The relationships between NIRv, SIF, and GPP were studied for the major biomes over Africa. The dominant biomes of the region are broadleaf evergreen forest (BLEF), C3 grasses (C3), shrubs (Sh), and C4 grasses (C4) (Fig. e). Northern and southern parts of Africa operated in anti-phase in their summer insolation, precipitation, and other environmental stresses. Therefore, we subdivided the shrubs and C4 grasses into a Northern Hemisphere (NH) part and Southern Hemisphere (SH) part. However, we did not split up the BLEF, which is at the tropical rain belt and shows weak symmetry between the north and south of the Equator, or the C3, due to its smaller coverage over the northern part of Africa. The vegetation type distribution is based on the terrestrial biosphere model SiBCASA .

The seasonal movement of the Inter-Tropical Convergence Zone (ITCZ) drives the climate of Africa in response to changes in the location of maximum solar heating in the region. The ITCZ seasonally migrates north and south of the Equator as the latitude of maximum solar insolation varies, causing equatorial Africa to be characterized by double-rainfall-maximum rainy seasons . However, in East Africa, local topography leads to spatially variable temperatures and a complex distribution of rainfall . The seasonal dynamics of vegetation are strongly controlled by these climatic conditions through the key processes of photosynthesis, respiration, and transpiration. Primarily, the length of the dry season has often been emphasized as a major factor controlling the vegetation structure and patterns in the tropics . November–April are the wettest months, whereas June–September are the driest months for the regions with C3 and Sh-SH biomes (Fig. d). On the other hand, July–September are the wettest months, while December to March are the driest months Sh-NH and C4-NH regions. For BLEF regions the precipitation is higher than 100 mm per month throughout the year.

In this study we use Level-3 GOME-2 Sun-Induced Fluorescence of Terrestrial Ecosystems Retrieval SIF (v2.0) with 737 nm monthly data at a spatial resolution of 0.5∘×0.5∘ covering the period from 2007 to 2016. For a fair comparability of the SIF with vegetation indices we normalized SIF by the cosine of the solar zenith angle. We used the Royal Netherlands Meteorological Institute (KNMI)–Wageningen University & Research (WUR) SIF retrieval, which is particularly suited for tropical conditions . The retrieval code uses a much larger dataset to construct the reference atmospheric spectra used to distinguish the small SIF signals from the complex structure of transmittance and reflectance from other atmospheric constituents such as water vapor .

The NIRv represents the fraction of reflected near-infrared reflectance (NIR) of light that originates from vegetation. NIRv was first described as a proxy for photosynthesis by . Recent studies have suggested using NIRv and computed NIRv-based SIF for a more robust GPP estimation under wide land cover with a varied canopy structure and soil brightness . In this study we used NIRv at a spatial resolution of 0.5∘×0.5∘ and a monthly temporal resolution for the years 2007–2016. We also used NIRv at a higher resolution (0.05∘×0.05∘, daily) in comparison with flux tower GPP. NIRv data used here were calculated using surface reflectance data from MODIS collection MCD43C4 v006 from 2007 to 2016 as 1NIRv=ρnir×ρnir-ρredρnir+ρred-0.08, where ρnir and ρred are reflectances acquired in the near-infrared (841–876 nm) and red (620–670 nm) portions of the electromagnetic spectrum, respectively . A constant 0.08 is subtracted to reduce the effects of the bare soil .

Monthly global ecosystem estimation of terrestrial GPP was made available by the Max Planck Institute's Biogeochemical Integration Group . This GPP estimation is constructed using a machine-learning method to upscale information from flux towers up to a 0.5∘×0.5∘ grid, aided by gridded meteorological and remote-sensing co-variables, and hereafter we refer to this product as MPI-BGC GPP. This product is available at a monthly temporal resolution (downloaded from https://www.bgc-jena.mpg.de/geodb/projects/Data.php, last access: 5 January 2020) covering the period from 2007 to 2011.

The enhanced vegetation index (EVI) and the normalized difference vegetation index (NDVI) are the two most widely used vegetation indices for monitoring vegetation conditions and have significant relationships with GPP . The monthly EVI and NDVI from the MODIS collection of MOD13C2 at a spatial resolution of 0.05∘×0.05∘ in the years 2007–2015 were used in the study. Compared with the NDVI, the EVI is less sensitive to soil background variations and remains sensitive over dense vegetation . For that reason we focus on comparison with the EVI rather than with the NDVI. Moreover, MODIS does not provide NIRv in the MOD13C2 dataset, so we calculated it using the BRDF-corrected surface reflectances from MCD43C4, following the steps outlined in .

Monthly precipitation data from the Global Precipitation Climatology Centre (GPCC) at a 0.5∘×0.5∘ spatial resolution were employed to show the dry and wet months of the region for each major biome. The GPCC Full Data Monthly Product Version 2018 covers the period from 1891 to 2016; this new extended product version using the new GPCC climatology as analysis background was generated in May 2018 and can be accessed from https://www.dwd.de/ (last access: 15 January 2020). In addition we used monthly temperature data from European Centre for Medium-Range Weather Forecasts (ECMWF) atmospheric reanalysis ERA-Interim (with a 0.5∘×0.5∘ grid). Furthermore, we used monthly soil moisture (SM) and incoming downward shortwave radiation (SWR) from the Global Land Data Assimilation System Version 2.1 (GLDAS2.1) of the National Aeronautics and Space Administration (NASA) Goddard Space Flight Center (GSFC) at a spatial resolution of 0.25∘×0.25∘ covering the period from years 2007–2016 .

Standardized eddy-covariance flux data are available under the fair-use data policy of the FLUXNET2015 dataset. The data processing of the FLUXNET2015 dataset ensures inter-comparison and quality assurance and control across sites . A collection of eddy-covariance flux data from six regions in Africa were used in this study to assess the correlation between SIF, NIRv, and GPP at an ecosystem level, specifically, available monthly GPP products from the daytime partitioning of fluxes (GPP_DT_VUT_REF) from Tier-2 FLUXNET2015 synthesis https://fluxnet.org/ (last access: 15 January 2020). These variables were screened using quality flags so that only samples that are either measured (flag = 0) or good quality (flag = 1) were retained. An overview of the selected towers is given in Table . In addition to these six African flux towers we used data from the Brazilian BR-Sa1 flux tower for a better representation of GPP in broadleaf evergreen forests.

To compare the gridded datasets (SIF, NIRv, MPI-GPP) with the GPP measurements from flux towers, we extracted data from a 4∘×4∘ window surrounding each flux tower. The flux towers have a footprint of about 1 km2, and it is hard to compare them to areas that are 200 km2, centering around the towers, which include many vegetation types. However, we use the vegetation mask to exclude grid cells with different vegetation from the tower's vegetation; this will account for the land heterogeneity of the regions. The variation in climate condition was addressed by splitting up the shrubs and C4 grasses over the Northern and Southern hemispheres. We used good quality data as recommended by each data source, and we further processed these datasets to match the spatiotemporal grids. For the reason that only a small fraction of emitted SIF signal can be sensed from space, which also depends on the direction of observation, the empirical relationship between SIF and GPP is complicated . Even if the mechanistic link between remotely sensed vegetation reflectance and GPP is complex , and showed that a simple linear relationship between SIF- and tower-based GPP is a reasonably convenient framework for presenting and evaluating arguments and counterarguments for the SIF–GPP relationship. Here, we predict GPP from this remotely sensed vegetation reflectance using a linear regression between GPP and these signals as 2y=ax+b, where y is the GPP obtained from SIF or NIRv signals, x is the SIF or NIRv signal, and a and b are the slope and y intercept of the fitting line, respectively. The conversion of SIF and NIRv to GPP is achieved by applying this fitting to all monthly SIF and NIRv values for each vegetation type separately.

The linear relationship between SIF and GPP in Eq. () may be rationalized with the formulation based on the concept of light use efficiency (LUE) in a simple parametric LUE-model GPP as follows: 3GPP=APAR×LUEp, where APAR is the absorbed photosynthetically active radiation expressed in radiance units and LUEp is the light use efficiency of photosynthesis, which represents the efficiency of energy conversion for gross CO2 assimilation. Similarly, SIF can be expressed as 4SIF=APAR×LUEf×fesc, where LUEf is the effective light use efficiency of SIF and fesc is the fraction of SIF photons escaping the canopy . These equations can be combined making the dependence on light implicit: 5GPP≈SIF×LUEpLUEf×fesc.

SIF has negative values due to noise in its retrieval, and a zero value of SIF may not result in a zero value; therefore we do not force the regression to pass through the origin, and as result, there will be intercepts “b” as in Eq. (). Further, the linear relationship between NIRv and SIF can be rationalized by the fact that both SIF and NIRv are jointly dependent on the flux of the fractional interception of vegetation, incoming solar radiation, and the fraction of photons that escape from the canopy; these last two are strongly related measurable fluxes .

Spatial patterns in climatological NIRv, SIF, and MPI-BGC GPP are very similar across large scales, with maximum annual mean productivity in tropical broadleaf forests. Figure a–c show that productivity changes strongly at the borders of the plant-functional types (PFTs), which is interesting because only the MPI-BGC product was actually informed by a PFT map in its machine learning, while the satellite observations provide independent spatial views on productivity. Both products suggest additional variations within PFTs not represented in the MPI-BGC map, as would be expected based on the higher volume of observed data in the remote-sensing products.

NIRv, SIF, the EVI, and MPI-BGC GPP generally capture seasonal patterns of tower GPP well, except at the Ghana GH-Ank flux tower where from all products only SIF yields the expected positive correlation with eddy-covariance-derived GPP observations. Figure  shows the observation-derived and simulated seasonal cycles of GPP and the generally high (R>0.8) seasonal correlations. SIF shows more rapid changes in signal during the transitions from wet-to-dry periods than other proxies. The May–June–July period at the savanna site CG-Tch is an example and indicates that SIF responds more rapidly to the decline in photosynthesis in wilting grasses, which are still green and reflective enough to affect the NDVI, EVI, and NIRv. Moreover, time series analysis for the years covering 2007–2016 around the CG-Tch tower resulted in correlations of 0.77, 0.89, and 0.88 between SIF, NIRv, and the EVI and soil moisture (SM) and 0.72, 0.64, and 0.64 between SIF, NIRv, and the EVI and precipitation, respectively (Fig. S1). The low correlation between SIF and SM is due to SIF responding before the soil gets too dry, whereas the correlations of the EVI and NIRv with SM show the same pattern which suggests that they respond when the vegetation loses its green color. An immediate response of SIF to water stress of a similar type has also been observed by others .

SIF and NIRv have a higher monthly correlation with the eddy-covariance (EC) GPP at most of the towers than the EVI and MPI-BGC GPP (Fig. S2). also found this and suggest it is because chlorophyll content seen through the NDVI and EVI adapts slowly to stress and it can take weeks for leaves to lose their green color . The EVI and NDVI not only had generally much weaker correlations but also show saturation when GPP is high (Fig. S2). SIF and NIRv had the strongest correlations (R>0.90) with the EC GPP in C4 grass vegetation sites (SD-Dem and SN-Dhr), while weak or no relationship was found for broadleaf evergreen vegetation (GH-Ank). This lack of relation was also found previously by over rainforest regions and is further discussed in Sect. . Even over evergreen forests where GPP is high throughout the season (see Fig. c), NIRv and SIF track seasonality of GPP well.

For a more detailed look, we also compare EC GPP to daily NIRv signals, derived from high-resolution (0.05∘×0.05∘) MODIS radiances instead of coarse averaged MODIS NIRv (0.5∘×0.5∘). At this daily timescale, we again find a very strong correlation over northern Africa, while this correlation decreases for the equatorial sites (Fig. ). And again, we see a weak correlation at GH-Ank, Ghana (Fig.  Gh-Ank). The high-resolution NIRv mostly improves the comparison for sites with more heterogeneous vegetation cover (GH-Ank and ZM-Mon, ZA-Kru), whereas there was no significant improvement for less heterogeneous sites (SN-Dhr and SD-Dem). In contrast to the low climatology correlation for tropical evergreen broadleaf forest (GH-Ank, R=-0.44), the correlation in the interannual variation in NIRv and GPP is higher (R=0.21) (see Table S1). These results are illustrative of tropical ecosystems, where GPP variations are irregular and strongly coupled to photosynthetic capacity changes of vegetation .

We next extend this view from the level of individual towers to the scale of six major biomes in Africa (Fig. e) by spatially averaging our productivity products. Also then, wet months show higher SIF and NIRv values than dry months for all biomes of the region. Both SIF and NIRv show a strong linear relation with those of the MPI-BGC GPP estimates. Signals from both SIF and NIRv were correlated well with MPI-BGC GPP over these biomes with a correlation of R2>0.85 for NIRv and R2>0.77 for SIF. However, the correlation was moderate for broadleaf evergreen forests with a correlation of 0.38 for NIRv and 0.16 for SIF (see Table S2). This seasonality shown in Fig.  agrees closely with that seen for similar biomes in the Amazon and confirms the known strong water control over GPP in tropical vegetation . Correlations with shortwave radiation are thus strongly negative, especially over short vegetation like shrubs and grasses. Over evergreen forests, SIF and NIRv show a double-peaked seasonality and a decrease in productivity during the dry seasons despite high SWR and a high leaf area index (see Fig. , Broadleaf evergreen), suggesting an influence of photosynthetic capacity on GPP that has been noted before to be not yet represented by most biosphere models and light use efficiency models . Clouds can strongly reduce direct solar radiation during the wet season which increases the ratio of diffuse versus direct solar radiation, possibly increasing productivity .

Mostly, the EVI and MPI GPP closely agree on the satellite-observed seasonality at these larger scales, but the EVI appears late in simulating the wet–dry season (September–December) decline in signal for C3 grasses and shrubs (see SIF-vs.-EVI hysteresis plots in Fig. S4). Similar differences in EVI- and SIF-derived seasonal cycles of photosynthesis were described in , for short vegetation at high latitudes. Our response is opposite in the sense that we see photosynthesis decline already before the seasonal brown-down of the savanna. To further investigate this, we therefore turn to the main drivers, water and light, of the African seasonal cycle in GPP.

The biome-integrated productivity in Africa is seasonally strongly controlled by soil moisture, with a weak influence of light availability superimposed. In Fig.  this is recognized by the positive correlation between SIF/NIRv (which independently display the exact same patterns) and soil moisture in both hemispheres. Peak productivity coincides with peak soil moisture that occurs in September in the NH and in March in the SH. Interestingly, the lead-up to peak productivity occurs more slowly than its subsequent decline even at the same soil moisture levels, evident when comparing the pairs of months (1) August and October and (2) July and November for the NH, or pairs (1) January and April and (2) December and May for the SH. Translating these points to the SWR diagrams in panels (a) and (c), a notable difference between the hemispheres appears: in the SH peak productivity occurs while light availability continues to decline, creating the elliptical shape of the hysteresis diagram. In the NH however, peak productivity happens at minimum light availability becoming larger at the same soil moisture levels past the peak productivity.

A SIF- or NIRv-based GPP estimate across each biome compares independently quite well to MPI-BGC-estimated GPP. Figure  shows the GPP estimated by applying the SIF/NIRV-vs.-GPP relation derived at EC sites to biome-wide satellite observations.

We chose the relation between the remotely sensed reflectances and the EC GPP from Senegal SN-Dhr, Congo CG-Tch, and Sudan SD-Dem and ZA-Kru towers to estimate GPP of the northern shrub, the southern shrub, and C4 grass and C3 grass, respectively, to represent different biomes over northern and southern Africa (see Fig. S2 for the fitting equations applied). Due to the weak relationship between SIF/NIRv with Ghana GH-Ank GPP, a flux tower in tropical rainforest of Africa with broadleaf evergreen vegetation, we used a linear relation between SIF/NIRv and EC GPP data from the Brazilian BR-Sa1 flux tower which is also in a tropical rainforest region with a broadleaf evergreen vegetation and shows a better relation with these signals than the GH-Ank site (see Fig. S3). The good correspondence to MPI-BGC GPP partly results from the EC sites being part of the training algorithm for that product , but we note that the spatiotemporal drivers in our product (SIF/NIRv) are very different from those in MPI-BGC (PAR, T, PFT, precipitation, etc.). Mostly, the SIF- or NIRv-based GPP estimate shows that a reasonable first estimate of spatiotemporal GPP patterns can be based on SIF and NIRv without the need for the more complex and data-intensive machine-learning approach. At least, it captures the large differences between the major biomes of Africa, allowing further study of their seasonal dynamics, drought response, and contribution to tropical GPP.