Andean watersheds present important snowfall accumulation mainly during the
winter, which melts during the spring and part of the summer. The effect of
snowmelt on the water balance can be critical to sustain agriculture
activities, hydropower generation, urban water supplies and wildlife. In
Chile, 25 % of the territory between the region of Valparaiso and
Araucanía comprises areas where snow precipitation occurs. As in many
other difficult-to-access regions of the world, there is a lack of
hydrological data of the Chilean Andes related to discharge, snow courses,
and snow depths, which complicates the analysis of important hydrological
processes (e.g. water availability). Remote sensing provides a promising
opportunity to enhance the assessment and monitoring of the spatial and
temporal variability of snow characteristics, such as the snow cover area
(SCA) and snow cover dynamic (SCD). With regards to the foregoing questions,
the objective of the study is to evaluate the spatiotemporal dynamics of the
SCA at five watersheds (Aconcagua, Rapel, Maule, Biobío and Toltén)
located in the Chilean Andes, between latitude 32.0 and 39.5
Snowmelt-driven watershed systems are highly sensitive to climate change, because their hydrologic cycle depends on both precipitation and temperature, and because water is already a scarce resource subject to ever-increasing pressure for its use (Barnett et al., 2005; Vicuña et al., 2011; Meza et al., 2012; Valdés-Pineda et al., 2014). Snowmelt controls the shape of the annual hydrograph, and affects the water balance at monthly and shorter timescales (Verbunt et al., 2003; Cortés et al., 2011). The effect of snowmelt on the water balance can be critical to sustain agriculture activities, hydropower generation, urban water supply and wildlife habitat quality (e.g. Vicuña et al., 2012, 2013).
The Andean watersheds present an important snowfall accumulation mainly during the austral winter; snow melts during spring and usually also during part of the summer, depending on relative altitude and ambient temperature. At higher elevations, a snowpack stores significant volumes of water, which are released to the surface runoff and groundwater when solar radiation increases.
In particular, 25 % of the Chilean territory between the Valparaiso and Araucanía regions is contained in areas where snow precipitation occurs (DGA, 1995). As in many other difficult-to-access regions of the world, the Chilean Andes – unlike western North America or the European Alps – have limited availability in temporal and spatial extent of hydrological data like discharge data, snow courses, and snow depths (Ragettli et al., 2013), which complicates the analysis of important hydrological processes and the validation of water quantity prediction models.
In this regard, remote sensing provides a promising opportunity to enhance the assessment and monitoring of the spatial and temporal variability of different variables involved in the precipitation–runoff processes in areas where data availability for hydrological modelling is scarce (Simic et al., 2004; Boegh et al., 2004; Melesse et al., 2007; Montzka et al., 2008; Milzowa et al., 2009, Er-Raki et al., 2010; Stehr et al., 2009, 2010).
Satellite-derived SCA from products like NOAA-AVHRR or MODIS can be used to enhance the assessment and monitoring of the spatial and temporal variability of snow characteristics (Lee et al., 2006; Li and Wang, 2010; Marchane et al., 2015; Wang et al., 2015), especially when it is combined with field data and snowpack models (Kuchment et al., 2010).
Specific spectral reflectance of snow (higher reflectance in the visible spectrum, compared to the mid-infrared electromagnetic spectrum) allows SCA to be accurately discriminated from snow-free areas using optical remote sensing methods (in the absence of clouds or vegetation canopies). Compared with other remote sensing techniques such as microwaves – which can be used to map snow water equivalent (SWE) – optical remote sensing, which is used to map snow areal extent (SAE), has a much higher spatial resolution (Zhou et al., 2005; Zeinivand and De Smedt, 2009).
Previous a Previous studies have compared MODIS snow maps with ground observations and snow maps produced by the National Operational Hydrologic Remote Sensing Center, USA (NOHRSC) (Hall et al., 2002; Klein and Barnett, 2003; Tekeli et al., 2005; Aulta et al., 2006). Klein and Barnett (2003) compared MODIS and NOHRSC products with ground observations (SNOTEL measurements), obtaining an overall accuracy of 94 and 76 %, respectively. The MOD10A2 snow product is capable of predicting the presence of snow with good precision (over 90 %) when the sky is clear (Zhou et al., 2005; Liang et al., 2008a; Wang et al., 2008; Huang et al., 2011; Zheng et al., 2017). All the previous studies were done in watersheds with a surface topography not as complex as in the Chilean Andes, where we have very steep slopes and a great oceanic effect due to the short distance between the coastline and the mountains.
Many of the studies on SCA changes and variability were done in the Northern Hemisphere, where topographic and orographic conditions are different from the ones we have in Chile. Research studies generally show negative trends in snow cover extent and snow water equivalent across both North America and Eurasia.
In Europe, Krajcí et al. (2016) analysed the main Slovak watersheds; for the 2001–2006 period, they obtained an increase in mean SCA; however, a significantly lower SCA is observed in the next 2007–2012 period. Their results indicate that there is no significant change in the mean watershed SCA in the period 2001–2014. Dietz et al. (2012), in their study of snow cover characteristics in Europe, found some abnormal events when comparing the mean conditions with single snow cover seasons. For the season 2005/2006 in particular, an increased snow cover with a later snow cover melt was detected. Marchane et al. (2015) studied the SCD over the Moroccan Atlas mountain range from 2000 until 2013, concluding that SCA has a strong inter-annual variability and that there is no statistical evidence of a trend in that period.
In Asia, two important regions have been studied: the Tibetan Plateau and the Himalayan region. In the case of the Tibetan Plateau, Zhang et al. (2012) have studied the SCD of four lake watersheds for the period 2001–2010; results indicate that spatial distribution and patterns of snow-covered days are very stable from year to year, and that there is no trend of snow cover change for each watershed. For the same period, Tang et al. (2013) found a high inter-annual variability of SCA, with parts of the studied area showing a declining trend in SCA, and other parts showing increasing trends in SCA. Wang et al. (2015) evaluated trends in SCA, showing a decrease in snow-covered areas from 2003 until 2010. For the Himalayan region, Maskey et al. (2011) studied the trends and variability of snow cover changes above 3000 m during the 2000–2008 period, showing a decreasing trend of snow cover during January and increasing trends during March. Snehmani et al. (2016) studied the 2001–2012 winter period (November–April) in order to obtain clear negative snow cover trends in the basin. Azmat et al. (2017) indicated that the watershed shows a consistent or slightly decreasing trend of snow cover, particularly over the high-altitude parts of the watershed during 2000–2009, and in the 14-year analysis (2000–2013), a slight expansion in the snow-covered area was observed in the whole basin. Gurung et al. (2017), based on the analysis SCA data between 2003 and 2012, also obtained a decline in SCA, with a statistically significant negative correlation between SCA and temperature, which indicates that this trend is partly a result of increasing temperatures.
In North America, Pederson et al. (2013) found that, since 1980, in the northern and southern Rocky Mountains the 1 April snow water equivalent (SWE) has been changing synchronously, and generally declining, associated with springtime warming. Fassnacht and Hultstrand (2015) examined trends in snow depth and SWE from three long-term snow course stations in northern Colorado from 1936 to 2014. They found negative trends at two of the stations for all the period and, at the third station, a positive trend was found for the first half of the record, and a decrease over the second half. Harpold et al. (2012) analysed SNOTEL SWE for the central and southern Rocky Mountains for the period 1984–2009; they found widespread decreases in maximum SWE and duration of snow cover. After a review of various studies in North America, Kunkel et al. (2016) concluded that all of them are consistent in indicating a decrease in snow on the ground, with some of the most extreme low values occurring in the last 10–15 years.
Studies of snowpack variation done in the central Andes of Chile and Argentina show that the average regional maximum value of the SWE series displays a positive (though non-significant) trend with marked interannual variability ranging from 6 to 257 % of the 1966–2004 mean (Masiokas et al., 2006). Results from Cornwell et al. (2016), who made a SWE reconstruction between 2001 and 2014, indicate that the years 2002 and 2005 stand out for displaying large positive anomalies throughout the entire model domain. The northern part shows above-average accumulation in only 3 out of the 14 simulated years (2002, 2005 and 2007), whereas the other part of the study area shows above-average accumulation for 6 years (2001, 2002, 2005, 2006, 2008 and 2009). In particular, the years 2007 and 2009 show a bimodal spatial structure, with excess accumulation (deficit) in the northern (southern) area and the inverse pattern in the latter year.
From the literature review, it is evident that SCD is site-specific and
exhibits high variability among the different sites around the world. The
Andes mountain range (on the Chilean side) has a great oceanic effect due to
the short distance between the coastal line and the mountains, unlike the
Northern Hemisphere, which has a big continental influence. This makes the
isotherm 0 higher than in the Northern Hemisphere; in addition, the
topographic features characterized by mean heights above 3000 m, with very
steep slopes that produce a huge orographic effect that forces the rise of
the western winds and condensation of moisture, strongly affect the regime of
precipitation and temperatures. This causes small-scale spatial variations in
weather, which are sometimes difficult to identify in satellite imagery. All
the previously mentioned aspects make it necessary to do a validation of
MOD10A2 before using it for other analyses. Indeed, one of the main sources
of error in the classification of satellite images is the interpretation of
topographic features, especially in areas with rugged slopes. The irregular
topography produces shading and lighting effects that change the radiometric
response of the surface, which depends on the local slope and its orientation
(Riaño et al., 2003). As the climate-controlling hydrologic processes in
the Andes are influenced by El Niño events (Escobar and Aceituno, 1993;
Ayala et al., 2014), as well as by warm winter storms (Garreaud, 2013),
assessment of SCD is of special interest in this region. Pioneer studies have
been presented in recent years for the Mataquito River watershed
(Demaría et al., 2013; Vicuña et al., 2013). The aim of the study is
to evaluate the spatiotemporal dynamics of the SCA in the southern–central
Chilean Andes, and to analyse its relationship with the precipitation
regime/pattern and El Niño–Southern Oscillation (ENSO) events. To prove
this, we selected five watersheds located in the Chilean Andes, between
latitude 32.0 and 39.5
Study sites.
The study site includes five watersheds located in central and southern Chile: Aconcagua, Rapel, Maule, Biobío and Toltén (Fig. 1). The watersheds were chosen considering their population and dependence on water-resourced economic activities.
The Aconcagua watershed is located in the Valparaiso region, between the
parallels 32
The Rapel watershed is located in the General Libertador Bernardo O'Higgins
region between the parallels 33
The Maule watershed is located in the Maule region, between the parallels
35
The Biobío watershed is located in the Biobío region, between the
parallels 36
The Toltén watershed is located in the Araucanía region, between the
parallels 38
For validation of MOD10A2, ground observations were performed, including continuous monitoring of snow depth with meteorological stations, snow courses, and 1-day observations of snow presence and depth at some mountain trails. Figure 2a shows the locations of measuring sites.
In order to perform a continuous measurement of snow depth, three meteorological stations were installed in the upper part of the Biobío watershed. The stations are Parque Tolhuaca, located at 900 m a.s.l., Termas Malleco at 1190 m a.s.l., and Laguna Verde at 1410 m a.s.l. Also, snow depth data from three DGA meteorological stations (Portillo, 3005 m a.s.l., Laguna Negra, 2709 m a.s.l., and Volcan Chillan, 1964 m a.s.l.) were used. Figure 2b shows the location of the meteorological stations.
Snow depth was measured using acoustic snow depth sensors (Campbell
Scientific, SR50A) with a frequency of 15 min. Corrections for variation of
the speed of sound in air were made considering the air temperature measured
at the same time intervals as the snow depth, using the temperature and
relative humidity probe (Vaisala, HMP60). The data were collected from
April 2010 to December 2011 at the Termas Malleco station, and from July 2011
to December 2011 at the Parque Tolhuaca and Laguna Verde stations. In the
case of the DGA station, available data from April 2013 to December 2015 were
used. Snow data were grouped considering the average snow depth over 8 days
(same 8-day period of MOD10A2) and then reclassified as snow (1) if the
average snow depth was > 0 cm and as no-snow (0) for snow
depth
Dates and repetition times for each snow route during winter 2011.
To cover a large spatial domain and to achieve a better spatial and temporal representation of SCA, a total of 124 different single measurements of snow depth were conducted during field campaigns at different mountain trails. Field measurements were taken from the end of June until the beginning of October 2011. The location of each observation point was recorded with a GPS and snow depth was measured with a Black Diamond QuickDraw Tour Probe 190. Figure 2c shows the location of the observation points and the date and number of observations for each day.
During 6 days of the 2011 winter season, 11 snow courses were conducted in the upper Malleco watershed. Figure 2d shows the location of snow courses. Each route was recorded with a GPS. Snow depth and density were measured with a Black Diamond QuickDraw Tour Probe 190 and Snow Sampling Tubes (3600 Federal Snow Tubes, Standard-Metric), respectively. Table 1 shows the dates and repetitions for each snow route performed.
The Moderate Resolution Imaging Spectroradiometer (MODIS) is on the Earth
Observing System, which employs a cross-track scan mirror and a set of
individual detector elements to provide imagery of the Earth's surface and
clouds in 36 discrete and narrow spectral bands ranging in wavelength from
0.405 to 14.385
Confusion matrix and percentage correct (overall accuracy), user accuracy and producer accuracy.
MODIS estimates of SCA were validated for the period between April 2010 and December 2011 by comparing them with ground observations. Eighty snow maps from MOD10A2 were analysed. Only cloud-free observations for each cell of the map grid were used.
To validate the correspondence between the image classification and ground observations, a confusion matrix was used (Congalton and Mead, 1983; Story and Congalton, 1986, 1991; Foody, 2002). The confusion matrix is a simple cross tabulation of classified data versus observed ones, providing a basis for accuracy assessment (Campbell, 1996; Canters, 1997). Figure 3 presents the confusion matrix and how the different indexes were calculated.
Precipitation stations and availability of data.
SCA variation in the five selected watersheds was evaluated using MOD10A2 for the period covering the years 2000–2016 using a total of 774 images. All images were processed in ArcGis, reprojecting them to WGS84 19S and cut according to the desired area, in this case for the Aconcagua, Rapel, Maule, Biobío and Toltén watersheds. SCA and clouds were quantified for each available image, and annual and seasonal averages of snow cover were calculated for each watershed.
SCD was analysed through a number of statistical tests to explore observed trends in SCA during the study period. “Non-parametric” statistical tests were chosen because they are more robust than the “parametric” test. A Mann–Kendall test was applied to determine the existence of monotonic trends. Sen's method (Gilbert, 1987) was applied to determine the rate of observed changes.
Confusion matrix and indexes of agreement for Parque Tolhuaca, Termas Malleco and Laguna Verde stations.
An analysis of the correlation between SCA and mean annual precipitation for the upper part of the watershed was done. Considering the spatial location of SCA, only precipitation stations located over the 700 m a.s.l. were considered. Availability of daily precipitation was evaluated for each year at each station; only years with more than 80 % of data were used in order to obtain the mean annual precipitation. The precipitation for the watershed was obtained considering the arithmetic mean between available stations. Table 2 shows the stations that were used and the availability of data.
SCD for the 2000–2016 period at Aconcagua, Rapel, Maule, Biobío and Toltén watersheds. Red boxes indicate Niña years, blue ones Niño years.
Mean seasonal and annual SCA (%) in Aconcagua, Rapel, Maule, Biobío and Toltén watersheds during 2000–2016. Red boxes indicate Niña years, blue ones Niño years.
Relation between % of mean annual SCA and annual precipitation at Aconcagua, Rapel, Maule, Biobío and Toltén watersheds.
Spatial representation of SCA for the period 2000–2016 in Aconcagua, Rapel, Maule, Biobío and Toltén watersheds. Black indicates the year with maximum SCA, dark grey the year with average SCA and grey the year with minimum SCA.
Annual and seasonal trends of SCA at Aconcagua, Rapel, Maule, Biobío and Toltén watersheds.
The MOD10A2 product was validated for determination of SCA in the watersheds under study and the SCD was analysed.
The composite images MOD10A2 were compared with ground observations. For the study period 23, 75, 26, 119, 123 and 32 images were available for comparison with observations at Parque Tolhuaca, Termas Malleco, Laguna Verde, Portillo, Laguna Negra and Volcan Chillan stations, respectively. Clouds were present in 2 (9 %), 1 (1 %), 1 (4 %), 3 (3 %), 3 (2 %) and 2 (6 %) images for each aforementioned station, respectively. Table 3 presents the confusion matrix and the indexes of agreement. Overall accuracies of 86, 81, 88, 92, 83 and 97 % were observed at Parque Tolhuaca, Termas Malleco, Laguna Verde, Portillo, Laguna Negra and Volcan Chillan stations, respectively, which reached the target of 85 % (Thomlinson et al., 1999).
A total of 117 images corresponding to the study period were available for comparison with 1-day observations, with none of them classified as covered by “clouds”. All ground observations were done over areas with snow presence. When comparing it to MOD10A2 images, the agreement was 97 %. MODIS that did not coincide with ground observations occurred during the beginning of the snowfall season or at the end of the melting period. In both cases, snow patches were observed in the field, covering areas at the subpixel scale of the MOD10A2 image.
For comparison with snow courses, a total of 282 images were available during the study period, all of them without clouds. Ground observations were done on areas covered by snow only. The overall accuracy of MOD10A2 for predicting SCA was 98 %.
Results indicate that MOD10A2 has a satisfactory agreement with ground observations, and therefore the 8-day composite images are suitable for analysis of SCD in the Andean watersheds.
Figure 4 shows the SCD in Aconcagua, Rapel, Maule, Biobío and Toltén watersheds for the period 2000–2016. All watersheds show the same SCD with more SCA in winter than the other seasons. It can be appreciated that in the Maule, Biobío and Toltén watersheds the maximum SCA in 2016 is considerably lower than in the 5 previous years, and is one of the lowest of the whole study period.
Figure 5 shows mean SCA for each season and annual precipitation in the
Aconcagua, Rapel, Maule, Biobío and Toltén watersheds for the period
2000–2016. Northern watersheds, i.e. Aconcagua and Rapel, present a higher
SCA percentage. It can be appreciated that years with more precipitation do
not necessarily have more SCA; instead, years with lower precipitation have a
higher SCA. Figure 6 displays the relation between SCA and annual
precipitation for the different watersheds; it is clear that only at
Aconcagua is there a good adjustment to the regression line (
Figure 7 shows the variability of spatial distribution of maximum yearly SCA for the five watersheds. Years with maximum and minimum SCA of yearly maximum SCA are not coincident between watersheds. A considerable spatial variability of maximum SCA can be appreciated. There are no trends in the data.
Figure 8 shows annual and seasonal trends of SCA at the Aconcagua, Rapel,
Maule, Biobío and Toltén watersheds. Trend analysis of the SCA
series was performed with the non-parametric Mann–Kendall test. Decreasing
trends in annual mean SCA are observed (
Considering snow accumulation and melt seasons, data were grouped in two
periods, i.e. autumn–winter (accumulation) and spring–summer (melt). For
these two periods, a trend analysis was also done. Results indicated that
there is a negative trend in average SCA during both seasons; the Aconcagua
watershed has a decreasing slope (
In general terms, we can say that MOD10A2 has good agreement with ground observations, with always over 80 % precision when the sky is clear. Our results are a bit lower than the 90 % obtained in studies of the Northern Hemisphere (Zhou et al., 2005; Liang et al., 2008a; Wang et al., 2008; Huang et al., 2011; Zheng et al., 2017); this difference can be attributed to the complex topography of the Chilean Andes. Greater disagreements between MOD10A2 and ground observations were found at Termas Malleco and Laguna Negra stations. These stations are located close to the snowline where most of the time the soil is not covered by snow or is covered by snow for small periods of time; in this case, the presence of small snow patches is expected to be highly variable in time; it can even change at the subdaily scale and therefore it could reduce the capacity of images to capture snow coverage in the proximity of the snowline. Regarding SCA, we can see in the results presented in Figs. 4 and 5 that SCA has a similar behaviour in the different watersheds during the study period, i.e. all watersheds follow the same snow accumulation and melt dynamics and the maximum and minimum % of SCA are also coincident in time. SCA has a strong inter-annual variability, which is in agreement with the results obtained by other studies (Masiokas et al., 2006; Marchane et al., 2015; Tang et al., 2013; Zhang et al., 2012). Comparing the result with those obtained by Karici et al. (2016) in Slovak watersheds, we can appreciate similar results for the Aconcagua, Rapel and Maule watersheds and opposite ones for Biobío and Toltén, i.e. for the 2001–2006 period a decrease in mean SCA with respect to the mean of the whole period, and more SCA for the period 2007–2012.
Comparing our results of SCA with those of SWE obtained by Cornwell et al. (2016), we can appreciate that they are in agreement only with respect to the years 2002 and 2005 displaying the largest SCA of the entire study period. The other results are in dissension, as we have more years with above mean SCA during the beginning of the study period, and below throughout the end of the period (2012–2016).
In the case of trend analysis only the watersheds located in the northern part of the study area have a significant decreasing trend during accumulation and melt season, which is in agreement with results from different studies in Asia and North America (Tang et al., 2013; Maskey et al., 2011; Gurung et al., 2017; Fassnacht and Hultstrand, 2015; Kunkel et al., 2016). During wintertime we could not see any significant trend, which is consistent with results from Dietz et al. (2012).
The SCA magnitude for each watershed was contrasted with historical data from
El Niño/La Niña–Southern Oscillation (ENSO) episodes from 2000 to
2016
Niña years are coincident with maximum SCA in wintertime in all watersheds, with similar amounts of SCA between different Niña episodes. The 2007 Niña event is coincident with the highest % of SCA during winter for the study period (2000–2016) in all watersheds. Winter SCAs for Niña and Niño episodes are all above or just at the mean SCA for the 14 years under study. Years under the mean of the SCA area are all not coincident with Niña and Niño episodes, i.e. normal years are all under the mean SCA value. There is also no linear relation between the amount of precipitation and SCA, with the exception of the Aconcagua watershed (Fig. 6). This study only analysed snow coverage and not snow depth, i.e. volume of snow, which implies that a bigger SCA is not necessarily related to more water stored as snow. Considering the existing data availability in time and space of the DGA regarding snow depth, it was not possible to carry out an analysis of changes in snow depth in this study. During the course of the last 5 years, an important decline in SCA was perceived in the entire watershed, with the exception of the Biobío watershed, where no difference can be seen until the last year (2016). This decline is bigger at the Toltén watershed, where the average for the last 5 years is 50 % of the mean for the entire period. In the analysis of the precipitation data we can observe that in all watersheds there was a rainfall deficit during 2012–2016: Aconcagua 26 %, Rapel 29 %, Maule 24 %, Biobío 19 %, and Toltén 27 % (DGA, 2012, 2013, 2014, 2015, 2016; DMC, 2013, 2014).
Masiokas et al. (2006) studied snowpack variation between 1951 and 2005 in
Chile between latitude 30 and 37
In the watersheds located south of 35
The first validation of MODIS snow product MOD10A2 for estimation of snow covered areas (SCAs) via remote sensing in watersheds located in the Southern Hemisphere was presented. Ground observations of SCA were conducted during the years 2010 and 2011 at six study sites including six meteorological stations, 124 1-day single-observation points, and 11 snow courses. The SCA was determined for 636 days from MODIS snow products and compared with the SCA measured in situ. The SCA estimated from MOD10A2 presented an overall agreement from 81 to 98 % with SCA determined from ground observations, showing that the MODIS snow product can be taken as a feasible remote sensing tool for SCA estimation in southern–central Chile.
On the other side, we have analysed SCA trends for the period 2000–2016 in
five of the most important watersheds of southern–central Chile in the
context of water use and inhabitants, covering a longitudinal gradient from
32
Results were compared with the ENSO episode during 2000–2016. From the latter comparison we can conclude that in the previously mentioned period, Niña years are coincident with maximum SCA at wintertime in all watersheds, with similar amounts of SCA between different Niña episodes.
In summary, the results presented in this work are highly relevant and can be used as one feasible approximation to obtain SCA, particularly in the Chilean Andes because of the lack of hydrological data such as discharge data, snow courses, and snow depths. These can become an outstanding tool to improve the analysis of important hydrological processes and the validation of water quantity prediction models.
MOD10A2 are available from
(
The authors declare that they have no conflict of interest.
The present research was conducted in the framework of the FONDECYT 11100119 project. Edited by: Stefan Uhlenbrook Reviewed by: three anonymous referees