Melt Detection Process The process shown in Figure S1 was found necessary after discovering areas of anomalously long melt duration in maps classified using a simple, single 2 dB threshold below the winter mean. The primary cause for this false detection is thought to be associated with snow accumulation acting to attenuate the temporal backscatter signature of a pixel. Setting larger thresholds for areas of high snow accumulation follows other threshold-based methods employed over Greenland [e.g., Wang et al., 2007]. Rather than manually identify these areas, we developed a decision tree to automatically detect and correct areas of false melt detection by using one of three larger threshold "cases". Although this process was necessary, it is important to note that the vast majority of melting was classified using a single, 2dB threshold (Case A in Figure S1). On average over 2000-2009, 94% of melt used Case A, 2% used Case B, 2% used Case C, and 2% used Case D. The characteristic temporal backscatter pattern of the Case A situations can be seen in Figure 2 of the paper (i.e., large and abrupt reductions in backscatter during melting, large and abrupt increases during freeze-up, and generally stable winter backscatter). The typical signature of snow accumulation is a very slow temporal reduction in backscatter, making it easily distinguishable from backscatter reductions in response to liquid water. Four different threshold configurations ("cases") were run as shown in Figure S1, and the decision tree process determined which case was used to calculate melt in each pixel for each year. The threshold case used for each pixel was determined independently each year of the study. The logic in the decision tree was determined through careful manual examination of many instances of false melt and modifying the decision tree logic until these instances were remediated (i.e., actual melt signatures detected as melt but snow accumulation signatures dp not contribute to the pixel's melt duration). The winter mean backscatter (July to September) was used to characterize "pre-melt" conditions. "Post-melt" conditions were calculated as the mean of days 122 through 152 (approximately the month of May) of each year. These particular pre- and post-melt values were calculated for each pixel of each year. In some areas, heavy snow accumulation during austral summer or directly thereafter results in false melt detection even in places that experience true surface melting. Thus, the output of a single 2 dB threshold (Case A) was compared to a single 4.5 dB threshold (Case B). If Case B produced 10x more melt than Case A, the larger threshold of Case B was used for that pixel in that melt year. One instance that necessitated Case A and Case B was as follows: (1) clear, discernible melting and freeze-up, (2) followed by very high backscatter values after melting (a signature of ice layer/lens formation), (3) followed by decreasing backscatter over a period of days to weeks to at least 2 dB below the winter mean again. Thus, this last step (3) results in false melt detection after the true melt events (1) earlier in the season. Elsewhere, in instances of autumn/winter/spring snow accumulation, backscatter reductions can sometimes exceed 2 dB from the winter mean, causing a pixel to be incorrectly classified as melt. In these cases, if true melt existed, freeze-up could be misidentified because of reduced backscatter after the melting events. Thus, if a 1 dB difference existed between pre- and post-melt conditions (signifying snow accumulation), a dual threshold approach was used given that the backscatter was generally decreasing (>1 dB) through the austral summer. For these cases we used two annual thresholds: (1) a smaller value to identify any true melt events at the beginning of austral summer, and (2) a larger value to prevent false melt associated with snow accumulation later in the austral summer. The first threshold is effective from day 201 of year one to day 90 of year two (mid July to March) and the second threshold is effective from days 91 to 200 of the second year (April through mid-July). Two cases using this method are labeled as "Case C" and "Case D" in Figure S1. If the Case C thresholds (defined as 2 and 4.5 dB) produced 2.1x more melt (a value empirically determined by examining backscatter time series) than the Case D thresholds (defined as 4.5 and 6 dB), Case D was chosen assuming that some of the Case C melt detection was incorrectly identified. As a final step, some clearly false melting (owing to its location in the high elevation continental interior) not corrected by this method was manually masked out each year. Some of these areas are subject to somewhat higher backscatter noise owing to surface conditions particularly sensitive to satellite observations at multiple azimuth angles, potentially resulting in SIR imaging artifacts (noise) [e.g., Long, 2010]. Sensitivity Analysis A sensitivity analysis was performed to evaluate how varying threshold values affected correlation with air temperature data at the AWS sites shown in Figure 1 of the manuscript. Figure S4 shows how the strength of the linear regression between QuikSCAT melting days and AWS positive temperature days vary as a function of the threshold chosen. This plot reveals that comparable regression strengths exist for thresholds ranging between 1.5 and ~6 dB. Simply, this fact is attributed to the large and sudden decrease in dB at melt onset. Figure S4 also reveals that the MDD-PDD regression is less sensitive to threshold used relative to the MD-PD regression. The reason for this is also simple: As the MDD algorithm subtracts a melting day backscatter value from the winter mean of that pixel, the result is a daily MDD value very close to the winter mean (i.e., ~0 dB) if the day is erroneously detected as melting. As such, an erroneous MDD value contributes minimally to the annual MDD melt intensity term. Figure S5 shows how the parameters of the linear regression (slope and intercept) change with varying threshold. This plot reveals that our primary 2 dB threshold produces a slope near 1 and an intercept near the origin. Thresholds greater than 2 dB produce shallower slopes indicating underestimation of melt, whereas thresholds below 2 dB produce larger intercepts that indicate overestimation of melt. References Long, D. G. (2010), Standard BYU QuikSCAT and Seawinds Land/Ice Image Products, BYU Center for Remote Sensing, Microwave Earth Remote Sensing Laboratory. Wang, L., M. Sharp, B. Rivard, and K. Steffen (2007), Melt season duration and ice layer formation on the Greenland ice sheet, 2000-2004, J. Geophys. Res., 112(F4), doi:10.1029/2007JF000760.