In our study, we conducted an analysis covering a total of 293 prefecture-level cities in China (excluding Taiwan, autonomous prefectures, and other special administrative regions), based on data from the year 2020 (http://xzqh.mca.gov.cn/map, last access: 19 August 2026) (Fig. 1). We hypothesized that the urban warming effect is most pronounced in urban centers and gradually diminishes towards rural areas (Zhang et al., 2004). To determine the urbanization intensity of each city, nighttime light data from NPP-VIIRS (Table 1) were used (Elvidge et al., 2017), employing the comparison method. The optimal segmentation threshold was established based on the statistical yearbook (Fan et al., 2019). Moreover, urban built-up areas were identified using the comparison method and annual average data derived from NPP-VIIRS nighttime light data between 2013 and 2020. All data processing procedures were performed using ArcGIS 10.2. We created ten buffer zones extending to a radius of 20 km from each city center, with each zone representing a 2 km annulus (Fig. 1).

This 20 km boundary was chosen because our preliminary analysis revealed that the urban warming footprint on phenological metrics (SOS and EOS) gradually diminishes to near zero beyond 15 km (Zhang et al., 2004; Figs. 1 and 2). Thus, defining the outermost 20 km buffer zone as the “Rural” baseline ensures a background environment largely free of urban heat island interference. The subscript (10) refers to the 10th buffer zone at 20 km from the urban center. Equation notations use a standardized buffer index (0 for urban center, 10 for rural reference) to avoid confusion between distance and buffer number.

Phenology data spanning the years 2010 to 2020 were derived from MODIS Enhanced Vegetation Index (EVI) 16 d composite data with a spatial resolution of 250 m (MOD13Q1) and processed with TIMESAT (Jönsson and Eklundh, 2004). To mitigate challenges such as cloud cover and snow, the raw EVI time series data were smoothed using the Savitzky-Golay (S-G) filter method (Cai et al., 2017). Subsequently, a dynamic threshold technique was employed to extract the phenology data, utilizing a seasonal amplitude of 20 %, a parameter proven effective for phenology extraction in urban environments (Zhou et al., 2016; Buyantuyev and Wu, 2012; Cong et al., 2012; White et al., 2009). The datasets underwent preprocessing and quality control to eliminate pixels with minimal vegetation EVI (e.g., bodies of water and fully paved surfaces). To exclude unrealistic phenological values, we retained only SOS and EOS values within valid ranges (30–180 d for SOS and 210–360 d for EOS), consistent with previous studies (Zhang et al., 2006). To reduce the influence of outlier years, we averaged the phenology data for each pixel over 2010–2020. After applying these quality control thresholds and land-cover filtering criteria, approximately 48.5 % of the initial pixels within the study buffer zones were excluded.

Furthermore, we employed land cover data from the MCD12Q1 dataset, which has a 500 m resolution, to mitigate the impact of agricultural practices on the mainland during our analysis. The MCD12Q1 dataset categorizes land into 17 distinct classes. These include various types of forested areas such as Evergreen Coniferous Forests, Deciduous Needle-leaf Forests, Evergreen Broad-leaf Forests, Deciduous Broad-leaf Forests, Mixed Forests, as well as Closed Shrub lands, Open Shrub lands, Savannas, Woody Savannas, Grasslands, and Croplands.

Given the frequent human-induced disturbances in rural croplands and the regular manual pruning of urban grasslands, we excluded savannas, woody savannas, grasslands, and croplands from our analysis. The remaining land types were consolidated into broader categories: broad-leaf forests (BF, deciduous and evergreen broadleaf forests), coniferous forests (CF, deciduous and evergreen coniferous forests), mixed forests (MF), and shrubs (closed and open shrublands). This classification approach allowed us to focus our analysis on land types that are less influenced by human activities.

Building upon the methodology of a previous study (Piao et al., 2006), we defined pre-season temperature (T) as the mean temperature for the three months preceding phenological onset. To align the temperature data with the phenological data, the average temperature dataset was resampled to a 250 m resolution using the nearest neighbor interpolation technique. To evaluate the performance of the downscaling procedure, the observed long-term monthly temperature and precipitation across China were obtained from the National Meteorological Information Center of China (http://data.cma.cn/en, last access: 19 August 2026). This dataset included observations from 496 national weather stations during 2010 to 2020. Upon conducting a fitting analysis with station data, we observed that the coefficient of determination (R2) for monthly mean temperature and remote sensing data reached 0.93, while that for monthly precipitation and remote sensing data achieved 0.85. These results indicated that the downscaled meteorological data sufficiently met the requirements of this study.

Background climate was characterized using land surface temperature (LST) and the aridity index (AI) as indicators. The LST and AI data for each pixel were averaged from 2010 to 2020. Specifically, MODIS LST datasets exhibiting an absolute bias of less than 1 K were selected as optimal for our investigation, in line with previous studies (Gow et al., 2016; Pablos et al., 2016; Qiao et al., 2013; Wan, 2008). Collection 6 LST products from the Terra satellite (MOD11A2, 8 d composite, 1 km spatial resolution, covering the period from 2010 to 2020) were used. Additionally, daily temperature (T) was calculated as the mean of daytime and nighttime temperatures, following the methodology outlined by Kang et al. (2016). AI was expressed as: 1AI=PETP where PET denotes the annual potential evapotranspiration and P denotes the annual precipitation. The PET values spanning from 2010 to 2020 were extracted from the MOD16A2 monthly synthetic product, characterized by an 8 d temporal resolution and a spatial resolution of 500 m (Mu et al., 2011), as delineated in Table 1. The annual precipitation (P) dataset was derived from the monthly precipitation records of China encompassing the timeframe from 1901 to 2020, at a spatial resolution of 1 km. Subsequently, both PET and P datasets were resampled to a spatial resolution of 250 m to facilitate the ensuing spatial analysis (Peng, 2020).

To robustly quantify these responses, we defined the distinct analytical roles of two critical thermal dimensions: pre-season air temperature (T) and land surface temperature (LST). As the direct biological driver of plant physiology (e.g., chilling requirements), T was used exclusively to quantify the pure biological phenological sensitivity (Rt) and identify true physiological thresholds. In contrast, as the standard metric for quantifying the urban heat island (UHI) effect, LST was used to characterize the spatial “background climate” of each city across the continent.

Because pre-season air temperature (T) is the direct driver of phenological sensitivity, we first derived the biological thresholds using T (Fig. 5). However, to address our core objective of quantifying how background LST shapes these phenological shifts, we defined the quantitative relationship between the background LST and T. We fitted the relationship between LST and T across all cities. To minimize data matching biases during this conversion, we used the spatiotemporally aligned 10-year climatological means (2010–2020) for both variables at a unified 250 m resolution. We observed a highly robust linear correlation (Fig. S4). Based on this high goodness-of-fit, we applied the empirical conversion formula between the two variables to mathematically replace T with LST in the sensitivity curves (Wan, 2008). This empirical translation allowed us to systematically convert the T-driven physiological thresholds into the specific LST-determined background climate thresholds (12.5 °C for SOS and 4 °C for EOS), explicitly projecting pure biological limits onto the urban map. Although T and LST represent different physical properties, they are highly correlated in urban settings (R2=0.93 for monthly mean temperature), confirming the reliability of this empirical conversion between physiological response and climatic background.

The phenology dates of each vegetation type were extracted in each buffer zone. Finally, the mean phenology in each buffer zone was calculated as a weighted mean of the phenology of each vegetation type. To ensure comparability in phenological analyses across buffer zones, we first identified the common vegetation types present in all zones and excluded the non-shared types. We then calculated the phenological means for these common types within each buffer zone and averaged these values to determine the overall mean vegetation phenology for the entire buffer zone, as shown in Eqs. (2) and (3). 2SOSi=∑SOSijn3EOSi=∑EOSijn SOS_i and EOS_i arethe mean SOS and EOS of the ith buffer zone; SOS_ij and EOS_ij are the mean SOS and EOS of the jth vegetation type in the ith buffer zone; n is the number of shared vegetation types across all buffer zones.

To investigate the impact of urbanization on phenology, mean values of start of season (SOS) and end of season (EOS) in each buffer zone were analyzed during 2010–2020, and fitted with respect to their distance from the urban center. In this study, the outermost buffer zone (20 km from the urban center) was defined as the rural reference region, and the urban center was defined as the urban region. This delineation was justified by the observation that phenological change rates approached zero at 15 km and beyond (Fig. S1), and similar patterns were observed for all four vegetation types (Fig. S2). Following prior investigations, the phenological sensitivity (Rt) to temperature was quantified by calculating the partial correlation coefficient between T and the phenology dates (Rt-SOS and Rt-EOS) (Meng et al., 2020). The phenological difference between urban and rural regions was calculated to quantify the magnitude of the urban warming effect on phenology: 4ΔSOS=SOS(0)-SOS(10)5ΔEOS=EOS(0)-EOS(10) where SOS₍₀₎ and SOS₍₁₀₎ are the mean SOS of the urban center (buffer index 0) and the 10th buffer zone (located 20 km away), respectively. Similarly, EOS₍₀₎ and EOS₍₁₀₎ are the mean EOS of the urban center and the 10th buffer zone. Biologically, a negative ΔSOS indicates an advanced start of the season (spring) in the urban center relative to the rural area, while a positive ΔEOS indicates a delayed end of the season (autumn senescence) in the urban center.

Similarly, the difference in phenological sensitivity to temperature between urban and rural regions was defined as: 6ΔRt-SOS=Rt-SOS(0)-Rt-SOS(10)7ΔRt-EOS=Rt-EOS(0)-Rt-EOS(10) where Rt-SOS(0) and Rt-SOS(10) are the mean Rt-SOS of the urban center (buffer index 0) and the 10th buffer zone (located 20 km away), respectively. Similarly, Rt-EOS(0) andR_t-EOS(10) are the mean Rt-EOS of the urban center and the 10th buffer zone.

To elucidate the distinct impacts of LST and AI on urban and rural components, we performed a partial correlation analysis. Specifically, the average LST or AI values from 2010 to 2020 for each of the 293 cities were correlated with their respective ΔSOS, ΔEOS, ΔRt-SOS, and ΔRt-EOS values, while controlling for the influence of the other climatic factor.

Path analysis (PA) (Doncaster, 2007) was used to examine both the direct and indirect effects of background climate variables (LST and AI) on phenological responses (SOS and EOS), as well as their corresponding discrepancies (ΔSOS and ΔEOS) and rates of change (ΔRt-SOS and ΔRt-EOS).

In the path analysis, we first assessed model fit using fit indices, including the chi-square (χ2) statistic, Comparative Fit Index (CFI), and Root Mean Square Error of Approximation (RMSEA). Then, path coefficients were derived for both the direct and indirect effects of each independent variable on the dependent variable from the model. Squared path coefficients were used to estimate the respective direct and indirect contributions of each independent variable to the dependent variable. The sum of these direct and indirect contributions represents the total contribution, that is, the relative contribution.

To determine if there was a temperature saturation effect on vegetation phenology, we conducted a model comparison. Rather than assuming a single relationship a priori, we evaluated the fit of phenology versus pre-season temperature using competing models: a linear model, a quadratic model, a segmented (piecewise) model, and a logistic model. We specifically selected the logistic function as our primary model because quadratic models biologically imply a symmetric reversal (a continuous steep decline) rather than a plateau, and piecewise models assume an abrupt, angular change in biological rates, which is less physiologically realistic than the smooth, asymptotic transition captured by the logistic curve (Pinheiro and Bates, 2000). The asymptote of the logistic function mathematically represents the biological “saturation point” – the maximum physiological limit of thermal responses.

Statistically, the model comparison results (Table 2) confirmed that the logistic model consistently outperforms the other three formulations for both spring SOS and autumn EOS, providing quantitative support for the non-linear saturation effect interpretation. For SOS, the logistic model yields an AIC value 2778 units lower than the linear model and 1424 units lower than the piecewise model, along with an R2 of 0.37, which is approximately 19 % higher than that of the piecewise model. For EOS, the logistic model still achieves the lowest AIC and highest R2, but the performance improvement is more modest: its R2 (0.40) is only marginally higher than that of the piecewise model (0.39), consistent with a weaker but stable saturation pattern.

We then extracted phenological information from all vegetation pixels across the country and generated scatter plots of phenology versus pre-season temperature. Next, we fitted logistic growth curves to these scatter plots and evaluated the assumption using the goodness-of-fit metrics, p-value and R2. Subsequently, we applied the same logistic curve to characterize the phenological responses of urban and rural vegetation in the 293 different cities against their pre-season temperatures, testing whether the phenological response to temperature in urban areas also follows a logistic growth pattern. Finally, we replaced the pre-season temperature with the background temperature LST to determine how the phenological saturation effect is regulated by the background temperature.

Among the 293 cities analyzed, 84 % (N=246) exhibited an advanced SOS in urban areas compared to rural regions, while 82 % (N=241) showed a delayed EOS (Fig. S3). Spatially, cities located in high-latitude regions generally demonstrated more negative ΔSOS values than those in low-latitude regions (Fig. 2), indicating that urban warming advanced spring phenology most strongly in northern and western China (Fig. 2a, b). This pattern is consistent with the conventional understanding of earlier SOS in colder regions. Additionally, cities in coastal provinces exhibited larger ΔSOS values than inland cities (Fig. 2).

A strong positive correlation was observed between ΔSOS and the rural rate of SOS change (Rt-SOS), suggesting that cities with more negative Rt-SOS values exhibited larger urban-rural disparities (Fig. 3). Notably, cities with lower temperatures tended to have more negative Rt-SOS and ΔSOS values.

ΔEOS also showed a significant increasing trend with latitude (Fig. 2), with northern cities displaying more positive ΔEOS values, indicating a delayed autumn phenology in urban areas (Fig. 2c, d). In contrast, the majority of cities in southern regions demonstrated negative ΔEOS values, signifying an advancement in autumn phenology. Furthermore, cities in the coastal provinces also exhibited larger ΔEOS values.

Similarly, a positive correlation was observed between ΔEOS and the rural rate of EOS change (Rt-EOS). Positive and negative ΔEOS values were predominantly found in cold and warm regions, respectively (Fig. 3). Moreover, when ΔEOS > 0, both the absolute value of ΔEOS and ΔRt-EOS increased with temperature, while for ΔEOS < 0, the absolute value of ΔEOS and ΔRt-EOS decreased with temperature.

As we compared urban and rural regions, logistic trends were observed for both SOS and EOS in relation to LST (Fig. 4). In urban areas where LST was below 12 °C, the SOS occurred notably earlier compared to rural areas, while urban areas exhibited a delayed EOS when LST was below 18.5 °C. At the low-LST (cold) end, SOS plateaued at 101.3 d in urban areas and 129.7 d in rural areas, while the EOS stabilized at 321.3 d for urban areas and 310.8 d for rural areas. In contrast, when LST exceeded the specific thresholds for urban-rural differences (LST = 18 °C for ΔSOS and LST = 18.5 °C for ΔEOS), the phenological disparities between urban and rural areas were no longer significant and sometimes even reversed (Fig. 4c, d). As LST increased, the urban and rural SOS stabilized at 71.7 d (Fig. 4a), and the urban and rural EOS stabilized at 327.4 and 332.4 d, respectively. Fitted curves of ΔSOS and ΔEOS against LST revealed a positive and a negative relationship, respectively, intersecting at LST = 18 and 18.5 °C, indicating a shift in SOS and EOS dynamics with increasing temperatures beyond these thresholds.

To determine whether the temperature sensitivity influenced the logistic trends of phenology with LST, we conducted a nationwide fitting of phenology to T and derived logistic curves for SOS and EOS (Fig. 5). The first-order derivatives of the fitted curves, representing the phenological sensitivity to T, revealed warming-induced phenological effects on SOS and EOS, ranging from -3.5–10 °C and 2.0–13.5 °C, respectively. The highest sensitivity was observed at T=4 °C for SOS and T=6 °C for EOS. Notably, a decreasing trend for EOS was observed when T > 15 °C, corresponding to the advanced EOS when T > 18.5 °C (Fig. 5d).

To assess the temperature sensitivity of phenological variations between urban and rural settings across vegetation types, we established regression models of ΔSOS and ΔEOS against LST for each type (Fig. 6). Results showed that shrublands had the highest temperature sensitivity for both metrics, whereas CF displayed the lowest. The temperature response of MF and BF did not differ significantly.

To obtain LST-determined phenological sensitivity, the T in the first-order derivative curves shown in Fig. 5c and d was converted to LST based on the strong correlation between T and LST (Fig. S4). Three distinct LST thresholds were identified: phenological date stabilization thresholds, urban-rural disparity reversal thresholds, and phenological sensitivity saturation thresholds. Specifically, phenological date stabilization and urban-rural disparity reversal occur at LST = 12 °C (SOS) / 18.5 °C (EOS), while phenological sensitivity saturates at LST = 12.5 °C (SOS) / 4 °C (EOS). Based on the first-order derivatives, we identified the specific thresholds for phenological sensitivity. We observed that when background LST exceeded 12.5 °C for spring and 4 °C for autumn, the temperature sensitivity (Rt) began to weaken significantly. This indicates a saturation of the physiological response rate, even though the absolute urban-rural difference (ΔSOS / ΔEOS) might still persist until higher temperatures are reached.

To characterize how phenological responses vary with background aridity, we fitted linear regressions of Rt-SOS, Rt-EOS, ΔSOS, and ΔEOS against the aridity index (AI) (Fig. 8). The zero-crossing points of the fitted regression lines (i.e., the AI value at which the regression-predicted response variable equals zero) were used to demarcate contrasting phenological response regimes. Because these response variables are signed, these zero-crossing points separate opposing phenological patterns (e.g., from warming-induced advancement to warming-induced delay of SOS).

Rt-SOS was positively related to AI (Fig. 8a), with its fitted regression line crossing zero at AI = 1.7. When AI < 1.7, the absolute value of Rt-SOS decreased with increasing AI, indicating that increasing aridity weakened the temperature sensitivity of SOS. In contrast, when AI > 1.7, most urban areas exhibited positive Rt-SOS (resulting in delayed SOS), with higher AI values corresponding to greater Rt-SOS, suggesting that severe drought markedly delayed the onset of SOS.

Rt-EOS was negatively related to AI (Fig. 8c), with its fitted regression line crossing zero at AI = 1.4. When AI < 1.4, most Rt-EOS values were positive and decreased with AI, i.e., intensified drought counteracted the EOS-delaying effect induced by urban warming. When AI > 1.4, most cities exhibited negative Rt-EOS, and higher AI values corresponded to more negative Rt-EOS. That is, extreme drought reversed the warming-delaying effect on EOS.

When urban-rural phenological differences were fitted to AI, ΔSOS was positively related to AI (Fig. 8b), indicating that increasing aridity progressively reduced the advanced green-up typically induced by urban warming. ΔEOS was negatively related to AI (Fig. 8d), with its fitted regression line crossing zero at AI = 2.0. Notably, negative ΔEOS tended to appear when AI > 2, suggesting that under extreme drought, urban EOS occurred earlier than rural EOS.

The same relationships were evident across all four vegetation types (Fig. 9): increasing AI was associated with a reduced advance of SOS (less negative ΔSOS) and an earlier urban EOS (more negative ΔEOS). Notably, ΔSOS for CF showed less sensitivity to AI than the other vegetation types. Both BF and MF demonstrated lower sensitivities to AI for both ΔSOS and ΔEOS than shrublands.

To further quantify the individual and combined contributions of LST and AI to phenological variations, a variance decomposition analysis was conducted (Legendre and Legendre, 2012; Fig. 10). The Venn diagram revealed that AI and LST explained 75.05 %, 76.21 %, 71.72 % and 82.14 % of the variance in ΔSOS, ΔEOS, Rt-SOS, and Rt-EOS, respectively (Fig. 10), and LST always contributed more to these phenological indicators. Additionally, the interaction effect of LST and AI ranged from 3.29 % to 5.66 %.

Our multi-city assessment confirms a predominant urban-rural phenological divergence across China, characterized by advanced SOS and delayed EOS within a concentrated 10 km urbanization footprint. This consistent spatial pattern confirms that urban warming acts as a primary driver of phenological shifts. Beyond mere thermal forcing, this convergence across 293 cities reflects the profound impact of urban microclimates on plant life cycles. The observed 10 km threshold is particularly significant as it suggests a localized but intense ecological footprint that necessitates specialized urban greening management strategies, especially in high-density northern cities where these shifts are most pronounced.

Spatially, the pronounced north-south gradient in ΔSOS and ΔEOS across China, as shown in Fig. 2, closely mirrors the spatial pattern of LST, reflecting stronger temperature sensitivity of vegetation phenology in northern regions than in the south. This finding aligns with previous research documenting greater phenological shifts in northern areas (Jia et al., 2021; Yang et al., 2025), and further highlights that background climate acts as a key modulator of urban phenological impacts. The more pronounced urban warming effects observed in coastal provinces relative to inland cities (Fig. 2) likely stem from the buffering effect of the maritime climate on seasonal temperature constraints, which interacts with urban heat islands to amplify urban-rural phenological divergence. Moreover, the strong correlations between ΔSOS / ΔEOS, temperature sensitivity indicators (Rt-SOS and Rt-EOS), LST, and AI, as presented in Fig. 3, confirm that background temperature and aridity are critical drivers shaping the magnitude of urbanization-induced phenological shifts across the 293 cities, rather than just passive correlates, validating the core hypothesis of this study.

Notably, the pattern of urban warming's effects on vegetation phenology was consistent among the four vegetation types (Fig. S2). This consistency could be attributed to the ecological convergence of plants coexisting in shared local habitats within each city (Wang et al., 2016). Ecologically, the consistent 10 km effective footprint validates that urban environments function as distinct microclimatic islands. This spatial boundary conceptually highlights that urban ecological engineering and greening strategies must account for highly localized, rather than regional, thermal forcings.

When we quantified phenological responses to temperature across the climatic gradient, clear logistic trends were observed for both SOS and EOS (Fig. 5). Consistent with this logistic trajectory, we observed that in urban areas, the growth rate for SOS / EOS began to plateau when temperatures exceeded 4 °C/6 °C, respectively (Fig. 5). Given the significant correlation between LST and T (Fig. S4), we anticipated a diminished sensitivity in phenology with higher LST (Fig. 7). These findings lend credibility to the “saturation” effect hypothesis, suggesting that the accelerating influence of warmth on vegetation phenology reported in earlier studies has its limits (Badeck et al., 2004; Li et al., 2021). Indeed, plants in southern regions exhibit not just a dampened phenological response but also a decelerated growth rate when faced with increasing temperatures (Li et al., 2021; Büntgen et al., 2019). As illustrated by the first-order derivatives in Fig. 5c and d, the phenological sensitivity to temperature does not increase linearly, but peaks at T=4 °C for SOS and T=6 °C for EOS before rapidly declining. Specifically, the stabilization of urban SOS at 101.3 d and rural SOS at 129.7 d when LST was low (Fig. 4a) further demonstrated that plants reach a required physiological threshold that cannot be bypassed by additional thermal forcing. Collectively, these results highlight that threshold-based non-linear models rather than simple linear relationships are more suitable for predicting phenological responses under scenarios of continuous environmental warming, especially in urban settings. Conceptually, these saturation thresholds challenge the traditional linear paradigm of “warming-driven phenological advancement”. By identifying explicit thermal boundaries, our framework establishes that the phenological response of urban vegetation to additional warming is bounded.

The observation of temperature saturation for the SOS in our study likely stems from the chilling requirements for vernalization (Kim et al., 2009; Hanninen et al., 2019). Typically, vegetation needs to undergo a period of exposure to low chilling temperatures to break dormancy. This mechanism is consistent with our spatial observations in Fig. 2a, where cities in lower latitudes (warmer regions) exhibited significantly smaller ΔSOS values. This suggests that in these warmer regions, the higher background LST may fail to satisfy the necessary chilling requirements, leading to the diminished phenological sensitivity we quantified in Fig. 3. Moreover, an earlier onset of greening increases the risk of damage from freezing events, posing a survival threat to plants (Duan et al., 2011; Wang et al., 2015). This phenomenon explains the delayed SOS in approximately 16 % of the cities under study, mainly distributed in low-latitude regions (Fig. 4).

Regarding the EOS, a moderate increase in temperature can enhance the activity of photosynthetic enzymes (Shi et al., 2014), slow down the degradation of chlorophyll (Fracheboud et al., 2009), reduce the risk of frost exposure in autumn, and increase the potential for growth and photosynthetic production, ultimately promoting the extension of the growing season. However, the positive effects of increased temperature can be offset by the negative impacts of heat waves and excessive thermal stress (Mathur et al., 2014). Our data in Fig. 5b and d provide direct evidence for this: when pre-season temperature (T) exceeded 15 °C, a decreasing trend for EOS was observed. This aligns with the intersection point identified in Fig. 4d at LST = 18.5 °C, where the urban-rural difference in EOS reverses, confirming that excessive heat in urban centers triggers premature leaf senescence rather than further delaying it. In fact, when T > 15 °C, EOS tends to be reduced, with an increased sensitivity to temperature changes (Estiarte and Peñuelas, 2015; Fig. 5b).

Despite the existence of “saturation” effects, lower background temperatures do not necessarily correspond to higher phenological sensitivity (Fig. 4b, c). For SOS, when T < 4 °C, the phenological sensitivity approached zero with decreased T (Fig. 5c), which led to a stable SOS of 120 d (Fig. 5a, c). In fact, phenological events such as bud burst could not be triggered when the threshold for high forcing temperatures was not reached (Hanninen et al., 2019). For EOS, when T > 6 °C, the increased T sensitivity produces more EOS advancement in warmer regions. This reversal may be associated with abscisic acid synthesis induced by low temperatures, which accelerates the progress of defoliation.

Beyond thermal controls alone, our findings regarding the aridity-driven modulation of phenological responses provide a critical geographic context that complements previous observations. While earlier studies primarily emphasized thermal forcing as the dominant driver of urban phenology (Meng et al., 2020), our results demonstrate that AI acts as a significant “environmental filter” that can override or even reverse warming-induced phenological effects. As shown in Figs. 8 and 10, the high explanatory power of AI (17.33 % for ΔEOS) suggests that in water-limited regions of northern and western China, the physiological stress induced by drought prevents plants from utilizing the additional heat provided by the Urban Heat Island (UHI) effect. This phenomenon is consistent with the “resource limitation” theory, where plant development is constrained by the most limiting resource (water; Harpole et al., 2011), regardless of optimal temperatures. Compared to coastal cities where water is abundant, the weakened warming-induced phenological effects in arid inland cities highlight the need for regionally differentiated urban irrigation strategies to maintain ecological productivity. However, when LST < 12.5 °C, warming regions exhibited higher urban-rural differences in ΔRt-SOS, which contradicted the weakened urban-rural difference in ΔRt-SOS reported in the previous study. For EOS, the warming-induced phenological effects only occurred when 0 < LST < 18 °C. Moreover, the weakened sensitivity was only detected at LST > 4 °C. The weakened ΔRt-SOS and ΔRt-EOS reported in the USA may reflect the fact that the LST range there was mostly within 8–22 °C.

In addition to background temperature (LST), the aridity index (AI) drives significant spatial variation in urban-rural phenological disparities (ΔSOS, ΔEOS). Spatially, the coastal cities experienced more pronounced warming-induced phenological effects (Fig. 2). Our statistical decomposition in Fig. 10 further elucidates this aridity-driven modulation, showing that AI and LST together explain more than 75 % of the variance in ΔSOS and ΔEOS. Notably, the fact that AI explains 17.33 % of the variation in ΔEOS (Fig. 10c) supports our inference that water availability is a critical limiting factor in autumn. This is further substantiated by the positive correlation between ΔSOS and AI in Fig. 8b, which quantifies how increasing aridity progressively offsets the advanced green-up typically induced by urban warming. These results highlight that the warming-induced phenological effects on vegetation phenology will be weakened in those drought-dominated regions. In fact, the delayed SOS has already been demonstrated in previous studies (Ji et al., 2021; Yu et al., 2003).

Water deficit inhibits leaf growth via chemical signals (Knauer et al., 2017), and urbanization exacerbates drought stress (Zhang et al., 2019; Pollastrini et al., 2019), further weakening or even reversing the positive effects of urban warming on phenology. Furthermore, in water-stressed regions, temperatures are already close to the optimum for photosynthesis, thus fewer positive physiological effects are expected (Peñuelas et al., 2004). The weakened temperature sensitivity in drought-prone cities is likely to lead to less extension of the growing season.

Notably, the influence of AI on autumn phenology response was higher than that on spring phenology response (Fig. 10). During spring leaf development, plants may use water stored in their bodies to support the growth of new leaves (Shi et al., 2023). Supporting evidence comes from the synchrony between plant water storage and leaf sprouting for the boreal and temperate forests (Tian et al., 2018). In this case, plants may have relatively little demand for external water supply, especially in some wetter areas. By comparison, in summer and autumn, plant transpiration is higher, and water demand is greater (Wu et al., 2022), so the influence of drought severity on plant phenology is more pronounced. The overarching ecological implication here is a profound vulnerability: as background climates dry under global change, the compounding stress of UHI and background aridity may lead to “phenological compression”. In arid urban ecosystems, this moisture deficit effectively neutralizes the anticipated carbon sequestration benefits of thermally extended growing seasons, rendering them highly sensitive to future climate extremes.

The phenological response to LST and AI varies significantly across vegetation types, reflecting the influence of distinct ecological strategies. Our results (Figs. 6 and 9) demonstrate that shrublands are the most sensitive to both warming and aridity, likely due to their opportunistic resource utilization strategy and higher leaf turnover rates (Xiong et al., 2024). In contrast, coniferous forests exhibited a muted response to both drivers (Malla et al., 2023). While our analysis relies on broad land-cover categories and does not directly measure species-specific functional traits, these discrepancies likely reflect generalized divergent ecological strategies among the vegetation groups.

This insensitivity is substantiated by our AI-sensitivity analysis (Fig. 9a), suggesting that the conservative hydraulic traits of conifers – such as needle morphology and lower stomatal conductance (Sperry et al., 2002; Brodribb and McAdam, 2013) – act as a buffer against the “aridity-driven inhibition” that more severely affects broadleaf species. Based on generalized ecological principles, this might be associated with typical drought-tolerant adaptations found in many conifers (e.g., needle morphology and stricter stomatal control), which could limit the potential positive impact of warming on their phenology. In contrast, the heightened sensitivity of shrublands observed in our dataset aligns with literature suggesting more opportunistic resource utilization strategies in shrubs (Xiong et al., 2024).

By considering these functional differences, our study provides a more nuanced understanding of urban ecosystem resilience, suggesting that urban greening programs should prioritize drought-tolerant coniferous or mixed forest structures in regions where the warming-saturation effect is most pronounced. However, we note that these functional interpretations remain broad generalizations; future studies incorporating field-based, species-level trait measurements are required to mechanistically validate these large-scale categorical patterns. Recognizing these trait-mediated responses is a crucial conceptual step for urban forestry. It indicates that building climate-resilient urban ecosystems requires shifting from generic greening to deploying specific functional traits – such as the conservative hydraulic strategies of conifers – to buffer against the coupled heat-aridity saturation effects.