Introduction

The Congo Basin, located in the equatorial Africa, hosts the world's second largest, contiguous rainforests (Figure 1). While its annual rainfall (1,500-2,500 mm) is lower than that of the Amazon (1,800-3,200 mm) (Alsdorf et al., 2016), its rainy seasons in boreal spring (March-April-May, denoted MAM) and boreal fall (September-October-November, denoted SON) limit the dry seasons to about 3-4 months in summer (June-July-August, denoted JJA) and winter (December-January-February, denoted DJF), and thus help to sustain the rainforests (e.g., Staver et al., 2011;Mayer & Khalyani, 2011). These two rainy seasons are associated with the north-south migration of the rain belt over tropical Africa that crosses the Congo Basin (Nicholson & Dezfuli, 2013); however, the complex interactions between large-scale atmospheric circulation, mesoscale convective processes, and moisture availability from ocean and terrestrial sources drive their onset and demise (e.g., Nicholson, 2018). These mechanisms that control the variability and changes of the rainy seasons over the Congo Basin are thus poorly understood (e.g., Alsdorf et al., 2016;Nicholson, 2018), leading to large uncertainties in representing its current and future rainfall in climate models (e.g., James et al., 2018;Washington et al., 2013).

Observations have shown that a decrease of rainfall and reduced terrestrial water storage in the Congo Basin have likely led to a decrease in vegetation greenness as well as widespread water deficits between 2003 and 2012 (e.g., Reager et al., 2016;Samba & Nganga, 2012;Zhou et al., 2014). Furthermore, the boreal summer dry season length has likely increased since the 1980s, mainly due to an earlier ending of the spring Abstract The Congo Basin hosts the world's second largest rainforest and is a major rainfall center.

However, the primary sources of moisture needed to maintain this forest, either from evapotranspiration (ET) or advection from the ocean, remain unclear. We use satellite observations of the deuterium content of water vapor ( D  ), solar induced fluorescence (SIF), precipitation, and atmospheric reanalysis to examine the relative contribution of ET to moisture in the free troposphere. We find that SIF, an indicator of photosynthesis, covaries with D  in early rainy seasons, suggesting that ET is an important contributor to atmospheric moisture in both the spring and fall rainy seasons. However, the relative contribution of ET to the free tropospheric moisture varies between the two rainy seasons. Observed D  relative to a range of observationally constrained, isotopic mixing models representative of water vapor coming from land suggests that 83% 9%  of the free tropospheric moisture come from ET in February, and 45% 13%  in April, versus 59% 12%  in August and 31% 12%  in October. Reanalysis indicate that this difference between seasons is due to increased advection of ocean air during the fall season, thus reducing the relative contribution of ET to the Congo Basin in the fall. In addition, ET is the primary atmospheric moisture source in the winter and summer dry seasons, consistent with estimates reported in literature. Our results highlight the importance of ET from the Congo rainforest as an important source of moisture for initiating the rainy seasons.

Plain Language Summary

The Congo Basin hosts the world's second largest rainforest and is a major rainfall center. It has been unclear whether moisture from the adjacent oceans or from evapotranspiration of the rainforests is its main moisture source, especially for initiating the rainy seasons. Using a suite of satellite data, we show that evapotranspiration, especially transpiration, is important for atmospheric moisture throughout the year. However, the relative contribution of ocean moisture from transport increases prior to fall rainy season and not prior to the spring rainy season, indicating that the relative contribution of ET to atmospheric moisture is higher in spring than in fall. Our finding highlights the importance of the rainforests in maintaining and modulating atmospheric humidity. rainy season (Jiang et al., 2019). Enhanced water stress could subsequently alter the composition and structure of the evergreen rainforests over the Congo Basin, which are already vulnerable to abrupt transitions to savanna ecosystem as annual rainfall over much of the Congo Basin is low relative to other rainforests (Alsdorf et al., 2016;Jiang et al., 2019;Staver et al., 2011). These changes in both rainy seasons and rainforest composition thus highlight the need for understanding the mechanisms that control rainy season variability, especially whether and how reduced and degraded rainforests could affect the rainy seasons over the Congo Basin through a change in the supply of moisture from evapotranspiration (ET), as well as changes in latent and sensible heating. For example, numerical model simulations have shown that deforestation, as well as the alteration of the composition of the Congo rainforest toward more drought resistant species, can lead to decreased ET, clouds and rainfall (Bell et al., 2015). Understanding the source of moisture for rainfall over Congo Basin should therefore allow us to better project future changes in the water cycle and its interaction with vegetation over the Congo Basin.

Moisture availability in the lower troposphere is central to rainfall frequency and intensity over the tropics in general (e.g., Bretherton et al., 2004;Holloway & Neelin, 2009;Schiro et al., 2018;Sobel et al., 2004), including the Congo region (Taylor et al., 2018). Previous studies have attributed rainfall variability on interannual to interdecadal timescales over the Congo Basin to sea surface temperature anomalies (SSTA) over the tropical Pacific, Atlantic, and Indian oceans through their influences on moisture transport (Balas et al., 2007;Dai, 2013;Diem et al., 2014;Hoerling et al., 2006;Hua et al., 2016;Nicholson & Dezfuli, 2013;Pokam et al., 2014;Tamoffo et al., 2019). Other studies, using methods such as global models or back trajectories, have found large recycling ratios (over 50%) in the Congo Basin, indicating ET is an important source of moisture for the area (e.g., Risi et al., 2013;van der Ent et al., 2010). It is important to distinguish that here we consider the fraction of water vapor in the atmosphere over the Congo Basin originating from land ET anywhere (e.g., Risi et al., 2013;van der Ent et al., 2010;Yoshimura et al., 2004), rather the typical recycling ratio that calculates the fraction of water vapor originating from land ET only within the domain (e.g., Eltahir & Bras, 1996;Trenberth, 1999).

Determining the contributions of free-tropospheric moisture from ocean evaporation versus ET from vegetation in the Congo Basin is therefore key for developing a better understanding of the relative influences on precipitation from external SSTA, internal land vegetation, and land-use changes. However, most recycling rates heavily rely on model or reanalysis to determine the contribution of ET to precipitation (e.g., Risi et al., 2013;van der Ent et al., 2010). Measurements of the isotopic composition of rainfall and rivers, when combined with reanalysis-based wind fields and satellite observations of rainfall have been shown to be useful for quantifying the sources of precipitation for different regions of Africa. For example, Levin et al. (2009) has used in situ measurements of oxygen 18 ( 18 O  ) and D  from rivers, as well as Tropical Rain- fall Measuring Mission (TRMM) precipitation and wind fields, to show that oceanic moisture is the primary source of precipitation for Kenya, while ET is the primary source of precipitation for Ethiopia. However, in situ measurements of isotopes are virtually unavailable over the Congo Basin. Satellite measurements of D  are sensitive to their oceanic versus terrestrial sources over the global tropics and can therefore be used to identify moisture sources and processes controlling atmospheric humidity (e.g., Brown et al., 2008;Risi et al., 2010;Worden et al., 2007). We use such satellite observations of D  , rainfall, and photosynthesis, along with reanalysis to determine the relative importance of the ET versus moisture transport to the Congo Basin atmospheric moisture using an approach similar to Wright et al. (2017). These data therefore allow us to test a hypothesis that ET is the primary moisture source for the Congo Basin rainy seasons.

Data

The deuterium content of water is expressed as the relative ratio of the number of HDO molecules to the total number of H 2 O molecules in parts per thousand (‰) relative to the isotopic composition of ocean water as shown below:

where R is the ratio of the number of HDO molecules to the total number of H 2 O molecules and R std is the corresponding ratio in a reference standard, taken here to be the Vienna Standard Mean Ocean Water:

 (e.g., Wright et al., 2017 and references therein). The isotopic composition of water vapor in the free troposphere is due to a mixture of air parcels originating from different sources (Galewsky, 2018;Galewsky & Hurley, 2010;Galewsky et al., 2016). Free-tropospheric D  measurements are from the Na- tional Aeronautics and Space Agency (NASA) Aura Tropospheric Emission Spectrometer (TES) satellite instrument. The analysis period of this study is 2005-2011 when the quality of TES D  data are suitable for our analysis. We use the monthly mean TES version 6 (v006_Litev01.00) Level 2 volume mixing ratios of D  in this study. The following quality flags were used when retrieving the data from the TES satellite: "Species Retrieval Quality = 1," "Degrees of Freedom for Signal > 1," and "Average Cloud Optical Depth < 0.4" in order to ensure good quality data as suggested by Worden et al. (2012). The accuracy of these data is ∼6 per mil with a precision of 20 per mil (Worden et al., 2012) for the vertical range used in this analysis (∼900-420 hPa, or about 1-6 km above sea level). For this study, we use the average D  over this vertical range.

Measurements of solar induced fluorescence (SIF) provide a nearly direct estimate of photosynthesis, a prerequisite for transpiration (e.g., Frankenberg et al., 2011). We can therefore use SIF to indicate the occurrence of transpiration, one of the main components of ET. A caveat is that transpiration also depends on plant water use efficiency, vapor pressure deficit, and radiation (e.g., Boese et al., 2017) so that we might not expect a one-to-one relationship between SIF and ET. We use SIF estimates from the GOME-2 V26 740 nm data products (Joiner et al., 2013) as their observational period overlaps with most of the TES record (2007)(2008)(2009)(2010)(2011). The precipitation estimates taken from TRMM are from the 3B43 gridded monthly average estimate at a horizontal resolution of 0.25 0.25   . TRMM precipitation estimates are generated using a combination of microwave and radar sensors on the instrument that are calibrated with gauge data from the Global Precipitation Climatology Center (GPCC) (Huffman et al., 2007).

ET data comes from a combination of reanalysis and the Moderate Resolution Imaging Spectroradiometer (MODIS) as described in Fisher et al. (2009). While there are a variety of choices of ET and precipitation, we are using these data primarily for qualitative comparisons to give an understanding of what is likely known. A full error analysis of ET and precipitation products is beyond the scope of this study, but described in the literature (Fekete et al., 2004;Fisher et al., 2009;Munier & Aires, 2018;Pan et al., 2020;Rauniyar et al., 2017).

We also use the Fifth Generation of the European Center for Medium-Range Weather Forecasts (ECMWF) Reanalysis (ERA5) monthly means of daily means of the longitudinal (u) wind component, latitudinal (v) wind component, and vertically integrated moisture flux divergence on a 0.25° × 0.25° grid (Hersbach et al., 2020, https://www.ecmwf.int/en/forecasts/datasets/reanalysis-datasets/era5). The reanalysis is a four-dimensional (i.e., space and time) data assimilation product that combines observations with model forecasts to estimate the dynamic and thermodynamic structures of the global atmosphere. Although the quality of reanalysis in general has been open to debate (Dee et al., 2011;Pan et al., 2020;Thorne & Vose, 2010), using this data in conjunction with observation-based data allows us to corroborate information of the moisture source that cannot be directly provided by observations.

Comparison of Isotope Observations With Theoretical Mixing and Rayleigh Models

Deuterium Content of Water

We can use the isotopic composition of an air mass to trace its source to either vegetation or ocean because D  values contributed by ocean evaporation are distinctively different from those by rainforest ET. Rainwa- ter will generally be more isotopically enriched (or heavier) than the source vapor because heavier isotopes preferentially condense (Risi et al., 2020;Tremoy et al., 2014;Worden, et al., 2007). Since the deuterium content of transpired water is relatively unchanged from that of the original source, for example, the isotopically heavier rainwater (Risi et al., 2013), it will be more enriched than the vapor evaporated from its oceanic source. Therefore, by examining the differences in deuterium content of the water vapor over a heavily vegetated surface, one can separate water coming from the ocean from water coming from plants (Wright et al., 2017).

The observed values of free-tropospheric deuterium content also depend on the type of convection. Some form of mixing between the surface and free troposphere is required to transport surface fluxes into the air parcels observed by the satellite. For example, shallower convection detrains near the midtroposphere, enriching D  of the midtropospheric water vapor (Lacour et al., 2018). Deeper convection, while also mixing air between the surface and free troposphere, is associated with other processes beside condensation that affect the isotopic composition of vapor such as rainfall evaporation (Lacour et al., 2018;Worden et al., 2007). The latter is a confounding factor in our conclusions. For example, Nlend et al. (2020), using a back trajectory method, suggests that D  of rainfall is primarily controlled by upstream mesoscale convective systems in- stead of ET in West-Central Africa. On the other hand, other research such as from Moore et al. (2014) and Bailey et al. (2017) suggest that the free-tropospheric signal is primarily related to moisture convergence and hence the balance between evaporation and precipitation, even during times of deep convection. Our interpretation of the moisture sources is less influenced by these uncertainties during the months before and after the peak of each rainy season when deep convection is expected to influence D  .

Mixing and Rayleigh Models

We first quantify the seasonality of the isotopic composition of water vapor in the free-troposphere and then compare such parcels with two models, a mixing model and a Rayleigh model, to identify which air parcels are likely influenced by land versus ocean (Noone, 2012). A mixing model describes what happens to a mixture of two air masses with different water vapor isotopic compositions:

where 0 q and 0  are the specific humidity and its D  value of the dry air mass in the upper troposphere, and F q and F  are the specific humidity and D

 value of the air mass at the surface.

0 F q q q   is the specific humidity of the mixed air mass between dry air mass in the upper troposphere and humid air mass from its surface source (Noone, 2012). This model considers two possible moisture sources for the air mass sampled by TES D  measurements: air with water vapor transpired from the rain- forest, and air with water vapor evaporated from the ocean. Generally, the isotopic composition of vapor sourced from vegetation over the tropical land ( F  for vapor from ET) is between 75‰  and 0‰ while the isotopic composition of vapor sourced from the ocean ( F  for vapor from ocean) is lower than 75‰  (Risi et al., 2013).

We also examine the observed D  values in relation to a Rayleigh model, which describes the change of D

 with water vapor mixing rate as liquid water evaporates in equilibrium with temperature:

where  is set to equal the temperature-dependent equilibrium fraction- ation factor between liquid and water vapor (Majoube, 1971). Under the Rayleigh distillation model, as an air mass moves upward (or toward cooler conditions), condensate is completely removed immediately after it forms under the assumption of pseudo adiabatic process (Galewsky & Hurley, 2010;Wright et al., 2017). During deep convection, air parcel moisture is more isotopically depleted than expected from Rayleigh models because of a combination of processes that occur during convection such as rainfall evaporation and entrainment of isotopically depleted air (Worden et al., 2007).

This combination of mixing and rainfall processes compared to observed We assess if our assumptions for the isotopic mixing models and Rayleigh models significantly affect our estimates of the relative contribution of ET versus moisture transport from ocean to the atmospheric moisture. For example, Risi et al. (2013) suggests that convection and large-scale circulation can influence continental recycling estimates based on these free-tropospheric deuterium data. Our approach for setting the bounding conditions for these models is based on Noone (2012): we first found the range of specific humidities at 421 mb ( 0 q ) and near the surface at 1,000 mb ( F q ) in Equations 2 and 3 as observed by TES by (a) finding their mean and standard deviation and (b) choosing 10 values between the standard deviation about the mean and the standard deviation added to the mean. Then, we determined how the fraction of D  above the uppermost, land-based, mixing (solid green) model line (hereby denoted as f) changed between February and August for the different values of 0 q and F q obtained. The mean difference between f estimated by these initial values in February and that in August is 0.199 0.005  , where the error is the standard deviation. Though there is a chance that the range of our initial values do not capture the true initial values, the small range of uncertainty of the f difference suggests that our results are not too sensitive to initial values. Therefore, while we show a range of mixing models in Figure 2, hereafter, we only consider the mixing model with the mean values of observed 0 q and F q when calculating f.

In February (Figure 2a), 83% 9%  of the data lies above the upper-most, land-based (solid green) mixing line, indicating that this water vapor largely originated from land and then was transported into the free troposphere. We should consider this fraction to be a lower bound on the amount of ET contributing to the Congo Basin atmospheric moisture because rainfall processes as well as vapor originating from oceans will decrease D  . The density contours of D  further suggests a strong con- tribution from terrestrial sources, as most of the parcels are concentrated between ∼-60‰ and -100‰, values well above the mixing and Rayleigh models for the ocean (dark gray lines) but within the ranges of landbased mixing models. Finally, a linear line of best fit shows an increasing trend toward more enriched D  values as specific humidity increases, consistent with a strong contribution from a terrestrial source (e.g., Risi et al., 2013). In August, 59% 15%  of the data lies above the land mixing line (Figure 2b), indicating that ET contributes to more than half of the moisture in the atmosphere during the transition month to the fall rainy season, though less than that in the spring. The density contours of D  further overlap mixing and Rayleigh models from both ocean and land sources, also suggesting a mixed contribution from terrestrial and oceanic sources. Finally, a linear line of best fit reveals declining D  values with increasing specific humidity, consistent with reduced continental recycling (Risi et al., 2013). While it is challenging to uniquely identify what combination of moisture sources and rainfall and mixing processes affects the observations below the ocean mixing line using this method, Figure 2 highlights the change in relative importance of ET and oceanic moisture from transport as moisture sources for the boreal spring and fall rainy seasons in the Congo Basin.

Results

The Relative Contributions of ET and Ocean Evaporation to Atmospheric Moisture

Comparison of the isotopic data to the mixing and Rayleigh models (Figure 2) shows that ET likely contributes more moisture in the transition month to the spring rainy season (February) than in the transition month to the fall rainy season (August). To investigate the relative contributions of ET to moisture throughout the year, we examine D  , precipitation, and , f the fraction of the observational samples for which D  exceeds the upper-most land-based mixing model (the top solid green line in Figure 2) in Figure 3.

Precipitation and D

 (Figure 3a) show opposing trends in general: increased rainfall corresponds with de- pleted D

 values in rainy seasons and vice versa in dry seasons. This could indicate one of three possibil- ities: (a) condensation plus rainfall evaporation decreases the deuterium content as discussed previously;

(b) increased transport of ocean evaporation into the Congo Basin during peak times could deplete D  measurements and signify an increased contribution of ocean evaporation to atmospheric moisture; or (c) some combination of both are responsible for depleted D  measurements. It is important to note that in the cases of (a) or (c), ET could still have a relatively significant contribution to atmospheric moisture, but precipitation also has a substantial impact on deuterium content.

Compared to the two rainy seasons, the relationship between precipitation and D  is less consistent during the two dry seasons. During the summer dry season, precipitation is generally low, while D

 is enriched WORDEN ET AL.

10.1029/2020JG006024 6 of 14 compared to the rainy seasons, but lower compared to the winter dry season. Because precipitation is lower, this could indicate that rainout has less impact on depleting D  during the summer dry season and instead it is possible that ocean evaporation advected into the region plays a larger role in contributing to moisture in the summer dry season compared to the winter dry season. During the winter dry season, D  is enriched max- imally, despite increased precipitation from January to February. This could indicate that ET is the main contributor to moisture for the winter dry season and the onset of the spring rainy season in February.

In general, f follows a similar seasonal pattern to D  in Figure 3a: f is higher during the two dry seasons in winter and summer and lower during the two rainy seasons in spring and fall (Figure 3b). Maximum f (0.83 0.09  ) occurs during February, suggesting that increasing vapor from the land areas with high ET is the main reason for increasing free tropospheric specific humidity ( 0 q ) prior to and during the onset of the spring rainy season. In contrast, f is lower during August (0.59 0.15  ) than during February, suggesting that increasing advected ocean evaporation contributes to the increasing 0 q during the transition to fall rainy season. However, throughout the winter and early spring, f remains above  , meaning that most of the observed D  measurements dur- ing those months must come from land. This indicates that ET is the main source of moisture for the dry seasons. As the rainy seasons reach their peaks, f reduces to 0.45 0.13  in mid-spring (April), and 0.31 0.12  in midfall (October). Throughout the rainy seasons, f only remains above Unfortunately, the D  observations cannot directly quantify the relative contribution of ocean and land sources without further knowledge from other observations. Reanalysis and remotely sensed data, on the other hand, can provide quantitative information on the Congo Basin moisture budget. Due to a lack of observational data, many previous studies have used reanalysis to analyze the moisture sources and dynamics of the Congo Basin. However, monthly reanalysis estimates of tropical precipitation and ET are known to be erroneous due to lack of observational constraints in tropical Africa (e.g., Fekete et al., 2004;Hua et al., 2019). As we do not have independent data sets to validate the current ET datasets and provide the most accurate ET and precipitation data sets over the Congo for this study, we choose to use moisture flux convergence (MFC) from ERA 5 reanalysis as well as precipitation from TRMM, and ET from a MODISbased product (Fisher et al., 2009) for the purpose of examining the seasonality of the moisture budget (e.g., Shi et al., 2019). We examine the seasonal cycles of the moisture budget components, as well as calculate the fractional contributions of ET and MFC to precipitation, to compare their relative contributions to precipitation in Figure 4.

Figure 4a shows that precipitation varies between 3.0 and 5.4 mm/day over the winter-spring period, and between 3.0 and 6.6 mm/day in summer-fall period, respectively. ET from the MODIS-based data set contributes about 3.0 mm/day of the moisture to atmospheric moisture during winter and summer dry seasons and increases slightly to 3.3 mm/day in spring and fall rainy seasons. The net moisture transport generally contributes much less to the atmospheric moisture than ET: from negative (moisture export) 0.5 mm/day in winter to positive (moisture import) 1.5 mm/day in spring, and from negative 0.5 mm/day in summer to positive 2.6 mm/day in fall. Note that there is a general imbalance of 0.5 mm/day between precipitation estimated by TRMM and the sum of ET and net moisture transport estimated by MODIS and ERA5, although their seasonal variations are consistent. Figure 4b shows the contributions of ET and MFC to precipitation within our Congo domain, as indicated by ET/P and MFC/P. ET/P shows consistently high values throughout the year and MFC/P is always lower than that of ET/P (e.g., Burnett et al., 2020;Crowhurst et al., 2020). These data support our conclusion that there is a larger relative contribution of ET to the moisture for the spring versus fall rainy seasons. We also compare the seasonal cycle of f to that of ET/P (Figure 4b). f represents the fractional ET contribution from both our Congo domain and upstream regions to the free tropospheric moisture within our Congo domain, whereas ET/P represents the local ET contribution to precipitation all within our Congo domain.

In addition, f represents a lower bound of the contribution of land to atmospheric water vapor due to choosing a high D  threshold for moisture from land as well as the potential D  depletion from precipita- tion. However, we believe it is still useful to compare as their common component is ET and ET/P is widely cited in literature as a local precipitation-recycling indicator. The seasonal variations of f and ET/P can be similar during the dry seasons when they are dominated by the change of local ET as moisture transport from ocean and precipitation are moderate. During the peak rainy seasons, variations of f and ET/P are dominated by a significant increase of precipitation and moisture transport from ocean. However, in the early spring rainy season (February-March), the variations of f and ET/P are distinctively different. This could be for a variety of reasons. For example, it is possible that the difference between f and ET/P originates from scale differences: f is a metric of the relative contribution of ET to atmospheric moisture from anywhere, whereas ET/P represents the fraction of precipitation that originates from ET in the same grid box. ET/P could decrease as the increase in P outpaces the local increase in ET during February-March, but the increase in P could be outpaced by the increase both regional and local ET, as represented by f . Furthermore, the disagreements could be due to the quality of the data. Studies show seasonal ET cycles from models, reanalysis, and other remotely sensed products that vary greatly in their magnitude of seasonality (e.g., Burnett et al., 2020;Crowhurst et al., 2020). The decreased sensitivity of remotely sensed ET during the wet seasons (Fekete et al., 2004;Pan et al., 2020), or the uncertainties of ET estimates based on MODIS in evergreen tropical rainforests (Paca et al., 2019) could lead to underestimates of the seasonal variations of ET.

Interpretation of These Relative Contributions Based on Spatial Patterns

As discussed previously, the deuterium-based data provides a lower bound estimate on how ET contributes seasonally to atmospheric moisture over the Congo. In this section, we discuss additional evidence based on analysis of moisture flux convergence and winds as well as satellite-based photosynthesis measurements. To examine the role of winds in bringing in moisture to the region, we investigate the spatial patterns of moisture flux convergence (MFC), precipitation, 800-875 hPa averaged wind, D  , and SIF over the Congo Basin for the following months: (a) during the transition month (February) to the spring rainy season, and the peak month of the spring rainy season (April) and (b) during the transition month (August) to the fall rainy season the peak month of the fall rainy season (October).

Figure 5a shows that in February, enriched

) is concentrated within the Congo Basin, with the exception of enriched values to the northeast where high mountainous ranges penetrate further into the atmospheric boundary layer and push high D  values from the atmospheric boundary layer into the free troposphere. D  is highest in the middle and east of the Congo Basin (20 30 E    ), corresponding with relatively high SIF and rainfall (Figure 5b). This covariation of D  values and SIF indicates that transpiration is one of the main components of ET contributing to atmospheric moisture during this time as photosynthesis co-varies with transpiration (e.g., Boese et al., 2017 and references therein). The increase of rainfall does not contribute to the increase of D  , as it would preferentially remove D  in the atmosphere. The covariance between SIF and rainfall suggests that vegetation photosynthesis also increases with an increase of rainfall.

Furthermore, no winds from either the Atlantic Ocean or Indian Ocean bring depleted D  into the Congo Basin (Figure 5b), and there is little moisture flux convergence the basin (Figure 5a). These conditions all point toward ET being the primary source of moisture in the spring rainy season, generally consistent with Sorí et al. (2017). While March D  values (not shown) remain similar to its distribution in February, D  values in April (Figure 5c), the peak of the boreal spring rainy season, show that enriched D  values occur mostly in the northern part of the basin, which corresponds with relatively higher SIF. However, in general, D

 is more depleted than in February. This is likely due to a combination of (a) winds bringing in moisture from the Indian Ocean and (b) the rainout process as described in Section 3, as indicated by high rainfall within a larger portion of the basin. This does not necessarily mean that the contribution of ET to atmospheric moisture decreases, only that its relative contribution drops as moisture from the ocean or the monsoon region increases. Overall, Figure 5 shows enriched D  , high corresponding SIF, and lack of mois- ture transport by winds, in February, indicating that ET is the main moisture contributor for the transition to spring rainy season, while in April, the presence of winds from the Indian Ocean combined with depleted D  indicate a higher relative contribution of advected ocean evaporation.

In August (Figure 6a southern part of the basin correspond with relatively high SIF. While the highest SIF is located in the northern part of the basin, D  values there are depleted likely because winds carry moisture from the Atlantic Ocean into that part of the region, which mixes with the enriched D  from plants. In October, during the peak of the boreal fall rainy season (Figure 6d), highest SIF concentrates in the eastern part of the basin, mostly likely as a result of increased precipitation. Since Figure 6d indicates a lack of low-level westerlies reaching the Congo Basin from the Atlantic, moisture is most likely brought in by strong meridional flow related to the African Easterly Jet (AEJ) from areas with mesoscale convective systems (Cook & Vizy, 2016;Pokam et al., 2012). These winds bring depleted D  to the Congo Basin (Figure 6c), which partially offsets the positive influence of ET on D  . Rainfall processes also likely decrease

Overall, in Figure 6, depleted D  values (compared to February) spatially correspond to moisture transport from the Atlantic Ocean in August, while rainout and possible moisture brought in by the AEJ deplete D  over the entire Congo Basin in October.

Discussion

Recycling ratios from previous studies indicate significant contributions of ET (greater than 50%) to atmospheric moisture throughout the year (e.g., Nicholson et al., 1997;Risi et al., 2013); however, while it is desirable to compare our assessment of the role of ET in contributing to atmospheric moisture to that suggested by previous studies that calculate continental recycling ratios, a direct comparison is not currently feasible. This is because the recycling rate in literature is mostly defined either as the fractional contribution of ET to precipitation or by local ET versus moisture advected from outside regions, whereas D

 is linked to the ratio of water vapor from ET to the total water vapor in the atmosphere (Risi et al., 2013). In addition, D  is influenced by ET both locally and along the path of the air mass, including ET from areas outside the Congo Basin when wind, and thus moisture advection, is strong. Therefore, we can only compare previous studies that quantify the ET contribution to water vapor considered from regions both inside and outside the Congo Basin domain (e.g., Risi et al., 2013;van der Ent et al., 2010;Yoshimura et al., 2004). For example, Risi et al. (2013) used a combined D  observations by TES and a water vapor tagging approach in a climate model. Their result suggests that ET provides more than half of the atmospheric moisture (about 65%) in the winter and summer dry seasons in the Congo region. Our estimates are qualitatively consistent with their result for these two dry seasons, although they suggest that the ET contribution could be somewhat higher than their model values based on comparison with the TES data. Furthermore, Pokam et al. (2012) suggests that ET influences spring rainy season changes more than fall rainy season changes. Our results clarify that this stronger influence in spring is likely because ET is the primary source of atmospheric moisture during the spring rainy season; therefore, changes in ET likely impact rainfall in spring.

Is it possible that the relatively higher ET contribution to atmospheric moisture for spring versus fall is not just due to increased contribution from advected ocean evaporation to atmospheric moisture for the fall rainy season? Previous studies have found a higher ET contribution to the moisture budget during the spring rainy season compared to that of the fall rainy season, which could also help explain why ET contributes more moisture for the spring rainy season. On the other hand, Burnett et al. ( 2020) calculated basin-scale ET in the Congo using water-balance methods, positing that increased radiation, as well as the availability of soil moisture, can explain increased ET in the spring: Higher direct photosynthetically active radiation (PAR) fractions combined with higher net solar radiation increase water use efficiency, while increased terrestrial water storage increases the amount of water available for transpiration. Crowhurst et al. (2020), using global climate models, also calculated ET over a similar domain and found that it was higher in the spring rainy seasons compared to the fall rainy season. They attributed changes in leaf area index and vapor pressure deficit to changes in transpiration between the spring and fall rainy season.

How does the relative contribution of ET to atmospheric moisture over Congo Basin compare to that over the Amazon basin? D  values are on average more enriched over the Congo Basin than over the Amazon rainforest.

In the Congo Basin, D  values are on average 88 ‰  annually, 87 ‰  during the boreal spring rainy season, and the summer dry season, 94 ‰  during the boreal fall rainy season, and 79 ‰  during the boreal winter dry season (Figure 3a). In contrast, D  values over the Amazon are on average 130 ‰   annually, 139 ‰  during the wet season (October-May), and 126 ‰ 

during the dry season (June-September). These differences are broadly consistent with a higher recycling rate in the Congo than in the Amazon as suggested by some studies (e.g., Nicholson et al., 1997;Risi et al., 2013). The recycling rate over the Amazon ranges from about 30% (e.g., Staal et al., 2018) based on reanalysis data to about 50% (Salati et al., 1979) based on isotopic composition of the stream flows. A higher ET contribution to atmospheric moisture in the Congo Basin versus the Amazon basin, as suggested by D  data, is also qualitatively consistent with the ET estimated from water balances using rainfall and runoff data. The annual ET is estimated to contribute to 75%-85% of the annual rainfall in Congo Basin versus 50%-62% in the Amazon basin (Alsdorf et al., 2016;Fernandes et al., 2008;Molion, 1975). Therefore, ET is central in determining the climate variability and change of the water cycle in Congo Basin which has a rainfall regime drier than the Amazon. Our results imply that the loss of rainforests due to land use, biomass burning, and climatic drying in Congo (Bell et al., 2015;Staal et al., 2016) will likely have a greater impact on rainfall in the Congo than in the Amazon basin, especially for the spring rainy season.

Conclusions

Most previous studies of the Congo Basin rainy seasons have focused on clarifying the impacts of moisture transport from oceans on rainfall variability and changes. In contrast, the role of vegetation and its modulation from land-use and deforestation have been far less clear due to lack of adequate observations. Using a suite of satellite measurements, including the deuterium content of water vapor and SIF, we show that plant-transpired water, lifted into the free troposphere, is a primary moisture source for the atmosphere during boreal winter (DJF) and summer (JJA). However, both moisture advected from oceans and water transpired from the Congo rainforests are important moisture sources during the rainy seasons (March-May and September-November).

Specifically, the climatology of D  points toward ET being the main initial contributor to atmospheric mois- ture during February, the transition month to the spring rainy season but less so in August, the transition month to the fall rainy season. f, the fraction of the observational samples for which D  exceeds that of the uppermost land-based mixing model line (Figure 2), is 83% 9%  in February versus 59% 15%  of total water vapor samples in the Congo domain in August. Enriched D  in February corresponds with relatively high SIF and a lack of winds bringing moisture in from the Atlantic Ocean, while relatively depleted D

 in

August is due to winds bringing in moisture from the Atlantic Ocean and mixing with the enriched D  from areas with high SIF within the Congo Basin. As the rainy season reaches its peak in both spring and fall, D  decreases and the fraction of the observed water vapor samples most likely from ET reduces to 45% 13%  in April and 31% 12%  in October as expected from deep convection and precipitation, and increased moisture transport from the ocean. In general, the climatological seasonal cycles of ET, precipitation and moisture flux convergence (MFC) derived from satellite observations and reanalysis are consistent with the D  results in that both show a higher ET contribution to atmospheric moisture during dry seasons than during the wet seasons. However, neither of these estimates of ET can capture the high ET contribution in the winter/early spring as shown by the D  data.

Our results imply the need to evaluate the possible change of photosynthesis and ET seasonality as a potential contributor to changes to the spring rainy season, in addition to other potential external forcings, such as SSTA, and changes to moisture transport and the African Easterly Jet. For example, observations have shown an earlier than normal onset and demise of the spring rainy season (Jiang et al., 2019;Taylor et al., 2018) over this region due to increased rainfall in February and decreased rainfall in May-June. Furthermore, our results raise several questions: Would the spring rainy season disappear or substantially weaken if ET were substantially reduced by rainforest loss? Are the mechanisms for the onset of spring rainy season significantly different from those of the fall rainy season? What roles do shallow convection versus lower tropospheric circulation play in lifting plant transpired water vapor from the surface to the free troposphere? Why does ET contribute more to the atmospheric moisture in the Congo than in the Amazon basin? Further study is needed to better elucidate these different mechanisms and discover how they might change with climate and land use.