Satellite Indices for Primary Production Estimates Evaluation of Satellite Based Indices for Primary Production Estimates in a Sparse Savanna in the Sudan Bgd Satellite Indices for Primary Production Estimates

One of the more frequently applied methods for integrating controls on primary production through satellite data is the Light Use Efficiency (LUE) approach. Satellite indices such as the Enhanced Vegetation Index (EVI) and the Shortwave Infrared Water Stress Index (SIWSI) have previously shown promise as predictors of primary production in 5 several different environments. In this study, we evaluate EVI and SIWSI derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) satellite sensor against in-situ measurements from central Sudan in order to asses their applicability in LUE-based primary production modelling within a water limited environment. Results show a strong correlation between EVI against gross primary production (GPP), demonstrat-10 ing the significance of EVI for deriving information on primary production with relatively high accuracy at similar areas. Evaluation of SIWSI however, reveal that the fraction of vegetation apparently is to low for the index to provide accurate information on canopy water content, indicating that the use of SIWSI as a predictor of water stress in satellite data-driven primary production modelling in similar semi-arid ecosystems is limited.


Introduction
With the emergence of international environmental treaties such as United Nations Framework convention on Climate Change (UNFCCC) and its Kyoto Protocol, there is an urgent request to quantify the global carbon budget and its temporal and spatial variations (IPCC, 2007). One of the least well-covered regions by studies on carbon made with the use of sophisticated process-based ecosystem models driven mainly by climate data (Sitch et al., 2003). However, during the last decade, rapid developments in satellite sensor technology have allowed remote sensing based primary production models to emerge as an attractive approach. Considering the spatial and temporal variations of the processes related to plant growth, repetitive and accurate satellite based 15 measurements may contribute significantly to our knowledge on vegetation dynamics and responses to changing environmental conditions. The improved spatial and spectral resolution of satellite sensors such as the Moderate Resolution Imaging Spectroradiometer (MODIS), in comparison to the extensively used Advanced Very High Resolution Radiometer (AVHRR), has further allowed a reassessment of the potential 20 for modelling primary production solely by the use of satellite data.
One of the more widely applied concepts within the domain of satellite based primary production models is the Light Use Efficiency (LUE) approach, first described by Monteith (1972Monteith ( , 1977. They are generally formulated as: 25 where GPP is gross primary production, the carbon assimilated by plants, ε is the conversion efficiency, PAR is incoming photosynthetically active radiation between the wavelengths 400-700 nm and fPAR represents the fraction of PAR absorbed by the canopy. FPAR is generally estimated through the use of spectral vegetation indices such as the Normalized Difference Vegetation Index (NDVI) (Daughtry et al., 1983;Asrar et al., 1984;Sellers et al., 1994) and the Enhanced Vegetation Index (Xiao et al., 2004) whereas the PAR received at the Earths surface can be estimated by interpolating point measurements from light sensors or be derived from the use of Earth 5 observation data (Eck and Dye, 1991;Seaquist and Olsson, 1999;Van Laake and Sanchez-Azofeifa, 2005;Liang et al., 2006;Olofsson et al., 2007b). The conversion efficiency factor ε was originally regarded as a fixed empirical constant when first introduced (Monteith, 1972) but during the past decade it has become more common to estimate or scale ε using models or satellite retrievals. Numerous stress factors control ε and estimating this critical parameter can be difficult due to high spatial and temporal variability inherent to site specific and meteorological conditions (Hilker et al., 2007).
In semi-arid environments such as the Sahel, water is generally the most limiting factor for growth and numerous satellite sensor-based primary production studies have focused on the derivation of information related to plant water stress in order to scale 15 ε with the use of modelling (Nemani and Running, 1989;Field et al., 1995;Prince and Goward, 1995;Seaquist et al., 2003). But progress has recently been made using satellite data to detect canopy water stress (Ceccato and Flasse, 2002;Ceccato et al., 2002;Fensholt and Sandholt, 2003;Xiao et al., 2004). However, efforts are still required to strengthen our knowledge regarding indices related to water stress as well 20 as their applicability for estimating primary production using satellite remote sensing.
In this paper, we aim to test satellite-based variables that can be used to upscale estimates of carbon in semi-arid Sahel by comparing these against site-specific measurements of CO 2 fluxes from central Sudan. We further investigate the applicability of these variables by including them in a simple parametric LUE-model, a model that Introduction

Eddy covariance data
The carbon budget is described by three components: i) gross primary productivity (GPP), the carbon captured through photosynthesis; ii) net ecosystem exchange 5 (NEE), the net exchange of carbon between the ecosystem and atmosphere and; iii) ecosystem respiration (R eco ), which is the sum of plant and heterotrophic respiration. NEE, GPP and R eco were derived from half-hourly values of CO 2 from July to December for the 2007 season (data prior to July were not available).
In order to obtain seasonal estimates of CO 2 exchange eddy covariance data was gap filled according to the method used by Reichstein et al. (2005) (amount of gaps was 39% for the period). This method considers both the covariance between fluxes and meteorological drivers and temporal structure. Previous studies from the Sahel have shown that soil respiration not only depends on temperature but also on soil moisture content (e.g. Friborg et al., 1997). R eco was 15 thus estimated by using the exponential regression model of Lloyd and Taylor (1994) in combination with a soil water content factor (F w ) derived from volumetric soil moisture (Wang and Leuning, 1998): where θ is the actual soil water content in the upper soil layer (5 cm), θ min and θ max is 20 the minimum and maximum soil water content. Using typical values for sandy soils, F w is thus scaled between wilting point and field capacity (i.e. approximately 5 and 15%) and then integrated with the Lloyd and Taylor (1994)  Daytime GPP was estimated as GPP=R e -NEE whereas nighttime GPP=0 based on a global radiation threshold of 20 Wm −2 .

Satellite data
We used satellite data from the MODIS/EOS Terra product MOD09A1, which provides an estimate of surface spectral reflectance in seven bands as it would have been mea-5 sured at the ground (Vermote et al., 2002). The MOD09A1 surface reflectance product includes correction for the effects of aerosols, thin cirrus clouds and atmospheric gases and serves as an input for several higher order land products such as FPAR/LAI and vegetation indices. The spatial resolution of the MOD09A1 data set is 500 m and data is composed of the best observations during an 8-day period with regards to overall 10 pixel quality and observational coverage (Justice et al., 2002). The Enhanced Vegetation Index (EVI) was developed to enhance the vegetation signal by reducing influences from the atmosphere and canopy background (Huete et al., 1997;Huete et al., 2002). As with the NDVI, EVI is estimated from surface reflectance in the red (ρ red ) and Near Infrared (NIR, ρ nir ) bands but it also uses reflectance in the 15 blue band (ρ blue ) to correct for effects of aerosols. Using surface reflectance data from MODIS, EVI is calculated as: Several studies have previously revealed a good general relationship between EVI and GPP (Xiao et al., 2004;Rahman et al., 2005;Sims et al., 2006b;Olofsson et al., 20 2007a). Building on this, Sims et al. (2008) developed a temperature and greenness model solely based on the MODIS Land Surface Temperature (LST) and EVI products. Modeled GPP estimates were in good agreement with measured values, highlighting EVI's potential for use as a predictor of GPP, and further demonstrating that GPP can be estimated with relatively high accuracy using only remote sensing data. 5,2008 Satellite indices for primary production estimates In the Sahel region, water is generally assumed to be the limiting factor for photosynthesis and several attempts have previously been made at developing satellite based stress factors that account for canopy or soil water deficits. The Short wave Infrared Water Stress Index (SIWSI) (Fensholt and Sandholt, 2003) is such an attempt. The Short-wave Infrared (SWIR, ρ swir ) and NIR bands available on the MODIS sensor 5 present opportunities to quantify equivalent water thickness (EWT), a term denoting spectral absorption resulting from the leaf water content. SIWSI is based on NIR and SWIR surface reflectance information, and has previously been shown to correlate well with soil moisture (Fensholt and Sandholt, 2003). SIWSI is estimated as:

BGD
10 SIWSI is thus a normalized index with values theoretically ranging between -1 and 1.
As the water content increases, the reflectance in the SWIR diminishes. Thus, a high SIWSI value, would indicate sufficient amounts of water whereas a low value would indicate water stress (Fensholt and Sandholt, 2003). We acquired the MOD09A1 8-day product from Earth Observing Systems Data 15 Gateway (EDG, http://edcimswww.cr.usgs.gov/pub/imswelcome/) for the year 2007. Data for surface reflectance were extracted from the site pixel, centred on the flux tower and climate station, and its 8 surrounding pixels. The extracted data were then used to calculate EVI and SIWSI according to Eqs. (4) and (5). In order to minimize the effects of sensor disturbances estimated EVI was seasonally adjusted by an adaptive 20 Savitzky-Golay filtering method using the TIMESAT program package Eklundh, 2002, 2004). TIMESAT fits a function to the upper envelope of the satellite time series data, effectively filtering out negatively biased noise due to, for instance, atmospheric effects (Eklundh and Olsson, 2003;Olofsson et al., 2007a). 5,2008 Satellite indices for primary production estimates

LUE-model
In order to further assess the applicability of satellite-based measurements we incorporated indices in a simple parametric LUE-model formulated as follows: where ε p represents the maximum biological efficiency of PAR conversion to dry matter 5 (g C mol −1 PAR).
Most herb layers in the Sahel region consist of a combination of C 3 and C 4 species, where C 3 often dominates in the early part of the growing season (Hanan and Prince, 1997). We therefore prescribed a fixed value of 0.489 g C mol −1 PAR for ε p (Seaquist et al., 2003;Seaquist et al., 2006), which indicates a mixture of C 3 and C 4 plants with 10 the assumption that C 4 grasses dominate for the greater part of the growing season.
SIWSI was linearly scaled as ε using maximum and minimum values during the 2007 season to assess its potential as a scalar of maximum light use efficiency. The index has previously been shown to increase predictions of above ground net primary production in the semi-arid Sahel (Fensholt et al., 2006).

15
EVI has been shown to be highly correlated with processes that are dependent on light absorption (Huete et al., 2002;Xiao et al., 2004;Rahman et al., 2005;Olofsson et al., 2007a). It therefore replaces fPAR in Eq. (1) whereas the PAR component, in turn, represents 8-day sums of measurements from the climate station in order to be consistent with the temporal resolution of the MOD09A1 product. Interactive Discussion production (AGNPP) assessments of herbs and grasses (negative values denote net ecosystem CO 2 uptake). It is apparent in these data that net uptake increases as conditions gradually become more favourable for plant growth with the onset of the rain in June (Fig. 2c). Total annual rainfall for the 2007 season was 364 mm, which is slightly higher than the annual average of 320 mm and noticeably higher compared to 5 previous season as Demokeya received a total of 273 mm in 2006. Peak uptake occurs in late August and early September. GPP and R eco decline rapidly as soil water content decreases with the last rains falling on 22 September (Fig. 2c).
The seasonal progression of TIMESAT-adjusted EVI is shown in Fig.'2d. Although vegetation index values are relatively low, the site shows a distinct seasonal variation 10 in EVI which corresponds rather well to the seasonal dynamics of 8-day sums of GPP.

GPP relationships and modelling
Linear regressions between EVI and SIWSI against GPP were computed and results are shown in Fig. 3. A strong linear relationship was observed between EVI and GPP ( Fig. 3a) with a coefficient of determination, R 2 =0.89 and RMSE=4.90 C m −2 8d −1 (Ta-15 ble 1). We also performed a linear regression analysis between EVI and tower GPP for the central 500 m pixel. Although the analysis did not show a considerable difference with R 2 =0.85 and RMSE=5.73 g C m −2 8d −1 we continued to use 3×3 pixel averages throughout our analyses. Nonetheless, the high correlations between tower GPP and EVI suggest that the latter may be usable in estimating seasonal variation of fPAR for 20 semi-arid environments. Furthermore, a strong relationship between SIWSI and GPP was detected (Fig. 3b). This is not surprising as an overall relation between vegetation intensity and water availability as observed by satellite could be expected. However, Fensholt and Sandholt (2003) found SIWSI to be highly correlated with NDVI, indicating a redundancy 25 between SIWSI and EVI. A comparison between EVI and SIWSI was subsequently performed (Fig. 4a) We further assessed SIWSI's potential as a scalar by testing the index against measurements of soil moisture at a depth of 5 cm (Fig. 4b). Although it can be argued that superficial soil moisture can be regarded as a surrogate measure of EWT, only a weak relationship (R 2 =0.45) with a RMSE of 1.57% was observed. It is also apparent in 5 Fig. 4b that the correlation is mainly attributed to dry season measurements. Additionally, it was also observed that EVI was stronger correlated to soil moisture with a higher R 2 of 0.57 and a lower RMSE of 1.46%. As previously stated, EVI and SIWSI preformed well when compared against 8-day sums of GPP (Fig. 3). However, results acquired through multiple linear regression, 10 with LUE-model parameters EVI and ε as independent variables, showed that ε did not add further explanation to the observed variance of GPP values (Table 1). Figure 5 shows 8-day modeled GPP at Demokeya with observed 8-day sums from July to December 2007, where modeled GPP was calculated every 8-day according to Eq. (6). As shown in Fig. 6, seasonal dynamics of modeled GPP agreed well with 15 measured GPP (R 2 =0.83), with cumulative sums over the period from model start on 12 July to the end on 27 December differing by 3.58% (Fig. 5b). The data points are shown by the equation GPP obs =0.93 GPP mod +0.68, which is close to a 1:1 line (Fig. 6). However, the model fails at the beginning of the growing season with the unexplained variance in the measured versus modeled data (6.15 g C m −2 8d −1 , Table 1) mainly 20 originating from the effects of ε.

Discussion
The aim of this study involved assessing the applicability of two satellite-based indices from MODIS data (EVI and SIWSI) for primary production modelling in a semi-arid environment in the Sahel. Indices were tested through comparison with site-specific 25 measurements of CO 2 fluxes from central Sudan. In order to evaluate the applicability of satellite indices, tower measurements of NEE were used to derive GPP. Soil moisture 2995 Introduction

Tables Figures
Back Close

Full Screen / Esc
Printer-friendly Version Interactive Discussion and soil temperature are usually regarded being the primary environmental factors controlling R eco in semi-arid environments and as such R eco was estimated by using a soil water content factor, representing the relative availability of soil water for plants (Wang and Leuning, 1998) together with the exponential regression model of (Lloyd and Taylor, 1994). This step is critical, as accurate estimation of R eco is important 5 with regards to validation of terrestrial carbon models as an erroneous estimate of R eco in turn will result in an error in the estimation of GPP. However, a full evaluation of the environmental factors driving respiration at this site and the applicability of the multiplicative model used to derive R eco is beyond the scope of this study. A strong relationship between EVI and tower GPP was observed at the Demokeya 10 flux site (Fig. 3a). This clearly indicates the usefulness of EVI in terrestrial carbon modelling. The good correlation was not unexpected, as the utility of this index in satellite-driven primary production modelling has previously been demonstrated for several different biome types (Xiao et al., 2004;Rahman et al., 2005;Sims et al., 2006b;Olofsson et al., 2007a). The highly linear relationship further suggests that 15 GPP can be estimated through a linear regression model for similar environments with relatively high accuracy, using only EVI as independent variable. We further detected a high correlation between SIWSI and measured tower GPP (Fig. 3b). Although linear regression parameters differed in comparison to the EVI-GPP relationship (Fig. 3a), a similar value range between SIWSI and EVI was ap-20 parent. Additionally, the strong relationship between the two MODIS indices (Fig. 4a) indicates that they both may measure the same process at this specific site. Similar results have been reported with NDVI. For instance, Ceccato et al. (2001; designed a spectral index, the Global Vegetation Moisture Index (GVMI), using NIR, blue and SWIR reflectance data from the VEGETATION sensor onboard the SPOT 25 (Satellite Pour l'Observation de la Terre). The authors demonstrated that GVMI was fully capable of predicting EWT for complete canopy cover, but comparisons between GVMI and NDVI for savanna regions showed a highly linear relationship (Ceccato and Flasse, 2002). Although Ceccato and Flasse (2002)  Interactive Discussion different values of GVMI, the authors mention that NDVI could be used to retrieve vegetation water content for some types of species, mainly for those where the degree of senescence is proportional to moisture content, which seems to be the case at the site used in our study. It is however important to mention that SIWSI does require a certain amount of vegetation to be present in order to be useful (Fensholt and Sandholt, 2003). 5 The rather weak correlation between SIWSI and soil moisture (Fig. 4b) indicates that the sparse vegetation cover at the site causes the index to fail as an indicator of water stress. Furthermore, the stronger, but still rather weak relationship observed between EVI and soil moisture could, to some extent, indicate that the effects of water stress are already manifested through the EVI signal. Cheng et al. (2006) compared several 10 indices, including SIWSI and EVI, to retrievals of EWT from Advanced Visible Infrared Imaging Spectrometer (AVIRIS) imagery. The authors showed that EVI had the highest correlation among indices for an agricultural site and a semi-arid savanna shrub site. However, it would have to be further investigated if EVI can mimic the temporal dynamics of EWT at these specific sites as Cheng et al. (2006) further concluded that 15 errors due to soil background reflectance and canopy architecture where inherent in the retrievals of EWT in both the AVIRIS and MODIS data. Results derived using a simple parametric LUE-model were shown to agree rather well with measured tower GPP over the 2007 season (Fig. 5). Although correlations showed that the relationship between EVI and GPP was highly linear (Fig. 3a), there 20 are still a number of factors that influence the vegetation signal recorded at the sensor that in turn can greatly influence the 1:1 relationship assumed between fPAR and EVI. Even though TIMESAT minimizes negatively biased noise due to the interference of clouds and atmospheric constituents, effects of varying solar zenith angles on satellite vegetation index data has previously been shown to be considerable at intermediate Interactive Discussion though the site did show a distinct seasonal variation in EVI (Fig. 2d), the applicability of vegetation indices to estimate GPP can still be greatly reduced due to solar angle effects, specifically for sites in semi-arid regions where vegetation is sparse (Sims et al., 2006a). But none of the reflectance values in the 3×3 window during 2007 were acquired at a solar zenith angle of more than 45 • , and as vegetation indices estimated 5 by using red and NIR reflectance are relatively unaffected at solar zenith angles less than 50 • (Goward and Huemmrich, 1992), the effects of solar angle on derived EVI may be minimal. The large differences between modeled GPP and observed GPP, specifically for a few 8-day periods early in the model run (Fig. 4a), can be attributed to ε (i.e. scaled 10 SIWSI) which primarily fails due to the low vegetation cover at the beginning of the growing season. As the water content of the soil gradually increases over time with the first rainfall, the overall albedo of the soil decreases. The reflectance in the SWIR, in turn, rapidly diminishes, causing a peak in the modeled GPP during the middle of the growing season. This rather deceptive temporal pattern of ε in the early vegetative 15 stage is further enhanced due to scaling. Finally, multiple linear regression analysis confirmed that no significant improvement was obtained by adding ε to the model, suggesting that, in this case, the parameter is of limited predictive use.

Conclusions
In order to test the applicability of MODIS EVI and SIWSI in primary production mod-20 elling for semi-arid areas in the Sahel, tower measurements of CO 2 fluxes from central Sudan were partitioned into R eco and GPP using a model that incorporates the Lloyd and Taylor (1994) equation together with a soil water factor (Wang and Leuning, 1998). Both indices showed consistent agreement with GPP with EVI having the highest correlation. The strong GPP-EVI relationship observed demonstrates that EVI show sig-25 nificant promise for efficient determination of primary production at similar ecosystems.
SIWSI was compared against data on soil moisture to assess its applicability as a BGD 5,2008 Satellite indices for primary production estimates Interactive Discussion measure of water stress. A rather weak correlation was observed and comparison between SIWSI and EVI illustrated that there is a considerable redundancy between the two indices. Results obtained through multiple linear regression as well as through implementing a simple parametric LUE-model demonstrated that SIWSI did not add further explanatory power to measured GPP values. The index broke down early in the 5 season due most likely to low vegetation cover, indicating that its use as a predictor of water stress in similar ecosystems, where vegetation fraction is low, is restricted. Research using multi-year and site-wide flux tower and climate data sets is however required to further test the use of EVI and SIWSI in satellite data-driven primary production modelling over semi-arid areas in the Sahel. 5,2008 Satellite indices for primary production estimates