The Coexistence of Trees, Shrubs, and Grasses Creates a Complex Picture of Land Surface Phenology in Dry Tropical Ecosystems

Abstract

The use of digital cameras to monitor vegetation phenology (phenocams) has become increasingly common as a means of ground truthing estimates of land surface phenology from Earth observation data. Whilst the relationship between phenocam and Earth Observation-derived indices of land surface phenology has been examined across many temperate land cover types, our understanding of these relationships across the seasonally dry tropics is limited. Here we examined phenological time series derived from coarse-scale MODIS and fine-scale phenocam data across four seasonally dry tropical sites in Brazil to determine their correlation and how phenological metrics derived from these time series differed. While MODIS-derived vegetation indices showed seasonal patterns, we found a poor correlation with vegetation indices from phenocams at sites with a high proportion of evergreen vegetation and a poor correlation of MODIS indices with specific vegetation components. The high spatial and temporal resolution of phenocams allowed us to demonstrate differences in phenological metrics among different components of the vegetation which were obscured in the coarser MODIS data. This study highlights the potential of phenocam data to improve our understanding of complex vegetation leaf phenology and its drivers within mixed tree–shrub–grass systems in the seasonally dry tropics. This could help improve the representation of the savanna, grass, and shrubland biomes within terrestrial biosphere models, and lead to better predictions of the impact of climate change on carbon dynamics via shifting vegetation phenology.

1. Introduction

The phenological behaviour of vegetation is an important determinant of ecosystem structure and function, and a key mediator of land–atmosphere mass and energy exchange [1,2,3]. There is a well-established link between leaf phenology and climate, where environmental factors including temperature, water availability, and photoperiod act as cues for phenological events like leaf bud burst and leaf senescence [4,5,6,7,8]. As terrestrial gross primary production (GPP) represents the largest global carbon flux [9], this provides strong justification for long-term monitoring of leaf phenology as an opportunity to track the responses of plants and ecosystems to a changing climate [10,11] and thereby potential changes to the global carbon flux [3,12]. Within temperate regions, plant phenology and environmental cues are well studied with long-term datasets [3]. Within the tropics, however, long-term phenological time series are still scarce [13,14], including in seasonally dry tropical forests and savannas [15,16,17,18], although the number of long-term studies has increased over the last decade [14].

With satellites increasingly collecting diverse data across the globe [19], it has become possible to investigate plant phenology at larger spatial scales [20]. Land surface phenology (LSP) captures these recurring changes in spectral reflectance observed via satellites [20,21]. An accurate understanding of LSP and its drivers is important, as they are incorporated into Earth system models (ESMs), which couple the land surface and atmospheric processes [2]. Earth observation (EO) instruments, such as the Moderate Resolution Imaging Spectroradiometer (MODIS) can be used to produce vegetation indices like the Enhanced Vegetation Index (EVI) and the Normalised Difference Vegetation Index (NDVI), both of which can be used to describe LSP [22,23,24].

While EO data provide insight on land surface phenology across large spatial scales, within areas of high spatial vegetation heterogeneity it has proven difficult to reliably detect phenological patterns due to the relatively coarse spatial and/or temporal resolution of EO data [25,26,27,28]. Semi-humid tropical savannas, such as the Cerrado in Brazil, often comprise a mixture of growth forms, namely trees, shrubs, and grasses, each with divergent phenological patterns [18,29]. Within-ecosystem variation in phenology can cause LSP derived from coarser spatial resolution data to inadequately represent the seasonal dynamics of the vegetation community for such sites [26,30,31]. This may partly explain why LSP within mixed tree–grass–shrub systems across the world is less well understood [30,32]. In American tropical dry forests, such as the Brazilian Caatinga, EO products may more accurately capture the phenological behaviour [32]. Within the American tropics, dry forests usually occur in areas with stronger water deficits during the dry season than tropical savannas [33]. This may drive the leafing phenology across vegetation components to be more tightly tied to water availability and thus synchronised [17,18,34]. As a result, LSP from coarser EO products may better represent vegetation phenology in these systems.

More recently, near-surface remote sensing has become increasingly widespread through the use of digital cameras or phenocams and the establishment of phenocam networks [3,11]. An important advantage of phenocams is that they allow for the collection of local data with high temporal and spatial resolution, permitting data collection on different species and vegetative components of the ecosystem [11,17,18,35,36,37,38]. In addition, phenocams are less sensitive to atmospheric effects such as pollution and cloud coverage, which can present major issues when using data obtained from satellites [3,11]. Previous studies have used phenocam data to evaluate MODIS products for temperate deciduous forests in the USA and found good agreement between the phenological timings derived from EO and phenocam data [39,40]. Within more heterogenous systems where vegetation components differ in their phenological dynamics, such as the oak and grass savannas in California, the relationship between phenocam data and EO has been shown to be less well correlated [26,41].

Phenocam data have been shown to be useful for analysing seasonal changes in GPP and net ecosystem exchange (NEE) in systems such as boreal and temperate forests and grasslands [36,38], crop fields [42], semi-arid shrublands [41,43], and seasonally dry tropical forests [17]. However, there are still limited studies of phenocams across the dry tropics, where plant phenological behaviour may be highly variable, and the characterisation and understanding of environmental drivers is restricted to few sites [15,17,44,45]. As a first step, it is imperative to evaluate the correspondence between phenological behaviour and timing as quantified by phenocams versus that quantified by EO products, to know if we can scale up our knowledge of phenological behaviour over geographic and environmental gradients to better dissect drivers.

In this study, we investigate covariation in time series of two vegetation indices, the Green Chromatic Coordinate (GCC) and Enhanced Vegetation Index (EVI), derived from EO data, and GCC derived from phenocams, as well as compare phenological metrics derived from these time series, across seasonally dry tropical vegetation. We use data from four sites in Brazil: three in different savanna landscapes of the Cerrado and one in short-statured tropical dry forest of the Caatinga [18]. Here, GCC time series were generated for different components of the vegetation, including grass, trees, and the overall vegetation community, from phenocam data across a four-year study period [18]. We compare the phenocam data with EVI data from the MODIS MOD13Q1 product and GCC data from the MODIS MCD43A4 v006 product, to investigate how phenological metrics obtained through these different methods vary. We hypothesise that the phenological behaviour of the MODIS products will match phenocam estimates well in the tropical dry forest vegetation, but not in the savanna vegetation where different components of the ecosystem have divergent phenological behaviour [18].

2. Materials and Methods

2.1. Study Sites

The four study sites are distributed across the two main seasonally dry biomes in Brazil [18]: the Caatinga site, in northeast Brazil, represents short-statured tropical dry forest within the Tropical and Subtropical Dry Forests and Thickets ecosystem functional group [46], while the three savanna sites, located in the southeastern of Brazil, represents three different Cerrado vegetation formations [47]; shrubland (campo sujo), open woodland (Cerrado sensu stricto) and closed woodland (dense cerrado) (Figure 1), within the Pyric Tussock Savannas ecosystem functional group [46]. Henceforth, these two biomes will be referred to as tropical dry forest (TDF) and savanna, respectively.

At each study site, environmental data such as rainfall, temperature, evapotranspiration, and photoperiod were also collected across the monitoring period. For the closed woodland Cerrado and Caatinga site, data were gathered at nearby eddy flux towers using a CR1000 datalogger (Campbell scientific Inc., Logan, UT, USA) to calculate averages of 30 s measurements at each 30 min interval for the monitoring period at each site. At the shrubland and open woodland Cerrado sites a u30-hobo meteorological station was installed on smaller-scale towers. The precipitation patterns for the sites differ, with the Caatinga site having a longer dry season, as well as only receiving around 20% of the precipitation of the Cerrado sites, resulting in a more constrained growing season (

Table 1. Description of study sites, adapted from Alberton et al. [18].

Site	Latitude, Longitude	Location (City, State, Region)	Elevation (m)	Biome	Monitoring Period	Mean Annual Precipitation (Monitoring Period) (mm)	Length of Dry Season (Months)
Caatinga	−9.05, −40.32	Petrolina, PE, Northeast Brazil	390	Tropical dry forest	10 May 2013 to 31 December 2015	260	8
Shrubland Cerrado	−22.26, −47.88	Itirapina, SP, Southeast Brazil	700	Savanna	28 March 2013 to 28 May 2015	1478	6
Open woodland Cerrado	−22.18, −47.87	Itirapina, SP, Southeast Brazil	700	Savanna	2 October 2011 to 3 February 2015	1478	6
Closed woodland Cerrado	−21.62, −47.63	Santa Rita do Passe Quatro, SP, Southeast, Brazil	649	Savanna	26 August 2013 to 31 October 2015	1150	6

, [18]). Across all four sites, the average annual temperature measured during the study period was higher than that of historical climatic data, and a lower-than-usual annual precipitation was also measured. This was especially pronounced at the Caatinga site where mean annual precipitation is ~510 mm based on historical data (1970–2014) but during the study period (2013–2015) was ~260 mm [18].

2.1.1. Caatinga (Tropical Dry Forest)

The Caatinga site is located at a Reserve of the Empresa Brasileira de Pesquisa Agropecuária (EMBRAPA), within a semi-arid region of Northeast Brazil (Figure 1), and is classified as a tropical dry forest, albeit a short-statured shrubby one [49,50]. The Caatinga vegetation is adapted to xeric conditions and tends to have highly deciduous vegetation, where only 0–30% of species keep their leaves during the dry season [18,51].

2.1.2. Cerrado (Savanna)

The three Cerrado sites are located in São Paulo state, South-eastern Brazil, within a humid seasonal climate with hot, wet summers and cool, dry winters [52]. The Cerrado is a mosaic of different vegetation [53], varying physiognomically from grassy, shrubby vegetation through to woodland and forest [53,54]. The three Cerrado vegetation sites studied can be distinguished by variation in woody cover, including shrubland, open woodland and closed woodland [18].

2.2. Phenocam Data

Alberton et al. [18] collected a four-year dataset of leaf phenology using digital hemispherical Mobotix Q 24 lens cameras (Mobotix AG-Germany, Winnweiler, Germany) from the Brazilian e-Phenology network [11], across the four sites, with varying lengths of data collection (Table 1). The frequency at which photos were obtained followed Alberton et al. [44], with cameras set to take five JPEG images (at 1280 × 960 pixel resolution) during the first 10 min of each hour, from 6 am to 6 pm. Cameras in the Caatinga, open woodland and closed woodland Cerrado were installed on towers at 16 m, 18 m, and 26 m above the ground, respectively, whilst the shrubland Cerrado camera was placed 4 m from the vegetation with a landscape, rather than hemispherical, field of view [11,18].

Phenocams enable the detection of leaf phenological events through analysis of changes in the red, green, and blue (RGB) colour channels of images captured by the cameras [6,36]. The RGB channels can then be used to calculate vegetation indices (VIs) such as the widely used Green Chromatic Coordinate (GCC) [3,25,37]. GCC is a normalised index, defined as in Equation (1) [11,55]:

$$\mathrm{G}\mathrm{C}\mathrm{C}={\displaystyle \frac{\mathrm{G}}{\mathrm{R}+\mathrm{G}+\mathrm{B}}}$$

(1)

where R, G, and B are the red, green and blue intensities of the image, respectively. For each field of view for a given phenocam, regions of interest (ROIs) were defined, where the percentage coverage of each ROI was determined by the percentage of pixels within an ROI with respect to the community ROI (Figure 2) [44]. Community ROIs were defined as the entire image, excluding bare soil patches, as well as the tower structure and bare sky for the camera with a landscape view in the Cerrado shrubland [18]. Deciduous ROIs represent all deciduous woody plant crowns identified within the community ROI, with Grassy ROIs including all herbaceous cover in the Cerrado shrubland (

Table 2. Proportion of vegetation identified as deciduous, grass or evergreen/semi-deciduous across the three Cerrado sites. The Caatinga site only has deciduous vegetation. List of plant species identified in the FOV of each phenocam can be found in Appendix A (Table A1, Table A2, Table A3 and Table A4).

Site	% of Community ROI Which Is Grass	% of Community ROI with Deciduous Strategy	Remaining % of Community ROI (Evergreen and Semi- Deciduous Strategy)
Shrubland Cerrado	23.7	16.2	60.1
Open woodland Cerrado	0	5.4	94.6
Closed woodland Cerrado	0	21.5	78.5

) [18]. Within the Caatinga site, all identified trees in the phenocam field of vision are deciduous, with the Community ROI additionally capturing parts of the ground layer and unidentified tree crowns.

Initial quality control of the raw images by eye allowed us to remove images where the field of view was obstructed. For each ROI within each image, GCC was calculated (Equation (1)) [18]. Due to issues with energy supply at the open woodland Cerrado site, there were periods with no images recorded. If these periods were more than seven days, an algorithm based on an auto-regressive moving average model (ARMA) was used to fill the GCC values within these gaps. This followed a three-step process: (1) selecting the optimal ARMA order using physical principles, (2) fitting ARMAs to data segments before and after missing values, and (3) interpolating gaps using a weighted combination of forward and backward predictions. See Alberton et al. [18] for further details. To minimise interference from lighting changes caused by weather, season, and time of day, the 90th percentile of GCC within a rolling three-day window was extracted, following Sonnentag et al. [37].

2.3. Earth Observation Data

To compare data from the phenocams with EO data, we used bands 1, 4 and 3 from the MODIS MCD43A4 v061 (Raytheon/Santa Barbara Remote Sensing, Goleta, CA, USA) satellite product, representing the red, green and blue channels, respectively, to calculate the GCC at its finest resolution of 500 m at daily intervals [56], using the same method as for the phenocams (Equation (1)). For each site, we used scenes from the monitoring period during which the phenocams were active (Table 1) with less than 20% cloud coverage, as higher cloud coverage can result in inaccurate estimates of land surface phenology [57]. We further excluded scenes where any of the corresponding mandatory quality bands had a quality other than zero [56]. Due to issues with the band quality for the closed woodland Cerrado site, the end of the monitoring period shifted from 31 December 2015 to 31 October 2015, as the last two months had no measurements of good quality. A visual inspection of the placement of the open woodland Cerrado phenocam revealed that it was located close to the woodland edge, adjacent to farmed cropland, with associated MCD43A4 pixel capturing >50% cropland. To ensure that the MODIS data products captured reflectance from natural vegetation only, we used the nearest neighbouring pixel to the one associated with the phenocam placement with >98% woodland coverage. The distance between the site and the edge of the nearby MCD43A4 pixel was 40 m.

We used the EVI time series from the MODIS MOD13Q1 v061 (Raytheon/Santa Barbara Remote Sensing, Goleta, CA, USA) satellite data product at its finest resolution of 250 m at 16-day intervals [58], using the same approach as the MCD43A4 regarding cloud coverage and quality checks. For the quality check, we used the layer containing information on pixel reliability, where a value of 0 indicates “good data” [59]. Due to the finer spatial scale, moving to a neighbouring pixel for the open woodland Cerrado site was unnecessary for the MOD13QI data product. The MODIS MOD13Q1 v006 product provides both the Normalised Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI). We used EVI, rather than NDVI, as previous studies have shown that EVI more closely matches the seasonal dynamics of GCC, while NDVI can show divergent end-of-season patterns compared to both GCC and EVI [25,40,60]. EVI performs better due to an optimised soil factor value and additional blue band which leads to a higher spectral sensitivity of EVI for multiple canopy leaf layers. EVI is calculated as in Equation (2) [58,61].

$$\mathrm{E}\mathrm{V}\mathrm{I}=2.5{\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{r}\mathrm{e}\mathrm{d}}{\mathrm{N}\mathrm{I}\mathrm{R}+(6\mathrm{R}\mathrm{e}\mathrm{d}-7.5\mathrm{B}\mathrm{l}\mathrm{u}\mathrm{e})+1}}$$

(2)

2.4. Data Analysis

All data analysis was conducted in R version 4.4.0 [62]. To facilitate comparison between GCC from phenocams and MODIS, as well as EVI from MODIS, each time series was scaled by converting the original value to a z-score [63]. We then fitted simple generalised additive models (GAMs) using the mgcv package [64] using only date as a smoother variable, as previous studies have shown that GAM estimates were accurate estimators of phenological patterns [8,18,65]. For each time series, we used the predicted values from the GAMs to perform a pairwise Pearson correlation test, which allowed us to quantify the correlation between the EO and phenocam time series, following Thapa et al. [66] and Tian et al. [67]. As the pairwise correlations showed large variability in how well the different vegetation products were correlated, we performed a maximum cross-correlation function (CCF) analysis to quantify the extent to which time lags were present between different phenological time series [68].

To investigate the difference in phenological metrics derived from the different VIs, we fitted individual GAMs to each growing season for each MODIS EO vegetation index (EVI and GCC) and each phenocam ROI (community GCC, deciduous GCC, and grassy GCC). For each GAM we then defined the following phenological metrics:

• Start of growing season (SOS). Represents the beginning of the growing season and the start of the green-up period. This was estimated using the first derivatives of the GAMs, where the start of the growing season was identified as the day when the model slope exceeds half of the maximum positive model slope for a continuous period of 27 days or more, using only backwards-looking data, following both Archibald and Scholes [69] and White et al. [70].
• End of growing season (EOS). Represents the end of the growing season and the end of senescence. Similarly to SOS, this was estimated using the first derivatives of the GAMs. However, the EOS was identified as the day where the model slope exceeds half of the maximum negative slope for a continuous period of 27 days or more, using only backwards-looking data.
• Length of growing season (LOS). The duration of the full growing season was calculated as the difference between SOS and EOS.
• Peak of growing season (POS). Measured as the highest value of the seasonal curve for each VI.

To enhance the reliability of the phenological metrics, we constrained the POS so that estimates had to have a z-score value greater than 0 (i.e., above the mean). Additionally, POS estimates were restricted from occurring within the first or last 15 days of the monitoring period. This measure was implemented to prevent the endpoints of the datasets from being erroneously identified as the highest seasonal values if seasons were cut short. For each phenological metric we calculated the mean and standard error (SE) across years, the latter representing interannual variability.

3. Results

3.1. Can Data from MODIS EO Products Be Used to Understand Land Surface Phenology in Dry Tropical Vegetation?

The time series for the MODIS EO and phenocam VIs, fitted with GAMs, are illustrated in Figure 3. EVI and GCC time series from the MODIS data exhibited clear seasonal cycles across all four sites. In the Caatinga and shrubland Cerrado sites, all time series were strongly positively correlated (Figure 4). In the open and closed woodland Cerrado sites, however, greater variation was observed among time series, both between MODIS and phenocam time series and among phenocam time series from the community and deciduous GCC.

In the Caatinga and shrubland Cerrado sites, the onset of the green-up period coincides well with the beginning of the wet season (Figure 3a,b). Conversely, in the open woodland Cerrado site, phenocam community GCC green-up occurs prior to the start of the wet season. MODIS EVI, MODIS GCC and phenocam deciduous GCC followed the precipitation pattern more closely (Figure 3c). In closed woodland Cerrado, phenocam community GCC, phenocam deciduous GCC and MODIS GCC indicated green-up at the start of the precipitation period, while MODIS EVI showed a lag (Figure 3d).

Generally, the GCC times series from both phenocams and MODIS EO showed a rapid green-up rate, which then either plateaued or dropped slightly and then plateaued before the senescence period began. In contrast, the MODIS EVI time series generally showed a slower green-up rate, followed by a peak and subsequent decline (Figure 3).

Pearson’s correlation analysis showed that the Caatinga phenocam community GCC exhibited the highest correlation with MODIS GCC time series, although MODIS EVI also showed a strong correlation. Both MODIS GCC and MODIS EVI were more strongly correlated with phenocam community GCC than with deciduous GCC (Figure 4a). In the shrubland Cerrado, all phenocam GCC time series correlated better with MODIS EVI than with MODIS GCC, with the highest correlation observed between phenocam grassy GCC and MODIS EVI (Figure 4b).

In the open woodland Cerrado site, the strongest positive correlation between phenocam and EO data was found between MODIS GCC and phenocam deciduous GCC, although this correlation was only moderately strong (Figure 4c). Notably, a strong negative correlation was detected between phenocam community GCC and deciduous GCC. In the closed woodland Cerrado site, the highest positive correlation between EO and phenocam data was between the MODIS GCC and the phenocam community GCC, though this was also only moderate (Figure 4d). MODIS EVI exhibited negative correlations with both phenocam community and deciduous GCC in this site.

To further assess the similarity between the time series derived from different VIs, a maximum cross-correlation function (CCF) analysis was conducted. This analysis revealed that within the Caatinga site, there was little to no lag between the MODIS EO and phenocam time series, and an overall high degree of similarity in the form of high maximum cross correlation (Figure 5a). In the shrubland Cerrado site, the MODIS GCC time series exhibited the greatest lag, trailing the three phenocam time series by 15–21 days. The MODIS EVI time series also lagged behind the phenocam time series by 6–14 days. Despite these lags, a high degree of correlation was detected among all time series, with the lowest maximum cross-correlation being 0.70 for the comparison between the phenocam community GCC and the MODIS GCC (Figure 5b).

The open and closed woodland Cerrado sites showed higher lags and lower similarity between the VI time series compared to the Caatinga and shrubland Cerrado sites. Within the open woodland, only the comparison between the two MODIS products showed a similarity above 0.5, with a 6-day lag between them. In the open woodland, the phenocam community GCC led the MODIS EVI time series by 57 days and the MODIS GCC by 78 days, while the deciduous GCC lagged 69 days behind the MODIS EVI and 24 days behind the MODIS GCC (Figure 5c).

In the closed woodland Cerrado, the GCC from community and deciduous ROIs exhibited a 3-day lag, with a maximum CCF above 0.7, while the MODIS GCC led the MODIS EVI by 18 days, resulting in a maximum CFF of 0.75. The MODIS EVI time series followed 75 days behind the phenocam GCC from community and deciduous ROIs; however, even when implementing lags, only a low degree of similarity was detected between the time series. The MODIS GCC showed fewer lags and a higher degree of similarity with the two phenocam ROIs, lagging 33 and 45 days behind the GCC from community and deciduous ROIs, respectively (Figure 5d).

3.2. Comparison of Phenological Metrics Across Sites and VIs

The phenological metrics derived from the individual seasonal GAMs exhibited similar trends for SOS as observed in Figure 4, with the season consistently starting later, after the rainy season began, in the Caatinga site compared to Cerrado vegetation sites. There were no discernible differences in the SOS estimates across VIs within the Caatinga site (Figure 6a). At all three Cerrado sites, the SOS for MODIS EVI occurred later than that of phenocam community GCC. Additionally, within both the open and closed woodland Cerrado, the MODIS GCC showed a later SOS than the one determined by the phenocam community GCC.

The EOS across the sites did not follow the same trend as SOS (Figure 6b), although the MODIS EVI in the shrubland Cerrado site did show a later EOS than the other VIs. The open and closed woodland Cerrado showed the greatest variability between estimates, mainly driven by earlier EOS for phenocam community GCC and phenocam deciduous GCC, respectively.

Due to the monitoring period, it was not possible to calculate the mean and SE for some VIs within the shrubland and closed woodland Cerrado sites. Despite the later SOS, the Caatinga site did not have a shorter LOS than the three Cerrado sites, and no clear trends were observed across or within the sites (Figure 6c).

Across all sites, the MODIS EVI exhibited a later POS than the phenocam community GCC (Figure 6d). In the closed woodland Cerrado, the MODIS EVI POS was 100 days later than the POS detected from the phenocam ROIs. Additionally, except for the open woodland Cerrado, the MODIS EVI POS was also later than the phenocam deciduous GCC. The MODIS GCC also showed a later POS than the phenocams within the shrubland and closed woodland Cerrado sites.

4. Discussion

The limited number of long-term, high-resolution near surface remote sensing measurements of plant phenology in the dry tropics has historically made it difficult to validate vegetation indices (VIs) derived from Earth observation (EO) instruments [15,17,18,71]. In this study, we compared GCC and EVI time series obtained from MODIS EO data products with GCC time series from phenocams across four seasonal dry tropical vegetation sites, using data from the Brazilian e-phenology network [11]. Across sites, we observed a mixed relationship between the phenological time series derived from phenocam and MODIS data. Sites with a predominance of woody deciduous vegetation and a more constrained growing season generally showed greater agreement between VIs derived from MODIS and phenocams, as well as higher agreement in phenological metrics. In contrast, sites with a higher proportion of evergreen trees and less constrained growing seasons exhibited lower agreement. Additionally, there is no consistent pattern across sites regarding which MODIS product aligns best with on-the-ground GCC measurements. The bias in EO MODIS products, whether it tends to estimate phenological transition dates earlier or later, varied between sites, making it difficult to identify a general trend in accuracy or errors.

4.1. Are EO MODIS Reliable Data Products to Describe Land Surface Phenology Across Dry Tropical Vegetation?

MODIS-derived VIs showed seasonal cycles at all sites, confirming their ability to track land surface phenology (LSP) in seasonally dry vegetation, consistent with findings from temperate forests [27,72], and similar vegetation types like African savannas [73] and tropical dry forests in Brazil [15,71]. However, the correlation between MODIS-derived VIs and phenocam-derived VIs varied significantly between sites and ROIs. Sites with a high proportion of woody deciduous vegetation had stronger correlations between MODIS and phenocam time series, consistent with findings from temperate deciduous forests [40,74], whilst sites with a high proportion of evergreen vegetation showed poorer agreement between MODIS and phenocam time series, following other studies in temperate areas [75,76].

At the Caatinga site, the MODIS EO time series exhibited a high degree of correlation with all phenocam ROIs. In tropical dry forests such as the Caatinga, most plant species are fully deciduous, shedding all their leaves during the dry season to reduce maintenance costs under water-limited conditions [17,18,77]. Leaf flushing in the Caatinga occurs rapidly in response to the first precipitation events of the wet season [17,18,71], closely following seasonal rainfall patterns, which is evident in both phenocam and MODIS time series (Figure 3). For the 2013–2014 growing season, the temporal resolution of the MODIS EVI appears to have been insufficient to fully capture the vegetation response to the precipitation pattern. While the other VIs showed two distinct peaks during the growing season, matching peaks in precipitation, MODIS EVI showed only one peak during the growing season. This highlights the importance of high temporal resolution data, as the MOD13Q1 EVI data product has a 16-day resolution whilst both MCD43A4 and phenocam GCC data have a daily resolution. Recent studies evaluating deciduousness across tropical vegetations using PlanetScope imagery has made advances in this area [78], suggesting that combining long time series from phenocams and high-resolution EO-derived indices may improve the accuracy of such comparisons and better validate products for LSP.

Across Cerrado sites, a complex relationship between MODIS EO and phenocam time series was found. Although a strong correlation generally was found between the different VIs and ROIs within the Cerrado shrubland, contrary to our expectations MODIS VIs showed a weaker correlation with phenocam community GCC, and the stronger correlation with phenocam grassy GCC. Given that only about 20% of the phenocam’s field of view (FOV) consisted of grass [18], it is surprising that the EO data seem to track this part of the community best. A potential explanation for this could be that whilst only ~20% of the phenocam FOV is classified as grass, the larger MODIS pixels might contain a larger proportion of grass than what is found within the phenocam FOV. If the area surrounding the phenocam FOV has a different vegetation composition than within the FOV, the size of the MODIS pixel might have a larger impact on the relationship between VIs. We found that the MODIS EVI time series showed a slightly higher correlation with the different phenocam ROI than the MODIS GCC time series, potentially due to the higher spatial resolution of the MODIS EVI data, compared to MODIS GCC, despite the lower temporal resolution. Furthermore, MODIS EVI is less influenced by soil and atmospheric reflectance compared to other VIs [25], which might also help explain why it showed higher correlations with phenocam data, as bare soil patches were not included in the ROIs. A study within African savannas showed that interannual variability in NDVI signals from EO was mainly driven by interannual variability in the seasonal growth patterns of grasses within the area [79], demonstrating how grassy layers can have a significant impact on the time series derived from EO in mixed tree-grass ecosystems. Furthermore, grass and trees within dry tropical vegetation tend to show different green-up patterns [69,79]. Alberton et al. [18] demonstrated how the leaf phenology of the grassy layer within the shrubland Cerrado is driven by different environmental factors compared to the overall woody community. This discrepancy highlights the value of separating different components of the vegetation community with different growth patterns when evaluating vegetation phenology [69].

Within the open and closed woodland Cerrado, a higher level of discrepancy was found between MODIS and phenocam time series compared to shrubland Cerrado and the Caatinga. A potential reason for this could be the higher proportion of vegetation with an evergreen or semi-deciduous leaf strategy found within these two sites. Whilst changes in colouration due to bud burst are relatively easy to detect in woody deciduous plants, this is challenging in areas with high proportions of evergreen trees, as changes in surface reflectance properties tend to be smaller and thus harder to detect [76], possibly explaining how the proportion of deciduous trees vegetation could impact the MODIS time series. The open woodland Cerrado also demonstrated a high level of discrepancy among phenocam ROIs. Previous studies have demonstrated that the deciduous part of the open woodland Cerrado responds differently to environmental cues than the overall community, with a positive relationship found between community GCC and temperature amplitude whilst the deciduous vegetation showed a negative relationship with temperature amplitude [18].

The spatial resolutions of the two MODIS products are coarser than the area captured by the phenocam FOV, resulting in the signal detected by those to be an average of likely a wide range of vegetation mixtures proportions. This could explain why the MODIS GCC and EVI time series appear to show a pattern which is intermediate between the phenocam community and deciduous ROIs within the open woodland cerrado, as well as differences between time series shown within the closed woodland cerrado. Previous studies have demonstrated how resampling MODIS data to coarser spatial scales can impact time series and estimates of phenological events [80]. For example, some vegetation types show earlier SOS in fine than in coarse resolution products, whilst other vegetation types show the opposite pattern. Determining the effect of spatial and temporal scale on phenological estimates across seasonally dry tropical vegetation was outside the scope of this study, but further studies on the effect of scale could help improve our ability to extrapolate findings across scale [81].

Differences in spectral bands between VIs also could influence their correlation and estimates of phenological events (SOS, EOS, POS). At the closed woodland site, we observed a substantial lag between the phenocam GCC and MODIS EVI time series, which led to a negative correlation and later POS for MODIS EVI compared to the other indices. Filippa et al. [82] demonstrated that across a temperate evergreen canopy, GCC began to increase when needles turned from brownish to green during the early spring, whereas NDVI first showed a drastic increase when new shoots occurred and biomass increased during late spring/early summer. Given EVI’s sensitivity to biomass and canopy structure [61], this could potentially influence the time series observed here. However, it is surprising that we do not see the same trend for both sites with a high proportion of evergreen vegetation, although that could potentially be due to vegetation heterogeneity and difference in areas captured by the phenocam FOV and the spatial resolution of the two MODIS data products as discussed above.

For vegetation sites with strong deciduous cycles and high phenological synchronicity, such as the Caatinga, MODIS VIs seem to effectively capture ecosystem phenological patterns. However, in ecosystems with heterogeneous vegetation and less synchronised phenological events, such as the Cerrado, MODIS data is less reliable for accurately representing phenological behaviour [83,84]. Another limitation for using EO products, particularly the 16-day MODIS EVI data product, is the larger uncertainty in estimation of phenological timing (Figure 3). This poses a challenge when studying phenological shifts due to climate change, as these tend to shift at a rate of a few days per decade [85,86]. Although daily MODIS data can be used to generate higher temporal resolution time series [87], similar to those of the phenology products generated from phenocam GCC data, this improvement in temporal resolution comes at a trade-off in data quality, as increased cloud contamination reduces the accuracy [88]. The launch of more recent satellites, such as Sentinel and PlanetScope, offers the potential for higher spatial and temporal resolution. However, they are still limited by cloud cover [89], and for commercial platforms like PlanetScope, accessibility and cost may further restrict their widespread use [78,90]. These factors, along with MODIS’s longer data record and pre-calculated vegetation VIs, likely contribute to the continued use of MODIS products despite their temporal and spatial limitations.

4.2. Difference in Phenological Metrics Between Sites and Sensors

Vegetation phenology remains a large source of uncertainty in Earth system models (ESMs) [2,91,92], although attempts have been made to improve the representation of leaf phenology, and its climate sensitivity, in models estimating ecosystem carbon balances [91]. Due to their global scale, ESMs often use simplified descriptions of vegetation, typically with a limited number of attributes, to define plant functional types as well as poorly capturing drivers of phenology [93,94,95]. Understanding unique characteristics of phenological patterns for different species and functional groups within the community can help improve the interpretation of LSP from EO data [96], and incorporation into ESMs.

Our study demonstrated that high-resolution phenocam data can identify distinct phenological patterns between grassy and deciduous tree vegetation components within the Cerrado shrubland site—patterns that were not detectable using coarser MODIS EO time series data. The phenocam grass GCC showed a later SOS and POS compared to the phenocam deciduous and community GCC. These findings align with prior findings from direct visual observations of grass and woody components in savanna ecosystems [69]. At the same time, the EOS occurred earlier for the phenocam grass GCC, reflecting an overall shorter growing season. These differences may be explained by the temporal niche separation hypothesis [97], which suggests that trees exploit early-season growth windows to minimise competition with grasses. Improvement of our understanding of leaf phenology and its drivers are key for better representing biomes like savannas and dry forests [14,17,18,44,45]. One limitation of standard RGB phenocams is their inability to calculate VIs that require near-infrared reflectance, such as EVI. This potentially limits the ability to reduce background noise from non-vegetated surfaces like soil or litter. To minimise such effects, we excluded visible bare soil patches from our phenocam ROIs. However, using multispectral phenocams with near-infrared channels could allow for the calculation of vegetation indices less sensitive to soil and litter reflectance, potentially improving the robustness of phenological metrics. This would also allow for more accurate comparison between satellite and phenocam-derived VIs.

The limitations of MODIS data in detecting seasonal changes in evergreen vegetation, demonstrated by later SOS and EOS than phenocam community GCC, reinforces the importance of combining ground-based and EO to improve the accuracy of phenological estimates. Within temperate evergreen forests, Richardson et al. [76] also found that MODIS EO systematically biassed toward later green-up detection. These delays may be due to changes in reflectance properties being harder to detect within evergreen vegetation [82,98]. In our study, this was particularly evident in the open and closed woodland Cerrado, which had higher proportions of evergreen vegetation compared to the shrubland Cerrado and the Caatinga [18]. Those differences likely contributed to variability between EO MODIS and phenocam-derived phenology estimates. Further studies into the specific drivers of phenology within seasonally dry tropical vegetation could help improve the currently crude representations of physiological processes of seasonally dry tropical vegetation in ESMs [99]. For example, Alberton et al. [18] showed that within the Cerrado, the overall community exhibits leaf flushing before the onset of the rainy season, similar to what has been detected within African savannas [73], indicating that vegetation models which model SOS as depending on rainfall might not accurately capture the actual phenological patterns of the vegetation.

Our findings show that modelling seasonal changes in vegetation reflectance using EVI data from the MOD13Q1 data product introduces greater uncertainties than models created from GCC time series from either phenocams or the MCD43A4 data product, likely due to lower temporal resolution. Given that phenology tends to shift by only a few days per decade [85,86], we propose the use of phenocam data to derive more precise phenological metrics to detect potential climate-driven changes. Furthermore, the high spatial resolution of phenocams allows users to define specific ROIs, such as individual vegetation components (e.g., trees versus grasses) or species [44], and thus investigate the phenology of specific parts of the vegetation, whilst still being able to examine larger ecosystem trends [36,37]. Integrating data from multiple sensors, such as Landsat, Sentinel-2, RapidEye, and Planet, could enhance the temporal and spatial resolution and extent of EO time series and potentially improve models of reflectance changes and estimates of phenological timings [100], especially if high-resolution phenocam data is incorporated.

5. Conclusions

Here, we explored if EO MODIS data products adequately represent the leaf phenology of tropical dry vegetation in Brazil. Although EO can track seasonal changes to vegetation indices within the Caatinga dry forest and Cerrado savanna vegetation, EO phenological time series tend to be associated with larger uncertainties than phenocam time series, as well as being unable to detect differences in phenological behaviour within heterogenous vegetation. This can result in large variations in the estimation of phenological metrics such as SOS and EOS, potentially impacting the detected length of the growing season. We propose that further studies into the drivers of vegetation phenology in mixed tree–shrub–grass ecosystems would benefit from using fine-scale phenocam data. It allows for the separation of grass, shrub, and tree signals within savanna systems, as well as between deciduous and evergreen vegetation (e.g., [18]), which in turn allows us to draw valid conclusions regarding the environmental drivers of phenology within these systems [17,18]. Better understanding of vegetation phenology could improve the rudimentary representations of physiological processes of seasonally dry tropical vegetation in Earth system models, improving predictions of the effect of climate change on phenology and estimations of ecosystem carbon balances.