Spatially distributed snow-cover extent can be derived from remote sensing data with good accuracy. However, such data are available for recent decades only, after satellite missions with proper snow detection capabilities were launched. Yet, longer time series of snow-cover area are usually required, e.g., for hydrological model calibration or water availability assessment in the past. We present a methodology to reconstruct historical snow coverage using recently available remote sensing data and long-term point observations of snow depth from existing meteorological stations. The methodology is mainly based on correlations between station records and spatial snow-cover patterns. Additionally, topography and temporal persistence of snow patterns are taken into account. The methodology was applied to the Zerafshan River basin in Central Asia – a very data-sparse region. Reconstructed snow cover was cross validated against independent remote sensing data and shows an accuracy of about 85 %. The methodology can be used in mountainous regions to overcome the data gap for earlier decades when the availability of remote sensing snow-cover data was strongly limited.
Water resources from remote mountain catchments play a crucial role in the development of regions in or in the vicinity of mountain ranges (Pellicciotti et al., 2012). Seasonal snow is an important water resource in many of Earth's semiarid regions (Durand et al., 2008). Particularly in Central Asia, seasonal snowmelt decisively contributes to the total runoff volume (Ososkova et al., 2000; Unger-Shayesteh et al., 2013).
Information on snow cover and snow depth and ideally on snow water equivalent in Central Asian catchments is crucial for seasonal forecasts of water availability and for calibration and validation of hydrological models. However, the available sparse station-based data are insufficient to represent the snow-cover variability over the large and remote mountain areas (Erickson et al., 2005). The development of remote sensing techniques during recent decades allows the derivation of snow cover spatially (Liu et al., 2012). Widely used remotely sensed snow-cover products are from Advanced Very High Resolution Radiometer (AVHRR), Landsat and Moderate Resolution Imaging Spectroradiometer (MODIS) missions. Whereas AVHRR (launched 1978) and Landsat (launched 1972) offer remote sensing data for a longer period, MODIS is available only after 2000. However, snow cover from Landsat and AVHRR needs to be derived by the end user themselves, whereas MODIS offers already-compiled snow-cover product. The above-mentioned snow products are extremely useful to study snow cover worldwide; however, they are strongly limited by the presence of clouds. Recently, the reconstruction of snow-cover time series from AVHRR data for Central Asia has been reported by Zhou et al. (2013), but they are also limited in time starting in 1986 at earliest.
Uzhydromet snow observation stations with indication of elevation (Elev), and snow predictability index (SPI) and land predictability index (LPI) values for the study area (see Sect. 4). The entries “records on snow reconstructed days” indicate whether a station was snow covered (0) or snow free (1) during a day for which snow-cover reconstruction was conducted and Landsat scenes were available for validation.
Several studies used remotely sensed snow cover either as input to hydrological models (Tekeli et al., 2005; Immerzeel et al., 2008; Li et al., 2008; Wang et al., 2010) or for calibration and validation purposes (Parajka and Blöschl, 2008a; Corbari et al., 2009; Liu et al., 2012; Duethmann et al., 2014). Particularly for hydrological model calibration, spatially distributed snow-cover data offer high information content required to constrain model parameters (Finger et al., 2011; Duethmann et al., 2014).
In Central Asia, continuous hydrometeorological records are widely available from the 1960s and earlier until the collapse of the Soviet Union in 1991. In contrast, continuous remote sensing snow-cover data from MODIS are readily available after 2000, when station data are very scarce. We present a methodology which enables reconstructing historical snow-cover pattern using long-term, point-based observations from existing meteorological stations and recent remotely sensed snow-cover data. By merging high-resolution spatial satellite data with long-term station data, snow-cover patterns can be reconstructed for several decades into the past.
Only a limited number of studies on snow-cover reconstruction have been conducted in the past that use long-term station observations and recent remote sensing data (Robinson, 1991; Brown, 2000; Frei et al., 1999; Brown and Robinson, 2011). These studies are, however, conducted at the continental scale under conditions of dense station network availability and neglecting the effect of topography. Robinson (1991) and Frei et al. (1999) conducted reconstruction of snow cover based on regression analysis between snow characteristics and snow-cover area (SCA) derived from AVHRR satellite observations. As snow characteristics both studies used snow-cover duration derived from interpolated station records. Another study by Brown (2000) conducted reconstruction of snow cover for “pre-satellite era” interpolating snow-depth data from station network. For grid cells of nearly 200 km, the interpolation of snow cover was done using different thresholds for snow depth and compared against NOAA snow-cover extent during “satellite era”. The calibration showed 2 cm to be most appropriate snow depth for accurate snow-cover reconstruction based on station data. Brown and Robinson (2011) updated and extended the snow-cover reconstruction of Brown (2000) to the period 1922–2010 and used these data for trend analysis of snow-cover extent in the Northern Hemisphere. These studies can be helpful in assessing climate-related variations of snow cover but are hardly transferable to smaller catchment scale with moderate resolution and limited station data availability.
Different to those studies mentioned above, we present a methodology for snow-cover reconstruction (1) with moderate spatial resolution (500 m), (2) suitable for catchment scale hydrological studies, (3) accounting topography and (4) delivering spatially distributed snow-cover maps. The methodology takes advantage of the strong control of topography on the spatial snow-cover distribution. Hence, measurements from snow observation stations at different elevations can be interpreted as representative sites to predict snow-cover patterns. The methodology consists of five successive steps which make use of topographic information and correlations between station records and spatial snow-cover patterns. In order to test the presented methodology, snow-cover reconstruction was conducted for 4 days (Table 1) for which independent Landsat data were available.
The methodology for snow-cover reconstruction was developed and tested for the area containing the upper Zerafshan River basin, Central Asia (Fig. 1).
Location of the upper Zerafshan River basin in the Gissaro-Alai mountain range, Central Asia. Snow-cover reconstruction was conducted for the entire area of Fig. 1b and validated for the area with Landsat footprint.
The upper Zerafshan River basin is located in the Gissaro-Alai mountain
range. Elevation ranges from 658 to 5402 m a.s.l. and basin area is about
12 000 km
We used (1) daily in situ snow-depth data, (2) daily MODIS snow-cover data, (3) a digital elevation model (DEM) and (4) Landsat data. The first three data sets were used for snow-cover reconstruction whereas Landsat data were used as an independent data set to validate the results.
Daily snow-depth data in the period from 1964 to 2012 were available for seven climate stations located at different elevations (Fig. 1, Table 1). These data contain continuous snow-depth measurements including records on no-snow conditions. Snow depth in these stations are recorded at 1 cm threshold depth. The data were provided by Uzbek Hydrometeorological Service (Uzhydromet).
Monthly average air temperature (T), cumulative precipitation (P),
discharge (Q) and SCA dynamics in the Zerafshan basin. T, P and Q means are
based on data for the period from 1930 until 2008. Daily SCA is for 2004
obtained from MODIS and cloud eliminated using Gafurov and Bárdossy (2009).
Temperature (0
Original Landsat scenes (top row) and derived snow-cover maps used for validation (bottom row). Black outlines show the validation domain for Zerafshan basin.
MODIS daily snow-cover data from the Terra satellite with 500 m spatial
resolution (MOD10A, version V005) were employed for the time period of 2000
to 2012. We used MODIS Terra snow-cover data due to the longer time series
compared to the Aqua satellite, which delivered snow-cover data only after
2002. The MODIS snow-cover product is based on the normalized difference
snow index (NDSI) algorithm (Hall et al., 2002). Its accuracy was tested in
different parts of the world showing good agreement with in situ data (Klein
and Barnett, 2003; Tekeli et al., 2005; Parajka et al., 2006; Ault et al., 2006;
Wang et al., 2008; Liang et al., 2008; Huang et al., 2011; Gafurov et al.,
2013; Parajka et al., 2012). The main drawback of MODIS snow-cover data is
the limitation due to cloud cover. There have been several studies on
filtering methods for reducing cloud cover or even removing it completely
(e.g., Parajka and Blöschl, 2008b; Gafurov and Bárdossy, 2009; Tong
et al., 2009; Hall et al., 2010; Lòpez-Burgos et al., 2013). We used
original MODIS snow-cover data to exclude any uncertainty that may be
introduced by cloud filtering. The data were obtained from the National
Aeronatics and Space Administration (NASA) Earth Observing System Data and
Information System (EOSDIS) Reverb platform. MODIS data are distributed as
tiles with the size of 10
The void-filled DEM with 90 m spatial resolution from NASA Shuttle Radar Topography Mission (SRTM) was used. SRTM DEM data were obtained from the CGIAR CSI (Consultative Group on International Agricultural Research, Consortium for Spatial Information) database (Jarvis et al., 2008). To have the same resolution as the MODIS data (500 m), the 90 m SRTM DEM was aggregated to 500 m.
Optical remote sensing data from the Landsat Thematic Mapper sensor were used to validate the reconstructed snow-cover maps. The Landsat data have a spatial resolution of 30 m and a temporal resolution of 16 days. Landsat data from 4 nearly clear-sky days in the snow season (10 April 1998, 20 November 1998, 20 April 1999 and 15 November 1999) were used for validation purposes. Snow-cover maps for the Landsat footprint (see Fig. 1) were prepared using the NDSI methodology. For a detailed description of the algorithm used for deriving snow-cover maps from Landsat refer to Gafurov et al. (2013). Figure 3 shows raw and processed Landsat snow-cover maps for the study area.
Since Landsat has a spatial resolution of 30 m and snow reconstruction was performed for 500 m pixels based on MODIS resolution, the processed Landsat snow-cover maps were spatially aggregated to 500 m resolution. This was done by classifying each of the 500 m pixels as snow covered or snow free, based on the majority of the 30 m Landsat pixels within the 500 m pixel.
The methodology presented hereafter is based on similarity of different
locations in terms of presence or absence of snow at a given time. The idea
is to use the information about the presence of snow from one location in
order to estimate the presence of snow at another location. The similarity
between different locations was assessed using both observed snow cover at
meteorological stations (i.e., records of snow depth
pixel to station CP fields temporally persistent monthly probability fields pixel to pixel CP fields usage of elevation information pixel to station CP for CP
In the first step, we consider the CP of each pixel, given the observed data
from a set of snow stations. We compute the CP of each pixel as follows:
CPs were computed for each MODIS pixel in the study area (total of 169 776 pixels)
using over 12 years of available MODIS data and observed snow-depth
measurements. Hence, the daily snow-cover maps from MODIS are treated as
snow observations for each 500 m grid cell, giving rise to a very dense
“observation network”. An example for a CP map for snow and land
conditions for Chimgan station is given in Fig. 4. In total, 14 maps were
derived (two maps for every of the seven stations: one for
CP maps for snow (top) and land (bottom) conditions of Chimgan station (see Fig. 1) for the study area. The figure shows the same domain as Fig. 1b.
The number of pixels with
Temporally persistent spatial snow (MP
Figure 4 shows that
This step leads to a partially reconstructed snow-cover map which is further enhanced in the next steps.
Snow-cover extent is a seasonally variable parameter. Accordingly, the
probability of a certain pixel to be covered by snow or land varies with
time. The second step for reconstructing snow cover is based on the
observation that during different months, certain pixels are snow covered or
snow free with high confidence. The spatial distributions of such temporally
persistent patterns can be identified using the available MODIS snow-cover
data in the period 2000–2012. A monthly probability (MP) of each pixel to be covered by
snow or land in a certain month is computed according to
Pixels with MP
The main idea in this step is to transfer these temporally persistent
monthly spatial snow/land patterns (SPI
Monthly SPI, LPI,
These
CP maps for snow (top) and land (bottom) conditions for the pixel
In step 1, CPs of each pixel in accordance to station records were computed,
and any pixel that had
The computation of
ELA records of Abramov glacier (WGMS, 2001; Pertziger, 1996).
The pixel with coordinates
The maximum SPI value (Fig. 7, top) is 46 %, meaning that, according to the
observations of the period 2000–2012, 46 % (78 189 pixels) of the study
area was always snow covered when that particular pixel was snow covered.
The maximum LPI value (Fig. 7 bottom) is 88 %, meaning that this
particular pixel is able to predict snow-free conditions for 88 % (149 685
pixels) of the basin. These two pixels with maximum SPI and LPI values
are located within an area which has high predictive power for snow and
land, respectively. When interpreting Fig. 7, three further features are
worth noting: (1) pixels with SPI
The SPI and LPI maps were used for classifying pixels that are still
undefined after the previous steps:
Since in this step SPI and LPI maps were generated for every pixel in the basin, this step tends to classify a significantly larger area than the first step where only seven stations were used for constructing CPs.
This step is adapted from Gafurov et al. (2009) and is based on the
information of neighboring pixels. Let us consider a pixel that has not
been classified as snow covered or snow free in any of the previous steps.
If any of the adjacent eight pixels is covered by snow and the elevation of
that snow-covered pixel is lower than the pixel that is still undefined,
then the undefined pixel is classified as snow covered. The same idea is
applied for snow-free pixels. Hence, this step can be formalized as follows:
SPI (top) and LPI (bottom) values of each pixel (in %) in the study area defined in Fig. 1b.
This step takes only elevation of neighboring pixel into account. However, in areas where factors others than elevation have an influence on neighboring pixel condition (e.g., pixels located near to water surfaces or forests), additional information such as a land cover map could be introduced into this step.
In the last step, the
Contingency table (in %) for the reconstructed snow-cover maps validated against four aggregated Landsat snow-cover images. Four cases are distinguished: SS, LL, SL and LS. The first (second) letter indicates the classification according to the presented algorithm (Landsat). “S” stands for snow and “L” for land. “Total” indicates the percentage of pixels classified after each step. Results refer to the Landsat domain (dashed line) shown in Fig. 1b.
Taking maximum lower bound CI values for still-undefined pixels in the last
step allows us to complete the classification for all pixels. However, since in
this step
Reconstructed (top) and Landsat (bottom) snow-cover maps for 10 April 1998.
Fraction of reconstruction in steps 1–4 and maximum CI values for snow and land in step 5 for the study area illustrated in Fig. 1b.
Applying the five steps described above, snow-cover maps for the area containing Zerafshan basin were reconstructed for 4 days in 1998 and 1999. The maps contain binary information showing whether a given pixel was covered by snow or not. The accuracy of the reconstructed snow-cover maps was assessed by comparing against independent snow maps derived from four Landsat images from the same days. The validation could be done only for recent years (1998 and 1999) due to the availability of cloud-free Landsat images during the snow period only for those days. The comparison was performed on a pixel-to-pixel basis, and the accuracy was assessed in a contingency table (Table 4). In Table 4, the sum of percentages of “SS” and “LL” columns represent the degree of accuracy after each reconstruction step, related to the total share of reconstructed pixels. Accordingly, the sum of “SL” and “LS” indicate the error in relation to the total percentage.
As an example, Fig. 8 shows the reconstructed and Landsat-derived snow-cover
maps for 10 April 1998. The comparison of these maps results in 85.7 %
of correct reconstruction (cases SS
In order to better illustrate snow reconstruction in step 5, Fig. 9 shows
the areal fraction for which the reconstruction was performed in steps 1–4
and maximum lower bound CI obtained in step 5 under CP
Figures 8, 9 and 10 also demonstrate that the methodology provides two types
of results for snow-cover reconstruction: deterministic and probabilistic
snow-cover maps. Deterministic maps result from the complete classification
of pixels (Fig. 8) with binary information (snow/snow free), taking
CI
Trade-off between erroneous reconstruction (ER, dashed lines) and reconstruction fraction (RF, solid lines) after step 4 as a function of CI.
The validation days for this study were chosen deliberately from the snowmelt
and snow accumulation season (transition period) when snow-cover estimation
is particularly challenging. For the time outside the snowmelt or snow
accumulation period, higher accuracies can be expected since a higher
fraction can be reconstructed in the first four steps already. During the snow
transition period, snow-cover conditions such as ephemeral snow cover can
occur which exacerbates snow-cover estimation. However, in the
reconstruction process using steps 1, 2 and 3 the conditions with ephemeral
snow cover (in the period 2000–2012) are accounted for as well. Under such
conditions, station or MODIS data may see different snow cover than the
reality. In such cases (e.g., MODIS sees “land” although there is ephemeral
snow, whilst the station sees “snow” since it is a manual point recording
with a certain threshold), and CP and MP of the pixel produce the value of
The validation of reconstructed snow-cover maps were done using independent
Landsat data in this study. Alternatively, the AVHRR snow-cover data, which
are also available beyond the MODIS data availability in the past, can be
used for validation purposes. However, AVHRR snow-cover data have a coarser
spatial resolution (
The predictive power of the observations at meteorological stations for snow-cover reconstruction is limited by the elevation range of the stations. If all meteorological stations are located at high elevations, they will be good predictors during summer for snow-free conditions but will perform poorly when predicting snow-covered areas during winter due to their elevation and correspondingly lower SPI values. Conversely, low-elevation stations are better indicators for snow-covered pixels at higher altitudes than they are for snow-free ones. Hence, a wide spread in station elevation is optimal for accurate snow-cover reconstruction. In our case study, the application of the presented methodology suffered from the small number of station data (only seven stations). A higher number of stations would lead to a higher number of SPI and LPI maps and would allow us to reconstruct a larger areal fraction of snow cover in the first four steps with high accuracy. Noticeably, the stations do not need to be located inside the area of interest.
Reconstruction of the snow cover for the past is based on the assumptions that (1) the calibration period, i.e., the MODIS data period, is representative for the past period, and (2) the relationship between station records and spatial snow patterns derived from MODIS data is stationary, i.e., does not significantly change in time. A calibration period which lacks extreme conditions, e.g., snow-rich or snow-scarce years, might lead to larger errors in the reconstruction. A longer calibration period is expected to lead to more robust relationships for reconstructing snow cover.
The problem of representativity of the MODIS period in the reconstruction step 2 is tackled by the introduction of the elevation buffer to capture the effect of interannual temporal variability of snow-line elevation. For this the temporal variability of the recorded ELA from the neighboring Abramov glacier was used as a proxy. Through changes in climatic conditions of the calibration period going beyond temporal variability of the snow-line elevation in the reconstruction period, the relationships between station records and some pixels (step 1) and between pixels (step 3) may become non-representative. This occurs if the snow line in the future/calibration period more often separates the station of the pixels compared to the reconstruction period. Hence, an analysis of temperature and precipitation trends and comparison of climatology between calibration and reconstruction periods may provide some confidence on representativeness of the relationships used.
The statistical relationship (CP) between point measurements and aerial patterns computed in this study highly depend on topography. Since the Zerafshan basin has a very heterogeneous topography with high elevation range, good predictive power (SPI and LPI) of individual stations could be obtained. This is important to estimate initial snow cover in the first step, which is a base input for next steps (except step 2). Thus, we can conclude that the methodology is well applicable for mountainous areas where high SPI and LPI values can be obtained. However, it might be difficult to exploit statistical relationships between point measurements and aerial pattern in lowland areas; this is a subject to be tested.
In this study, a methodology for reconstructing past snow cover using historical in situ snow-depth data, recent remote sensing snow-cover data and topographic data was presented. The methodology is based on (1) constructing relationships between station observations and remote sensing data, (2) estimating the monthly variation of snow cover from remote sensing data, (3) deriving pixel-to-pixel relationships using remote sensing data and (4) using neighborhood relations. Once the dependence between individual pixels and station records is derived, this dependence is used to reconstruct past snow cover based solely on station records.
The methodology was applied to a study area containing the Zerafshan River basin – a basin with high topographic gradients – in Central Asia and showed correct classification in the range of 83.3–85.7 % when compared to four Landsat snow-cover scenes. This high agreement is noteworthy, given that only seven stations and 12 years of remote sensing data were available. Moreover, snow-cover reconstruction was done for snowmelt and onset season when snow classification is challenging compared to outside snowmelt and onset seasons when large areas are easy to classify as snow or land. The agreement is only slightly less than that of original MODIS snow-cover product with accuracy of about 92 % for Central Asia when compared to Landsat-derived snow-cover maps (Gafurov et al., 2013). Just 12 years of MODIS data was sufficient to extract stable patterns of snow cover and relate them to station records in the Zerafshan basin with heterogeneous topography. Hence, we conclude that the developed methodology is suitable to derive past snow cover in remote mountainous regions such as the Zerafshan basin with very limited data availability. Reconstructed snow-cover patterns can be used for hydrological model calibration/validation and for understanding snow-cover dynamics over large areas prior to the age of satellite observations. The performance of methodology presented here for non-mountainous areas remains an open question.
This work was carried out within the framework of the CAWa (Water in Central Asia)
project (