Vegetation growth in semi-arid regions is an important sink for fossil fuel emissions. Mean carbon dioxide (CO2) uptake by terrestrial ecosystems is dominated by highly productive lands, mainly tropical forests, whereas semi-arid regions are the main biome driving its inter-annual variability (Ahlström et al., 2015; Poulter et al., 2014). Semi-arid regions contribute to 60 % of the long-term trend in the global terrestrial C sink (Ahlström et al., 2015). It is thus important to understand long-term variability of vegetation growth in semi-arid areas and the response of vegetation to environmental conditions to better quantify and forecast effects of climate change.

The Sahel is a semi-arid transition zone between the dry Sahara desert in the north and the humid Sudanian savanna in the south. The region has experienced numerous severe droughts over the last decades, which resulted in region-wide famines in 1972–1973 and 1984–1985 and localized food shortages across the region in 1990, 2002, 2004, 2011 and 2012 (Abdi et al., 2014; United Nations, 2013). Vegetation production is thereby an important ecosystem service for livelihoods in the Sahel, but it is under threat. The region is experiencing strong population growth, increasing the demand on ecosystem services due to cropland expansion, increased pasture stocking rates and fuelwood extraction (Abdi et al., 2014).

At the same time as we have reports of declining vegetation production, we have contradicting reports of the greening of the Sahel based on earth observation (EO) data (Dardel et al., 2014; Fensholt et al., 2013). The greening of the Sahel has mainly been attributed to alleviated drought stress conditions due to increased precipitation since the mid-1990s (Hickler et al., 2005). Climate is thus another important factor regulating vegetation production. Semi-arid regions, such as the Sahel, are particularly vulnerable to climate fluctuations due to their dependency on moisture.

Estimation of gross primary production (GPP), i.e. uptake of atmospheric CO2 by vegetation, is still a major challenge for the remote sensing of ecosystem services. Gross primary production is a main driver of ecosystem services such as climate regulation, carbon (C) sequestration, C storage, food production and livestock grassland production. Within EO, spatial quantification of GPP generally involves light use efficiency (LUE), defined as the conversion efficiency of absorbed solar light into CO2 uptake (Monteith, 1972, 1977). It has been shown that LUE varies in space and time due to factors such as plant functional type, drought and temperature, nutrient levels, and physiological limitations of photosynthesis (Garbulsky et al., 2010; Paruelo et al., 2004; Kergoat et al., 2008). The LUE concept has been applied through various methods, either by using a biome-specific LUE constant (Ruimy et al., 1994) or by modifying a maximum LUE using meteorological variables (Running et al., 2004).

An example of a LUE-based model is the standard GPP product from the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor (MOD17A2). Within the model, absorbed photosynthetically active radiation (PAR) is estimated as a product of the fraction of PAR absorbed by green vegetation (FPAR from MOD15A2) multiplied with daily PAR from the meteorological data of the Global Modeling and Assimilation Office (GMAO). A set of maximum LUE parameters specified for each biome are extracted from a Biome Properties Look-Up Table (BPLUT). Then maximum LUE is modified depending on air temperature (Tair) and vapour pressure deficit (VPD; Running et al., 2004). Sjöström et al. (2013) evaluated the MOD17A2 product (collection 5.1) for Africa and showed that it underestimated GPP for semi-arid savannas in the Sahel. Explanations for this underestimation were that the assigned maximum LUE from BPLUT was set too low and that there were uncertainties in the FPAR product (MOD15A2). Recently, a new collection of MOD17A2 at a 500 m spatial resolution was released (MOD17A2H, collection 6) with an updated BPLUT, updated GMAO meteorological data, improved quality control and gap-filling of the FPAR data from MOD15A2 (Running and Zhao, 2015).

It has been shown that the LUE method does not perform well in arid conditions and at agricultural sites (Turner et al., 2005). Additionally, the linearity assumed by the LUE model is not usually found as the response of GPP to incoming light follows more of an asymptotic curve (Cannell and Thornley, 1998). Investigating other methods for remotely determining GPP is thus of great importance, especially for semi-arid environments. Therefore, instead of LUE, we focus on the light response function of GPP at the canopy scale, and spatial and temporal variation of its two main parameters: maximum GPP under light saturation (canopy-scale photosynthetic capacity, Fopt) and the initial slope of the light response function (canopy-scale quantum efficiency, α; Falge et al., 2001; Tagesson et al., 2015a). Photosynthetic capacity is a measure of the maximum rate at which the canopy can fix CO2 during photosynthesis (µmol CO2 m-2 s-1), whereas α is the amount of CO2 fixed per incoming PAR (µmol CO2 µmol PAR-1). To clarify the difference in LUE and α in this study, LUE (µmol CO2 µmol APAR-1) is the slope of a linear fit between CO2 uptake and absorbed PAR, whereas α (µmol CO2 µmol PAR-1) is the initial slope of an asymptotic curve against incoming PAR.

It has been proven that Fopt and α are closely related to chlorophyll abundance due to their coupling with the electron transport rate (Ide et al., 2010). Additionally, in semi-arid ecosystems, water availability is generally considered to be the main limiting factor affecting intra-annual dynamics of vegetation growth (Fensholt et al., 2013; Hickler et al., 2005; Tagesson et al., 2015b). Several remote sensing studies have established relationships between remotely sensed vegetation indices and ecosystem properties such as chlorophyll abundance and equivalent water thickness (Yoder and Pettigrew-Crosby, 1995; Fensholt and Sandholt, 2003). In this study, we will analyse whether EO vegetation indices can be used to upscale Fopt and α and investigate whether this could offer a promising way to map GPP in semi-arid areas. This potential will be analysed by the use of detailed ground observations from six eddy covariance (EC) flux tower sites across the Sahel.

The three aims of this study are

to investigate whether the recently released MOD17A2H GPP (collection 6) product is better at capturing GPP for the Sahel than collection 5.1. We hypothesize that the MOD17A2H GPP (collection 6) product will estimate GPP well for the six Sahelian EC sites because of major changes made in comparison to collection 5.1 (Running and Zhao, 2015);

Our hypothesis that vegetation indices closely related to SIWSI would be most strongly coupled with intra-annual dynamics in Fopt and α was not rejected for Fopt since this was the case for all sites except for Wankama Millet (Table 2). However, our hypothesis was rejected for α, since it was more closely related to vegetation indices of chlorophyll abundance (RDVI and EVI). In the Sahel, soil moisture conditions in the early rainy season are important for vegetation growth, and during this phase vegetation is especially vulnerable to drought conditions (Rockström and de Rouw, 1997; Tagesson et al., 2016a; Mbow et al., 2013). Photosynthetic capacity (Fopt) peaked earlier than α did in the rainy season (Fig. 3), thereby explaining the close relationship of Fopt to SIWSI. Leaf area index increased over the growing season and leaf area index is closely coupled with vegetation indices related to chlorophyll abundance (Tagesson et al., 2009). The increase in leaf area index increased canopy level quantum efficiency (α), thereby explaining the closer relationship of α with RDVI.

Our hypothesis that vegetation indices closely related to chlorophyll abundance would be most strongly coupled with spatial and inter-annual dynamics in Fopt and α was not rejected for either Fopt or α; NDVI, EVI and RDVI all correlated with spatial and inter-annual dynamics in Fopt and α (Table 5). However, it was surprising that NDVIpeak had the strongest correlation with spatial and inter-annual variability in Fopt (Table 5). Both EVI and RDVI should be less sensitive to saturation effects than NDVI (Huete et al., 2002; Roujean and Breon, 1995), and based on this it can be assumed that peak values of these indices should have stronger relationships with peak values of Fopt and α. However, vegetation indices with a high sensitivity to changes in green biomass at high biomass loads become less sensitive to green biomass changes at low biomass loads (Huete et al., 2002). The peak leaf area index for ecosystems across the Sahel is generally ∼ 2 m2 m-2 or less, whereas the saturation issue of NDVI generally starts at a leaf area index of about 2–5 m2 m-2 (Haboudane et al., 2004).

The Fopt_peak estimates from Agoufou, Demokeya and the Wankama sites were similar, whereas Dahra and Kelma values were high in relation to previously reported canopy-scale Fopt_peak from the Sahel (∼ -8 to -23 µmol m-2 s-1; Hanan et al., 1998; Merbold et al., 2009; Moncrieff et al., 1997; Boulain et al., 2009; Levy et al., 1997; Monteny et al., 1997). These previous studies reported much lower Fopt at canopy scale than at leaf scale (e.g. Levy et al., 1997: 10 vs. 44 µmol m-2 s-1; Boulain et al., 2009: 8 vs. 50 µmol m-2 s-1). The leaf area index at Dahra and Kelma peaked at 2.1 and 2.7, respectively (Timouk et al., 2009; Tagesson et al., 2015a), and it was substantially higher than at the above-mentioned sites. A possible explanation for high Fopt estimates at Dahra and Kelma could therefore be the higher leaf area index. Tagesson et al. (2016b) performed a quality check of the EC data due to the high net CO2 exchange measured at the Dahra field site and explained the high values by a combination of moderately dense herbaceous C4 ground vegetation, high soil nutrient availability and a grazing pressure resulting in compensatory growth and fertilization effects. Another possible explanation could be that the West African monsoon brings a humid layer of surface air from the Atlantic, possibly increasing vegetation production for the most western part of the Sahel (Tagesson et al., 2016a).

Our model substantially overestimated GPP for Wankama Millet (Fig. 7f). Being a crop field, this site differed from the other sites in its species composition and ecosystem structure, as well as land and vegetation management. Crop fields in southwestern Niger are generally characterized by rather low production, resulting from decreased fertility and soil loss caused by intensive land use (Cappelaere et al., 2009). These specifics of the Wankama Millet site may cause the model, parameterized with observations from the other study sites without this strong anthropogenic influence, to overestimate GPP at this site. Similar results were found by Boulain et al. (2009) when applying an upscaling model using leaf area index for Wankama Millet and Wankama Fallow. It worked well for Wankama fallow, whereas it was less conclusive for Wankama Millet. The main explanation for this difference was low leaf area index in millet fields because of a low density of millet stands due to agricultural practice. There is extensive savanna clearing for food production in the Sahel (Leblanc et al., 2008; Boulain et al., 2009; Cappelaere et al., 2009). To further understand impacts of this land cover change on vegetation production and land–atmosphere exchange processes, there is an urgent need for more study sites covering cropped areas in this region.

In Demokeya, GPP was slightly underestimated for 2008 (Fig. 7c) because modelled Fopt was much lower than the actual measured value in 2008 (the thick black line in Fig. 4). An improvement of the model could be to incorporate some parameters that constrain or enhance Fopt depending on environmental stress. Indeed, the regression tree analysis indicated that incorporating hydrometeorological variables increased the ability to predict both Fopt and α. Conversely, for spatial upscaling purposes, it has been shown that including modelled hydrometeorological constraints on LUE decreases the ability to predict vegetation production due to the incorporated uncertainty in these modelled variables (Fensholt et al., 2006; Ma et al., 2014). For spatial upscaling to regional scales, it is therefore better to simply use relationships with EO data. This is particularly the case for the Sahel, one of the largest dryland areas in the world, which includes only a few sites of hydrometeorological observations.

The pattern seen in the spatially explicit GPP budgets (Fig. 5c) may be influenced by a range of biophysical and anthropogenic factors. The clear north–south gradient is expected given the strong north–south rainfall gradient in the Sahel. The West African monsoon mentioned above could also be an explanation of high GPP values in the western part of the Sahel, where values were relatively high in relation to GPP at similar latitudes in the central and eastern Sahel (Fig. 5c). The areas with the highest GPP are sparsely populated woodlands or shrubby savanna with a relatively dense tree cover (Brandt et al., 2016). However, the maps produced here should be used with caution as they are based on upscaling of data collected at only six EC sites available in the region, especially given the issues related to the cropped fields discussed earlier. Still, the average GPP budget for the entire Sahel 2001–2014 was close to an average annual GPP budget estimated at these six sites (692 ± 89 g C m-2 yr-1; Tagesson et al., 2016a). The range of GPP budgets in Fig. 5c is also similar to previous annual GPP budgets reported from other savannas across the world (Veenendaal et al., 2004; Chen et al., 2003, 2006; Kanniah et al., 2010).

Although MOD17A2 GPP has previously been shown to capture GPP in several ecosystems types well (Turner et al., 2006, 2005; Heinsch et al., 2006; Sims et al., 2006; Kanniah et al., 2009), it has been shown to underestimate it in others (Coops et al., 2007; Gebremichael and Barros, 2006; Sjöström et al., 2013). Gross primary production of Sahelian drylands have not been captured well by MOD17A2 (Sjöström et al., 2013; Fensholt et al., 2006), and as we have shown, this underestimation persists in the latest MOD17A2H GPP (collection 6) product (Fig. 2). The main reason for this pronounced underestimation is that maximum LUE is set to 0.84 g C MJ-1 (open shrubland, Demokeya) and 0.86 g C MJ-1 (grassland; Agoufou, Dahra, Kelma; Wankama Millet and Wankama Fallow) in the BPLUT, i.e. much lower than maximum LUE measured at the Sahelian measurement sites of this study (average: 2.47 g C MJ-1; range: 1.58–3.50 g C MJ-1; Sjöström et al., 2013; Tagesson et al., 2015a), a global estimate of ∼ 1.5 g C MJ-1 (Garbulsky et al., 2010) and a savanna site in Australia (1.26 g C MJ-1; Kanniah et al., 2009).

Several dynamic global vegetation models have been used for decades to quantify GPP at different spatial and temporal scales (Dickinson, 1983; Sellers et al., 1997). These models are generally based on the photosynthesis model of Farquhar et al. (1980), a model particularly sensitive to uncertainty in photosynthetic capacity (Zhang et al., 2014). This and several previous studies have shown that both photosynthetic capacity and efficiency (both α and LUE) can vary considerably between seasons as well as spatially vary, and both within and between vegetation types (Eamus et al., 2013; Garbulsky et al., 2010; Ma et al., 2014; Tagesson et al., 2015a). This variability is difficult to estimate using broad values based on land cover classes, yet most models apply a constant value, which can cause substantial inaccuracies in the estimates of seasonal and spatial variability in GPP. This is particularly a problem in savannas that consist of several plant functional types (C3 and C4 species, and a large variability in tree and/or herbaceous vegetation fractions; Scholes and Archer, 1997). This study indicates the applicability of EO as a tool for parameterizing spatially explicit estimates of plant physiological variables, which could improve our ability to simulate GPP. Spatially explicit estimates of GPP at a high temporal and spatial resolution are essential for environmental change studies in the Sahel and can contribute to increased knowledge regarding changes in GPP, its relationship to climatic change and anthropogenic forcing, and simulations of ecosystem processes and hydro-biochemical cycles.