Glacier Variations in the Fedchenko Basin, Tajikistan, 1992–2006: Insights from Remote-sensing Images

Abstract The Fedchenko glacier system harbors the longest glacier (77 km) outside of the polar areas. It has a great influence on the local economy and ecology as a water resource. The glacier area in the Fedchenko basin was investigated using Landsat TM images in 1992, Landsat ETM+ images in 2000, and ALOS/AVNIR2 images in 2006. The total area was 835.9 km2 in 1992, 830.0 km2 in 2000, and 864.8 km2 in 2006, increasing by 3.46% from 1992 to 2006 (an average rate of 0.25% per year). The Fedchenko glacier system is relatively stable when considering glacier surging and data errors. Reanalysis data (eg for temperature and precipitation) from the US National Centers for Environmental Prediction and National Center for Atmospheric Research suggest that increasing precipitation since 1990 probably caused the slight change in glacier area.

Most glaciers (surging and nonsurging) on the Karakoram Mountains are advancing or stable, which is consistent with the gain of mass during 2000 to 2008 (Gardelle et al 2012). An increase in winter precipitation since 1961 is a potential source of greater accumulation in the upper parts of glaciers (Gardelle et al 2012). The eastern Pamir also show positive mass balance, which has been attributed to increasing precipitation since 1979, due to strengthened westerlies (Yao et al 2012). Glacier investigations in the Pamir Plateau are still scarce, especially for the Fedchenko glacier system (Kutuzov and Shahgedanova 2009).
Generally, the foothills receive less precipitation, while mountains with high elevation can intercept water vapor carried by westerlies, store it in the form of glaciation, and release water by glacier melting. Water from glaciers irrigates farmland and supplies other anthropogenic activities in the arid area of the Pamir region (Barnett et al 2005). Thus, there is an urgent need to investigate the variation of glaciers in the region under ongoing climate change and accurately evaluate its effect on local water resources (Kutuzov and Shahgedanova 2009). This article investigates variations in glacier area in the Fedchenko basin from 1992 to 2006 and discusses the relationship between glacier variation and climate change.

Study area
The extent of the Fedchenko glacier system (38u509N; 72u159E) ( Figure 1A) has been estimated in different ways, including 649 km 2 by Aizen et al (2009) and 992 km 2 by the Japan Aerospace and Exploration Agency (2006). This study defines the Fedchenko glacier system based on a specific pour point ( Figure 1B). The system contains

Systems knowledge
Mountain Research and Development (MRD) An international, peer-reviewed open access journal published by the International Mountain Society (IMS) www.mrd-journal.org many tributary glaciers, such as the Bivouac and Nalifikin glaciers, as well as many glaciers that have been disconnected since the Little Ice Age. The longest alpine dendrite glacier outside of the polar regions, the Fedchenko glacier system has a length of 77 km, an elevation of 2900-6300 m above sea level (asl) (Aizen et al 2009), and an average ELA of 5000 m asl. In the system, many dark-brown debris-covered glaciers and surging lobes can be found on Landsat Enhanced Thematic Mapper plus (ETM+) remote-sensing images in a composite of bands 5, 4, and 3 ( Figure 1C). In winter, moisture comes from the Atlantic Ocean and contributes FIGURE 1 The Fedchenko glacier system: (A) location; (B) constitution; and (C) partial map of delineation on an ETM+ image. Threshold for NDSI is 0.5; 1 and 2 refer to the debris-covered glaciers, which were manually delineated; 3 and 4 refer to seasonal snow, which was erased from the image. A false-color composite image (bands 5, 4, and 3) of ETM+ appears in the background. The melt water of glaciers flows through the Muksu, Surkhob, Vakhsh, and Amu Darya Rivers, eventually discharging into the Aral Sea. It feeds densely populated and industrial areas along the rivers. Thus, glacier variations in the Fedchenko basin have a great influence on ecology, agriculture, and economic development in these areas. Sustainable strategies, such as encouraging reasonable water consumption and establishing effective irrigation systems, have been carried out along the Amy Darya River (Micklin 2006).

Data and methodology
Landsat and Advanced Land Observing Satellite (ALOS) images were used in this study (Table 1). Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper plus (ETM+) images were available from the Global Land Cover Facility (2013); these images have already been orthorectified to Level 1G products (UTM projection Zone 43N, WGS 1984 datum). The ALOS images with rational polynomial coefficients (RPC) files were bought from the Remote Sensing Technology Center of Japan; they were L01B products and coarsely geometrycorrected. Cloud cover obscured less than 1% of the total image areas; the acquisition time was at the end of the ablation season, and seasonal snow covered the lowest yearly area. Thus, snow and clouds did not greatly affect our ability to extract glacier outlines.
The Shuttle Radar Topography Mission (SRTM) Cband digital elevation model (DEM) with resolutions of 90 m, downloaded from the Consultative Group for International Agriculture Research's Consortium for Spatial Information (2013), was used. SRTM DEM voids were filled and updated to version 4, and the problems of radar shadow, layover, and foreshortening from synthetic aperture radar (SAR) were resolved (Le Bris et al 2011). The Fedchenko basin was defined using SRTM DEM by the hydrology tools in ArcGIS 9.3 software, and then it was used to clip the glacier polygons.
Glaciers react to climate by changing in area; glacier area is easily detectable on remote-sensing images (Paul 2002;Paul et al 2002). Exposed glaciers and snow have high reflectance in the visible and near-infrared (VNIR) band and strong absorption in the shortwave infrared  band (SWIR) (Warren 1982;Dozier 1989;Hall et al 1995); therefore, these two bands can be used to extract glacier outlines. TM3/TM5 or TM4/TM5 plus an additional green or blue band and the normalized difference snow index (NDSI) are widely used; this ratio algorithm is a robust method for delineating glaciers in shadow areas and removing cloud covers (Winther 1992;Paul 2002;Paul et al 2002;Racoviteanu, Arnaud, et al 2008). However, water bodies are bracketed, and postediting is necessary (Paul et al 2007;Andreassen et al 2008;Svoboda and Paul 2009). The threshold value for the band ratio can be obtained using a histogram map (Gupta et al 2005;Silverio and Jaquet 2005). The band value can be a raw digital number value (Bayr et al 1994) or reflectance at the top of the atmosphere (Jacobs et al 1997); this digital number value is easier to use and leads to more accurate results (Paul 2000). Small seasonal snow patches and gaps due to debris cover can be reduced using a median filter (Andreassen et al 2008;Racoviteanu, Arnaud, et al 2008). Moraine-and debris-covered glaciers cannot be extracted by the band ratio method (Paul et al 2004); the best way is to manually delineate them with the help of DEM and band composite (Bolch, Menounos, and Wheate 2010). Methods used in this paper are shown in Figure 2. For Landsat TM/ETM+, strong shadows occurred on the images because of complicated relief, and mapping of the glacier boundaries had to be improved using NDSI 5 (band 2 2 band 5)/(band 2 + band 5). NDSI is robust enough to diminish the topographic effect (Racoviteanu,FIGURE 2 Processing method used in this study.  (Table 2); the resulting glacier margins match well with those on the false-colorcomposite images ( Figure 1C). Water bodies were identified using the normalized difference water index (NDWI), but turbid lakes and shadowy areas have the same spectral property, and so a logical algorithm of NDWI $ 0.34 and slope angle , 6u was used to extract water pixels on the image, and then the total area of water body inclusions within the glacier features was subtracted. ETM+ images have a panchromatic band with 15 m resolution; this band can be fused with VNIR of 30 m resolution to get 15-m-resolution images, which can improve the accuracy of delineating debris-covered glaciers.

MountainResearch
For the ALOS image, ground-control points for orthorectification were selected directly from the ETM+ image. In total, 14 control points were selected. The ALOS/AVNIR2 image was orthorectified with RPC in ENVI 4.7; the root-mean-square error (RMSE) was 3.6 m within 1 pixel error. NDSI is not applicable to the ALOS/ AVNIR2 image because it has no SWIR band, so the B4R2/ 4 was used from reference to B3R1/3 (threshold for band 3 and the ratio of band 1 to band 3) for Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) images (Kargel et al 2005), because the spectral ranges of band 2 and band 4 of ALOS/ AVNIR2 were in agreement with band 1 and band 3 of ASTER, respectively. This algorithm can isolate water bodies and overcome satellite sensor saturation over snow (Raup et al 2007). A two-dimensional scatter plot was drawn using 10 3 (band 2/band 4) and band 4, and glacier threshold values were visually inspected on the twodimensional scatter plot that was linked to the false-color image of ALOS/AVNIR bands 4, 3, and 2. The constraints for dirty ice were 10R2/4 $ 16.6 and band 4 $ 58, whereas for illuminated snow and clean ice, they were 10R2/4 , 16.6 and B4 $ 121.5 ( Table 2).
The glaciers in shadow areas could not be discriminated automatically by logistic algorithm; therefore, the ALOS/AVNIR2 image was topographically corrected with the Sandmeier model (Sandmeier and Itten 1997), which was run in ATCOR3 for ERDAS IMAGINE 2011 software. The key input parameters for the model are scene visibility (km) and water content. Water content was substituted by middle-latitude summer rural provided by the ATCOR3 software, and scene visibility was calculated with aerosol optical depth (t) obtained by Song's algorithm (Song et al 2001). The path radiance in Song's algorithm was determined by the method described by Chavez (1988). A t of 0.078122 and scene visibility of 82 km were obtained when the satellite overpassed. Finally, glaciers in shadow were delineated manually on the corrected map.
Postediting was necessary for both images. Small snow patches and gaps were removed using a 5 3 5 median filter. Glaciers with area smaller than 0.01 km 2 , large seasonal snow patches, and remnant water bodies were eliminated. Debris-covered glaciers, missed glaciers in shadow or under cloud, and supraglacial water bodies were manually appended. Figure 1C shows the study area for which postediting was carried out.
Additionally, in order to analyze the relationship between topographical parameters and glacier area, the watershed was delineated based on the SRTM DEM using the hydrology analysis tool in ArcGIS 9.3 and modified manually to obtain the drainage divides. Individual glaciers can be obtained by intersecting glacier and drainage polygons and then calculating topographic parameters (mean elevation, mean slope, and mean aspect) for each glacier entity by zonal statistics.

Error analysis
There are three main types of errors in mapping glacier outlines: technical errors, methodological or interpretation errors, and algorithm errors (Paul and Andreassen 2009). Assessing these errors is difficult, primarily because of the lack of sufficient ground-control points measured by differential global positioning system (DGPS) (Racoviteanu et al 2009). Methods of evaluating error include the error matrix (Gjermundsen et al 2011).
In this study, it was difficult to select the threshold for NDSI, because a glacier in shadow has a reflectance similar to that of bright rock. If the threshold is too low, shaded glaciers can be delineated automatically, but many rocks are included; 0.5 for NDSI is a balance trick. Seasonal snow is another factor producing error. If snow is not connected to a glacier, much of it can be removed by filter for areas smaller than 0.01 km 2 . However, large snow patches classified as part of glaciers cannot be removed automatically, and deleting them manually would involve arbitrary choices. Although the remotesensing images were acquired at the end of the ablation season, it is possible that it snowed some hours before the satellite passed over the area. In a positive mass balance year, seasonal snow can lead to greater error in delineating glacier areas (Gjermundsen et al 2011). A debris-covered glacier can also only be delineated with great uncertainty; it is very difficult to define its boundary, even if it is investigated on the spot. Because ALOS/AVNIR2 images have no SWIR band, errors were also caused by laborious manual delineation and editing. Considering all these factors, three areas were selected on a classification map for Landsat TM/ETM+; each area included 400 pixels. There were 1, 2, and 4 pixels that could not be classified with certainty as glacier; thus, the glacier areas in 1992 and 2000 had an error of 0.5% at the confidence level of 95% based on the De Moivre-Laplace theorem in probability theory. Using the methods described earlier herein, we identified 5, 5, and 7 pixels in three areas on the ALOS classification map that could not clearly be classified as glaciers; thus, the glacier area in 2006 had a 1.1% error at the confidence level of 95%. The results were confirmed by hypothesis testing at the 0.05 significance level.

Results and discussion
There are 287 glaciers in the Fedchenko basin. Figure 3A shows the distribution of glacier areas and number of glaciers in 1992 per area class in the basin. Glaciers smaller than 1 km 2 make up 90.6% of the total number but only 4.1% of the total area. In contrast, glaciers larger than 5 km 2 make up 2.1% of the total number and 90.7%  of the total area. The area class is bigger, the number of glaciers is smaller, and the total area is larger. The largest five are the Fedchenko, Kosinenko, Kalinina, Ulugbeka, and Seismologov glaciers, with areas of 697.7 km 2 , 23.2 km 2 , 17.7 km 2 , 9.1 km 2 , and 5.6 km 2 , respectively (Table 3). Figure 3B presents the area-elevation distribution in 1992 by 50 m elevation intervals. The longterm ELA is approximately equal to the elevation for total maximum glacier area (Braithwaite and Raper 2009). The hypsography of glacierized areas indicates that the largest area lies in the range from 5000 m to 5050 m; thus, the ELA in the glacier system is close to 5000 m. The relationship between glacier area and mean slope ( Figure 4A) suggests that glaciers with a larger area have a smaller mean slope. Glaciers with a smaller area possess both more and less pronounced mean slopes, indicating a larger variability of slope for small glaciers than for large ones. The variation of glacier mean elevation against mean aspect ( Figure 4B) suggests that glacier mean elevation is dependent on the aspect. Generally, the elevation of south-facing glaciers is higher; this may be mainly caused by solar radiation. Figure 5 compares changes in glacier area between 1992 and 2006 to initial area for 127 selected glaciers, suggesting that both the relative change in area and the variation in that change increase in inverse proportion to the original glacier area. There was a larger scatter for glacier areas smaller than 1 km 2 . Many small glaciers grew by more than 20%; some even increased by more than 100%, possibly by merging with other small glaciers. The finding of more than 500% change over 14 years is very unrealistic and was more likely caused by the presence of seasonal snow. The change in area of most glaciers ranged from 251% to +48%. Glaciers smaller than 1 km 2 either lost (up to 1.80 km 2 ) or gained (up to 7.08 km 2 ) area (the net gain was 5.28 km 2 ), whereas glaciers larger than 1 km 2 almost all gained area. The Fedchenko, Kosinenko, Kalinina, Ulugbeka, and Seismologov glaciers gained 11.16 km 2 , 4.06 km 2 , 0.50 km 2 , 0.66 km 2 , and 0.6 km 2 , respectively ( Table 3). The area gain for the Kosinenko glacier is prominent because of its surging. Absolute area gain was higher for larger glaciers. This implies that relative area change depends on glacier size (Bolch, Yao, et al 2010).
The total glacier areas in 1992, 2000, and 2006 are presented in Table 3. The uncertainty is 60.5% (or 64 km 2 ) for 1992 and 2000 and 61.1% (or 610 km 2 ) for 2006. The area diminished by 0.71% from 1992 to 2000 but enlarged by 4.19% from 2000 to 2006, for a gain of 3.46% from 1992 to 2006 ( Figure 6). Two areas in particular changed remarkably: 1. The Vasilevskogo glacier advanced to connect with its trunk glacier in 2006, forming the shape of a paw, with a surging area of 0.2 km 2 . The upper part of the main trunk of the Bivouac glacier surged after 1992 (Kotlyakov et al 2008) and is now in a quiescent phase; the surging belt can be discerned by comparing images In addition to the surging area of 3.7 km 2 , glacier area increased by another 25.2 km 2 from 1992 to 2006. Taking into account the delineating errors, glaciers in the Fedchenko basin are in a stable state. This result is in agreement with the mass balance of 0.00 6 0.08 m y 21 water equivalent for the Fedchenko glacier between 2000 and 2011, and with the mass balance of surge-type and non-surge-type glaciers in Pamir, with +0.20 6 0.24 m y 21 water equivalent and +0.13 6 0.037 m y 21 water equivalent, respectively, during the same time (Gardelle et al 2013).
The glacier fluctuations could be linked to climate change. The Fedchenko glacier meteorological station has been closed since 1994; thus, reanalysis data (eg for temperature and precipitation) from the US National Centers for Environmental Prediction and National Center for Atmospheric Research (NCEP/NCAR) were used to investigate climate change in the region. The climate changes recorded by NCEP/NCAR (Figure 7) before 1994 follow the same trend as observed by the Fedchenko glacier meteorological station (Khromova et al 2006) for that period, which gives us confidence in using NCEP/NCAR data to interpret glacier changes.
Summer temperatures have not displayed a clear warming trend since 1970 ( Figure 7); however, precipitation in winter decreased markedly from 1975 to 1990 and increased from 1990 to 2010. Glacier area responds more slowly to annual meteorological conditions than mass balance does (Cuffey and Paterson 2010). The adaptation of temperate glaciers to a mass change takes several years to several decades (Lopez et al 2010). With increasing precipitation in winter and almost stable temperatures in summer since 1990, the glaciers in the Fedchenko basin shrank slightly from 1992 to 2000, and then expanded after 2000. Therefore, increased winter precipitation is probably the major reason for slight glacier expansion in the last decade, which is consistent with the report of Yao et al (2012).

Conclusion
We used the NDSI method for TM/ETM+ images and the B4R2/4 method for an ALOS/AVNIR2 image to extract the area of the Fedchenko glacier system, and we divided this into individual glaciers through the drainage divide method. Glacier area increased and glacier numbers decreased with increasing glacier area. The mean slope decreased when the area increased, and smaller glaciers showed a larger variability in slope. The altitude for maximal area ranges from 5000 to 5050 m asl; thus, the ELA in the glacier system is around 5000 m. Glacier elevation is larger when facing south. Results also show that larger glaciers have less relative change in area, and smaller glaciers have a greater range of variability. Total glacier area in the basin shrank 5.9 km 2 (0.71%) from 1992 to 2000 and expanded 34.8 km 2 (4.19%) from 2000 to 2006, for an increase in total area of 28.9 km 2 (3.46%) from 1992 to 2006-a slight expansion after excluding errors and glacier surging. With a continuous increase in winter precipitation since 1990, glaciers retreated slowly at first, and then advanced beginning in 2000, when they adjusted to positive mass balance. Increasing precipitation probably represents the controlling factor for glacier variations in this region.