Data Processing
To facilitate zonal analysis, the NSIDC EASE-Grid product was remapped to a linear projection (Plate Carée) with a 1° grid cell resolution. The total number of 25 km input cells that corresponded to each one-degree output cell ranged between eight and twenty. An individual 1° output cell was classified as snow-covered if 30% or more of the input cells were snow-covered. A nearest-neighbor procedure was employed for the remapping. The weekly snow cover data files were each assigned a sequential week number (1-52) within the calendar year, where the week number corresponds to a fixed sequence of Julian days (1-7, 8-15, etc.).

Quantifying Snow Cover Timing and Duration
The temporal characteristics of snow cover in each annual period was quantified on a grid cell basis with respect to two key variables: the week number of the last observed snow-cover in spring (C_dis) and the week number of the first-observed snow cover in autumn (C_ons ). For this purpose the spring and autumn periods were defined as weeks in the January-July and August-December periods, respectively. The annual duration of the snow free period (C_dur , units in weeks) was computed as
C_{dur, n}=C_{ons, n}- C_{dis, n}- 1             (1)
where n is the year and 1971£n£1994.

Estimation of fPAR and APAR
The monthly composite NDVI data was corrected for calibration differences between satellite sensors following the approach of Myneni et al. (1997). The global correction was applied based on observed deviations in the average NDVI for the Sahara Desert region. In this analysis, fPAR was assumed to be equal to the NDVI. Monthly APAR (MJ m^-2) for May of 1983 and 1993 was computed as
APAR = fAPAR X PAR            (2)
Statistical Analysis
Descriptive statistics (mean, standard deviation) were employed to quantify the average timing of snow cover and the magnitude of interannual variability between 41° N and 75° N. Aggregations representing three spatial scales were examined: local (per pixel), continental (North America, Eurasia), and hemispheric (combined land areas except Greenland). Zonal statistics were also computed for 1° latitudinal bands. Least-squares linear regression was used to quantify any temporal trends over the 24-year study period. Trends were considered to be statistically significant when p < 5.0 (95% confidence level). Zonal averages were determined by computing the mean for each row of grid 1° grid cells. Individual 1° zonal trends were computed by linear regression of the 24-year time series of zonal average values.

Results

Zonal Trends for Snow Cover
Zonal trend analysis indicates that the duration of snow cover in northern hemisphere land areas between 45° N and 75° N increased at a zonal average rate of 8.8 (1.7) days per decade between 1971 and 1994 (Fig. 1). The increased duration is a consequence of observed advance in the timing of snow cover disappearance in spring (-6.5 [0.7] days per decade) and delay of snow cover onset in autumn (+4.5 [0.9] days per decade). These average rates of change represent the mean of individual 1° zonal trend values between 45° N and 75° N that are statistically significant.


Figure 1. Zonal average trends (1971-1994) in the timing of snow cover disappearance, snow cover onset, and the duration of the annual snow-free period. All trends shown are significant at the 95% confidence level.