Biological processes dominate seasonality of remotely sensed canopy greenness in an Amazon evergreen forest

Summary

• Satellite observations of Amazon forests show seasonal and interannual variations, but the underlying biological processes remain debated.
• Here we combined radiative transfer models (RTMs) with field observations of Amazon forest leaf and canopy characteristics to test three hypotheses for satellite-observed canopy reflectance seasonality: seasonal changes in leaf area index, in canopy-surface leafless crown fraction and/or in leaf demography.
• Canopy RTMs (PROSAIL and FLiES), driven by these three factors combined, simulated satellite-observed seasonal patterns well, explaining c. 70% of the variability in a key reflectance-based vegetation index (MAIAC EVI, which removes artifacts that would otherwise arise from clouds/aerosols and sun–sensor geometry). Leaf area index, leafless crown fraction and leaf demography independently accounted for 1, 33 and 66% of FLiES-simulated EVI seasonality, respectively. These factors also strongly influenced modeled near-infrared (NIR) reflectance, explaining why both modeled and observed EVI, which is especially sensitive to NIR, captures canopy seasonal dynamics well.
• Our improved analysis of canopy-scale biophysics rules out satellite artifacts as significant causes of satellite-observed seasonal patterns at this site, implying that aggregated phenology explains the larger scale remotely observed patterns. This work significantly reconciles current controversies about satellite-detected Amazon phenology, and improves our use of satellite observations to study climate–phenology relationships in the tropics.

Introduction

A fundamental unanswered question for global change ecology is the degree to which tropical forests are vulnerable to climate change. Increasingly, satellite remote sensing is being used to tackle this question by investigating how forests respond to climatic variations at multiple spatial and temporal scales (Saleska et al., 2007; Xu et al., 2011; Lee et al., 2013; Saatchi et al., 2013; Hilker et al., 2014; Zhou et al., 2014; Guan et al., 2015). Many remote sensing products – such as the moderate-resolution imaging spectroradiometer (MODIS) vegetation indices (VIs), which are spectral transformations of two or more reflectance bands – provide estimates of canopy greenness. These products are composite indices of both leaf biochemistry (leaf cellular structure, Chl content and biochemical composition) and canopy structure (leaf area, crown geometry, leaf demography) (Huete et al., 2002; Doughty & Goulden, 2008; Brando et al., 2010; Lopes et al., 2016; Wu et al., 2016, 2017). If accurate, they can reveal important mechanisms regulating the response of tropical forests to seasonal and interannual climatic variability, the same mechanisms which we rely on to validate and improve Earth system model simulations.

However, several key issues remain in the understanding and biophysical interpretation of satellite remote sensing. Recent studies of satellite observations of Amazon phenology show a significant seasonality in tropical evergreen forests (e.g. Jones et al., 2014; Bi et al., 2015; Maeda et al., 2016; Saleska et al., 2016). However, seasonal variation in leaf optical characteristics (Roberts et al., 1998; Toomey et al., 2009; Chavana-Bryant et al., 2017; Wu et al., 2017), canopy leaf area index (Brando et al., 2010; Samanta et al., 2012b), canopy leaf demography (Doughty & Goulden, 2008; Brando et al., 2010; Lopes et al., 2016) and canopy structure (Anderson et al., 2010; Tang & Dubayah, 2017) may complicate the biophysical interpretation of true canopy reflectance seasonality. Furthermore, some studies suggest that some standardized MODIS products (e.g. enhanced vegetation index, or EVI), which are processed to provide a coherent data product across space and time, may still be sensitive to seasonally varying artifacts – particularly atmospheric cloud/aerosol contamination (Samanta et al., 2010, 2012a), or sun–sensor geometry (Galvao et al., 2011; Morton et al., 2014) – that may confound a sensor's ability to accurately capture the real phenology and functional response of tropical forests to climate variability. To help untangle these confounding factors we need a complementary bottom-up approach that considers physically based canopy radiative transfer models (RTMs) and upscales observable leaf- and canopy-level properties.

Canopy RTMs, such as PROSAIL (Baret et al., 1992; Jacquemoud et al., 2009), GEOSAIL (Huemmrich, 2001; Ustin et al., 2012) and forest light environmental simulator (FLiES; Kobayashi & Iwabuchi, 2008; Kobayashi et al., 2012), which encompass the interactive effects of surface optical elements (leaf, bark, litter and soil) and canopy structural properties within a forest community, sun–sensor geometry and atmospheric radiation condition, can potentially be used to differentiate the roles of biophysical processes from potential artifacts in canopy-scale reflectance and greenness seasonality. These modeling approaches, parameterized by field-observed leaf and canopy properties, have been assessed previously across diverse ecosystems (e.g. boreal, temperate and agricultural), showing good agreements between models and observations (e.g. Verhoef & Bach, 2003; Koetz et al., 2007; Kobayashi et al., 2012; Schneider et al., 2014). However, these modeling approaches have rarely been applied and tested in tropical evergreen forests over seasonal timescales. Previous studies have investigated the sensitivity of remote sensing to certain biological effects (e.g. Toomey et al., 2009; Morton et al., 2014, 2016), or focused on retrieval of important biophysical variables (e.g. Chl concentration, Hilker et al., 2017), but few studies have integrated field observations with RTMs to comprehensively assess the contributions of competing biological processes to the aggregated canopy-scale reflectance and greenness seasonality.

Here we use canopy RTMs to connect field-observed leaf and canopy characteristics with satellite remote sensing to mechanistically interpret canopy-scale reflectance and greenness seasonality in Amazonian evergreen forests. Specifically, we examine three phenological factors that could drive seasonality in satellite-observed tropical forest canopy reflectance, including (1) the leaf area index (LAI) effect, that is, the change in reflectance due to the seasonality of LAI and light-scattering by multi-leaf-layer (Verhoef, 1984; Samanta et al., 2012b), (2) the leafless crown fraction effect, that is, the change in reflectance due to seasonal change in whole-canopy optical properties, specifically the fraction of leafless crowns, which is strongly negatively correlated with LAI seasonality (e.g. Lopes et al., 2016), and (3) the leaf demography effect, that is, the change in reflectance due to the seasonality of leaf age distributions, as leaf reflectance and transmittance show strong dependence on leaf age (e.g. Roberts et al., 1998; Wu et al., 2017).

We used two canopy RTMs, PROSAIL and FLiES. Their main difference lies in FLiES being a three-dimensional (3-D) canopy RTM, allowing for more sophisticated and realistic representation of canopy structure, relative to PROSAIL, a 1-D canopy RTM. Comparison of the two models permits an assessment of the more realistic 3-D canopy structure on canopy-scale reflectance and greenness seasonality.

To assess these phenology-related effects and to compare PROSAIL and FLiES, we used field-observed leaf and canopy characteristics collected over the annual cycle in a central–eastern Amazonian evergreen forest. These characteristics include monthly LAI and litterfall of a 1 ha control plot (Brando et al., 2010); monthly leafless crown fraction of the upper canopy derived from a tower-camera (Wu et al., 2016); field-observed leaf reflectance spectra at different leaf ages (Wu et al., 2017); and an airborne LiDAR survey of 3-D canopy structure (Stark et al., 2012, 2015; Hunter et al., 2015). PROSAIL and FLiES were parameterized and driven by the above three phenological factors separately and in combination to assess the relative contribution of each factor responsible for canopy-scale reflectance and greenness seasonality. Specifically, we posed two questions: When driven by all three phenological factors (i.e. LAI, leafless crown fraction and leaf demography), can PROSAIL and FLiES capture tropical forest canopy reflectance seasonality? What is the relative contribution of each phenological factor towards explaining canopy-scale reflectance seasonality? By answering these questions, we provide a benchmark for scaling leaf and canopy characteristics to landscapes and for broad application across multiple sites, and ultimately increase our understanding of fundamental biophysical processes that regulate tropical canopy reflectance seasonality, enabling more accurate use of existing and future remote sensing platforms in the tropics.

Materials and Methods

Satellite observations

We used satellite observations targeted at the k67 tower site where ground observations were made. The k67 site (54°58′W, 2°51′S) is located in the Tapajós National Forest, near Santarém, Pará, Brazil. It is an evergreen tropical forest on a well-drained clay-soil plateau (Rice et al., 2004), with a mean upper canopy height of c. 40 m (Hutyra et al., 2007). Mean annual precipitation is c. 2000 mm yr⁻¹ with a 5-month dry season when evapotranspiration exceeds precipitation from approximately mid-July to mid-December (Restrepo-Coupe et al., 2013).

Two kinds of satellite observations were used to evaluate RTM performance, namely a WorldView-2 (WV-2) image and two versions of the collection six time-series MODIS land surface reflectance and VI products.

The WV-2 image was acquired on 28 July 2011 (Supporting Information Fig. S1). The image has 2 m spatial resolution, 2° off-nadir view, 42° solar zenith angle and 2.6% cloud cover. It includes eight spectral bands (see Table S1 and Fig. S2 for spectral response functions). For details on data availability, image pre-processing, atmospheric correction and reflectance calculation refer to Methods S1 and Meng et al. (2017). Based on our knowledge of the study site and vegetation spectroscopy, we used a true-color RGB composite to visually identify three phenophases for fully illuminated upper-canopy crowns within the 1 × 1 km² footprint surrounding the k67 site. These were young (bright green in color), old (dark green) and leafless (largely occupied by bare branches) (Fig. S1d). Thirty crowns of each phenophase were selected, totaling at least 300 pixels per phenophase, to derive landscape average reflectance. We subsequently used the derived phenophase-specific canopy reflectance to validate the RTMs.

We primarily used the collection six MODIS Multi-Angle Implementation of Atmospheric Correction product (MAIAC; Lyapustin et al., 2012) for years 2000–2014 to investigate remotely sensed vegetation seasonality. For details on data acquisition, processing and quality control refer to Methods S2. MAIAC incorporates a new bidirectional reflectance distribution (BRDF; corrected to nadir view and 45° solar zenith angle) and strict atmospheric corrections for clouds and aerosols (Lyapustin et al., 2012). It includes four MODIS bands (blue, green, red and near-infrared (NIR); see Spectral response functions in Fig. S2) and associated VIs (i.e. normalized difference vegetation index (NDVI) and EVI; see Table S2 for equations). Here we focus on the MAIAC data because the data have become commonly used for phenology monitoring in the tropics, with several empirical validations from site-level phenology data (Lopes et al., 2016; Wagner et al., 2016; Wu et al., 2016), eddy covariance data (Jones et al., 2014; Guan et al., 2015; Saleska et al., 2016), and satellite data of other sources (Maeda et al., 2016) and alternative approaches (Bi et al., 2015; Saleska et al., 2016). We expect that the multi-year (2000–2014) average annual cycle of monthly BRDF-corrected MAIAC data with robust removal of cloud/aerosols can be used as a benchmark to assess the integrated phenology effects (i.e. the three phenological factors above) on modeled canopy reflectance seasonality. We also tested the robustness of the MAIAC results through retrievals of different target areas (3 × 3 km² and 5 × 5 km²) and through comparisons with an alternative MODIS product, the BRDF and Albedo product (MCD43A1; Wang et al., 2015). We analyzed the MCD43A1 product (5 × 5 km²) for the same sun–sensor geometry as MAIAC and two levels of quality control flags (QA3 and QA5; Methods S2). The results (Figs S3, S4) indicate that the relative seasonality of MAIAC data of 5 × 5 km² is very robust, especially in the most critical products, NIR and EVI. We observed some differences in the visible bands between the MAIAC and MCD43A1 products (Fig. S4), presumably due to corresponding differences in associated atmospheric corrections. These differences, however, did not greatly affect the overall seasonality in EVI, which was dominated by NIR and is the main focus of the study.

Ground observations

To parameterize canopy RTMs and to assess the relative contribution of the three phenological factors to canopy-scale reflectance seasonality, we used several field observations at the k67 site, including measurements of tissue optics (i.e. leaves, bark and litter) and canopy characteristics (i.e. phenology and structure).

Tissue optics

Reflectance spectra of leaves, bark and litter were measured using a portable spectrometer (ASD FieldSpec Pro; Analytical Spectra Devices, ASD Inc., Boulder, CO, USA; spectral range: 350–2500 nm; spectral resolution: 3 nm at 350–1000 nm and 8 nm at 1000–2500 nm). Spectral data were interpolated to 1 nm before analysis using the default ASD output. For each tree, we obtained spectra of leaves at young (≤ 2 months), mature (3–5 months) and old (6–14 months) age classes (Wu et al., 2017). Seven tree species with all three leaf age classes were used (see Tables S3 and S4 for species identification and number of leaf replicate): four species were from the upper canopy (accounting for 22.2% of the local basal area), and the other three species were from the 20–30 m mid-canopy stratum. Leaves were sampled to represent three incident canopy light conditions: upper canopy sunlit, upper canopy shaded and mid-canopy shaded. For details on this dataset (Fig. 1a) and the protocols used for spectral measurements and leaf age classification see Wu et al. (2017). Bark reflectance spectra were measured in 2002 by T.M. Bark samples were harvested from 13 canopy trees at c. 1.3 m above the ground (see Table S5 for species identification). These samples were kept in sealed plastic bags in a cool-box, and reflectance measurements (Fig. 1b) were made within 24 h of sampling. The litter spectra (Fig. 1b) were measured in March 2014, by randomly collecting 40 leaf litters over various locations across the forest floor. Reflectance measurements were made within 1 h after sampling.

Figure 1

Plant tissue optics at the k67 site: (a) field observed leaf reflectance of three age classes, young (n = 11 tree–canopy light conditions, see Supporting Information Table S3; ≤ 2 months, blue), mature (n = 11; 3–5 months, green) and old (n = 11; 6–14 months, red); (b) field observed bark reflectance at 1.3 m above the ground (n = 13 tree species, see Table S5; black) and litter reflectance (n = 40 tree species, see the Materials and Methods section; orange); (c) PROSPECT model inverted leaf transmittance of three age classes (n = 11 tree–canopy light conditions; see the Materials and Methods section); and (d) modeled leaf absorptance of three age classes (absorptance = 1 − reflectance-transmittance; n = 11 tree–canopy light conditions). Color lines for the mean; shading for ± 1 SD.

Together with reflectance, leaf transmittance regulates the leaf single scattering albedo and is an important component of canopy RTMs and process models (Sellers et al., 1997; Pinty et al., 2004). However, acquiring accurate and reproducible measurements of leaf transmittance is challenging, primarily due to limitations inherent to integrating sphere instrumentation (Shiklomanov et al., 2016). Instead, we estimated leaf transmittance (Fig. 1c) and subsequent absorptance (Fig. 1d) by inverting the leaf reflectance model PROSPECT (Jacquemoud & Baret, 1990; Feret et al., 2008) after it was optimized to closely match the field-observed leaf reflectance. Shiklomanov et al. (2016) showed that PROSPECT-inverted leaf transmittance is highly consistent with values obtained with an integrating sphere for fresh leaves. As a check, we compared our estimated transmittance against a set of independent measurements made in 2002 by T.M. (see Methods S3 for more details). Similar to Shiklomanov et al. (2016), we found strong agreements between measured and modeled transmittance, in terms of both mean values and standard deviation (Fig. S5a). Furthermore, similar trends across leaves of multiple species were found for field-observed leaf NIR reflectance and transmittance and for field-observed NIR reflectance and PROSPECT-inverted NIR transmittance (Fig. S5b), except that modeled transmittance was biased low in Manikara huberi leaves with high NIR reflectance (> 0.6). This might be because M. huberi has thick, waxy leaves which are not currently well represented/constrained in PROSPECT, but a more in-depth understanding is still needed. The age-dependent leaf reflectance, transmittance and absorptance for each of all 11 tree–canopy light conditions (i.e. four canopy sunlit, four canopy shaded and three mid-canopy shaded) are shown in Fig. S6.

Canopy characteristics (phenology and structure)

Three components of canopy phenology were available at the k67 site. These were, first, the mean annual cycle of monthly field-derived LAI (or LAI_field) obtained with the LAI-2000 instrument at ground level (January 2000–December 2005). See Brando et al. (2010, fig. 4) and Fig. 2(a) for more details. Second, we used the mean annual cycle of monthly leafless crown fraction (1 minus the green crown fraction; Fig. 2a) from tower-mounted camera image timeseries (January 2010–December 2011). The timeseries of the green crown fraction was derived using a camera-based tree inventory approach, and see Wu et al. (2016, fig. S8) for more details. Third, we used leaf age fractions in three age classes (Fig. 2b): young (≤ 2 months), mature (3–5 months) and old (≥ 6 months). These were derived from a leaf age demography model described by Wu et al. (2016).

Figure 2

Field derived canopy-scale phenological and structural indices at the k67 site: (a) ground measurements of mean annual cycle of leaf area index (LAI; monthly measurements from January 2000 to December 2005; green squares) and tower-camera-derived mean seasonality of leafless crown fraction (daily measurements from January 2010 to December 2011, adapted from Wu et al., 2016; and see the Materials and Methods section; black circles); (b) field-estimated seasonality of leaf age fraction for three leaf age classes (adapted from Wu et al., 2016: Fig. 3a), using a leaf demography model as in Wu et al. (2016), constrained to sum to total ground-observed LAI (green squares in a), with young (1–2 months, blue), mature (3–5 months, green) and old (≥ 6 months, red); and (c) a 2012 airborne LiDAR-estimated and gap-filled mean canopy area density (i.e. the sum of leaf and woody area density, which were assumed as a constant fraction, 0.89, for leaves; m² m⁻³) along the entire vertical canopy profile. Error bars in (a) indicate ± 1 SD; shadings in (a, b) indicate the dry season; shading in (c) indicates 95% confidence interval on the mean of gap-filled data.

Canopy structure was derived from an August 2012 airborne LiDAR survey at k67 (Hunter et al., 2015). Details of the LiDAR sensor and airborne survey are given by Stark et al. (2012, 2015). We estimated canopy area density (CAD) in 3-D canopy voxels with a 2 × 2 × 2 m³ grain, following approaches developed by Stark et al. (2012, 2015). For details on how we derived 3-D voxel data from the LiDAR survey refer to Methods S4. Additionally, a constant was adjusted to set the vertically integrated landscape-average LiDAR canopy area to match empirical estimates from LAI_field (Fig. 2a) and to extend one-time LiDAR-derived 3-D canopy structure to the seasonal scale (see Methods S5). The LiDAR-derived 3-D canopy structure is shown in Fig. S7, and the landscape average canopy height–CAD relationship is shown in Fig. 2(c).

Canopy radiative transfer models

The two canopy RTMs used are PROSAIL and FLiES. The former has the advantage of being simple and less computationally demanding, while FLiES uses more realistic 3-D canopy structure, in this case derived from airborne LiDAR. Here, we parameterized the two models (see Methods S5) to explore whether the relative contributions of the three phenological factors to canopy reflectance seasonality are consistent between the two, and to assess the impact of using a more realistic 3-D canopy structure on modeled canopy reflectance by cross-model comparisons.

PROSAIL

PROSAIL (Jacquemoud et al., 2009) is a combination of the PROSPECT leaf optical properties model (Jacquemoud & Baret, 1990) and the SAIL canopy bidirectional reflectance model (Verhoef, 1984, 1985). Details of PROSPECT and SAIL are shown in Methods S6. Coupling of PROSPECT and SAIL as implemented in PROSAIL allows simulation of the joint effect of leaf biochemistry (morphological and chemical parameters), canopy characteristics (LAI and crown geometry), sun–sensor geometry (sun angle and sensor angle) and clear/diffuse sky on canopy-scale reflectance. We used MATLAB version PROSAIL_5B_Matlab, available at http://teledetection.ipgp.jussieu.fr/prosail/.

FLiES

FLiES consists of a 1-D atmospheric RTM and a 3-D canopy RTM, based on the Monte Carlo ray tracing method (Kobayashi et al., 2012). The original FLiES model (e.g. Kobayashi & Iwabuchi, 2008) used geometric objects such as cone, cylinder and spheroid to delineate individual tree structure. Here we extended FLiES for LiDAR-based voxel representation of 3-D canopy structure, which enables a more realistic depiction of forest canopies. Each voxel contains leaf and woody elements, parameterized by using the LiDAR-derived 3-D canopy area, with 89% assigned to leaf elements and 11% assigned to woody elements, following a recent field survey in a Costa Rican tropical evergreen forest (Olivas et al., 2013). Voxel grains of 2 × 2 × 2 m³, representing a 3-D forest landscape of 600 × 600 m² surrounding the k67 site (Fig. S7), were used for FLiES simulations. Additionally, shoot-scale clumping, a metric quantifying foliage clumping within a shoot (Chen et al., 1997), is an important parameter in FLiES (Kobayashi et al., 2012). It increases with increasing clumping, and was set equal to 1 (or no shoot-scale clumping) for the default model simulations. The FLiES code in Fortran and the voxelized data for the k67 site are available from the authors upon request.

Model experiments

We used PROSAIL and FLiES as our main tools (see Method S5 for more details), which were set to the same sun–sensor geometry as MAIAC, and performed a suite of model experiments to assess the relative contributions of the three phenological factors (P) on canopy-scale reflectance seasonality. Details are shown as below:

• (P1) assess the LAI effect due to light-scattering by multi-leaf layers. The models were run under different monthly LAI_field (Fig. 2a) for PROSAIL and the interpolated LiDAR 3-D canopy structure for FLiES (Method S5), with fixed tissue optics (i.e. annual mean leaf demography-weighted leaf reflectance and transmittance, and multi-species average bark/litter reflectances).
• (P2) assess the leafless crown fraction effect due to the distinct canopy reflectance characteristics of leafless and green canopy phenophases. This is described as:

(Eqn 1)

where R_canopy,t is canopy-scale reflectance at given month t, R_green and R_leafless are modeled canopy-scale reflectance for green and leafless phenophases, respectively, and f_leafless,t is the tower camera-derived leafless crown fraction at month t (Fig. 2a). R_green was modeled under fixed LAI_field (i.e. annual mean LAI_field) for PROSAIL and the corresponding LiDAR canopy structure for FLiES, and fixed tissue optics (annual mean leaf demography-weighted leaf reflectance and transmittance, and multi-species average bark/litter reflectances). R_leafless was modeled under the same fixed LAI and multi-species average bark/litter reflectances, while leaf optics were set equal to bark optics (field-observed bark reflectance and bark transmittance = 0).

• (P3) assess the leaf age effect through seasonally varying leaf demographics by scaling the age-dependence of leaf optics to the canopy. We described this as:

(Eqn 2)

where R_Y, R_M and R_O are modeled canopy-scale reflectance of young, mature and old phenophases, respectively, and f_Y,t, f_M,t and f_O,t are their relative abundances at month t (Fig. 2b). R_Y, R_M and R_O were again modeled under fixed annual average LAI_field for PROSAIL and the corresponding LiDAR canopy structure for FLiES, multi-species average bark/litter reflectances and age-dependence of leaf reflectance and transmittance.

(P1 + P2 + P3) assess the joint effects of the three phenological factors (i.e. P1–P3) on canopy reflectance seasonality. We described this as:

(Eqn 3)

(Eqn 4)

where R_green,t and R_leafless,t are modeled canopy reflectance at month t for green and leafless phenophases, respectively. R_Y,t, R_M,t and R_O,t were modeled under seasonally varying LAI_field for PROSAIL and the interpolated LiDAR 3-D canopy structure for FLiES, age-specific leaf reflectance and transmittance, and multi-species average bark/litter reflectances.

We performed the above model experiments parametrized by the leaf optics from each of the 11 tree–canopy light conditions (Fig. S6), and the mean and SD of the total 11 modeling runs were calculated for the final analysis (see Method S5). To assess the relative importance of three phenological factors (i.e. P1–P3) on the ‘comprehensive’ model (i.e. P1 + P2 + P3) results, we used a relative importance of regressors in R (package relaimpo, using the method called ‘betasq’), following the approach of Grömping (2015).

Note that the interaction term, or the scattering between green and leafless crowns, or among crowns dominated by different canopy phenophases (i.e. young, mature, old and leafless), was not considered in this study. At nadir view angle (i.e. WV-2 and MAIAC data), this interaction term might be a second-order effect in controlling canopy-scale reflectance seasonality, compared with the three phenological factors already explored here; future analysis is needed to fully understand such interaction terms, which is beyond the scope of the current paper.

Results

Age-dependence of leaf optics

Despite marked variation in leaf optical properties across all 11 tree–canopy light conditions, within each tree–canopy light condition, field-measured leaf reflectance and PROSPECT-inverted leaf transmittance and absorptance show a strong dependence on leaf ages (Figs 1, S6, S8). The mean visible (400–700 nm) reflectance shows a continuous decline with leaf age, from 0.08 in young leaves to 0.05 and 0.04 in mature and old leaves, respectively (Fig. 1a). The mean NIR (700–1100 nm) and shortwave infrared (SWIR) (1100–2500 nm) reflectance initially increase with leaf ages from 0.44 (NIR) and 0.19 (SWIR) in young leaves, and then reach a peak at 0.46 and 0.22 after full leaf expansion (Fig. 1a). By contrast, the mean values of visible, NIR and SWIR transmittance all decline with leaf ages (Fig. 1c), from 0.10 (visible), 0.47 (NIR) and 0.26 (SWIR) in young leaves, to 0.03, 0.39 and 0.22 in mature leaves, and 0.01, 0.36 and 0.21 in old leaves. Our derived leaf absorptance (=1 − reflectance-transmittance) also shows a strong age-dependence, and the mean values of leaf absorptance continuously increase with leaf ages throughout the full spectral range (Fig. 1d).

Phenophase effects on canopy reflectance

We used PROSAIL and FLiES, together with observations from the WV-2 image, to explore the phenophase effects (i.e. young, mature, old and leafless) on canopy reflectance. Our results show that canopy reflectance varies in concert with canopy phenophases with a fixed LAI (i.e. LAI_field = 5 m² m⁻²; Fig. 3) and depends strongly on the combined variation in LAI and phenophases (Figs S9, S10). Specifically, canopy visible reflectance, although typically being of low magnitude (< 0.1), shows strong dependence on canopy phenophases, with the young phenophase displaying a much higher value than that of either mature or old phenophases. The leafless phenophase has a similar mean visible reflectance to that of the young phenophase. Canopy NIR reflectance shows a much broader magnitude of variation (> 0.2) across all phenophases, and displays a continuous decline from young to old, with minimum value at the leafless stage. Such phenophase effects on canopy reflectance are consistent across the two models (Fig. 3a,b for PROSAIL and Fig. 3c for FLiES), although PROSAIL results show consistently higher intra-phenophase variability than FLiES. Additionally, we performed a sensitivity analysis (with and without the thick-leaved species M. huberi, which has bias in modeled leaf transmittance), and our results show that the observed phenophase effects on canopy reflectance are also consistent (Fig. S11). Finally, both PROSAIL and FLiES show similar phenophase effects on canopy reflectance compared with the WV-2 observations (Fig. 3d), suggesting that the two RTMs can reproduce similar phenophase effects to observations.

Figure 3

Canopy radiative transfer models (RTMs) simulated and WorldView-2 (WV2) observed canopy-scale reflectances at different canopy phenophases at the k67 site. (a) PROSAIL simulated canopy-scale reflectance spectra for young (leaf age ≤ 2 months; blue), mature (leaf age: 3–5 months; green), old (leaf age: 6–14 months; red) and leafless (black) phenophases; (b) PROSAIL simulated canopy-scale reflectance convolved to WorldView-2 (WV-2) spectral bands (see Supporting Information Table S1); (c) FLiES simulated canopy-scale reflectance convolved to WV-2 spectral bands (see Table S1); and (d) the WV-2 image (acquired on 28 July 2011 with near-nadir view; see the Materials and Methods section) observed canopy-scale reflectances for each of three phenophases. All RTM results shown here are based on the scenario when canopy LAI = 5 m² m⁻², tree–environment-specific (n = 11 tree–canopy light conditions), age-dependent leaf optics for young, mature and old phenophases, respectively, multi-species (n = 13 tree species) average bark optics for leafless phenophase, solar zenith angle = 45° and nadir view. Shading indicates ± 1 SD. See Fig. S2 for WV-2 band spectral response functions; see Figs S9 and S10 for model simulations of other canopy LAIs.

Phenology effects on canopy reflectance seasonality

We used PROSAIL and FLiES to explore how including model representation of three phenological factors influences modeled canopy-scale reflectance seasonality, as evaluated against the MAIAC data. The seasonal pattern of MAIAC NIR and EVI shows an initial decline in the late wet season and then an increase in the dry season (Fig. 4). Our results show that the models driven by all three phenological factors (P1 + P2 + P3; or the ‘comprehensive’ model; Fig. 4) are best able to capture the MAIAC data, respectively explaining 87 and 75% of seasonal variation in MAIAC NIR and EVI using PROSAIL, and 64 and 69% using FLiES, although PROSAIL has larger model bias than FLiES.

Figure 4

Comparisons between canopy radiative transfer models (RTMs) simulated and MAIAC version of MODIS observed canopy-scale seasonality of (a) near-infrared (NIR) reflectance, and of (b) enhanced vegetation index (EVI). The models were parameterized by ‘comprehensive’ monthly phenological components (leaf area index (LAI), leafless crown fraction and leaf demographics, i.e. P1 + P2 + P3 in the Materials and Methods section), with PROSAIL (grey) and FLiES (black). RTM results shown here are the mean of their respective 11 RTM simulations (driven by tree–environment-specific leaf optics; see Methods S5). MAIAC data (in red) represent monthly means for 2000–2014, spatially averaged over a 5 × 5 km² window, and fully account for sun–sensor geometry and cloud/aerosol contamination (see the Materials and Methods section). r², coefficient of determination, is based on the linear regression between model and observations with intercept. Error bar indicates ± 1 SD; shading indicates the dry season.

We further ran the models separately parameterized by each of the three phenological factors (P1–P3), aiming to quantify their relative contributions. The results are shown in Fig. 5 (all 11 tree–canopy light conditions), Fig. S12 (all, but excluding M. huberi) and Table 1. Our results show that both PROSAIL and FLiES attributed very comparable phenological contribution in NIR and green reflectances, but with some variations in blue and red reflectances, and VIs. Additionally, both models show that canopy-surface leafless crown fraction (P2) dominated the magnitude of the ‘comprehensive’ modeled NIR and EVI seasonality; canopy-surface leafless crown fraction (P2) and leaf demographics (P3) jointly determined the relative seasonality of the ‘comprehensive’ modeled NIR and EVI; canopy LAI phenology (P1) barely contributed to the ‘comprehensive’ modeled NIR and EVI seasonality. Furthermore, when the models were run with and without M. huberi, the simulated relative seasonalities were almost identical to each other (Fig. S12), suggesting that our results are very robust, despite some transmittance bias associated with M. huberi (e.g. Fig. S5b).

Figure 5

Relative roles of different phenological components in accounting for ‘comprehensive’ model-simulated canopy reflectance seasonality in (a, c) near-infrared (NIR) reflectance, and (b, d) enhanced vegetation index (EVI), convolved to MODIS spectral bands (see Supporting Information Fig. S2). Upper panel for the PROSAIL model and lower panel for the FLiES model. Canopy radiative transfer models (RTMs) results shown here are the mean of their respective 11 RTM simulations (driven by tree–environment-specific leaf optics; see Methods S5). The models that include only the direct leaf area index (LAI) phenology effect (P1) are shown in blue, the models that include only canopy-surface leafless crown fraction seasonality (P2) are shown in red, the models that include only the leaf demography seasonality effect (P3) are shown in green, and the ‘comprehensive’ models which include all three phenological components (P1 + P2 + P3) are shown in black. Shading indicates the dry season.

Table 1. Relative contributions of the three phenological factors in accounting for ‘comprehensive’ canopy radiative transfer models (PROSAIL and FLiES) simulated seasonalities of canopy reflectance and vegetation indices (see Figs 5, S13)

	Spectra\relative contribution	P1	P2	P3
PROSAIL	Blue	0.00	0.95***	0.05**
	Green	0.00	0.03	0.97***
	Red	0.01	0.63**	0.37**
	NIR	0.02***	0.31***	0.67***
	EVI	0.03***	0.68***	0.29***
	NDVI	0.00	0.83***	0.17**
FLiES	Blue	0.00	0.48***	0.52***
	Green	0.00	0.04***	0.96***
	Red	0.00	0.52***	0.48***
	NIR	0.00	0.30***	0.70***
	EVI	0.01	0.33***	0.66***
	NDVI	0.00*	0.46***	0.54***

• PROSAIL- and FLiES-simulated canopy reflectance and vegetation indices were convolved to MODIS spectral bands (see Supporting Information Fig. S2); the three phenological factors are: P1, the leaf area index (LAI) effect; P2, the canopy-surface leafless crown fraction effect; and P3, the leaf age effect. Relative contributions were assessed by using a relative importance of regressors in R (package relaimpo; see the Materials and Methods section); asterisks indicate levels of significance (*, P = 0.05; **, P = 0.01; ***, P = 0.001).

We also analyzed both modeled and observed canopy reflectance seasonality for blue, green and red bands and reflectance-derived NDVI. The results are summarized in Figs S13 and S14. Overall, the modeled canopy visible reflectances were low (< 0.05) and had similar magnitudes to the observations, although their relative seasonalities were different. We also compared the canopy RTM results with the MCD43A1 product (Fig. S15), and the results are practically the same as with MAIAC.

Comparing PROSAIL with FLiES, we found that the relative seasonalities of observed EVI and NIR reflectance were more closely simulated by PROSAIL than by FLiES, although FLiES, with less of an offset, more closely matched the absolute value (Fig. 4). This was somewhat unexpected: if the 3-D structure of the forest matters for reflectance dynamics, we would in principle expect that a model (like FLiES) representing forest structure in three dimensions should be able to do better than a 1-D model (like PROSAIL). This suggests that some aspects of the 3-D dynamics (which are less constrained than the 1-D version) are mis-parameterized. For example, the FLiES 3-D structure is more capable of representing woody elements than is PROSAIL, but if the woody fraction (which is not well constrained by observation but has less seasonality) is overrepresented, this would artifactually reduce modeled canopy NIR reflectance seasonality. A more complete understanding of the role of 3-D vs 1-D structure in driving temporal variation in canopy reflectance is needed, but for the purposes of the present seasonality study, it is sufficient to note that both models appear to capture well the dynamics of vegetation indices over seasons.

Model sensitivity to the clumping effect

We used the ‘comprehensive’ model (P1 + P2 + P3) to explore the extent to which the clumping effect within a voxel in FLiES (approximated by shoot-scale clumping; Kobayashi et al., 2010) affected modeled canopy-scale reflectance seasonality. We found that although the absolute values varied greatly, the relative seasonalities of modeled reflectance remained consistent across a wide range of shoot-scale clumping (from 1.0 to 1.39) (Fig. S16). Because the value of shoot-scale clumping shows a strong biome dependence (He et al., 2012), and is difficult to measure in the field, our model sensitivity results suggest that the shoot-scale clumping effect can be an important source of uncertainty that causes the absolute magnitude difference between models and observations as shown in Fig. 4.

Discussion

This study provides two main results. First, when field-observed leaf and canopy characteristics are used to drive canopy RTMs, they simulate canopy-scale reflectance seasonality patterns that closely match satellite observations, robustly vetted to remove known artifacts, of the same region where the ground observations were made (Fig. 4). Second, simulated reflectance seasonality is shown to arise in the model directly from two main phenological factors: changes in canopy-surface leafless crown fraction and changes in leaf demography, with only a very minor contribution from changes in LAI (Figs 5, S12, S13; Table 1). We discuss three broad implications of these results.

Assessing differential phenological effects on canopy reflectance seasonality should help to correctly interpret climate–phenology relationships in the Amazon

We start by discussing the implications of our second main finding that seasonal variations in canopy-surface leafless crown fraction and leaf demography are the two main phenological factors driving modeled reflectance seasonality (Table 1). This is because canopy reflectance shows strong dependence on canopy phenophases (i.e. young, mature, old and leafless) (Fig. 3) and their seasonally changing fractions (Fig. 2a,b). Although LAI_field and the canopy-surface leafless crown fraction were highly negatively correlated at our site (Fig. 2a), our partitioning analysis attributes very little of the reflectance seasonality to changes in LAI (see Table 1): canopy-scale NIR reflectance increases with LAI through light scattering by the multi-leaf layer (Verhoef, 1984; Samanta et al., 2012b), but LAI-induced NIR increase saturates when LAI exceeds 4 m² m⁻² (Figs S9, S10). This implies that small seasonal changes detected in LAI (c. 1 m² m⁻²; Fig. 2a) in a dense tropical forest canopy, where LAI exceeds 5 m² m⁻² all year, will have little direct impact on canopy reflectance seasonality (Table 1). This also implies that the previously reported correlation between LAI and EVI at several sites in the Amazon (e.g. Brando et al., 2010; Wu et al., 2016) is probably driven by the associated correlated decline in canopy-surface leafless crown fraction (Fig. 2a), which significantly regulates canopy-scale reflectance seasonality (Fig. 5; Table 1).

Interestingly, although leaf age strongly regulates leaf reflectance (Fig. 1a; see also Chavana-Bryant et al., 2017; Wu et al., 2017), NIR reflectance and EVI had opposite trends with age at the canopy vs leaf scales, with young canopy phenophase having highest NIR reflectance, while young leaves have the lowest NIR reflectance (Fig. 1a vs Fig. 3a; Fig. 6). Canopy-scale reflectance is a joint consequence of overall canopy structure (LAI and crown geometry) in addition to leaf optical properties (reflectance and transmittance) (Roberts et al., 1998; Ollinger, 2011). Young leaves have much higher NIR transmittance (due to their low absorptance and reflectance; Fig. 1) leading to more multiple-scattering of NIR light within a very dense canopy (e.g. LAI > 3 m² m⁻²) dominated by young leaves. The effect of more multiple-scattering events in a canopy dominates that of lower NIR reflectance in individual leaves, leading to higher NIR reflectance at the canopy scale (Figs 3, S9, S10). This highlights the importance of connecting leaf optics of single scattering albedo (reflectance + transmittance) to canopy structure to correctly represent or interpret canopy-scale reflectance signatures.

Figure 6

Different leaf age effects on leaf- and canopy-scale (a) near-infrared (NIR) reflectance and (b) enhanced vegetation index (EVI), convolved to MODIS spectral bands (see Supporting Information Fig. S2). Leaf-scale patterns (red lines) are based on the field observations shown in Fig. 1(a) and canopy-scale patterns (black lines) are based on the PROSAIL model results shown in Fig. 3. Note the reversal of the age ranks for both NIR reflectance and EVI when going from leaf scale to canopy scale. Since the FLiES model simulates an age-dependency of canopy reflectance spectra similar to that of PROSAIL, we expect that FLiES has a similar change in the ranking. Error bars are 1 SD among 11 tree–canopy light conditions (leaf-scale) and associated PROSAIL simulations (canopy-scale).

By identifying the two phenological factors (leafless crown fraction and leaf demography) that explain satellite-detected seasonality (Figs 4, 5, S15), our work has direct implications for characterizing both spatial and temporal variation in satellite-detected tropical phenology (e.g. using MAIAC EVI), as these two factors represent different ecophysiological strategies for tropical tree responses to seasonal/inter-annual resource availability: leafless crowns are a manifestation of deciduous or near-deciduous habit, related to the hydrological sensitivity of tropical trees. Higher water stress in dry seasons can lead to higher abundance of deciduous trees over space (e.g. Bohlman, 2010; Guan et al., 2015; Xu et al., 2016) or increased drought-induced leaf shedding and mortality over time (e.g. Nepstad et al., 2007; Xu et al., 2016), resulting in an increased canopy-surface leafless crown fraction (and thus ‘brown-down’); leaf demography is more regulated by evergreen trees, which are not water limited and show sensitivity to dry-season increased sunlight, and associated light-induced new leaf flushing (‘green-up’) in response to seasonal and/or inter-annual drought (e.g. Wright & Van Schaik, 1994; Brando et al., 2010; Doughty et al., 2015; Lopes et al., 2016; Wu et al., 2016). Leaf demography, especially dry-season leaf turnover, might also have evolved to avoid herbivory for tropical trees (e.g. Aide, 1988, 1992). Promisingly, these two phenological factors exert different effects on seasonal variation in visible (especially in the green band; Fig. S13; Table 1) and SWIR (Fig. 3a) reflectances. It would therefore be technically feasible and scientifically important to use remote sensing products from multi-spectral or hyperspectral sensors that include visible and NIR (and ideally also SWIR) reflectances to differentiate these two processes.

Canopy RTM-modeled seasonality helps reconcile the Amazon phenology debate

Our findings also have important implications for recent debates about the mechanisms of satellite-observed ‘green-up’ of Amazon forest canopies (Huete et al., 2006; Brando et al., 2010; Galvao et al., 2011, 2013; Morton et al., 2014; Bi et al., 2015; Saleska et al., 2016). Several studies (Galvao et al., 2011; Morton et al., 2014) have suggested that satellite-observed seasonality in vegetation greenness (as captured by EVI) is an artifact of sun–sensor geometry and not representative of biophysical factors in forests measured on the ground. However, our modeling of biophysical factors from the bottom up, by simulating EVI seasonality (with wet-season declines followed by dry-season green-up, Fig. 4b) that is consistent with top-down satellite observations corrected for artifacts of clouds/aerosols and of sun–sensor geometry (e.g. Brando et al., 2010; Bi et al., 2015; Maeda et al., 2016; Saleska et al., 2016), suggests that these biophysical factors are in fact driving the observed vegetation seasonality, at least at this site.

Although derived from one site, these findings reveal general mechanisms for phenology that effectively rule out the interpretation that remotely sensed patterns are artifacts of changing sun-sensor geometry, as suggested by Galvao et al. (2011, 2013 and Morton et al. (2014). Like this work, Morton et al. (2014), in particular, sought to use sophisticated RTM simulations to identify causal mechanisms for observed seasonality, concluding that forest biophysical factors (e.g. phenology) were dominated by sun–sensor geometry artifacts in the satellite observations. What accounts for the different conclusions between the RTM-based study of Morton et al. (2014) and the RTM-based study presented here?

First, the BRDF correction applied by Morton et al. (2014) to the satellite observations appears to overestimate the size of the artifact that needs correction (Bi et al., 2015) and in any case does not eliminate detectable seasonality in greenness (Saleska et al., 2016), as determined both by comparison with the BRDF-corrected EVI used by Morton et al. (2014) (Saleska et al., 2016), and with the more rigorously validated BRDF-corrected MAIAC product (Lyapustin et al., 2012).

Second, in investigating the potential effects of vegetation phenology, although Morton et al. (2014) anticipated that individual leaf reflectances may change with age, they did not account for the fact that leaf transmittance also changes with age, or for the possibility of seasonally changing canopy structure caused by the changing fraction of leafless crowns. As discussed above, the lower NIR absorptance of young leaves (and consequent higher multiple-leaf scattering) is what causes canopies composed of young leaves to have higher overall canopy-scale NIR reflectance (Fig. 3) and hence higher EVI. Furthermore, leafless crown fraction is observed, in tower-based camera images, to change dramatically (by a factor of four, Fig. 2b), with substantial effect on the modeled reflectance seasonality.

This work thus clarifies the relationships between tropical canopy phenology and satellite-observed seasonality, helping to resolve debates regarding satellite-detected Amazon phenology (e.g. Morton et al., 2014 vs Bi et al., 2015, and Saleska et al., 2016). In particular, we note that not accounting for higher NIR transmittance in young leaves, or leafless crown fraction dynamics, will substantially diminish the simulated effect of vegetation phenology on reflectance seasonality.

Phenological impacts on canopy reflectance, with implications for hyperspectral retrievals of biophysical traits

By assessing the effects of leaf phenology on canopy reflectance spectra (Figs 3-5), our results help extend leaf-to-canopy scaling with RTMs – widely applied to scaling across space (e.g. Wessman et al., 1988; Weiss et al., 2000; Clark et al., 2005; Baret & Buis, 2008; Asner & Martin, 2011; Asner et al., 2011, 2016; Singh et al., 2015) – to the temporal domain. This extension is enabled by recent studies that demonstrate strong relationships of leaf traits and spectra with leaf age across diverse growth environments and species in tropical forests (Chavana-Bryant et al., 2017; Wu et al., 2017) and other biomes (e.g. Yang et al., 2014, 2016; Meerdink et al., 2016). Our work builds on these leaf-scale studies to provide support for the importance of remotely detectable temporal convergent relationships in traits and spectra at the canopy scale as well, suggesting the potential for deriving temporal variability in plant traits using canopy reflectance spectroscopy techniques.

Importantly, our work also provides a practical guide for effective trait retrieval using canopy reflectance spectroscopy techniques, in the context of both model inversion (Weiss et al., 2000; Baret & Buis, 2008) and empirical statistical approaches, such as partial least squares regression (PLSR; Asner et al., 2011; Singh et al., 2015). These are:

• Model inversion: in our work, although RTM-simulated canopy reflectance captured the seasonal dynamics of satellite observations (which was our main focus), there was a significant offset between the two, especially in the PROSAIL simulation of NIR and EVI (Figs 4, S14). These differences are probably associated with uncertainty in the foliage clumping effect (e.g. Fig. S16), which causes uncertainty in the absolute magnitude of canopy reflectance. More fine-scale measurements and modeling experiments are needed to fully understand such model–observation magnitude differences to better constrain the RTMs and reduce potential bias in model-inverted key plant traits.
• Empirical statistical approaches (e.g. PLSR): because leaf age affects both leaf traits and spectra (Wu et al., 2017), as well as canopy-scale reflectance spectra (Fig. 3), our work highlights the importance of sampling leaves of different representative leaf ages and of accounting for leaf demography within canopies, which will allow development of more reliable canopy-scale spectra–trait relationships in the tropics. Beyond the tropical forests studied here, this consideration is also probably important for any biome that has seasonally changing traits and spectra (e.g. Meerdink et al., 2016; Yang et al., 2016).

Conclusion

This study demonstrates that different components of leaf phenology (primarily leafless crown fraction and leaf demography) contribute significantly to satellite-observed seasonal variations in canopy greenness (i.e. MAIAC EVI). These findings effectively reconcile current controversies about satellite-detected vegetation seasonality in the Amazon, and provide a robust basis for using satellite remote sensing (after being properly processed, like MAIAC EVI used here) to monitor phenology and study climate–phenology relationships in the tropics. This work thus lays the foundation for the study of how these phenological factors, which represent different ecophysiological strategies by which tropical trees respond to varying resource availability, structure the large-scale satellite observations of forest dynamics across space and over time, thus offering new insight into the study of tropical climate–phenology relationships.

Acknowledgements

This work was supported by NASA (#NNX11AH24G, #NNX17AF56G and NESSF to J.W.), NSF (PIRE #0730305), and the Next-Generation Ecosystem Experiments–Tropics project supported by the US DOE, Office of Science, Office of Biological and Environmental Research and through contract #DE-SC0012704 to Brookhaven National Laboratory. K.H. was supported by JAXA GCOM-C (RA6 #111) and JSPS KAKENHI (#16H02948). S.C.S. was supported by NSF (EF-1550686 and EF-1340604). A.R.H., N.N.T. and S.G. were supported by an Australian Research Council – Discovery Project (ARC-DP140102698, CI Huete). DigitalGlobe data were provided by NASA's NGA Commercial Archive Data (http://cad4nasa.gsfc.nasa.gov) under the National Geospatial Intelligence Agency's NextView license agreement. We also thank the three anonymous reviewers for their constructive comments that helped to improve the scientific rigor and clarity of the manuscript.

Author contributions

J.W., H.K., S.C.S., S.R.S. and A.R.H. designed the project. J.W. and T.M. performed field-based tissue optical measurements, with help from R.C.O. H.K. and W.Y. developed the FLiES code. S.C.S. processed and analyzed the airborne LiDAR data. S.P.S. and R.M. provided and processed the WorldView-2 image. A.R.H., N.N.T and S.G. downloaded and processed the MODIS data. J.W., H.K. and S.C.S. performed the synthetic data analysis. J.W. drafted the manuscript, and H.K., S.C.S., S.R.S., A.R.H., S.P.S., B.W.N., K.G., D.G.D., A.R., T.M., N.R-C., R.M., N.N.T. and S.G. contributed to the final version.