID 2

604 U.S. Geological Survey (Flagstaff, AZ), U.S.A.; Ella LEE,


Paul GEISSLER 2A

606 University of Arizona, U.S.A.; Jeff KARGEL 2A

—a University of Alaska, U.S.A.; Christopher LARSEN 2A

602 University of Colorado, U.S.A.; Bruce RAUP 2A, 3A

602 University of Colorado, U.S.A.; Christopher HELM 2A

—a U.S. Geological Survey (Reston, VA), U.S.A.; Bruce MOLNIA 2A

—a University of Colorado, U.S.A.; Mark DYURGEROV 2A

—a University of Otago, New Zealand, Shelley MACDONELL 3A


a
Dashes indicate no assignment yet to a Regional Center.

Vous aimerez peut-être aussi