**4. Discussion**

#### *4.1. Dynamics Changes in LSP in the QMs*

Understanding the interannual variations in vegetation phenology and its trends is important for recognizing the patterns of vegetation growth dynamics as a response to climate warming. Our study showed that there is an advanced trend (1.5 days/decade) for SOS, a delayed trend (2.4 days/decade) for EOS, and an overall extended trend (3.9 days/decade) for LOS in the QMs during 2001–2019. In comparison with previous studies on phenology changes, there were different degrees of advanced SOS, delayed EOS, and lengthened LOS in different study areas and periods [32,41]. For example, during 1982–2006, the SOS was advanced by 0.56 days/decade, while the EOS delayed trend rate was 5.5 days/decade, and the growing season was significantly longer by 6.06 days/decade in North America [42]. In the Tibetan Plateau region, SOS was advanced at a rate of 0.17 days/decade, EOS was delayed at a rate of 5.29 days/decade, and LOS was lengthened at a rate of 5.46 days/decade for the period 1981–2017 [41]. These results are not entirely consistent with those of studies conducted for the QMs, and the differences may be mainly due to their different target periods and the different methods of phenology extraction. However, other investigations reported that, compared to 1982–1999, the phenology trend slowed down in 2000–2008 and the changes were not highly significant [14,43]. Meanwhile, our results also show a slower change in phenology trends over the last 20 years in the QMs, and the magnitude of SOS advance is also smaller than that of EOS delay. This observation is similar to the results of Wang et al. and Xia et al. [33,44], which indicates that the satellite-observed phenology change rates slowed down during a global warming hiatus between 1998 and 2012.

We also found that the trend in phenology changes of different vegetation is also highly variable, which is related to the microclimate of different regions or the geographical variation of plant origin [45]. Our results show a significantly greater trend of SOS changes in SL than other vegetation types, indicating that the earlier the SOS, the more significant the trend in SOS variations, and that this difference may be related to plant pollination type, life type, phylogenetic and wood type, etc. [46]. However, the trend of SOS changes in CL was weaker than in other vegetation types, and this trend was insignificant because farmers controlled the sowing time in each year, resulting in significantly smaller variability in crop phenology than in field observed plants [47]. The results show that the trend of delayed EOS is significantly stronger for EBF than other vegetation types, mainly because EBF is mainly located in the region south of the QMs, with a humid northern subtropical climate. Some researchers have shown that in subtropical mountainous and hilly areas, broadleaved forests can grow longer under the same similar climatic conditions compared to coniferous forests [48]. The study also found a significantly stronger trend in the lengthening of growing season for trees than for shrubs and herbs, which is the same as the findings of Zhu et al. [49] but in contrast to those of Ge et al. [50], who reported that the interannual variation trend for trees in China from the 1960s to the 2000s was significantly weaker than for herbaceous plants, and this difference in trend was due to differences in the study area and the species of the plants themselves.

#### *4.2. Asymmetry in Contributions of SOS and EOS Trends to LOS*

We found the asymmetry in contributions of the SOS and EOS trends to LOS variations by counting the percentage of pixels with positive and negative C-index values. The results show that SOS trends control 48.4% of LOS variations and EOS trends control 51.4% of LOS variations, which shows a stronger association between EOS trends and LOS variations compared to SOS (Figure S1). Previous studies illustrated that the lengthened growing season was mainly driven by delayed autumn phenology, which is consistent with our results [14,38,49]. However, other researchers found that it is the changes in SOS, and not EOS, that dominate the changes in growing season length [51,52]. It can be seen that there are differences in previous studies regarding the attribution of LOS variations. To investigate the reasons for such differences, we divided the trends of SOS, EOS, and LOS into six change patterns (Table 2 and Figure 7). Our results show that in addition to the main growth pattern of Type I (SOS advanced, EOS delayed, and LOS lengthened), 15.2% of the regions had Type V (SOS delayed, EOS delayed, and LOS lengthened), and the delayed EOS was significant in this pattern. Another 12% of the regions showed Type II (SOS advanced, EOS advanced, and LOS lengthened) growth pattern, and advanced SOS was significant. However, since the percentage of the region of the growth pattern Type II is smaller than that of Types I and V, it is still the trend of delayed EOS that dominates the variation in LOS for the whole study area, leading to asymmetry of the relative contribution of SOS and EOS to LOS.

In addition, we found that ENF, EBF, MF, SL, and CL were all controlled by EOS trends, while the variations in LOS for two vegetation types, DBF and GL, were controlled by SOS trends (Table 5). As Figure 4e shows, the growing season lengths of DBF and GL are short, and there are previous studies demonstrating that the effect of EOS shifts on vegetation with short growing season cycles is insignificant [53]. The percentage of growth pattern type II is higher than other vegetation types in DBF and GL, and the advance in SOS is also significant, resulting in SOS dominating the variation in LOS (Table 4). Therefore, we sugges<sup>t</sup> that the asymmetry in SOS and EOS trends contributing to LOS is related to vegetation types, and that future studies should focus on vegetation types to accurately model and predict vegetation phenology periods.

#### *4.3. Analysis of the Drivers of Interannual Variations in LSP*

Previous studies showed that the interaction of meteorological, soil, and biological factors influenced the interannual variability of LSP [6,54]. Our results sugges<sup>t</sup> that SWP is the most important driver of interannual variations in SOS and EOS across the QMs (Figure 8). This is mainly because shortwave radiation compensates for the lack of chilling demand during plant physiological dormancy through day length, i.e., longer daylight hours, and has an critical effect on SOS by delaying the accumulation of abscisic acid and slowing down the rate of leaf senescence, and also on EOS [21]. We also found that SWP and MD contributed a total of 42.3% to the interannual variations in SOS, and besides SWP, MD was also an important driver of the interannual SOS variation (Tables 6 and S1). This is due to the fluctuation of the time interval between SOS and MD, which depends on the different developmental stages of the phenology events to a large extent and on the specific differences in the life history of the plant, and needs to be explained by the phenotypic plasticity of the individual and the adaptation to the environment [55]. Therefore, we sugges<sup>t</sup> that the effect of MD on interannual SOS variation varies considerably among vegetation individuals. STP also contributed 12.0% importance in explaining the interannual SOS variation (Table S1). This is mainly due to the increased soil temperature, which accelerated the rate of leaf tip emergence and whole leaf expansion, thus promoting SOS [56]. Moreover, SWP and PP together explain the importance of 43.2% of the interannual EOS variation (Table S1). This is mainly because preseason shortwave radiation and precipitation control the availability of sunlight and water in vegetation, respectively, and reduced precipitation affects water transport capacity, which limits the photosynthetic rate of leaf, leading to lower utilization of light and water conditions by plants and affecting the interannual variation in EOS [20,57]. The combined contribution of MD, SWP, and PP to interannual EOS variation was also found to be as high as 54.0% (Table S1), suggesting that the lifecycle of vegetation is strongly regulated by its own rhythms under improved hydrothermal conditions, and that biological rhythms play a critical role in interannual EOS variation [7].

Furthermore, our study shows that the effect of each driver on interannual variations in LSP was varied for different vegetation types (Figure 9 and Table 6). For example, SWP is the most important driver for ENF, MF, GL, and CL; MD is the most important driver for EBF and DBF; and STP is the most important driver for SL (Table 6). This difference is mainly due to the diversity of plant physiological structures and the different adaptive strategies of plants to environmental changes [58]. SWP has the greatest effect on the interannual SOS variation in GL, which is mainly distributed in higher parts of the QMs and receives abundant solar radiation. The strong solar radiation promotes root activity and advances SOS [59]. The greatest contribution of MD to interannual SOS variation in EBF is related to the fact that EBF grows mainly in the southern part of the QMs, where its deeply rooted system and water conservation adaptations combine to reduce water stress under the influence of a humid northern subtropical climate. This adaptation to environmental changes is strongly regulated by its own rhythms, such that EBF is most affected by MD [60]. The effect of STP in SL is mainly due to the preseason accumulation of soil temperatures susceptible to specific thresholds that accelerate soil thaw and vegetation wake, triggering SOS [61]. SWP was the main driver of interannual EOS variation for all vegetation types, with GL being most influenced by SWP (Table 6). This is because abundant solar radiation increases surface evaporation and reduces water availability in grasslands, which subsequently inhibits vegetation growth, resulting in the EOS of GL being most influenced by SWP [62]. PP and MD have the strongest effect on interannual EOS variation in SL (Table 6). To our best knowledge, SL is mainly distributed in semi-humid and semiarid areas, and the control of plant metabolism by water stress affects its transpiration and photosynthesis, resulting in impaired ATPase synthesis and accelerated chlorophyll degradation. Meanwhile the adaptation of vegetation to such adversity changes also affects the interannual variations in EOS [63]. The relationship between regional climate and vegetation phenology growth will be further explored in future studies.

#### *4.4. Evaluation of RF Model*

We validated the accuracy of our RF model (Figure S2). The R<sup>2</sup> values of the linear regression formed by the predicted values and the observations inversions are both greater than 0.900 and 0.911, respectively. When predicting the dates of SOS or EOS, both RMSE and MAE are relatively small, showing that our RF model has good predictive performance. Machine learning, as a nonparametric multivariate approach, can integrate complex relationships between multiple spatial and temporal LSP dates and climate into a single model for predicting SOS and EOS. Then, the main drivers of SOS and EOS can be identified by estimating the importance of each variable [6]. However, it should be noted that although we have tried our best to adjust the hyperparameters of the algorithm to prevent overfitting in the model, some errors still appear in the test datasets (Table S2). Therefore, to obtain a better fit, it is necessary to further refine the study area and tree species in the future. Further study should compare different algorithms to better simulate the phenology period.
