1. Introduction
Glacier mass balance is a key measure to determine the contribution of glaciers to regional hydrology [
1,
2] and global sea level [
3,
4,
5]. As the number of glaciers with direct measurements of mass balance (using the glaciological method) is limited (e.g., [
6]) and their representativeness for the larger mountain ranges often unknown [
7], it is appealing to obtain glacier mass changes from remote sensing data. Over larger regions the differencing of digital elevation models (DEMs) from two points in time (geodetic method) has been applied widely at approximately decadal resolution to determine total volume and mass changes of glaciers [
8,
9,
10,
11]. However, for several geoscientific and socio-economic applications, the availability of region-wide glacier mass changes with annual resolution would be beneficial. As direct measurements reveal a high correlation of glacier mass balance with (a) the snow cover on a glacier and (b) the elevation of the snow line altitude (SLA), mapping snow cover (SC) on glaciers from satellite images offers a proxy for glacier mass balance (e.g., [
12,
13,
14]), whereby the remotely sensed snow cover ratio (SCR) is taken as a proxy for the accumulation area ratio (AAR) [
12,
13,
15] and the elevation of the snow line at the end of the ablation period [
12,
16] as a proxy for the equilibrium line altitude (ELA). To compensate for local variability in elevation, the ELA is defined mathematically as the elevation where the vertical mass balance profile obtained by direct measurements crosses zero [
17]. Mapping SC extent has also been applied to reconstruct missing mass balance measurements [
18] on the basis of manually selected satellite scenes coinciding with the end of the ablation season, revealing a high correlation between satellite-derived snowlines and ELAs measured in the field.
Snow and (clean) glacier ice together can be easily mapped from multispectral satellite images by application of various band ratios like near infrared (NIR) / shortwave infrared (SWIR) [
19], red / SWIR [
20], or the normalized difference snow index (NDSI), which is given as (green – SWIR) / (green + SWIR) [
21], and a manually selected threshold. In contrast, the classification of snow on glaciers is difficult with these band ratios, as the shape of the spectral curves of ice and snow is very similar and ratios thus result in about the same values for both facies. A robust and widely used approach for mapping SC on glaciers is based on the use of a threshold on either single-band reflectance values, preferably in the NIR to avoid saturated values over snow with 8-bit Landsat data [
12,
19,
22,
23,
24,
25,
26,
27,
28,
29], or albedo products that are readily available [
25]. Mapping snow extent on glaciers has thus come a long way, starting in the 1970s with early tests using contrast-enhanced Landsat multispectral scanner (MSS) satellite images [
22,
23], to the recent use of near-daily MODIS data to derive time series of surface albedo and SC maps for even small glaciers (e.g., [
24,
25,
30]).
For repeat application over large regions, it would be beneficial to derive the snow-covered area on glaciers automatically, including determination of the snow line altitude (SLA) near the end of the ablation season (as a proxy for the ELA). This is made challenging because of the following: (i) Due to the strong impact of terrain orientation (defined by its normal vector) on surface reflectance in mountain topography, deriving any measure of SC from reflectance requires implementation of a topographic normalisation using a DEM; (ii) SC on glaciers is often patchy and snow lines are in general not parallel to elevation contour lines; and (iii) highly variable atmospheric conditions (e.g., due to cloud shadows) imply that the same threshold might not work equally well on all glaciers. It is thus crucial for any automated algorithm to find a threshold that is adjusted to the local conditions, i.e., for each glacier individually, and as a consequence, for smaller glacier samples, snow lines have often been digitized manually (e.g., [
13,
31,
32,
33]).
To map SC on comparably small glaciers such as in the Alps and for analysis over long time periods, the 30 m resolution Landsat data are the best choice as they span 1984 up to the present and have higher spatial resolution than MODIS data (250 m). On the other hand, Landsat data have the disadvantage of a much coarser temporal coverage (every 16 days), so autumn snowfall and cloud cover can make it difficult to obtain useable images from the end of the ablation period in certain years. As surface elevation of glaciers changes over time, there is a need for multi-temporal DEMs to correctly determine SLA values for long time series.
In this study, we present the fully automated snow mapping on glaciers (ASMAG) tool to (a) map SC on glaciers, and (b) derive the SLA from the 30 year time series of multi-temporal Landsat data and a DEM. The method is applied to the glaciers in the southern Ötztal Alps as three glaciers with long-term mass balance measurements are located there (Hintereisferner, Kesselwandferner, and Vernagtferner) providing data for validation of the tool. As explained above, the specific challenges for an automated approach are the discrimination of snow from glacier ice, consideration of clouds (or their shadows) affecting parts of a glacier, terrain shadow, the often patchy nature of snow cover, glacier surface debris cover, and methodological constrains such as the impact of a time invariant DEM. Our method has been designed to handle these and automatically finds a glacier-specific reflectance threshold to separate snow from ice. As an introduction to our methodological approach, we present the workflow, discuss the main challenges, and offer a comprehensive assessment of the accuracy of the approach using three independent methods of validation. We also present the results for the study region (SC ratio and SLA time series) in comparison to field data and investigate potential reasons for deviations.
2. Study Region
The tool is tested in the Ötztal Alps, Tyrol, Austria (
Figure 1) covering 26 glaciers in three different basins (Niedertal, Rofental and Gepatsch), all draining northward. The southern part of the study region includes the main Alpine divide, which often acts as a cloud barrier for weather patterns driven by northerly and southerly flow. In summer, orographic convective clouds are frequent above mountain ridges, often casting shadows on glaciers. The glaciers span ~2150 to ~3750 m a.s.l., cover several aspect directions, have different slopes and shading conditions, and variable hypsometries, collectively resulting in a locally variable distribution of SC.
The total area of all selected glaciers is 71 km
2 in 2011 (year of the glacier mask). The largest glacier is Gepatschferner (no. 21 in
Figure 1) with 16.3 km
2, followed by Schalfferner (6.5 km
2, no. 5), Vernagtferner (6.1 km
2, no. 13), Taschachferner (5.5 km
2, no. 18), Hintereisferner (5.2 km
2, no. 11), Marzellferner (4 km
2, no. 6) and Kesselwandferner (no. 13) with 3.6 km
2. Of these, Hintereisferner (HEF), Kesselwandferner (KWF), and Vernagtferner (VNF), have long and continuous time series of mass balance measurements, ELAs and AARs [
18,
34,
35,
36,
37], providing a valuable dataset for comparison with the automatically generated SLA and SCR estimates. The highest glacier elevation is reached at Taschachferner (3745 m a.s.l.) and the lowest is at Gepatschferner (2148 m a.s.l.), which also has the largest elevation range (~1350 m). The glaciers often have steep slopes exposed to the north (ten glaciers) or northeast (six glaciers). The average value of the ELA
0 (balanced-budget ELA) is found between 3050 m and 3150 m a.s.l. (1953–2013) [
34]. At this elevation range, the mean annual air temperature is around −5 °C and the annual precipitation (solid and liquid) about 1500 mm [
34]. The ablation period usually lasts from May to September. Large area and mass losses have been observed in the Ötztal Alps since the mid-1980s [
38], indicating a continuous glacier shrinkage over the investigated period.
7. Conclusions
In this study, we presented a method (ASMAG) for the multi-temporal (1985–2016) retrieval of snow cover ratios (SCR) and snow line altitudes (SLA), and the results of its application to selected glaciers in the Ötztal Alps as well as Abramov Glacier in the Pamir Alay, Kyrgyzstan. The SC mapping is applied to each glacier individually using an automatically selected top of atmosphere reflectance (TOAR) threshold based on the Otsu method. Subsequently, the SC map is intersected with 20 m elevation bins of the DEM and its SLA is retrieved.
The overall performance of ASMAG can be seen as good to very good (90% accuracy), but local outliers exist. The SLA detection capability is better than 80% with an accuracy of ±19 m. If the acquisition date of the satellite scene is close to the date of the highest snowline, a good agreement of the derived SCR (SLA) values with the measured AAR (ELA) values is achieved, otherwise SCR (SLA) values are systematically higher (lower). However, glacier specific differences are still correctly captured.
Problems arise when clouds and their shadows are covering parts of glaciers or when fresh snow is present. A small amount of debris cover has no impact on the result, but for large debris covered glaciers a separate ice mask could be useful. A DEM from one point in time is important and useful for the topographic correction of the NIR band. When analysing a time period of more than ten years, a DEM of one point in time introduces errors of the SLA values especially in low laying areas of the glacier. In contrary, the unchanged glacier mask only has a minor impact on the SCR in long time series.
We conclude that ASMAG is a promising tool to derive SCR and the SLA for large glacier samples. The derived snow cover maps can be used to calibrate/validate modelled snow cover evolution and as a proxy for glacier mass balance if cloud-free satellite scenes acquired close to the date of the highest snowline are available. However, Landsat data alone is too sparse for accurate retrieval of maximum snow line elevations, particularly in mountain regions with frequent cloud cover. The inclusion of higher frequency optical satellite data such as the forthcoming abundance of Sentinel-2 imagery will help overcome this restriction and could be relatively readily incorporated into the tool.