Topographical effects on the timing of growing season in alpine grasslands

Context : It is important to understand the responses of alpine vegetation to recent anthropogenic climate change. The mountainous landscapes with high climatic heterogeneity are good locations to investigate the effects of microclimatic variation on alpine ecosystems. Objectives : a) To what degree do topographical factors (aspect and elevation) affect the timing of growing season in alpine grasslands? b) Are these topographical effects different on alpine and non-alpine grasslands? Methods : We extracted five annual growth phenology indices (Start, End, Length, Peak and Peak-NDVI) in alpine and non-alpine grasslands in the Clutha river catchment, New Zealand with a near-daily NDVI (Normalized Difference Vegetation Index) dataset through 16 years (2001-2016). The shifting rates of these phenology indices were quantified with two topographical factors (aspect and elevation). Results : The start of season was delayed by 7.5, 5.1 and 3.7 days per 100 m higher of elevation in three grassland types (Alpine, Tall Tussock and Low Producing) respectively, and the end of season was advanced by 1.7, 1.3 days and delayed by 0.3 days per 10-degree-south on slopes individually. The longer season length was observed at lower elevation and on north-facing (sunny) slopes. The later season peak occurred at higher elevation and on north-facing slopes. The lower peak NDVI was detected at the higher elevation. Conclusions : In the studied grasslands, aspect and elevation were correlated to different phenological indices, and they affect phenology independently. The topographical effects are more pronounced in alpine ecosystems at the elevation above 1,300 m than in non-alpine ecosystems at lower elevation. Abstract Context : It is important to understand the responses of alpine vegetation to recent 4 anthropogenic climate change. The mountainous landscapes with high climatic heterogeneity 5 are good locations to investigate the effects of microclimatic variation on alpine ecosystems. 6 Objectives : a) To what degree do topographical factors (aspect and elevation) affect the 7 timing of growing season in alpine grasslands? b) Are these topographical effects different on 8 alpine and non-alpine grasslands? 9 Methods : We extracted five annual growth phenology indices (Start, End, Length, Peak and 10 Peak-NDVI) in alpine and non-alpine grasslands in the Clutha river catchment, New Zealand 11 with a near-daily NDVI (Normalized Difference Vegetation Index) dataset through 16 years 12 (2001-2016). The shifting rates of these phenology indices were quantified with two 13 topographical factors (aspect and elevation). 14 Results : The start of season was delayed by 7.5, 5.1 and 3.7 days per 100 m higher of 15 elevation in three grassland types (Alpine, Tall Tussock and Low Producing) respectively, 16 and the end of season was advanced by 1.7, 1.3 days and delayed by 0.3 days per 10-degree- 17 south on slopes individually. The longer season length was observed at lower elevation and 18 on north-facing (sunny) slopes. The later season peak occurred at higher elevation and on 19 north-facing slopes. The lower peak NDVI was detected at the higher elevation. 20 Conclusions : In the studied grasslands, aspect and elevation were correlated to different 21 phenological indices, and they affect phenology independently. The topographical effects are 22 more pronounced in alpine ecosystems at the elevation above 1,300 m than in non-alpine 23 ecosystems at lower elevation. 24

Researches showed that topographical features are substantial gradients which affect the 37 growing season of alpine species. For example, the regional difference in slope and aspect 38 affects the timing and duration of snow cover (Tennant et  2) Are there any differences of the effects of aspect / elevation on the growth phenology 71 between in alpine and in non-alpine grassland ecosystems?  (Wardle 2008). Low Producing grassland can be found at mean elevation of 640m, and more 87 than 80% of this type lives between 338m and 929m elevation. This type is recognised as the 88 non-alpine region in this study (Fig.1a). 90 Five annual growing phenological indices (Table 1)  year (P-NDVI) and the base level (the mean of the two lowest NDVIs before and after the POS 103 of a year) is defined as the "seasonal amplitude". We used 50% of the seasonal amplitude as    The SAR-lag model:

Growth phenology extraction
The SAR-error model: The SAR-sac/sarar model: In above formulas, Y represents the vector of response variable (one of the five phenological 137 indices); X is the matrix of predictor variables (Southness, Elevation, and both); β is the 138 coefficient of a linear regression model; W is the spatial weights matrix; ρ is the autoregressive 139 coefficient for the spatial lagged dependent variable Y; λ is the autoregressive coefficient for 140 the spatial weighted error term; ε is a spatially dependent error term; u is a non-spatial error term. In SAR-lag model, λ is assumed to be 0, while in SAR-error model ρ is assumed to be 0.

142
In SAR-sac/sarar model, ρ and λ are both non-zeros. In formula (3) the Ws for spatial lagged 143 item (ρWY) and spatial error item (λWε) could be different, however, we used the same W for 144 both items in this study. 145 We used an inverse-distance-weight algorithm ("dnearneigh" function in the package "spdep" 146 on R platform (Bivand and Wong 2018)) to generate a specific spatial weights matrix (W) for 147 each SAR model. The main idea is that: In an image, one pixel's spatial proximity to its 148 surrounding pixels can be described by certain weights which are negatively associated with 149 the distances of pixels (Fortin 2005). The hypothesis is that spatial autocorrelation affects  Table 4.  The trend of later EOS on more northeast-facing slopes was also observed in Tall  grassland showed a reverse trend that the north-facing habitats had 11.6 days shorter of LOS.

185
The peak of growing season (POS) exhibited an identical pattern as EOS in the three grassland  (3.7 days later per 100m upwards) but higher when above this elevation (8.2 days later per 203 100m upwards).

204
The season end was less sensitive to elevation changes, especially in Alpine grassland (Fig.4).

205
There was a weak trend of later EOS as elevation climbs in the Alpine grassland under 1,300 206 m. However, when above the elevation of 1,300m, EOS in Alpine grassland stayed the same 207 at about 312-315 mDOY (Table 3). For Tall Tussock grassland, the EOS happened 4.9 days 208 later per 100m upwards when below 1,300m elevation, and EOS also kept at 316-320 mDOY 209 when above 1,300m. In Low Producing grassland the EOS was delayed by 5.8 days per 100m 210 higher of elevation when below 1,300 m, but it fixed at 320 mDOY when above this elevation.

211
The length of growing season showed a strong negative correlation with elevation in the 212 grassland above 1,300 m (Fig.4) grassland above 1,400m showed a shorter LOS of 11.8 days per 100m upwards.

219
The peak of season was positively correlated with elevation (Fig.4). In the Alpine grassland    265 We displayed the distributions of five annual phenological on an aspect-elevation coordinate 266 to illustrate whether interactive topographical effects existed in our study (Fig.6). It showed 267 that SOS was mainly correlated to elevation only in grasslands (Fig.6a), except the SOS in the

277
The POS in Alpine grassland was correlated to both aspect and elevation (Fig.6c). POS Generally, P-NDVI in three grassland types decreased as elevation rises (Fig.6e). P-NDVI was 290 strongly correlated with elevation in both alpine and non-alpine ecosystems. Above 1,000m 291 elevation, P-NDVI declined at the same rate along elevation in all aspects. While in the Tall

292
Tussock grassland under 1,000m and in the Low Producing grassland under 800m, P-NDVI 293 was higher on south-facing slopes.

294
The delta AICs of SAR models showed that the additive and interactive formulas are more 295 effective than any single variable formulas when explaining the changes of the five annual 296 phenological indices ( Overall, the delta AICs of models (Table 4) Table 1 The five annual growth phenology indices investigated in this study (see Methods

519
and Fig.2 for details of how these indices are calculated). End of season EOS The day of the season on which NDVI drops to 50% of the season peak NDVI after the peak day of season. 3 Length of season LOS The number of days between start and end of the season. 4 Peak Day of season POS The day of the season on which NDVI reaches its peak of this growing season. 5 Peak NDVI P-NDVI The fitted NDVI maximum of a growing season 521