Carbon response of tundra ecosystems to advancing greenup and snowmelt in Alaska

The ongoing disproportionate increases in temperature and precipitation over the Arctic region may greatly alter the latitudinal gradients in greenup and snowmelt timings as well as associated carbon dynamics of tundra ecosystems. Here we use remotely-sensed and ground-based datasets and model results embedding snowmelt timing in phenology at seven tundra flux tower sites in Alaska during 2001–2018, showing that the carbon response to early greenup or delayed snowmelt varies greatly depending upon local climatic limits. Increases in net ecosystem productivity (NEP) due to early greenup were amplified at the higher latitudes where temperature and water strongly colimit vegetation growth, while NEP decreases due to delayed snowmelt were alleviated by a relief of water stress. Given the high likelihood of more frequent delayed snowmelt at higher latitudes, this study highlights the importance of understanding the role of snowmelt timing in vegetation growth and terrestrial carbon cycles across warming Arctic ecosystems.


Supplementary Note: Spatial representativeness assessment and MODIS greenup and snowmelt timing evaluation
We assessed the spatial representativeness of the landscape within the tower footprint (diameter of 200-300 m; Fig. S7) for the Moderate resolution Imaging Spectroradiometer (MODIS) gridded spatial scales (500×500 m 2  range value indicates the average patch size of the landscape heterogeneity 1 ; therefore, we assumed that the landscape around the tower is likely to be representative of the MODIS in the pixel window (e.g., 1×1 pixel or 3×3 pixels) when the range value at the pixel window is smaller than (or close to) the tower footprint size (approximately 250 m) 3 . Table S3 show that the range values of the flux tower sites are smaller than (or close to) the tower footprint size in the 1×1 pixel windows over the seasons (i.e., snowmelt, early GS, and peak GS) but higher than the tower footprint size in the 3×3 pixel windows at most sites. This indicates that the landscapes around the towers are likely to be representative for the MODIS 1×1 pixel windows but less (or even not) representative for the 3×3 pixel windows. PhenoCam sites, except imcrkridge0 and NEON-D19-HEAL, are spatially representative of the MODIS 1×1 pixel windows during the early GS. Four NCDC stations are likely representative in the 1×1 pixel windows during the snowmelt season, but the USC00503585 station may not be spatially representative in either pixel window given that the range values are far above 250 m.

Fig. S8 and
For snowmelt timing validation, we collected ground data, including snow depth measurements at the KOPRI site (available upon request from S. Park) and five NCDC stations (https://www.ncdc.noaa.gov/cdo-web/, Table S2), and incoming and outgoing short wavelength radiation data at the US-EML site (https://doi.org/10.17190/AMF/1418678). Snowmelt timing was estimated as the day when the amplitude of the ground data (i.e., snow depth or the ratio of incoming to outgoing short-wavelength radiation) dropped below 10% of the wintertime mean value.
We also calculated the MODIS snowmelt timing when a logistic fit to the MODIS snow cover (MOD10A1.V006, quality flags of good and best) passed 0.1 each year and evaluated it against ground-based estimates (Fig. S9). Our results show that the MODIS snowmelt timings agree well with those from ground data where the site was spatially representative during the snowmelt season at the MODIS scale (Fig. S9A, p < 0.001). We also found that the agreement is higher at a more representative scale (i.e., higher 2 at the 1×1 pixel window, Fig. S9B). Poor agreement was also found at the USC00503585 station (which was spatially not representative in either pixel window, Fig. S9C).
PhenoCam dataset v2.0 2 (https://phenocam.sr.unh.edu/webcam/, Table S2) was used to validate the MODIS greenup timing (MCD12Q2.V006, quality flag of best). In the dataset, we used the 10% amplitude threshold date during "greenness rising", which was However, it should be noted that the number of data points is still very limited (N ≤ 11); therefore, further analysis with longer data periods including more sites should be performed.
Our results imply that it is critical to consider the implication of spatial representativeness on remote sensing-based timing estimates 5,6 . Given that the flux tower sites are mostly representative in the MODIS 1×1 pixel windows during both the snowmelt season and early GS (Table S3), MODIS snowmelt and greenup timings were applied for further analysis in this study. Due to the low availability of leaf area index (LAI) measurements at the tower sites, MODIS LAI data (MCD15A3H.V006) evaluation in terms of spatial representativeness was not performed in this study. However, given that the landscapes around the towers are mostly representative of the MODIS 1×1 pixel windows during the early and peak GS, there would be less uncertainty in the MODIS LAI data resulting from landscape heterogeneity; therefore, we used the MODIS LAI values to calibrate and evaluate ED2.               19,20 Table S5. Summary of ED2 calibration and validation. The dataset, period, and statistical measures (r 2 and root-mean-square-error, RMSE) for each variable used for calibration at the US-Atq site (c) and evaluation at the seven study sites