SEARCH

SEARCH BY CITATION

Keywords:

  • snow cover;
  • AVHRR;
  • Northern Hemisphere

Abstract

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Summary and Discussions
  7. Appendix A
  8. Appendix B
  9. Acknowledgments
  10. References

[1] The Global Climate Observing System has identified the need for systematic global daily snow cover data sets over land. Current in situ snow cover data sets have limited spatial coverage while satellite-based snow cover records have either limited historical extent or limited temporal and spatial resolution because of cloud cover or specific sensor availability. NOAA Advanced Very High Resolution Radiometer (AVHRR) data offers nearly complete daily global coverage of the Northern Hemisphere, extending back to the early 1980s with successors slated to continue into the next decade. In this paper, we apply a new algorithm, Snowcover, to estimate daily snow cover, including periods of cloudy conditions, from AVHRR Polar Pathfinder (APP) data over Northern Hemisphere land surfaces. This new snow cover product is compared to snow cover estimates derived from long-term in situ snow depth measurements over Canada and the northern Eurasia. The APP snow cover maps showed an 80% agreement rate or better at 95% of the in situ sites. This performance was comparable to the agreement of MODIS 0.05 degree snow cover products over the same sites; although the MODIS product was only retrieved ∼20% of the time corresponding to clear sky conditions in contrast to over 95% of the time with the APP snow product. The almost continuously temporal and spatial coverage for the past 23 years from 1982 to 2004 over Northern Hemisphere makes the new daily snow cover product especially suitable for analysis of large-scale patterns of spring snowmelt in association with variability of circumpolar climate and ecological parameters.

1. Introduction

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Summary and Discussions
  7. Appendix A
  8. Appendix B
  9. Acknowledgments
  10. References

[2] Snow cover (SC) is one of the most dynamic and ephemeral terrestrial features on the Earth's surface. It is believed to be a sensitive factor of climate change exerting important influence on the regional and global climate [Massom, 1991] in addition to its important economic and societal impacts. The Global Climate Observing System (GCOS) has identified an immediate need for generating continuous global historical daily SC products at relatively high (≤1 km) spatial resolution that would benefit the estimation of planetary albedo and improvement of regional and global climate model performance [Mason, 2006]. Available in situ snow depth data sets have limited spatial coverage and the snow courses are often specific to their surrounding environment [Dyer and Mote, 2006]. In contrast, prior to 2000, satellite-based SC maps produced by National Oceanic and Atmospheric Administration (NOAA) offer at best 25 km spatial resolution [Armstrong and Brodzik, 2002] with daily temporal resolution only available after 1997 [Ramsay, 1998] and a higher spatial resolution at approximately 4 km available since 2004 (detailed in National Snow and Ice Data Center website). These satellite-based products tend to exhibit larger uncertainties during the ablation period [Wang et al., 2005]. Wang et al. [2005] reported that the current NOAA weekly SC data set consistently overestimated SC extent during the spring melt period, with delays of up to 4 weeks in melt onset.

[3] The large-scale spatial distribution and temporal variability of SC over Northern Hemisphere high-latitude lands has attracted considerable attention in recent years. This is because SC conditions significantly affect land cover characteristics and land-atmosphere energy exchange through changes in planetary albedo [United Nations Environment Programme, 2007]. Therefore accurate mapping SC is critical for understanding snow-albedo feedbacks as well as improving climate models [e.g., Robinson, 2003; Robinson and Kukla, 1985], especially under the present global warming. Previous studies concluded that the maximum snow-albedo feedback occurs during spring [e.g., Hall and Qu, 2006; Déry and Brown, 2007]. As such, long-term changes in SC during spring play a significant role in the surface radiation budget [Déry and Brown, 2007; Qu and Hall, 2005; Robinson, 2003]. On the basis of the NOAA weekly SC data set over Northern Hemisphere continents, Robinson [2003] identified an abrupt decrease in SC duration in the spring and early summer around the early 1980s. There are only a couple of days per decade in change of snowmelt date in the Northern Hemisphere [Dye, 2002]. It is not likely that the NOAA weekly SC data set is sufficient for identifying such small changes in spring snowmelt date. On the basis of model simulations with greenhouse gas induced climate change, Raisanen [2008] has suggested that variations in snow amount at a geographical point depend on both warming during winter and spring and increased snowfall as well. Having consistent information regarding the melt date will reduce uncertainty regarding the impact of these two potentially counteracting processes on SC dynamics.

[4] Recent studies have shown that the Arctic regions are experiencing rapid changes in ecosystem associated with the global warming over recent decades [Comiso et al., 2008]. Wang and Key [2005] have suggested that the Arctic has warmed and become cloudier in spring and summer and the surface broadband albedo has decreased significantly in autumn on the basis of the NOAA AVHRR Polar Pathfinder data set over the period of 1982–1999 [Wang and Key, 2005]. GCOS has specifically identified that SC maps should be derived from archived NOAA AVHRR data extending back to the early 1980s once this data has been reprocessed to meet requirements as a fundamental satellite-based climate data record. In anticipation, we have already developed and tested an algorithm, Snowcover [Fernandes and Zhao, 2008], for historical SC mapping from full resolution (1 km) AVHRR imagery in Canada. However, the existing APP data set should also allow us to derive daily SC maps, albeit at 5 km gridded resolution. Significantly, this APP data set has been the basis for global albedo trend studies [Wang and Key, 2003] that would immediately benefit with an associated daily SC map to assess global snow-albedo feedbacks [Qu and Hall, 2005]. Furthermore, the 5 km resolution will likely still be useful for assessing global and regional climate models.

[5] The objectives of this paper are to describe the application of the Snowcover algorithm to the APP 5 km data set over the Northern Hemisphere land surfaces with an emphasis on the approach used to generate consistent continuous SC estimates over multiple sensors and to evaluate this new APP SC product using SC estimates derived from long-term in situ snow depth measurements over northern Eurasia and Canada. Finally, spatial and temporal distributions in spring snowmelt date for the circumpolar zone are presented as an example of new information, relevant to climate studies, contained in this data set.

2. Data and Methods

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Summary and Discussions
  7. Appendix A
  8. Appendix B
  9. Acknowledgments
  10. References

[6] The APP daily 5 km Equal Area Scalable Extent (EASE)-grid composites are a collection of products for circumpolar regions [Fowler et al., 2000], derived from twice-daily calibrated and gridded satellite measurements. In this study, the Northern Hemisphere daytime observations are used. Data include top-of-atmosphere (TOA) reflectance of AVHRR Channels 1 (0.58–0.68μm) (CHN1) and 2 (0.725–1.10μm), normalized difference vegetation index (NDVI), derived clear sky surface broadband albedo based on J. R. Key (The Cloud and Surface Parameter Retrieval (CASPR) Version 4.0.3, National Environmental Satellite, Data, and Information Service, 1999), skin temperature (ST), solar zenith angle (SZA), satellite view angle (SVA), sun-satellite relative azimuth angle, and cloud mask. This data set extends poleward from 48.4°N, available from 24 July 1981 through 30 June 2005. Only valid daytime observations corresponding to SZA < 85° and SVA < 55° were used to ensure sufficient signal at the detector as well as to reduce the contribution from surfaces outside the nominal pixel.

[7] Quality controlled daily snow depth measurements, between 1982 and 2004, over 67 standardized in situ sites in Canada (Figure 1), falling in the APP spatial domain, were extracted from a database provided by the Meteorological Service of Canada (MSC) [Braaten, 1998]. Another 260 station sites in the northern Eurasia (Figure 1), in the APP domain, were also extracted directly from the Historical Soviet Daily Snow Depth Version 2 (HSDSD) data set [Armstrong, 2001]. Canadian snow depth data correspond to the average of up to twenty daily ruler measurements of snow depth, taken within a site area of ∼50 m diameter by MSC observers and archived as daily element in the Canadian National Climate Data Archive. Sites typically correspond to flat open fields in the vicinity of airports. Measurements were performed with a view of avoiding snow drifts. Average snow depths less than 1 cm but with snow visibly present were labeled as trace depth. A detailed quality control procedure was applied by MSC to flag suspicious values and filled as missing data [Braaten, 1998]. This data set represents a subsample of the entire MSC in situ snow depth measurement network stratified to provide a relatively uniform geographic coverage over Canada. The HSDSD data set is based on observations from 284 World Meteorological Organization (WMO) stations throughout Russia and the former Soviet Union, spanning from 1881 to 1995. Information on this data set [Armstrong, 2001] is detailed in the National Snow and Ice Data Center website (http://nsidc.org/cgi-bin/get_metadata.pl?id=g01092). In this study, SC was assumed present if snow depth is greater than or equal to 2 cm [Braaten, 1998]. Snow-free conditions were assumed present if snow depth was 0 cm. As a result, no in situ label was assigned for intermediate values.

image

Figure 1. Locations of the 67 in situ sites in Canada and the 260 in situ sites in the northern Eurasia.

Download figure to PowerPoint

[8] To address the concern with both the difference in spatial support of the in situ snow depth measurements and the 5 km APP grid-based SC maps and the assumption that complete SC is present with a depth equal to and greater than 2 cm, we also evaluated the well characterized daily Moderate Resolution Imaging Spectroradiometer (MODIS) SC product at the Canadian in situ sites. Previous work over Canada [Simic et al., 2004] found the MODIS 500 m MOD10A1 product has generally an agreement in excess of 90% with SC estimates derived from the same in situ data set. To determine the influence of spatial scaling effects, we compared the daily 0.05 degree spatial resolution SC product (MOD10C1) derived from MODIS, distributed through National Snow and Ice Data Center (NSIDC), over the in situ sites. The 0.05 degree grid size is of similar spatial support as the 5 km APP product, hence we expect that major differences in agreements between MOD10C1 and in situ SC estimates versus APP and in situ estimates should reflect differences in either the quality of the Snowcover and MODIS algorithms or the input satellite data rather than scale differences with in situ data. The MOD10C1 products were reprojected from geographic to 5 km resolution EASE-grid coordinates. Only dates where the MOD10C1 confidence mask was 100 (highest level), corresponding to clear sky conditions, were retained for further analysis of this product. A land cover database of North America at 1 km resolution [Latifovic et al., 2004], provided by the Global Land Cover 2000 website, (available at http://www-gem.jrc.it/glc2000/ProductGLC2000.htm), was used to identify land cover types for the in situ snow sites.

[9] The Snowcover algorithm [Fernandes and Zhao, 2008], originally developed for SC mapping with 1 km resolution AVHRR imagery over Canada, is used to estimate SC over each APP 5 km land grid cell irrespective of cloud cover. The algorithm consists of three procedures applied on a grid cell basis to annual daily time series of APP measurements: temporal filtering and interpolation, normalization of CHN1 to standard acquisition geometry, and SC detection.

[10] The temporal filtering and interpolation stage estimates continuous daily “clear sky” time series of annual CHN1, ST and NDVI measurement time series (CHN1s, STs and NDVIs respectively) on a per-grid cell basis. A series of moving window temporal rank filters, adaptive to local cloud conditions and grid cell wise calibration of snow and snow free end-members, are applied to identify clear sky time series. The rank filters assume that clouds result in substantially higher (lower) TOA CHN1 (ST and NDVI) estimates compared to measurements under corresponding clear sky conditions [Viovy et al., 1992]. This filter differs from conventional rank filters [e.g., Jedlovec et al., 2008] in that it is specifically designed to preserve locally monotonic noise free time series. Cloud shadows are assumed to substantially reduce TOA estimates for CHN1, ST and NDVI in comparison to corresponding clear sky conditions. Then, given an a priori estimate of the probability of clouds and shadows and assuming the clear sky time series is locally monotonic, the rank of a sample within a local time window can be associated with a probability that it is free of cloud or shadow contamination (Appendix A).

[11] The a priori cloud probability was estimated by a 30 day moving average frequency of APP cloud flags. A constant cloud shadow probability of 10% was applied on the basis of average statistics derived from AVHRR data processed over Canada that include shadow masks [Latifovic et al., 2005]. Successive passes of the moving window rank filter, with window sizes decreasing from 40 days to 10 days, were applied. The 40 day upper window size was used to limit computational requirements while the lower size corresponded to the smallest window allowing a 0.95 clear sky probability for a priori cloud probability over 0.5.

[12] At each pass of the rank filter, measurements satisfying both the local monotonicity conditions and a rank corresponding to a clear sky probability greater than 0.95 were identified. In addition, gaps in the clear sky time series falling in the local filter window containing each clear sky grid cell were filled with the clear sky value. By applying successively smaller filter windows, gaps between clear sky grid cells were filled using the smallest valid interpolation interval. This form of nearest neighbor gap filling was required to preserve the associated acquisition geometry for the normalization stage.

[13] The Roujean bidirectional reflectance distribution function (BRDF) model [Roujean et al., 1992] was used to normalize the continuous CHN1 measurements to a nominal nadir view at the solar zenith angle corresponding to the observed overpass time at the local summer solstice. This reference geometry was selected since it minimized the range of solar zenith angles to be normalized at each grid cell. As with Romanov et al. [2003] we assumed the volume and geometric kernel coefficients were invariant for each calendar year. However, in contrast, grid cell specific coefficients were used to account for land cover variability. At each grid cell, the volume scattering coefficient of the Roujean model was assigned as a function of NDVI on the basis of a relationship fitted using the data from Latifovic et al. [2005]. A relationship between median August NDVI and the geometric kernel coefficient was calibrated using paired CHN1s and in situ snow depth measurements during deep snowpack conditions over Canada between 1985 and 2000. At each site and for each year, the coefficient minimizing the absolute magnitude of the slope of the CHN1s versus solar zenith angle curve was selected.

[14] Automated SC detection algorithms often employ a threshold derived from measurements over either spatial neighborhoods [Metsämäki et al., 2005] or snow free land cover condition [Hall et al., 2002; Khlopenkov et al., 2006; Romanov et al., 2003; Wang et al., 2005]. These approaches may not be applicable for APP data given the potential for substantial subgrid cell mixing in both snowpack and topographic and vegetation properties. Our examination of CHN1s time series over in situ snow depth sites during spring melt periods indicated a strong discontinuity in their derivative versus both time and STs at the date of complete spring snowmelt (to) (e.g., Figure 2). This behavior was confirmed using a simple theoretical model relating surface conditions during melt to top of canopy CHN1 albedo (Appendix B). Our SC detection approach used the CHN1s value corresponding to this discontinuity of the annual CHN1s versus STs curve during the spring melt period, at each grid cell, as the annual CHN1s snow free threshold.

image

Figure 2. (a) Nonsmoothened (dots) and smoothened (lines) time series of Band 1 BRDF- corrected reflectance (b01brdf/b01brdfs), NDVI (ndvi/ndvis), and surface temperature (red). (b) Observed snow depth (0.01 cm), the fractional APP snow cover (snowlabel), and smoothened albedo (albds) for the open canopy site in Canada.

Download figure to PowerPoint

[15] Given the noisy nature of the CHN1 time series, we relied on nonsmooth optimization theory [Borwein and Lewis, 2006] to identify to. Grid cell specific STs thresholds were applied to identify the start (Tmin) and end (Tmax) of a broad interval bracketing the spring melt period (limited to between 1 January and 31 August of each year). The to was initially estimated as the point on the CHN1s versus STs curve with minimum projection on the normal to the chord subtending this interval. The interval was then reduced by shrinking the STs thresholds to the halfway point between the current interval boundary and to and the estimation of SC updated until only one other point remained. Dates when CHN1s <= CHN1s(to) or with STs > Tmax were mapped as snow free. Dates when CHN1s > CHN1s(to) or when STs < Tmin were mapped as having snow present. Additionally, a linear mixture model was fitted between the two remaining dates in the fitting interval and applied to the CHN1s annual time series at the grid cell. The fractional value served as an estimate of the precision of the SC detection for other days during the year (Figure 2b).

[16] Comparisons between satellite and in situ SC estimates were conducted over the year of 2002 in Canada, and the year of 1993 in the northern Eurasia. These periods correspond to the years, other than data used for calibration of the algorithm, with the greatest temporal coverage of MODIS and APP SC data, respectively. Comparison of in situ and satellite-based SC estimates is complicated by differences in both temporal and spatial sampling. Satellite estimates of a given grid cell correspond to the local overpass time (for MOD10C1) or the typical overpass time (for the interpolated APP product) while in situ estimates could be measured anytime during the day. Simic et al. [2004] documented up to 5% difference between 500 m MODIS SC maps and the same in situ sites due to spatial scale difference. Even larger differences are likely when 5 km resolution satellite products are assessed. To minimize differences in maps due to spatial and temporal sampling effects we only compared satellite and in situ estimates over periods with at least k continuous days with the same in situ SC status. Comparisons were performed at k = 3 days and k = 10 days with the former aimed at reducing temporal scaling effects but still capturing ephemeral SC conditions (and also including some spatial scaling uncertainties) and the latter aimed at also reducing spatial scaling effects at the cost of missing ephemeral SC conditions and the latter stage of melt as well. To test seasonal dependence of the Snowcover algorithm, the estimations are also made for full year period and for the partial period from day of year (DOY) 1 to 244 (1 January to 31 August).

[17] We also evaluated a simpler method for producing continuous SC time series based on filling gaps between only clear sky snow cover estimates using the most recent preceeding clear sky label [Parajka and Blöschl, 2008]. The MOD10C1 SC product was used for this purpose since, in theory, it should produce more accurate estimates of clear sky labels in comparison to our input temporally interpolated APP measurements. This “postlabeling” gap filling method was tested by evaluating agreement with in situ SC over Canada during 2002.

3. Results

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Summary and Discussions
  7. Appendix A
  8. Appendix B
  9. Acknowledgments
  10. References

[18] We examined all stages of the Snowcover algorithm at the in situ sites specifically to ensure it performed robustly as a function of land cover and SC duration. For brevity we discuss a typical site corresponding to an open canopy mixed forest region near Bagotville, Quebec, Canada (48.3°N, 71.0°W). Figure 2 presents the original and BRDF normalized filtered time series of APP measurements for this site together with derived SC and in situ snow depth. The temporal filtering captures the spring snowmelt transition well but also results in a “blocky” time series because of the nearest neighbor gap filling. CHN1s varies substantially during winter with snow depth prior to snowmelt. The sensitivity of snow covered end-members to land cover agrees with similar findings by Metsämäki et al. [2002] in that grid cell specific calibration for vegetation attenuation is required for large area SC mapping. As Figure 2 indicates, the CHN1s trend exhibits a rapid decrease prior to complete snowmelt allowing the Snowcover algorithm to accurately identify the melt date and associated snow free threshold. At the same time, the rapid decrease results in some uncertainty in identifying snow covered areas as indicated by the range of derived fractional SC. The threshold, calibrated to snowmelt conditions, together with the applied temporal filtering, also results in uncertainties in estimation of snowpack status during ephemeral onset conditions.

[19] For Canadian sites, the level of agreement of the daily SC product extracted from the APP 5 km parameters using the Snowcover is shown in Figure 3, sorted from best to worst in situ site over the period of 2002. In comparison to snow courses during 3-day continuous snow status, the 80 percentile of the total sites exceeds an accuracy of 90% for the full year period and the 95 percentile of the sites for the partial period (January–August). This improvement over the latter period assessment indicates that the most frequent disagreements occurred mainly during the snow onset season. This could be due to (1) the temporal filtering and zeroth-order interpolation applied to the satellite measurements, (2) large differences in visible reflectance between fresh snow albedo during onset and the level prior to spring melt used to define the SC end-member in the Snow cover algorithm, and (3) the increase in visible reflectance during fall senescence over vegetated areas resulting in a bias in the snow-free end-member. For 10-day continuous snow status (not shown), agreement between the APP SC and in situ SC increased to 90 for 85 percentile of the sites for the full year and for 95 percentile of sites for the partial period. This supports the suggestion that the majority of the errors in the APP product occur during the onset period in fall and early winter. We will show later that some of these disagreements may be attributed to the aforementioned scale mismatch between in situ SC and satellite SC estimates when water bodies fall within the satellite footprint.

image

Figure 3. Snow mapping agreement rates for Canadian sites using Snow cover based on the 5 km APP data set. (a) An estimation for a partial period from 1 January to 31 August (244 days) and (b) an estimation for the full period of the year.

Download figure to PowerPoint

[20] Figure 4 presents the level of agreement between the MODIS daily snow with in situ measurements at Canadian sites over 2002. For the partial period (prior to onset), the 85 percentile agreement rates for the MOD10C1 products are only marginally (<5%) better than the APP results shown in Figure 3. For full year with 3-day continuous snow conditions, the MOD10C1 product shows ∼5% higher agreement rates, for the 80 percentile or higher ranked sites, in comparison to the APP product. This likely reflects the additionally uncertainty in the temporal filtering and interpolation used in the Snow cover algorithm. The benefit for this slight decrease in level of agreement in the APP SC product is its nearly continuous SC mapping versus under 20% retrieval rates for the MOD10C1 product, corresponding to clear sky condition, as indicated in Figure 5. In contrast, the APP SC product only shows missing data during high-latitude winters when even skin temperature measurements are not available in the APP data set. This benefit is significant if the SC product is to be used as a climate data set. Figure 6 illustrates the resulting snow maps for a given day, displayed on the APP 5 km EASE-grid net, demonstrating that the APP SC product has much higher spatial coverage ability than MODIS snow product at the same 5 km resolution over lands while both products agree well in the mapped areas. Figure 7 presents the agreement rates of both APP and MODIS snow products with the in situ observations during the snowmelt period (defined as a period from 50% of the maximum snow depth to zero depth over the partial period from DOY 1 to 244). The APP product retrieves an estimated ∼70% of the time versus ∼20% for the MODIS product. Furthermore, it is found that there is no significant difference in agreement of the APP product versus in situ sites across major land cover types (not shown for brevity).

image

Figure 4. Same as Figure 3 except for MODIS.

Download figure to PowerPoint

image

Figure 5. Comparison of available number of days retrieved per site over Canada between APP (dark) and MODIS (light) snow product in 2002.

Download figure to PowerPoint

image

Figure 6. Snow coverage maps of the (a) APP and (b) MODIS on 30 April 2002. Note that white indicates snow cover, green indicates snow free land, black indicates missing or no data, and blue indicates water bodies or areas outside the APP domain.

Download figure to PowerPoint

image

Figure 7. Same as Figure 3 but a comparison of snow mapping agreement during melting period between the APP and the MODIS products with available snow and no snow sample size in percentage of the 244 days superimposed. Note that the legend applies to both Figures 7a and 7b.

Download figure to PowerPoint

[21] The postlabeling gap filling approach resulted in a increase from ∼20% to ∼57% in retrieval rates for the MODIS product when limiting the interpolation to a seven day period as in the work of Parajka and Blöschl [2008]. At the same time, the agreement rate with in situ measurements dropped to below 50% for the 85 percentile of sites.

[22] For completeness, Figure 8 presents the agreements of the APP 5 km SC product with those extracted from the HSDSD data set [Armstrong, 2001] over the northern Eurasian continent, which shows very similar results to those over Canada (Figure 3) although there is a little bit lower agreement (80%) for the full year of 1993 but similar agreement rates to the Canadian sites (90%) from DOY 1 to 244 at 95% of the in situ sites.

image

Figure 8. Same as Figure 3 except over the northern Eurasian sites.

Download figure to PowerPoint

[23] A scaling problem occurs when comparing the SC product derived from the APP 5 km resolution data set with station-based in situ snow depth observations, especially over a grid cell with multiple land cover types, including water bodies. For example, Figure 9 depicts the northern most in situ site on the coast of Baffin Island, Canada. Figure 9, corresponding to a Landsat 5 Thematic Mapper (TM) image on 7 August 2000 clearly indicates multiple land covers with water bodies within the nominal 5 km APP grid cell although the land surface is snow free. Sea ice generally covers this area until July. In such a case the Snowcover algorithm, when applied to the APP data, determines the snow end-member to correspond to the date when both sea ice and SC are absent as indicated in Figure 10 resulting in a later estimated snow free date following spring melt in comparison to in situ measurements (Figure 10b). In contrast, application of Snowcover to 1 km AVHRR data over the same period results in good agreement between in situ and satellite SC (Figure 10c). This further supports the suggestion that our current assessment is likely conservative over areas where the in situ site is not located in the same land cover as the rest of the APP nominal grid cell footprint.

image

Figure 9. The Landsat 5 TM image superimposed with the APP 5 km EASE grid points (white points connected by white lines) in the vicinity of Site #1 (the star). The nominal grid cell of the APP EASE-grid point corresponding to Site #1 in fact represents an area including water bodies, which results in a scaling problem when comparing with the averaged snow measurements at Site #1. The small box in the center of the image represents approximately an area of nominal 1 km pixel. Note that blue colors are sea/lake ice, dark colors are water bodies, and red colors are lands.

Download figure to PowerPoint

image

Figure 10. (a and b) Same as in Figure 2 but for Site #1. (c) One km snow product using Snow cover based on the northern most site.

Download figure to PowerPoint

[24] In our study snowmelt date (Smtd) is defined as the date from which each grid cell has no snow at least for three continuous days during the spring melting season from DOY 90 to 244 (1 April to 31 August). Figure 11 presents the climatological mean Smtd during 1982–2004 for each grid cell at 5 km resolution over the APP domain. To better observe the spatial distribution of the mean Smtd, different colors were assigned to the Smtd by 20-day interval from DOY 90 to 250. Smtd generally increases with latitude consistent with the seasonal march of solar radiation during spring and early summer in the Northern Hemisphere. High-altitude areas such as the western cordilleras and regions mapped as having permanent ice or snow cover represent clear outliers in this pattern. The zonally distributed Smtd belts also show earlier melt dates at the margins of continents consistent with warm Pacific and Atlantic currents.

image

Figure 11. The mean Snowmelt date (Smtd) for 1982–2004 over Northern Hemisphere APP domain, shown by 20-day interval. The white regions are assigned to permanent snow.

Download figure to PowerPoint

[25] Figure 12 presents the time series of Smtd averaged over the two continents, the northern Eurasia and North America, and the Northern Hemisphere over the APP domain as well. The three time series do not show statistically significant trends over the period of 1982–2004. There is large interannual variability that is not related to specific satellite sensors indicating our data set does not have large-scale sensor artifacts. The peaks and valleys in the time series of Smtd match to some extent between the two continents for the period of pre-1998, but diverge since 1998, which may be associated with interdecadal variability in the atmospheric circulations.

image

Figure 12. The time series of Smtd averaged over northern Eurasia (EAmtd), North America (NAmtd), and Northern Hemisphere (Gmtd). Dashed lines are the mean Smtd over the period of 1982–2004 as shown at the left of these time series with digital values.

Download figure to PowerPoint

4. Summary and Discussions

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Summary and Discussions
  7. Appendix A
  8. Appendix B
  9. Acknowledgments
  10. References

[26] In this paper, we described how the Snowcover algorithm was applied to the new APP daily 5 km Northern Hemisphere Polar Pathfinder data set, which extends poleward from 48.4 degrees north, available from 24 July 1981 through 30 June 2005, to generate a new continuous daily circumpolar SC data set. This study demonstrated that the Snow cover is able to map continuous daily SC using APP data with almost continuously temporal and spatial coverage ability. The APP SC maps showed an 80% agreement rate or better at 95% of the in situ sites on the basis of the estimation of 3–10-day continuous snowpacks over the main snow variation seasons. Agreement rates during the melt period increased to over 90% at the 95 percentile level suggesting that this product is suitable for tracking rates of melt and the covariation of snowmelt with other climate variables.

[27] Some of the observed differences, especially for the 5% of sites with large disagreement rates, may be due to differences in the spatial support of the APP 5 km grid cells and the in situ measurements. Scale differences were especially problematic for sites with mixed land cover that include water bodies. Other studies [Hall and Riggs, 2007] have relied on Landsat imagery to produce spatially extensive SC maps but this approach can only validate retrievals during clear sky conditions and may also suffer from similar biases related to vegetation and terrain masking of SC that the APP satellite data experiences since they both use visible and near-infrared wavelength measurements. In the future, nested spatial sampling approaches using in situ sensor webs with automated SC or depth measurements (see the Cold Land Process Experiment website, available at http://www.ist-world.org/ProjectDetails.aspx?ProjectId=3c6db4a581d54dec9ac6527876f2881a) may offer one means of validating GCOS compliant 1 km SC products.

[28] The Snowcover algorithm provided almost continuous retrievals at the cost of less than 5% additional disagreements in comparison with MODIS product. The fact that the state of the art MODIS algorithm does not appreciably result in higher agreement with in situ estimates suggests that there is sufficient temporal sampling in the APP data set to approximate seasonal melting trends in clear sky reflectance and temperatures to the level required for SC mapping. There are still a number of improvements that could be made to the Snowcover algorithm including the use of advanced temporal interpolators and the calibration of separate melt and onset period end-members. More importantly, the algorithm is limited by errors and uncertainties in the APP data including geometric and radiometric accuracy and spatial resolution.

[29] It is important to note that the adaptive temporal filtering applied to the APP data set serves both to interpolate gaps and to produce a temporal profile from which a per grid cell threshold can be derived. In contrast to studies using 500 m resolution MODIS snow product [Parajka and Blöschl, 2008], the application of postlabeling gap filling approaches to the 5 km MODIS snow cover maps resulted in significant decreases in agreement with in situ snow cover was observed. This suggests that such a gap filling strategy is not suitable for the 5 km APP data.

[30] SC conditions significantly affect land cover characteristics and land-atmosphere energy exchange through changes in planetary albedo. Previous studies concluded that the maximum snow-albedo feedback occurs during spring [Hall and Qu, 2006; Déry and Brown, 2007; Robinson and Kukla, 1985], as such, long-term declines in SC during spring play a significant role in the surface radiation budget [e.g., Qu and Hall, 2005; Déry and Brown, 2007]. On the basis of this new APP snow data set, we produced spring snowmelt dates at each grid cell for each year and observed an expected but previously not explicitly mapped latitudinal patterns in spring melt dates together with areas where orographic or coastal climate effects are expected to delay snowmelt. The large interannual variability in snowmelt date together with the absence of obvious trends would be counterintuitive if one hypothesized that spring temperatures alone drive snowmelt. Rather our data provides evidence in support the more complex hypothesis that winter precipitation and spring temperatures interact, in perhaps opposing directions, in their impacts on snowmelt trends in the Northern Hemisphere [Ye et al., 1998; Brown, 2000; Robinson, 2003; Dyer and Mote, 2006; Raisanen, 2008].

[31] This study demonstrates that there is sufficient temporal sampling in the APP data set to estimate continuous daily SC during the melt season as long as a snow mapping algorithm, like Snow cover, carefully performed temporal interpolation on reflectance rather than clear sky SC data. It provides an opportunity to consistently map Northern Hemisphere daily and 5 km SC over 1982–2004, with validation using in situ measurements, that approaches GCOS requirements other than the spatial resolution threshold and hopes to improve spatial and temporal resolution of current climate snow cover records over 1982–2004. Finally, it shows that our approach to SC mapping allows for derivation of snowmelt date that is useful both for assessing snow-albedo feedbacks and as a climate indicator in its own right. We hope that the scientific community will make use of this new information as additional evidence to test hypothesis regarding SC, climate interactions in the Northern Hemisphere.

Appendix A

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Summary and Discussions
  7. Appendix A
  8. Appendix B
  9. Acknowledgments
  10. References

[32] This appendix provides a definition of the rank filtering used for identifying clear sky measurements meeting a prespecified probability of commission error. A detailed discussion of the motivation for the filter and its verification is given by Fernandes et al. [2008].

[33] Noise sources in satellite measurements include both impulse noise, corresponding to large noise processes assumed independent of the signal level, as well as noise from “continuous disturbances” corresponding to smaller magnitude additive or multiplicative noise processes. The former could include contamination from clouds, haze, shadows, sun glint, cosmic ray effects or detector dropouts while the latter could include clear sky atmospheric effects, variability due to acquisition geometry, the noise equivalent delta temperature of the detectors and digital quantization. The signal, free of impulse noise, can then be considered as a realization of random process X(t) sampled in the presence of continuous disturbances. The measured signal, including impulse noise, is a realization of a related random process equation image(t).

[34] Let x(t) correspond to a realization of X(t) and equation image(t) correspond to noisy measurements of x(t) that falls above or below x(t) with a priori probabilities

  • equation image

and

  • equation image

respectively. For example p+ and p are the a priori cloud and shadow probabilities respectively.

[35] If X(t) is stationary overΔ(t) the probability that the rank K sample in this interval, assuming no ties in ranks and that impulse noise is statistically independent from sample to sample, is noise free is given by the product of two cumulative binomial distributions corresponding to the probability that the first NK samples are free of positive noise and the last KJ of the remaining K samples are free of negative noise

  • equation image

B(n, N, p) is defined as the probability of at most n successes (samples with impulse noise) in N trials where the probability of an individual success is p according to the cumulative binomial distribution.

[36] This model is restrictive since X(t) is not always stationary over any selected interval. Instead, if x(t) is locally monotonic decreasing the positive impulse noise will only increase (decrease) preceeding (subsequent) measurements of a noise free sample within Δ(t). Then, if we restrict the rank K to correspond to the Kth last sample in Δ(t), equation (A3) still holds. A similar argument applies for the Kth earliest sample in Δ(t) when x(t) is locally monotonic increasing.

[37] We were unable to develop necessary and sufficient conditions for the identification of locally monotonic intervals Δ(t) of the underlying x(t) given noisy measurements equation image(t). However, some necessary conditions include (1) with P(t, K) constant then for all of the J samples satisfying equation A1 within any given Δ(t), at least P(t, K)J must themselves match the local monotonicity condition; (2) for each of the J samples satisfying equation A1 within any given Δ(t), at least [1 − p(t)]K(J) of the preceding (following) samples be greater (less) than the noise free samples if x(t) is locally monotonic decreasing (increasing); and (3) for each of the J samples satisfying equation A1 within any given Δ(t), at least [1 − p+(t)](NK(J)) preceding (following) samples containing negative noise must be less (greater) than the noise free samples if x(t) is locally monotonic increasing (decreasing).

Appendix B

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Summary and Discussions
  7. Appendix A
  8. Appendix B
  9. Acknowledgments
  10. References

[38] A simplified model relating surface conditions to directional upper hemispherical canopy albedo, αc[UPWARDS ARROW], was developed to assess the qualitative nature of the temporal trend in CHN1 during the spring melt period. The top of canopy directional upper hemispherical albedo, assuming lambertian surfaces, can be approximated as [Lewis and Disney, 2007]

  • equation image

where

αL

is the foliage single scattering albedo;

τ[DOWNWARDS ARROW]

is the downwelling uncollided canopy direct transmittance;

equation image[UPWARDS ARROW]

is the upwelling uncollided canopy diffuse transmittance;

αb and equation imageb

are the directional hemispherical and bihemispherical background albedo respectively;

αb and equation imageb

are the directional hemispherical and bihemispherical background albedo respectively;

equation imagec[DOWNWARDS ARROW]

are the directional hemispherical and bihemispherical background albedo respectively;

equation imagec[DOWNWARDS ARROW], equation imagec[UPWARDS ARROW]

are the bihemispherical canopy scattering albedo in downwelling and upwelling directions respectively;

p, q

are the canopy mean recollision and upward escape probabilities.

[39] Following Mõttus and Stenberg [2008] we gave

  • equation image
  • equation image

We assume for now that the time derivative of canopy plant area index is ∼0 during the melt period and immedeatly following the date of complete melt. Then, the derivative of canopy directional hemispherical albedo in this period is given by

  • equation image

Assuming nonzero background albedo at the date of complete melt, to, the derivative of the canopy albedo at the melt date is proportional to the derivative of the background albedo at the melt date

  • equation image

The background albedo is approximated as a linear mixture of fraction of snow covered (fs) and snow free areas with respective directional upper hemispherical albedos αs and αg

  • equation image

The fraction of snow-covered background is related to mean snowpack depth Ds after Romanov et al. [2003]

  • equation image

where a will in general be a scale and location specific constant.

[40] The fraction of snow covered background is then related to snowpack water equliavlent, θs, assuming a constant snowpack density gives ρs

  • equation image

The directional upperemispherical albedo of the snow covered areas is modeled as a linear mixture of the fraction of exposed litter (fl) and exposed open snow with respective directional upper hemispherical albedos αl and αp

  • equation image

For a given surface area density of litter, Al, we can define a litter area index l

  • equation image

The fraction of exposed litter on the snowpack is related to the litter area index using a modified Beers law with extinction coefficient Kl

  • equation image

The derivative of the backround albedo with time is then

  • equation image

where the derivative of the snowpack albedo with time is given by

  • equation image

Assuming a degree day based daily snowmelt rate gives

  • equation image

where Tm is a base temperature and C is the melt rate coefficient.

[41] Then, the time derivation of the background (and proportionally top of canopy) directional hemispherical albedo as the date of complete melt is approached in a causal and anticausal directions are given by

  • equation image

[42] In general exposed time derivative of leaf area index (or plant area index in general) will be nonzero prior to and following complete melt due to evergreen vegetation and woody matter. Furthermore, the derivative of exposed leaf area index will likely increase during snowmelt as vegetation (and even microtopographic relief) emerges from the snowpack. However, this increasing trend will only serve to increase the rate of decrease of canopy directional hemispherical visible albedo during melt due to increased shadowing for a given illumination condition. It is possible that, over very large regions (possibly 5 × 5 km APP grid cells), microclimate effects will also result in some continued increase in exposed leaf area index immediately following complete melt due to vegetation growth. Rapidly emerging vegetation will tend to decrease the slope of the post-melt period visible albedo trend slightly. However, the derivative of albedo will still be a discontinuous function of temperature at the complete melt date and the albedo versus temperature curve will still be convex in the vicinity of the discontinuity.

Acknowledgments

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Summary and Discussions
  7. Appendix A
  8. Appendix B
  9. Acknowledgments
  10. References

[43] The authors would like to thank Rasim Latifovic and Yi Luo for their important comments that have improved this manuscript significantly. The authors wish to thank the anonymous reviewers for their insightful critiques that led to significant changes and improvements of the manuscript. This research is supported by Climate Change Program, Natural Resources Canada and the Canadian International Polar Year Programme. Authors would like to thank NSIDC, MSC, GLC, and GeoBase for providing their data products.

References

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Summary and Discussions
  7. Appendix A
  8. Appendix B
  9. Acknowledgments
  10. References