Slower changes in vegetation phenology than precipitation seasonality in the dry tropics

The dry tropics occupy ~40% of the tropical land surface and play a dominant role in the trend and interannual variability of the global carbon cycle. Previous studies have reported considerable changes in the dry tropical precipitation seasonality due to climate change, however, the accompanied changes in the length of the vegetation growing season (LGS)—the key period of carbon sequestration—have not been examined. Here, we used long‐term satellite observations along with in‐situ flux measurements to investigate phenological changes in the dry tropics over the past 40 years. We found that only ~18% of the dry tropics show a significant (p ≤ .1) increasing trend in LGS, while ~13% show a significant decreasing trend. The direction of the LGS change depended not only on the direction of precipitation seasonality change but also on the vegetation water use strategy (i.e. isohydricity) as an adaptation to the long‐term average precipitation seasonality (i.e. whether the most of LGS is in the wet season or dry season). Meanwhile, we found that the rate of LGS change was on average ~23% slower than that of precipitation seasonality, caused by a buffering effect from soil moisture. This study uncovers potential mechanisms driving phenological changes in the dry tropics, offering guidance for regional vegetation and carbon cycle studies.

Understanding the response of the dry tropics to climate change is therefore critical to our assessment of global carbon sequestration ability and climate mitigation efforts.
The dry tropics are characterized by stronger seasonal and interannual rainfall variation than the moist tropics.They are located in the outer layer of the Inter-Tropical Convergent Zone rain belt, occupying ~40% of the tropical land surface.The ecosystems in the dry tropics (e.g., deciduous forests, shrublands, and grasslands) have visible seasonal variations due to periodic leaf shedding from water stresses, in contrast to the evergreen rainforest in the moist tropics (Xu et al., 2022).Multiple studies have reported that rainfall regime-characterized by the total precipitation (Huber et al., 2011), intra-annual (Good & Caylor, 2011;Knapp et al., 2002) and inter-annual (Gherardi & Sala, 2015) variability of precipitation amount-is a main controlling factor for vegetation growth and land-atmosphere carbon exchange in the dry tropics.
However, one essential aspect of the rainfall regimes (i.e., precipitation seasonality) has rarely been assessed for its impacts on vegetation dynamics, though several case studies have suggested strong correlations between the timing of rainfall and dry tropical vegetation phenology (Guan et al., 2014;Philippon et al., 2005).
This issue has become more prominent as some studies have reported that precipitation seasonality in the tropics has significantly changed in the past few decades, such as lengthened dry seasons over southern Amazonia and the Congo rainforest (Fu et al., 2013;Jiang et al., 2019;Xu et al., 2022) and shortened dry seasons in the Sahel, central Africa and northern Brazil (Feng et al., 2013;Xu et al., 2022).With the changes in precipitation seasonality, we expect to see associated changes in vegetation phenology, which would further influence land-atmosphere CO 2 exchange (Churkina et al., 2005;Xia et al., 2015), and the global terrestrial carbon cycle (Ahlström et al., 2015;Poulter et al., 2014).Indeed, the impact of the length of the wet season (LWS) on plant productivity and soil carbon flux has been noted at a dry-deciduous caatinga ecosystem (Rohr et al., 2013).Therefore, it is critical to assess the historical changes in vegetation phenology across the dry tropics and their dependence on the precipitation seasonality.
Our current understanding of the changes in dry tropical vegetation phenology is inconsistent and incomplete.Several remote sensing-based studies have examined the trends of vegetation phenology metrics (i.e., the start, end and length of the vegetation growing season, or in acronyms SGS, EGS and LGS) in the dry tropics of Africa (Adole et al., 2018), South America (Medeiros et al., 2022), Southeast Asia (Suepa et al., 2016), Northern Australia (Ma et al., 2013), or across the pan tropics over a short period (Tong et al., 2019).However, they reported inconsistent results on phenological changes.For example, in the Sub-Sahel SGS was reported to have either advanced (Tong et al., 2019) or delayed (Adole et al., 2019), and there was a large difference in the magnitude of reported trends.Meanwhile, only few short-term ground phenology observations are available in the dry tropics, which are inadequate to detect phenological trends and investigate spatial variations in phenological trends (Chambers et al., 2013).The studies so far highlight two issues in studying dry tropical phenology-(i) we face substantial uncertainties in the phenology extracted from satellite data, which is partly linked to the lack of ground observations to validate satellite estimates, and (ii) there is larger spatial heterogeneity in phenology trends in the dry tropics, compared to the phenology of the mid-high latitudes (i.e., prevailing lengthened growing season) (Piao et al., 2007), and the mechanism underlying the high spatial heterogeneity is unknown.
In this study, we aim to fill in the knowledge gap on the historical change in dry tropical vegetation phenology and examine its synergies with the trend in precipitation seasonality.In particular, we will address the following questions: (i) how did the LGS and LWS change in the dry tropics over the past four decades?(ii) did the changes in LWS drive the changes in LGS? and (iii) if so, what are the potential mechanisms by which LWS impacts LGS?To do so, we selected and updated a long-term satellite dataset of Normalized Difference Vegetation Index (NDVI) (i.e., VIP15 NDVI from 1981 to 2019; see Materials and Methods) that is suitable for dry tropical phenology study, and developed a customized and robust algorithm for phenology extraction (Figure 1a).We validated satellite-derived phenology metrics against 56 site-years of carbon flux records from 14 eddy covariance tower sites in the dry tropics (see Section 2 and Table S1).
Precipitation seasonality and soil moisture seasonality were obtained from multiple sources of climate datasets (see Section 2 and Figure 1b,c) for examining their synergies against vegetation phenology.Our study provides a pan-tropical view on how dry tropical ecosystems cope with shifts in precipitation seasonality, and critical insights into the mechanism of climate-vegetation interaction in the dry tropics.
We used a dynamic threshold method to determine three key vegetation phenological metrics, i.e., SGS, EGS and LGS (Wang et al., 2019(Wang et al., , 2022)), following the algorithm from MODIS phenology products (e.g., MCD12Q2).Specifically, we calculated the VI ratio of vegetation index (VI) time-series data annually in each pixel using Equation 1: where VI t represents the value of satellite VI time-series observations (e.g., NDVI and LAI) on a given day.We tested and employed different thresholds (i.e., from 10% to 50%) to determine vegetation phenology metrics (Figure S1).Based on validation against ground observations (Figure S2), we selected the threshold of 15% for our main analysis, and the 15% can also produce highly consistent results with other phenology extraction methods (e.g., curvature method) (Wang et al., 2022).Thus, the SGS and EGS were defined as the dates when VI ratio increased to 15% and decreased to 15% of the VI amplitude above the VI minimum, respectively, and the LGS was then defined as the difference between EGS and SGS (Figure 1a).Because vegetation phenological dates may span calendar years in the dry tropics, especially in the Southern Hemisphere, we applied a 3-year moving window to accurately estimate the phenological dates and avoid overlooking phenological events (Figure 1a).Consequently, the vegetation phenological metrics were not extracted for the first and last year of each record.
To determine the main satellite-based vegetation phenological metrcis for our analysis, we conducted a comprehensive intercomparison of phenological metrcis extracted from four long-term satellite remote sensing datasets (i.e., VIP15 NDVI, GIMMS NDVI3g, GLASS LAI and GLOBMAP LAI) (Figure S3a-j).We validated these metrics against ground-based phenology results obtained from eddy covariance measurements (Figure S3k-n).Additionally, we compared phenological metrics derived from four long-term and three shortterm satellite time-series products (i.e., MOD13C1 NDVI, CSIF, and VOD) over their overlapped period (i.e., 2002 to 2013) (Figures S4   and S5).Based on the intercomparison between remote sensing products and validation against eddy covariance observations (see Section 4), we selected the phenology metrics derived from the VIP15 NDVI for our main analysis.
To minimize the influence of vegetation with two growing seasons on our study, we performed a harmonic analysis to determine the number of seasons through Fourier transform of the entire NDVI time series (Dunning et al., 2016;Xu et al., 2022).Regions with a ratio greater than 1.0 were identified as having two vegetation growth seasons, accounting for only 9.5% of the dry tropics (Figure S6a).After excluding pixels with two vegetation growing seasons, we further eliminated irrelevant land types, including evergreen forests (22.43% of all pixels), urban and bare areas (16.23%), and mosaics (11.71%), classified by ESA CCI land cover classification (Figure S6c).To mitigate the effects of land cover changes over the past few decades, we removed pixels with identified land cover changes according to the ESA CCI land cover classification.Additionally, we excluded sparse vegetation pixels (i.e., annual NDVI mean < 0.1), and those showing weak seasonality (i.e., multi-year mean NDVI max < 0.2 and annual NDVI max less than 1.2 times of the NDVI min ), following a protocol established in previous satellite-based phenology studies (Shen et al., 2014).Finally, we excluded pixels with less than 50% clear observations each year to eliminate potential contamination from cloud cover on NDVI signals.

| Eddy covariance carbon flux phenology
To validate the results of satellite-derived phenology detection, we employed ground-based eddy covariance measurements from the FLUXNET2015 dataset (Pastorello et al., 2020).We identified 14 available flux sites in the dry tropics with a total of 56 siteyears records (Table S1; Figure S6c).We used daily gross primary productivity (GPP) from the night-time partitioning method to determine SGS and EGS.Although GPP is not equivalent to NDVI, previous studies have widely reported that they correlate quite strongly (Huang et al., 2019).We applied the same extraction method to GPP time series as satellite-derived vegetation phenology.It is essential to note that phenological metrics extracted from GPP reflect the dynamics of ecosystem productivity, while those from NDVI represent the dynamics of canopy structure.
Although ecosystem productivity and canopy structure are often highly correlated, and NDVI is frequently used as proxy for productivity (Piao et al., 2014), there are instances where the slower development in leaf physiology might decouple the two, particularly during seasonal transition periods (Croft et al., 2015).We acknowledge that this decoupling is commonly observed in temperate and boreal forest ecosystems.However, for the dry tropics, where changes in leaf physiology are relatively small or progress in pace with canopy structure (Luo et al., 2019), the decoupling of productivity and NDVI is expected to be minimal.

| Precipitation seasonality
In this study, two independent precipitation datasets (i.e., gridded CRUJRA V2.1 and satellite-based TRMM 3B42) were used.(1) CRUJRA V2.1 dataset  has a six-hourly measurement frequency and 0.5° spatial resolution, which was constructed by regridding data from the Japanese Reanalysis data (JRA), incorporating 10 meteorological variables (Harris et al., 2020).et al., 2007).To simplify our analysis, we initially excluded pixels with two dry seasons (9.5% of the dry tropics) using the harmonic method mentioned in the vegetation phenology section (Figure S6b).
Subsequently, we obtained three precipitation seasonality metrics, i.e., the start, end and length of the wet season (i.e., SWS, EWS, and LWS, respectively).The method to define SWS and EWS is as follows: we first calculated the cumulative daily precipitation anomaly using Equation 2: where Q i represents the daily mean precipitation for the calendar year, ranging from 1st January to 31st December (d), and Q is the mean precipitation from all months in all years (Dunning et al., 2016;Jiang et al., 2019).The dates of C min and C max correspond to the beginning and end of the climatological water season, respectively.Subsequently, we calculated the daily cumulative precipitation anomaly using Equation 3: where R i represents the precipitation on date j from d s − 60 to the considered date (D).A(D) is calculated for each day from d s − 60 to d s + 60 for each year.Finally, the dates of A min and A max were determined as SWS and EWS, respectively, and LWS represents the difference between SWS and EWS (Figure 1b) (Dunning et al., 2016;Jiang et al., 2019).The length of the dry season (i.e., LDS) is equal to the remaining days of LWS in each calendar year.As climatological wet seasons can potentially span across calendar years, the metrics of precipitation seasonality are not determined for the first and last year of our study period.

| Soil moisture seasonality
We used version v07.1 of the ESA CCI soil moisture product, which reports soil water content in the 0.5-2 cm soil layer from 1978 to 2019, to retrieve soil moisture seasonality.The ESA CCI soil moisture product contains daily soil moisture time-series retrievals (SM) from active microwave, passive microwave, and combined active-passive sensors on a 0.25° spatial grid (Gruber & Scanlon, 2019).Three key soil moisture seasonality metrics, including the start, end and length of the soil moisture wet season (i.e., SSWS, ESWS and LSWS, respectively), were determined using the same method as vegetation phenology extraction.SSWS and ESWS were defined when the SM ratio increased to 50% and decreased to 50%, respectively, and LSWS was determined as the difference between ESWS and SSWS (Figure 1c).The choice of a 50% threshold allows for close alignment of soil moisture seasonality dates with those of precipitation seasonality but does not affect the interannual variations of extraction results (Figure S7).
The length of the soil moisture dry season (LSDS) is equal to the remaining days of LSWS in each calendar year.

| Vapor pressure deficit seasonality
To assess vapor pressure deficit (VPD) seasonality, we first calculated the VPD time series (Shaman & Kohn, 2009) following the Equations 4-6: where e s is the saturation vapor pressure and T i is the air temperature (unit: K) on a given day i. e is partial water vapor pressure, while q and p are specific humidity and atmospheric pressure, respectively.The air temperature, specific humidity, and atmospheric pressure are derived from the CRUJRA V2.1 dataset.Subsequently, we extracted VPD seasonality using the same extraction method and threshold as applied to soil moisture seasonality.

| Vegetation phenology from dynamic global vegetation models
To evaluate the effectiveness of current vegetation models in detecting phenological changes in the dry tropics, we extracted the phenological metrics from monthly LAI estimates of 12 dynamic global vegetation models (DGVMs) in TRENDY v9 project (Table S2) (Friedlingstein et al., 2022)

| Analyses
To assess the ratio of the LWS accounting for the LGS for each pixel in the dry tropics, we calculated Wet ratio using Equation 7: (2)  (Sabatini et al., 2022).Due to inconsistent spatial resolutions among the various datasets, we resampled all metrics and results into 0.5° using the pixel aggregation (PA) method for further analysis.We used t-test to analyze whether the differences in the three key plant hydraulic traits are significant between the C-wet and C-dry regions.

| Spatial patterns of vegetation phenology and precipitation seasonality in the dry tropics
Across the dry tropics, multi-year average of LGS varied from 180 to 310 days, while LWS varied from 60 to 210 days (Figure 2a,c,g).
The average LGS and LWS exhibited similar spatial patterns-i.e., southern South America and central Africa had relatively longer LGS and LWS while the Sahel and India showed relatively shorter LGS and LWS (Figure 2a,c).Although LWS was generally shorter than LGS (Figure 2g), we found that 30% to 70% of the vegetation growing season overlapped with the wet season period (referred to as Wet ratio hereafter, indicating the proportion of LGS in the wet season) (Figure 2e).The Wet ratio demonstrated a clear latitudinal gradient- the ratio was the largest near the equator and gradually decreased towards the mid-latitude (Figure 2e).We tested seven commonly used satellite remote sensing products, including NDVI, LAI, CSIF and VOD (Figure S4), and thoroughly assessed the phenological metrics derived from them against ground-based phenology derived from eddy covariance data (Figure S3k-n).Our results revealed that the phenological metrics derived from the VIP15 NDVI time series (i.e., main dataset used in this study) had the best agreement with ground observations (R 2 = 0.66, p ≤ .01 for SGS and R 2 = 0.45, p ≤ .01 for EGS; Figure 2i), and demonstrated higher cross-sensor consistency and less artefacts than other products (Figure S4).
Over the past 40 years, we found that 54.4% of dry tropical vegetation showed an increase in LGS (18.0% had significant positive LGS trends, p ≤ .1,Mann-Kendall test), while 45.6% showed a decrease in LGS (12.6% had significant negative LGS trends, p ≤ .1)(Figure S8a).
The LGS trends varied from −9.70 to 9.98 days/decade (mean: −0.28 days/decade), with India and the belt above the Congo Basin exhibiting the largest increasing trend and Central-Southern Africa and Australia displaying the largest decreasing trend (Figure 2b).In parallel, we found that 71.7% of the dry tropics showed an increase in LWS (20.4% had significant positive LWS trends, p ≤ .1),while 28.3% of the region had a decrease in LWS (2.9% had significant negative LWS trends, p ≤ .1)(Figure S8b).The LWS trends ranged from −8.94 to 13.33 days/decade (mean: 2.08 days/decade), with almost all of Africa and India showing an increasing trend, while the largest decreasing trend was in the southern Amazon and eastern and western Australia (Figure 2d).Overall, the trends in LGS were smaller than those of LWS (Figure 2h).
To examine the divergence in the direction of LGS trends, we divided the dry tropics with significant trends (p ≤ .1,Mann-Kendall test) based on the consistency of trend directions of LGS and LWS (see Section 2).25.6% of the dry tropics had LGS trend direction that was consistent with the LWS trend direction (noted as C-wet regions; Figure 2f), while 17.9% showed LGS trend direction that was consistent with the trend direction of LDS or opposite to LWS trend direction (noted as C-dry regions; Figure 2f).The existence of C-dry regions suggests that increases in LWS did not necessarily lead to increases in LGS in the dry tropics.

| The direction and magnitude of trends in vegetation phenology and precipitation seasonality
As the vegetation in C-wet and C-dry regions showed contrasting changes in LGS in response to changes in LWS, we hypothesized that this inconsistency was related to the vegetation types in the two regions and their long-term adaptation to precipitation, which can be described by the Wet ratio .We found a significant difference (p ≤ .001,t-test) in the Wet ratio between C-wet and C-dry regions (Figure 3).
Specifically, we found that a region was more likely to be classified as C-wet if the Wet ratio was greater than 51%, indicating that for regions where most of the growing season (>51%) is in the wet season, the direction of LGS trend tends to be consistent with the direction of LWS trend.The Wet ratio thresholds to separate C-wet and C-dry regions varied across PFTs.They were 55% for deciduous broadleaved forests (DBF), 52% for croplands rainfed (CLR), 52% for shrublands (SHL) and 45% for grasslands (GRL) (Figure 3b,d-f).Meanwhile, for the croplands irrigated (CLI), there was no significant distinction between the Wet ratio of C-dry and C-wet regions (Figure 3c).This was expected as the irrigation changed local water supply and invalidated our assumption of natural vegetation adaptation to longterm precipitation regimes.Our result also indicated that PFTs alone cannot explain the difference in trend direction, as within each PFT, there existed both C-wet and C-dry portions.
We examined the magnitudes of LGS trends and LWS/LDS trends and further analyzed whether their pairwise differences are significant.We found that for the C-wet and C-dry regions of all vegetation types except CLI, the trends in the LGS and LWS/LDS were significantly different (p ≤ .05,t-test) (Figure 3g-l), and the trends in vegetation phenology were smaller than that of precipitation seasonality (Figure 3g). the dry tropics, the change in LGS was on average 23.1% slower than that of LWS/LDS.

| Comparison of vegetation phenology, precipitation seasonality and soil moisture seasonality
The slower changes in LGS than LWS/LDS resulted from the slower changes in SGS and EGS than their counterparts in precipitation seasonality (Figure 4; Figure S9).For the C-wet regions, we found that The result suggested that all phases of vegetation phenology show a degree of resistance to precipitation seasonality.This resistance implied a potential process that buffers the variation in water supply from the shifted precipitation seasonality, which directed us to examine the changes in plant water reservoir-soil moisture.
Soil serves as the medium through which plants take up water and has been reported a close relationship with plant growth in recent studies (Liu et al., 2022;Lozano-Parra et al., 2018).We used a remote sensing-based soil moisture product to extract key metrics of soil moisture seasonality (see Section 2 and Figure 1c), includ- where the changes in vegetation phenology were insignificant (i.e., the rest of the dry tropics we studied, accounting for 56.5% of the dry tropics).We found that there were also no significant changes (p > .05) in the precipitation seasonality and soil moisture seasonality in those regions (Figure S10), indicating that the lack of phenology trend is associated with the absence of the trend in the precipitation seasonality.

| DISCUSS ION
In this study, we examined vegetation phenological changes across the dry tropics over the past 40 years, using long-term satellite observations and a robust algorithm for detecting dry tropical vegetation phenology (Figure 1a).We found that vegetation in the dry tropics show divergent changes, with 18.0% of the dry tropics showing LGS increased significantly and 12.6% showing LGS decreased significantly (Figure S8a).The changes in LGS cannot be exclusively explained by previously reported changes in the LWS/LDS (Fu et al., 2013;Jiang et al., 2019;Xu et al., 2022), as we found that the direction of LGS change was also dependent on the long-term adaptation of vegetation to local precipitation, and the magnitude of LGS change was smaller than that of LWS/LDS.Additionally, we identified a potential role of soil moisture in buffering the rate of response of dry tropical vegetation phenology to changes in precipitation seasonality (Figure 4).
Considering the numerous satellite datasets and algorithm thresholds available for large-scale vegetation phenology studies, we tested an ensemble of seven remote sensing vegetation products (Figures S3 and S4) and employed five different algorithm thresholds to extract vegetation phenology metrics in the dry tropics (Figure S1).To select the main remote sensing product for our analysis, we compared phenological metrics derived from four long-term remote sensing products and validated their results using ground observations (Figure S3).We found that the phenological results derived from the VIP15 NDVI had the highest agreement with ground observations (Figure S3k-n).Additionally, to select the suitable threshold for dry tropical phenology extraction, we tested five thresholds (i.e., 10%, 15%, 20%, 30% and 50%) on the VIP15 NDVI dataset and examined incurred uncertainty.We found that the trends of LGS were similar when we used either 15% threshold (i.e., the threshold used in our main analysis) or 10%, 20%, and 30% thresholds to extract phenological metrics (Figure S1g-i).While there was a slight upward shift in the LGS trend from −0.28 days/ decade (when using the 15% threshold) to 1.66 days/decade (when using the 50% threshold) in the mean trend values (Figure S1f), this slight change did not alter the direction of trends (Figure S1b,e).
Additionally, we compared the phenological results derived from remote sensing and ground observations using the 15% and 50% thresholds (see Section 2).We found that the phenological metrics derived from 15% (i.e., R 2 = 0.66, p ≤ .01 for SGS and R 2 = 0.45, p ≤ .01 for EGS) are better correlated with ground observations than those of 50% (i.e., R 2 = 0.49, p ≤ .01 for SGS and R 2 = 0.04, p > .05for EGS) (Figure S2).Based on the ground validation and the intercomparison between datasets and thresholds (Figures S1-S4), we selected the phenological metrics derived from the 15% threshold of VIP15 NDVI for our main analysis.
To assess whether a dataset is merged or not as a determining factor impacting the LGS trend detected, we compared the phenological trends of all long-term and short-term remote sensing products under their overlapped time frame (i.e., 2003-2013), and found that the phenological metrics derived from MOD13C1 and the merged VIP15 NDVI showed the highest agreement in spatial correlation (i.e., pixel-wise correlation of LGS trends derived from the two products; correlation coefficient: r = .76)among all products (Figure S5).Meanwhile, both the non-merged GIMMS product and the merged GLASS product showed lower agreements with MOD13C1, that is, r = .40and r = .27,respectively.Therefore, merged data from different sources might not be a key factor determining the accuracy of phenology in the dry tropics, while the smoothing algorithm (i.e., for GLASS) (Ma & Liang, 2022) and the previously reported sensor degradation issue (i.e., for AVHRR used in GIMMS after 2000) (Wang et al., 2015;Zhang et al., 2013) could be key potential factors.
Previous studies have uncovered the critical role of long-term mean precipitation seasonality in shaping the vegetation distribution in the tropics (Ciemer et al., 2019;Hirota et al., 2011) and recognized that long-term mean precipitation conditions are critical to the photosynthetic seasonality of tropical forests (Guan et al., 2015).
Our results further advanced the understanding of precipitation in determining tropical vegetation phenology.Interestingly, we found regional differences in how vegetation phenology responded to precipitation seasonality (i.e., C-wet versus C-dry) (Figure 2f).For the C-wet regions, LGS increased with the increase in LWS, while for the C-dry regions, LGS decreased with the increase in LWS (or decrease in LDS) (Figure 5g,h).In particular, the existence of C-dry regions seems to defy the presumption on how wet season impacts phenology, as the increases in LWS and wet-season precipitation and soil moisture (Figure 5b,e) did not lead to increases in LGS (Figures 4f   and 5h).Interestingly, we also noted that the SGS normally occurs earlier than that of the SWS, while the EGS occurs later than the EWS (Figures 4 and 5g,h).The observation was consistent with previous reports-many species initiate their growth in the middle or at the end of the dry season and shed their leaves well into the dry season (Seghieri et al., 2012).That is because a small rainfall event of 10-20 mm in the dry season is enough to initiate growth even before the start of wet seasons in the dry tropics.
Recent evidence highlighting the significance of dry-season precipitation on vegetation activities in the dry tropics (Wang et al., 2018;Zuidema et al., 2022)  The difference in the main sources of water supply between C-wet and C-dry regions can lead to the development of different plant functional traits (Vico et al., 2015).In-situ studies and experiments have previously reported variations in species composition (Fauset et al., 2012;Prevéy & Seastedt, 2014) and plant hydraulic traits (Xu et al., 2016) between ecosystems adapted to wet and dry conditions.Here, we briefly examined three key traits relevant to ecosystem water use-effective plant rooting depth (Yang et al., 2016), isohydricity (Konings & Gentine, 2017) and plant alpha diversity (Sabatini et al., 2022) (Figure 6; Figure S11).
We found significant differences in isohydricity and plant diversity between C-wet and C-dry regions, but no significant differences in effective plant rooting depth (Figure 6).C-wet regions exhibit higher isohydricity (i.e., more anisohydric) than C-dry regions (i.e., more isohydric) (Figure 6b), suggesting that the C-wet regions impose weaker stomatal control on transpiration than the C-dry regions (Novick et al., 2019).This difference allows C-wet vegetation to use water aggressively for growth, leading them to have their growing season during water-abundant wet seasons (Figure 5g).In contrast, C-dry vegetation uses water conservatively to better suit water-limited conditions.Therefore, their growing season occurs during the dry season, relying on dry-season rainfall (Figure 5h).
In fact, a recent study in the extra-tropical Northern Hemisphere (Wu et al., 2022) demonstrated that vegetation adopting an isohydric water use strategy (i.e., strict regulation of water loss) tended to senesce earlier in autumn during droughts.This mechanism aligns with the behavior of C-dry regions in our study, which are isohydric (Figure 6b) and exhibit earlier EGS with decreased precipitation in the dry season (Figure 5d-f).Additionally, we found that the C-wet regions have higher biodiversity (i.e., species richness) than C-dry regions (Figure 6c), which is consistent with previous studies suggesting an increase in biodiversity with total precipitation (Yan et al., 2015).
To assess the ability of current vegetation models in capturing phenological changes in the dry tropics, we examined phenological trends estimated from the monthly LAI estimates of 12 DGVMs (see Section 2 and Table S2) (Friedlingstein et al., 2022;Sitch et al., 2015).We found a substantial inter-model variation in phenology trends (Figure S12), with r (i.e., correlation coefficient) between models ranging from −0.08 to 0.14 (Figure S13).phenological trends derived from the remote sensing product (i.e., VIP15 NDVI) and DGVMs exhibited low agreement, with r ranging from −0.04 to 0.28 (Figure S13).The disagreements among models as well as between models and remote sensing suggest that for accurately detecting dry tropical vegetation phenology, current vegetation models lack considerations of essential mechanisms, such as plant isohydricity and the diversity of hydraulic traits (Xu et al., 2016).
Across the dry tropics, we found that vegetation phenology changed more slowly than precipitation seasonality for most PFTs (e.g., CLR, DBF and SHL), however, we noted that CLI and GRL showed slightly different results.In particular, we found that GRL shows a faster change in vegetation phenology compared to precipitation seasonality, especially in the C-dry regions.Meanwhile, the trends of LGS (−9.35 ± 8.04 days/decade) and LDS (−6.88 ± 4.49 days/ decade) for GRL are the most pronounced across all PFTs (Figure 3l).
Unlike DBF and SHL, GRL often possesses shallower root systems (Figure S11a-f), limiting its access to deep soil moisture and making it more sensitive to changes in precipitation (Currier & Sala, 2022).It may lead to a rapid change in vegetation phenology.Previous studies have reported that GRL productivity is responsive to rainfall intensity and variability (Gherardi & Sala, 2015).Thus, we suspected that the trends in rainfall intensity and variability may have a stronger impact on GRL phenology than rainfall seasonality, especially considering that most GRL are in the C-dry regions (Figure 3l).Additionally, since we did not distinguish pasture from natural grasslands in our land cover classification (Figure S6c), it is possible that intense grazing on GRL (Erb et al., 2007;Fetzel et al., 2017) could also contribute to some of the impacts on vegetation phenology detection.We also noted that the phenology of CLI was not impacted by precipitation seasonality in the same way as CLR (Figure 3h,i).Considering that the water supply to CLI is modulated by irrigation, the additional water supply buffers the shift in precipitation seasonality and ensures crops have sufficient water to grow across the year.Previous studies have also suggested that the timing of rainfall is critical to the phenology of rainfed crops (Fukai, 1999), and irrigation can make crop growth less susceptible to changes in future rainfall (Ishaque et al., 2023).
While previous studies suggest that the prevailing trend of precipitation seasonality in the tropics is the lengthening of the dry season and the shortening of the wet season (Fu et al., 2013;Jiang et al., 2019), our study suggested that this may often not be the case for the dry tropics.We found that in over 71.7% of the dry tropics there was an extension of the LWS, and in 28.3% there was a decrease in the LWS (Figure S8b).A closer examination, however, reveals that our results align with previous studies.We also found that the Southern Amazonia and Congo basins are hotspots of the LWS decrease (Fu et al., 2013;Jiang et al., 2019), while the Sahel is a hotspot of LWS increase (Xu et al., 2022).The primary region where our study diverges from a recent pan-tropical study on precipitation seasonality (Xu et al., 2022)

is central and southern
Africa which we found a dominance of increases in LWS.However, a station-based analysis indicated that LWS in central Africa also increased from 1930 to 1990 (Feng et al., 2013).Due to a high level of inconsistency across precipitation products regarding the spatial pattern of LWS trends, we used an independent satellite-based precipitation dataset (i.e., TRMM) to validate the precipitation seasonality extracted from the gridded CRUJRA dataset (i.e., the main precipitation dataset in our study).We found a high level of agreement between TRMM and CRUJRA regarding the spatial variation of mean LWS (R 2 = 0.85, p ≤ .01)and LWS trend (R 2 = 0.17, p ≤ .01)during 1999-2018 (Figure S14).Additionally, we note that apart from the timing and duration of precipitation seasonality, some studies have also reported changes in rainfall concentration (e.g., how uniformly rainfall is distributed across months) (Feng et al., 2013).This aspect of change in precipitation seasonality may also lead to alterations in vegetation dynamics such as shifts in rainfall concentration towards or away from the growing seasons, thus warranting further examination.
Soil moisture, unlike discrete rainfall events, can provide a relatively continuous water reservoir for plant use.Many recent studies have demonstrated the importance of soil moisture in modulating vegetation growth at seasonal and inter-annual scales across various biomes (Lozano-Parra et al., 2018;Méndez-Barroso et al., 2009).
Some studies have even directly reported that soil moisture has positive impacts on vegetation phenology.For example, soil moisture controls the magnitude and direction of SGS response to warming (Liu et al., 2022), and increased soil moisture deficit results in a significant negative trend for the vegetation growing season (Luo et al., 2021).However, we also noted that in the dry tropics, soil moisture seasonality alone was inadequate to explain vegetation phenology.This limitation arises because soil moisture seasonality changed considerably more slowly than vegetation phenology (Figure 4), raising questions about phenology models in certain land surface models (Hufkens et al., 2016;Kim et al., 2020;Sutanudjaja et al., 2014) that are solely based on soil moisture dynamics.Instead, we proposed considering the joint role of precipitation seasonality and soil moisture seasonality in explaining dry tropical vegetation dynamics.Additionally, we noted that VPD (i.e., vapor pressure deficit) seasonality, a factor controlling stomatal openness and photosynthesis (Green et al., 2020;Yuan et al., 2019), also showed slower trends than precipitation phenology in some cases (Figure S15).
However, the buffering effects from VPD are not as prevalent as those from soil moisture.

| CON CLUS ION
In conclusion, we found that 54.4% of the dry tropics show an in- is, VIP15 NDVI, GIMMS NDVI3g, GLASS leaf area index (LAI) and GLOBMAP LAI, along with three short-term satellite time-series products, that is, MOD13C1 NDVI, contiguous solar-induced fluorescence (CSIF) and vegetation optical depth (VOD) to study the change in vegetation phenology across the dry tropics.(1) VIP15 NDVI dataset (1981-2014) has a spatial resolution of 0.05°.It harmonized the daily measurements of Advanced Very High Resolution Radiometer (AVHRR) from 1981 to 1999 and Moderate Resolution Imaging Spectroradiometer (MODIS) C5 from 2000 to 2014 by the rigorous band adjustment to produce 15-day composites (Didan et al., 2015).We further extended the dataset to 2019 using MODIS NDVI and the VIP15 algorithm, generating a new 40-year long NDVI series for phenology study.(2) GIMMS NDVI3g dataset (1982-2015), which has a 15-day temporal resolution and 1/12° spatial resolution, was exclusively produced from daily AVHRR surface reflectance (Pinzon & Tucker, 2014).(3) GLASS LAI dataset (1981-2018) has 8-day and 0.05° resolutions.It is a combination of AVHRR LAI (1981-1999) and MODIS LAI (2000-2018) using the bidirectional long short-term memory (Bi-LSTM) ) MOD13C1 NDVI dataset (2000-ongoing) has 0.05° and 16-day cloud-free resolutions.(6) CSIF clear-sky dataset (2000-2020) has a spatial resolution of 0.05° and a temporal resolution of 4 days, which was produced by surface reflectance from MODIS and SIF , we resampled each satellite time-series dataset into a uniform 0.05° grid and 15-day resolution using the Nearest Neighbor (NN) and Maximum Value Composite (MVC) methods, respectively.To further remove the impacts of pixels contaminated by clouds, we applied the iterative Savitzky-Golay F I G U R E 1 The concept diagrams illustrate the extraction algorithms for vegetation phenology, precipitation seasonality, and soil moisture seasonality.(a) Vegetation phenology metrics including the start of the vegetation growing season (SGS), the end of the vegetation growing season (EGS), and the length of the vegetation growing season (LGS); (b) precipitation seasonality metrics including the start of the wet season (SWS), the end of the wet season (EWS), and the length of the wet season (LWS); and (c) soil moisture seasonality metrics including the start of the soil moisture wet season (SSWS), the end of the soil moisture wet season (ESWS), and the length of the soil moisture wet season (LSWS) for a specific pixel (central coordinate: −66.35 E, 8.85 N).DOY, day of the year; SM, soil moisture.
(2) TRMM 3B42 dataset has a three-hourly measurement frequency and 0.25° spatial resolution, and was generated by gauge adjustment and blending microwave and infrared satellite data (Huffman (1) VI ratio = VI t − VI min VI max − VI min , For the C-wet regions, LGS increased at a rate of 2.40 ± 4.54 (mean ± standard deviation) days/decade while LWS increased at a rate of 3.49 ± 5.23 days/decade; for the C-dry regions, the LGS decreased at a rate of 3.93 ± 6.42 days/decade while LDS decreased at a rate of 4.62 ± 5.16 days/decade.On average, the changes in LGS were 1.09 days/decade slower than those of LWS for C-wet regions, and 0.69 days/decade slower than those of LDS for C-dry regions.Over the 40-year study period, these differences resulted in roughly widened gaps of 4.36 days between LGS and LWS for C-wet regions and 2.76 days for C-dry regions.We found similar patterns for all PFTs except GRL, which shows a negative trend in LGS in both C-dry and C-wet regions (Figure3h-l).The largest trend differences were observed in CLR (the difference in rate was −1.45 days/decade) for the C-wet regions, and in DBF (−2.00 days/ decade) and SHL (−1.32 days/decade) for the C-dry regions.Across F I G U R E 2 Spatial patterns of vegetation phenology and precipitation seasonality in the dry tropics.(a) Average of the length of the vegetation growing season (LGS) from 1982 to 2018, (b) trend of LGS, (c) average of the length of the wet season (LWS) from 1982 to 2018, (d) trend of LWS, (e) proportion of LGS in the wet season (i.e., Wet ratio ) and (f) distribution of C-wet regions, that is, defined as regions where the direction of LGS trend is consistent with the direction of LWS trend and C-dry regions, that is, defined as regions where the direction of LGS trend is consistent with the direction of the length of dry season (LDS) trend.For the regions with statistically significant LGS and LWS trends (p ≤ .1,Mann-Kendall test) see Figure S8.The histograms summarize the average LGS and LWS (g), and LGS and LWS trends (h).(i) Validation of the start of the vegetation growing season (SGS) and the end of the vegetation growing season (EGS) derived from the satellite remote sensing (i.e., VIP15 NDVI) against the SGS and EGS extracted from 14 ground eddy covariance tower sites (i.e., FLUXNET).The red solid lines and black dotted lines in (i) represent the best linear fit of the data and the 1:1 reference lines, respectively.**p ≤ .01.DOY, day of year.Map lines delineate study areas and do not necessarily depict accepted national boundaries.F I G U R E 3 The direction and magnitude of trends in vegetation phenology and precipitation seasonality.(a-f) Histograms and violin plots summarize the Wet ratio of C-wet regions and C-dry regions for all pixels (a), croplands rainfed (CLR) (b), croplands irrigated (CLI) (c), deciduous broadleaved forests (DBF) (d), shrublands (SHL) (e) and grasslands (GRL) (f).The solid curves of histograms are distributed fitted lines, and red fork symbols suggest the interaction points of fitted lines.(g-l) Violin plots summarize the trends of LGS and LWS/LDS of C-wet regions and C-dry regions for all pixels (g) and different vegetation types (h-l).The red and black solid lines in violin plots are median and mean values, respectively.n represents the number of sample points.D indicates the difference of trends between vegetation phenology and precipitation seasonality (i.e., D 1 = |LGS| − |LWS|; D 2 = |LGS| − |LDS|).Negative D represents a slower trend of vegetation phenology than that of precipitation seasonality, while positive D means a faster trend of vegetation phenology than that of precipitation seasonality.***p ≤ .001;**p ≤ .05;non-significance (ns): p > .05.
SGS advanced by 1.04 days/decade and EGS delayed by 1.50 days/ decade, while SWS advanced by 1.72 days/decade and EWS delayed by 1.80 days/decade.For the C-dry regions, SGS delayed by 1.86 days/decade and EGS advanced by 1.18 days/decade, while the start of the dry season (SDS) delayed by 2.37 days/decade and the end of the dry season (EDS) advanced by 1.33 days/decade.
ing SSWS, ESWS, LSWS, and the start, end and length of the soil moisture dry season (i.e., SSDS, ESDS and LSDS).Soil moisture seasonality across the dry tropics showed much more muted changes than that of precipitation seasonality (Figure4).For the C-wet regions, the trends of SSWS, ESWS and LSWS were −0.56 days/ decade (p: non-significance, hereafter ns), 0.16 days/decade (p: F I G U R E 4 Comparison of vegetation phenology, precipitation seasonality, and soil moisture seasonality from 1982 to 2018 in the dry tropics.For the C-wet regions, (a) shows the start of the vegetation growing season (SGS), the start of the wet season (SWS), and the start of the soil moisture wet season (SSWS); (b) shows the end of the vegetation growing season (EGS), the end of the wet season (EWS), and the end of the soil moisture wet season (ESWS); (c) shows the length of the vegetation growing season (LGS), the length of the wet season (LWS), and the length of the soil moisture wet season (LSWS); for the C-dry regions, (d) shows SGS, the start of the dry season (SDS), and the start of the soil moisture dry season (SSDS); (e) shows EGS, the end of the dry season (EDS), and the end of the soil moisture dry season (ESDS); (f) shows the LGS, the length of the dry season (LDS), and the length of the soil moisture dry season (LSDS).The LDS is equal to the remaining days of LWS in each calendar year, and the LSDS is equal to the remaining days of LSWS in each calendar year.The shaded error bars indicate mean ± 0.1*standard deviations for (a, b, d, and e) and mean ± 0.3*standard deviations for (c, f).The solid lines represent the best linear fit of the data.**p ≤ .01;non-significance (ns): p > .05.DOY, day of year; Slope unit, days/decade.ns)and 0.65 days/decade (p: ns), respectively; for the C-dry regions, the trends of SSDS, ESDS and LSDS were 0.85 days/decade (p: ns), 0.65 days/decade (p: ns) and −0.18 days/decade (p: ns), respectively.While vegetation phenology changes were slower than precipitation seasonality, they were faster than those of soil moisture seasonality, indicating that the muted changes in soil moisture seasonality make vegetation more resistance to precipitation seasonality changes.Furthermore, we also examined the changes in precipitation seasonality and soil moisture seasonality in regions offers a possible explanation for the existence of C-dry regions.We found a significant decrease in dry-season precipitation and soil moisture in the dry tropics (Figure 5c,f).Considering that vegetation in C-dry regions have most of their growing season coincide with the dry season, the decrease in precipitation and soil moisture could have shortened LGS.If this assumption holds true, we can infer that vegetation in C-wet regions relies on the water supply (i.e., precipitation and soil moisture) during wet seasons, while vegetation in C-dry regions relies on water supply during dry seasons (Figure 5g,h).

F
Differences in the precipitation and soil moisture between C-wet and C-dry regions.The average precipitation across the dry tropics for the (a) whole year, (b) wet season and (c) dry season from 1982 to 2018.The average soil moisture for the (d) whole year, (e) wet season and (f) dry season from 1982 to 2018.The shaded error bars indicate mean ± 0.5*standard deviations.The solid lines represent the best linear fit of the data.**p ≤ .01;non-significance (ns): p > .05.(g) and (h) are schematic diagrams showing the general changes in vegetation phenology, precipitation seasonality, total precipitation, and total soil moisture in the C-wet and C-dry regions, respectively.EGS, the end of the vegetation growing season; EWS, the end of the wet season; SGS, the start of the vegetation growing season; SM, soil moisture (slope unit: m 3 m −3 /decade); SWS, the start of the wet season; TP, total precipitation (slope unit: mm/decade).
Furthermore, the F I G U R E 6 Spatial patterns of plant physiology traits (i.e., effective plant rooting depth, isohydricity and plant alpha diversity) in the dry tropics.(a-c) are spatial distributions of effective plant rooting depth, isohydricity slope σ and plant alpha diversity in the dry tropics, respectively.Larger isohydricity slope σ means that the plants tend to be more anisohydric, and larger plant alpha diversity means higher species richness.Violin plots summarize their values of plant physiology traits for the C-wet and C-dry regions.The red and black solid lines of violin plots are median and mean values, respectively.Non-significance (ns): p > .05;***p ≤ .001.Map lines delineate study areas and do not necessarily depict accepted national boundaries.| 13 of 17 TIAN et al.
crease in LGS (significant in 18.0% of the dry tropics, p ≤ .1,Mann-Kendall test), while 45.6% show a decrease in LGS (significant in 12.6% of the dry tropics, p ≤ .1)over the past 40 years.The changes in vegetation phenology did not often align with previously reported changes in precipitation seasonality.Our findings indicate that the adaptation of vegetation water use (i.e., isohydricity) to long-term precipitation seasonality influenced the direction of the vegetation phenological trend.Moreover, soil moisture serves as a buffer, slowing down vegetation phenological changes compared to precipitation seasonality.This study contributes new insights into vegetation dynamics in the dry tropics and sheds light on how dry-tropical ecosystems may respond to future changes in hydroclimate.AUTH O R CO NTR I B UTI O N S Jiaqi Tian: Conceptualization; data curation; formal analysis; methodology; visualization; writing -original draft; writing -review and editing.Xiangzhong Luo: Conceptualization; funding acquisition; methodology; writing -original draft; writing -review and editing.Hao Xu: Data curation; methodology; resources; writing -review and editing.Julia K. Green: Methodology; writing -review and editing.Hao Tang: Methodology; writing -review and editing.Jin Wu: Data curation; funding acquisition; methodology; resources; writing -review and editing.Shilong Piao: Methodology; writing -review and editing.
The valid range of Wet ratio is from 0 to 100%.A higher Wet ratio indicates a larger proportion of the vegetation growing season occurring in wet seasons.For example, Wet ratio >50% implies that more than 50% of the growing season is in the wet season, and the dry season constitutes less than 50%.
(Yang et al., 2016)S) − max(SGS, SWS)LGS × 100 % rection) and C-dry (i.e., LGS trend direction consistent with the LDS trend direction).To analyze the differences in water use strategy between Cwet and C-dry regions, we collected three key plant hydraulic traits, that is, effective plant rooting depth, isohydricity slope σ and plant alpha diversity.(1)Effectiveplantrootingdepthdatasetata spatial resolution of 0.5° was estimated using Guswa's Carbon Cost-Benefit Model(Guswa, 2008), and revealed that tropical and subtropical regions have deeper roots (i.e., generally >1.5 m) than other regions(Yang et al., 2016).(2)Isohydricity slope σ dataset (0.25°) was produced by Advanced Microwave Scanning Radiometer for EOS (AMSR-E) satellite microwave observations (Konings & Gentine, 2017).Larger isohydricity values suggest that plants tend to be more anisohydric (i.e., weak stomatal control), while lower values indicate a tendency toward more isohydric behavior (i.e., strong stomatal control).(3) Plant alpha diversity dataset (i.e., local species richness) has a spatial resolution of 0.5° at a global scale, which was created by numerous local plant assemblages