Nonlinear Impacts of Interannual Temperature and Precipitation Changes on Spring Phenology in China’s Provincial Capitals

Highlights

What are the main findings?

• For 31 provincial capitals and municipalities in mainland China (2001–2023), the start-of-season (SOS) date, derived from MODIS MCD12Q2 Greenup_1 as the spring green-up transition metric, is primarily distributed within approximately DOY 73.55–138.11 and advances by 0.81 d·yr⁻¹ on average, exhibiting earlier timing in the south, later timing in the north, and delayed timing over plateau regions.
• Among pixels with significant or marginally significant SOS trends identified by the Mann–Kendall test (MK p < 0.10), advancing and delaying SOS trends commonly coexist within most cities, while advancing pixels dominate overall (75.02% vs. 24.98%). SHAP dependence relationships further reveal generally nonlinear and piecewise effects of interannual temperature- and precipitation-change rates (tem_slope, pre_slope), delineating spatially differentiated climate-sensitive intervals via tipping points.

What are the implications of the main findings?

• The identified tipping points and associated sensitive ranges provide quantitative, interpretable climate thresholds that can support urban greening design and climate-adaptive management under sustained warming and precipitation changes.
• The unified 500 m grid workflow integrating MK–Sen trend characterization with XGBoost–SHAP attribution and threshold identification offers a transferable framework for diagnosing spatially heterogeneous, nonlinear climate–phenology relationships across cities.

Abstract

Spring vegetation phenology is highly sensitive to climate change; however, climate drivers and their threshold responses at the urban scale remain insufficiently and systematically quantified. Focusing on 31 provincial capitals and municipalities in mainland China, this study integrated MODIS MCD12Q2-derived start-of-season (SOS) for spring green-up and TerraClimate climate data (2001–2023) at a 500 m grid resolution. SOS trends were characterized using the Mann–Kendall test and the Theil–Sen slope estimator. Building on these trend metrics, we developed an XGBoost–SHAP framework using the interannual rate of temperature change (tem_slope) and the interannual rate of precipitation change (pre_slope) as input features, to quantify the nonlinear contributions of climate-change rates to SOS trends and to identify key thresholds. Results indicate that the multi-year mean SOS across China’s provincial capitals and municipalities is primarily distributed between approximately DOY 74 and 138, exhibiting a clear spatial pattern of earlier green-up in the south, later green-up in the north, and delayed green-up on plateaus, with pronounced shifts in distribution centers and dispersion among climatic zones and cities. At the city level, the mean SOS trend shows an overall advancing rate of 0.81 d·year⁻¹ (i.e., the average of city-mean Sen slopes across the 31 cities). Pixel-level trend analyses show that advancing and delaying trends commonly coexist within most cities; among pixels with significant or marginally significant SOS trends identified by the Mann–Kendall test (MK p < 0.10) across all cities, advancing and delaying SOS pixels account for 75.02% and 24.98%, respectively. At the city scale, the proportions of advancing versus delaying pixels vary markedly among cities, forming directional structures characterized by advance-dominant, delay-dominant, or bidirectional coexistence patterns. SHAP dependence relationships further reveal that the effects of tem_slope and pre_slope on SOS trends are generally nonlinear and piecewise, with substantial heterogeneity across climate zones and cities. The identified tipping points and associated sensitive ranges collectively delineate spatially differentiated climate-sensitive intervals, which define the nonlinear response boundaries of spring SOS to sustained warming and precipitation changes. This study provides quantitative evidence for regional differences in urban spring phenological responses to climate change across major Chinese cities and offers a methodological reference for identifying actionable climate thresholds in urban greening design and climate-adaptive management.

1. Introduction

Vegetation phenology refers to the seasonal timing of a series of periodically occurring biological activities in plants and has long been regarded as a sensitive indicator of climatic variability and ecosystem responses [1,2,3]. As an important component of land–atmosphere interactions, phenological processes influence ecosystem productivity as well as carbon and water cycles and associated biogeochemical feedbacks [4]. Among phenological metrics, the start of season (SOS) marks the onset of spring greening and is particularly critical for large-scale monitoring [5]. Remote sensing techniques enable efficient extraction of SOS from continuous time series of vegetation indices and overcome the spatial limitations of ground-based observations [5,6]. With the advent of long-term satellite datasets such as MODIS, phenological research has gradually expanded from local case studies to global and continental-scale assessments, substantially improving our understanding of vegetation response patterns across heterogeneous landscapes [7]. Over recent decades, changes in spring SOS across the Northern Hemisphere have been widely documented [8,9,10]. An earlier SOS often lengthens the growing season, alters seasonal-scale photosynthetic carbon uptake, and may affect ecosystem carbon sink capacity [11,12]. Meanwhile, advanced green-up does not always produce positive effects: in some regions, late frost events following earlier green-up can damage newly emerged tissues and reduce productivity, implying that SOS shifts may entail ecological risks [13,14]. Importantly, phenological trends are not governed by temperature alone; processes including chilling accumulation, soil moisture availability, frost frequency, and radiation conditions can jointly determine the timing of leaf emergence and exert opposite regulatory effects under different environmental contexts [4,15,16,17,18]. These patterns are broadly consistent with warming, highlighting the triggering role of thermal conditions [18,19,20,21]. In parallel, under arid–semiarid and other water-limited conditions, the regulatory role of precipitation and water availability in spring phenology often intensifies; studies in temperate grasslands, mountainous drylands, and controlled experiments consistently demonstrate that changes in water supply can markedly alter phenological processes [19,20,21,22]. Consequently, SOS responses to climate frequently exhibit nonlinear, threshold-dependent, or inter-seasonally coupled characteristics, making it difficult for correlation analysis or linear regression alone to adequately capture the underlying response structure [18,23,24].

Urban environments provide a unique context for investigating phenological phenomena because cities can substantially modify local microclimates [25,26]. Urban surfaces store heat, reduce evaporation and transpiration, and alter surface albedo, thereby leading to persistent warming and a reduced diurnal temperature range [25,26,27,28]. Remote-sensing analyses indicate that vegetation in urban environments often exhibits earlier timing of early-warning signals than in rural environments, particularly in areas with high impervious surface density and pronounced urban thermal effects [29]. The heterogeneity of urban patches further increases system complexity. Vegetation in parks, along streets, within residential neighborhoods, or adjacent to high-rise buildings may differ in energy balance, soil moisture conditions, and management regimes, thereby giving rise to distinct phenological patterns [30,31]. In this study, “urban phenology” is interpreted as the phenology of vegetated patches within urban administrative boundaries rather than a city-wide uniform signal, because vegetation is spatially fragmented and embedded in heterogeneous impervious matrices in cities. Therefore, a pixel-resolved analysis is necessary to preserve within-city heterogeneity and to avoid over-generalizing from city-mean metrics.

Therefore, even within the same city, vegetation phenology can exhibit pronounced spatial heterogeneity [30,31]. Despite substantial progress, current research remains constrained by several key limitations that impede the detection of climatic thresholds in urban phenology [32]. First, many analyses focus on a single city or a small subset of large cities, which restricts comparisons across different climatic settings; consequently, it remains unclear whether the reported responses are broadly generalizable or region-specific [32,33,34]. Second, prior work has often relied on linear or binary approaches to relate the start of season (SOS) to climate; while these methods can capture overall trends, they are not well suited to revealing nonlinearities, plateau phases, or abrupt thresholds [29,31,32]. Third, most studies estimate sensitivity slopes rather than identifying the magnitude of climatic change required to trigger pronounced shifts in SOS; for planning and adaptation, actionable thresholds are generally more desirable than correlations [35,36]. Accordingly, addressing these gaps is essential for robust detection of climate thresholds in urban phenology [32,35,36].

A city-scale perspective can provide information that is not directly available from broad-scale analyses, because urban vegetation is spatially fragmented and strongly shaped by local microclimate and management heterogeneity [33,34]. By using standardized grid units across many cities and focusing on within-city pixel heterogeneity, it becomes possible to test whether nonlinear response intervals are consistent across climate zones and whether threshold behavior differs among urban contexts under comparable statistical conventions [35,36,37,38,39].

Recent advances in machine learning and explainable artificial intelligence (AI) offer powerful solutions to these challenges [37,38]. Gradient-boosted decision-tree models such as XGBoost are more effective than traditional approaches for handling nonlinear relationships, high-dimensional data, and interaction effects [39,40]. However, machine-learning models are often criticized as black boxes, which limits their interpretability [41]. The emergence of SHAP (Shapley Additive Explanations) enables transparent and quantitative interpretation of model outputs by decomposing predictions into the additive contributions of individual variables [42,43]. By integrating XGBoost with SHAP-based analyses, researchers can both rank key climatic drivers and identify the climatic ranges over which drivers influence the direction or magnitude of phenological change—capabilities that are particularly well suited for detecting climatic thresholds [42,43,44]. Nevertheless, to our knowledge, multi-city and city-comparable applications of explainable machine learning to urban vegetation phenology that rely on consistent remote-sensing products and standardized spatial units remain relatively scarce. China’s provincial capitals and municipalities provide a useful multi-city testbed because they span diverse climatic regimes and urbanization contexts [32]. Importantly, we do not assume uniform vegetation cover among cities; instead, differences in vegetation fraction and composition are treated as part of the urban context that the pixel-level framework is designed to characterize and compare. Understanding how urban vegetation responds to climate not only advances theory on ecological seasonality but also provides practical guidance for greening configuration, species selection, irrigation scheduling, and climate-adaptation planning [27,30].

Therefore, this study selected 31 provincial-level administrative centers in mainland China (provincial capitals and municipalities) as the study objects. As regional cores of population, economic activity, and policy implementation, these cities are highly representative and well suited for evaluating urban ecosystem responses to climate change. Using a 500 m regular grid as the standardized analytical unit, we integrated the spring start-of-season (SOS) time series (2001–2023) derived from MODIS MCD12Q2 (Greenup_1) with TerraClimate meteorological data and introduced an XGBoost–SHAP framework. The interannual rates of change in temperature and precipitation (tem_slope and pre_slope) were used as core climatic drivers to quantify, at the urban grid scale, their marginal contributions to differences in SOS trends and to characterize potential piecewise nonlinear response patterns. The specific objectives were to: (1) construct and extract a long-term, temporally stable spring SOS time series within cities; and (2) under consistent grid resolution and statistical conventions, focus on the key driving roles of temperature and precipitation changes to identify and quantify city-scale climatic thresholds and sensitive intervals, and further compare their spatial heterogeneity across different climatic regions and urban environmental contexts. By adopting a threshold-oriented analytical strategy rather than merely estimating linear sensitivities, this study aims to delineate the critical magnitudes of climatic change that trigger advances or delays in spring green-up, and to conduct cross-city comparisons along geographic and climatic gradients to test the spatial consistency of threshold behavior in urban vegetation phenology and to explore potential mechanisms underlying observed differences.

2. Materials and Methods

2.1. Study Area

This study investigated 31 provincial-level administrative centers in mainland China, including 27 provincial capitals (of which five are capitals of autonomous regions) and four municipalities directly under the central government. These cities concentrate population, economic activities, and policy implementation, and thus provide a representative environmental context for regional urban ecosystem processes [29]. The selected cities span a broad geographic extent (20.0–45.8°N, 87.6–126.6°E). According to the Chinese national standard GB/T 17297-1998 (Climatic regionalization and the names and codes of climatic zones in China), these 31 cities were classified into seven climatic zones: the mid-temperate, warm-temperate, northern subtropical, middle subtropical, southern subtropical, marginal tropical, and plateau climatic zones (Figure 1) [45]. The number of cities in each zone is summarized in

Table 1. Research sites classified by climate zone. Cities are ordered according to the GB/T 17297-1998 Climate Zone Code (in increasing order).

City	Province-Level Unit	Climate Zone Code	Climate Zone Name
Harbin	Heilongjiang Province	12	Middle temperate zone
Shenyang	Liaoning Province	12	Middle temperate zone
Changchun	Jilin Province	12	Middle temperate zone
Yinchuan	Ningxia Hui Autonomous Region	12	Middle temperate zone
Urumqi	Xinjiang Uygur Autonomous Region	12	Middle temperate zone
Lanzhou	Gansu Province	12	Middle temperate zone
Hohhot	Inner Mongolia Autonomous Region	12	Middle temperate zone
Beijing	Beijing Municipality	13	Warm temperate zone
Jinan	Shandong Province	13	Warm temperate zone
Shijiazhuang	Hebei Province	13	Warm temperate zone
Taiyuan	Shanxi Province	13	Warm temperate zone
Tianjin	Tianjin Municipality	13	Warm temperate zone
Xi’an	Shaanxi Province	13	Warm temperate zone
Zhengzhou	Henan Province	13	Warm temperate zone
Hangzhou	Zhejiang Province	21	Northern subtropical zone
Hefei	Anhui Province	21	Northern subtropical zone
Shanghai	Shanghai Municipality	21	Northern subtropical zone
Changsha	Hunan Province	21	Northern subtropical zone
Nanjing	Jiangsu Province	21	Northern subtropical zone
Wuhan	Hubei Province	21	Northern subtropical zone
Chengdu	Sichuan Province	22	Middle subtropical zone
Chongqing	Chongqing Municipality	22	Middle subtropical zone
Fuzhou	Fujian Province	22	Middle subtropical zone
Guiyang	Guizhou Province	22	Middle subtropical zone
Nanchang	Jiangxi Province	22	Middle subtropical zone
Kunming	Yunnan Province	22	Middle subtropical zone
Nanning	Guangxi Zhuang Autonomous Region	23	Southern subtropical zone
Guangzhou	Guangdong Province	23	Southern subtropical zone
Haikou	Hainan Province	31	Marginal tropical zone
Xining	Qinghai Province	43	Plateau climate zone
Lhasa	Tibet Autonomous Region	44	Plateau climate zone

. Urban boundary data were obtained from the 2018 Global Urban Boundary (GUB) dataset. GUB is a vector urban boundary product, with boundaries delineated based on the Global Artificial Impervious Area (GAIA) product [46]. In ArcMap, the GUB layer was overlaid with the administrative division layer of the 31 cities, and the built-up polygons belonging to each city were extracted to define the urban extent used in this study. Therefore, the urban extent refers to the GUB-based built-up area within each city rather than the full administrative boundary. Subsequently, a regular 500 m grid was overlaid on each city polygon, and grid cells located within the city were used as standardized spatial analysis units. For cartographic readability, the cities shown in Figure 1 are represented by point symbols corresponding to the geometric centroids of the extracted urban polygons, rather than the actual urban boundaries used in the spatial analyses. For each grid cell, the start of season (SOS) and its associated climatic drivers were extracted and subsequently used for model fitting and cross-city comparisons.

2.2. Datasets and Processing

2.2.1. Urban Boundaries and Definition of Analysis Units

All spatial data preprocessing was conducted on the Google Earth Engine (GEE) cloud platform. GEE integrates a global, multi-source, multi-temporal archive of remote-sensing and geospatial datasets and provides cloud-based parallel computing capabilities for long time-series and large-area analyses [47]. For the 31 cities, the urban extent was defined as the GUB built-up polygons extracted within the corresponding city administrative boundaries, rather than the full administrative boundary polygons themselves.

Within each urban extent, a regular 500 m × 500 m grid was generated in GEE. The grid resolution was selected to match the native 500 m spatial resolution of the MODIS MCD12Q2 product, thereby ensuring consistency between the observations and the spatial analysis units. Grid cells whose centroids were located within the urban area were retained and used as standardized spatial units for subsequent data extraction and statistical analyses.

2.2.2. Vegetation Phenology Data

The MODIS Land Cover Dynamics product MCD12Q2 (Collection 6.1), distributed by the NASA Land Processes Distributed Active Archive Center (LP DAAC), USGS Earth Resources Observation and Science (EROS) Center, Sioux Falls, South Dakota, USA, was used to characterize vegetation growth cycles. This product provides global land-surface phenological metrics at 500 m spatial resolution and annual temporal resolution from 2001 onward [48,49]. MCD12Q2 derives phenological transition dates from time series of the two-band Enhanced Vegetation Index (EVI2) calculated from Nadir BRDF-Adjusted Reflectance (NBAR). For each product year, up to two vegetation growth cycles (cycle 1 and cycle 2) are characterized. For each cycle, the product reports Greenup (EVI2 first crossing 15% of the segment amplitude), MidGreenup (first crossing 50%), Maturity (first crossing 90%), Peak (segment maximum), Senescence (EVI2 last crossing 90%), MidGreendown (last crossing 50%), and Dormancy (last crossing 15%), together with ancillary metrics including EVI2 minimum, seasonal amplitude (maximum minus minimum), and the integrated EVI2 area (sum of daily interpolated EVI2 from Greenup to Dormancy), as well as overall and detailed quality-assurance (QA) layers [48,49]. The product has been widely used for land-surface phenology monitoring and for examining vegetation responses to climate variability, and its performance has also been systematically evaluated against phenology algorithms such as TIMESAT [50,51].

In this study, annual MCD12Q2 images from 2001 to 2023 were used, and the Greenup_1 band of the first growth cycle was adopted as the indicator of the spring start of season (SOS). To ensure data reliability, we first applied the overall QA and cycle-specific QA layers to mask low-quality pixels affected by cloud contamination, snow cover, or retrieval failures [49]. We then further filtered SOS dates using a range-based criterion, consistent with recent studies on urban spring ecology and with remote-sensing phenology studies that improve time-series robustness by excluding unrealistically early or late pixels [52,53]. Specifically, pixels with SOS earlier than day of year (DOY) 30 or later than DOY 150 were treated as outliers and discarded, thereby retaining only SOS values within a climatologically plausible spring window for the study region [54]. The adopted DOY 30–150 range was used as a conservative plausibility filter for built-up urban vegetation within the study cities, rather than as a universal phenological range for all natural ecosystems across the corresponding climatic zones.

After quality control, annual SOS images were clipped to the urban extent of each city and overlaid with the 500 m analysis grid. For each grid cell and each year, SOS values were extracted from the corresponding MCD12Q2 pixels, yielding a 23-year (2001–2023) SOS time series for each 500 m unit within each city. Grid-cell-year SOS records flagged by QA screening or removed by the DOY-based plausibility filter were treated as missing observations and excluded from subsequent analyses; no temporal gap filling or imputation was applied. In addition, SOS was not stratified by vegetation functional type within cities; therefore, the reported SOS represents an integrated urban vegetation phenology signal at the MODIS 500 m scale.

2.2.3. Climate Data

Climatic drivers were derived from the TerraClimate dataset. TerraClimate provides monthly climate and climatic water-balance variables on a geographic latitude–longitude grid at 1/24° × 1/24° spatial resolution (approximately 4 km, with the exact ground distance varying with latitude) [55]. TerraClimate was generated by integrating high-resolution WorldClim climatological layers with time-varying fields from the Climatic Research Unit Time-Series dataset (CRU TS, version 4.0) and the Japanese 55-year Reanalysis (JRA-55) using a climatically aided interpolation framework, and it has been widely applied in ecological and hydrological studies from regional to global scales [55,56].

From TerraClimate, we selected monthly climate variables for 2001–2023, including the near-surface air temperature variables tmmx and tmmn, which represent monthly mean maximum and minimum temperature, respectively, as well as the precipitation variable pr, which represents monthly precipitation accumulation (mm). For each calendar year, we aggregated the monthly data to construct annual-scale climatic indicators. Specifically, annual mean air temperature (tem) was calculated as the average of annual mean maximum temperature (T_max) and annual mean minimum temperature (T_min), while annual precipitation (pre) was calculated as the sum of the 12 monthly precipitation accumulations within that year.

$$\mathit{tem}={\displaystyle \frac{T_{max}+T_{min}}{2}}$$

(1)

To spatially align TerraClimate with the 500 m urban analysis grid, annual climate variables were extracted to each 500 m grid cell in Google Earth Engine using a region-based mean reducer at a nominal scale of 500 m. The climate value assigned to the i-th grid cell was calculated as

$$C_i={\displaystyle \frac{\sum _{k=1}^{m_i}{w}_{ik}{c}_{ik}}{\sum _{k=1}^{m_i}{w}_{ik}}}$$

(2)

where C_i is the climate value assigned to the i-th 500 m grid cell, c_ik is the climate value of the k-th pixel included in the reduction for that grid cell, w_ik is the corresponding pixel weight, and m_i is the number of pixels contributing to the reduction. This formulation summarizes the weighted regional mean used to assign TerraClimate information to each 500 m urban grid cell in the GEE workflow. No additional downscaling method (e.g., bilinear interpolation, kriging, or inverse distance weighting) was applied. Because the requested analysis scale (500 m) is finer than the native TerraClimate resolution (approximately 4 km), and no alternative resampling method was explicitly specified, any required reprojection in Google Earth Engine followed the default nearest-neighbor resampling behavior. Note that the climate predictors retain their effective spatial support at the native TerraClimate grid scale; values were assigned to the 500 m phenology grids for spatial alignment and modeling convenience and should therefore be interpreted as grid-scale climatic forcing rather than true 500 m microclimate observations. The primary datasets used in this study are summarized in

Table 2. Description of Data Sources.

Data Type	Data Product	Native Resolution	Time Range	Source
Urban boundary	Global Urban Boundary	Vector (no native pixel size; boundary delineation derived from 30 m GAIA impervious surface; effective mapping precision approximately 30 m)	2018	http://data.ess.tsinghua.edu.cn (accessed on 17 March 2026)
Spring phenology (SOS)	MCD12Q2 (Collection 6.1)	500 m	2001–2023	https://www.earthdata.nasa.gov/data/catalog/lpcloud-mcd12q2-061 (accessed on 17 March 2026)
Climate variables	TerraClimate	1/24° (approximately 4 km)	2001–2023	https://www.climatologylab.org/terraclimate.html (accessed on 17 March 2026)

.

2.2.4. Construction of Annual Panel Datasets for Temperature and Precipitation

After independent preprocessing and spatial harmonization of the phenological and climatic datasets, we constructed an annual SOS–climate panel dataset using the 500 m regular grid as the unified analysis unit. Specifically, for each 500 m grid cell within each city, one observation was compiled for each year from 2001 to 2023, including SOS and the corresponding annual climate indicators (tem and pre). This yielded a “grid cell–year” indexed data structure, in which each row represents observations for a given grid cell in a given year, and each column corresponds to a phenological or climatic attribute (SOS, tem, and pre), thereby enabling comparable representations within and across cities at a consistent spatial granularity.

This panel dataset provides a consistent data basis for subsequent trend characterization and climate attribution. On the one hand, based on the multi-year SOS series for each grid cell, trend significance was assessed using the Mann–Kendall (MK) nonparametric trend test, and long-term change rates were quantified using the Theil–Sen median slope estimator. These methods rely on rank-based statistics and median-based estimation, require no specific distributional assumptions, and are relatively insensitive to extreme values [57,58,59,60]. On the other hand, the same data structure supports establishing linkages between SOS trend metrics and the long-term change rates of climate variables within a machine-learning framework, enabling contribution decomposition and threshold identification using the XGBoost–SHAP framework described in the following Methods subsection. By organizing data at the grid-cell scale, phenological and climatic information for all 31 cities was expressed on consistent 500 m analysis units, facilitating cross-city comparisons of SOS spatial patterns, long-term trends, and climatic sensitivity under contrasting urban environmental and climatic contexts.

2.3. Methods

2.3.1. Pixel-Scale SOS Trend Testing and Significance Classification

At the 500 m grid scale, we used SOS time series from 2001 to 2023 and jointly applied the Mann–Kendall (MK) nonparametric test and the Theil–Sen median slope estimator to characterize long-term trends in SOS and their statistical significance. The MK test is a rank-based, distribution-free approach that does not rely on the assumption of normality; it is robust to outliers and to limited missing observations and has been widely used for time-series analyses in hydrology, climatology, and vegetation phenology [57,58]. The Theil–Sen median slope estimator proposed by Sen, hereafter referred to as the Sen slope, estimates the overall trend as the median of slopes computed from all pairwise combinations of observations. Rooted in Theil’s rank-invariant regression concept, it is among the most commonly used robust slope estimators for trend analysis in environmental time series [59,60].

For an SOS time series of length n, where SOS_i denotes the SOS value in year i, the Mann–Kendall test statistic S is defined as:

$$S=\sum _{i=1}^{n-1}\sum _{j=i+1}^n\mathrm{s}\mathrm{g}\mathrm{n}\left({\mathrm{S}\mathrm{O}\mathrm{S}}_j-{\mathrm{S}\mathrm{O}\mathrm{S}}_i\right)$$

(3)

where the sign function sgn(${\mathrm{S}\mathrm{O}\mathrm{S}}_j-{\mathrm{S}\mathrm{O}\mathrm{S}}_i$) is given by:

$$\begin{array}{r}\mathrm{s}\mathrm{g}\mathrm{n}\left({\mathrm{S}\mathrm{O}\mathrm{S}}_j-{\mathrm{S}\mathrm{O}\mathrm{S}}_i\right)=\{\begin{array}{ll}1& {\mathrm{S}\mathrm{O}\mathrm{S}}_j-{\mathrm{S}\mathrm{O}\mathrm{S}}_i>0\\ 0& {\mathrm{S}\mathrm{O}\mathrm{S}}_j-{\mathrm{S}\mathrm{O}\mathrm{S}}_i=0\\ -1& {\mathrm{S}\mathrm{O}\mathrm{S}}_j-{\mathrm{S}\mathrm{O}\mathrm{S}}_i<0\end{array}\end{array}$$

(4)

This formulation evaluates the direction of pairwise differences across all year pairs and aggregates them to quantify the monotonic trend.

The Mann–Kendall (MK) test adopts the null hypothesis H₀ that no monotonic trend exists in the time series. Under H₀, the expected value of the statistic S is zero, and its variance, denoted as Var(S), can be computed following the formulations derived by Mann [57] and Kendall [58]. The standardized test statistic Z is then constructed as:

$$Z=\{\begin{array}{ll}{\displaystyle \frac{S-1}{\sqrt{\mathrm{V}\mathrm{a}\mathrm{r}\left(S\right)}}}& S>0\\ 0& S=0\\ {\displaystyle \frac{S+1}{\sqrt{\mathrm{V}\mathrm{a}\mathrm{r}\left(S\right)}}}& S<0\end{array}$$

(5)

Trend significance was determined based on the two-sided p-value associated with Z. Given that environmental time series may exhibit autocorrelation, which can affect the ability of the MK test to correctly assess trend significance, we interpreted the significance results with reference to MK modification approaches and related discussions developed for autocorrelated series [61,62].

After completing the MK significance test, we estimated the long-term rate of change in SOS using the Sen slope. This method computes slopes for all year pairs (i,j) satisfying j > i and takes their median as the overall trend slope, where tᵢ and tⱼ denote the observation years corresponding to SOSᵢ and SOSⱼ, respectively:

$$\beta =\mathrm{m}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{n}\left({\displaystyle \frac{{\mathrm{S}\mathrm{O}\mathrm{S}}_j-{\mathrm{S}\mathrm{O}\mathrm{S}}_i}{t_j-t_i}}\right)\hspace{1em}j>i$$

(6)

This is the Theil–Sen slope estimator [59,60]. In this study, β is expressed in d·year⁻¹: β < 0 indicates that SOS advances by |β| days per year on average, whereas β > 0 indicates that SOS is delayed by β days per year on average.

Based on the MK-test p-value and the sign of the Sen slope, each 500 m pixel was classified into three categories. Specifically, pixels with p < 0.05 were identified as exhibiting a significant change, those with 0.05 ≤ p < 0.10 were considered marginally significant, and those with p ≥ 0.10 were treated as non-significant. Pixels in the significant and marginally significant groups were further categorized as advancing or delaying according to the sign of β. This combined MK–Sen approach has been widely used in remote-sensing phenology studies to identify pixel-scale long-term phenological trends and regional differences [63,64].

In the descriptive analyses, to comprehensively characterize the overall distribution of spring SOS within each city, we used all pixels that passed quality control, including pixels with non-significant trends, and accordingly produced city-scale SOS distribution plots and pixel-count summaries. In contrast, for subsequent analyses focusing on trend signals and their climatic drivers, only pixels with p < 0.10 were retained as valid samples. For these pixels, the Sen slope was used as the response variable for constructing SOS trend histograms and for XGBoost–SHAP modeling and climatic threshold identification.

2.3.2. Estimation of Long-Term Change Rates in Climatic Drivers

To characterize long-term changes in the climatic background using a statistical framework consistent with SOS trend estimation, we constructed annual climate series for 2001–2023 from TerraClimate monthly data and estimated long-term trend slopes of temperature and precipitation at the 500 m grid scale. Here, “change rate” refers to the Theil–Sen trend slope of the annual series (units per year), rather than year-to-year changes between consecutive years. First, for each year, we calculated the arithmetic mean of the annual mean maximum temperature (tmmx) and the annual mean minimum temperature (tmmn) and then averaged these two quantities to obtain annual mean air temperature (tem). For precipitation, because TerraClimate pr is a monthly accumulated variable, annual precipitation (pre) was defined as the sum of the 12 monthly pr values, yielding annual total precipitation for each year. Next, annual climate variables were assigned to each 500 m grid cell using the region-based mean extraction procedure implemented in Google Earth Engine, consistent with the spatial harmonization described in Section 2.2.3. Finally, the Mann–Kendall test was applied to assess the significance of monotonic trends, and the Theil–Sen median slope estimator was used to quantify long-term change rates, producing tem_slope (°C·year⁻¹) and pre_slope (mm·year⁻¹, annual total precipitation basis). These two rate-of-change indicators were estimated under the same nonparametric trend framework as the Sen slope of SOS, thereby facilitating comparable interpretation in subsequent XGBoost–SHAP attribution and threshold analyses. Only 500 m grid cells with valid SOS values after QA screening and plausibility filtering were retained for subsequent trend estimation and attribution analyses.

2.3.3. XGBoost Model Development and Evaluation

To identify the nonlinear response structure of SOS trends to warming and precipitation changes and to provide a regression basis for subsequent threshold detection, we adopted the XGBoost regression model to analyze the nonlinear impact of climate factors on spring phenology. XGBoost is an efficient and scalable implementation of gradient-boosted decision trees that can fit complex nonlinear relationships without prespecifying a functional form, and it has been widely applied in remote-sensing and ecological applications [39,40,65]. Pixels identified as significant or marginally significant by the MK screening were used as modeling samples. The Sen slope of SOS was used as the response variable, and tem_slope and pre_slope were used as input features. Here, tem_slope represents the long-term rate of change in annual mean air temperature, whereas pre_slope represents the long-term rate of change in annual total precipitation. This setup allowed us to focus on the marginal contributions of the two climate-change rate metrics to SOS trends and their potentially piecewise response behaviors. The core model was intentionally restricted to tem_slope and pre_slope to preserve a parsimonious and directly interpretable city-comparable framework centered on the two primary climate-change-rate dimensions addressed in this study, whereas the extended model including vpd_slope was used as a robustness check on threshold consistency rather than as the main analytical basis. Notably, the model target is the long-term SOS trend slope (DOY·year⁻¹), rather than SOS timing (DOY) or detrended anomalies, so that the inferred response regimes reflect climate-change-rate effects on phenological change rates rather than baseline geographic gradients. We additionally assessed collinearity between tem_slope and pre_slope using Pearson correlation and the variance inflation factor (VIF), defined as

$${\mathrm{V}\mathrm{I}\mathrm{F}}_j={\displaystyle \frac{1}{1-R_j^2}}$$

(7)

where R_j² is the coefficient of determination obtained by regressing predictor x_j on the remaining predictor [66,67]. Because the baseline model included only two predictors (tem_slope and pre_slope), R_j² reduces to r², where r is the Pearson correlation between the two predictors; thus, VIF_tem = VIF_pre = 1/(1 − r²). In our data, r = −0.702, yielding VIF = 1.97 for both predictors, which indicates that severe multicollinearity was not present in the baseline two-driver setting.

Model training followed a train–test split and incorporated cross-validation to tune key hyperparameters, reducing overfitting risk and improving generalization performance. The train–test split, 5-fold cross-validation, and repeated 80/20 stability tests were performed randomly across pixels rather than using a spatially blocked design. Model performance was evaluated on the test set using the coefficient of determination and error metrics [40,68]. In addition, given that the XGBoost–SHAP approach has been widely used in ecological and remote-sensing attribution studies to quantify driver contributions and characterize nonlinear response structures, the modeling configuration and evaluation workflow in this study were aligned with established practice in the relevant literature [69,70].

Hyperparameters of the XGBoost regressor were tuned using GridSearchCV with 5-fold cross-validation and R² as the scoring metric. The search space covered tree complexity, learning rate, subsampling, and regularization parameters (n_estimators, max_depth, learning_rate, subsample, colsample_bytree, reg_lambda, and gamma). For each feature set (main model: tem_slope + pre_slope; extended model: tem_slope + pre_slope + vpd_slope), the final model adopted the best-ranked hyperparameter configuration according to the cross-validated test score returned by GridSearchCV, as reported in Table S1.

To quantify robustness to random sampling, we conducted a stability experiment by repeating an 80/20 train–test split 30 times with different random seeds and summarized performance as mean ± SD of R² and RMSE, as shown in Figure S1. In addition, to evaluate whether the identified turning points were sensitive to the driver set, we repeated the same tuning and stability procedure after adding vpd_slope; the resulting turning points for tem_slope and pre_slope remain qualitatively consistent between the main and extended models, as illustrated in Figure S2.

Missing values in predictors were retained and handled internally by the tree-based XGBoost algorithm, while samples with invalid values in the response variable (SOS Sen slope β) were excluded prior to model fitting.

2.3.4. SHAP Interpretation Framework and Threshold Extraction

To enhance the interpretability of machine-learning attribution results and to identify key tipping intervals in climatic responses, we applied SHapley Additive exPlanations (SHAP) to interpret the XGBoost model. SHAP is grounded in the Shapley value concept from cooperative game theory and decomposes a model prediction into additive contributions from individual input features, thereby characterizing, at the sample level, the direction and magnitude of the marginal effect of feature values on the predicted outcome [42]. For tree-based models, TreeSHAP provides an efficient approach for computing feature-level marginal contributions and is well suited to gradient-boosted tree frameworks such as XGBoost [71]. We compute SHAP values using TreeSHAP, where the base value corresponds to the expected model output under the training distribution; our interpretation focuses on how feature contributions shift the prediction relative to that expected output. It should be particularly noted that the SHAP value represents the deviation of the model’s expected output value. Mathematically speaking, the predicted SOS slope of a given pixel is equal to the baseline value (the average predicted value of the entire dataset) plus the sum of the SHAP values of all features. In the context of this study, the baseline value reflects the regional background trend of SOS changes. Therefore, the reported SHAP values quantify the specific marginal contribution of local temperature or precipitation changes to the deviation of the phenological trend from the regional average.

In implementation, SHAP values were computed for grid-cell samples separately within each climatic zone, and SHAP dependence relationships were used to characterize the nonlinear response structure of the SOS Sen slope β along gradients of climatic change rates. Because the model included only two input features, tem_slope and pre_slope, these features served as direct interpretive axes for diagnosing how the marginal contribution varies across the feature-value range for each climate-change rate metric. Accordingly, SHAP dependence plots were generated for tem_slope and pre_slope, and their overall functional forms were summarized using locally smoothed fits.

Climatic threshold identification was based on the piecewise characteristics of SHAP dependence relationships. For each city and each climatic driver (tem_slope or pre_slope), we fitted a local weighted regression scatter smoothing curve (LOWESS; fraction = 0.4) to the SHAP–feature scatter plot to capture nonlinear trends while suppressing random noise. The climatic threshold was defined as the zero-crossing point, i.e., the intersection between the LOWESS-smoothed curve and the SHAP = 0 baseline. Within our framework, this point indicates the driver value at which the marginal contribution to the SOS trend changes sign, reflecting a transition between advancing and delaying effects [47]. Within this framework, the threshold should be interpreted as an empirical response-transition point in the modeled SHAP dependence relationship, rather than as a directly observed physiological tipping point.

To reduce spurious thresholds caused by sparse observations at the distribution tails or by small-amplitude oscillations of the smoothed curve around zero, we applied a reproducible support-based screening procedure. Candidate zero-crossings were first merged when they formed dense sequences along the x-axis, yielding representative turning-point candidates. For each candidate, local data support was quantified as the number of samples within a symmetric neighborhood window around the candidate value; candidates with insufficient support (local sample size < 50) were discarded, with a fallback to the unfiltered candidate set if no candidates remained. Candidates were then ranked using a support-weighted score that combines the crossing strength of the LOWESS curve and local sample density. For each city, we reported at most two thresholds per driver: the dominant threshold is always retained, whereas a secondary threshold is reported only when it is comparably supported (score ratio ≥ 0.85) and sufficiently separated from the dominant threshold. This strategy avoids tail-driven thresholds and ensures that reported thresholds represent robust turning points of the SHAP dependence pattern rather than artifacts of sparse samples.

To further assess uncertainty associated with curve fitting and sampling variability, we selected one representative city from each climatic zone and performed a bootstrap validation (n = 1000 iterations) to derive the threshold distribution and confidence interval, as shown in Supplementary Figures S3 and S4. The bootstrap medians were consistent with the zero-crossing thresholds reported in the main analysis, supporting the statistical robustness of the extracted thresholds. Threshold locations and associated sensitive intervals were then summarized and compared across climatic zones and cities to reveal zonal differences and nonlinear response boundaries of urban spring phenology to sustained warming and precipitation changes. Here, “sensitive intervals” refer to the driver-value ranges delimited by the retained threshold(s), within which the LOWESS-smoothed SHAP curve shows a relatively coherent response direction; they are therefore defined operationally by the threshold-delimited piecewise response structure rather than by a separate numerical cutoff in slope magnitude or SHAP amplitude.

To improve the transparency of the analytical procedure, the overall workflow of the XGBoost–SHAP framework from input variables to climatic threshold identification is summarized in Figure 2.

3. Results

3.1. Spatial Pattern of Mean SOS Across Provincial Capitals and Municipalities in Mainland China

Based on the 500 m gridded SOS dataset for the 31 provincial capitals and municipalities over 2001–2023, we first compared the spatial patterns of the multi-year mean spring green-up timing across cities (Figure 3). After QA screening and DOY-based plausibility filtering (SOS within 30–150), the multi-year mean SOS exhibits a pronounced south–north and low–high elevation gradient, with earlier SOS in southern cities and later SOS in northern and high-elevation cities. City-level mean SOS ranges from 73.55 DOY in Xi’an to 138.11 DOY in Hohhot, yielding a span of 64.56 DOY. This wide range reflects the joint influence of large-scale thermal gradients and elevation-related constraints on urban spring phenology.

When stratified by climatic zones, southern subtropical and marginal tropical cities generally show early green-up, with city-mean SOS mostly in the early-to-mid 80 s DOY (e.g., Nanning: 82.29 DOY; Guangzhou: 83.97 DOY; Haikou: 83.56 DOY). Northern subtropical cities display an overlapping but wider window (approximately 77–88 DOY; e.g., Hefei: 77.21 DOY; Nanjing: 77.98 DOY; Hangzhou: 82.63 DOY; Changsha: 88.00 DOY), indicating that city-scale SOS timing can overlap substantially across adjacent climatic zones. The middle subtropical zone exhibits the largest inter-city dispersion, spanning from relatively early timing in Chongqing (81.61 DOY) and Guiyang (82.16 DOY) to much later timing in Nanchang (100.73 DOY) and Kunming (100.96 DOY). This zone-wise comparison highlights that, although a large-scale gradient exists, city-level mean SOS patterns are not strictly monotonic across zones and show pronounced within-zone heterogeneity, particularly in the middle subtropical and warm-temperate regions.

Compared with the southern and northern subtropical cities, several warm-temperate and most mid-temperate cities exhibit substantially later green-up, while warm-temperate cities show strong within-zone dispersion. Warm-temperate cities range from relatively early timing in Xi’an (73.55 DOY) and Zhengzhou (79.85 DOY) to much later timing in Beijing (109.76 DOY), Tianjin (124.21 DOY), and Taiyuan (130.24 DOY). Mid-temperate cities such as Harbin (136.71 DOY), Changchun (136.03 DOY), and Hohhot (138.11 DOY) show the latest SOS among the study sites. Cities in plateau and semi-arid settings also show delayed green-up, including Xining (127.91 DOY), Lhasa (125.56 DOY), Lanzhou (124.01 DOY), and Urumqi (118.01 DOY), suggesting that elevation and regional thermal limitations can markedly delay spring phenology beyond what would be expected from latitude alone. Detailed city-level statistics on multi-year mean SOS values and their dispersion are reported in Table S2. For a more compact comparison,

Table 3. Summary statistics of multi-year mean SOS and SOS trends by climate zone.

Climate Zone Code (A–G)	Climate Zone	n (Cities)	Mean SOS (DOY) ± SD	Mean SOS Trend (d·Year−1) ± SD	Significant Grids (%) (p < 0.05)
A	Plateau climate zone	2	126.73 ± 1.66	0.04 ± 0.46	30.79
B	Southern subtropical zone	2	83.13 ± 1.19	−0.28 ± 0.29	5.84
C	Middle subtropical zone	6	89.92 ± 8.78	−1.75 ± 0.74	28.24
D	Northern subtropical zone	6	81.20 ± 3.92	−0.67 ± 0.65	27.79
E	Middle temperate zone	7	130.72 ± 7.45	−1.04 ± 0.30	57.74
F	Warm temperate zone	7	100.47 ± 21.62	−0.44 ± 1.56	62.25
G	Marginal tropical zone	1	83.56 ± —	0.12 ± —	6.16

Notes: ① Mean ± SD for SOS and SOS trend are computed across city-level averages within each climatic zone (n = number of cities). ② SD is not applicable for zones with n = 1 city (shown as ‘—’). ③ Significant grids (%) is computed as the proportion of valid 500 m urban grids with Mann–Kendall p < 0.05 within each climatic zone (pooled across all cities in the zone). ④ SOS trend is Sen’s slope (d·year⁻¹); negative values indicate advancing SOS and positive values indicate delaying SOS.

summarizes SOS and SOS-trend statistics by climatic zone, including mean ± SD across city-level averages and the proportion of significant trends (MK p < 0.05).

Overall, the multi-year mean SOS across China’s provincial capitals and municipalities exhibits a clear large-scale gradient, with early green-up in southern low-latitude cities, intermediate timing in the northern subtropical belt, and late green-up in northern, arid, and high-elevation cities. At the same time, notable within-zone dispersion (especially in the middle subtropical and warm-temperate zones) highlights the role of geographic setting and intra-urban heterogeneity in shaping city-mean SOS. This structured pattern provides a basis for subsequent climate-zone–based analyses of SOS distributions and for investigating the climatic drivers of spring phenological change across cities.

3.2. Distributional Characteristics of City-Level SOS Within Different Climatic Zones

Based on the GB/T 17297-1998 climatic regionalization, the 31 provincial capitals and municipalities were classified into seven climatic zones, namely the mid-temperate, warm-temperate, northern subtropical, middle subtropical, southern subtropical, marginal tropical, and plateau climatic zones (Figure 1; Table 1). We then compared pixel-level SOS distributions across cities within each climatic zone (Figure 4).

The results show clear differences in both the central tendency and dispersion of SOS distributions among climatic zones. In general, climatic zones at lower latitudes with more favorable thermal conditions exhibit earlier SOS distributions, whereas higher-latitude zones and the plateau climatic zone show overall later SOS distributions, revealing a pronounced climatic gradient. At the same time, noticeable inter-city differences persist within the same climatic zone, indicating that the spatial pattern of urban SOS is jointly shaped by the broader climatic background and city-scale heterogeneity.

It should be noted that the kernel density curves in Figure 4 were used to depict the overall distributional shapes of urban SOS and were generated using all pixels that passed quality control, including pixels with non-significant trends. Because cities within the same climatic zone differ substantially in urban extent and in the number of valid pixels, the corresponding pixel counts are reported below the kernel density plots to facilitate interpretation of how sample-size differences may influence the observed distributional characteristics.

3.3. Pixel-Scale and City-Scale SOS Trend Patterns and Spatial Heterogeneity

Based on the Mann–Kendall test and Theil–Sen slope estimation applied to the 2001–2023 SOS time series at the 500 m grid resolution, we identified long-term change types of spring SOS within cities. Pixels were classified by significance level into significant change (p < 0.05), marginally significant change (0.05 ≤ p < 0.10), and non-significant change (p ≥ 0.10). Among significant and marginally significant pixels, trend direction was further distinguished using the sign of the Sen slope, yielding advancing (β < 0) and delaying (β > 0) types.

Pixel-level results indicate that advancing and delaying trends commonly coexist within individual cities, highlighting pronounced spatial non-uniformity in urban spring phenological change (Figure 5). Meanwhile, the proportion of significant and marginally significant pixels relative to all valid pixels varies among cities, suggesting that the strength of trend signals and their spatial coherence differ across urban areas (Figure 5).

At the city scale, we summarized and compared the directional composition of trends among significant and marginally significant pixels (Figure 6). The results show marked inter-city differences in the proportions of β < 0 and β > 0 pixels: some cities are dominated by advancing pixels, others by delaying pixels, and some exhibit comparable shares of both, reflecting within-city coexistence of contrasting trend directions. This city-level synthesis provides directional context for subsequent analyses of climatic drivers and threshold heterogeneity.

3.4. Key Climatic Drivers and Threshold Differences Revealed by XGBoost–SHAP

After identifying pixels with statistically significant and marginally significant SOS trends, we used the Sen slope of SOS for these 500 m grid pixels as the response variable β (d·year⁻¹). Two predictors were used as input features: the interannual rate of change in temperature (tem_slope) and annual total precipitation (pre_slope). We trained an XGBoost regression model and interpreted outputs using SHAP. For a pixel at its observed feature values, the SHAP value quantifies the marginal contribution of that feature to the predicted β: positive SHAP values indicate an increased β (a stronger tendency toward SOS delay or a weakened SOS advance), whereas negative values indicate a decreased β (a stronger SOS advance or a mitigated SOS delay). We then translate SHAP dependence structures into city-scale thresholds and sensitive intervals and compare their ranges and regime structures across climatic zones for tem_slope and pre_slope. For visualization and inter-city comparability, the dependence plots for each driver adopt a shared x-axis range across all cities (Figure 7 and Figure 8). The main climatic threshold for each city is defined as the zero-crossing point of the LOWESS-smoothed curve relative to the SHAP = 0 baseline; an additional threshold is reported only when it passes strict screening criteria (sample density, relative support, and separation distance).

To characterize nonlinear responses, we generated SHAP dependence plots and applied LOWESS smoothing to the SHAP–feature scatterpoints (Figure 7 and Figure 8). For each city and each driver (tem_slope or pre_slope), the climatic threshold was primarily defined as the zero-crossing point, i.e., the intersection between the LOWESS-smoothed curve and the SHAP = 0 baseline, which indicates where the marginal contribution to the SOS trend changes sign. To avoid spurious crossings caused by small-amplitude fluctuations around zero, we applied a city-level screening criterion: only crossings associated with a clear change in curve direction and an evident contrast in SHAP levels on the two sides were retained, whereas weak oscillations near zero were discarded. The retained turning points delineate piecewise response regimes, and we define climatic sensitive intervals as feature-value ranges where the LOWESS curve is relatively steep and SHAP values deviate clearly from zero. To facilitate cross-city comparison, all subplots in Figure 7 and Figure 8 share the same x-axis range for the same driver variable.

We first summarize tem_slope thresholds across climatic zones (Figure 7). Thresholds are positive for most cities, implying that SOS-trend responses to temperature change generally shift from near-neutral to a more temperature-sensitive regime only after warming reaches a certain rate. In the plateau climatic zone, thresholds are 0.0217–0.0501 °C·year⁻¹ (Xining: 0.0501; Lhasa: 0.0217). In the southern subtropical zone, thresholds are relatively concentrated at moderate positive values (Nanning: 0.0492; Guangzhou: 0.0664). Middle subtropical cities show the broadest spread, spanning from negative to clearly positive values (−0.0267 to 0.1000 °C·year⁻¹; e.g., Guiyang: −0.0267; Chongqing: −0.0098; Kunming: −0.0051 versus Fuzhou: 0.1000), suggesting that temperature sensitivity can emerge under slight cooling/weak warming in some mountainous subtropical settings but requires stronger warming in others. Northern subtropical thresholds fall mainly within 0.0234–0.0920 °C·year⁻¹ (e.g., Hangzhou: 0.0546; Shanghai: 0.0920). In mid-temperate cities, thresholds range from 0.0193 to 0.0801 °C·year⁻¹ (e.g., Lanzhou: 0.0193; Changchun: 0.0779; Harbin: 0.0507), indicating that sustained warming rates on the order of ~0.02–0.08 °C·year⁻¹ are typically required to produce an evidently non-neutral marginal contribution to SOS trends. Warm-temperate thresholds span −0.0233 to 0.0566 °C·year⁻¹ (e.g., Xi’an: −0.0233; Jinan: 0.0566), and the marginal tropical city Haikou shows a positive threshold at 0.0275 °C·year⁻¹.

Beyond threshold locations, the number of tem_slope turning points is predominantly one per city under the adopted screening strategy, indicating a two-regime structure in most cases: a near-neutral regime below the main threshold and a more temperature-sensitive regime above it. Only a small subset of cities exhibits an additional turning point after screening; for example, Hohhot shows two retained thresholds (0.0801 and 0.1163 °C·year⁻¹), suggesting a possible three-stage pattern where the marginal temperature effect changes again under relatively stronger warming. Overall, the dominance of a single main threshold across cities indicates that the temperature-related contribution to SOS trends is generally characterized by one major regime transition at the city scale, with secondary regime shifts being comparatively rare and restricted to cases with strong statistical support.

We then summarize pre_slope thresholds to characterize moisture-related sensitivity (Figure 8). Overall, pre_slope thresholds, expressed on an annual total precipitation basis, remain concentrated around relatively modest wetting or drying trends rather than extremely large-magnitude moisture changes. In the plateau climatic zone, thresholds are positive and small (2.2129–2.6248 mm·year⁻¹; Xining: 2.2129; Lhasa: 2.6248). In the southern subtropical zone, thresholds differ in sign between the two cities (Nanning: 4.6989 versus Guangzhou: −0.4609 mm·year⁻¹), suggesting heterogeneous moisture sensitivity even within the same climatic background. Middle subtropical cities span the widest range (−13.4286 to 14.9929 mm·year⁻¹; e.g., Kunming: −13.4286; Nanchang: −2.1743; Guiyang: 0.4572; Chengdu: 2.3373; Chongqing: 14.1318; Fuzhou: 14.9929), implying that both intensified drying and wetting can be associated with non-neutral SHAP contributions depending on local setting. Northern subtropical thresholds also straddle zero (−8.9312 to 9.4372 mm·year⁻¹; e.g., Shanghai: −8.9312; Wuhan: −2.8960; Changsha: −5.5045; Nanjing: −0.0520; Hefei: 7.6678; Hangzhou: 9.4372). Mid-temperate thresholds tend to be negative overall but still include sign heterogeneity (−8.0515 to 4.4671 mm·year⁻¹; e.g., Hohhot: −8.0515; Urumqi: −5.4774; Lanzhou: −5.4804; Yinchuan: −5.4290; Shenyang: −2.4543; Harbin: 0.4565; Changchun: 4.4671), suggesting that slight to moderate drying is more often associated with marked marginal contributions to SOS trends in these cities. Warm-temperate thresholds also concentrate near zero with both signs (−4.1455 to 4.6941 mm·year⁻¹; e.g., Taiyuan: −4.1455; Jinan: −2.2237; Zhengzhou: −1.3766; Xi’an: −0.3928; Beijing: 3.5390; Tianjin: 4.5648; Shijiazhuang: 4.6941). The marginal tropical city Haikou shows two retained thresholds (−11.3841 and −1.8239 mm·year⁻¹), indicating that sensitivity may be confined to a drying interval rather than extending monotonically across the full moisture-change spectrum. In most cities, pre_slope yields a single dominant threshold separating near-neutral and moisture-sensitive regimes, whereas multi-threshold patterns remain uncommon under the adopted screening criteria.

Finally, substantial inter-city and within-city heterogeneity is evident even within the same climatic zone, helping explain why threshold locations and counts vary. Differences among cities in topographic setting, underlying surface characteristics, and urbanization intensity can shift sensitive intervals and partition responses into multiple stages. Within cities, SHAP dependence plots often show elongated or bifurcated scatter distributions near thresholds, suggesting that response direction and magnitude to the same climatic-change rate are not fully consistent across urban sectors, consistent with pronounced within-city spatial heterogeneity.

Overall, the XGBoost–SHAP interpretation indicates that tem_slope and pre_slope link to SOS trend β through piecewise response structures, and their turning points delineate city-scale climatic thresholds and sensitive intervals. Cross-zone contrasts reflect spatial differences in thermal versus moisture constraints, while within-zone differences in threshold locations and counts highlight pronounced city-scale heterogeneity. Collectively, these thresholds and sensitive intervals form a comparable “climatic sensitive-interval” framework for interpreting regional differences in urban spring phenological change from a nonlinear, threshold-oriented perspective.

4. Discussion

4.1. Spatial Pattern of Urban Spring SOS and the Climatic Background

This study, based on MCD12Q2 and TerraClimate data for 2001–2023, provides a systematic city-scale characterization of the spatial patterns and long-term dynamics of spring SOS across 31 provincial capitals and municipalities in mainland China. The multi-year mean SOS becomes progressively later from south to north and from low to high elevations. Cities in the southern subtropical and marginal tropical zones generally exhibit early green-up (mostly in the early-to-mid 80s DOY), whereas the middle subtropical zone shows pronounced inter-city dispersion, extending from the low 80 s to around 101 DOY. In contrast, cities in the mid-temperate and plateau climate zones predominantly green up after approximately 110 DOY, forming a large-scale pattern characterized by earlier green-up in southern low-latitude cities, later green-up toward northern regions, and generally delayed green-up in plateau and high-elevation settings. Similar spatial gradients have been reported in previous studies focusing on multiple Chinese cities or urban–rural gradients; however, most of these studies either emphasized urbanization effects per se or were limited to a small set of cities [72,73,74]. In contrast, under a unified 500 m grid framework and climate-zone grouping, our analysis further resolved this gradient at the cross-level of inter-city comparisons within each climatic zone, thereby more clearly revealing spatial-structural differences in urban spring SOS.

Kernel-density–based distribution summaries grouped by climatic zone (Figure 4) show pronounced differences in both the central tendency and dispersion of SOS distributions among climatic zones. In lower-latitude zones with ample thermal conditions, green-up not only occurs earlier overall, but within-city pixel-level dispersion is also relatively small. By contrast, in thermally constrained regions such as the mid-temperate and plateau climate zones, SOS distributions are generally later and exhibit greater within-city dispersion, indicating that even under similar large-scale climatic backgrounds, factors such as topographic variability, underlying-surface structure, and local microclimates can amplify intra-urban phenological heterogeneity. This pattern is consistent with phenological evidence from the Mongolian Plateau, Xinjiang, and the north–south transition zone in China, where regional climatic constraints and local environmental conditions jointly regulate phenological spatial patterns [75,76,77].

Pixel-scale MK–Sen trend results indicate that advancing and delaying SOS trends commonly coexist within most cities, and the proportion of significant and marginally significant pixels among all valid pixels varies markedly across cities. This suggests substantial geographic differences in both the strength of phenological trend signals and their spatial coherence. When summarized at the city scale, some cities are dominated by pixels with β < 0, indicating an overall acceleration toward earlier green-up, whereas others are dominated by pixels with β > 0, indicating an overall delay. Additional cities show comparable proportions of the two types, implying that responses to climatic change are not directionally uniform across intra-urban subregions. Compared with prior studies that relied primarily on station observations or averaged urban–suburban transects [72,73,74], our grid-based characterization within cities more clearly reveals “heterogeneous responses within the same city” and provides contextual support for subsequent interpretation of spatial differences from threshold-based and nonlinear-response perspectives [75,76,77,78,79].

4.2. Climatic-Zonal Attribution of Spring SOS and Differences in Relative Constraints

At the level of climatic drivers, previous studies have consistently indicated that the dominant constraints on spring phenology exhibit pronounced spatial zonation. In energy-limited regions, spring green-up is highly sensitive to warming, and daytime high temperatures or pre-season thermal conditions can exert strong triggering effects on leaf emergence [18,80,81]. In contrast, in water-limited regions, precipitation and moisture availability play a non-negligible regulatory role in spring phenology, and some studies further suggest that the relative importance of precipitation in explaining spring phenological change may increase in certain regions and periods [19,20,21], consistent with evidence from controlled experiments showing that enhanced water supply can advance spring phenological progression [22]. Building on this understanding, we used tem_slope and pre_slope as the core driver metrics to quantify differences in marginal contributions at the urban grid scale and to discuss their relative constraint characteristics across climatic zones.

Using an XGBoost–SHAP framework including only tem_slope and pre_slope, we characterized response differences in spring SOS trends to the rate of warming and the rate of change in annual total precipitation from two complementary perspectives, namely marginal contributions and sensitive intervals. Overall, the results indicate that urban spring phenology responds more directly to the direction and rate of climatic change than to static multi-year mean climate states. The SHAP dependence relationships for tem_slope and pre_slope generally exhibit nonlinear and piecewise features, with systematic differences among climatic zones. This is consistent with synthesis assessments across the Northern Hemisphere and diverse regions of China showing strong climatic sensitivity of spring phenology to thermal and moisture conditions [75,78,79].

From a spatially zonal perspective, plateau climatic zone and mid-temperate cities show more concentrated response thresholds for tem_slope and often more clearly defined sensitive regimes, indicating that under thermally constrained backgrounds, variation in warming rate more readily triggers stage-like adjustments in SOS trends, whereas precipitation changes are more likely to act as a background modulator. This pattern aligns with findings that temperature constraints are particularly prominent for spring phenology in cold and/or arid environments such as the Mongolian Plateau and Xinjiang [75,77]. By contrast, in the southern subtropical, middle subtropical, and parts of the warm-temperate cities, pre_slope thresholds are broader and sensitive intervals are more likely to occur under sustained precipitation decreases or within value ranges associated with moisture redistribution. This suggests that, under humid–warm or relatively dry backgrounds, the marginal influence of moisture-change rates on SOS trends is not negligible and may jointly shape the direction and magnitude of urban phenological change together with warming effects [76,77]. Overall, the strengths of marginal contributions of tem_slope and pre_slope and their sensitive intervals exhibit stable and comparable structural differences among climatic zones, indicating that the primary climatic controls on urban spring SOS trends adjust with regional thermal and moisture backgrounds. This conclusion is consistent with the broader understanding that phenological responses are jointly constrained by energy and water availability [78,79]. Beyond shifts in mean climate, the change-rate metrics tem_slope and pre_slope can also be associated with altered persistence of anomalies and the frequency of extreme events, which often elicit nonlinear phenological responses. Under accelerated warming, reduced chilling accumulation and episodic late-spring cold events can partially offset thermal forcing, contributing to saturation or multi-stage responses in cold/high-elevation regions. Conversely, sustained drying trends can increase drought probability and atmospheric moisture demand, so that warm–humid cities may enter a high-sensitivity regime only after precipitation decreases exceed a certain magnitude, whereas transitional zones may exhibit two-sided sensitivity intervals because both dry and wet anomalies can reorganize soil-moisture regimes and water availability. These variability- and event-mediated mechanisms provide an additional interpretation for the zonal differences observed in tem_slope and pre_slope sensitive ranges.

4.3. Climatic Thresholds and Nonlinear Responses of SOS

Based on the SHAP dependence plots and LOWESS smoothing, we quantified the nonlinear response structures of tem_slope and pre_slope with respect to SOS trends and identified city-scale climatic thresholds (defined as SHAP zero-crossing points on LOWESS-smoothed curves). Under the adopted screening strategy, each city is assigned one main turning point by default, while an additional turning point is retained only when it passes strict criteria (sample-density support, high relative support compared with the main turning point, and a minimum separation distance). For visualization and inter-city comparability, the dependence plots for each driver adopt a shared x-axis range across all cities (Figure 7 and Figure 8).

Across cities, tem_slope thresholds are positive for most cases, indicating that the marginal temperature contribution to SOS trends tends to shift from near-neutral to more temperature-sensitive regimes only after warming reaches a certain rate. At the climate-zone scale, plateau cities show thresholds of 0.0217–0.0501 °C·year⁻¹ (Lhasa and Xining). Southern subtropical cities exhibit moderately positive thresholds (0.0492–0.0664 °C·year⁻¹; Nanning and Guangzhou). Middle subtropical cities display the widest spread, ranging from negative to clearly positive thresholds (−0.0267 to 0.1000 °C·year⁻¹), suggesting heterogeneous temperature sensitivity in mountainous subtropical settings. Northern subtropical thresholds mainly fall within 0.0234–0.0920 °C·year⁻¹, while mid-temperate thresholds concentrate around 0.0193–0.0801 °C·year⁻¹, implying that warming rates on the order of approximately 0.02–0.08 °C·year⁻¹ typically correspond to the emergence of non-neutral marginal temperature contributions at the city scale. Warm-temperate thresholds span −0.0233 to 0.0566 °C·year⁻¹, and the marginal tropical city Haikou shows a positive threshold at 0.0275 °C·year⁻¹.

In contrast to temperature, pre_slope thresholds, expressed on an annual total precipitation basis, remain concentrated within relatively modest long-term wetting or drying trends, while still showing strong inter-city heterogeneity. Plateau thresholds are small and positive (2.2129–2.6248 mm·year⁻¹; Xining: 2.2129; Lhasa: 2.6248). In the southern subtropical zone, thresholds differ in sign between the two cities (Nanning: 4.6989 versus Guangzhou: −0.4609 mm·year⁻¹), highlighting non-uniform moisture sensitivity even under broadly humid climatic conditions. Middle subtropical cities span the widest range (−13.4286 to 14.9929 mm·year⁻¹; e.g., Kunming: −13.4286; Nanchang: −2.1743; Guiyang: 0.4572; Chengdu: 2.3373; Chongqing: 14.1318; Fuzhou: 14.9929), implying that both intensified drying and wetting can be associated with non-neutral SHAP contributions depending on local setting. Northern subtropical thresholds also straddle zero (−8.9312 to 9.4372 mm·year⁻¹; e.g., Shanghai: −8.9312; Wuhan: −2.8960; Changsha: −5.5045; Nanjing: −0.0520; Hefei: 7.6678; Hangzhou: 9.4372). Mid-temperate thresholds tend to be negative overall but still include sign heterogeneity (−8.0515 to 4.4671 mm·year⁻¹; e.g., Hohhot: −8.0515; Urumqi: −5.4774; Lanzhou: −5.4804; Yinchuan: −5.4290; Shenyang: −2.4543; Harbin: 0.4565; Changchun: 4.4671), suggesting that slight to moderate drying is more often associated with marked marginal contributions to SOS trends in these cities. Warm-temperate thresholds also concentrate near zero with both signs (−4.1455 to 4.6941 mm·year⁻¹; e.g., Taiyuan: −4.1455; Jinan: −2.2237; Zhengzhou: −1.3766; Xi’an: −0.3928; Beijing: 3.5390; Tianjin: 4.5648; Shijiazhuang: 4.6941). Notably, Haikou retains two pre_slope turning points (−11.3841 and −1.8239 mm·year⁻¹), implying that sensitivity may be confined to a drying interval rather than changing monotonically across the full moisture-change spectrum.

Under the adopted screening criteria, most cities exhibit one main turning point for each driver, consistent with a two-regime interpretation (near-neutral vs. sensitive). Multi-threshold patterns are comparatively rare and only retained when strongly supported; for example, Hohhot shows two tem_slope thresholds (0.0801 and 0.1163 °C·year⁻¹), suggesting a potential three-stage temperature response under relatively strong warming. Despite the predominance of a single main threshold, substantial within-city dispersion is evident in many SHAP scatter distributions around the zero-crossing region, implying that marginal responses to the same climatic-change rate are not fully uniform across urban sectors. This pattern is consistent with strong intra-urban heterogeneity in vegetation composition, underlying surfaces, microclimate, and management intensity, and it motivates future work to incorporate explicit heterogeneity indicators (e.g., impervious fraction, vegetation cover, topographic relief, or management proxies) to better explain the dispersion and potential mixing of response regimes.

Our findings are broadly consistent with recent evidence that spring phenology is constrained by both thermal and moisture conditions, while the magnitude and even the sign of climatic sensitivity can vary markedly across regions due to background climate, elevation, and ecosystem/land-surface heterogeneity. For example, studies on the Mongolian Plateau and Xinjiang highlight strong climatic constraints and pronounced spatial variability in spring phenology responses under arid–semiarid and high-elevation contexts, which is consistent with the heterogeneous pre_slope regimes and the within-zone dispersion observed here. Similarly, multi-regime or changing temperature sensitivity of spring phenology has been reported across northern temperate systems and transitional mountain regions, supporting the notion that temperature effects may shift across climatic ranges rather than remaining constant. Together, these studies reinforce that the thresholds derived from SHAP dependence structures should be interpreted as aggregated city-scale response transitions under heterogeneous climatic and surface settings rather than as a single uniform physiological tipping point [79,80,81,82,83,84,85].

4.4. Methodological Implications and Limitations

From a methodological perspective, this study contributes two main innovations to the quantitative analysis of urban phenology–climate relationships. First, we coupled pixel-scale MK–Sen trend analysis with machine-learning attribution to jointly characterize the long-term direction of SOS change and the rates of change in the climatic background within a unified 500 m grid system. This design avoids overreliance on simplified representations based on city averages or station-level indicators and is consistent with recent methodological developments that emphasize spatially explicit, grid-based characterization of driving processes in studies of ecosystem service attribution, landscape-change diagnosis, and the extraction of spatial effects [82,83,84]. Second, we introduced an XGBoost–SHAP framework and further identified climatic thresholds using SHAP dependence curves, enabling nonlinear responses of SOS trends to temperature and precipitation changes to be quantified through a combined strategy of contribution decomposition and threshold extraction. This approach partially addresses the limitations of traditional linear regression and correlation analyses in capturing nonlinear and piecewise responses, and it highlights the broader potential of explainable machine learning for driver attribution of ecosystem services, mechanism-oriented analyses of landscape dynamics, and spatial-effect characterization [82,83,84,85]. Compared with existing SHAP-based studies on ecosystem service attribution, landscape dynamics, and spatial effects [82,83,84,85], the distinctive feature of this study is the use of the Sen slope of urban SOS trends as the response variable, allowing sensitive intervals of temperature- and precipitation-change rates to be identified directly at the joint comparison scale of cities and climatic zones. As in other studies adopting XGBoost–SHAP or related interpretability approaches [83], the thresholds identified here are statistical response turning points rather than strictly physiological thresholds or parameters of process-based models. Nevertheless, they provide informative empirical references for defining plausible parameter ranges in process models and for conducting cross-threshold sensitivity analyses in scenario simulations.

Several limitations should also be acknowledged. First, SOS extraction relied on the Greenup_1 metric from the MCD12Q2 product. Although its temporal and spatial resolutions are suitable for large-scale comparisons, mixed pixels and classification errors are unavoidable over heterogeneous urban surfaces and may influence local threshold locations. Second, TerraClimate represents a relatively smoothed, large-scale climatic background, and its spatial resolution cannot fully capture urban heat-island effects and microclimatic variability. Consequently, the climatic thresholds derived in this study are more representative of regional climate–urban phenology relationships than of fine-scale, neighborhood-level thresholds. We also note that the MK test can be sensitive to serial autocorrelation (e.g., AR(1)) in annual sequences, which may inflate nominal significance if not variance-corrected; therefore, MK-based significance levels in this study are primarily used as a screening criterion, and applying trend-free pre-whitening or related sensitivity analyses is left for future work. In addition, because MK significance is assessed at the pixel level, multiple testing may increase false positives under nominal p-values; thus, pixel-wise significance is interpreted here as a screening indicator, while formal multiplicity control (e.g., FDR correction) and spatial field-significance approaches (e.g., clustering-based tests) are left for future work. Third, XGBoost–SHAP remains an association-based interpretability framework. Threshold identification depends on the sample distribution and LOWESS smoothing, and robustness may be limited in regions of the feature space where samples are sparse. Although we assessed the reliability of threshold extraction via bootstrap-based validation for representative cities across climatic zones (Supplementary Figures S3 and S4), future work could extend this uncertainty analysis to all study cities to comprehensively quantify statistical robustness and city-specific variability in threshold identification. A full parameter-sensitivity analysis of the LOWESS fraction and support-window settings was not exhaustively conducted in the present study; although the threshold-consistency and bootstrap results support overall robustness, further evaluation of parameter-dependent variability remains a useful direction for future work.

In the present study, the representative-city bootstrap results shown in Supplementary Figures S3 and S4 are intended as zone-level illustrations rather than as a complete 31-city uncertainty census; they indicate that threshold uncertainty is not uniform across climatic contexts, with some representative cities showing broader confidence envelopes than others.

In addition, because the modeling samples are spatially structured 500 m grids, spatial autocorrelation among neighboring pixels may lead to optimistic estimates under conventional random cross-validation. Therefore, the reported cross-validation metrics should be interpreted primarily as an internal stability check rather than strict out-of-area predictive generalization. Accordingly, the SHAP-derived thresholds in this study are intended to characterize response patterns within the standardized multi-city framework and to support cross-city comparability and hypothesis generation; future work may further assess predictive performance and SHAP stability using spatially structured validation strategies for representative cities. Moreover, we did not explicitly control for static geographic gradients (e.g., latitude and elevation) or baseline vegetation/land-cover structure, nor did we include proxies of urbanization intensity and management (e.g., built-environment metrics, greening configuration, or irrigation). These factors can modulate climatic sensitivity and may lead to shifts in the locations of SHAP-derived response turning points; incorporating such covariates and evaluating robustness under correlated predictors are therefore important directions for future work. In particular, we did not explicitly incorporate intra-urban heterogeneity metrics (e.g., impervious fraction, vegetation cover, topographic relief, or management proxies) to quantitatively explain the within-city coexistence of opposite SOS trends or the dispersion near thresholds; this represents an important direction for future work. Expected directions of influence. Latitude and elevation generally strengthen energy limitation (lower preseason temperatures, delayed snow/soil thaw) and thus are expected to be associated with stronger temperature control, potentially lowering the effective tem_slope threshold at which SOS becomes highly responsive and increasing the magnitude of warming-related contributions. Baseline vegetation/land-cover structure may shift threshold locations because plant functional types differ in chilling/forcing requirements and water–energy constraints (e.g., woody vs. herbaceous, evergreen vs. deciduous), thereby altering the curvature and turning points of SHAP–feature relationships. Urbanization intensity (e.g., impervious fraction, built-up density) can modify local thermal and moisture environments via urban heat-island effects and evaporative demand, which may amplify apparent warming rates while also enhancing moisture limitation; conversely, irrigation and greening management may buffer moisture deficits and partially decouple SOS from background pre_slope changes. These factors may therefore shift both the number and locations of SHAP-derived turning points. Finally, this study focused primarily on climatic drivers and did not explicitly incorporate anthropogenic factors such as land-use change, greening management, irrigation, and urban expansion. These processes can indirectly affect the sensitivity of spring SOS to climatic change by altering local energy and water conditions [72,73].

Future work could be advanced in three directions. First, integrating higher-resolution remote-sensing products with refined urban climate datasets would improve the ability of threshold identification to resolve intra-urban heterogeneity, particularly in areas with strong urban heat-island intensity and rapid urbanization. Second, building on the current climate-driven framework, incorporating non-climatic factors such as urban morphology metrics, management practices, and air pollution through multi-source data fusion or joint modeling could extend climatic thresholds toward coupled climate–human-activity thresholds. In parallel, future work may extend the hydroclimatic driver set (e.g., soil moisture, evapotranspiration, vapor pressure deficit, and downward shortwave radiation) to evaluate whether the identified response regimes remain robust under coupled climatic controls. Such extensions should explicitly address multicollinearity and interpretability (e.g., correlated predictors and downstream variables) to avoid conflating marginal contributions in attribution analyses. This would better explain the combined effects of urban expansion and climate change on urban phenological gradients [69,70] and connect with recent methodological progress in characterizing spatial effects using local interpretation methods [83]. Third, linking the tem_slope and pre_slope thresholds identified here with process-based phenology models and ecosystem process simulations would allow mechanistic evaluation of these empirical thresholds and enable assessments under future climate scenarios of potential phenological reorganization after crossing key thresholds, together with associated ecological implications [75,76,77,78,79,82,83,84,85].

5. Conclusions

Based on MCD12Q2 and TerraClimate data for 2001–2023, this study developed a 500 m urban-grid framework integrating MK–Sen trend analysis with XGBoost–SHAP–based threshold identification and systematically assessed the spatiotemporal patterns of SOS and their relationships with climatic change-rate metrics across 31 provincial capitals and municipalities in mainland China. The main conclusions are as follows:

(1)

The multi-year mean SOS exhibits a clear large-scale spatial gradient characterized by earlier green-up in the south, later green-up in the north, and delayed green-up over plateau regions. City-level mean SOS values are mainly distributed within approximately 74–138 DOY, with pronounced differences among climatic zones. Cities in the southern subtropical and marginal tropical zones generally show earlier green-up with relatively smaller within-city dispersion, whereas the middle subtropical zone exhibits the largest inter-city dispersion, spanning from the low 80 s to around 101 DOY. These patterns are consistent with the combined influences of large-scale climatic gradients and local geographic conditions (e.g., topography and underlying-surface heterogeneity). Pixel-scale MK–Sen results further show that advancing and delaying pixels commonly coexist within most cities. When trend directions of significant and marginally significant pixels are summarized at the city scale, marked inter-city differences emerge in the proportions of β < 0 and β > 0 pixels, manifesting as cities dominated by advancing trends, cities dominated by delaying trends, or cities in which the two types occur at comparable proportions. Overall, these results highlight substantial geographic heterogeneity in both the direction and magnitude of SOS changes within and across cities.

(2)

XGBoost–SHAP attribution reveals that the impacts of climatic-change rates on SOS trends are predominantly nonlinear and piecewise at the city scale. For temperature, most cities exhibit a positive tem_slope threshold (typically ~0.02–0.08 °C·year⁻¹), indicating that marginal temperature contributions become evidently non-neutral only after warming exceeds a minimum rate; threshold ranges vary across climatic zones, with plateau cities clustering around 0.02–0.05 °C·year⁻¹ and middle subtropical cities spanning from slightly negative to clearly positive values (−0.0267 to 0.1000 °C·year⁻¹). In contrast, precipitation-related thresholds, expressed on an annual total precipitation basis, span an overall range of −13.4286 to 14.9929 mm·year⁻¹, indicating that transitions into moisture-sensitive regimes can occur under relatively modest long-term wetting or drying trends, with substantial inter-city heterogeneity. Under our strict screening criteria, most cities are characterized by a single main turning point for each driver, while multi-threshold patterns are rare and retained only when strongly supported (e.g., Hohhot for tem_slope and Haikou for pre_slope). Collectively, these turning points delineate comparable climatic thresholds and sensitive intervals that enable cross-city comparisons of nonlinear climate–SOS linkages across climatic zones.

(3)

By integrating gridded MK–Sen trend analysis with XGBoost–SHAP attribution and threshold identification, this study provides a transferable and interpretable framework for examining urban phenology–climate relationships that can be extended to other phenological indicators and ecological response processes. Meanwhile, given the spatiotemporal resolutions of MCD12Q2 and TerraClimate and the lack of explicit representation of anthropogenic influences (e.g., land-use change, greening management, and urban heat-island effects), the thresholds reported here should be interpreted as statistical response turning points within an observational, model-interpretation framework under a regional climatic background rather than as strict causal tipping points. Future work should combine higher-resolution phenological products with refined urban climate datasets and incorporate urban morphology and management factors to better resolve intra-urban heterogeneity, and to evaluate the potential ecological risks and management implications associated with crossing these statistically identified response regimes using process-based models and scenario simulations.