Greenland ice sheet melt area from MODIS (2000–2014)

The Greenland ice sheet is an excellent observatory for global climate change. Meltwater from the 1.8 million km2 large ice sheet infl uences oceanic temperature and salinity, nutrient fl uxes and global sea level (IPCC 2013). Surface refl ectivity is a key driver of surface melt rates (Box et al. 2012). Mapping of diff erent ice-sheet surface types provides a clear indicator of where changes in ice-sheet surface refl ectivity are most prominent. Here, we present an updated version of a surface classifi cation algorithm that utilises NASA’s Moderateresolution Imaging Spectroradiometer (MODIS) sensor on the Terra satellite to systematically monitor ice-sheet surface melt (Fausto et al. 2007). Our aim is to determine the areal extent of three surface types over the 2000–2014 period: glacier ice, melting snow (including percolation areas) and dry snow (Cuff ey & Paterson 2010). Monthly 1 km2 resolution surface-type grids can be downloaded via the CryoClim internet portal (www.cryoclim.net). In this report, we briefl y describe the updated classifi cation algorithm, validation of surface types and inter-annual variability in surface types.


Greenland ice sheet melt area from MODIS (2000-2014)
Robert S. Fausto, Dirk van As, Jens A. Antoft, Jason E. Box, William Colgan and the PROMICE project team* Th e Greenland ice sheet is an excellent observatory for global climate change. Meltwater from the 1.8 million km 2 large ice sheet infl uences oceanic temperature and salinity, nutrient fl uxes and global sea level (IPCC 2013). Surface refl ectivity is a key driver of surface melt rates (Box et al. 2012). Mapping of diff erent ice-sheet surface types provides a clear indicator of where changes in ice-sheet surface refl ectivity are most prominent. Here, we present an updated version of a surface classifi cation algorithm that utilises NASA's Moderateresolution Imaging Spectroradiometer (MODIS) sensor on the Terra satellite to systematically monitor ice-sheet surface melt (Fausto et al. 2007). Our aim is to determine the areal extent of three surface types over the 2000-2014 period: glacier ice, melting snow (including percolation areas) and dry snow (Cuff ey & Paterson 2010). Monthly 1 km 2 resolution surface-type grids can be downloaded via the CryoClim internet portal (www.cryoclim.net). In this report, we briefl y describe the updated classifi cation algorithm, validation of surface types and inter-annual variability in surface types.

Classification algorithm
Th e algorithm uses normalised thresholds (Th ) from calibrated radiances (MOD021KM) between the near-infrared band 5 (1230-1250 nm) and the visible band 10 (483-493 nm). Th is updated classifi cation improves on Fausto et al. (2007) by implementing new surface type thresholds: Th dry snow ≤0.86, 0.86<Th melting snow <0.94 and Th glacier ice ≥0.94. Data gaps can be caused by cloud cover, which is identifi ed from the MOD35_L2. Th e ratio between band 1 (620-670 nm, red visual) and band 7 (2105-2155 nm, shortwave infrared) as well as the magnitude of band 1 are used to identify non-glacierised terrain (Fausto et al. 2007).
Daily classifi cation scenes are aggregated to yield a monthly Greenland surface type (GST) product. For the GST product, pixels are classifi ed according to their most frequently occurring (cloud-free) colour during a month. Th e maximum melt extent (GSTmax) is defi ned by a given pixel classifi ed as melting or bare ice. Th e pixel class must occur at least one day per month. Th e GSTmax product thus contains information about brief melt events. Conversely, a minimum melt extent (GSTmin) is calculated from pixels classifi ed as dry snow at least once per month. Pixels from areas not cov- ered by ice are excluded from Greenland melt area calculations employing the melting snow and glacier ice types in the GST, GSTmax, and GSTmin products (Fausto et al. 2007).

Validation
Th e PROMICE automatic weather station (AWS) network currently consists of eight regions each with two or three stations at a variety of low elevations on the Greenland ice sheet ( Fig. 1; Ahlstrøm et al. 2008;Citterio et al. 2015). Most stations are located in the ablation area, and are thus transitioning from snow-covered to bare ice surfaces through the melt season. Each station records a suite of meteorological and glaciological measurements, including ice temperature to 10 m depth and surface height changes due to accumulation or ablation (Citterio et al. 2015).
To validate the three surface types we classify, we use daily mean surface albedo (α), the ratio between incoming and outgoing solar radiation and surface temperature (T) data for the 2008-2014 period from the AWS. Based on α and T, we defi ne three classifi cations for AWS sites: dry snow (α > 0.7 or T < -7°C), melting snow (0.7 > α > 0.55), and glacier ice (α < 0.55; Cuff ey & Paterson 2010). Employing the T criterion acknowledges the infl uence of diurnal cycles at the beginning of the melt season. As a validation example, the in situ albedo and nearest-pixel classifi cation at KAN_L in 2009 are presented in Figs 2 and 3, including a visual comparison with the passive microwave melt area product (PMP) by Mote (2007).
Th e KAN_L station, located c. 10 km from the ice sheet margin at 680 m elevation, transitions through all three surface types during the melt season, from dry snow to melting snow to glacier ice. Relative to the 2000-2014 period, the 2009 surface melt was normal in west Greenland, with maximum melt areal extent in August. At KAN_L, the surface melted from May to August, with a daily mean albedo generally between 0.5 and 0.6 (Fig. 2). Th e algorithm accuracy for the KAN_L site may be assessed by an error matrix (Table 1). Th e diagonal represents successful classifi cations, the total number represents all classifi cations and the ratio between the sum of the diagonal and total is the accuracy. Th e algorithm yields 79% successful classifi cations at KAN_L, with an overall accuracy of 71%. Th e classifi cation algorithm performs best in the south and worst in the north, with accuracies of 87% (NUK_L) and 61% (KPC_U), respectively. Figure 2 illustrates changes in surface type during summer 2009, between 15 May and 14 September, according to the AWS data; all but two classifi cations were successful.

Results and discussion
Th e melt area from this algorithm and the PMP of Mote (2007), illustrated in Fig. 3 for 12 July 2012, are consistent with the reported melt area by Nghiem et al. (2012), who documented that 98.6% of the ice-sheet surface had melting. Th e GST also demonstrates close visual correspondence with PMP for the 2000-2014 MODIS period (Fig. 4). In Fig. 4 we have plotted the yearly maximum values of the GST, GSTmax and GSTmin products, as well as the PMP maximum extent of Greenland melt area. Th e increasing trends of GSTmax and GSTmin indicate a rising frequency of melt events and increasing summer melt, which is corroborated by the PMP which is comparable with GSTmax. Th e trend for PMP between 1979 and 2000 and 2000 and 2012 are almost identical making the PMP and GSTmax trends comparable. Overall, an expansion of the melt area to higher elevations is apparent (Fig. 4). Fausto et al. (2007) suggested that a sub-monthly GST product is non-optimal, because missing data due to cloud cover is the primary problem in determining the melt area. When trying to characterise all of Greenland, Hall et al. (2012) also found clear-sky, day-count problems, and also suggested that a sub-monthly product would have signifi cant uncertainty. However, uncertainties associated with the dif-  ferent surface types are assessed with the number of observations and standard deviation for each cloud-free pixel of the GST product (Fausto et al. 2007). Th e MODIS data have the advantages of high spatial resolution (1 km 2 ), pan-ice sheet coverage and quasi-daily temporal coverage, while the footprints of the in situ measurements are small. Th e AWS surface type classifi cations are therefore not an ideal ground truth for the surface classifi cation. Furthermore, whereas both GST and PMP melt area products can give daily results, the PMP surface microwave emittance originates not only from the surface but the top metre of the snow or fi rn, and is infl uenced by the water content in the snow during the previous days (Mote 2007). MODIS classifi cation is sensitive to cloud cover, but the spatial resolution of PMP is 625 times coarser than GST. During the melt period, exposed glacier ice in the ablation zone can have sub-zero temperatures. Such areas are included in the melting area, because the algorithm only makes use of the visual and near-infrared spectrum. Hence the melt area that we map might be more representative of the cumulative melt area during the melt period. However, if exposed, glacier ice in the ablation zone is covered by snow it will be mapped as non-melting areas. An August anomaly in monthly GST is evident during the 2010-2014 period. All August images indicate a noisy melting snow classifi cation in the northern ice sheet (not shown), which is most likely due to false classifi cation. However, anomalous, high concentrations of dust or reddish material have been observed on the ice sheet during recent late summers (Dumont et al. 2014). Increasing dust concentra-tions are problematic for the fi xed threshold algorithm we employ, because of enhanced absorption in near infrared wavelengths. Despite this possible biased source, an increasing trend in the melt area for the MODIS and PMP periods (Fig. 4) is consistent with increasing Greenland mass loss due to surface processes (Tedesco et al. 2013). Both independent, remotely sensed observations  and in situ observations (McGrath et al. 2013) show that the Greenland melt area is expanding to higher elevations.

Conclusions
Th e MODIS data can yield daily, automated classifi cations of the Greenland ice sheet into bare ice, melting and dry snow areas. Validation indicates that the surface classes are useful as ice-sheet climate indicators. Th e surface-type products are complementary to existing ice-surface temperature ) and melt-area (Mote 2007) products.