Data description and spectral-difference method

Passive-microwave data from the multispectral radiometer SSM/I, projected on a polar stereographic grid and corresponding to the daily mean brightness temperatures recorded at the time of the U.S. Defense Meteorological Satellite Program (DMSP) satellite repeated overpasses (DMSP SSM/I brightness-temperature grids), were obtained from the U.S. National Snow and Ice Data Center (NSIDC). Two dual-frequency algorithms derived from the spectral-gradient method of Reference Künzi, Patil and RottKünzi and others (1982) and based on the difference between brightness temperatures measured at 19 and 37 GHz in horizontal polarization were developed and applied to the data. Although research groups in Finland (Reference Hallikainen and JolmaHallikainen and Jolma, 1992) and in Canada (Reference Goodison and WalkerGoodison and Walker, 1993) have demonstrated the superiority of the vertical polarization for retrieving snow-cover parameters, the horizontal polarization was chosen because it has shown greater sensitivity to surface conditions (Reference Chang, Foster and RangoChang and others, 1991) and it is theoretically better able to detect thickness variations of snow covers that have undergone several metamorphic transformations (Reference Hall, Foster and ChangHall and others, 1984), such as frequently occur in Churchill. The presence or absence of snow on the ground can be identified by averaging, on a 5 day basis (pentads), the spectral difference between brightness temperatures measured at 19H and 37HGHz. In homogeneous areas, such as the Canadian Great Plains and the Russian Steppes, it has successfully been demonstrated that this algorithm usually represents a valuable snow water-equivalent index when snow cover is dry (Reference Walker and GoodisonWalker and Goodison, 1993). In addition, snowmelt occurrences (presence of liquid water within the snowpack) can efficiently be detected through the difference between maximum and minimum spectral differences recorded over a pentad. Beside the identification of the melt and refreeze cycles, the use of pentads additionally allows a decrease of the volume of data when long time series are analyzed.

At the same time, we also carefully inspected the pentads in order to ensure that working with these would not lead to any significant information loss, especially in relation to brief but significant events affecting the seasonal evolution of snow cover. Because they correspond to brightness temperatures acquired twice daily (ascending and descending overpasses) over our study area, SSM/I data provided by NSIDC in the Equal-Area SSM/I Earth Grid (EASE-Grid) format permit examination of the seasonal behaviour of snow cover at a fine temporal resolution. This provides a means of improving our understanding of the interactions between the passive-microwave signal and snow-cover conditions. Specific information on the EASE-Grid data product can be found in Reference Armstrong and BrodzikArmstrong and Brodzik (1995).

Seasonal brightness temperatures at19H and 37HGHz and fluctuations in spectral differences for a single SSM/I gridcell: an example from winter 1997/98

The ability of SSM/I data to identify various stages during the snow-cover period (accumulation, date of appearance and disappearance of snow on the ground, snowmelt period) can be evaluated by comparing the spectral differences between 19H and 37HGHz computed for the winter 1997/98 pentads and daily brightness temperatures for a single SSM/I gridcell, together with meteorological data (air temperature and precipitation) recorded at Churchill airport.

During the first part of winter (STAGE-I), the brightness temperatures at 37HGHz in ascending and descending overpasses decrease jointly with the amount of snow on the ground because of the higher microwave diffusion by the numerous snow particles (Fig. 2a). As the brightness temperatures at 19HGHz remain quite stable (Fig. 2a), the mean spectral difference increases gradually and concurrently with snow accumulation (Fig. 3). For dry-snow conditions, emissivities decrease with increasing frequency (Reference MätzlerMätzler, 1987). As long as air temperatures remain below approximately 0°C, the snow cover is dry and the difference between the maximum and minimum spectral differences remains low.

Fig. 2. Spectral signatures of snow cover for one SSM/I EASE-Grid cell centred on the Churchill area (winter 1997/98) at 19 and 37 GHz, in horizontal polarization, ascending and descending overpasses, during (a) the snow-accumulation period, (b) an episodic snowmelt event and (c) the snowmelt period before the disappearance of snow cover. Dates are yy/mm/dd.

Fig. 3. Temporal fluctuations of the maximum, mean and minimum differences between the brightness temperatures at 19H and 37H GHz for one SSM/I gridcell centred on the Churchill area vs the meteorological data (air temperature, precipitation and snow on the ground) recorded at Churchill airport during winter 1997/98. Dates are yy/mm/dd.

On occasion, snowmelt events can occur in the middle of the winter season (STAGE-II). On 18 February 1998, for example, the diurnal air temperature rose above 0°C (Fig. 3), and the surface temperature recorded froma tower located in the open forest in the vicinity of Churchill reached 3.75°C during the descending overpass. The presence of wet snow, confirmed by field observations on that day, strongly affects the microwave signal so that the brightness temperature measured at 37HGHz increases by 5K (Fig. 2b). The microwave signature of wet snow is similar to that of a snow-free surface and opposite to that of a dry snow cover; emissivities increase with increasing frequency (Mätzler, 1987). Therefore, the difference between 19H and 37HGHz frequencies tends to be considerably smaller. On 19 February 1998, due to a return to more severe climatic conditions, the liquid water contained in the snow cover refroze, inducing a great increase in the spectral difference. Finally, even on a 5 day basis, one can observe a substantial decrease of the mean spectral difference which corresponds to a drop in snow depths and an important lag between the maximum and minimum spectral differences, testifying to the presence of the brief snowmelt and refreezing events (Fig. 3). However, the decrease in brightness temperatures between 9 and 25 February 1998 (Fig. 2b) can be attributed to the 10 cm increase of the depth-hoar layer thickness at the bottom of the snow cover, as observed and measured at several places in the field (Reference PivotPivot, 2000). Depth hoar consists of sizeable snow particles (ϕ >4mm) with a high microwave scattering power, which is typically responsible for a decrease in snow-cover emissivity and thus brightness temperature. Beside the depth-hoar effect, a decrease in the polarization difference by about 20 and 9.5 Kat 19 and 37 GHz frequencies, respectively, has been computed from the available EASE-Grid data between the two dates. According to Reference Surdyk and FilySurdyk and Fily (1995), this is related to significant stratigraphic changes within the snow cover. The metamorphism due to melting and refreezing simultaneously modifies the spectral signature of the snow cover by altering its internal structure (increase in the number of snow layers). A thorough examination of the EASE-Grid data throughout this period also allows the influence of atmospheric effects on passive microwaves to be observed (Fig.2b). Indeed, on 14 and 15 February, the presence of a disturbance and snowy clouds over the Churchill area possibly generated a rise of brightness temperatures at 37HGHz. Atmospheric effects are usually stronger at higher frequencies. This is an important problem when snow studies are conducted using the 85 GHz channel of SSM/I. However, Reference Wang, Chang and SharmaWang and others (1992) suggest that atmospheric absorption takes place even at 37 GHz, so that radiometric corrections remain essential for obtaining accurate estimates of snow water equivalent with satellite passive-microwave data.

At the beginning of April (STAGE-III), both 19H and 37HGHz diurnal brightness temperatures (Fig. 2c) increase strongly because of the presence of liquid water in the snow cover. At night, during refreezing, the formation of crusts leads to an important increase of the spectral difference that significantly contrasts with the diurnal observations. These daily fluctuations are associated with variations in air temperature (Fig. 3) and in the thawing–freezing metamorphism process. Around 20 April, the snowpack probably reaches its isothermal condition, and the snow-cover decay succeeds themelting–freezing cycles. This stage coincides with a sustained decrease of the spectral differences (Fig. 3). As soon as the differences become negative, the ground is free of snow.

The spatial approach: normal snow-season cycle and regional contrasts

Earlier, we showed that the mean spectral differences for a single gridcell centred on Churchill could be regarded as a proxy of snow-cover conditions at the local scale. Thus, to regionalize our observations, we applied the mean spectral-difference method to a 65665 window centred on the western section of the Hudson Bay Lowlands, extending roughly from the 60th Parallel to Lake Winnipeg (see Fig. 1). Here, the aim was to extract abnormal spatial features beyond the global character of the identified system (High Subarctic snow cover), in order to get a better idea of the spatial variability of snow cover. Finally, a non-standardized PCA with a Varimax rotation was performed, where the mean spectral differences were treated as the variables (784 pentads over the 11 year series of SSM/I observations) and the 4225 gridcells (65665 window) were used as the observations. The first four PCs derived from this procedure summarize 92.2% of the information contained in the initial dataset.

PC1 accounts for 44.3% of the explained variance. It describes the typical summer situation, mainly characterized by a very clear opposition between the ice-free Hudson Bay and the continent (Fig. 4a). Positive scores correspond to the summer pentads (higher loadings and low negative spectral differences) and are representative of the snow-free gridcells on land. Values fluctuate with the surface roughness induced by the various vegetation covers (boreal forest, subarctic transition zone and tundra). Negative scores are related to the winter pentads (high positive spectral differences) and are spatially attributed to the open-water gridcells on Hudson Bay. PC2 explains 36.1% of the variance and is symmetrical to PC1. In other words, it illustrates the typical winter situation where the land is snow-covered and Hudson Bay is frozen (Fig. 4b). The different ice-cover types are responsible for the disparities in factor scores on Hudson Bay. On land, the subarctic transition zone is clearly visible because of its specific spectral signature (the higher scores correspond to the lower emissivities at 37HGHz and higher spectral differences). Here, the amount of snow cover is higher than for the tundra. In contrast to the boreal forest, the snow-cover–microwave interactions in the subarctic zone are not masked as much by the forest canopy.

Fig. 4. Maps of PCA scores for the first four components of the spatial PCA. (a) PC1 and (b) PC2 represent typical summer and winter situations, respectively, whereas (c) PC3 and (d) PC4 correspond to snowmelt then ice decay on Hudson Bay.

PC3 and PC4 are also symmetrical. They summarize 7.2% and 4.6% of the explained variance, respectively. On PC3, the strongest loadings correspond to the snowmelt period. The higher scores are observed both in the arctic tundra and in the higher parts of the subarctic zone. They indicate snowmelt (Fig. 4c). Elsewhere, snow cover has disappeared. It is very interesting to note the particular position of Churchill, just between those two contrasting zones. Finally, the fourth component succeeds in time the third component. It shows ice decay on Hudson Bay in June, at which time the latitudes below 55° N are snow-free (Fig. 4d). The northern sea-ice drift is also clearly visible. The lower factor scores correspond to open-water areas, whereas the higher scores characterize the residual ice cover.

The temporal approach: interannual variability of snow cover at tree line

Here, the sea-ice gridcells were carefully removed and a standardized PCA with a Varimax rotation was performed on an 11613 window centred on land, in the vicinity of Churchill (see Fig. 1). The first two components summarize 95.8% of the explained variance. PC1 explains 54.3% of the variance and illustrates the seasonal behaviour of snow cover. This component shows: (1) in a general way, pentads characterized by snow on the ground (positive scores) against pentads free of snow (negative scores), and (2) in detail, pentads corresponding to the maximum snow amount on the ground (higher positive scores) against pentads corresponding to the date of snow disappearance (higher negative scores).

PC 2, however, is probably the most interesting component (41.5% of the explained variance). It shows the interannual variability of snow-cover conditions and allows a rough estimate to be made of the snow-cover duration for each year within the period of observations. Observations corresponding to spectral differences above 0K (pentads with snow on the ground) contribute primarily to the formation of the second axis (PC2). Thus, annual snow-cover duration can be estimated through this component, since the reversal of the scores sign closely agrees with the dates of onset and disappearance of snow (Fig. 5a). In a further investigation, to be validated, this information will be compared with the meteorological data (e.g. daily snow-depth records) and the estimations issued from the weekly snow-extent maps for Northern Hemisphere land surfaces performed using visibleband U.S. National Oceanic and Atmospheric Administration (NOAA) Advanced Very High Resolution Radiometer (AVHRR) imagery. PC2 also depicts the interannual variability of snow cover, in a similar fashion to the mean spectral difference derived for the single SSM/I gridcell centred on Churchill (Fig. 5b). Indeed, the strength of the scores almost describes snow-depth fluctuations (Fig. 5c). Finally, two groups are opposed (Fig. 5a): (1) cold and/or snowy winters of 1988/89, 1991/92, 1993/94, 1994/95 and 1996/97 (positive scores) with mean spectral differences often greater than 30 K; and (2) warm and/or dry winters of 1989/90, 1990/91, 1992/93, 1995/96, 1997/98, and 1998/99 (negative scores) with mean spectral differences not exceeding 30 K.

Fig. 5. Temporal PCA outputs related to the second component and some elements of interpretation. (a) PC2 scores; (b) variations of the mean spectral difference recorded for one SSM/I gridcell around Churchill; (c) snow depth measured at Churchill airport. Dates are yy/mm/dd.

Negative scores are also noticeable in the middle of winter. They are associated with sudden changes affecting the spectral-difference values, sometimes due to an episodic snowmelt event (e.g.1992/93) and sometimes due to a significant accumulation of snow on the ground following a winter storm (e.g. 1995/96). Monthly air-temperature and snowfall anomalies with respect to a 30 year normal from the weather station of Churchill airport are presented in Figure 6a and b. Both graphs corroborate the interpretation of the second component by reproducing rather accurately the climatic conditions prevailing during winters 1988/89 to 1998/99. Except for the important amount of snowfall recorded in October, snow accumulation in 1988/89 was near normal. In March and April, however, the unusually low air temperatures largely contributed to the persistence of snow on the ground. Winters 1991/92 and 1996/97 were the snowiest of the 1990s. During those winters, a thick snow cover formed in the fall due to abundant snowfalls at the beginning of the snow-accumulation period as well as favourable temperatures. Interpreting winter 1994/95 with currently available information is more complicated. Most of the snow accumulated in November, and a snow cover 450 cm thick was maintained over much of winter and a good part of spring. During the snow season, however, monthly air temperatures were 2.5–5.5°C higher than normal, except in February. In winter 1993/94, very low air temperatures appear to account for the spectral-differences rise above 30 K, and consequently for the position of this winter on the second component.

Fig. 6. Climatic conditions, October 1988–June 1999, at Churchill airport. (a) and (b) correspond to the monthly air-temperature and precipitation anomalies with respect to a 1961–90 normal. Dates are yy/mm/dd.

Throughout winters 1992/93,1995/96,1997/98 and 1998/99, snowfall did not exceed 160 cm in total. These dry conditions explain the shallow snow depths and the lower scores for these winters compared to those described earlier. The lack of snowfall observed during winter 1997/98 is undoubtedly caused by the occurrence of the temperature anomaly encountered in the Equatorial seas off South America, commonly known as El Niño. Indeed, the Southern Oscillation Index (SOI) fluctuations modify the atmospheric circulation (at the 5 km altitude) by splitting the jet stream over North America. Thus, in Canada, a severe snow deficit and an abnormal rise in temperature are typically ascertained during El Niño years (Reference Brown and GoodisonBrown and Goodison, 1996; Brown, 1998). This is obviously what happened during winters 1997/98 and 1998/99. Winter 1991/92 also witnessed an El Niño event, but did not produce the expected climate configuration. That particular year, higher dust concentrations in the stratosphere following the eruption of Mount Pinatubo, Philippines (June1991), obscured the climatic effects of El Niño. According to Portman and Gutzler (1996), volcanic eruptions also play a significant role in interannual climate variability at global and regional scales.

Finally, a problem still remains. While the factor scores and the microwave signal are very similar in winters 1991/92 and 1993/94, snow-cover conditions were very different. We previously stated that in 1991/92 (Pinatubo year) the PCA outputs were explained mostly by abnormal snowfall, whereas in 1993/94 the main controlling factor was the abnormally cold temperatures. We know that the amount of snow vs the spectral-difference relationship is not unique. As the surface temperature also influences the brightness temperature, cold (warm) winters cannot be discriminated from snowy (dry) winters. Indeed, in the passive-microwave wavelengths, the brightness temperature measured for a given surface corresponds to the product of the emissivity of this surface and its physical temperature. Consequently, even if temperature only intervenes to the first power, compared to the thermal infrared wavelengths where it intervenes to the fourth power, its influence on the brightness temperature cannot be completely occulted. This problem will have to be taken into account in future improvements of passive-microwave snow algorithms.