Introduction

Monitoring the snow cover extent and its variability is important in terms of water resource management and climate change studies (Cohen and Entekhabi 2001; Brubaker et al. 2005). The importance of the snow cover extent increases as the ratio of snowmelt runoff in the annual river discharge increases (Sönmez et al. 2014). Euphrates and Tigris, two transboundary rivers originating in Eastern Turkey, are fed by spring snowmelt which accounts for 60–70 % of the annual discharge (Tekeli et al. 2005a). As Jacobs et al. (2005) reported, winter precipitation at higher elevations is particularly important since the cooler temperatures allow the accumulation and persistence of snow pack and the accumulated snow will appear as snowmelt runoff in spring. Thus, better understanding of snow patterns and their variation both in time and space becomes more important (Richer et al. 2013) for accurate snowmelt runoff estimations. Eastern Turkey, having an average elevation of 2200 m (Akyürek et al. 2011), contains the major snow-dominated basins and holds the main snow accumulation areas (Tekeli 2008) making snow monitoring over the region an important issue.

Snowmelt runoff model (SRM) and Hydrologiska Byråns Vattenbalansavdelning (HBV) model have been commonly used in operational runoff forecasting (Udnaes et al. 2007). SRM is among the few models that directly use the snow-covered area (SCA) as an input parameter. Successful applications of SRM over Turkey are discussed in Tekeli (2005), Tekeli et al. (2005a), and Ṣensoy and Uysal (2012). SCA is one of the input variables of SRM so that accuracy of the SCA directly affects the performance of SRM simulations/forecasts.

Due to the highly dynamic nature of snow, inadequacy of the ground-based snow field observations for continuous monitoring of snow in both temporal and spatial domain were discussed in several previous studies (Mognard 2003; Rees 2006; Tekeli and Tekeli 2012; Richer et al. 2013; Sönmez et al. 2014). In this context, satellite-based remote sensing is capable of providing snow information at regional and global scales (Hall et al. 2005; Brown and Armstrong 2010) especially for the remote and topographically complex terrains. Successful snow detection using satellite imagery is presented in various studies (e.g., Crane and Anderson 1984; Dozier 1989; Gessel 1989; Romanov et al. 2000; Hall et al. 2002). Advanced Very High Resolution Radiometer (AVHRR) data were used, among various satellite platforms, in SCA detection over Turkey by Tekeli (2000). Tekeli (2005) and Tekeli et al. (2005a) investigated the usability of Moderate Resolution Imaging Spectroradiometer (MODIS)-based snow cover maps over Turkey. The accuracy dependence of MODIS SCA with respect to snow depth, land cover, and presence of cloud was discussed in detail in several studies (Justice et al. 1998; Bitner et al. 2002; Klein and Barnett 2003; Simic et al. 2004; Zhou et al. 2005; Tekeli et al. 2005a; Hall and Riggs 2007; Burgos et al. 2013). The preference of MODIS over AVHRR in the later studies was due to its increased spatial resolution, 500 to 1000 m. However, both MODIS and AVHRR are optical sensors, and their usage is limited by cloud obstruction (Maurer et al. 2003; Tekeli and Tekeli 2012). Although some techniques were developed to reduce cloud obstruction (Parajka and Blöschl 2008; Gafurov and Bárdossy 2009; Gao et al. 2010; Tekeli and Tekeli 2012), cloud blockage still remains a major concern, especially when continuous datasets are needed for studies such as trend analysis (Sönmez et al. 2014).

Though cloud blockage can be solved to a certain degree either by using passive microwave imagery or by using some techniques, to the knowledge of the authors, there is no method available at the present time to forecast SCA maps needed by hydrological models. There are some studies (Tekeli et al. 2005b; Jain et al. 2011) in which cumulative mean air temperature (CMAT) is used to forecast SCA; however, these are site-specific and depend on the existence of air temperature data.

In this study, snow cover data from Interactive Multisensor Snow and Ice Mapping System (IMS) operated by the National Oceanic and Atmospheric Administration’s National Environmental Satellite Data and Information Service (NOAA/NESDIS) is used. The main advantages for the use of IMS snow product are as follows: (1) IMS provides the longest satellite-derived environmental data record for hemispheric snow cover monitoring (Robinson et. al. 1993); (2) erroneous snow detections, which can arise from interference of clouds, fog with ice and snow, bright surface features, boundaries between water bodies, are eliminated as much as possible by the interactive techniques (Chen et al. 2012); (3) interference of clouds, the major problem in snow monitoring, is eliminated in IMS product where multiple images are used along with microwave data (Helfrich et al. 2007).

The main objective of this study is to characterize the temporal and spatial variability of the snow cover based on long-term (2004–2012) IMS snow cover maps. The long-term analysis enable to define probability for each pixel to be snow covered in a particular day of year, which helps in understanding the snow cover patterns and their variation both in time and space. The probabilities can also be used to estimate the SCA for the upcoming year.

Study area and datasets

Study area

Turkey, the study area, is located in the transaction between Europe and Asia continents from 26° E to 45° E and from 36° N to 42° N. The area has a complex terrain with Mediterranean Sea, Aegean Sea, and Black Sea presenting the southern, western, and northern boundaries, respectively (Fig. 1).

Fig. 1
figure 1

Digital elevation model of Turkey and ground truth stations used in the analysis

Full size image

North and south coasts of Turkey are bounded by the mountains that are parallel to the coast, extending from west to east and merging in the east. Flat areas are mainly located around the coastal regions. Central Anatolian Plateau lies in between the northern and southern mountains. The average elevation of Turkey is 1130 m and increases from west to east with the highest areas located in Eastern Turkey (Deniz et al. 2011).

The climate and snow characteristics vary with respect to sub-regions of Turkey (Fig. 1). Black Sea (BLS) region has warm summers and cool to cold winters, both which can be classified as wet. Precipitation is observed throughout the year. Rize, a city in the Eastern BLS region, has the highest precipitation record with 2200 mm. Snowfall can be observed almost every year. However, over the coastal places, snow does not stay on the ground more than a few days. But, along the mountainous region, separating BLS from Eastern Anatolia (EAN) snow can stay for longer periods. Marmara (MAR) region has warm to hot summers and cool to cold winters. Summers are generally dry, and winters are wet. Snow is observed mostly in every year. However, similar to BLS, snow does not stay long over the coastal regions of MAR. Aegean (AEG) region has hot summers and mild to cool winters. Summers are mostly dry, and winters are wet. Opposite to BLS and MAR, snow is very rare over the coastal regions in AEG. However, high mountainous places get some snow during winter which does not stay long. Mediterranean (MED) region has hot and dry summers and mild to cold wet winters. Occurrence of snowfall is less than AEG. However, along the mountainous region extending parallel to the coast, snowfall is observed and snow can stay till end of June. Southeastern Anatolia (SAN) is the hottest and driest region in Turkey. Snowfall is not observed frequent other than the high mountains, even where, snow does not stay later than end of May. Eastern Anatolia (EAN) records the minimum air temperatures, and winters are mostly snowy. Snow can be observed till end of June. This region includes many snowmelt runoff-fed rivers like Euphrates and Tigris. Central Anatolia (CAN) has hot and dry summers with cold and wet winters. Snowfall is the main precipitation in winter period.

IMS snow cover data

NESDIS started producing weekly snow and ice charts on November 1966 (NSIDC 2014). These maps are considered to be the longest satellite-derived environmental data record for snow cover monitoring over the hemisphere (Robinson et al. 1993). Main data sources in IMS production are polar operational environmental satellites (POES) working in the visible region of the electromagnetic spectrum. Among those, AVHRR and MODIS are the primary ones. Geostationary satellites help snow monitoring by their high temporal resolutions. In this context, NOAA geostationary (GOES), European geostationary meteorological satellites (METEOSAT), and Japan’s geostationary meteorological satellites (GMS) are put into operation in 1975, 1988, and 1989, respectively (Ramsay 1998). After a 30-year operational period of weekly snow and ice maps at a spatial resolution of 190 km, the temporal and spatial resolutions of the product were improved to daily basis with 24 km in 1997 (Helfrich et al. 2007). The spatial resolution was further improved to 4 km in February 2004. Till then, IMS is produced on a daily basis at 4 km. The 4-km data is upscaled to 24 km to enable 24-km data continuity. Currently, 4-km daily IMS snow cover maps are provided for the Northern Hemisphere (NH) and classify the NH to a pixel type of “sea,” “land,” “sea/lake ice,” or “snow.” Further details about the IMS algorithm and products can be obtained from Ramsay (1998), Helfrich et al. (2007), and NSIDC (2014).

Despite the coarser spatial resolution of IMS (4 km) in comparison to MODIS snow cover products (0.5 km), Mazari et al. (2013) demonstrated that IMS accuracy is similar to MODIS product in clear skies and higher than that of MODIS in all sky conditions. The accuracy of IMS snow cover maps was found to be 80.79 % for Eastern Turkey (Sönmez et al. 2014) and was regarded as comparable to MODIS accuracy (62–82 %) for the Upper Euphrates River Basin that is located in Eastern Turkey (Tekeli et al. 2005a). Figure 2 presents a sample IMS snow cover image over Turkey for 1 March 2013.

Fig. 2
figure 2

Sample IMS snow cover map for 1 March 2013

Full size image

Ground truth data

Figure 1 shows the total 219 meteorological sites used in this study. Operation and maintenance of all these sites are under the responsibility of Turkish State Meteorological Services. For the purpose of accessibility and quick maintenance, most of these sites are located around the urban regions of the country. Table 1 summarizes the elevation groups and sub-regions of the 219 sites. The daily snow depth (SD) observations at these sites are used as the primary ground truth data in this study. No snow existence in the site is indicated by “0” value, and “−1” is used to indicate “trace” meaning presence of snow having a depth less than 1 cm. Any other positive integer number corresponds to the snow depth on the site. Range, step, persistence, and spatial quality control tests were applied to the SD observations, and observations with quality flag of “good” are used in the study. The details about the quality control tests and the flagging procedures are briefly introduced by Sönmez (2013).

Table 1 Number of the ground truth site variation with respect to selected elevation groups for Turkey (TUR) and sub-regions (adapted from Sönmez et al. 2014)
Full size table

Methodology

Long-term monitoring of the snow cover pattern provides data that can help in determining the spatial and temporal distribution of snow cover within the study region. For this purpose, IMS data covering the 2004–2012 periods were downloaded and processed for the study area. Based on data period, the probability of snow is defined, similar to Richer (2009), as the ratio of number of days that the pixel was snow covered to the total number of days under study. Although the definition is same as in Richer (2009), the equation used to calculate “probability of snow” (PS) in this study is rather different than that of Richer (2009) as IMS data is free of cloud blockage. Thus, PS is calculated using the equation

$$ {P}_s=\frac{{\displaystyle \sum_{i=1}^nS}}{n}, $$
(1)

where S is the total number of snow observations for the particular pixel during the study period and n is the total number of available images for the particular day of year considering the whole study period.

Results and analysis

The PS values are calculated on a Julian day basis using the daily IMS data for the 2004–2012 period. For each Julian day, Eq. 1 is used to compute the PS value for each IMS pixel within the study area (Turkey). PS variation over Turkey is presented in 15-day interval sequence starting from March 1, till June 30 in Fig. 3. The red pixels in the figure, pixels having a PS value of 1, were snow covered in every year for that Julian day. Gray pixels, pixels having PS value of 0, were snow free for that Julian day considering each available product in 2004–2012 period. Increased PS values indicate higher number of snow observations for that particular pixel for that Julian day.

Fig. 3
figure 3

PS values starting from March 1, till June 30 in 15-day intervals

Full size image

Similar to the findings in Ataṣ et al. (2015) and Akyürek et al. (2011), Fig. 3 indicates that the main snow depletion starts around the end of March. Furthermore, time sequence of PS maps is in agreement with the general snow cover pattern over Turkey. This is observed by the decreased (increased) PS values in the images as the time progress during the ablation (accumulation). The variation of PS images follows the general topography of Turkey, where higher PS values are observed mainly in the mountainous eastern regions and along high elevated regions parallel to the Black Sea and Mediterranean Sea. Lower PS values occur in the mid elevation regions mainly in the inner Anatolian region and no snow cover (zero PS value) in the coastal regions.

The effects of PS values on the SCA images are also investigated. Figure 4 represents the SCA change based on the selected different PS values (PS ≥ 0, 0.25, 0.50, and 0.75) for the Julian day of 60. Figure 4 shows that snow cover retreats toward the higher elevations as the PS value increases. This finding is similar to Richer et al. (2013) and is also in agreement with the general snow cover pattern where high-elevation regions tend to be more snow covered than the lower regions.

Fig. 4
figure 4

The effect of PS value on the SCA for Julian day 60

Full size image

As indicated in Burgos et al. (2013), PS provides an estimate of the probability that the pixel is actually snow or land. A threshold should be selected so that PS maps should be converted to binary SCA maps where snow is indicated by 1 and no snow areas should be indicated by 0.

Contingency tables (Table 2) are prepared for the whole period covering 2004–2012, and the omission (C / (C + D)) and commission (B / (A + B)) errors and their sums are calculated. The optimum threshold is determined following the approach of Burgos et al. (2013), in which the sums of omission and commission errors are minimized. Unlike Burgos et al. (2013), in this study, the computations are performed for a probability step size of 0.125 which yields nine classes from 0 to 1. For each Julian day and for each threshold, the summation of omission (OMI) and commission (COM) errors are determined. Then, for each Julian day, the PS is attributed to the threshold enabling the smallest summation of omission and commission errors. Figure 5 presents the variation of OMI and COM summation against PS values for Julian day 60 for the year of 2004. Figure 5 clearly indicates that PS value of 0.75 corresponds to the smallest OMI and COM summation.

Table 2 Contingency table representation for the IMS PS maps and ground truth data
Full size table
Fig. 5
figure 5

Variation of OMI and COM summation against PS values for Julian day 60 for the year of 2004

Full size image

For the same day, March 1, the variation of OMI and COM summations over 2004–2012 period are also calculated and presented in Fig. 6. It can be seen from Fig. 6 that minimum OMI and COM summation values changed from year to year. For this reason, the mode of the PS value that minimizes the OMI and COM summation is selected for that particular Julian day. Same procedure is repeated for each Julian day, and the optimum PS values are determined on a daily basis. Figure 7 shows the variation of PS values that yielded the minimum OMI and COM summation throughout the year. Figure 7 indicates that there is not a constant PS value that minimizes the OMI and COM summation; on the contrary, it changes seasonally, even daily.

Fig. 6
figure 6

Variation of OMI and COM summations against PS values for Julian day 60 over 2004–2012 period

Full size image
Fig. 7
figure 7

Variation of PS values that yielded the minimum OMI and COM summation throughout the year

Full size image

Figure 8 represents the time series of PS with respect to the elevation. Based on the time series of PS values, the entire study area seems to be snow free from Julian day 150 to 284. Also, it can be seen in the figure that regions with elevation higher than 600 m seems to be more snow covered than the other elevations. Although snow is common during winter time, PS values generally increase for the same date with increasing elevation. Starting with the first weeks of April (Julian day 90), snow cover retreats to 950 m and stays on the ground till Julian day 153 (1 June). With the first week of October (Julian day 283), snow cover reappears starting from elevations of around 1200 m. The reason for no snow-indicated regions in Fig. 8 for elevations above 2000 m after Julian day 100 may be due to the relatively coarse resolution of IMS (4 km) in comparison to the highly complex topography at these high altitudes.

Fig. 8
figure 8

Plot of PS time series with respect to elevation

Full size image

The optimally selected PS values (Fig. 7) are used to convert PS images into binary snow cover maps. To test the effectiveness of the proposed methodology in snow-covered area estimation, snow cover map for Julian day 60 of 2013 (corresponding to 1 March 2013) was downloaded and processed. Figure 9a presents the binary snow cover map that is obtained from the original IMS snow cover map for Julian day 60 of 2013. PS value that yielded minimum OMI and COM summation for the Julian day 60 was found to be 0.875 (Fig. 7). Figure 9b presents the binary snow cover map obtained from PS image of Julian day 60 by setting the PS threshold value to 0.875. Figure 9c represents the difference between two images (Fig. 9a, b). In Fig. 9c, gray areas indicate a match between Fig. 9a, b, meaning both indicated “snow” or “no snow.” Blue areas indicated “overestimation,” whereas red areas present the “underestimation” of snow-covered areas by PS methodology. Figure 10 shows the variation of match, overestimations/underestimations over the meteorological sites used in this study. Five (four) stations indicated under (over)estimation, 1 (−1). The majority of the stations matched between the two, indicating that both snow cover maps represented same land cover type for the pixels where the meteorological stations are located. Figure 11 represents the variation of overestimations (blue) and underestimations (red) with elevation. It is seen that overestimation follows a unimodal distribution over a narrower range. Underestimation also follows a unimodal distribution; however, values are spread over a wider range, and it is rather skewed to the right.

Fig. 9
figure 9

Binary snow cover obtained from IMS for Julian day 60 of 2013 (a) by PS > = 0.875 for Julian day 60 (b) Difference image between a and b (c)

Full size image
Fig. 10
figure 10

Variation of overestimations and underestimations over the ground truth stations for Julian day 60

Full size image
Fig. 11
figure 11

Distribution of overestimations (blue) and underestimations (red) with respect to elevation

Full size image

Conclusions

In this study, a methodology is proposed and implemented over Turkey to estimate the snow-covered area information. By the introduced methodology, IMS pixel-wise probability of the snow determination is possible and such information can enable determining the possible snow cover extent. As the snow-covered area information is one of the major parameters in hydrological applications such as snowmelt runoff estimations in hydrological models, the output of this study can improve the hydrological forecasting studies in the first hand. For such purpose, PS values were calculated using the long-term satellite-based snow cover maps as the first step. PS values enabled investigation of the temporal and spatial variation of snow cover. The variation of PS indicated the dependence of snow cover on altitude, with the higher PS values observed in high elevations. Furthermore, PS time series enable monitoring the retreat of snow cover.

To convert the PS images to binary snow-covered area maps, a threshold is needed. This optimum threshold was determined by minimizing the summation of commission and omission errors utilizing the contingency tables constructed for each day covering the 2004–2012 period using 219 ground stations. The needed binary snow cover maps were then obtained from the PS maps by using the optimum threshold for the day under consideration. To test the efficiency of the methodology, the binary snow cover maps obtained from PS maps using optimum PS value and original IMS snow cover maps are compared. Two hundred ten ground stations (95.89 %) out of 219 indicated the same land cover type (snow/no snow). Five stations (2.28 %) indicated underestimation, and the remaining four (1.83 %) showed overestimation of the snow cover.

Investigation of PS time series (Fig. 8) indicates that some of the high-elevation regions could not be mapped as snow covered. This points out that 4-km spatial resolution of IMS images should be increased to higher resolution for topographically complex regions.

An add-on value of the present study is that as PS time series showed variation of snow cover in both temporal and spatial cases, they can also be used in selecting new snow station locations or updating the present ones. As a follow-on of the present study, the effect of the minimum and maximum snow cover extents on the optimum threshold values can be investigated. Moreover, all the territory of Turkey is treated as a single region in the present study. Performing the analysis for sub-regions which can be based on either geographical considerations (regions in Table 1) or elevation bands may enable better region-specific analysis and results.