**4. Discussion**

#### *4.1. Does Seasonal Snow Seasonality Metrics Affect Land Surface Phenology Metrics?*

Our findings demonstrated that land surface phenology at high elevations responds to snow cover seasonality. However, this response relationship is complicated because different snow season metrics affect different phenological metrics (Figure 8, Table 1). The magnitude asymmetry in the significant correlation between the snow seasonality and land surface phenology metrics suggests that a more snow-prone non-growing season (earlier first snow, later snowmelt, longer snow season and more snow cover days) may benefit a more flourishing vegetation growing season the following year (earlier start and later end of the growing season, longer growing season). The area of significant positive correlation between FSD and SOS is more than two times the area of significant negative correlation. In contrast, the area significantly positively correlated with LOS is less than half of the area significantly negatively correlated. These magnitude asymmetries may indicate the direction of the effect of FSD on LSP [33]. That is, FSD is positively correlated with SOS and negatively correlated with EOS and LOS. This suggests that earlier snowfall in the autumn triggers an earlier growing season the following spring and delays the end of the vegetation growing season the following autumn, which also extends the vegetation growing season.

Qiao and Wang [38] found no or negative correlation between FSD and SOS when exploring winter snowpack and spring grassland vegetation phenology in Inner Mongolia, which is not consistent with our findings. This may be attributed to elevation. Significant negative correlations could be observed only in central Inner Mongolia, where the elevation is below 3000 m. However, at higher elevations in the southwest, the correlations are not significant. In QLMA, there are relatively few areas below 3000 m in elevation, resulting in the negative correlation that is not widely observed by us. The findings of Wang et al. [35] in the Tibetan Plateau (TP) are close to ours. They extensively observed a positive correlation between FSD and SOS, especially in the central TP and northwest of the QLMA. The negative correlation between FSD and EOS is consistent with the findings of Qi et al. [34] in QLMA; moreover, we observed a larger area (Figure 8). This finding is equally relevant in other areas with continuous seasonal snowpack [66]. However, the impact of late season transient snowfall events may be limited, which could explain the small impact of FSD in some areas [58,67].

We considered that a later snowmelt may result in a delayed end of the growing season and a longer growing season length. Longer snow seasons and more snow days lead to earlier growing seasons and also prolong growing season length. Our study confirmed the findings of Wang [35] and Qi [34] in the same study area, but differences in elevation and hydrothermal conditions cause this finding to change in other areas. In fact, the effect of different snow seasonality metrics can still be observed even within a region, which may also be due to the elevation and climatic conditions of the different regions [36]. Moreover, different vegetation indices, LSP estimation methods and thresholds for estimating snow seasonality have an impact on the conclusions. However, these effects are usually reflected in intensity and significance. The apparent magnitude asymmetry of correlations within the study area points to the main direction of snow seasonality effects, in which we are able to corroborate each other. Even in different areas, similar effects of snow seasonality can still be observed if elevations and climates are similar, such as at higher elevations in Inner Mongolia and Nepal [38,66].

#### *4.2. Why Do the Effects of Snow Seasonality Metrics Vary with Elevation?*

Our study confirmed the high-elevation dependence between snow cover and phenology (Figure 9). The effect of earlier FSD on SOS shifted from delaying to facilitating with increasing elevation. The effect on LOS changes from shortening to lengthening, with the turning point occurring roughly at 3500 m. The correlation between FSD and EOS changes from a nonsignificant positive correlation to almost no correlation. If an earlier and longer growing season is considered to be better, then an earlier FSD is detrimental to vegetation growth at lower elevations and beneficial at higher elevations. This effect of first snow on vegetation phenology with elevation is not unique to QLMA. Paudel and Andersen [66] found no correlation between FSD and EOS at low elevations, but a strong positive correlation at very high elevations. Obviously, the elevation of our study area is far from 'very high', above 5000 m, so we could only observe an insignificant correlation between FSD and EOS. The findings in the Qinghai–Tibetan Plateau and QLMA are more comparable. Qi [34] divided QLMA into four elevation intervals, and the correlation between FSD and EOS is consistent with our findings on the trend of weakening with elevation. Our more detailed division makes the results more obvious. Wang et al. [35] found a positive correlation between FSD and SOS on the TP that gradually increased between 2500–5000 m and then weakened, which is similar to our findings. We further found a change in this correlation not only in intensity but also in direction, that is, a significant negative correlation between FSD and SOS below 3500 m. The study area of Wang et al. [35] is much larger than ours, and the seasonality of snow varies greatly at lower elevations, which may lead to their inability to accurately count correlations below 3500 m. However, that earlier first snowfall at higher elevations favors earlier vegetation phenology is what we all agree on.

We did not observe an effect of LSD on LSP metrics with elevation. Regardless of elevation, later snowmelt is beneficial for earlier vegetation growing season initiation and a longer growing season. Snow melt directly provides the necessary water for vegetation to sprout, and the spring snowpack maintains soil temperatures. No matter what the elevation, accumulated temperature and water are necessary for vegetation to sprout. Paudel and Andersen [66] observed the same conclusion as ours in the low elevation arid zone, and Wang et al. [35] also found a negative correlation between LSD and SOS in the TP, but it was not significant at low elevation. In contrast, Qi et al. [34] found a significant positive correlation between LSD and SOS in the high-elevation interval of QLMA, which may be related to the selection and treatment of the vegetation indices. Compared to estimating SOS based on NDVI, the NDPI we used is shown to better eliminate the effect of pre-season snowpack and avoid misclassification of snow and vegetation pixels [40–42]. In addition, our reconstructed NDPI with a temporal resolution of 4 d also helps to obtain a more accurate SOS.

We also found a similar effect of SSL and SCD on LSP metrics. Vegetation below 3300 m is barely affected by them, while vegetation above 3500 m SOS is significantly negatively correlated with them, and LOS is significantly positively correlated with them. This is consistent with the findings of many other studies, at least at the same elevation [34,35,58]. This may be due to different climatic conditions at different elevations, which could drive differences in correlation [68]. Although the melting of snow will always provide moisture

to the soil, different temperature conditions may lead to different effects on the presence of snow. More snow is needed at higher elevations to cover the soil than at relatively warm lower elevations because the snow acts as an insulator [36,69]. Soils are protected from the harsh climatic conditions and severe solar radiation to which they would otherwise be subjected by a snow cover [69–71]. Soil temperatures at high elevations with snow cover are usually higher than in areas without snow [72].

We described for the first time at QLMA the process of reversal of the direction of influence of snow seasonality metrics on LSP metrics with elevation. However, we estimated that the thresholds for LSP (0.3 and 0.7) and the use of seasonal vegetation filter (NDPI > 0.1) may have influenced the strength of the relationship. The vegetation constitute that varies with elevation may be another factor affecting the correlation. Changes in other vegetation proportions can cause fluctuations in correlations (LSD\_FSD below 3300 m), but grass is always the predominant vegetation type in QLMA, which could ensure the relative stability of correlations.

#### *4.3. Why Do the Effects of Snow Seasonality Metrics Vary with Vegetation Type?*

The response of vegetation to snowpack varies considerably between biomes, and similar phenomena have been observed in the QLMA (Figure 10). We found that the LSP metrics of shrub and grass respond most significantly to LSD compared to other vegetation, namely that later snowmelt extends the growing season of both vegetation species. This is consistent with the findings of many other studies on grasslands [34,36]. It may be because shrub and grass have relatively simple structures and short size, so they are more likely to be completely covered by snow. In addition, due to the strong solar radiation, the snow on the grassland melts more easily and water can be supplied to the soil in a timely manner [73]. The LSP metrics of desert only respond to SCD. Compared to other snow seasonality metrics, SCD reflects not only the timing of snow presence, but also the frequency of snow presence, which is important for drought desert. Conversely, alpine vegetation is more sensitive to FSD and SSL and these may reflect the arrival and duration of cold air during the snow season as temperature is more important for alpine vegetation. We also found that the SOS of the forest responds differently to snow seasonality metrics than other vegetation types. Snow falling on branches does not have a direct and timely effect on the root system [74,75]. Due to the tall structure of the trees, the snow in the canopy and on the ground is exposed to different solar radiation, resulting in differences in snow melt [76,77]. Furthermore, some studies have demonstrated that different vegetation types have different temperature requirements for breaking dormancy [69,78]. Woody plants may require cold conditions to promote germination, while grasses require warmer conditions. Snow protects shallow underground root systems from the cold, resulting in different effects on different vegetation types.

In summary, our study illustrated that different vegetation types have different reflections of LSP metrics on snow seasonality metrics, which explains the spatial variation in the impact of snow seasonality metrics from another perspective than elevation.

#### *4.4. Prediction of Vegetation Phenology from Satellite Data Is Beneficial for Future Research*

The concept of using satellite data to estimate vegetation phenology metrics was born long ago. With the continuous improvement of remote sensing image accuracy and cloud computing capability, the estimation of phenology based on satellite data has gradually become reliable. Unlike the phenology metrics obtained from field observations, the remote sensing-based estimation of phenology metrics focuses on the variation of regional greenness. Important time points are calculated from the interannual variation curves of vegetation indices, and thus vegetation phenology is estimated. For stakeholders, satellites can quickly provide long time series and large-scale data, which could also avoid timeconsuming field work. Characterizing vegetation phenology at a larger scale is beneficial for studies such as vegetation response in the context of global climate change.

However, the validation for the remote sensing-based phenological indices and in situ phenological indices is also important. Although there may be temporal differences in the phenological indices obtained by the two methods, both can indicate consistent phenological trends. More in situ data is an important way to improve the accuracy of satellite predictions. This is also the objective of our future study.

#### *4.5. Study Limitations and Future Work*

The medium-resolution satellite data and complex topographic conditions of the Qilian Mountains make it difficult to estimate land surface phenology. Although we used the NDPI, which is least affected by pre-season snowpack, as a vegetation index, the accuracy of such threshold-based extraction remains uncertain. Land surface phenology is a complex parameter that is influenced by a combination of factors. In addition to snow seasonality, vegetation is influenced by snow depth, pre-growing season temperature and precipitation and light conditions. In addition to elevation, slope and aspect play important roles. Thus, assessing the response of land surface phenology requires more accurate models. As mentioned earlier, the differences in water conditions across the study area may be an important factor influencing the conclusions. The content and depth of groundwater, meltwater from permafrost and glaciers, may also have an impact. These would be good to include in future study. Finally, although the vegetation distribution data we used are very reliable, changes are inevitable over long time series, especially in low-elevation areas that are inherently more susceptible to human activities.
