Large diurnal compensatory effects mitigate the response of Amazonian forests to atmospheric warming and drying

Abstract

Photosynthesis and evapotranspiration in Amazonian forests are major contributors to the global carbon and water cycles. However, their diurnal patterns and responses to atmospheric warming and drying at regional scale remain unclear, hindering the understanding of global carbon and water cycles. Here, we used proxies of photosynthesis and evapotranspiration from the International Space Station to reveal a strong depression of dry season afternoon photosynthesis (by 6.7 ± 2.4%) and evapotranspiration (by 6.1 ± 3.1%). Photosynthesis positively responds to vapor pressure deficit (VPD) in the morning, but negatively in the afternoon. Furthermore, we projected that the regionally depressed afternoon photosynthesis will be compensated by their increases in the morning in future dry seasons. These results shed new light on the complex interplay of climate with carbon and water fluxes in Amazonian forests and provide evidence on the emerging environmental constraints of primary productivity that may improve the robustness of future projections.

Remote sensing data indicates that the peak evapotranspiration and productivity in the Amazon are shifting to mornings.

INTRODUCTION

Amazonian rainforests play an important role in the climate system by exchanging large amounts of energy, water, and carbon dioxide (CO₂) with the atmosphere and acting as a major carbon stock. Climate change is expected to make the Amazon forest substantially warmer and potentially drier over the 21st century (1) with potential warring impacts on primary productivity and the regional carbon budget. Photosynthesis and evapotranspiration (ET) may decline under future atmospheric warming and drying due to the stomatal closure, hydraulic failure, and Rubisco enzymatic capacity impairment (2, 3). Large-scale decreases in ET might further feed back to the regional climate by reducing cloud formation and subsequent rainfall events (4), causing a higher risk of drought and tree mortality (5). Moreover, the negative impact on primary productivity of Amazon forests could substantially weaken their carbon uptake (6).

Given these potential climate impacts on Amazon rainforests, it is becoming critical to understand and predict the direction and magnitude of photosynthesis and ET changes in response to atmospheric warming and drying in Amazonian forests (7, 8). Dynamic global vegetation models were used to predict forest response to long-term climate change, whereas there were substantial uncertainties in the prediction due to their poor responses to short-term variations in environmental drivers (9). For example, because land-surface models do not accurately account for the daily hysteretic response of stomatal conductance to compound atmospheric and soil dryness, most of them underestimate and overestimate latent heat flux (LE; equivalent to ET) in the morning and afternoon, respectively (10). These mechanisms also highlight difficulties in predicting the responses of photosynthesis and transpiration to warmer and drier conditions.

In addition to model simulations, numerous studies based on ground and satellite measurements have quantified the underlying mechanisms (biotic or abiotic) of the Amazon seasonality (11–13). The photosynthetic seasonality in Amazonian forests is strongly related to phenology (that is, leaf development and demography) (11), rather than directly to the physiological response to environmental drivers that explain only 3% of the photosynthesis variability at the monthly scale (14). As a result, the response of Amazonian forests to atmospheric warming and drying remains controversial, because conventional assessments of these environmental effects on photosynthesis and ET, such as the changes in the phenology of leaves and their age demography (young or mature) at seasonal scale, often masked out their responses to warming or drying at the diurnal time scale.

Nevertheless, environmental drivers explain most of the diurnal variations in photosynthesis when leaf phenology barely changes within a day (14). Air temperature (T_air) and atmospheric dryness [or vapor pressure deficit (VPD)] are higher in the afternoon than in the morning (15, 16), while the foliar water potential decreases from morning to afternoon due to the loss of water through transpiration and then recovers overnight (17–19). These processes can induce different responses of photosynthesis and ET to T_air and VPD between morning and afternoon. The difference in the fluxes of carbon and water between morning and afternoon provides key information on the physiological responses of plants to environmental drivers and potential constraints on primary productivity (15). For example, an afternoon depression of photosynthesis is a typical indicator of stomatal stress due to high VPD, whereas a symmetric pattern between morning and afternoon suggests that photosynthesis is predominantly light limited (15).

Diurnal variations in photosynthesis and ET have been measured at a few eddy covariance (EC) flux sites across the Amazon forest, where afternoon reductions in photosynthesis and ET have been observed (16). However, these flux sites are sparsely distributed across the region and may not fully represent the complex geographical patterns in vegetation and climate of the Amazon forest (20), resulting in a clear gap in scaling up from site observations to regional estimates of photosynthesis and ET. At present, large-scale observations of diurnal variations in photosynthesis or ET of Amazon forests are lacking, limiting our understanding of the physiological regulation of carbon-water fluxes and their feedbacks to the climate.

The International Space Station (ISS) hosts a series of instruments that observe the Earth at different times throughout the day (21). Among these instruments, the Ecosystem Spaceborne Thermal Radiometer Experiment (ECOSTRESS) has been retrieving ET measurements since 2018 (22), and the Orbiting Carbon Observatory-3 (OCO-3) has been measuring solar-induced chlorophyll fluorescence (SIF) since 2019 (23). SIF is electromagnetic radiation emitted by leaf chlorophyll during the light reaction of photosynthesis (24) and can be retrieved from high spectral resolution atmospheric and meteorological space missions (25, 26). Over the past decade, SIF has been shown to be a promising indicator of terrestrial photosynthesis (25, 27, 28) and for studying the seasonality of Amazonian forests (29, 30) and their responses to drought (31, 32) and aridity (33). However, because traditional sun-synchronous satellites pass over a given location on Earth at the same time each day, they cannot provide observations for different times of day over the Amazon.

Hence, the combination of ECOSTRESS ET and OCO-3 SIF provides a unique opportunity for studying the large-scale fluxes of carbon and water at the diurnal time scale. The diurnal photosynthesis and ET at regional scales are informative as how land-atmosphere interactions are regulated by changes in physiological responses to varying environmental conditions (34). In addition, the high T_air and VPD in the afternoon can mimic future conditions compared with current status. Thus, an analysis of the diurnal patterns of photosynthesis and ET from morning to afternoon can offer a unique view of their future trends and responses to climate change. Therefore, we aim to investigate the large-scale diurnal patterns of photosynthesis and ET in the Amazon forest and their responses to atmospheric warming and drying by combining ECOSTRESS ET with OCO-3 SIF.

Building on these ata streams, we investigated the large-scale diurnal patterns of photosynthesis and ET in the Amazon forest and SIF-based water use efficiency (WUE_SIF = SIF/ET) (34), which is a crucial indicator of the water use strategy in photosynthesis (35). We also analyzed the sensitivities of SIF, ET, and WUE_SIF to perturbations in environmental drivers [T_air, VPD, and soil moisture (SM)] and their differences between morning and afternoon. We further generated temporally and spatially continuous hourly total canopy SIF emission (denoted as SIF from here onward) and ET during 2015–2021 using random forest (RF) models trained with the response variables of OCO-3 SIF and ECOSTRESS ET and with the drivers of T_air, VPD, SM, radiation, and fraction of photosynthetic active radiation (FPAR) (see Materials and Methods). Using the RF models, we also projected changes in future photosynthesis and ET from 2015 to 2100 under a scenario of climate warming and increasing VPD based on the sixth phase of the Coupled Model Intercomparison Project (CMIP6). The limited biases related to the spatial extrapolations, temporal extrapolations, and future extrapolations does not affect our results and conclusions using RF models (see Material and Methods).

RESULTS

Diurnal patterns of photosynthesis and ET in Amazon forests

First, we explored the diurnal patterns of clear-sky SIF, ET, and WUE_SIF across the Amazon forest (fig. S1) based on the RF modeling from 2015 to 2021 (Fig. 1). Both carbon and water fluxes typically peaked around noon (11:00 to 13:00) (Fig. 2, A and D), and their diurnal variations were driven mainly by solar radiation and air temperature. SIF slightly increased from 7.69 mW · m⁻2 · nm⁻1 in the morning to 7.70 mW · m⁻2 · nm⁻1 in the afternoon in the wet season but decreased from 8.12 to 7.76 mW · m⁻² · nm⁻¹ in the dry season (Fig. 2A), which was consistent with CMIP6 earth system model (ESM) simulations: slightly increasing in the morning and decreasing in the afternoon (Fig. 2C). Site-based gross primary production (GPP) also showed a larger afternoon decrease in the dry season than in the wet season (Fig. 2B), which can be attributed to the much higher VPD and T_air in the dry season afternoon (fig. S2). Such a higher water stress condition of the dry season afternoon is further observed to accompany with a reduction in ET by RF models (Fig. 2D) and a weak ET increase by site level flux observations (Fig. 2E), both of which are in large contrast to the substantial afternoon increase in wet season ET. CMIP6 ESMs simulated a similar afternoon increase in ET between the wet and dry seasons (Fig. 2F), showing that ESMs could not fully capture the effects of water stress on ET. In addition, the morning WUE_SIF was always higher than the afternoon WUE_SIF, consistent with WUE_GPP from both site-based flux observations and CMIP6 ESM simulations (Fig. 2, G to I). Because we simulated clear-sky SIF and ET using the symmetrical radiation around local solar noon, the differences in SIF and ET between morning and afternoon were not due to light effects but rather to increased T_air and VPD and the related stomatal regulation on photosynthesis and transpiration.

Fig. 1. Diurnal patterns of SIF (mW · m⁻² · nm⁻¹), ET (W · m⁻²), and WUE_SIF (mW · m⁻² · nm⁻¹/W · m⁻²) in Amazonian forests throughout 2015–2020.

W · m⁻² Diurnal (A to E) SIF, (F to J) ET, and (K to O) WUE_SIF (=SIF/ET) at 2-hour intervals: 7:00 to 9:00, 9:00 to 11:00, 11:00 to 13:00, 13:00 to 15:00, and 15:00 to 17:00. The mean and SD are shown in the bottom right of each panel. The spatial resolution is 1° × 1°.

Fig. 2. Diurnal patterns of energy, carbon, and water fluxes in Amazonian forests for wet and dry seasons.

The RF models generated (A) SIF (mW · m⁻² · nm⁻¹), (D) ET (W · m⁻²), and (G) WUE_SIF (mW · m⁻² · nm⁻¹/W · m⁻²) averaged over 2015–2021. The shaded area represented the one SD across the Amazon forest. Site-based eddy flux-derived (B) GPP (μmol · m⁻² · s⁻¹), (E) ET (W · m⁻²), and (H) WUE_GPP (μmol · m⁻² · s⁻¹/W · m⁻²). The shaded area represented the one SD across three sites (BR-Sa1, BR-Sa3, and GF-Guy). CMIP6 ESM simulated historical (C) GPP (μmol · m⁻² · s⁻¹), (F) ET (W · m⁻²), and (I) WUE_GPP (μmol · m⁻² · s⁻¹/W · m⁻²) averaged over 2000–2014. The shaded area represented the one SD across ESMs. More details about the sites and CMIP6 ESM can be found in tables S1 and S2, respectively. The averages of morning (a.m.) and afternoon (p.m.) were shown in blue and orange for the wet and dry seasons, respectively.

Next, we investigated the diurnal and seasonal patterns of carbon and water fluxes across the Amazon forest and the hysteretic responses of WUE to VPD, which have not yet been observed at the regional scale before. The amplitudes of the diurnal variations in both SIF and ET were greater than their respective seasonal variations (Fig. 3, A and B and fig. S3), and the relationships between SIF and ET differed significantly (P < 0.001, using one-way analysis of covariance) between the morning and afternoon in both the wet and dry seasons (Fig. 3, C and D). The results of the linear regression modeling showed that SIF was lower in the late afternoon than in the early morning for the same level of ET, supporting the idea that WUE_SIF declined during the day as VPD increased (Fig. 3, E and F). Moreover, the regional diurnal WUE_SIF exhibited an oblique U-shaped, hysteretic relationship with VPD (Fig. 3, G and H). The hysteresis of WUE_SIF-VPD suggests that the relationship between photosynthesis and ET as quantified by WUE_SIF changes during the day, most likely because of a decline of water content in the plant system and an increased sensitivity of stomatal conductance to VPD in the afternoon.

Fig. 3. Seasonal and diurnal patterns (shown as fingerprints) of SIF (mW · m⁻² · nm⁻¹), ET (W · m⁻²), WUE_SIF (mW · m⁻² · nm⁻¹/W · m⁻²), and VPD (kPa) throughout the Amazon forest.

(A) SIF, (B) ET, (E) WUE_SIF, and (F) VPD. The red vertical lines represent the dry season from June to October. The SIF and ET were derived on the basis of the RF models during 2015 and 2021, and VPD was from the ERA5 reanalysis data during 2015 and 2021. The linear relationships between SIF and ET in the (C) wet and (D) dry seasons, with crosses (morning) and squares (afternoon) representing average hourly values at a monthly temporal resolution in (A) and (B). The relationships between hourly WUE_SIF and VPD averaged over the (G) wet and (H) dry seasons. The black arrows represent the directions of variations in diurnal WUE_SIF and VPD.

To assess how the hysteresis of WUE_SIF-VPD varied spatially, we selected three subregions in eastern, northwestern, and southwestern Amazon (three boxes in fig. S4) that differ in canopy structure, phenology, and rainfall and found that the hysteretic responses of WUE_SIF to VPD varied spatially across the three regions (Fig. 4), indicating that the limited site observation cannot adequately represent all the forests in the Amazon region. The strongest hysteretic response of WUE_SIF to VPD was observed in drier regions, such as the eastern Amazon in the dry season (Fig. 4E). The oblique U-shaped relationships between WUE_SIF and VPD across different subregions were consistent with surface flux observations at three Amazonian sites (BR-Sa1, BR-Sa3, and GF-Guy) (figs. S5 to S7). These results indicated that our regional hourly SIF and ET captured well the diurnal, spatial, and seasonal variations in photosynthesis and ET and their responses to environmental drivers.

Fig. 4. Relationships between WUE_SIF and VPD for the different regions.

(A to C) Northwestern, eastern, and southwestern Amazon Basin (red, blue, and black boxes, respectively, in fig. S4). The dots represent the seasonal averages, with the colors indicating the time of day. (D to F) similar to (A) to (C) but for the dry season.

To quantify the Amazonian diurnal variations in the ecosystem fluxes of carbon and water, we compared SIF, ET, and WUE_SIF between morning (7:00 to 12:00) and afternoon (12:00 to 17:00). Clear spatial variations were observed for SIF, ET, and WUE_SIF in both the wet and dry seasons (fig. S8). For example, WUE_SIF during the dry season was generally lowest in the east with relatively younger forest (fig. S9), where VPD was high and SM was low (fig. S2). Marginal diurnal changes in SIF were observed during the wet season (Fig. 5A), indicating that photosynthesis was mostly light limited when T_air and VPD were low (fig. S2, A and F) and SM was high (fig. S2K). However, during the dry season, SIF decreased from morning to afternoon (Fig. 5B), especially in the southeastern part of the Amazon where rainfall was low and T_air and VPD were high. ET increased and decreased diurnally over much of the Amazon in the wet and dry seasons, respectively (Fig. 5, C and D), with a shift from a demand-limited regime to a supply-limited regime. Therefore, these results revealed varying effects of T_air and VPD on photosynthesis and ET between the wet and dry seasons.

Fig. 5. Difference in SIF, ET, and WUE_SIF between morning (7:00 to 12:00) and afternoon (12:00 to 17:00) in Amazonian forests.

(A and B) Spatial distributions of percentage changes in SIF from morning to afternoon in the wet and dry seasons, respectively. The mean and SD are shown in the bottom right of each panel. (C to F) are similar to (A) and (B) but for ET and WUE_SIF, respectively.

Sensitivities of photosynthesis and ET to environmental drivers

The afternoon depressions of photosynthesis and ET in the dry season (Fig. 5, B and D) provide unique insights on how land-atmosphere interactions will respond to future climatic warming and atmospheric drying, but it is unclear how individual environmental drivers (such as T_air, VPD, and SM) affect the fluxes of carbon and water at the diurnal scale. Particularly, the effects of T_air, VPD and incoming solar radiation were commonly mixed due to their high correlations (fig. S10). To distinguish between the interdependent effects of environmental drivers on the fluxes of carbon and water, we analyzed the sensitivities of SIF, ET, and WUE_SIF using RF models (see Materials and Methods).

The results showed positive sensitivity of SIF to VPD (in other words, SIF would increase with VPD) in the morning more than 98.7% and 70.9% of the region in the wet and dry seasons, respectively (Fig. 6, A, B, and D), after factoring out the effects of other drivers. The negative sensitivity of photosynthesis to VPD in the dry season morning mainly occurred in the nonforest regions (such as grass and savanna). The sensitivity of SIF to VPD also showed a clear northwest-southeast gradient from being positive to being negative in the dry season morning, indicating that the super-wet northwest Amazon was not VPD stressed. However, the afternoon SIF exhibited negative sensitivity to VPD more than 74.5 and 67.4% of the region in the wet and dry seasons, respectively (Fig. 6, A, C, and E). Moreover, asymmetric responses of SIF to T_air between morning and afternoon were observed in the dry season (Fig. 7, A, D, and E). Different from SIF, ET showed consistently negative sensitivities to VPD (Fig. 6, F to J) in the morning or afternoon in both dry and wet seasons but positive sensitivities to T_air (Fig. 7, F to J). These results stressed the importance of stomatal control on the ET of the Amazon that is able to reverse the sensitivity of ET to VPD (i.e., plants are transpiring less when the atmospheric demand is higher). Consequently, WUE_SIF was positively sensitive to VPD (Fig. 6, K to O) but negatively sensitive to T_air (Fig. 7, K to O); therefore, forests varied their water use strategies under different temperature and atmospheric dryness stresses. Given the small variations in SM during the course of the day, SIF and ET exhibited weaker sensitivities to SM (fig. S11) than to VPD and T_air. The combined negative sensitivities of SIF to T_air and VPD in the dry season afternoon led to a decrease in SIF, which is a sign of what could happen in Amazonian forests under the projected atmospheric warming and drying in the future.

Fig. 6. RF-based sensitivity analysis of SIF, ET, and WUE_SIF to VPD.

(A) Box plots of the sensitivity of SIF to an increase in VPD by one SD for the wet season morning (WM), wet season afternoon (WA), dry season morning (DM), and dry season afternoon (DA, the period when foliar and soil water potentials are lowest and stress is highest). (B to E) The spatial distribution of the sensitivity of SIF to VPD. (F to J) and (K to O) are similar to (A) to (E) but display the sensitivities of ET and WUE_SIF to VPD, respectively. These sensitivities were derived as partial dependences, assuming that other drivers did not vary with VPD. The red line in each box plot indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. The stippling on the maps represents regions where at least 95% of the RF models agree on the sign of the sensitivity. The numbers in the bottom right corner of each map panel are the area percentages of positive (+, red) and negative (−, blue) sensitivities.

Fig. 7. RF-based sensitivity analysis of SIF, ET, and WUE_SIF to .

T_air (A) Box plots of the sensitivity of SIF to an increase in T_air by one SD for the wet season morning (WM), wet season afternoon (WA), dry season morning (DM) and dry season afternoon (DA, the period when foliar and soil water potentials are lowest and stress is highest). (B to E) Spatial distribution of the sensitivity of SIF to T_air. (F to J) and (K to O) are similar to (A) to (E) but display the sensitivities of ET and WUE_SIF to T_air, respectively. These sensitivities were derived as partial dependences, assuming that other drivers did not vary with T_air. The red line in each box plot indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. The stippling on the maps represents regions where at least 95% of the RF models agree on the sign of the sensitivity. The numbers in the bottom right corner of each map panel are the area percentages of positive (+, red) and negative (−, blue) sensitivities.

Next, we quantified the large-scale afternoon depressions of SIF, ET, and WUE_SIF following increases in T_air and VPD. For this purpose, we applied the afternoon RF model to the morning conditions during 2015–2021 (see Materials and Methods). The results of this process were considered in the “model_PM + condition_AM” simulations and were compared with the “model_AM + condition_AM” simulations. SIF and ET decreased by 4.4 ± 1.5 and 2.1 ± 2.8%, respectively, from the model_AM + condition_AM simulations to the model_PM + condition_AM simulations in the wet season (Fig. 8, A and C); these decreases were stronger in the dry season (Fig. 8, B and D). The stronger reduction in SIF than ET led to decreases in WUE_SIF by 2.2 ± 3.6 and 0.6 ± 3.5% in the wet and dry seasons, respectively (Fig. 8, E and F). Therefore, the strong afternoon reductions in SIF and ET point to the stronger asymmetric responses of photosynthesis and ET to atmospheric warming and drying in the dry season than in the wet season in the Amazon forest.

Fig. 8. Simulated percentage changes in SIF, ET, and WUE_SIF from the model_AM+ condition_AM simulation to the model_PM+ condition_AM simulation due to the increase in and VPD and decline in leaf water potential.

T_air The model_AM + condition_AM and model_PM + condition_AM simulations were obtained by driving the morning and afternoon RF models using the morning inputs (including radiation, FPAR, VPD, T_air, and SM), respectively. (A and B) Percentage change in SIF from the model_AM + condition_AM to model_PM + condition_AM simulations in the wet and dry seasons, respectively. (C to F) are similar to (A) and (B) but for ET and WUE_SIF, respectively. The mean and SD are shown in the bottom right of each panel.

Projections of photosynthesis and ET under atmospheric warming and drying

The RF models were used to predict future changes in Amazonian morning and afternoon photosynthesis and ET from 2015 to 2100 using the CMIP6 model simulations as inputs for the shared socioeconomic pathways (SSPs) SSP245, SSP370, and SSP585 (36). We did not consider the CO₂ fertilization effect and changes in forest composition or structure. The morning SIF in the wet season increased by 4, 6, and 8% from 2015 to 2100 under SSP245, SSP370, and SSP585, respectively (Fig. 9A), but the increases in the afternoon SIF were weaker (Fig. 9B). Accordingly, the daily SIF in the wet season increased (Fig. 9C), which was consistent with the CMIP6 model simulations of GPP, although there was a difference in magnitude between our results and CMIP6 simulations (Fig. 9D). The changes from 2015 to 2100 in ET in the wet season were much smaller than those in SIF or GPP (Fig. 9, E to H). In the dry season, an increase in the morning SIF (Fig. 9I) was offset by the decrease in the afternoon SIF (Fig. 9J). Consequently, only marginal changes in the daily SIF in the dry season were projected (Fig. 9K). However, CMIP6 models predicted increasing trends of primary productivity under future climate warming in the Amazon (Fig. 9L), which could be dominantly driven by the CO₂ fertilization effects. The morning ET in the dry season increased by 3, 4, and 5% from 2015 to 2100 under SSP245, SSP370, and SSP585, respectively, but the afternoon ET was nearly constant. As a result, daily ET increased (Fig. 9, M and N). The widespread reductions in SIF (89%) and ET (77%) in the dry season afternoon were predicted by the end of this century under SSP585 (figs. S12 and S13). On the basis of our data stream of diurnal SIF and ET, our results revealed stronger sensitivity to atmospheric dryness leading to overall a neutral response of Amazonian photosynthesis to atmospheric warming and drying in the future dry season.

Fig. 9. Projections of relative changes in SIF and GPP from 2015 to 2100 across the Amazon forest.

Projected changes in SIF in the wet season morning (A), wet season afternoon (B), wet season daily average (C), and CMIP6 daily average GPP in the wet season (D). The gray areas are the overlapped regions of one SD across different CMIP6 models for three SSP scenarios. To reduce the interannual variation, these data were smoothed using a 5-year running average. (E to H) similar to (A) to (D) but for ET. (I to P) similar to (A) to (H) but for the dry season.

DISCUSSION

Differences in diurnal patterns of photosynthesis and ET between wet and dry seasons

This study uses a combination of satellite-based data from two instruments (ECOSTRESS and OCO-3) aboard the ISS to study the large-scale diurnal patterns of photosynthesis and ET over the Amazon forest and their responses to environmental drivers. We reveal different large-scale diurnal patterns of photosynthesis and ET between wet and dry seasons in the Amazon forest (Fig. 3), which have not been observed at regional scale. The photosynthesis and ET across the Amazon forest did not decrease from morning to afternoon in the unstressed wet season but decreased in the dry season (Fig. 2, A and D), which could be attributed to the decreases in mesophyll (37) and stomatal conductance (15) in response to high VPD (fig. S2J) and decreased leaf water potential during the day. The stronger stresses on photosynthesis and ET in the dry season than in the wet season were also supported by site-based flux observations: Photosynthesis decreased more, and ET increased less from morning to afternoon in the dry season than in the wet season (Fig. 2, B and E).

The different diurnal patterns of ET between wet and dry seasons could be partly attributed to the strong interplay between SM and VPD on leaf water potential and therefore on stomatal behavior (15). Because the decrease in SM from morning to afternoon was very small, SM had a weak direct effect on ET at the diurnal scale. However, the difference in SM between wet and dry season regulates predawn leaf water potential (38). In the wet season (SM is high), the high predawn leaf water potential resulted in low stomatal sensitivity to VPD. In the dry season with low SM, leaf water potential could reach a critical threshold, inducing partial stomatal closure at high VPDs and hence the decrease in ET. A recent study also pointed out that plants could either positively or negatively regulate water use depending on the water potential of the leaves (39). In other words, ET increased when water stress was limited and ET was demand limited (e.g., in the wet season) following an increase in evaporative demand. In contrast, ET decreased due to stomatal closure in response to substantial water limitation, leading to a decrease in ET (e.g., in the dry season).

Asymmetric responses of photosynthesis to VPD and T_air

These results raise a question of whether the response of photosynthesis to VPD would be different between morning and afternoon in the humid topics; for example, photosynthesis could be positively and negatively sensitive to atmospheric dryness in the morning and afternoon, respectively (Fig. 6). The positive sensitivity of photosynthesis to VPD in wet environment could not be attributed to the correlation between VPD and radiation in the morning because radiation remained unchanged using the sensitivity analysis (see Materials and Methods). Rainfall interception by leaves and morning dew can limit photosynthesis by reducing light availability and slowing carbon diffusion (40). Drier air can facilitate the rate of evaporation from wet foliar surfaces, thus reducing morning fog and increasing light absorption earlier in the day (41) and hence photosynthesis and SIF. The positive sensitivity of morning SIF to VPD in this work is consistent with a recent study reporting that SIF derived from Global Ozone Monitoring Experiment–2 (~9:30 overpass time) was positively sensitive to VPD in the Amazon rainforest (33). However, we further found that the positive effect of VPD on photosynthesis mostly disappeared in the afternoon, by then the dew would already have evaporated. In addition, we also found the asymmetric responses of SIF to T_air in the dry season but not in the wet season. This could depend on whether T_air reached the optimum temperature for photosynthesis (T_opt, estimated at around 30°C for tropical forests) (42), indicating that T_air exceeded T_opt in the dry season afternoon and hence photosynthesis reduced.

The asymmetric responses of photosynthesis to VPD and T_air between morning and afternoon in the humid tropics were not observed at regional scale based on SIF from sun-synchronous satellites that passed over at a fixed time, thus highlighting the importance of OCO-3 SIF and ECOSTRESS ET observations for studying the physiological responses of Amazonian forests to atmospheric warming and drying and their compound effects with soil drying, offering a window onto future physiological changes in the Amazon. Although these results indicate that SIF from existing sun-synchronous satellites cannot fully reveal the physiological responses to varying environmental drivers at the diurnal scale, it is still feasible to use SIF based on sun-synchronous satellites to study the seasonality of photosynthesis and phenology (29, 30, 43) because the diurnal bias does not affect the overall seasonality of photosynthesis.

Different water use strategies under temperature and atmospheric dryness stresses

On the basis of the sensitivity analysis, this study identified the different large-scale water-use strategies of Amazonian forests under temperature and atmospheric dryness stresses: WUE_SIF increased with VPD (being conservative; Fig. 6, K to O) and decreased with T_air (Fig. 7, K to O). Therefore, ET decreased to reduce water losses to reduce the risk of hydraulic failure under high VPD (44), leading to increasing WUE_SIF. However, under high temperature and ample water supply, humid tropical forests reduced their WUE_SIF (less conservative) by increasing ET to lower foliar temperature (for cooling) for maintaining high photosynthetic rates (45). The decreases in WUE_SIF from morning to afternoon indicate a dominant role of the cooling mechanism in regulating WUE_SIF for humid tropical forests. The decrease in WUE_SIF is opposite to what would be predicted by the widely used model of stomatal conductance based on optimality principles (46), i.e., minimizing water losses while maximizing photosynthesis. Such an inconsistency with optimal stomatal models was also observed from a warming experiment (47). Our results suggest that stomatal models based on optimality principles may underestimate the role of ET as a key cooling mechanism in the humid tropical forests in response to temperature stress.

Neutral changes of Amazon photosynthesis and ET in the dry season

We projected that both diurnal (morning and afternoon) and seasonal (wet and dry) variations of photosynthesis and ET in the Amazon forest will change under future atmospheric warming and drying conditions (Fig. 9). Assuming a similar species composition and no major CO₂ fertilization effect, photosynthesis will likely increase in the unstressed wet season and in the mornings of the dry season in the humid tropics, supporting the positive effect of warming on photosynthesis under adequate soil water supply (48). However, this positive effect will not occur in the dry season afternoon due to increasing temperature (beyond T_opt) and atmospheric dryness, which will reduce the photosynthetic capacity and stomatal conductance (3, 44). Therefore, we predict large-scale afternoon depressions of photosynthesis in the dry season. These diurnal variations do not contradict the dry season increase of photosynthesis and ET (12) because the seasonality of Amazon ecosystems is largely controlled by phenology (i.e., leaf quality and demography) (14), although the global seasonal and interannual variations in photosynthesis are also related to radiation (49) and water availability (50). Our results highlight the importance of incorporating more accurate representation of physiological responses driving carbon and water fluxes in “next-generation” Earth system models.

Diurnal CO₂ data have not been available at the regional scale, and therefore, atmospheric CO₂ concentrations were not included as a driver in our RF model, and the potential impact of increasing atmospheric CO₂ concentration on our simulations was therefore not considered. Because we mainly investigated the diurnal patterns of photosynthesis and ET and their responses to environmental drivers, ignoring CO₂ effects has little effect on the diurnal patterns. Therefore, the projections still present valuable information on the large-scale diurnal patterns of photosynthesis and ET. For example, the future increases in photosynthesis in the wet season are higher in the morning than in the afternoon. Even if considering the CO₂ fertilization effects, our projection still holds because of the strong and negative influences from the high afternoon T_air and VPD. As for the dry season, it is unclear whether the CO₂ fertilization effect could mitigate afternoon decreases of photosynthesis. This issue is beyond the scope of this study and deserves future study. CMIP6 ESMs include additional processes (such as CO₂ fertilization effects and atmospheric feedbacks), and they predict increasing GPP in both the wet and dry seasons. Given the notable atmospheric drying in the future dry season, the simulated similar increase in the dry season photosynthesis (Fig. 9L) as the wet season photosynthesis (Fig. 9D) by the CMIP6 ESMs suggests that ESMs did not fully accurately represent the negative VPD effects on photosynthesis, and likely overestimate photosynthesis and potentially also the terrestrial sink strength of the Amazon forest under future climatic conditions.

In summary, we used a combination of satellite observations to assess the large-scale diurnal patterns of photosynthesis and ET and their drivers over the entire Amazon forest. We identified reductions in regional photosynthesis and ET from morning to afternoon in the dry season and quantified this effect at an unprecedented spatial scale through an observation-driven analysis. Our large-scale diurnal patterns of photosynthesis and ET provided a strong observational basis for modeling the exchanges of carbon and water fluxes between the land and atmosphere in the Amazon forest. This study would also improve our understanding of the direction and magnitude of ecosystem fluxes in response to climate change in other tropical forests. We envisage that these findings will pave the way for the use of high-frequency observations made by satellite-based sensors in studies of plant functioning and ecosystem processes to better project future changes in the terrestrial carbon and water cycles.

MATERIALS AND METHODS

OCO-3 SIF

SIF is tightly linked to photosynthesis and is a good proxy of GPP at the ecosystem scale (25, 27, 28). The varying overpass time of the ISS allowed us to derive the diurnal pattern of SIF over the Amazon forest from OCO-3 SIF data (v10) (23, 51), which has been used in a recent perspective article (34). Figure S1B shows the footprints of OCO-3 from January to February 2020 as an example, which present the variations in observation time (local solar time) for each path. We used OCO-3 SIF located in the Amazon forest from August 2019 to December 2021. OCO-3 provides SIF data at both 757 and 771 nm, but we used only the 757-nm SIF data due to their stronger signal (52). We also retained only the observations at nadir (viewing zenith angle near zero) and removed low-quality data, which were characterized as having a quality flag of >1.

We investigated the diurnal patterns of SIF over the Amazon forest using observations from OCO-3 on the ISS. However, the sun-target-viewing geometry effects on satellite SIF observations are particularly relevant for OCO-3, given the wide range of solar zenith angles (SZA) observed during the diurnal cycle. Therefore, we corrected the sun-target-viewing geometry effects in SIF data acquired by the OCO-3 sensor on the ISS (SIF_obs) by deriving the total canopy SIF emission (SIF_total) based on a reflectance-based method, which has been widely used in previous studies (52–54)

$${\mathrm{SIF}}_{\mathrm{total}}={\mathrm{SIF}}_{\mathrm{obs}}/f_{\mathrm{esc}}\times \mathrm{\pi }$$

$$f_{\mathrm{esc}}={\mathrm{NIR}}_v/i_0\times K$$

$${\mathrm{NIR}}_v=(\mathrm{NIR}-R)/(\mathrm{NIR}+R)\times \mathrm{NIR}$$

$$i_0=1-\exp (-G\times \mathrm{CI}\times \mathrm{LAI}/\cos \mathrm{SZA})$$

where f_esc is the probability of the escape of fluorescence from the canopy in the direction of observation, NIR_V is the near-infrared reflectance of vegetation, R and NIR are the reflectance at red and near-infrared bands, respectively, i₀ is canopy interception, ω is foliar leaf albedo at the band of 757 nm, G is the value of a G function, CI is clumping index, LAI is leaf area index, and cosSZA is the cosine of SZA. R and NIR were simulated with the RossThick-LiSparseR (RTLSR) Bidirectional Reflectance Distribution Function model using the sun-target-viewing geometry of OCO-3 SIF. The parameters provided by the MCD19A3 BRDF/albedo product were used to drive the RTLSR model. For Amazonian forests, K was set at 1.2, G was set at 0.5, and CI was set at 0.66. LAI data were obtained from Copernicus LAI products (see below). SIF_total is referred to as SIF for simplicity elsewhere in the manuscript.

ECOSTRESS ET and T

The ECOSTRESS, a thermal radiometer built by the Jet Propulsion Laboratory of the National Aeronautics and Space Administration, is also aboard the ISS. ECOSTRESS measures thermal infrared radiation in five bands, which are used to generate higher level products, such as surface temperature and ET. ECOSTRESS and OCO-3 follow similar acquisition schedules at different times of day, which enabled us to explore the diurnal cycle of ET (fig. S1C). We used the ECOSTRESS ET product computed with the ET_PT–JPL algorithm (55). We used ECOSTRESS ET located in the Amazon forest from July 2018 to December 2021. Observations under cloudy conditions were excluded from the analysis based on the quality control flag. A preliminary validation study demonstrated that ECOSTRESS ET performed well against site measurements at 82 EC sites around the world (22). In addition to ET, ECOSTRESS level 3 products also provide the fraction of transpiration (T) in ET (T/ET), but the retrieval accuracy of T/ET has not been extensively assessed due to the lack of ground validation data. In addition, the T/ET provided by ECOSTRESS (fig. S14) was lower than previously reported values as it is less constrained by observations (56). Therefore, we did not use T.

Calculation of SIF-based WUE

The coupling of carbon and water was studied by developing a metric, SIF-based water-use efficiency (WUE_SIF = SIF_total/ET) (34), which is a proxy for GPP-based WUE (WUE_GPP = GPP/ET). ECOSTRESS’ level 4 WUE was not used because it was calculated from ET and eight-day MODIS GPP. WUE_GPP is an indicator of the adjustment of photosynthesis to the loss of water, and has been studied many times at the seasonal and yearly scales globally (35). The magnitude of WUE often corresponds to the strategy of water use by plants, with higher WUEs indicating a more conservative strategy and lower WUEs indicating a less conservative strategy. The derived patterns of the relationship between VPD and WUE_SIF (Fig. 3, G and H) were consistent with those between VPD and WUE_GPP based on the surface EC flux measurements (figs. S5 to S7), and uncorrected SIF_obs-based WUE (WUE_SIFobs = SIF_obs/ET) had reverse relationships with VPD (fig. S15), demonstrating the necessity of correcting the sun-target-viewing effects of SIF_obs.

Copernicus LAI

We used LAI data from the Copernicus Global Land Service LAI (version GEOV2). The Copernicus LAI dataset was derived using a near real-time algorithm based on an artificial neural network. Copernicus LAI was released at 10-day intervals and was linearly interpolated into a daily interval. Copernicus LAI was used to estimate i₀ in Eq. 4 above and then used to estimate diurnal FPAR (57)

$$\mathrm{FPAR}=(1-0.05)-(1-0.06)\times (1-i_0)$$

ERA5 reanalysis

We acquired air temperature (T_air), dew-point temperature (T_dew), downward surface solar radiation (DSSR), and clear-sky DSSR from the fifth ECMWF ReAnalysis (ERA5) hourly data, which cover the global surface and have a spatial resolution of 0.25°. T_air and T_dew were used to calculate the VPD using the equation described in a previous study (58). To match satellite SIF and ET observations, linear interpolation was used to interpolate these meteorological data based on the latitude, longitude, and overpass time of the satellite observations. The DSSR and clear-sky DSSR, as proxies of PAR, were used to calculate the absorbed PAR (APAR). Because both the satellite SIF and ET data were filtered for cloudy conditions, they were better correlated with clear-sky APAR than with all-sky APAR (fig. S16), and hence clear-sky DSSR was used.

Eddy flux data

To compare with results from satellite observations, we used EC measurements from three flux sites (BR-Sa1, BR-Sa3, and GF-Guy) located in Amazonian forests (red dots in fig. S1A and table S1) and acquired the data from the FLUXNET2015 TIER ONE dataset (59). We picked up half-hourly or hourly GPP, LE, shortwave radiation (Rg), and VPD and removed low-quality data using quality control flag. The clearness index, which is the ratio of shortwave radiation to potential radiation at the top of the atmosphere, was calculated to remove data collected under cloudy skies (clearness index < 0.3). GPP-based WUE (WUE_GPP) was calculated as the ratio of GPP to LE. The local time was converted to local solar time based on the time zone and longitude to enable comparisons with data acquired by the OCO-3 and ECOSTRESS instruments. All data for the available period were averaged into hourly data from 7:00 to 17:00 and per month to represent the climatology.

RF model

The available OCO-3 SIF and ECOSTRESS ET data were sparse in time and space and had substantial gaps (figs. S17 and S18) due to the high cloud cover over the study region and the ISS orbit. Discrete sampling of SIF and ET might produce biased diurnal patterns and spatial distributions after simple aggregation if data are not properly processed. However, the thousands of satellite-based SIF and ET data points we acquired were sufficient to train a machine learning model, such as the RF model (60). RF models can represent the nonlinear interactions among various variables and are relatively insensitive to outliers. Accordingly, we used RF models to generate temporally and spatially continuous SIF and ET fields. OCO-3 SIF and ECOSTRESS ET are released at different spatial resolutions. To maintain consistency, we averaged instantaneous OCO-3 SIF and ECOSTRESS ET into 0.1° × 0.1° grids and then used the gridded SIF and ET as response variables in the RF model. FPAR derived from Copernicus LAI and clear-sky DSSR, T_air, VPD, and SM from the ERA5 reanalyzed data were used as predictor variables. Note that SM was 0- to 100-cm volumetric SM, which was weighted average from three levels of ERA5 SM (0 to 7, 7 to 28, and 28 to 100 cm) according to their respective depth.

Because of the hysteretic and asymmetric responses of photosynthesis and ET to environmental drivers, particularly between morning and afternoon (15), we trained the RF model individually for the wet season morning, wet season afternoon, dry season morning, and dry season afternoon. Precipitation was used to define the wet and dry seasons at the grid cell level in each month (fig. S19): Months with precipitation rates higher than the annual mean were defined as the wet season, and months with precipitation rates lower than the annual mean were defined as the dry season (33). For the whole Amazon forest, the dry season was defined from June to October when >50% grid cells were in the dry season. The 100 RF models were used to predict SIF and ET in 1-hour intervals each month from 2015 to 2021 at a spatial resolution of 1° × 1°. The RF-based SIF and ET values were averaged to represent the climatological values over 2015–2021 without the interannual variations. The climatological SIF and ET values in the grid cells where the three ground flux sites (BR-Sa1, BR-Sa3, and GF-Guy) were located were compared with the tower GPP and ET data averaged from multiple years. The RF-based SIF and ET were well correlated with the GPP and ET data from the towers, with the correlation (r) of 0.86 and 0.92, respectively (fig. S20, A and B). Although the magnitude of RF-based ET was higher, such a bias had little effect on the diurnal trends or interannual trends expressed as percentages (%). The GPP-based WUE (WUE_GPP = GPP/LE) was also moderately correlated with the SIF-based WUE (WUE_SIF = SIF/ET) (R² = 0.44; fig. S20C). These results supported the use of RF-based SIF and ET for the diurnal and regional analysis of carbon and water fluxes. The interpolation and spatial and temporal extrapolation abilities of the RF models were evaluated in text S1. Only a small absolute relative bias (|rBias| < 0.2%) was observed for interpolation ability of RF models (fig. S21). In addition, the |rBias| was less than 0.8% for most cases in terms of spatial (fig. S22) and temporal extrapolation (fig. S23). We used the changes in SIF, ET, and WUE_SIF from morning to afternoon as the diurnal variations and from wet to dry season as the seasonal variations (Fig. 2).

Sensitivity analysis

The RF model was also used to calculate the sensitivities of SIF, ET, and WUE_SIF to perturbations in T_air, VPD, and SM, separately. We use the calculation of the sensitivity of ET to VPD as an example below. On the basis of the trained RF model described above, the FPAR and ERA-5 climatic data were used to produce ET (denoted as ET_O). Then, we produced ET (denoted ET_VPD) again by artificially adding one standard deviation to VPD, representing a disturbance, while keeping the other variables unchanged (33). Last, the sensitivity of ET to the disturbance in VPD was calculated using Eq. 6. The process was repeated using 100 RF models, and the means were analyzed to increase the robustness of the results (Figs. 6 and 7 and fig. S11). We also calculated the sensitivity of SIF and ET to T_air by adding one SD and to SM by decreasing one SD.

$$\mathrm{Sens}.(\mathrm{ET})\text{to VPD}=\mathrm{mean}\left[\frac{{\mathrm{ET}}_{\mathrm{VPD}}-{\mathrm{ET}}_{\mathrm{O}}}{{\mathrm{ET}}_{\mathrm{O}}}\right]\times 100\mathrm{\%}$$

Comparison of model_AM + condition_AM and model_PM + condition_AM simulations

The foliar water potential decreases from morning to afternoon due to the loss of water and then recovers during nighttime (17–19). To evaluate the afternoon reductions in photosynthesis and ET in response to the drop in foliar water potential and increases in T_air and VPD, we performed the model_AM + condition_AM and model_PM + condition_AM simulations, in which the morning and afternoon RF models were driven using the same morning inputs (including radiation, FPAR, T_air, VPD, and SM), respectively. By using the same morning inputs for both the morning and afternoon RF models, the effects of varying environmental drivers between morning and afternoon can be factored out. As an example, Eq. 7 was used to calculate the percentage change in ET from the model_AM + condition_AM simulations to the model_PM + condition_AM simulations

$$\text{Change in ET}(\mathrm{\%})=\frac{{\mathrm{ET}}_{\mathrm{PM}}-{\mathrm{ET}}_{\mathrm{AM}}}{{\mathrm{ET}}_{\mathrm{AM}}}\times 100\mathrm{\%}$$

where ET_AM and ET_PM are the predicted ET using the model_AM + condition_AM and model_PM + condition_AM simulations, respectively.

Projection of photosynthesis and ET

The RF-based SIF and ET models were also used to predict diurnal SIF and ET from 2015 to 2100 under different combined SSPs and representative concentration pathways, including SSP245, SSP370, and SSP585, from CMIP6 (36). The climatic data (T_air, SM, specific humidity, surface pressure, and clear-sky radiation) with a 3-hour interval from CMIP6 models were resampled to hourly values (CMIP6 models can be found in table S2). T_air, specific humidity, and surface pressure were used to calculate VPD. Changing species composition could have a large impact on SIF and ET but is beyond the scope of this study and thus is not considered here. FPAR was kept constant at the values observed in 2015. Then, the hourly climatic data and FPAR were used in the trained RF model to simulate hourly SIF and ET. We analyzed relative changes in morning (7:00 to 12:00) and afternoon (12:00 to 17:00) SIF and ET from 2015 to 2100 across the Amazon forest. The morning and afternoon SIF and ET were averaged as daily averages, which were also compared with daily averaged GPP and ET predicted by CMIP6 models (see table S2) These comparisons will improve our understanding of the differences in photosynthesis and ET between satellite observation-based and model-based simulations. In addition, we used different datasets of training (mimicking the current conditions) and validation (mimicking the extrapolation to future higher VPD conditions than those used for calibration) to assess the uncertainty of RF models when making predictions of future scenarios (text S1 and fig. S24). Under the future conditions of SSP245 and SSP370, the |rBias| was less than 2% in all cases (figs. S25 and S26). Under the future conditions of SSP585, the highest |rBias| (4.03%; fig. S27E) was obtained for predicting ET in the wet season morning, followed by the dry season afternoon with the |rBias| of 3.17% (fig. S27H). In other cases, the |rBias| was less than 2% (fig. S27). Overall, these results indicate that the limited biases related to the future extrapolations would not affect our results and conclusions using RF models.

Supplementary Materials

This PDF file includes:

Tables S1 to S3

Figs. S1 to S27

Text S1