Content uploaded by Andrew G. Klein
Author content
All content in this area was uploaded by Andrew G. Klein on Sep 05, 2014
Content may be subject to copyright.
64th EASTERN SNOW CONFERENCE
St. John’s, Newfoundland, Canada 2007
85
A Reassessment of the Satellite Record of
Glacier Change in the Rwenzori Mountains, East Africa
ANDREW G. KLEIN1 AND JONI L. KINCAID2
ABSTRACT
Three massifs in Eastern Africa currently have glaciers. Glaciers in two of these, Mt. Kenya and
Mt. Kilimanjaro, have received extensive study. The third, the Rwenzori Mountains on the border
of Uganda and the Democratic Republic of the Congo has been the focus of less research, but
some recent studies have been undertaken. Glaciers on all three massifs have been shrinking in
recent decades. A recently published study has examined glacier retreat in the Rwenzori from
1987 to 2003 using Landsat satellite images. Our analysis of the images used in this study,
however, reveals that the 1995 and 2003 images contain significant snow outside of the glaciers
and therefore are unreliable indicators of glacier extent. Using a combination of Landsat, ASTER
and SPOT images of the Rwenzori glaciers, ice areas have been reevaluated for the period 1987 to
2006. The Normalized Difference Snow Index (NDSI) and visual mapping were used to determine
the glacier areas. Our analysis indicates that the glaciers in the Rwenzori have decreased in area
from 2.55 km2 in 1987 to 1.31 km2 in 2006. These areas, like previous estimates, are not without
their own uncertainties.
Keywords: Remote Sensing, Glaciers, Tropics, Africa
INTRODUCTION
There are currently three massifs that support glaciers in Africa. Two of these ranges, Kibo on
Kilimanjaro and Mt. Kenya have been the site of exhaustive studies. The third, the Rwenzori
Range, which lies on the border of the Democratic Republic of the Congo and Uganda (Figure 1)
has received much less study though some work has recently been published (Kaser and
Osmaston, 2002; Mölg et al., 2003; Taylor et al., 2006). However, despite recent work, the
measured glacier retreat, as well as the interpretation of the climatic factors responsible for the
retreat since the 1980s remains controversial.
Recently, Taylor et al. (2006) published a glacier retreat time series for the Rwenzori Range of
Eastern Africa based on analysis of Landsat Thematic Mapper images and published estimates of
glacier extent from field observations (Kaser and Osmaston, 2002). To extract glacier extent from
Landsat images in their study Taylor et al. (2006) used both a supervised classification and the
Normalized Difference Snow Index (NDSI). Based on their analysis, Taylor et al. (2006) assert
that glaciers in the Rwenzori have experienced a constant retreat rate since 1905 when the first
mapping of the glaciers was undertaken.
1 Department of Geography, MS 3147, Texas A&M University, College Station, TX 77853-
3147, tel: 979 845-5219, email: klein@geog.tamu.edu
2 Department of Geography, MS 3147, Texas A&M University, College Station, TX 77853-
3147, tel: 512-431-9963, email: jkincaid@geog.tamu.edu
86
Taylor et al. (2006) attributed the measured ice loss to increasing air temperatures without
significant changes in precipitation. The climatic interpretation, and to a lesser degree, the glacier
mapping has been challenged (Molg et al., 2006). Taylor et al. (2006) responded to the criticisms,
especially on the climatic interpretation, but did not address in detail the problems with the
satellite derived time series of glacier retreat.
Figure 1. 1987 Landsat TM image (bands 542) of the Rwenzori Mountain Range,
showing the three currently glaciated peaks of Mount Stanley, Mount Speke and Mount Baker.
In this study, we provide an extended analysis of the satellite image record that highlights
inaccuracies of previous analysis. While discussion of the climate interpretations drawn from their
analysis is beyond the scope of this research, this work provides an alternative analysis of the
satellite image record of glacier change in the Rwenzori. We believe our time series to be a more
accurate depiction than that in Taylor et al. (2006) of the actual changes in snow and glacier cover
captured in the satellite record.
STUDY AREA AND PREVIOUS GLACIER MAPPINGS
Glaciers in the Rwenzori range are currently confined to three peaks in the range: Mt. Stanley,
Mt. Speke and Mt. Baker (Figure 2). The highest altitude of these peaks is in excess of 4800 m
with Mt. Stanley exceeding 5000 m. (Mt. Stanley: 5111 m, Mt. Speke: 4891 m and Mt. Baker:
4873 m). Published estimates of glacier extent based on field observations are limited to three
times: 1906, 1955, and 1990. These glacier areas are presented, as revised by Kaser and Osmaston
(2002) in Table 1. The glacier extents for these dates are illustrated in Figure 3. In 1990, glaciers
in the Rwenzori extended from approximately 4400 m to over 5000 m. A thorough discussion of
the Rwenzori and its glaciers can be found in Kaser and Osmaston (2002).
87
Table 1. Glacier areas (km2) in the Rwenzori (Table 6.6.1 from Kaser and Osmaston, 2002).
Peak 1906 1955 1990
Mt. Baker 1.47 0.62 0.12
Mt. Speke 2.18 1.31 0.56
Mt. Stanley 2.85 1.88 1.00
Total 6.51 3.18 1.67
Figure 2. An ASTER false color composite image (bands 4, 3 and 2) of the Rwenzori showing the three
currently glacierized massifs: Mt. Stanley, Mt. Speke and Mt. Baker.
METHODS
Glacier extents were mapped using the NDSI method as well as through visual mapping. Linear
spectral unmixing was also attempted, but problems caused by shadowing made the technique less
reliable than the other two methods in determining glacier extent.
Several major limitations exist to using remote sensing to map glacier retreat in the Rwenzori.
Due to persistent cloud cover the image archive is quite limited and the limited series of images
selected for analysis is illustrated in Figure 4. It is evident from Figure 4, that several dates, most
notably those in 1995 and 2003, are impacted by transient snow cover.
88
Figure 3. Mapped glacier extents of the Rwenzori for 1906, 1955 and 1990 from Kaser and Osmaston (2002)
overlain on an ASTER false color composite (bands 4, 3 and 2) from 2005.
89
Figure 4. False color composites of all images selected for analysis. Each false color composite is composed
of a mid-infrared band (red), a near-infrared band (green) and a visible band (blue).
Image Processing
All Landsat, ASTER and SPOT scenes used in this study were processed in a similar manner to
assure as comparable glacier mapping results as possible using the applied mapping techniques.
Geometric Rectification
To facilitate comparisons between images, all images were geolocated. Initially, all Landsat and
the 2005 ASTER image were coregistered using a polynomial warp to the ASTER image.
Subsequently, this original geometric rectification, an improved digital elevation model and the
ASTER and Landsat images were shifted several tens of pixels in the Y direction to bring them
into better alignment with the Digital Elevation Model (DEM). To bring the SPOT image into
alignment with the other images it was necessary to orthorectify the image. Orthorectification was
accomplished using the ENVI 4.3 image processing software along with the DEM and ground
control points selected from the 2005 ASTER image.
Radiometric Preprocessing
For all image types, the original DN (digital numbers) values were converted to at-satellite
radiances (W m-2 sr-1 μm-1). This was accomplished using the gains and offsets provided by the
90
data providers. For ASTER images, this step was accomplished using the automated procedures
provided by ENVI 4.3 while for Landsat and SPOT the gains and offsets were individually input
for each sensor and band.
The radiance values were then converted to at-satellite reflectance using the general approach:
()
θ
π
λ
λ
λ
cos*
2
sun
E
dL
R= (1)
where Rλ is top of the atmosphere reflectance, Lλ is at-satellite radiance, d is the earth-sun distance
in astronomical units at the time of image acquisition, Esun is mean solar exoatmospheric
irradiance in each band, and θ is the solar zenith angle. Esun values were determined individually
for each sensor using published values.
Following the conversion to at satellite-reflectance, the modified black-body subtraction method
developed by Chavez (1988) was used to perform a simple atmospheric correction to each image
as detailed information on the state of the local atmosphere was not available. For ASTER and
Landsat the average reflectance from a 3x3 pixel box in high altitude Lake Bujuku were used to
determine the correction for the shortest wavelength band, while corrections for the remaining
bands were calculated based on a power law scattering model. For the SPOT image, the
reflectance from the smaller Lac du Speke was used due to haze over Lake Bujuku at the time of
image acquisition. For Landsat images a clear sky scattering model was employed (λ-3), for
ASTER a very clear model (λ-4) and for SPOT, that had extensive clouds, a hazy model (λ-1) was
used.
Normalized Snow Difference Index
Following the atmospheric correction, the Normalized Snow Difference Index (NDSI) was
computed for each image in the following manner (Hall et al., 1995) using appropriate bands for
each sensor:
(
)
()
Band A -Band B
NDSI Band A Band B
=+ (2)
The bands representing Bands A and B for each sensor are listed in Table 2
Table 2. Bands used for computing the NDSI..
Sensor Band A Band B
Landsat 2 (0.52-0.60 μm) 5 (1.55-1.75 μm)
ASTER 2 (0.63-0.69 μm) 4 (1.600-1.700 μm)
SPOT 2 (0.617-0.687 μm) 4 (1.580-1.750 μm)
Due in part to the larger atmospheric influences in Band 1 of ASTER and SPOT, Band 2 from
both of these instruments was used to compute the NDSI, rather than a shorter wavelength band
that corresponded slightly better to Band 2 of Landsat.
Thresholds were then selected to create a binary snow/ice map from each NDSI image. The
selected thresholds were based on a thorough visual analysis of each image including comparisons
of the binary snow maps with the original satellite images, our experience in other tropical
mountain ranges and in the case of Landsat, extensive testing of the approach for snow mapping
(Hall et al., 1995; Klein et al., 1998). The following thresholds were selected: 0.40 for Landsat and
0.10 for both ASTER and SPOT. To ascertain the impact of the selected threshold, additional
snow maps were also created by adjusting the threshold by 0.1 NDSI units above and below the
selected NDSI.
91
A mask was then created by buffering the 1906 glacier extents by 50 m. This mask was applied
to each snow map to remove erroneously mapped snow falling well outside the glacier boundaries.
For some image dates, some snow that did exist outside the buffered area may have been removed.
In addition, for the 2003-12-17 Landsat image clouds that were erroneously classified as snow by
the NDSI method were removed manually. The manual removal was conservative and snow totals
on Mt. Baker for this date may include some cloud cover pixels.
Visual Mapping
For the three dates with minimal snow cover, visual mapping of the glaciers of the Rwenzori
was also undertaken in a manner analogous to that used to map glaciers on Puncak Jaya in Papua,
formerly Irian Jaya (Klein and Kincaid, 2006). False color composites were created for the 1987
Landsat, 2005 ASTER and 2006 SPOT images from the following mid-infrared, near-infrared and
visible bands: Landsat 5, 4 and 2, ASTER 4, 3 and 2 and SPOT 4, 3 and 2.
The glaciers on the three main peaks comprising the Rwenzori were then mapped. Group
discussions of problematic areas resulted in the following process to produce the final maps. The
visual mapping was done in time sequential order and in some instances glacier extents from a
previous or subsequent image were used to help constrain the mapping. Following completion of
the mapping, small shifts were applied to the mapped polygons to bring the borders into alignment
with the boundaries mapped on the 2006 SPOT image for consistency. In order to place overly
conservative estimates of the error of our glacier mapping, the digitized glacier boundaries were
contracted and expanded by 15 m using a buffering operation within a Geographic Information
System (GIS). The areas of these shrunken and enlarged glacier extents are also listed in Table 3.
RESULTS AND DISCUSSION
The computed glacier areas for the three glaciated peaks in the Rwenzori are listed in Table 3 as
are the areas from Taylor et al. (2006) for all images studied. It is apparent that substantial glacier
retreat has occurred between the late 1980s and early 2000s in the Rwenzori. However,
successfully quantifying this retreat using the existing satellite image archive is problematic.
These problems are evident in the large differences between the mapped areas for the same images
between this study and Taylor et al. (2006). Slight differences in glacier areas are also found
between mapping methods.
Snow Cover Problems
Unfortunately, with optical images it is difficult, if not impossible, to distinguish between a
pixel covered with transient snow and one containing snow or ice on a glacier surface. As is
evident in Figure 4, three of the Landsat images in the time series (January 17, 1995, January 31,
2003 and December 17, 2003) all contain significant snow cover off of the glaciers. These images
are not suitable for analysis of glacier areas. However, images from two of these dates from
January 17, 1995 and January 31, 2003 were used by Taylor et al. (2006) to determine glacier
extents and are thus included in Table 3 although not used in our own analysis. While it is possible
to use these images to qualitatively infer glacier retreat of some of the main glaciers, given the
extensive snow cover existing off of the glaciers, the published glacier areas of Taylor et al.
(2006) for these two dates are highly suspect.
Comparisons with Previously Published Studies
As is evident in Table 4, there are large differences in the computed glacier areas for the
individual peaks between our analysis and the published glacier areas from Taylor et al. (2006).
Given the limited detail on the processing methods provided by Taylor et al. (2006) it is difficult
to assess the origin of these large differences. Nevertheless, there are large differences in areas
even though both studies employed one common mapping method – the Normalized Snow
Difference Index (NDSI).
92
Overall, the agreement between the Taylor et al. (2006) and our own analysis for the 1987
Landsat image is closer than for the later Landsat dates where snow cover is more extensive. The
smaller ice area on Mt. Baker (0.147 km2) and the larger ice area on Mt. Stanley indicated by our
NDSI methods are not within their stated error bounds (Table 3). Again the source of this
discrepancy is unknown.
For the 1995 and 2003 Landsat images, the NDSI estimated areas of Taylor et al. (2006) for Mt.
Speke and Mt. Stanley, as well as for the total glacier areas for the Rwenzori, are considerably less
than our own best estimates. For example, the total glacier areas for the Rwenzori determined by
Taylor et al. (2006) are much smaller than our own, 0.822 km2 (35%) and 1.194 km2 (55%) for the
1995 and 2003 Landsat images both groups analyzed.
With the exception of some snow covered pixels possibly falling outside of the 1906 glacier
boundaries, the total areas of this study include all pixels with NDSI values > 0.40. For Taylor et
al. (2006) to arrive at glacier areas considerably smaller than our own would require masking of a
considerable portion of the snow-covered pixels or use of a considerably more restrictive NDSI
threshold then the one we employed. Given the large differences in snow-covered areas using the
same technique, caution should be given to any glacier areas derived from 1995 and 2003 Landsat
images.
Our analysis of glaciers in the 2005 ASTER satellite image agree quite well with the published
areas for these glaciers by Molg et al. (2006) who mapped glaciers from the same ASTER image
using a combination of visible bands. The published areas, Mt. Baker 0.04±0.01, Mt. Speke
0.30±0.03 and Mt. Stanley 0.80±0.06, are in general, slightly larger on the order of 0.05 km2 for
each peak than our own (Table 3). The comparability of mapped areas gives us confidence in our
own mapped areas for other dates.
Visual Mapping
Of the four image dates, only three were considered minimally contaminated with snow to
accurately map glacier boundaries. From these three images: Landsat 1987-08-07, ASTER 2005-
02-21 and SPOT 2006-02-10 glacier extent was mapped visually (Table 3). The 2006 ASTER
image was not mapped because it was felt the 2006 SPOT image was of higher quality and little
would be gained by digitizing both images for 2006. The visually mapped glacier extents are
illustrated in Figure 5. The marked retreat of the glaciers from 1987 to 2005 is evident.
Comparing our visual mapping and our best estimates of glacier area for the same dates using
the Normalized Difference Snow Index, we can see that overall the areas are quite comparable. In
all but the glacier area mapped for Mt. Baker from the 2006 SPOT image, the visually mapped
areas are all slightly higher than our best estimates from NDSI. A major source of the differences
is that areas of high shadow are included in the visual maps but the NDSI values are too low for
the pixels to be mapped as snow using this method. This highlights the necessity of using different
methods to assess glacier areas in areas of complex topography and heavy shadows.
In order to identify potential problem areas with our visual mapping, our 1987 mapping was
also compared to the 1990 glacier extent maps from Kaser and Osmaston (2002) as is illustrated in
Figure 6. This comparison yields some interesting results. The first is that if snow/ice covered
pixels in the Landsat image do represent only pixels containing glacier ice, then considerable
retreat occurred between 1987 and 1990 on all peaks, especially on Mt. Stanley.
However, an alternative, and likely hypothesis, is that some of the snow covered pixels in the
1987 Landsat image represent snow in areas outside the glacier. To examine this we further
compared our visually mapped 1987 glacier extents to the reanalyzed 1955 extents from Kaser and
Osmaston (2002). Despite some geolocation errors, it is clear that the 1987 image contains snow
covered pixels that lie outside the 1955 glacier boundaries. For Mt. Baker and Mt. Speke, the total
areas are under 0.1 km2: 0.049 km2 and 0.072 km2, for Mt. Baker and Mt. Speke, respectively.
However, for Mt. Stanley, the area mapped as snow/ice in 1987, but not as glacier in 1955 is 0.335
km2. This would indicate that quite possibly our estimated glacier area for Mt. Stanley in 1987 is
too large.
It is also interesting to note that on all peaks, but primarily Mt. Stanley and Mt. Speke there are
considerable areas mapped as glacier in 1990, but not in 1987. In some areas such as the eastern
93
slopes of Mt. Stanley, it is clear that areas mapped as containing glacier in 1990 clearly did not
contain snow or ice in 1987 and may indicate problems in the mapped 1990 extents. The same is
true for Mt. Speke. In addition, there exists several small ice areas that while mapped as glaciers in
both 1987 and 1990 are offset geographically.
Table 3. Glacier areas computed for the individual peaks of the Rwenzori.
NDSI1 Visual Mapping2 Taylor3
−0.1 Center +0.1 −15m Mapped +15m Supervised NDSI
1987-08-07 – Landsat
Mt. Baker 0.253 0.233 0.206 0.172 0.273 0.388 0.43±0.13 0.38±.04
Mt. Speke 0.687 0.644 0.588 0.587 0.770 0.968 0.65±0.18 0.63±.0.2
Mt. Stanley 1.249 1.201 1.134 1.334 1.511 1.834 1.03±0.25 1.00±0.05
Total 2.189 2.081 1.928 2.093 2.554 3.19 2.11±0.56 2.01±0.11
1995-01-17 – Landsat
Mt. Baker 0.198 0.159 0.125 0.21±0.06 0.20±0.09
Mt. Speke 1.161 0.989 0.755 0.45±0.11 0.44±0.17
Mt. Stanley 1.358 1.174 0.986 0.69±0.15 0.86±0.10
Total 2.717 2.322 1.8666 1.35±0.32 1.50±0.36
2003-01-31 – Landsat
Mt. Baker 0.156 0.128 0.102 0.16±0.05 0.11±0.03
Mt. Speke 0.893 0.812 0.678 0.40±0.09 0.35±0.11
Mt. Stanley 1.321 1.214 1.103 0.53±0.09 0.50±0.20
Total 2.370 2.154 1.883 1.09±0.22 0.96±0.34
2003-12-17 – Landsat
Mt. Baker 0.038 0.017 0.008
Mt. Speke 1.326 1.1211 1.053
Mt. Stanley 1.818 1.723 1.578
Total 3.182 2.951 2.639
2005-02-21 – ASTER
Mt. Baker 0.035 0.027 0.018 0.014 0.033 0.058
Mt. Speke 0.261 0.208 0.173 0.147 0.231 0.330
Mt. Stanley 0.762 0.694 0.625 0.547 0.725 0.922
Total 1.058 0.929 0.816 0.708 0.989 1.31
2006-02-10 – SPOT
Mt. Baker 0.043 0.034 0.022 0.015 0.030 0.052
Mt. Speke 0.273 0.191 0.123 0.139 0.222 0.320
Mt. Stanley 0.858 .0689 0.558 0.521 0.723 0.936
Total 1.174 0.914 0.703 0.675 0.975 1.308
2006-09-04 – ASTER
Mt. Baker 0.086 0.056 0.044
Mt. Speke 0.450 0.335 0.261
Mt. Stanley 0.819 0.748 0.678
Total 1.355 1.139 0.983
1NDSI center values vary by sensor. The NDSI used to compute the center value for Landsat is 0.4, for
ASTER and for SPOT, while the -0.1 and +0.1 columns represent the areas computed for NDSI values
decreased and increased by 0.1 NDSI values for each sensor, respectively
2For the visual mapping the mapped column represents the area for the mapped boundaries. The -15m
and +15m represent the areas in which the entire boundary has been eroded and dilated by 15 m,
respectively.
3Values from Taylor et al. (2006) Table 1.
94
Geolocation errors are the cause of several of the observed differences. Such geolocation errors
and problems in combining satellite-based glacier maps with ground based maps has been noted in
other locations (e.g., Klein and Kincaid, 2006). These problems also point to the desirability of
orthorectification of the Landsat images to enable better comparisons. In general, this study also
highlights that glacier extents mapped from sources other than satellite images may have their own
problems.
Figure 5. Glacier extents visually mapped from 1987
Landsat, 2005 ASTER and 2006 SPOT images area.
Figure 6. Comparison of our visual mapping of glacier
extents in 1987 (this study) and the 1990 mapped extent
(Kaser and Osmaston, 2002).
Summary of Glacier Retreat in the Rwenzori
Based on previous studies and our own analysis, our interpretation of the glacier retreat of the
Rwenzori glaciers follows. Figure 7 illustrates both previously published estimates and our own
areas from visual mapping. Our visual mapping may contain inaccuracies, but we feel it is the best
estimate of glacier areas that can be derived from the satellite archive.
Our larger estimate of glacier extent in 1987 than Taylor et al. (2006) is due to our inclusion of
highly shadowed areas that appear to contain ice, but whose NDSI values are too low to be
classified as ice spectrally. In addition, the glacier areas indicated by any remote sensing method
in 1987 may be incorrect due to snow cover existing on non-glacier surfaces, especially on Mt.
Stanley. The apparent large decrease in total glacier area between 1987 and 1990 can thus be due
to an overestimated glacier area in 1987 and, possibly, with problems in the reanalyzed 1990
mapping as well.
The Taylor et al. (2006) glacier areas for 1995 and 2003 cannot be considered reliable as it is
obvious that snow cover exists off of the glaciers themselves. Our own analysis of these images
shows the actual are of snow ice is much larger than the areas published by Taylor et al. In
95
addition, it would appear based on the 2005 and 2006 images, that the Taylor et al. (2006)
estimates for Mt. Stanley are smaller than the areas measured in 2005 and again in 2006.
Figure 7. Retreat of the Rwenzori Glaciers with previously published information
from Kaser and Osmaston (2002) [white circles] and Taylor et al. (2006) [gray circles].
Our own estimates from visual mapping are shown as black circles.
CONCLUSIONS
From 1987 to 2005, considerable glacier retreat is evident in the Rwenzori Range in eastern
Africa based on satellite images. However, quantifying this retreat is difficult given that few
cloud-free images exist and some of those are severely impacted by cloud cover. We reiterate the
conclusion of Mölg et al. (2006) that the Taylor et al. (2006) analysis does not meet the accuracy
requirements for glacier mapping in the tropics.
ACKNOWLEDGMENTS
This work was funded by NASA Grant 02000-015. The authors wish to thank Georg Kaser,
Thomas Mölg and Thomas Gumbricht for providing us with data used in this reanalysis.
REFERENCES
Chavez PS, 1988. An improved dark-object subtraction technique for atmospheric scattering
correction of multispectral data. Remote Sensing of Environment 24: 459-479.
Hall DK, Riggs GA, Salomonson VV. 1995. Development of methods for mapping global snow
cover using moderate resolution imaging spectroradiometer data. Remote Sensing of
Environment 54: 127-140.
Kaser G, Osmaston H. 2002. Tropical Glaciers. Cambridge University Press, Cambridge 207 p.
Klein AG, Hall DK and Riggs GA. 1998. Improving snow-cover mapping in forests through the
use of a canopy reflectance model. Hydrological Processes 12: 1723-1744.
Klein AG, Kincaid JL, 2006. Retreat of glaciers on Puncak Jaya, Irian Jaya, determined from 2000
and 2002 IKONOS satellite images. Journal of Glaciology 52: 65-79.
96
Mölg T, Rott H, Kaser G, Fischer A, Cullen NJ. 2006. Comment on "Recent glacial recession in
the rwenzori mountains of east Africa due to rising air temperature'' by Richard G. Taylor,
Lucinda Mileham, Callist Tindimugaya, Abushen Majugu, Andrew Muwanga, and Bob
Nakileza. Geophysical Research Letters 33. DOI: 10.1029/2006GL027254.
Mölg T, Georges C, Kaser G. 2003. The contribution of increased incoming shortwave radiation
to the retreat of the Rwenzori glaciers, East Africa, during the 20th century. International
Journal of Climatology 23: 291-303.
Taylor RG, Mileham L, Tindimugaya C, Majuga A, Muwanga A, Nakileza B, 2006. Recent
glacier recession in the Rwenzori Mountains of East Africa due to rising air temperature.
Geophysical Research Letters 33: DOI: 10.1029/2006GL027606.