This study presents a spatio-temporal continuous data set for snow cover in Iceland based on the Moderate Resolution Imaging Spectroradiometer (MODIS) from 2000 to 2018. Cloud cover and polar darkness are the main limiting factors for data availability of remotely sensed optical data at higher latitudes. In Iceland the average cloud cover is 75 % with some spatial variations, and polar darkness reduces data availability from the MODIS sensor from late November until mid January. In this study MODIS snow cover data were validated over Iceland with comparison to manned in situ observations and Landsat 7/8 and Sentinel 2 data. Overall a good agreement was found between in situ observed snow cover, with an average agreement of 0.925. Agreement of Landsat 7/8 and Sentinel 2 was found to be acceptable, with
On a global scale snow cover has a strong interaction with the cryosphere and ocean systems and therefore the climate system of the Earth. The two main effects of snow on the cryosphere are its control on the reflection of radiation, reaching the surface of Earth and balancing its radiation budget
Overview of Iceland. Red outlines show the main catchment boundaries for large hydropower diversions, reservoirs and power plants. Manned sites for observed snow cover are shown in green points. Contours are shown for the 200 m elevation band. A solid blue contour represents 200 m elevation.
In the Northern Hemisphere the spring snow cover extent has decreased significantly, influencing the dynamics of spring melt intensity and timing in recent years
On regional scales, seasonal snow is a vital part of water budgets in mountain and highland catchments where precipitation falls as snow during winter
Iceland is an island with an area of 103 100 km
Elevation distribution for Iceland for both land and glaciers. Glaciers cover about 11 % of Iceland.
In Iceland runoff from snowmelt is critical for hydropower production and reservoir storage as the energy system is strongly dependent on snowmelt and glacier melt. Over 13 % of the highland area in Iceland is developed for hydropower generation which provides over 72 % of the total average energy produced in Iceland
Spaceborne sensors operating in the visible and near-infrared range of electromagnetic spectrum have proven to be useful in effectively mapping snow cover for large areas since the early 1980s
A range of snow cover products has been developed from the Aqua and Terra satellites carrying the MODIS sensor dating back to the early 2000s
MODIS snow cover products have been widely tested and validated for various land covers, topographic regions, and climates, with a typical average absolute accuracy of 93 %
The aim of this study was to create a gap-filled snow cover product for Iceland and extract snow cover characteristics for the period from 2000 to 2018. The first objective was a thorough validation of MODIS sensor-derived snow-covered maps over Iceland to validate the quality of the product and assess its limitations. Validation was an important and necessary step due to the annual and seasonal variability in climate, high average cloud cover and polar darkness during winter. The second objective of the study was to reduce the gaps to provide a spatio-temporal continuous product. By merging of data and temporal aggregation methods, data gaps are reduced and finally eliminated by using classification learners trained on topography and location of pixels. Based on the gap-filled data set snow cover characteristics on a regional scale over Iceland were derived showing relations to elevation, aspect and general trends in snow cover extent and duration.
In Iceland in situ snow cover and depth observations are sparse, especially in the highlands. Few sites have had automatic observations of properties of snow until recently. The Icelandic Meteorological Office (IMO) operates a network of synoptic meteorological observations including daily manned observation of snow cover at 09:00. Figure
MOD10A1 (Terra) and MYD10A1 (Aqua) Version 6 were obtained from the National Snow and Ice Data Center (NSIDC)
Data acquired by the Landsat 7 Thematic Mapper (TM), Landsat 7 Enhanced Thematic Mapper Plus (ETM+), Landsat 8 Thematic Mapper (TM) and Landsat 8 Thermal Infrared Sensor (TIRS) were used. The data were downloaded from the United States Geological Survey (USGS) (
Data acquired by a European Space Agency (ESA) Sentinel 2A and B multispectral instrument (MSI) sensor were also used. The data were downloaded from the ESA's data hub (
Digital elevation models (DEMs) and masking data for water bodies and glaciers were obtained from the National Land Survey of Iceland. The original DEM is a raster with a 10 m spatial resolution which is resampled to match the grid of the MODIS pixels using nearest neighbor sampling. From the resampled 500 m DEM the aspect data are calculated.
Manned observations of snow cover from the IMO are reported daily at 09:00. Observations are made both at the local site where the instruments are located as well as in mountains where applicable; these are reported as local snow cover (SNC) and snow cover in mountains (SNCM). For each observation the local snow cover is reported as snow free (Code 0), patchy snow cover (Code 2) and fully snow covered (Code 4)
From the MOD10A1 and MYD10A1 daily data tiles we extracted the MOD Grid Snow 500 m grid and the variable NDSI snow cover was used for further analysis of snow cover. It is based on the MOD10-L2 algorithm which selects the best observation of the day to write to the daily data set. The variable NDSI snow cover ranges from 0 to 100, but in addition various other classifications are provided with the tile. As a preprocessing step data were reclassified to (a) snow, (b) no snow (land) and (c) no data (clouds, missing data, no decision, saturated detector). As the spatial extent of the tile is Daily tile merging: daily tiles from Aqua and Terra are merged to a single data set to improve daily coverage with data. Data from Terra have priority over data from Aqua as previous studies have found data from Terra to be of higher accuracy Temporal aggregation: for the remaining unclassified pixels in the daily merged data tiles (MCDAT) we apply temporal aggregation to further reduce unclassified pixels due to clouds in the data. Each MCDAT tile from step 1 is given a center date as the date of acquisition ( Gap filling with classifiers: after the first two processing steps the remaining gaps are classified as snow or no snow with classification learners. For each data set the unclassified pixels are reclassified with four predicting variables, location (easting, northing), elevation (
Figure
A simple process flow diagram for the daily tile merging, temporal aggregation and gap filling.
Landsat 7 and 8 data were retrieved as L1TP surface reflectance products. These products have a terrain correction and are radiometrically calibrated. The U.S. Geological Survey (USGS) uses the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) for Landsat 7 Surface Reflectance generation, while Landsat 8 is processed with the Landsat Surface Reflectance Code system (LaSRC)
Sentinel 2 data were retrieved as an L1C ortho-image product. Images are in top-of-atmosphere reflectances in cartographic geometry and have undergone geometric transformation and radiometric interpolation with a constant GSD (ground sampling distance)
Comparison of observed snow cover and MODIS daily combined snow cover. For the whole data set the overall classification accuracy was 0.925.
To classify the data we need to select a classification method. In general terms a model is trained with the training data set and the trained fit applied to the data that need classification. Within the Matlab Classification Toolbox
In general snowfall in a region and formation of a snowpack are dependent on several climate and geographic factors such as latitude, longitude, elevation, distance from moist sources (ocean and lakes) and regional air mass circulation
Relationship of classified snow pixels for Landsat 7/8 and Sentinel 2 with the MCDAT product.
Confusion matrix for observations of snow compared to the Modis Aqua and Terra daily snow product combined.
Overall a good agreement was found between in situ observed snow cover and MODIS daily combined snow cover (MCDAT). Figure
Confusion matrix for snow cover derived from Landsat 7, Landsat 8 and Sentinel 2 compared to the Modis Aqua and Terra daily snow product combined.
Table
For each Landsat 7/8 and Sentinel 2 tile a classification map was constructed. The classification maps show the agreement of different satellite sources to the MCDAT product. A selected sample of the maps was manually screened to identify patterns in misclassification. The screening reveals that disagreement was mainly located at snow cover boundaries, i.e., where snow-free land meets snow-covered land as well as boundaries of clouds and land. Previous studies in snow-covered Arctic and alpine areas have revealed a similar effect when comparing MODIS to higher-resolution data
Average cloud cover over Iceland based on the MCDAT product for February to November each year from 2000 to 2018.
Figures
Average cloud cover distribution between months.
In general the advantage of temporal aggregation of data is reduced cloud-obscured pixels, which provides a spatio-temporal continuous product. The trade-off of temporal aggregation contrasts with the dampening of the response of the snow cover to rapid melt or snowfall events. This poses a limitation on the use of the data in real-time applications such as short-term flow forecasting for water resources.
Average gap-filling improvement with merging of daily data and temporal aggregation.
After applying a temporal aggregation to the data, unclassified pixels still remained in the data set. To classify the remaining pixels, various classifiers were tested to assess their classification accuracy. Various configurations of classification trees,
A daily gap-filled snow cover product was derived for Iceland based on MODIS sensor bi-daily overpasses at a temporal resolution from 1 March 2000 to 30 June 2018. All water bodies and glaciers were excluded from the gap-filled snow cover product. Based on the data set, various descriptive spatio-temporal dynamics of snow cover in Iceland can be derived. The main limitation to the data set was polar darkness during December and January that limits the continuous temporal structure of the data set. Snow cover duration within a season is a parameter that is often used to describe characteristics of snow cover. The duration of snow cover is a property that can be linked to many applications such as seasonal snowmelt magnitude for operational water resources and length of vegetation growing season.
Snow cover duration (SCD) in Iceland from 1 March 2000 to 31 December 2017; 100 % indicates full snow cover, while 0 % represents a snow-free area. Glaciers and water bodies are not included.
Figure
Figure
Normal distributions for extracted first snow-free date, number of snow-free days and last snow-free date.
In various studies of snow cover where polar darkness applies, a filter assuming that if a pixel has snow at the beginning of polar darkness (late November in Iceland) and the same pixel still has snow when polar darkness recedes (mid January in Iceland), it can be assumed that the snow cover is continuous for that time period
First column: mean snow cover duration as percentage of time for each period. Second column: standard deviation of days for each period. Third column: mean trend in snow cover duration as percentage of time for each period. Rows represent different combinations of monthly values and the bottom row is for the whole period from February to November.
Figure
Monthly mean results for snow cover. For each month mean snow cover extent is calculated and a linear fit applied. The results from the Mann–Kendall test are shown as 1 or 0, where 1 indicates a significant change in trend. A linear equation shows the results for a fitted linear model to the data set.
Figure
Average elevation distribution of snow cover for 2000–2018. Fractions of area for each elevation band are 23.9 % for 0–200 m a.s.l., 17.5 % for 200–400 m a.s.l., 21.7 % for 400–600 m a.s.l., 18.4 % for 600–800 m a.s.l., 8.2 % for 800–1000 m a.s.l. and 9.9 % for elevations over 1000 m a.s.l.
Figure
Figure
As previously mentioned, the data set only spans 18 years, so statistical interpretation such as trends should be treated with care. To evaluate whether these trends are significant, a linear trend test and a Mann–Kendall test were performed on monthly mean snow cover extents for
Average aspect distribution of snow cover for 2000–2018. Fractions of area for each aspect are 22.8 % for 0–90
Figure
In this study, a gap-filled satellite-observed snow cover was produced from daily MODIS Aqua
Average cloud cover in Iceland is high (75 % average), providing a significant limitation to the application of MODIS data and all optical remote-sensing instruments. No significant temporal patterns were found in cloud cover, while the central highland in general has lower average cloud cover. This was addressed with temporal aggregation of data where the tradeoff from temporal aggregation (7 d) could have implications for hydrological applications of the data set where onset of melt and melt events could be retained or smoothed out of the product. This was also a limitation for identifying rain on snow events during winter.
Availability of MODIS data during polar darkness was also a temporal limitation for the data set. From late November to mid January no data were available, which limits the application of the data set to identify rain on snow events that can cause flooding and deplete areas of snowpack. Due to the dynamics of Icelandic snow during winter, especially at lower elevations, this is challenging to solve without combining other data sources such as snow models or other sources of remote sensing, for example synthetic aperture radar such as the ESA's Sentinel 1, which has a frequent overpass over Iceland.
The changes over time (trend) analyzed for the 18 years showed a slight increase in average snow cover in spring, likely driven by cold springs in 2013, 2014 and 2015 and extended liquid-phase precipitation in the fall for the same years. This aligns with observations of winter mass balance of Icelandic glaciers in recent years with a slight significant positive trend for the past 20 years
Another influencing factor for onset of melt and melt enhancement is radiative forcing by light-absorbing particles
The gap-filled snow cover product provides a useful tool to monitor and analyze inter-annual variability and long-term trends in snow cover in Iceland. The methodology applied here can be applied to other satellite sensors such as Sentinel 3 or the Visible Infrared Imaging Radiometer Suite (VIIRS) to extend the temporal range of data beyond the MODIS mission.
Code used in the project to process data is available at
AG conceived and designed the study, performed the analyses, and prepared the manuscript. SMG and ÓGBS contributed to the study design, interpretation of the results, and writing of the manuscript.
The authors declare that they have no conflict of interest.
We would like to thank Jessica D. Lundquist, University of Washington, for discussion and valuable feedback during the design of the study. Special thanks to Helgi Jóhannesson, project manager at Landsvirkjun, for providing constructive feedback during review of the manuscript. The Valle Scholarship and Scandinavian Exchange Program at the University of Washington is also thanked for financial support during the academic year 2017–2018 at the University of Washington.
This paper was edited by Ryan Teuling and reviewed by two anonymous referees.