The light-deficient climates of western Central African evergreen forests

Rainfall thresholds under which forests grow in Central Africa are lower than those of Amazonia and southeast Asia. Attention is thus regularly paid to rainfall whose seasonality and interannual variability has been shown to control Central African forests’ water balance and photosynthetic activity. Nonetheless, light availability is also recognized as a key factor to tropical forests. Therefore this study aims to explore the light conditions prevailing across Central Africa, and their potential impact on forests’ traits. Using satellite estimates of hourly irradiance, we find first that the four main types of diurnal cycles of irradiance extracted translate into different levels of rainfall, evapotranspiration, direct and diffuse light. Then accounting for scale interactions between the diurnal and annual cycles, we show that the daily quantity and quality of light considerably vary across Central African forests during the annual cycle: the uniqueness of western Central Africa and Gabon in particular, with strongly light-deficient climates especially during the main dry season, points out. Lastly, using an original map of terra firme forests, we also show that most of the evergreen forests are located in western Central Africa and Gabon. We postulate that despite mean annual precipitation below 2000 mm yr−1, the light-deficient climates of western Central Africa can harbour evergreen forests because of an extensive low-level cloudiness developing during the June–September main dry season, which strongly reduces the water demand and enhances the quality of light available for tree photosynthesis. These findings pave the way for further analyses of the past and future changes in the light-deficient climates of western Central Africa and the vulnerability of evergreen forests to these changes.


Introduction
The climate conditions under which forests grow in Central Africa are still not well known, as it is one of the most under-studied climatic areas of the world. This relates mainly to the dramatic lack of in situ climatic measurements (Washington et al 2013, Bigot et al 2016.
As compared to Amazonian and southeast Asian forests, Central African forests grow under lower rainfall thresholds: mainly below 2000 mm yr −1 although noteworthy spatial variations exist between wetter west and east margins (>2000 mm), and drier north and south ones (∼1500 mm, figures S1(a), (e) is available online at stacks.iop.org/ERL/14/034007/mmedia). This explains that rainfall is acknowledged as the main environmental factor controlling forest photosynthesis and functioning in Central Africa (Gond et al 2013, Guan et al 2015, Cherrington et al 2016. Contrarily to Amazonian forests where the mean seasonality in photosynthetic activity is mostly driven by light availability, i.e. the highest photosynthesis levels are recorded during the luminous dry season (Huete et al 2006, Myneni et al 2007, Bi et al 2015, Wagner et al 2017, in Central Africa, the two seasons of highest (lowest) photosynthesis, i.e. March-May and September-November (December-February and June-August), are concomitant with the two rainy (dry) seasons (Gond et al 2013, Philippon et al 2016. Indeed, below 2000 mm yr −1 on average, the water stored during both rainy seasons is not expected to meet the main dry season water demand and to maintain evergreen forest. However, recent studies also show that the western part of Central Africa is one of the most cloudy region across the tropics (Wilson andJetz 2016, Dommo et al 2018), which accounts for its low mean incoming solar radiation (irradiance hereafter) at the surface (figures S1(b), (f)). Moreover an analysis of the diurnal cycles for northern Congo (Philippon et al 2016) shows that neither the two rainy seasons nor the two dry seasons resemble each other in terms of cloud cover, irradiance and rainfall, which translates into different levels of photosynthetic activity. In spite of that, a clear picture of the light conditions prevailing across Central Africa is still lacking.
It is also noteworthy that forests are usually seen as a uniform green block across the whole Central Africa (Hansen et al 2008, Verhegghen et al 2012. According to authors 'lowland rainforests' or 'dense moist forests' might encompass a huge variety of forest types from wet Atlantic forests toward moist semi-deciduous forests further inland (Caballé 1978, White 1983, Letouzey 1985, Fayolle et al 2014b. Further divisions have also been observed locally (Viennois et al 2013) according to the geological substrate and disturbance history (Fayolle et al 2012(Fayolle et al , 2014a. A forest typology, i.e. evergreenness versus deciduousness, at the regional scale is currently lacking despite its importance for the improvement of the modelling of land-atmosphere interactions.
Here, we examine for the first time light available for trees across Central Africa and its potential control on forest traits, particularly forest evergreenness.
To that aim and expanding on a previous study (Philippon et al 2016), we extract first the main types of diurnal cycles of irradiance. Then we explore their implication on variables important for forests such as potential evapotranspiration (PET) (i.e. water demand) and diffuse irradiance (i.e. quality of light). Lastly, accounting for scale interactions between the diurnal and annual cycles, we develop a novel irradiance-based climatic regionalization for Central Africa, that we cross with an original map of forests types. The main hypothesis of our study is that the low irradiance recorded in western Central Africa by reducing the water demand but associated with a better quality of light, enables the existence of evergreen forests under conditions which are drier than in Amazonia and southeast Asia.

Data and methods
Central Africa is defined here as the region encompassing latitudes 8°S-7°N and longitudes 8°E-30.5°E. Because we focus on the climate conditions under which forests grow, the non-forested land pixels and oceanic pixels are masked out. Only a rapid overview of data used is given hereafter as details are provided in the supplementary information. The study period is 2005-2013, dictated by the availability of irradiance data.

Data
The diurnal and seasonal cycles of light received at the surface are documented using mainly direct normalized irradiance data (DNI) produced by the CMSAF (Müller et al 2015) (Satellite Application Facility for Climate Monitoring).
To understand variations observed in DNI and their implications in terms of water budget and light quality, information on cloudiness, rainfall, land surface temperature, relative humidity, PET and photosynthetically active radiation (PAR) are analysed using various datasets. This enables the extraction of robust and coherent results across datasets, despite the uncertainties associated with the data. Cloudiness data come from the Satellite Application Facility for Support to Nowcasting and Very Short Range Forecasting Cloud Type (SAFNWC) product (Derrien and Le Gléau 2005). The TRMM 3B42v7 rainfall data is used as the reference rainfall dataset because its 3-hourly time resolution enables to document the diurnal cycle (Huffman et al 2007). Daily land surface temperatures come from the Berkeley Earth Temperature Study (Rohde et al 2013) while monthly relative humidity data originate from the Vmerge project database (Jones and Wint 2015). We also used daily PET provided by the (Senay et al 2008) National Oceanic and Atmospheric Administration's Global Data Assimilation System (GDAS) and monthly PET from World-Clim (Fick and Hijmans 2017). Total and diffuse PAR (400-700 nm) come from the Breathing Earth System Simulator (BESS) daily products (Ryu et al 2018).
Lastly, expanding on a forest typology proposed for northern Congo (Gond et al 2013), forests traits are documented using the enhanced vegetation index (EVI) at a 16-day temporal resolution issued from MODIS (Huete et al 2002), and crossed with forest inventory data and existing vegetation maps (see supplementary information).

Extraction of the four main types of DNI diurnal cycles
Following a methodology previously developed (Philippon et al 2016) and extending it for the whole of Central Africa, the detection of the four main types of diurnal cycles recorded over Central Africa for DNI relies on a classification of the whole set of 788 400 diurnal cycles of hourly DNI (365 days * 9 years from 2005 to 2013 and 2400 pixels), and not as commonly found on mean diurnal cycles computed on predefined seasons (Malhi et al 2002).
The extraction is done thanks to a k-means clustering algorithm which agglomerates data around randomly chosen seeds and iteratively finds the partition which minimizes the variance within clusters, for a given number of clusters (Michelangeli et al 1995). As preliminary steps, we (1) filter the diurnal cycles (lowpass band at 6 h), and (2) normalize them (by subtracting the annual hourly mean). Given the large number of observations available (N=2400 pixels * 365 days * 9 years for 24 variables), the k-means is first performed on a subset of 2000 randomly chosen observations. The number of iterations, i.e. the number of random seeds is set to 200. The number of clusters varies from 3-5 and the 4-cluster solution is retained as the partition for which the four types seem, as expected, mostly driven by the first and second harmonics of the diurnal cycle, that is (1) the opposition between anomalously dark and bright conditions all along the day (∼amplitude of the 1st harmonic) which is (2) superimposed on the temporal phase of the second harmonic (i.e. maximum near either the morning or the afternoon).

Regionalization of Central Africa according to DNI diurnal cycles
To delineate irradiance-based regions across Central Africa, a k-means clustering is applied to the daily frequency of the four types of DNI diurnal cycles across the 2400 forest pixels. First, the daily mean frequency over the 9 years of the study period, for the four types of diurnal cycles, is computed for each pixel. Then a 30-day low-pass filter is applied to remove the high frequency variability. We obtain four mean annual cycles, which are concatenated so that we attain a matrix of 365 days * 4 types as variables and 2400 forest pixels as observations to be clustered. We standardize the variables to zero mean and unit variance, then reduce the matrix dimensions and filter out the noise through principal component analysis, retaining ∼75% of the total variance, i.e. eight principal components. The clustering model is therefore built in the EOF space with only eight variables taken into account. The number of clusters is varied from 2-9 and a classifiability index is applied to determine the appropriate number of clusters to be retained. This index rules out partitions into 2-3 clusters as not significant at the 95% level and we retain the partition into six clusters.

Compositing of other climatic parameters according to the four types of diurnal cycle and/or the six regions
The mean diurnal evolution of rainfall, and cloudiness, and mean daily values of PET, temperature and relative humidity, associated with each type of diurnal cycle of DNI are obtained by averaging values over the corresponding days (obscure, obscure AM K). The relationships between daily rainfall (from TRMM) and PET (from GDAS) as well as the one between daily direct and diffuse irradiance (either from DNI or PAR) have also been analysed for the 4 types of days. The conditional and marginal frequencies have been computed for deciles of daily rainfall, PET, direct and diffuse irradiance and PAR computed from all 2400 forest grid-points and 3285 days. For daily rainfall, we added a class for zero rainfall, which accounts for roughly 47% of all forest grid-points and days. The counts of observations are normalized by the grand total of forest grid-points and days. Then, we computed the conditional and marginal frequencies for the four daily types of irradiance and computed the relative anomalies versus the ones computed on all days. If the four types of days would not discriminate in terms of rainfall/PET and direct/diffuse irradiance deciles, anomalies would be close to zero and not significant.
The mean annual cycles of the parameters relative to climate (rainfall K) observed within each of the six irradiance-based regions are obtained by averaging values over the corresponding pixels (southwest Gabon, Cameroon K). For variables provided at a daily resolution, a 30-day low-pass filter is applied on the computed mean annual cycles to remove the high frequency variability (e.g. figures S6, S7, S8).
2.2.4. Identifying and mapping forest types according to EVI mean seasonal cycles plus elevation, soils, inventory data and vegetation maps To identify and map forest types with similar trait and photosynthetic activity, we extended to the whole of Central Africa an approach previously developed for a restricted area in the Sangha River Interval, north of the Republic of Congo (Gond et al 2013).
The EVI value attributed to each 16-day period corresponds the best normalized difference vegetation index and the lowest zenith angle. Using the 'good' pixels only (i.e. pixels which are flagged 'no clouds') in the 12 years for each 16-day period, a mean seasonal cycle was computed from the 12-year database. An unsupervised iterative self-organizing data analysis technique (ISODATA) classification was applied on the EVI mean seasonal cycles. The ISODATA classification is a k-means algorithm which allows selecting clusters by splitting and merging the initial pixels dataset and has been previously applied for forests in Madagascar (Mayaux et al 2000).
Nine classes were finally retained which were crossed with elevation, soils and inventory data, and existing vegetation maps to corroborate their spatial extend, interpret them in terms of deciduousness and density, and label them. In particular we detected significant and consistent differences in forest structure and composition among the nine remotely-sensed forest types (see supplementary information for a description of datasets and of the nine types of forests).

Results
3.1. Daily variations in direct irradiance in Central Africa DNI strongly varies within a day as pictured by the four main types of diurnal cycles extracted (see Methods). Obscure days which are the most frequent (29%, figure 1(a)) display maximum DNI as low as <150 W m −2 . These days relate to a heavy cloud cover during daytime, from low to very high opaque clouds (figure S2) and above average hourly rainfall from midnight to noon (∼4 mm h −1 , figure S3). In contrast, bright days display the largest values of DNI, 700 W m −2 at 12-13LT. They are associated with a high frequency of clear skies and are the driest ( figure 1(d)). Days classified as obscure in the morning ( figure 1(b), 'Obscure AM') have a peak value usually above 500 W m −2 at 14LT. These days are characterized by a high frequency of ultra-low and medium clouds, and below normal rainfall in the morning. Days classified as obscure in the afternoon ( figure 1(c), 'Obscure PM') are associated with a high frequency of high to very high opaque clouds, and high rainfall intensities in the afternoon (more than 7 mm h −1 by 15LT), i.e. the typical end of afternoon, rain-bearing deep convective clouds, but they are nonetheless the less frequent (19%).
These four types of days translate into different relationships of PET versus rainfall (i.e. water demand versus supply) and direct versus diffuse irradiance (i.e. quantity versus quality of light) (figure 2, see also figure S4 for direct versus diffuse PAR). While obscure and bright days display the expected opposite rainfall versus PET and direct versus diffuse irradiance relationships (figures 2(a), (b) and (e), (f) respectively), i.e. higher rainfall and diffuse irradiance, and lower PET and direct irradiance recorded during obscure days as compared to bright days, this does not hold true for obscure AM and obscure PM days. Obscure AM days are usually characterized by low rainfall and low to medium PET whereas obscure PM days frequently record high rainfall and high PET (figures 2(c), (d)). High levels of diffuse irradiance associated with medium levels of direct irradiance are also more frequent during obscure AM days than obscure PM days (figures 2(g), (h)).

Seasonal and spatial variations across Central Africa
We identify wide spatial variations across Central Africa in the frequency of these four types of days across the mean annual cycle (see Methods). The western part of Central Africa ('SW Gabon' and 'Cameroon', figures 3(a)-(c)) display outstanding behaviour when compared to the other regions (figure S5): the cumulated frequencies of obscure and obscure AM days exceed 81% on average. In 'SW Gabon', the June-September season shows up with over 70% of obscure days, following a remarkable steep increase in May. Therefore, the forests of 'SW Gabon' grow under the lowest light levels recorded across Central Africa as attested by a DNI annual average of ∼109 W m −2 (figure S6), and below 80 W m −2 in JJAS ( figure 4(a)). Further inland, the four types of days are more evenly distributed, especially in 'Centre DRC' characterized by seasonal cycles of weak amplitude (figure 4(d)), and obscure PM days are relatively much more frequent ( figure S5).
We also observe that forests of 'SW Gabon' exhibit the driest and longest dry season: less than 1 mm d −1 in June-August (figure 4(a)) whatever the rainfall estimates used (figure S6). Indeed, most of the days in June-September are rainless in 'SW Gabon', especially the obscure days ( figure S7). Hence, we show that western Central Africa, and specifically 'SW Gabon', stands apart as the area in Central Africa where the main dry season is the most light-deficient season, whereas further inland the main dry season is, as usually expected, the brightest season ( figure S6).
This peculiar seasonality of sunlight and rainfall in 'SW Gabon' is related to the presence of a uniform and opaque cover of low-level stratiform clouds in June-September (Dommo et al 2018) (figure S8, mean frequency of ∼45%) which are mostly non-precipitating, and strongly filter the incoming solar radiation.
By decreasing the direct irradiance, a very direct effect of this June-September persistent low-level cloud cover is to lower the water demand (PET). In fact, the long dry season in 'SW Gabon' (June-September) records the lowest PET at the regional scale (<2.3 mm d −1 , figures 4(c) and S9). It is also the coolest (Tmax below 29°C) and the moistest (RHmax ∼90%, figure 4(d)) (concomitantly, Tmin and RHmin are the highest in 'SW Gabon', not shown). Comparatively, in the other regions, the long dry season (June-September for 'S DRC' and 'Centre DRC', and December-February for 'W&E Margins' and 'N DRC') is the season when the water demand and maximum temperature reach their highest level, and the maximum relative humidity its lowest level (note however that the high levels of PET maintain through the 1st part of the subsequent rainy season, i.e. September and March respectively).
A second effect of that persistent low-level cloud cover in 'SW Gabon' is a prominent modification of the quality of light available for trees photosynthesis ( figure 4(b), in agreement with findings for the four types of diurnal cycles): the ratio between the direct and the diffuse irradiance in June-September varies between 0.4-0.8 (meaning that diffuse irradiance levels are 20% to 60% and up higher than those of direct irradiance). The relationship between PAR and PARdiff (figure S10) shows that forests in 'Cameroon' and 'SW Gabon' grow under much more days with low PAR and medium to high PARdiff than forests further inland (note also in figure 4(b), the DRI/DFI ratio below 1 during part of the March-May and September-November rainy seasons).

Links with forest trait
The six irradiance-based regions identified ( figure 3(a)) are crossed with a regional map of terra firme forests ( figure 5(a)) developed independently based on EVI seasonality and forest inventory data (see Methods). Nine different types of forests are identified across Central Africa which differ in terms of photosynthetic activity (i.e. showing mean annual cycles of EVI with different levels, amplitudes and phases, figure S11), structure (i.e. dense/degraded/secondary) and composition (i.e. evergreen/semi-deciduous/deciduous, table S3).
We find that the spatial distribution of these nine types of forests is in good agreement with the six irradiance-based regions ( figure 5(b)). The 'SW Gabon', with its dark long dry season with a low water demand, and where diffuse radiation dominates during a large part of the year, harbours most of the evergreen and dense forests of Central Africa: the 'dense evergreen forests', 'dense evergreen and semi-deciduous forests' Figure 2. Water demand versus rain, and light quantity versus quality during the four types of diurnal cycles of irradiance. Conditional frequencies (expressed in relative anomalies versus all days) for deciles of PET and rainfall (in tenth of mm per day, panels (a)-(d)) and of diffuse and direct irradiance (in W m −2 , panels (e)-(h)). The marginal frequencies are given for reference in the rightmost column for PET or direct irradiance and upper row for rain or diffuse irradiance. Blank squares indicate non-significant anomalies at the twosided 95% level (see Methods). and 'dense semi-deciduous forests' of western Central Africa (i.e. forest types 1, 3 and 4) are the most represented. But it is also remarkable that the 'SW Gabon' region closely matches the range of the Okoumé (Aucoumea Klaineana Pierre, figure 3(a)), a pioneer evergreen tree encountered as mono-dominant stands after shifting cultivation or savannahs, and one of the most important timber-producing native tree of western Central Africa (White et al 2000).

Discussion
In this study, we propose an original analysis of light conditions prevailing in Central African forests, and their implication on variables relevant for forests' functioning and traits. This analysis relies on the diurnal cycles of irradiance, the associated levels of rainfall, ETP and direct versus diffuse light, and the evolution of their frequency along the annual cycle. It leads to the first irradiance-based regionalization of Central African forests.
Our irradiance-based regionalization highlights western Central Africa, and more specifically the southwestern Gabon, as an area standing apart from the rest of Central Africa in terms of mean climate functioning: indeed, an important finding is that the 'SW Gabon' is much darker than the other regions, especially during its main dry season which is also the driest and longest at the regional scale. This contrasts with most previous climatic regionalizations based on the mean annual cycles of rainfall only (Dezfuli 2011, Badr et al 2016: they mainly picked-up zonal patterns driven by the gradual lengthening, on each side of the   equator, of the main dry season (and the corresponding shortening of the secondary dry season) as a result of the latitudinal shift of the ITCZ, but fail to mention the darkness of the long dry season in western Central Africa, thus its low water demand and high levels of diffuse light.
A second key finding is that the light-deficient climates of western Central Africa harbor most of the evergreen forests of Central Africa. Despite the relatively low mean annual rainfall (<2000 mm against 2500 mm and up in Amazon and southeast Asia) and long dry season (3 months below 30 mm), evergreen forests might have been favoured by the limited water demand during the long dry season due to the extensive low cloud cover. Rainfall above 2000 mm yr −1 has indeed been shown to sustain evergreen forests worldwide though most tropical forests in Central Africa (including the evergreen forests) are found below this climatic threshold (spatial variations of mean annual rainfall among the six regions do not exceed 170 mm, i.e. ∼10% of the mean annual amount, figure S6). We further hypothesize that in western Central Africa, the June-September rainfall amounts recorded in raingauges or estimated through remote-sensing are far from being representative of the actual amount of water available to forest ecosystems, since they do not take into account cloud water interception and water foliar interception (Oliveira et al 2014). At the annual scale, water inputs resulting from fog are estimated to represent up to 20% of the total precipitation in the Gabon highlands (Bruijnzeel et al 2011). This ratio might be even higher in June-September given the high frequency of low clouds likely to be regularly trapped in the canopy and the high relative humidity levels, and might help at balancing the water demand.
Another finding is the particularity of the light quality in western Central Africa, specifically in June-September, when the diffuse component represents an important part of the total irradiance. Although high rainfall/low water demand and low soil fertility are recognized as the main determinants of forest evergreenness (Givnish 2002, Ouédraogo et al 2013, the specificity of irradiance in western Central Africa raises question on the impact of light quality on forest trait (and also on photosynthetic activity levels although their estimation in June-September is uncertain because of the large cloud cover). Many studies suggest that diffuse irradiance is favourable to photosynthesis because it irradiates more of the subcanopy leaves and decreases light-saturation of the canopy top leaves (Mercado et al 2009, Doughty et al 2010, Rap et al 2015. Therefore, evergreenness in western Central Africa might be the most cost-effective way to cope with light-deficiency as the presence of mature leaves through a large part of the annual cycle might be the best adaptive strategy for gaining carbon (Schwartz 1993, Aerts 1995, Givnish 2002.
A last key finding is that the four types of diurnal cycles of direct irradiance extracted translate into different PET versus rainfall and direct versus diffuse light relationships. This is particularly so for obscure AM and obscure PM days although they have comparable daily levels of irradiance. Such relationships are not trivial given the uncertainties in the data. They also demonstrate that direct irradiance at the diurnal scale is a particularly comprehensive parameter which deserves much more attention in climatological studies especially those dedicated to plant functioning, and which we believe is easier to work with than cloudiness and rainfall which are intrinsically discrete.

Conclusion
Because mean annual rainfall in Central Africa is lower than in Amazonia and therefore recognized as the main controlling factor of forests functioning, little attention has been paid to solar radiation levels.
An irradiance-based regionalization of Central Africa reveals the uniqueness of western Central Africa and Gabon in particular, as an area with a strongly light-deficient main dry season because of an extensive low-level cloud cover. The low water demand and high quality of light observed during the main dry season might explain that western Central Africa sustains most of the evergreen forests of Central Africa when semi-deciduous and deciduous forests are mainly observed further east.
The questions are now how these light-deficient climates of western Central Africa will be impacted by global warming, and whether they will still be able to sustain most of the evergreen forests of Central Africa, their biodiversity and the ecosystem services they provide. This implies an appropriate simulation of the June-September low cloud cover and of the diurnal cycles of irradiance by climate models which is unfortunately one of their weak points (Nam et al 2012).

Authors contribution
NP has designed the study and performed climate analyses; GC, VG and LM have produced the forest traits map; VM has analysed the rainfall/PET and direct/diffuse irradiance relationships; JP and GS have helped with the retrieval and analysis of cloud data; PC has analysed PET data; AN has provided data for the Okoumé range; SB, CD and AF have helped with the interpretation of the results with regards to their field knowledge. All authors have contributed to the writing of the paper. The authors declare no conflicts of interest.