Introduction

Land surface phenology (LSP) is a significant indicator for studying the recurring season of vegetated areas based on remote sensing (White et al. 2009; de Beurs and Henebry 2010; Richardson et al. 2013; Xie et al. 2021). The dynamics of LSP in vegetated areas in urban and rural ecosystems play a crucial role in affecting the human living environment, such as public health and ecological processes (Buyantuyev and Wu 2012; Donnelly et al. 2018). However, the changing urban built environment can affect the LSP and lead to its spatiotemporal alteration, such as advanced spring and delayed autumn vegetation phenology in many temperate cities (Walker et al. 2015; Krehbiel et al. 2016; Zhou et al. 2016; Gervais et al. 2017; Li et al. 2017a; Meng et al. 2020). These impacts are significantly amplified in urban ecosystems because of anthropogenic activities, particularly in urbanization and infrastructure development (Luder et al. 2018; Jeong et al. 2019).

LSP studies have revealed differences between urban and rural regions (e.g., city and countryside) by contrasting different land use and land cover types (Dallimer et al. 2016; Zhou et al. 2016; Krehbiel et al. 2017; Li et al. 2017b; Qiu et al. 2017; Yao et al. 2017). However, the comparisons of microclimates or micro-scale characteristics within a city (Parece and Campbell 2018) and between cities (Zipper et al. 2016; Xie et al. 2022) has received little attention. Notably, the response mechanism of vegetation phenology to temperatures in the built environment under various local climates in mega-urban areas has not been investigated (Melaas et al. 2016; Zipper et al. 2016; Krehbiel et al. 2017; Li et al. 2017c; Villalobos-Jiménez and Hassall 2017; Kabano et al. 2021).

Generally, the start of season and end of seasons (SOS and EOS, respectively) influence ecosystem processes in cities (Richardson et al. 2010; Zipper et al. 2016; Jeong et al. 2019). The SOS is documented as a critical determinant of the beginning of growth (Richardson et al. 2010; Trujillo et al. 2012; Barrio et al. 2013). Changes in both SOS and EOS determines the length of season (i.e., the growing period) (Walther et al. 2002; Richardson et al. 2013; Gill et al. 2015). Reports have indicated that the higher land surface temperatures (LSTs) lead to earlier SOS, thus, an extended season in urban regions than in rural or naturally vegetated areas (Jochner et al. 2011, 2012, 2013). Urban surface warming effects can influence the EOS. For example, temperate cities may experience interannual differences in the EOS, resulting in an extended length of the growing season, whereas tropical cities may encounter a shortened growing season (Zhou et al. 2016; Krehbiel et al. 2017; Kabano et al. 2020, 2021).

Urban regions with high coverage of impervious surfaces increase the sensible heat flux resulting in the urban heat island (UHI) that affects the local climate, contrasting the effect of natural ground cover (Avissar 1996). Unlike the UHI, which focuses on the urban–rural air temperature gradient, the surface UHI (SUHI) typically represents the most significant surface temperature differences along the urbanization gradient during daytime and the slightest discrepancies at night (Roth et al. 1989). The SUHI in urban regions is higher than that in the surrounding zones (Walker et al. 2015). The urban canopy layer air temperature and SUHIs are influenced by the morphology of streets and buildings, construction materials, the permeability of surfaces, and the anthropogenic heat of cities (Stewart and Oke 2012; Stewart et al. 2014; Leconte et al. 2015; Krehbiel et al. 2017; Villalobos-Jiménez and Hassall 2017; Lindh et al. 2018; Daramola and Balogun 2019; Backe et al. 2021). The local climate zone (LCZ) describes the morphology, structure, and cover of urban land surface with the comparability and consistency of cities’ land use and land cover classification (Stewart and Oke 2012). Local climate zones (LCZs) can be categorized into different classification levels (Stewart et al. 2014) (see the summary of values describing their surface properties in Table S1). SUHIs and their climatic effects differed among LCZ types (Goggins et al. 2012; Leconte et al. 2015; Ng and Ren 2018; Mushore et al. 2019; Wang et al. 2019a).

Over the past two decades, the Pearl River Delta (PRD) region has experienced rapid urban sprawl (Xu and Gong 2018), and its expansion is expected to intensify (Liang et al. 2018a, 2018b; Chen et al. 2020). On the one hand, accompanied by a high urbanization rate and high urban density, the region has observed an increasing trend of SUHI intensity (Wang et al. 2019a, 2019b) and greenness (Hu and Xia 2019), although the two are usually negatively correlated (Jochner et al. 2011, 2013). On the other hand, considering the expected urban growth in future land use (Liang et al. 2018a, 2018b), an investigation of SUHIs and urban ecosystems would benefit future urban heating mitigation and greenspace planning and optimization (Dallimer et al. 2016; Su et al. 2020) in the PRD region. In the subtropical region, the interannual variability of vegetation phenology was higher than that in temperate regions (Wang et al. 2015). It is sensitive to the local urban climate (Kabano et al. 2020), which can affect biodiversity change (Du et al. 2019) and water use efficiency (Guo et al. 2020).

However, there remains insufficient knowledge on the spatiotemporal dynamics of LSP associated with urbanization, its high-density urban setting, and complicated urban activities in the PRD region, which is critical for the understanding and strategic planning of the human living environment. Effects of combined fine urbanization characteristics and the corresponding SUHI (Wang et al. 2019a, 2019b) on urban LSP require further analysis related to the LCZ scheme. Studying LCZ combined with LSP is crucial for applications in studies of urban ecosystems and their response to SUHIs (Parece and Campbell 2018; Kabano et al. 2021). Moreover, understanding the feedback mechanisms that microclimate and micro-characteristics of urban areas generate in the environment and climate is critical (Latorre 1999; Solonen and Hildén 2014; Townroe and Callaghan 2014; Lindh et al. 2018; Katz et al. 2019; Su et al. 2020).

In this work, the study objectives are to (i) investigate the spatial relationship of LCZs with LSP and LST across the PRD, (ii) test the interannual changes of LSP and study its association with the urban–rural changes of LCZs combined with their LST for the period 2000–2019, and (iii) analyze the influence of urbanization intensification on LSP across LCZ gradients in the PRD region.

Methods

Study area

The PRD region is one of the global most urbanized, emerging mega-urban areas. With a total area of 55,374.79 km2 surrounding the estuary of the Pearl River (Fig. 1), the PRD region is on the southeast coast of China (111.35–115.40° E, 21.56–24.40° N) with both tropical and subtropical climate characteristics. Its physical condition is characterized by extensive plains and numerous hills ranging from 0 to 1585 m above sea level (m asl; Fig. 1). The PRD region comprises 11 cities and has experienced exponential economic and industrial development since the effective implementation of China’s Reform and Opening-Up policies in 1978 (Wong et al. 2003; Shen et al. 2006), accompanied by a substantial population increase and environmental changes (Gong et al. 2019; Yim et al. 2019; Zhang et al. 2019). Additionally, the study region has experienced high urban density and high rates of urban expansion (Lin et al. 2009; Liu et al. 2019; Wang et al. 2019a; Wong et al. 2019) and the degradation of vegetated areas that have converted to built areas since the 1970s (Wang et al. 2019a, 2019b).

Fig. 1
figure 1

Location, spatial elevation patterns, and administrative boundaries of cities in the Pearl River Delta

Overview of study steps

Figure 2 provides a flowchart of the steps involved in this analysis. Each step is outlined in more detail below, starting with LCZ determination, LSP identification, LST calculation, and then finally regression analysis (Fig. 2). In this study, we employed time series LSP metrics (i.e., start, end, and length of season) and LST metrics (i.e., daytime and nighttime LST in both spring and autumn). We analyzed the interannual covariation of LSP metrics with LST metrics depending on LCZ types of vegetated areas from 2000 to 2019. As explanatory variables, we used the LCZs (Table S1) of the PRD region.

Fig. 2
figure 2

Flowchart of the research steps for investigating vegetation phenology along urban–rural local climate zone gradients using Earth observation data

Local climate zones and urban–rural gradients

Open-access remote-sensing resources provide free data for classifying LCZ maps over decades. Multispectral and radar imaging sensors have systematically tracked and characterized urbanization and spatiotemporal variations in millions of locations across Earth for over a decade (Xu et al. 2017; Bechtel et al. 2019; Demuzere et al. 2019). We employed the LCZ classification in 2000 and 2019 to avoid the uncertainty in tracking continuous time series dynamics of vegetation phenology resulting from land cover conversion between LCZ built types (i.e., LCZ 1–10) and land cover types (i.e., LCZ A–H) over the investigated period of 2000–2019.

To map the LCZs of 2000 and 2019 in the PRD region (first column in Fig. 2), we presented an adjusted workflow using a cloud-based Google Earth Engine (GEE) platform: (1) various pre-processed Earth observation data in the GEE archive were selected and post-processed using GEE’s client library; (2) LCZ sampling data (Fig. S1) were collected and edited in Google Earth Pro and exported to the GEE platform using the.kml format; and (3) the random forest (RF) classifier was applied to produce the LCZ classification on the GEE platform, and an accuracy assessment was conducted, as described by (Hay Chung et al. 2021).

For the 2000 LCZ classification, USGS Landsat 5 TM orthorectified images at a 30 m spatial resolution for 01/01/2000–31/12/2000, NASA SRTM Digital Elevation 30 m (Farr et al., 2007) of 2000, and DMSP OLS: Global Radiance-Calibrated Nighttime Lights Version 4 and Nighttime Lights Time Series Version 4, Defense Meteorological Program Operational Linescan System at 30 arc seconds (approximately 1 km spatial resolution) of 2000 were collected. For the 2019 LCZ classification, the Sentinel-1 SAR GRD: C-band Synthetic Aperture Radar Ground Range Detected, log scaling bands at a 10 m spatial resolution and cloud-free Sentinel-2 MSI: MultiSpectral Instrument, Level-1C bands at 10 and 60 m spatial resolution, VIIRS Stray Light Corrected Nighttime Day/Night Band Composites Version 1 at 15 arc seconds (approximately 500 m spatial resolution) for 01/01/2019–12/31/2019, and GMTED2010: Global Multi-resolution Terrain Elevation Data 2010 at 7.5 arc seconds (approximately 250 m spatial resolution) were employed.

A total of 1972 (2000) and 2144 (2019) digitized samples (Fig. S1) were collected across the urbanized areas and their surrounding areas in the PRD region based on Google Earth Pro by using the World Urban Database and Access Portal Tools (WUDAPT) workflow (Bechtel et al. 2015) that typify the LCZ types (Fig. S1). They were digitized, saved in the KML format, and imported into the GEE platform using a Google Fusion table. The training and testing samples were selected randomly, and 70% of samples (1405 samples [2000] and 1459 samples [2019]) were used for classifier training. The remaining 30% of samples (567 samples [2000] and 685 samples [2019]) were used for the accuracy assessment. To create the LCZ maps of 2000 and 2019 for the study regions, an RF classifier (Breiman 2001) was employed because of its ability to balance the computational performance and achieve accuracy (Bechtel et al. 2015). The number of trees was set to 80, according to the error tests of the RF classifier in this study.

Temperature and its elevated local warming effect differ among LCZ types with urban–rural gradients (Goggins et al. 2012; Leconte et al. 2015; Ng and Ren 2018; Mushore et al. 2019; Wang et al. 2019a). Compact high-rise denotes a dense mix of tall buildings with the highest impervious but the lowest pervious surface fraction and the highest anthropogenic heat output among LCZ built types (1–10) (Stewart and Oke 2012). Moreover, most compact high-rise buildings are located in the centre of urban areas in the PRD region from 1999 to 2019 (Xie et al. 2022). Thus, we computed the distance gradients by selecting a distinct 250 m zone iteratively displaced every 250 m between the spatial borders of compact high rises (i.e., LCZ 1) of the LCZ classification map of 2019 and the administrative boundary of the PRD region. In this study, an urban–rural gradient represented the distance to the closest urban center (i.e., LCZ 1).

Land surface phenology

The normalized difference vegetation index (NDVI) and the enhanced vegetation index (EVI) can capture seasonal changes in vegetation properties and productivity, but EVI is more suitable for high biomass conditions (Pettorelli et al. 2005; Sesnie et al. 2012; Melaas et al. 2013). These indices enable local to broad-scale monitoring of vegetation phenological events (Tucker 1979; Pettorelli et al. 2005; Fisher et al. 2006; Cleland et al. 2007; Liang and Schwartz 2009; White et al. 2009; Zhao et al. 2022). Moderate-resolution imaging spectroradiometer (MODIS) data provide advantages in monitoring LSP derived from the vegetation index (VI) in long-term observations and broad coverage (Walker et al. 2012). In this study, 457 NASA MOD13Q1 (collection 6250 m resolution) products between 2000 and 2019 were employed to extract the NDVI and EVI (represented as VIs) (2nd column in Fig. 2), VI quality indicators, and day of year VI pixel layers by using the practice of (Xie et al. 2017). A pixel with annual mean VIs lower than 0.2 from the LSP estimation was excluded and thus not considered in this study, since the annual mean VI values within built areas of 2000 and 2019 are lower than 0.2 according to the LCZ 1–10 sample test.

Many approaches can determine the start of season (SOS) and end of season (EOS) from Earth observation data (de Beurs and Henebry 2010). However, the standard midpoint pixel threshold (White et al. 2009) applies well to time series phenology variations and at large scales such as in tropical/subtropical regions (Garonna et al. 2016, 2018). VIs were used to derive yearly LSP metrics for the SOS, EOS, and length of season (LOS) at the pixel level. First, the time series of the VIs were smoothed using harmonic analysis describe the seasonal change by sinusoidal functions effectively (Roerink et al. 2000). Second, SOS was estimated using the day of a year when VIs first arrived at the midpoint of their annual amplitude using the Midpointpixel approach (White et al. 2009). EOS was then computed using the day on which VIs declined below this mid-threshold, and LOS was estimated as the number of days of the year between the SOS and EOS. Day of Year (DOY) was expressed for the SOS and EOS, and number of days for LOS. Finally, pixels with the SOS before DOY 16 and EOS after DOY 352 were filtered out according to the collection date range of the VI layers. In this study, SOSNDVI, EOSNDVI, and LOSNDVI represent the LSP metrics derived from the NDVI, and SOSEVI, EOSEVI, and LOSEVI represent the LSP metrics derived from the EVI. Vegetated pixels that experienced land cover changes based on LCZ maps between 2000 and 2019 were masked off. Furthermore, an independent Global land-cover product (GLC_FCS30) of 2000 and 2020 (Zhang et al. 2021) was utilized to ensure the exclusion of agricultural vegetation information. VIs and LSP computation were conducted using R programming (version 3.4.1). Mapping visualizations of LCZs were conducted using ArcGIS (v10.8, ESRI, USA).

Land surface temperature metrics and the digital elevation model

We also analyzed the temperature controls on phenology in the PRD region (3rd column in Fig. 2). Remotely sensed LST is reported to correlate well with a SUHI, depending on urban–rural gradients (Roth et al. 1989; Walker et al. 2015). In this study, LST data over the period 2000–2019 were obtained from 3660 NASA MOD11A1 products with a spatial/temporal resolution of ~ 1 km/1 day. We converted MOD11A1 LST from Kelvin degrees to degrees Celsius (°C). The seasonal average values were calculated from the daily values for each temperature factor (day LST and night LST). A pixel with good quality, a cloud-free record, or a small error was employed in the seasonal average LST metric computation according to MOD11A1 layers of daytime and nighttime LST quality indicators. According to the varying range of DOY for SOS (SOSNDVI and SOSEVI) and EOS (EOSNDVI and EOSEVI) over the study period (Fig. S6), the selected spring season was defined as March, April, and May (92 days of the year), and autumn was defined as September, October, and November (91 days of the year) (Zhou et al. 2016; Xie et al. 2020). The employed LST metrics in this study were the spring day average LST (STday), spring night average LST (STnight), autumn day average LST (ATday), and autumn night average LST (ATnight) in vegetated areas (LCZ A–D).

Digital elevation information of the PRD region at an ~ 30 m scale was collected from the United States Geological Survey (https://lta.cr.usgs.gov/SRTM1Arc) and employed to generate a digital elevation model (DEM). DEM was applied to mask the pixels at a higher elevation because more than 99.0% of built areas were located at elevations lower than 200 m asl by 2019 (Xie et al. 2022). The time series of LST metrics and DEM were converted into the Universal Transverse Mercator projection in Zone 49 North (UTM 49N) and World Geodetic System 84 (WGS-84) and resampled as 250 m scale images by using the nearest neighbor approach in an ENVI/IDL environment (v5.3, EXELIS Inc., McLean, VA, USA) for subsequent statistical analysis matching coordinates and 250 m resolution of LSP metrics. Mapping visualizations were performed in ArcGIS (v10.8, ESRI, USA).

Statistical analysis

To identify relative changes in both SOS and EOS to the interannual dynamics of LOS over the study period in the PRD region, we computed the interannual differences (Δ) of SOS and EOS at the pixel level, which are the difference (expressed in days) between the current (observed) year and the preceding year, as Xie et al. (2017) described. To measure the relative determination of interannual variation in SOS and EOS to the interannual dynamics of LOS by employing interannual differences of SOS (ΔSOS) and EOS (ΔEOS), we adopted the C index approach used by Garonna et al. (2014) as follows:

$$C=\frac{\left|\Delta SOS\right|-\left|\Delta EOS\right|}{\left|\Delta SOS\right|+\left|\Delta EOS\right|}.$$

We evaluated each pixel’s mean C values (computed from the symmetry of 19 ΔSOS and 19 ΔEOS). The C index varies between − 1 and 1. A positive value of C indicates that LOS variation is more influenced by a shift of SOS than EOS between the current (observed) year and the preceding year. By contrast, a negative C index means that the EOS shift is more important than the SOS shift in the interannual variation in LOS. We expressed CNDVI to identify the C index derived from SOSNDVI, EOSNDVI, and LOSNDVI, and CEVI computed from SOSEVI, EOSEVI, and LOSEVI.

Linear regression was used to test the urban–rural change trends (with statistical significance defined as p < 0.001) in LST variables at the 250 m gradient level. Linear regression was also employed to examine interannual trends (with statistical significance defined as p < 0.05) in LSP metrics and LST variables at the pixel level over the 20 investigated years. Partial Spearman’s rank correlation coefficients (r) were used to test interannual associations (with statistical significance defined as p < 0.05) of LSP metrics with LST variables at the pixel level of the study period.

We combined 250 m scale maps of the statistical outcomes with the 250 m scale distance gradients and LCZ maps. They were then analyzed across distance gradients in the PRD region (lower panel in Fig. 2). Each distance gradient was characterized by the average values of consisting 250-m gradient zones of the statistical regression/correlation outcomes on statistical significance (defined as p < 0.05). Distance gradients with (i) a few vegetated pixels less than 5% proportion of total areas and (ii) a percentage of statistically significant regression/correlation lower than 5% were filtered out in the urban–rural gradient analysis. All statistical analyses were computed using R programming (version 3.4.1).

Results

Local climate zones, land surface phenology, and land surface temperatures and their urban–rural gradients

Based on the LCZ classification, the LCZ maps of 2000 and 2019 were generated (Fig. 3). The accuracy assessment was conducted using an independent set of validation samples for all 17 LCZ classes from 2000 to 2019 (Tables S2, S3). The overall accuracies were 96.3% and 94.7%, built class accuracies were 71.7% and 72.5%, and κ coefficients were 0.810 and 0.883 for 2000 and 2019, respectively. The areas of all LCZ built types (1–10) of 2000 and 2019 accounted for 10.5% and 13.4%, respectively, of the total areas (Fig. 3). Based on the LCZ 1 compact high-rise of 2019, the distance gradients were generated to identify the urban–rural local climate zone gradients (Fig. S2). According to the detection of LCZ changes between 2000 and 2019, 62.7% of the vegetated pixels experienced no changes and were employed in the following analysis.

Fig. 3
figure 3

Spatial patterns of LCZ types of 2000 and 2019 in the PRD region. The black lines indicate a 10,000 m distance to the spatial borders of compact high rises (i.e., LCZ 1) based on the LCZ classification map of 2019

More than 99.0% of LCZ built areas were located at gradients between 0 and 10,000 m (Fig. 4). More than 95.0% of LCZ built areas were distributed in the distance gradients between 0 and 5000 m of 2000 and 2019 (Fig. 4; Fig. S3). Compact high-rise (LCZ 1) is the most central among all other LCZ built types (i.e., 2–10). The areas of LCZ 2–10 varied with their distance gradient to LCZ 1 of 2019, especially notable is the decrease in distance gradients between 0 and 5000 m. However, LCZ built types were limited but regularly scattered at a distance farther than 5000 m in 2000 and 2019. Specifically, LCZ 1, 2, 4, and 5 (characterized as compact, high-rise, or mid-rise types) were located at distance gradients between 0 and 2000 m. LCZ 3, 6, 7, and 8 (characterized as open or low-rise types) were mainly distributed at distance gradients between 0 and 2000 m and secondarily between 2000 and 5000 m. More than 95.0% of LCZ built areas were distributed in the distance gradients between 0 and 5000 m of the period 2000–2019 (Fig. 4; Fig. S3). The LCZ 2–10 of each city showed a similar variation with its distance gradient to LCZ 1 in the PRD region of 2019 (Fig. S4). The investigated vegetation areas (LCZ A–D) between 2000 and 2019 are distributed around each city’s LCZ built types (Fig. 3) and varied regularly with their distance gradient to LCZ 1 of 2019 (Fig. S5). In most surveyed cities, the area of all LCZ built types decreased with distance gradients between 0 and 2000 m (Fig. S4). Compact and high-rise LCZ built types (2–4) were primarily located at distance gradients close to LCZ 1. In the following statistical analysis of distance gradients, we set 2000 and 5000 m as gradient thresholds to match the variation of LCZ types with distance gradients.

Fig. 4
figure 4

Distance gradient distribution of LCZ built types (1–10) of 2000 and 2019 (left vertical axis) and its percentage of the total areas (right vertical axis). Dashed grey lines represent the distance of 2000 and 5000 m

Over the study period of 2000–2019, the land surface phenology (LSP) metrics, i.e. SOSNDVI and EOSNDVI derived from NDVI and SOSEVI and EOSEVI derived from EVI, varied with vegetated spatial patterns across the PRD region (Fig. S6). More specifically, the mean value of start of season (SOS) (SOSNDVI and SOSEVI) ranged between 50 and 159 day of the year (DOY), the mean value of end of season (EOS) (EOSNDVI and EOSEVI) changed from 202 to 335 DOY.

The variations of LOS (LOSNDVI and LOSEVI) are determined by the changes of SOS (SOSNDVI and SOSEVI) and EOS (EOSNDVI and EOSEVI). C index analysis results (Fig. 5) show that CNDVI varied between − 0.13 and − 0.10, and CEVI varied between − 0.20 and − 0.16 with the distance gradients. The negative values of CNDVI and CEVI showed that the interannual variations of LOS (LOSNDVI and LOSEVI) are more determined by the interannual shifts of EOS (EOSNDVI and EOSEVI) rather than SOS (SOSNDVI and SOSEVI) over the study period. CNDVI and CEVI presented the most substantial values at distance gradients of 0–2000 m, and the values were secondary at 2000–5000 m. Spatial patterns of CNDVI and CEVI are shown in Fig. S7.

Fig. 5
figure 5

Distance gradient mean values of the C index of vegetated area for the PRD region. Dashed grey lines represent the distance of 2000 and 5000 m

Figure 6 shows that the day and night LST in spring and autumn decreased with an increase in distance to the urban built centre (LCZ 1) and their urban–rural decrease trends (with statistical significance defined as p < 0.001) with an increase in gradients in the PRD region. Taking Fig. 6a as an example, an increase of X meters (from X1 to X2) in distance is associated with a temperature decrease of 0.177(ln(X2) − ln(X1)) °C in STday. STday decreased from 13.5 to 12.8 °C, and STnight decreased from 13.9 to 12.7 °C (Fig. 6a, b). However, STnight was slightly higher than STday at the distance gradients between 0 and 3750 m but lower at distance gradients farther than 3,750 m (Fig. 6a, b; Fig. S8a, b). By contrast, at distance gradients to the urban built center (i.e., LCZ 1), the day LST in autumn (ATday) slightly decreased from 23.5 to 21.5 °C, and the night LST in autumn (ATnight) slightly decreased from 18.0 to 16.0 °C (Fig. 6c, d). Notably, ATday and ATnight in the vegetated regions decreased more steeply close to the built areas between 0 and 2000 m than in the farther regions (Fig. 6c, d; Fig. S8c, d). These urban–rural decrease trends were statistically significant in areas at distance gradients between 0 and 10,000 m. Overall, the daytime and nighttime LST in spring (STday and STnight) and autumn (ATday and ATnight) in built areas and their adjoining areas were higher than in other regions.

Fig. 6
figure 6

Distance gradient distribution of the average values of spring day and night (STday and STnight) and autumn day and night LST metrics (ATday and ATnight) of vegetated areas for the PRD region. Dashed grey lines represent the distance of 2000 and 5000 m

Interannual trends in land surface phenology and land surface temperature

Linear least-squares regression results (with statistical significance defined as p < 0.05) showed that LSP metrics experienced statistically significant trends in the PRD area from 2000 to 2019. Figure 7 shows that the average interannual trends (i.e., average linear least-squares regression slopes) of the LSP metrics varied with their distance to LCZ 1. These interannual trends were more pronounced in built areas at distance gradients between 0 and 5000 m than in other vegetated areas farther than 5000 m. Specifically, SOSNDVI and SOSEVI illustrated statistically significant postponing trends with less than 10% of the total pixels over the study period (Fig. 7a). The proportion of statistically significant pixels (% of total pixels) was more than 5% but less than 10% at the distance gradients between 0 and 5000 m and less than 5% at distance gradients farther than 5000 m. EOS (EOSNDVI and EOSEVI) also showed a significantly delayed trend with less than 15% of the total pixels over the investigated 20 years (Fig. 7b). The delayed trends of EOSNDVI occurred more significantly (with more than 7% of the total pixels) at distance gradients between 0 and 5000 than at farther distance gradients. More statistically significant pixels were observed closer to the central built areas, except for EOSEVI and LOSEVI at gradients between 0 and 2000 m. Compared with EOSNDVI, EOSEVI presented a similar case but with the less interannual delay and the corresponding proportion of statistically significant pixels at distance gradients of 0–5000 m (Fig. 7b). LOS (LOSNDVI and LOSEVI) also significantly tended to extend, with less than 25% of the total pixels from 2000 to 2019. The extension of LOSNDVI and LOSEVI increased between 0 and 2000 m (Fig. 7c). Areas with significant trends (with more statistically significant pixels) of LOSNDVI and LOSEVI were more pronounced at distance gradients between 0 and 5000 than at distance gradients greater than 5000 m.

Fig. 7
figure 7

Distance gradient mean values of an interannual trend (day year−1) of LSP metrics (a SOSNDVI and SOSEVI, b EOSNDVI and EOSEVI, and c LOSNDVI and LOSEVI) and the corresponding proportion of statistically significant areas (a % of SOSNDVI and % of SOSEVI, b % of EOSNDVI and % of EOSEVI, and c % of LOSNDVI and % of LOSEVI) of vegetated area for the PRD region. Dashed grey lines represent the distance of 2000 and 5000 m

Over the study period of 2000–2019, the spring LST (i.e., STday and STnight) tended to decrease. The interannual decreasing trend of STday was more substantial than that of STnight (Fig. 8). Notably, the interannual decreasing trends of STnight mainly occurred in the built areas of the distance gradients between 0 and 5000 m. The interannual decreasing trends of STday are especially pronounced in built areas and their adjoining regions. However, ATday and ATnight showed statistically significant trends, with fewer than 3% of the total area in the PRD region over the 20 investigated years.

Fig. 8
figure 8

Distance gradient mean values of an interannual trend (°C year−1) of spring day and night LST metrics (STday and STnight) and the corresponding proportion of statistically significant areas (% of STday and % of STnight) of vegetated area for the PRD region. Dashed grey lines represent the distance of 2000 and 5000 m

Association of land surface phenology with land surface temperature

Over the study period of 2000–2019, the average partial Spearman’s correlation coefficients (r) (with statistical significance defined as p < 0.05) of SOS (SOSNDVI and SOSEVI) with spring LST (STday and STnight) and of EOS (EOSNDVI and EOSEVI) with autumn LST (ATday and ATnight) varied with distance gradients (Fig. 9). In spring, correlations between SOS (SOSNDVI and SOSEVI) and STday and the corresponding proportion of statistically significant areas (% of total pixels) were more substantial and more significant than those observed between SOS metrics and STnight across the distance gradients. In autumn, only the proportion of statistically significant areas (% of total pixels) of correlations between EOS (EOSNDVI and EOSEVI) and ATday were more substantial than those observed between EOS metrics and ATnight across the distance gradients.

Fig. 9
figure 9

Variation in average partial Spearman’s correlation coefficients between spring day and night LST metrics (STday and STnight) and (a SOSNDVI and b SOSEVI) and between autumn day and night LST metrics (ATday and ATnight) and EOS (c EOSNDVI and d EOSEVI) and the corresponding proportion of statistically significant areas (% of total pixels) of each distance gradient for the PRD region. Dashed grey lines represent the distance of 2000 and 5000 m, and the dashed purple line denotes an average correlation coefficient of 0

The correlation of STday with SOSNDVI (with average moderate r approximately − 0.50) was stronger than that of STnight with SOSNDVI across the distance gradients (Fig. 9a). The corresponding proportion of statistically significant areas was the largest at closer distance gradients of 0–2000 m and was secondary at 2000–5000 m. The average correlations of SOSEVI with STday (mean r ranging between − 0.60 and − 0.22) were slightly reduced across distance gradients between 0 and 10,000 m (Fig. 9b). Correlations of SOSEVI with STnight (mean r ranging between − 0.14 and − 0.05) were insignificant across distance gradients. The corresponding proportion of statistically significant areas for the correlations of SOSEVI with STday decreased with distance gradients while for the correlations of SOSEVI with STnight increased with distance gradients but then decreased; expectedly, these correlations were the strongest at distance gradients between 0 and 2000 m and secondary between 2000 and 5000 m.

Correlations between EOSNDVI and ATday differed from positive to negative with distance gradients between 0 and 5000 m. Across distance gradients farther than 5000 m, correlations between EOSNDVI and ATday were negative, with a mean r of approximately − 0.60 (Fig. 9c). Absolute correlations of EOSNDVI with ATnight slightly reduced with distance gradients between 0 and 5000 m but tended to 0 and became insignificant at distance gradients farther than 5000 m. The corresponding proportion of statistically significant areas for the correlations of EOSNDVI with ATday and ATnight showed the highest values at distance gradients of 0–2000 m and presented secondary values of 2000–5000 m.

Negative correlations of EOSEVI with ATday and ATnight were stronger and more significant at a closer distance gradient of 0–5000 m than at farther distance gradients (Fig. 9d). Specifically, the correlation of EOSEVI with ATday has negative values with a mean r of approximately − 0.40 across distance gradients, and the correlation of EOSEVI with ATnight has negative values with a mean r of approximately − 0.35 on distance gradients. The corresponding proportion of statistically significant areas for the correlation between EOSNDVI and ATday and the correlation between EOSNDVI and ATnight presented the most considerable ratio at distance gradients of 0–2000 m and showed a secondary matter of 2000–5000 m.

Discussion

Local climate zones, land surface phenology, and land surface temperatures and their urban–rural gradients

Our results show that the zones closer to the centers of urban areas (LCZ 1: compact high-rise) have higher proportion of LCZ built types. Moreover, our results show that LCZ built types have changed from open, low-rise, and mid-rise LCZs to compact and high-rise across the urban–rural gradients from 2000 to 2019 (Fig. 4). This phenomenon was pronounced in zones close to the Pearl River estuary (Fig. 3) and the centers of the urban regions (Fig. 4), which confirms the urbanization in the PRD region happening as the growth of compact and high-rise buildings, particularly in proximity to the estuary and downtown areas (Liang et al. 2018a, 2018b; Wang et al. 2019a; Chen et al. 2020).

Vegetation phenology can be affected by temperature in both spring and autumn (Menzel et al. 2006; Körner and Basler 2010). However, phenological events are reported to be more variable in autumn than in spring (Walther et al. 2002; Richardson et al. 2013; Garonna et al. 2014; Donnelly et al. 2018). In this study, we observed that the EOS played a more important role than the SOS in determining LOS, based on the C index analysis (Fig. 5). This phenomenon was slightly pronounced (i.e., a greater absolute value of C index) in urban domains (distance gradients of 0–2000 m). This finding further suggests that variation in autumn vegetation phenology can play a significantly pronounced role in the vegetation of urban domains.

Our results show that the areas and proportions of urban–rural LCZs, as well as the spring and autumn LST metrics of vegetated areas, varied considerably and exhibited a statistically significant relationship with distance gradients from urban centers, namely, compact high-rise (LCZ 1; Figs. 4, 6). The LST transition depends on the LCZs and reveals that the SUHI showed a significant urban–rural decline across the PRD region. Specifically, we found that the interannual trends of spring and autumn LST metrics of vegetated areas were more significant in urban–rural gradients dominated by compact and high-rise urban areas compared to those dominated by open, mid-rise, or low-rise regions. Furthermore, these metrics decreased as the distance to compact high-rise (LCZ 1) areas increased. These results indicate that LCZs are linked to the distance gradients of the LST (Figs. 4, 6). This finding is in line with reports that the UHI and SUHI effects are more pronounced in the centre than in other areas of cities and metropolitan belts (Han and Xu 2013; Geletič et al. 2016; Krehbiel et al. 2016; Zipper et al. 2016; Ferreira and Duarte 2018; Parece and Campbell 2018; Richard et al. 2018; Brousse et al. 2019; Katz et al. 2019; Mushore et al. 2019).

Interannual trends in land surface phenology and land surface temperature

Although a small portion of pixels showed statistically significant trends, our findings illustrate the interannual trends of SOS (3.6% of SOSNDVI pixels and 4.1% of SOSEVI pixels) and EOS (5.9% of EOSNDVI pixels and 6.8% of EOSEVI pixels) were delayed. In contrast, the trends of LOS (10.9% of LOSNDVI pixels and 12.4% of LOSEVI pixels) extended over the period of investigation. These trends were more pronounced in vegetated areas located in urban–rural gradients dominated by LCZ 1–6 (i.e., urban domains of distance gradients between 0 and 2000 m) than in other vegetated areas (Figs. 4, 7). These findings confirm the dependency of spatiotemporal vegetation phenology on urbanization identification and expansion in metropolitan areas, which is consistent with reported findings (Gazal et al. 2008; Jochner et al. 2012; Han and Xu 2013; Walker et al. 2015; Dallimer et al. 2016; Liang et al. 2016; Zhou et al. 2016; Zipper et al. 2016; Gervais et al. 2017; Li et al. 2017b; Qiu et al. 2017, 2020; Yao et al. 2017; Parece and Campbell 2018; Singh et al. 2018).

Although an advanced spring start of vegetation growth over the past decades has been reported worldwide in urban domains in the field (e.g., Li et al. 2017c, 2019; Yao et al. 2017; Qiu et al. 2020), our analysis presents a notable finding that the delayed SOS is especially pronounced in regions close to urban centers (of distance gradients between 0 and 2000 m), although without considerably significant statistical results based on linear regression found in the PRD region over the investigated period (Fig. 7). This finding agrees with a report on delayed SOS trends between 2007 and 2013 in Guangzhou (Zhou et al. 2016). We found a delayed trend in EOS, and the delay was more pronounced in urban domains (of distance gradients between 0 and 2000 m) than in suburban (distance gradients between 2000 and 5000 m) and rural regions (distance gradients farther than 5000), which agrees with the widely reported delay of vegetation senescence (Li et al. 2017c). Furthermore, these findings show that the EOS played a role in extending the LOS in the PRD region because both SOS and EOS were delayed over the investigated years.

Our findings indicate that both spring day and night LST (STday and STnight) experienced decreasing trends and variation over the period 2000–2019 (Fig. 8). We did not find significant trends in the autumn day and night LST (ATday and ATnight) based on the time series LST variables and the statistical method employed in this study. Nevertheless, the differences in research periods must be considered when comparing these inconsistencies between reported findings because the direction and strength of a trend are highly dependent on the temporal scale of the investigation periods (Bertin 2008; Marty et al. 2017; Lindh et al. 2018).

Although the ranges of the proportion of pixels (% of total) with a significant interannual trend of LSP metrics (between 5 and 25%) and spring LST variables (between 5 and 40%) were not significant (Figs. 7, 8), our study shows that locations closer to the central urban regions (compact high-rise) showed a greater trend of vegetation phenology and spring LSTs than those that were further away.

Association of land surface phenology and land surface temperature

LST analysis based on urban–rural gradients revealed a significant influence of SUHIs on spatiotemporal LSP in vegetated areas of the PRD region (Fig. 9). That SOSNDVI and SOSEVI were negatively correlated with spring LST metrics, although a slight positive correlation coefficient was observed between SOSNDVI and STnight across distance gradients (Fig. 9a, b), which clearly indicates that spring day and night LSTs influenced the start of vegetation growth in the PRD region and thus affected the variation in vegetation, particularly in urban domains (of distance gradients between 0 and 2000 m). Studies have reported that elevated urban air or LSTs result in significant differences in vegetation phenology (Zhang et al. 2004; Gazal et al. 2008; Dallimer et al. 2016; Zipper et al. 2016; Krehbiel et al. 2017; Katz et al. 2019). Moreover, lower air temperatures can result in later-onset growth of plants (Jochner et al. 2011; Parece and Campbell 2018). In this study, based on the LCZ scheme, our findings show that urban LSTs have a gradient-dependent influence on spring vegetation phenology.

Most centrally, under the scenarios of a global warming climate (IPCC 2019), anthropogenic activities can significantly amplify the impacts of the changing urban built environment on LSP and lead to its spatiotemporal alteration (Luder et al. 2018; Jeong et al. 2019). Specifically, over the investigated period 2000–2019, we found several novel results: decreased spring LST metrics, delayed SOSNDVI and SOSEVI, and the corresponding negative correlation of spring LST metrics with SOSNDVI and SOSEVI across urban–rural gradients in the PRD region. These findings indicate that the PRD region experienced a decreasing trend in spring daytime LST across the entire urban–rural gradients (Fig. 8) and that a delayed SOS was relatively pronounced in the urban domains (Fig. 7a). This cooling trend might be due to the afforestation and vegetation coverage improvements in the PRD region over the past decades (Hu and Xia 2019) because the greening strategy effectively mitigates UHI effects (Venter et al. 2021). Correspondingly, a cooling temperature-influenced delay of spring phenology should be observed over the same period. Based on the space-borne observation findings, a similar case was reported in Fairbanks, USA (Gazal et al. 2008). In addition, the dynamics of a later bud break, longer leaf discolouration, and shorter growing seasons were reported to be affected by UHIs in Salzburg, Austria (Stanley et al. 2019).

We found that EOSNDVI and EOSEVI were negatively correlated with ATnight across distance gradients (Fig. 9c, d), indicating that autumn night LSTs influence the senescence of vegetation. However, ATday presented a positive correlation with EOSNDVI but was negatively correlated with EOSEVI at distances of less than 1500 m. This inconsistency may arise due to the varying sensitivities of the time series of NDVI and EVI to temperature differences.

Statistical analysis of the correlation of SOSNDVI and SOSEVI with spring LST metrics and the correlation of EOSNDVI and EOSEVI with autumn LST metrics (Fig. 9) show that the number of statistically significant pixels (% of total) decreased with distance to compact high-rise (i.e., LCZ 1, the central urban patterns), which confirms that the changing spring and autumn LST and SUHI affect vegetation phenology and thus the ecosystem. These effects are more pronounced in urban domains than in suburban and rural regions. Nevertheless, our findings suggest that LSP could also be employed as an indicator or proxy of SUHI intensity and its effect on the urban ecosystem based on the LCZ gradients.

Urban–rural gradients and variations in vegetation phenology might be related not only to SUHI gradients and LCZ types but also to elevational gradients, which have been widely reported in naturally vegetated areas (Xie et al. 2017, 2020). Nevertheless, our findings indicate a more pronounced sensitivity of urban ecological processes to the SUHI and LST effects associated with distance to centres of built areas than that of suburban and rural areas in the RPD region. In addition, spatial differences in the effects of SUHIs on the LSP of vegetated areas are especially pronounced in urban domains with higher built proportions in the PRD region.

Because of the relatively higher sensitivity of urban vegetation than suburban and rural areas in the PRD region (Fig. 9), our results suggest that the direction and strength of the response of urban vegetation to SUHIs depend on the distance of vegetated areas to the urban centres and compact and high-rise buildings (of distance gradients between 0 and 2000 m) over the study period.

Limitations and further research

The urban composition characteristics and medium spatial resolution of remote-sensing-derived vegetation indices in the PRD region make LSP metrics estimates complex (Li et al. 2017a, 2019). Over the observed period, missing data might have created bias in the curve-fitting estimation and the midpoint threshold determination, increasing uncertainties in the computation of vegetation phenology metrics (Fisher et al. 2006; Hmimina et al. 2013). A significant concern in heterogeneous urban areas is that the use of coarse satellite images may introduce bias when estimating vegetation phenology due to the greater diversity of phenology dates within a pixel (Tian et al. 2020). Consequently, finer resolution data holds the potential for greater biophysical accuracy compared to its coarse counterparts (Gao et al. 2021). Previous findings indicate that finer resolution datasets enable effective assessment of vegetation phenology at the landscape/ecosystem level (Gu et al. 2023) and even at the tree crown/canopy level (Zhao et al. 2022; Cao et al. 2023). While finer resolution datasets may contribute to more credible conclusions in heterogeneous urban regions, it is essential to acknowledge the potential for differences. Future studies should undertake comprehensive comparisons among these datasets in real-world conditions to provide a more nuanced understanding. Validation of remote-sensing modeled phenology at a fine resolution based on ground observations of phenology is also necessary in heterogeneous urban areas.

Heterogeneity of vegetation within LCZs could result in inaccuracies in LSP comparison at the vegetation type classes. Further LSP analysis should consider the lower hierarchical level of LCZs (Stewart and Oke 2012) to clarify the effect differences of LST between vegetation types. However, the LCZ classification results are also affected by various resolutions of the Earth observation dataset, such as MODIS, Sentinel-1 and 2, and DEM data. Limited by the spatial resolution of LCZ maps and LSP metrics, we cannot exclude all agricultural land covers from our investigation. Therefore, agricultural practices and management in the croplands surrounding urbanized areas, which affect the estimation of LSP metrics and their relationships with LCZs and the corresponding SUHI, were not considered in this study. This topic was beyond the scope of this research, for example, seeding and sowing in spring and harvesting in autumn. Nevertheless, LSTs are not entirely associated with air temperature (Roth et al. 1989; Arnfield 2003). Disentangling the climate and SUHI effects of agricultural activities on greening solely using Earth observation data would be difficult. In addition, SUHI assessments relying on satellite data were reported to have an upward bias over both day- and nighttime (Venter et al. 2021), which might limit the estimation confidence in analysis of the urban–rural SUHI gradient based on the LCZ scheme. To this end, further analyses involving additional auxiliary information are required. We focused our analysis of the LSP of vegetated areas on pixels that experienced no urbanization. Although we recognize vegetation transformations due to urban development and LCZ changes are going to affect LSP dynamics (Wang et al. 2019a, 2019b) such processes were beyond the scope of this study and require further research.

Remote sensing-based analysis of LSP of tropical/subtropical urban regions should focus on the relative influence of identified driving factors LSP dynamics across vegetated areas. Combining air temperature and LST in such analyses should facilitate, the detection of long-term trends in the start and end of greening and the role of precipitation, solar and artificial light, as well as urban water availability in explaining complex patterns of phenology. It should improve our understanding of the effect of urban expansion and changing climate on the vegetation in the PRD region.

Conclusion

Based on multi-source Earth observation data, classified LCZ maps and time series vegetation indices, we derived LSP metrics (SOS, EOS, and LOS) along the distance gradient to urban centres (i.e., LCZ 1 compact high-rise) in the PRD region. We found that more than 95.0% of built areas were distributed within a distance of 5000 m to urban centers, and the distances of compact and high-rise LCZ types to LCZ 1 were closer than those of open, mid-rise, and lower-rise LCZ types. The interannual trends of SOS and EOS were delayed, and the interannual trend of LOS extended over the investigated period and. The trends were pronounced in vegetated areas next to built-dominated regions (of distance gradients between 0 and 5000 m) than in other vegetated areas. Additionally, the interannual dynamics of LOS depended more on the EOS than the SOS. The influence of compact and high-rise areas on LSP in urban domains was more substantial than that of open, mid-rise, and lower-rise areas. Correspondingly, the association of LSP with LST was higher in the areas adjoining compact high-rise than in the open, mid-rise, and lower-rise areas; the effects of day LST on the LSP were slightly more substantial than those of the night LST.

This study provides a novel view in combining LSP and LCZ and reveals the spatiotemporal dependence of the vegetation on the LCZ and the LST. Taking forward the conservation of urban ecosystems in the PRD region (i.e., the Guangdong–Hong Kong–Macao Greater Bay Area), the LSP of vegetated areas along with LCZ-based urban–rural gradients might be critical for understanding vegetation phenology changes and for the planning and management of urban ecosystem services. Notably, a good understanding of urban vegetation and well-verified LSP and LST metrics is essential for vegetation protection and the resilience of ecosystems along with urban development. Our applications show the potential for using LCZ as a critical indicator in urban vegetation and its gradient identification.