International Journal of Remote Sensing Vol. 26, No. 12, 20 June 2005, 2561–2594

Evaluation of MODIS and NOAA AVHRR vegetation indices with in situ measurements in a semi-arid environment R. FENSHOLT* and I. SANDHOLT Institute of Geography, University of Copenhagen, Øster Voldgade 10, DK–1350 Copenhagen, Denmark (Received 9 January 2004; in final form 10 September 2004 ) Much effort has been made in recent years to improve the spectral and spatial resolution of satellite sensors to develop improved vegetation indices reflecting surface conditions. In this study satellite vegetation indices from the Moderate Resolution Imaging Spectroradiometer (MODIS) and the Advanced Very High Resolution Radiometer (AVHRR) are evaluated against two years of in situ measurements of vegetation indices in Senegal. The in situ measurements are obtained using four masts equipped with self-registrating multispectral radiometers designed for the same wavelengths as the satellite sensor channels. In situ measurements of the MODIS Normalized Difference Vegetation Index (NDVI) and AVHRR NDVI are equally sensitive to vegetation; however, the MODIS NDVI is consistently higher than the AVHRR NDVI. The MODIS Enhanced Vegetation Index (EVI) proved more sensitive to dense vegetation than both AVHRR NDVI and MODIS NDVI. EVI and NDVI based on the MODIS 16day constrained view angle maximum value composite (CV-MVC) product captured the seasonal dynamics of the field observations satisfactorily but a standard 16-day MVC product estimated from the daily MODIS surface reflectance data without view angle constraints yielded higher correlations between the satellite indices and field measurements (R2 values ranging from 0.74 to 0.98). The standard MVC regressions furthermore approach a 1 : 1 line with in situ measured values compared to the CV-MVC regressions. The 16-day MVC AVHRR data did not satisfactorily reflect the variation in the in situ data. Seasonal variation in the in situ measurements is captured reasonably with R2 values of 0.75 in 2001 and 0.64 in 2002, but the dynamic range of the AVHRR satellite data is very low—about a third to a half of the values from in situ measurements. Consequently the in situ vegetation indices were emulated much better by the MODIS indices than by the AVHRR NDVI.

1.

Introduction

Past decades of Earth observation (EO) data have improved our understanding of the seasonal dynamics of Sahelian vegetation and its response to both short- and long-term environmental stress. The payload of the satellites, however, has improved significantly with technological advances since vegetation was first observed regularly from space and thereby new enhanced methods for vegetation monitoring have developed. The Normalized Difference Vegetation Index (NDVI) derived from the Advanced Very High Resolution Radiometer (AVHRR) satellites *Corresponding author. Email: [email protected]. Tel: + 45 35 32 25 26; Fax: + 45 35 32 25 01 International Journal of Remote Sensing ISSN 0143-1161 print/ISSN 1366-5901 online # 2005 Taylor & Francis Group Ltd http://www.tandf.co.uk/journals DOI: 10.1080/01431160500033724

2562

R. Fensholt and I. Sandholt

emerges as one of the most robust tools for monitoring vegetation and since the early 1980s the NOAA AVHRR satellites have provided data vital to a range of land applications. An extensive literature covers the use of the AVHRR NDVI for intra- and interannual vegetation dynamic studies at a regional scale, as well as for monitoring climate-related phenomena at the continental and global scale (Myneni et al. 1998, Buermann et al. 2001, Nemani et al. 2003). Seasonal transects of NDVI have been used to monitor the propagation and contraction of the Sahara desert (Tucker et al. 1991, Olsson et al. in prep.). Land cover mapping using monthly composites of NDVI is the subject of a study by Loveland et al. (1991), while DeFries and Townshend (1994) and Malingreau et al. (1989) have demonstrated the utility of the AVHRR data for monitoring tropical deforestation on a global scale. As the AVHRR sensor was not originally designed to monitor vegetation, the research has revealed some disadvantages in the payload of the AVHRR platforms regarding the design of the channel 1 and 2 when formulating NDVI. Built on this knowledge a new and improved generation of EO data has emerged, including the polar orbiting Moderate Resolution Imaging Spectroradiometer (MODIS) sensor aboard the National Aeronautics and Space Administration’s (NASA) TERRA and AQUA platforms. These sensors hold promise for environmental monitoring in general and estimation of vegetation indices in particular, as they offer radiometric and spatial resolution superior to that of the AVHRR sensor (Huete et al. 2002, Trishchenko et al. 2002). The growing season time integral of NDVI has successfully been used as an estimator of net primary production (NPP) (Tucker and Sellers 1986, Diallo et al. 1991, Prince 1991, Ruimy et al. 1994, Prince and Goward 1995). Many studies have also related NDVI to biophysical vegetation variables like fAPAR (the Fraction of Photosynthetically Active Radiation absorbed by vegetation canopies) (e.g. Asrar et al. 1984, Sellers 1985, Running and Nemani 1988, Goward and Huemmrich 1992) and LAI (the one-sided green leaf area per unit ground area) (Asrar et al. 1984, Be´gue´ 1993). To obtain accurate estimates of new satellite derived biophysical variables (MODIS LAI, fAPAR and NPP) it is however necessary to evaluate the vegetation indices because fAPAR, LAI and NPP to some extent are based on the same spectral information, as used for deriving vegetation indices. Two full years of MODIS TERRA data (2001 and 2002) are now available and as a part of the MODIS Land products, two vegetation indices (NDVI and Enhanced Vegetation Index (EVI)) are produced and distributed by the MODIS Land Discipline Group (the MOD13 products at http://modis-land.gsfc.nasa.gov). These indices are both calculated from reflectance measurements in the visible and nearinfrared (NIR) wavelengths; however, the purpose and the application of each index is unique. The MODIS EVI is designed to be resistant to atmospheric water vapour and aerosol contamination by adding information from the blue wavelength (Justice et al. 1998). The effects of canopy background variability including differences in soil type, litter and water content are also expected to be reduced (Huete et al. 1999, Gao et al. 2000). The MODIS continuity vegetation index (MODIS NDVI) is, by contrast, calculated in the same manner as the well-known standard NDVI which was introduced over 20 years ago (Deering 1978, Tucker 1979). The purpose is to create a scientific link to the 20 years of NOAA AVHRR data (van Leeuwen et al. 1999) thus allowing making comparisons to the time series of AVHRR data even though the MODIS wavebands are much narrower than for AVHRR. The channels

Evaluation of MODIS and AVHRR VIs with in situ measurements

2563

Table 1. Different vegetation indices and wavelengths used. Vegetation index

Blue (mm)

NOAA AVHRR NDVI MODIS NDVI MODIS EVI

Channel 3 0.46–0.48

LANDSAT NDVI

Red (mm)

Near Infrared (NIR) (mm)

Channel 1 0.58–0.68 Channel 1 0.62–0.67 Channel 1 0.62–0.67 Channel 3 0.63–0.69

Channel 2 0.72–1.00 Channel 2 0.84–0.87 Channel 2 0.84–0.87 Channel 4 0.75–0.90

and corresponding wavelengths for the three vegetation indices discussed are given in table 1. Evaluation of VI in different biomes is an important and necessary ongoing process when developing new EO data products providing a global coverage. The fieldwork sites for this study form part of the global fluxnet (2003) sub-sites (http:// public.ornl.gov/fluxnet) and the aim of this paper is to compare the quality of the new vegetation indices from the MODIS sensor to the conventional AVHRR NDVI in the semi-arid ecosystem of western Africa. The remotely sensed indices were evaluated using in situ measured indices. Whether the MODIS indices are an improvement over the AVHRR NDVI will be determined by field measurements of vegetation indices using multispectral radiometers designed with identical wavelengths to the MODIS and AVHRR channels in use. This will highlight the importance of differences in sensor configuration in the MODIS and AVHRR with respect to vegetation monitoring. 2. 2.1

Theoretical background NOAA AVHRR NDVI and MODIS NDVI

Conventional ‘ratio-based’ VIs have been in use for decades as the approach reduces the influences of varying topography, illumination conditions and atmosphere to some extent. The NOAA AVHRR NDVI and MODIS NDVI are given by: NDVI~

NIR{RED NIRzRED

ð1Þ

Difference in the performance of the two indices originates from the difference in the spectral resolution of AVHRR and MODIS channel 1 and 2 response functions as can be seen from table 1 and figure 1. These differences alter the influence of atmospheric water vapour and the sensitivity to vegetation reflectance. The MODIS red channel is narrower than the AVHRR red channel. The sensitivity of spectral reflectance to chlorophyll content is very dependent on wavelength in the visible part of the spectrum (Gitelson and Merzlyak 1996) and therefore even small variations in the spectral location of channel in the red wavelength might influence the value of NDVI. Chlorophyll-A pigment absorption reaches a maximum at 670 nm and figure 1 illustrates that 670 nm is the upper wavelength limit for both the AVHRR and MODIS red channel. The narrower MODIS red band thus causes it to be more sensitive to small chlorophyll variations for low densities of chlorophyll content but on the other hand the MODIS red

2564

R. Fensholt and I. Sandholt

Figure 1. Illustration of the MODIS and AVHRR spectral response curves channels 1 and 2 with superimposed response curves for green grass measured in the laboratory (source: Aster Spectral Libraries 2003) and for a natural grass surface measured by AVIRIS hyperspectral data (source: Williams and Hunt 2002).

channel saturates earlier than the AVHRR red channel for intermediate to high densities of chlorophyll content (Gitelson and Kaufman 1998) whereby the MODIS NDVI no longer responds to variations in green biomass. In the NIR wavelength area the MODIS and AVHRR response functions are highly divergent (figure 1). Spectral reflectance from a vegetated surface is superimposed on the graph of response functions, illustrating how differences in NIR channel widths have implications for the derived NDVI signal in terms of the chlorophyll reflectance. In general the AVHRR NDVI tends to be lower than the MODIS NDVI because the NIR band of AVHRR includes the major part of the red edge area around 700 nm whereas the red edge area is not covered by the MODIS NIR band (Galva˜o et al. 2000). The two vegetation reflectance curves illustrated in figure 1 have very differing origins. The US Geological Survey (USGS) Digital ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer) Spectral Library (2003) green grass laboratory test is performed at leaf level while the AVIRIS (Airborne Visible/Infrared Imaging Spectrometer) is a pixel averaged field spectrum for grass (Williams and Hunt 2002) where each pixel represents an area of approximately 400 m2. Yet the two reflectance spectra are remarkably alike in the red edge area (690–750 nm) suggesting that the position of

Evaluation of MODIS and AVHRR VIs with in situ measurements

2565

the red edge generally is very stable for most types of unstressed vegetation in the vegetative phase. MODIS channel 2 consequently is centred in the middle of the NIR plant reflectance plateau whereas AVHRR channel 2 also covers an area where vegetation absorbs most of the radiation. A coarse resolution pixel like those recorded by MODIS and AVHRR is influenced by patches of bare soil, stems, senescent plant material and shadows. The effect of the difference between the broad- and narrow-banded channel 2 on non-photosynthetic vegetation is even stronger than on green vegetation, potentially causing larger variations between MODIS and AVHRR NDVI towards the end of the growing season (Galva˜o et al. 1999). With regard to atmospheric water vapour absorption, the MODIS sensor is designed more appropriately than the AVHRR sensor when calculating NDVI. This can be seen from figure 2, which compares the spectral response of MODIS and AVHRR channels 1 and 2 with the wavelength dependent absorption by atmospheric water vapour. The MODIS NIR band is placed in an atmospheric water vapour window where there is no absorption by water vapour in the atmosphere. The response curve of the NIR channel onboard the AVHRR 16 sensor covers the area from 0.72–1 mm including an area of considerable water vapour absorption

Figure 2. Illustration of the MODIS spectral configuration channels 1 and 2, the NOAA AVHRR channels 1 and 2 and the corresponding water vapour absorption bands.

2566

R. Fensholt and I. Sandholt

from 0.9–0.98 mm. Absorption by atmospheric water vapour is thus substantially reduced in the MODIS EVI and NDVI, thus yielding a more precise VI estimate (Huete et al. 1999). To overcome these differences Heidinger et al. (2002) suggest the use of MODIS channel 2 and the additional spectral information offered by MODIS to calibrate AVHRR channel 2. Landsat NDVI is also calculated from equation (1) and the spectral configuration of the Landsat band 3 and 4 used can be seen in table 1. From figure 2 it is clear that the Landsat NIR band overlaps an area of atmospheric water vapour absorption. However the Landsat NIR band does not cover the strong water vapour absorption area from 0.9–1 mm, so atmospheric influences on Landsat NDVI are not expected to be as severe as on AVHRR NDVI. The Landsat red band is slightly narrower than AVHRR (table 1), but marginally broader than the MODIS red band. 2.2

MODIS EVI

The MODIS EVI is designed to remove some of the influences inherent to the mixing of soil/vegetation reflectance signal (Gao et al. 2000). The algorithm corrects for canopy background contamination without introducing new influences from the atmosphere. Earlier attempts to minimize the effects of variable atmosphere had increased sensitivity to variations in soil background and vice versa (Leprieur et al. 1994, Qi et al. 1994). EVI is designed to approximate the vegetation index isolines in red–NIR space to the measured or modelled vegetation biophysical isolines (Huete et al. 1999) and is given by: EVI~

NIR{RED ð1zLÞ NIRzC1
RED{C2
BLUEzL

ð2Þ

All input channels are corrected for atmospheric Rayleigh scattering and ozone absorption. The constants C1, C2 and L are empirically determined to be 6.0, 7.5 and 1.0 (Huete et al. 1999). C1 and C2 are designed to compensate for residual aerosol influences in the red band by incorporation of the blue band. This method was originally developed as the Atmosphere Resistant Vegetation Index (ARVI) by Kaufman and Tanre´ (1992). L is a canopy background adjustment factor originally introduced in the Soil Adjusted Vegetation Index (SAVI) by Huete (1988). Background reflectance from soil can be considerable in partially vegetated areas due to nonlinear differences in absorption, reflectance and transmittance of photosynthetically active radiation (PAR) and NIR between soil and canopy. Consequently canopy background noise is most severe at intermediate levels of vegetative cover (Huete 1988). Correction for soil influence is vital for vegetation monitoring in the Senegal and the semi-arid West Africa: the Sahel is dominated by open canopies like grasslands, savannas, shrub lands and other sparsely vegetated areas. MODIS EVI is designed to be sensitive to variations in a dense vegetation cover (Justice et al. 1998) compared to MODIS NDVI, which is designed to capture changes in low to intermediate vegetation intensities. Because the nonlinear stretch of NDVI is larger than for EVI, the EVI values increase slower than NDVI (Huete et al. 1999). 3.

Study area

Senegal is located in the western part of the Sudano-Sahelian zone (figure 3) and is characterized by a pronounced rainfall gradient. The northern and central part of

Evaluation of MODIS and AVHRR VIs with in situ measurements

2567

Figure 3. The extent of the Sahelian region with Senegal located on the Atlantic seaboard. The northern boundary of Sahel is often defined by the 100 mm isohyet while the Southern boundary is designated by the 600–800 mm isohyets (Prince et al. 1995).

the country is a marginal area in respect to a stable presence of vegetation which is due to the erratic inter- and intra-annual variations in rainfall characterizing this zone. The rainy season is from July to September and the precipitation is sparse and fluctuates, with annual totals ranging from 100–500 mm. In southern Senegal the rains begin in early June and last until October, with annual totals exceeding 1000 mm. Rainfall is the limiting parameter for vegetation growth in most of the SudanoSahelian zone (Hendricksen and Durkin 1986, Hielkema et al. 1986, Malo and Nicholson 1990) and the significant rainfall gradient is clearly echoed in a corresponding gradient in the density (figure 4) and type of vegetation. In northern Senegal the vegetation is dominated by fine-leaved annual grasses and scattered bush steppe mixed with woody vegetation with a very sparse tree cover of less than 1% (Diallo et al. 1991). Annual grasses with a maximum height of 60 cm are abundant, but widely spaced perennial grasses with a maximum height of 80 cm are also found (Ridder et al. 1983). To the south the vegetation gradually changes to perennial grasses comprising a continuous herbaceous layer at least 80 cm high. Rain-fed crops (millet, groundnut and sorghum) are more common and there is a pronounced increase in tree-cover (generally above 30%) (Hanan et al. 1991). Fieldwork was carried out in the semi-arid central and northern part of the country (figure 4) where rainfall averages 300–450 mm per year. Intermittent water shortages during the growing season occasionally lead to severe droughts with food shortages and production deficiencies. The vegetation of the study area mainly consists of the annual grasses Schoenefeldia gracilis, Dactyloctenium aegupticum, Aristida mutabilis and Cenchrus bifloures. Tree and shrub canopy cover generally does not exceed 5% in the study area and is dominated by two species; Balanites aegyptiaca and Boscia senegalensis (Diallo et al. 1991).

2568

R. Fensholt and I. Sandholt

Figure 4. MODIS NDVI satellite image covering Senegal (September 2001). Bright areas indicate very sparse vegetation while darker areas towards the south indicate more dense vegetation. Locations of the fieldwork sites are indicated. Pixel resolution is 250 m.

4.

Data

The study is based on MODIS TERRA Land data, NOAA AVHRR 16 data and in situ measurements from 2001 and 2002. In addition, information from the Landsat Enhanced Thematic Mapper Plus (ETM + ) sensor is used in a point-to-pixel comparison between in situ measurements and MODIS/AVHRR to explore potentially scale related differences. This involves determining the relationship between Landsat ETM + and MODIS indices and subsequently the relationship between Landsat ETM + and in situ measurements, thereby establishing a relation between MODIS and in situ measurements. A direct comparison between the different vegetation indices measured from satellite data is presented by Fensholt (2004) and is therefore excluded from this article.

Evaluation of MODIS and AVHRR VIs with in situ measurements 4.1

2569

In situ measurements

Field measurements were made at one location (Dahra North) in 2001 and at four different locations in 2002 (two sites near Dahra 10 km apart and two near Tessekre, also 10 km apart) illustrated in figure 4. Average rainfall is 300–450 mm in the study areas; 2001 was a slightly drier year with an annual total of 288 mm, while 2002 was an extremely dry year with a total of 150–200 mm, less than 50% of the average. Furthermore the seasonal distribution of rain was uneven and characterized by long dry spells resulting in a very short growing season and extremely low vegetation cover. Incident and reflected radiation were measured in situ from June to November in order to calculate time series of VIs, at wavelengths similar to the wavelengths used for VI monitoring onboard the TERRA MODIS and NOAA AVHRR sensors (see table 1). Measurements were performed using the SKYE Instruments two-channel sensors (SKR 110 series) and four-channel sensors (SKR 1850 series) (http:// www.skyeinstruments.com/light.htm). All instruments were inter-calibrated at location before and after each growing season and sensor calibration using a standard reference was furthermore performed by SKYE instruments both years. Sensors were mounted on a 6 m aluminium mast and data were collected using Campbell CR10 dataloggers sampling every 10 min. Figure 5 shows the Dahra North site 2001 in the peak of the growing season. MODIS EVI, NDVI and AVHRR NDVI are derived from the radiometric measurements. Ideally MODIS and AVHRR data should be compared to 10:30 am and 2:30 pm in situ measurements, respectively; however, using instantaneous in situ measurements for comparison introduces considerable noise due to sub-pixel clouds causing

Figure 5. Photograph of the Dahra North site 1 September 2001. The picture illustrates the fieldwork equipment mounted on a grass savanna in the peak of the growing season where height of the grass is approximately 50 cm (UTM: 453484; 1703779, zone 28).

2570

R. Fensholt and I. Sandholt

Figure 6. In situ measured MODIS NDVI daily variations (ten minute values) for a cloudfree day (values corresponding to MODIS and AVHRR sensor overpass are highlighted).

variations in reflectance intensity and thereby VI values. On a diurnal basis NDVI was found to be highest in the morning and late afternoon due to large solar zenith angles and lowest around noon where the solar zenith angle is low. It was therefore found that daily averages of VI calculated from 9:00 am to 3:00 pm approximated the 10:30 am and 2:30 pm values as illustrated in figure 6 and table 2. An in-depth description of diurnal NDVI variations as a function of the Bidirectional Reflectance Distribution Function (BRDF) will be the subject of a forthcoming paper. The sensors incorporate a cosine diffuser whereby the capture of light depends on the cosine of the angle the ray of light makes to the axis down the length of the sensor. The footprint of the sensors is visualized by figure 7. The full line shows sensor response as a function of degrees from nadir while the dashed line gives the corresponding coverage at Table 2. Statistics of 10 min in situ measured MODIS EVI from 9:00 a.m. to 15:00 p.m. 18 September 2001. Mean EVI Standard error Median Standard deviation Sample variance Range Minimum EVI Maximum EVI

0.5655 0.0031 0.5647 0.0189 0.0004 0.0674 0.5330 0.6004

Evaluation of MODIS and AVHRR VIs with in situ measurements

2571

Figure 7. Graphical illustration of the footprint of the SKYE sensors from a 6 m mast. The full line is the sensor response as a function of degrees from nadir and the dashed line gives the corresponding coverage on the surface.

the surface. The line drawn from the secondary y-axis is matching the coverage of a 30 m Landsat pixel. At off-nadir view angles of 66% the sensors cover an area corresponding to a Landsat pixel and it can be seen that the percentage sensor response is still 40% with this angle. For homogeneous surfaces like the grass savanna as in figure 5, the instrument set-up can reliably imitate the signal from a 30 m Landsat pixel. When measuring reflectance from vegetation the sensor view angle is however an important determinant because the vegetation appears denser with increasing view angles. This is especially pronounced for savanna grasses characterized by an erechtophile leaf angle distribution (Goel 1989, Gao 1993). When comparing in situ NDVI measurements from sensors recording hemispherical reflectance to Landsat and MODIS NDVI recording directional reflectance, the in situ measurements therefore can be expected to yield higher values than the corresponding Landsat values and MODIS values derived from images recorded with low scan angles. Therefore, the absolute values of in situ and satellite measured VI are not directly comparable, but in situ and satellite data can be compared in a relative sense to assess their abilities to capture surface variations. The view geometry differences will be further discussed in the pointto-pixel comparison section. The footprint calculations presented in figure 7 do not account for anisotropy in the reflectance data, but working with daily averages is expected to reduce the influence of anisotropy. Table 3 summarizes the fieldwork set-up with respect to which sensors were used and consequently which types of VI can be derived. In both years a mobile mast was used to carry out measurements of different crop types at various phenological stages. In 2001 the mobile mast was equipped with a two-channel sensor and in 2002 with a four-channel sensor.

2572

R. Fensholt and I. Sandholt

Table 3. Overview of the distribution of the different SKYE sensors and the derived VIs. Field work location Dahra North 4-channel sensor Dahra South 2-channel sensor Tessekre North 2-channel sensor Tessekre South 2-channel sensor Mobile mast (used for crop measurements)

4.2

2001 In situ measurements

2002 In situ measurements

MODIS EVI, NDVI and AVHRR NDVI

MODIS EVI, NDVI and AVHRR NDVI MODIS NDVI MODIS NDVI MODIS NDVI

MODIS NDVI (2-channel sensor)

MODIS EVI, NDVI and AVHRR NDVI (4-channel sensor)

Satellite data

Table 4 outlines the specifications of the MODIS Land and the AVHRR satellite products. MODIS collection three data from 2001 and 2002 are provisional, meaning that the products are only partially validated and improvements are continuing. The MODIS 500 m VI products are based on the level 2 daily surface reflectance product (MOD09 series), which is corrected for the effect of atmospheric gases, thin cirrus clouds and aerosols. The correction scheme uses the MOD05 product to correct for water vapour, MOD04 for aerosols and MOD07 for ozone as well as MODIS band 26 for detection of thin cirrus clouds (Vermote and Vermeulen 1999). The MOD09 product employs MODIS bands 1–7 and is an estimate of the surface spectral reflectance for each band as would be measured at ground level if there were no atmospheric scattering or absorption (Vermote et al. 2002). The MODIS VI product is delivered as 16-day composites. NOAA AVHRR 2001 raw data were received at the Dartcom receiving station at Centre Suivi Ecologique in Dakar and were available for the authors at no cost. In 2002 the receiving station malfunctioned so instead data from the Mas Palomas ground station were acquired from Eurimage. To minimize costs, the 2002 AVHRR scenes were pre-screened to avoid poor view-angles and excessive cloud cover. Scenes recorded at large view angles (.40u) were rejected because of the undesirable effect of forward scattering (Cihlar et al. 1994, Lind and Fensholt 1999); scenes with extensive cloud cover were also rejected. These screening constraints are also applied to the processing of the AVHRR 2001 data as described below (NOAA AVHRR processing). The numbers of AVHRR scenes available in 2001–2002 are summarized in table 5. Both the MODIS TERRA and the NOAA AVHRR use a cross-track scan mirror with similar swath wide and range of view angles. The MODIS TERRA crosses the equator at 10:30 am and the AVHRR 16 at 2–3:30 pm, resulting in solar zenith angles of 5–15u for the MODIS data and 15–35u for the AVHRR 16 data. Solar azimuth angles also differ, and both conditions influence the BRDF. 5. 5.1

Data processing NOAA AVHRR processing

NOAA AVHRR data processing is performed automatically using standards developed at the Institute of Geography, University of Copenhagen (IGUC). The

Table 4. The MODIS and NOAA AVHRR data products. Year

MODIS/Terra Vegetation Indices 16-Day Level 3 Global 500m ISIN Grid V003 MODIS/Terra Surface Reflectance Daily Level 2 Global 500m ISIN Grid V003 NOAA AVHRR (NOAA 16) SHARP 1 Level 1 Daily Local Area Coverage 1.1 km LANDSAT 7 ETM + Level 1G product 30 m pixel resolution

2001–2002 2001–2002 2001–2002 2002 (4)

Table 5. The temporal availability of MODIS, AVHRR, Landsat ETM data products and in situ measurements. 16-day comp. period 11 12 13 14 15 16 17 18 19 20

Year 2001 Period 10/6–25/6 26/6–11/7 12/7–27/7 28/7–12/8 13/8–28/8 29/8–13/9 14/9–29/9 30/9–15/10 16/10–31/10 01/11–16/11

Landsat 7 ETM +

MODIS VIs No No Yes Yes Yes Yes Yes Yes Yes Yes

Year 2002

AVHRR number In situ of scenes measurements 9 10 7 11 8 15 12 10 15 13

Landsat 7 ETM + 30/6 1/8 17/8

Yes Yes Yes Yes Yes

4/10

MODIS VIs

AVHRR number of scenes

Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes

2 5 3 5 2 5 5 8 8 5

In situ measurements Yes Yes Yes Yes Yes Yes Yes Yes Yes

Evaluation of MODIS and AVHRR VIs with in situ measurements

Data product

2573

2574

R. Fensholt and I. Sandholt

processing employs software developed at IGUC for the WINCHIPS image processing software package (Hansen 2001). Dybkjaer et al. (2003) provide a detailed documentation of the processing chain from raw image to a final product which is radiometrically calibrated, cloud-masked and geometrically rectified. Radiometric calibration is performed following the NOAA NESDIS guidelines using the method suggested by Rao and Chen (1995, 1996) for channels 1 and 2. Images are rectified to UTM coordinates using a nearest-neighbour re-sampling algorithm, with geometrical registration achieving sub-pixel accuracy. All images are screened for high view angles with a view angle constraint of 42u corresponding to a view zenith angle of 49u. The commonly used maximum value composite (MVC) method is an efficient method of compositing AVHRR data not corrected for atmospheric water vapour (Holben 1986, Cihlar and Howarth 1994). MVC images of NDVI are transformed to 16-day composite periods matching the MODIS composite periods (table 5) to allow direct comparisons with MODIS data. 5.2

MODIS processing (compositing)

Even though the MVC approach is known to minimize atmospheric influence, many studies have shown that applying the MVC method on AVHRR NDVI data favours pixels with a high forward scatter, causing an overestimation of VI (Huete et al. 1992, Cihlar et al. 1994, Lind and Fensholt 1999). This is caused by the strong reflectance anisotropy within the red spectrum of many vegetated surfaces, whereas no reflectance anisotropy is observed in the NIR spectrum. When calculating VI from the MODIS data, surface anisotropy is expected to be even more pronounced than in the case of AVHRR data since MODIS reflectance values are atmospherically corrected prior to compositing and VI computation (Huete et al. 1999). For the MODIS compositing a two criteria composite technique is therefore used considering both NDVI and view angle. This technique is designed to constrain the forward scatter selection encountered in the MVC selection method and is called constrained view angle MVC (CV-MVC) (Huete et al. 2002). For a given 16-day period the algorithm considers the two highest VI values and selects the VI associated with the lowest view angle. The MODIS VI algorithms use daily surface reflectance data as input and composite these to generate 16-day 500 m products. For a given composite period an initial data screening divides the available data into three categories of decreasing quality as illustrated in figure 8. If two or more good quality images (category 1) are available the CV-MVC is applied. In the case of only one good (category 1 and 2) observation in a 16-day period, the composite VI value is computed from this observation. If no observations pass the initial data screening in a 16-day period, then the ordinary MVC technique is applied to the available data regardless of data quality (category 3 included). The highest VI pixel is thus selected even if it is influenced from clouds, opening up for potentially serious underestimates of VI. 5.3

MODIS post processing

The following processing is not part of the standard MODIS product package but is performed by the authors. MODIS Land data are originally projected onto an Integerized Sinusoidal (ISIN) mapping grid. These are rectified to UTM coordinates using the MODIS re-projection tool (MODIS Reprojection Tool distribution page, 1 June 2003). The re-sampling of image data is done by bilinear re-sampling

Evaluation of MODIS and AVHRR VIs with in situ measurements

Figure 8.

2575

MODIS VI composition flow chart. Modified after Huete et al. (2002).

algorithms and nearest-neighbour re-sampling algorithm is used for the quality assurance data. Image data are extracted using the WINCHIPS image processing software package (Hansen 2001). 5.4

Landsat processing

Landsat 7 ETM + data have been converted to top of atmosphere (TOA) reflectance using the equations and constants in the Landsat user’s handbook (2002). Landsat scenes having a 30 m pixel resolution are geometrically rectified to sub-pixel accuracy using band 8 (panchromatic, 15 m pixel resolution) and ground control points from the area. The Landsat acquisition dates listed in table 5 represent all the cloud-free scenes available for 2001 and 2002. No cloud-free Landsat scenes were available for the growing season 2001, making direct validation of the point-to-pixel comparison impossible for the 2001 data. 6.

Methodology

6.1 Point-to-pixel comparison (MODIS and AVHRR) using Landsat ETM + images The purpose of the scaling exercise is to test the validity of performing point-to-pixel comparison between in situ measurements and MODIS/AVHRR pixels. When comparing in situ measurements of NDVI with a MODIS 500 m NDVI pixel, information from Landsat 30 m NDVI pixels is a useful tool for doing point-to-pixel comparison. Expressed as a logical argument we have A (MODIS NDVI), B (Landsat NDVI) and C (in situ NDVI). By relations between A and B, and B and C it is possible to relate to C from A. In this way it is possible to evaluate whether the sites of in situ measured NDVI are representative of corresponding MODIS 500 m NDVI pixels. In this way Landsat NDVI is merely used for validation of performing point-to-pixel comparison between in situ measurements and MODIS/AVHRR pixels and the quality of the Landsat NDVI itself is not the subject of the paper. The

2576

R. Fensholt and I. Sandholt

Figure 9. Illustration of point-to-pixel comparison analysis. Step 1: comparison of 500 m Landsat ETM + NDVI (calculated from aggregated Landsat ETM + channel 3 and 4 reflectance) to MODIS NDVI (500 m resolution). Step 2: comparison of the Landsat ETM + pixel NDVI (the pixel covering the sites of in situ measurements) to 289 surrounding Landsat ETM + NDVI pixels encompassed in the MODIS 500 m pixel. Step 3: comparison of in situ measured NDVI to Landsat ETM + pixel NDVI (30 m resolution).

comparison approach is illustrated stepwise in figure 9 and a detailed description is given below. The first step of point-to-pixel comparison is to determine the relation between Landsat and ETM + NDVI values and MODIS NDVI/EVI 500 m values. This is done by aggregating the Landsat channel 3 and 4 reflectance pixels covered by each MODIS pixel into 500 m pixels and subsequently calculating NDVI. The re-sampled Landsat NDVI and MODIS NDVI are compared along the transect line illustrated in figure 10. Three different cloud-free dates in 2002 are analysed and resulting vegetation indices are given in figure 11. For the three dates the MODIS scan angles along the transect are below 5u and are thereby comparable with the Landsat ETM data. On 30 June vegetation was very sparse and the Landsat NDVI is almost similar to the MODIS EVI and approximately 0.04 NDVI units below MODIS NDVI. By 1 August the vegetative cover had increased along the northern part of the transect and maximum Landsat NDVI exceeds MODIS EVI; however the margin between Landsat and MODIS NDVI remains constant. By 4 October, towards the end of the growing season, NDVI reaches its highest values. Differences between Landsat and MODIS NDVI of approximately 0.04 NDVI units remain constant regardless of the vegetation density. However the MODIS EVI level compared to the other indices varies with vegetation intensity. For low vegetation cover the difference between MODIS EVI and NDVI is about 0.06 NDVI units; this difference grows to 0.11 units for the highest vegetation intensity. This is because of the increased EVI sensitivity to high vegetation densities (see theoretical background section) causing the EVI values to

Evaluation of MODIS and AVHRR VIs with in situ measurements

2577

Figure 10. Landsat scene from 4 October 2002 with transect line and the four fieldwork sites illustrated in figure 3; OTM of fieldwork sites (from north to south); 493483, 1757279; 493463, 1748779; 453483, 1703779; 456983, 1699279.

increase slower than NDVI. To verify whether the relation between the MODIS NDVI and re-sampled Landsat NDVI is sensitive to changes in vegetation intensity, scatterplots for the three dates are presented in figure 12. All slopes are very close to 1 and offsets remain remarkably constant around 0.04–0.05 NDVI units, suggesting the relation to be insensitive to variations in vegetation density. It furthermore elucidates that possible variations in atmospheric water vapour in between the three dates of acquisition do not influence Landsat NDVI, which otherwise might be a potential source of error in the comparison approach, though atmospheric influence on Landsat NDVI is not expected to be as severe as on AVHRR NDVI as outlined in the theoretical section. Thus the relation between Landsat NDVI and MODIS NDVI is fairly robust, making the Landsat NDVI scenes suitable when comparing in situ vegetation indices to coarse resolution satellite images. The second step is to analyse the relation between the Landsat NDVI value of the pixel covering the fieldwork site and the surrounding 289 pixels Landsat NDVI pixels encompassed by the area of the MODIS pixel. The results of this analysis are presented in table 6. Valid comparison between in situ measurements and MODIS pixels requires that Landsat pixel NDVI covering the site of in situ measurements represent the mean NDVI value for the area identical to the MODIS pixel. Table 6 reveals that the study areas are generally very homogeneous. For an area of 289

2578

R. Fensholt and I. Sandholt

Figure 11. Transects of MODIS NDVI, EVI and Landsat 500 m NDVI (calculated from 500 m aggregated channel 3 and 4 reflectance) for three dates 2002 (pixel-to-pixel comparison). All MODIS data are recorded with scan angles less than 5u.

pixels the spread of Landsat NDVI values for all sites and dates is low, and values for the study area agree very well with the area mean. The third step in the scale comparison is to establish a relationship between Landsat NDVI and in situ NDVI. In figure 13 time series of in situ measured

Evaluation of MODIS and AVHRR VIs with in situ measurements

2579

Figure 12. Scatterplot of MODIS NDVI and resampled Landsat 500 m NDVI (transect values) for three dates 2002 (pixel-to-pixel comparison).

MODIS NDVI are compared to the Landsat NDVI pixel covering the fieldwork site for the 2002 growing season. No scenes were available for the period of maximum vegetative cover, but three scenes cover the beginning and the end of the season respectively. Landsat NDVI for the first two scenes (1 and 17 August) is consistently 0.09–0.1 NDVI units below the in situ MODIS NDVI for all four sites. On 4 October Landsat NDVI for the two Dahra sites is only 0.03 units below in situ NDVI. This last discrepancy cannot be explained by variable water vapour influence in the Landsat scenes as this effect would then be reflected in figure 12 through a larger offset in the plot from October, which is not the case. Differences within the time series are therefore difficult to explain. The most consistent result from figure 13 is nevertheless that in situ MODIS NDVI is in fact 0.09–0.1 units higher than Landsat NDVI. A relation between in situ measurements of MODIS NDVI and satellite MODIS NDVI can now be obtained from the relation between satellite MODIS NDVI and

2580

R. Fensholt and I. Sandholt

Table 6. Landsat pixels for the field work sites and statistics of 289 pixels encompassed by the 500 m MODIS pixel covering the location of in situ measurements; grass savanna 2002. Field work site Dahra North

Dahra South

Tessekre North

Tessekre South

NDVI statistics Landsat pixel value for field Mean of area Standard deviation of area Minimum of area Maximum of area Landsat pixel value for field Mean of area Standard deviation of area Minimum of area Maximum of area Landsat pixel value for field Mean of area Standard deviation of area Minimum of area Maximum of area Landsat pixel value for field Mean of area Standard deviation of area Minimum of area Maximum of area

work site

work site

work site

work site

01/08 2002

17/08 2002

04/10 2002

0.150 0.146 0.008 0.130 0.180 0.100 0.123 0.011 0.090 0.170 0.150 0.144 0.010 0.120 0.200 0.160 0.164 0.008 0.140 0.210

0.210 0.214 0.017 0.180 0.260 0.210 0.197 0.028 0.150 0.320 0.120 0.119 0.013 0.100 0.190 0.140 0.140 0.017 0.110 0.190

0.180 0.181 0.010 0.160 0.230 0.140 0.162 0.021 0.130 0.280 0.220 0.217 0.023 0.180 0.350 0.350 0.353 0.023 0.270 0.410

Figure 13. Grass savanna 2002. Time series of in situ measured vegetation indices compared to the corresponding Landsat ETM + NDVI pixel (point-to-pixel comparison). (At Tessekre South in situ measurements fail from 7 September.)

Evaluation of MODIS and AVHRR VIs with in situ measurements

2581

Landsat NDVI (step 1), knowledge of the representativeness of the Landsat ETM + pixel NDVI covering the sites of in situ measurements to the 289 surrounding Landsat ETM + NDVI pixels encompassed in the MODIS 500 m pixel (step 2) and the relation between Landsat NDVI and in situ NDVI (step 3). From the above relations it can be inferred that in situ MODIS NDVI is expected to be about 0.05 units higher than the level of the satellite MODIS NDVI. If the satellite MODIS channel 1–7 reflectance values are to be viewed as surface reflectance as it would have been measured at ground level with no atmospheric scattering or absorption (Vermote et al. 2002), then in situ MODIS NDVI should theoretically only be 0.04–0.05 NDVI units higher than Landsat NDVI at the site of the in situ measurement, as was the case for satellite MODIS NDVI compared to Landsat NDVI in figure 12. However, because of differences in scale, the values of satellite MODIS NDVI (even if they are perfectly atmospherically corrected) will probably never be identical to in situ measured NDVI. Vegetation appears denser at increasing view angles and the in situ NDVI measurements are therefore expected to exceed the corresponding Landsat NDVI value and satellite MODIS NDVI recorded with low scan angles (as laid out in the fieldwork section). No cloud-free Landsat scenes are available for the growing season 2001. Direct validation of whether the fieldwork site is representative of the corresponding MODIS pixel in 2001 is therefore not possible. However, analysis of the Dahra site in 2002 indicates that the vegetation density in the area encompassed by the MODIS pixel is extremely homogeneous (table 6), which is also confirmed by knowledge of that particular location acquired during four months of fieldwork in 2001 (photograph of the site is given in figure 5). 7.

Results and discussion

7.1

In situ measurements

Figure 14 compares daily values of the two in situ measured MODIS vegetation indices with in situ measured AVHRR NDVI for a grass savanna. The 2002 values are generally lower than for 2001 because of the sparse rainfall in 2002; however the scatterplots represent different phenological stages. The 2002 data cover the entire growing cycle whereas the time series in 2001 begins just before peak of growing season and lasts until dry down. For both years MODIS NDVI values exceed AVHRR NDVI and MODIS EVI. R2 values are high between the different indices and the relations are highly significant. Despite the high coefficient of explanation the slope of the regression between MODIS EVI and AVHRR NDVI differs significantly between the 2001 and 2002 data. Vegetation suffered from severe water stress in 2002 compared to 2001 (causing the grasses to appear greyish in 2002), which is known to alter the leaf internal structure because it changes the number and size of intercellular air-spaces controlling the size of the reflection (Tucker and Sellers 1986). Variations in canopy structural properties (e.g. variations in intercellular air-spaces caused by reduced leaf water content) generally have strong influence on reflectance in NIR (Tucker 1980, Tucker and Sellers 1986, Hunt and Rock 1989). NIR reflectance is weighted differently between the two indices, NIR having the largest influence on EVI (Justice et al. 1998) and consequently the interannual slope difference in figure 14 can be explained by the different EVI and AVHRR NDVI response to variations in canopy structure.

2582

R. Fensholt and I. Sandholt

Figure 14. Scatterplot of daily in situ measured MODIS NDVI/EVI versus daily in situ measured AVHRR NDVI for grass savanna Dahra North 2001 and 2002 (point-to-point comparison).

Evaluation of MODIS and AVHRR VIs with in situ measurements

2583

The strong linearity indicates that all three indices respond similarly to phenological changes in the vegetation. Differences in MODIS and AVHRR channel 2 bandwidth do not affect the VIs in respect to changing phenology, which is in contradiction to Galva˜o et al. (1999). MODIS NDVI/AVHRR NDVI regression slope differs slightly from the MODIS EVI/AVHRR NDVI regression slope. This corresponds well with theoretical expectations of EVI sensitivity to dense vegetation as laid out in the theoretical section. The MODIS NDVI having higher values than AVHRR NDVI is also expected (offset of 0.07). This can be attributed to the narrow MODIS red band located in the maximum chlorophyll absorption area compared to the AVHRR red band (figure 1), combined with the fact that the MODIS NIR band does not overlap the red-edge as like AVHRR (figure 1). MODIS NDVI is also higher than MODIS EVI, which is a consequence of the weighted inclusion of the blue band in the EVI equation (2). The linearity between MODIS NDVI and AVHRR NDVI implies that the MODIS NDVI does not saturate at high chlorophyll content as a consequence of the narrow MODIS red band compared to the AVHRR broader red band as has been stated by Gitelson and Kaufman (1998). The linearity is a feature of this experiment covering MODIS NDVI values from 0.1 to 0.8, but it is possible that saturation differences between MODIS NDVI and AVHRR NDVI will gain influence above 0.8. The relation between in situ MODIS NDVI and in situ AVHRR NDVI for two different crop types (millet and groundnut) is given in figure 15. The grass, millet and groundnut regression slopes in figures 14 and 15 were tested for significant deviations to illuminate whether the MODIS NDVI/AVHRR NDVI relation is dependent on leaf angle distribution (LAD). The slopes of grass, millet and groundnut were not significantly different, indicating that both MODIS and AVHRR respond identically to variations in LAD. In situ data for millet and groundnut are, however, limited compared to the grass savanna, and more fieldwork is needed to confirm these findings.

Figure 15. Scatterplot of in situ measured MODIS NDVI versus in situ measured AVHRR NDVI for millet and groundnut (point-to-point comparison) for mobile mast 2002.

2584 7.2

R. Fensholt and I. Sandholt Comparing in situ measurements and satellite data (16-day composites)

In situ 16-day mean MODIS NDVI and EVI are plotted against 16-day CV-MVC satellite MODIS NDVI and EVI in figure 16. In situ mean values were chosen over the MVC approach because the latter is designed to compensate for atmospheric influences (clouds and water vapour). Using the MVC approach on in situ measurements would therefore lead to an overestimation of NDVI. In 2001 the R2 values are high for both indices, indicating that the satellite indices capture seasonal variations well. However, the general VI-level of in situ MODIS is higher than the satellite MODIS and the differences are substantially higher in the growing season. General statistics including significance levels for the different relations between in situ measured VIs and satellite derived VIs are shown in table 7. In 2002, where the growing season was much shorter than in 2001, in situ and satellite measured values are more uniform except for one composite value. VI quality flags are added for each composite period and the explanation for each class is given in table 8. For

Figure 16. Scatterplot of 16-day mean in situ measured MODIS NDVI/EVI versus satellite measured CV-MVC MODIS NDVI/EVI (point-to-pixel comparison) for grass savanna Dahra North 2001 and 2002. MODIS VI quality flags are added and explanation is given in table 6.

Year

Sensor

Variable comparison

Site of in situ measurements

Regression

r2 value

S.E. of the estimate

Significance level

2001

MODIS MODIS MODIS MODIS AVHRR MODIS MODIS MODIS MODIS AVHRR MODIS MODIS MODIS MODIS MODIS MODIS

NDVI CV-MVC EVI CV-MVC NDVI MVC EVI MVC NDVI MVC NDVI CV-MVC EVI CV-MVC NDVI MVC EVI MVC NDVI MVC NDVI CV-MVC NDVI CV-MVC NDVI CV-MVC NDVI MVC NDVI MVC NDVI MVC

Dahra North Dahra North Dahra North Dahra North Dahra North Dahra North Dahra North Dahra North Dahra North Dahra North Dahra South Tessekre North Tessekre South Dahra South Tessekre North Tessekre South

y50.68x + 0.02 y50.69x + 0.03 y51.00x 2 0.07 y50.95x 2 0.07 y50.37x + 0.07 y50.57x + 0.09 y50.60x + 0.06 y50.99x + 0.02 y51.05x + 0.02 y50.28x 2 0.01 y51.15x 2 0.02 y51.24x 2 0.11 y50.83x 2 0.04 y51.28x 2 0.02 y51.30x 2 0.11 y50.89x 2 0.03

0.95 0.97 0.92 0.98 0.75 0.59 0.55 0.74 0.78 0.74 0.83 0.94 0.92 0.85 0.98 0.96

0.04493 0.03366 0.05389 0.02849 0.10294 0.08209 0.06747 0.06589 0.07701 0.10401 0.04669 0.02415 0.02860 0.04866 0.01479 0.02186

0.000224 0.000060 0.000561 0.000026 0.026461 0.043088 0.055664 0.013360 0.008155 0.030257 0.001641 0.000074 0.010369 0.001130 0.000003 0.004039

2002

Evaluation of MODIS and AVHRR VIs with in situ measurements

Table 7. Statistics of the relations between in situ measured VIs and satellite derived VIs.

2585

2586

R. Fensholt and I. Sandholt Table 8. Explanation of the MODIS VI quality flags. MODIS VI quality flag 1 2 3 4 5 6 7 8 9

Description High quality Good quality Acceptable quality Fair quality Intermediate quality Below intermediate quality Average quality Below average quality Questionable quality

both years practically all composite values in the growing season are qualitatively sub-optimal due to an extended cloud cover in the growing season and in 2002 the quality of the composite values diverging from the one-to-one line are flagged as questionable. In lieu of the MODIS 16-day view angle constrained MVC product, a normal 16day MVC product was produced from the daily MOD09 data. In figure 17 in situ measurements are plotted against 16-day MVC data for MODIS NDVI and EVI. For MODIS 2001 indices the regression R2 values are still high and highly significant (table 7) but the slopes are much closer to 1, indicating that the MVC MODIS approach better matches in situ measurements than is the case for the CVMVC approach in terms of both NDVI and EVI. In situ measurements are still slightly higher than the satellite level, which is in accordance with the point-to-pixel analysis where a difference of approximately 0.05–0.06 NDVI units was found. R2 values for 2002 MODIS regressions are considerably improved with significance levels increasing from approximately 95% to 99% (table 7) and the slopes are also closer to 1. Unlike the 2001 data there is practically no offset in the regression equations meaning that in situ and satellite VI are the same in 2002. This might be because of the much shorter growing season captured by only two composite periods. Using mean values for the in situ compositing will thereby reduce the VI level compared to 2001 where the longer growing season maintains VI at a high level for more composite periods. As suggested by the point-to-pixel analysis, subtracting 0.06 NDVI units from in situ MODIS vegetation indices in 2001 (as a consequence of the different scale) results in a nearly perfect agreement of both gain and offset between VI measured on the ground and from satellite. The results presented in figures 16 and 17 indicate that using an MVC approach to the MODIS satellite data more precisely reflects in situ measurements than the CV-MVC approach does. This might indicate that the masks (clouds and aerosols) used in the MODIS VI processing mask out useful data in the Sahelian Africa. It can thus be argued that the price for using the CV-MVC in order to compensate for the forward scatter selection (described in the MODIS processing section) in areas with an extended cloud-cover might be too high. It is common for only two images to meet the screening criteria for CV-MVC, and uncritically choosing the scene with the lowest view angle without regard for potentially large NDVI differences might produce an underestimation. Figure 18 compares 16-day MVC AVHRR NDVI and in situ AVHRR NDVI for the two years. For both 2001 and 2002 the R2 values are lower than the

Evaluation of MODIS and AVHRR VIs with in situ measurements

2587

Figure 17. Scatterplot of 16-day mean in situ measured MODIS NDVI/EVI versus satellite measured MVC MODIS NDVI/EVI from daily MODIS data (point-to-pixel comparison) for grass savanna Dahra North 2001 and 2002.

corresponding MODIS analysis in figure 17. The 2001 relation is significant at a 95% level whereas all the MODIS relations are significant at a 99.9% level (table 7). Equally striking is the very low dynamic range in the AVHRR satellite NDVI data. The regression slopes for both years are very far from the one-to-one line and satellite AVHRR NDVI reflects in situ values of AVHRR NDVI rather poorly. In situ VI is generally 2–3 times higher than satellite AVHRR NDVI measurements. Earlier studies of AVHRR performance also revealed considerable difference between in situ and satellite AVHRR NDVI with differences reaching 0.25–0.3 NDVI units (Goward et al. 1991, Prince and Goward 1995). A study by Trischenko et al. (2002) comparing a MODIS NDVI scene to an AVHRR 14 NDVI scene found an average difference of 0.154 NDVI units. The superior performance of MODIS over AVHRR satellite data in approximating in situ measurements can have two possible explanations. The AVHRR NIR response function covers an area of considerable atmospheric water vapour absorption as illustrated in figure 2, which may cause a well-known depression in NDVI values. The Sahelian environment is

2588

R. Fensholt and I. Sandholt

Figure 18. Scatterplot of 16-day mean in situ measured AVHRR NDVI versus satellite measured MVC AVHRR NDVI from daily AVHRR data (point-to-pixel comparison) for grass savanna Dahra North 2001 and 2002.

characterized by high values of water vapour in the rainy season and large day-today fluctuations controlled by the movements of the Intertropical Convergence Zone (ITCZ) but the question is whether atmospheric water vapour alone can cause these differences. The second possible reason for the divergent performance of MODIS NDVI and AVHRR NDVI could be the different overpass time of the sensors. Thermal convection causes considerable cloud build-up during the day, which means that the cloud cover is often much more extensive during the AVHRR sensor’s afternoon pass than during the MODIS sensor’s morning pass. This influences the number of cloud-free scenes in a given composite period in the rainy season and could partly explain the poor AVHRR performance. The different overpass time also causes differences in the solar zenith angles (5–15u for the MODIS data and 15–35u for the AVHRR 16 data) and solar azimuth angles. Both conditions influence the BRDF and variations in BRDF are known to have a large influence on vegetation indices (Huete et al. 1999). For the three sites equipped with two-channel sensors in 2002 (table 3), satellite MODIS NDVI has been compared to in situ MODIS NDVI. Figure 19 is a scatterplot of the 16-day satellite CV-MVC MODIS NDVI versus in situ MODIS NDVI. R2 values are generally high and the quality flags indicate that most of the satellite data are of high quality. In the case of Dahra South, however, the extremely short growing season means only one composite point is available, making a high R2 less informative. At the Tessekre South site, instrument failure (see figure 13, lower right) also left us with only one composite point to represent vegetation coverage. The slopes/offsets vary for the different sites but any interpretation is tenuous due to the limited number of composite points representing high NDVI values. MODIS standard MVC calculated from daily MODIS data from 2002 has also been tested for the three sites and the results are presented in figure 20. R2 values and significance levels (table 7) are slightly improved for all three sites but the differences are small compared to the results from Dahra North (e.g. figure 17). The quality flags in figure 19 also indicate less severe cloud contamination during the growing season than at the Dahra North site.

Evaluation of MODIS and AVHRR VIs with in situ measurements

2589

Figure 19. Scatterplot of 16-day mean in situ measured MODIS NDVI versus satellite measured CV-MVC MODIS NDVI (point-to-pixel comparison). Grass savanna Dahra South, Tessekre North and South 2002.

8.

Conclusions and perspectives

For the semi-arid grass savannas of western Africa, the quality of two new vegetation indices from the MODIS sensor (NDVI and EVI) and conventional AVHRR NDVI have been evaluated against vegetation indices measured in situ by multispectral radiometers designed for wavelengths identical to the relevant MODIS and AVHRR channels. The in situ indices perform as expected with a nearly one-toone MODIS NDVI/AVHRR NDVI relation but with an offset of 0.06–0.1 NDVI units. In situ MODIS EVI is more sensitive to dense vegetation which is reflected by slope values less than 1 when MODIS EVI is plotted against in situ AVHRR NDVI. The in situ MODIS and AVHRR indices respond identically to variations in LAD; this was verified using measurements over two different crop types characterized by different LAD. Point-to-pixel comparisons between in situ measurements and MODIS and AVHRR pixel resolutions have been validated by statistical analysis of Landsat

2590

R. Fensholt and I. Sandholt

Figure 20. Scatterplot of 16-day mean in situ measured MODIS NDVI versus satellite measured MVC MODIS NDVI (point-to-pixel comparison). Grass savanna Dahra South, Tessekre North and South 2002.

ETM + NDVI data. The analysis revealed the grass savanna areas used for the in situ measurements to be very homogeneous, which facilitates and justifies comparing point measurement to MODIS and AVHRR pixel resolutions. A comparison of in situ MODIS indices with the MODIS 16-day constrained view angle MVC product showed that the satellite MODIS NDVI and EVI satisfactorily capture seasonal dynamics (R2 from 0.55 to 0.97). A standard 16-day MVC product was calculated from daily MODIS MOD09 data and found to reflect seasonal dynamics of in situ measurements more accurately (R2 from 0.74 to 0.98); moreover the MVC regressions furthermore approaches a 1 : 1 line with in situ measured values compared to the CV-MVC regressions. This indicates that useful information is masked out in the processing of the 16-day CV-MVC MODIS VIs. Extensive cloud cover is likely to have a greater impact on data retrieval under view angle constraints. The majority of in situ measurements were carried out in 2002 characterized by an extremely short growing season and low vegetation density. The

Evaluation of MODIS and AVHRR VIs with in situ measurements

2591

2002 field campaign should ideally be repeated for a ‘normal’ year for further evaluation. The relation between 16-day AVHRR MVC data and in situ AVHRR data was poor. AVHRR data capture the general seasonal variation in the in situ measurements with an R2 of 0.75 in 2001 and 0.64 in 2002 but the dynamic range of the AVHRR satellite data is very low. VI from in situ measurements is generally 2–3 times higher than the satellite measurements, indicating that the atmosphere has a large influence on the signal due to the broad NIR channel 2. The AVHRR afternoon overpass is also more likely to encounter cloud cover than the MODIS morning overpass, thereby reducing the number of cloud-free pixels in the rainy season. From this evaluation of the MODIS and AVHRR NDVI composites it can be concluded that the MODIS data are more accurate than the AVHRR data in capturing both the seasonal dynamics and the overall level of vegetation intensity. Creating a scientific link between the two data sources will require more research into the reasons for this difference. The influence of atmospheric water vapour must be studied further, for instance using the MODIS water vapour 1 km resolution product. The influence of BRDF on VIs also needs to be better understood and by analysing daily variations of in situ vegetation indices it might be possible to obtain information on the magnitude of the variations caused by BRDF. Acknowledgments This study was funded by the Danish Research Councils, ESA related research, grant no. 9902490. Scientific equipment for fieldwork is sponsored by the Danish Research Agency, Danish Agricultural and Veterinary Research Council grant no. 23-01-0153. The authors would like to thank the MODIS Land Discipline Group for creating and sharing the MODIS LAND data. The Centre de Suivi Ecologique in Dakar is warmly acknowledged for providing logistic support and NOAA AVHRR data. Thank you also to the staff at Institut Se´ne´galais de Recherches Agronomiques (ISRA) in Dahra for support during the fieldwork. Finally Jørn Torp Pedersen, Mette Wolstrup Petersen and Anette Nørgaard deserve thanks for their fieldwork assistance. References ASRAR, G., FUCHS, M., KANEMASU, E.T. and HATFIELD, J.L., 1984, Estimating absorbed photosynthetic radiation and leaf area index from spectral reflectance in wheat. Agronomy Journal, 76, 300–306. ASTER SPECTRAL LIBRARIES, 2003, http://speclib.jpl.nasa.gov (accessed February 2003). BE´GUE´, A., 1993, Leaf area index, intercepted photosynthetically active radiation, and spectral vegetation indices: a sensitivity analysis for regular-clumped canopies. Remote Sensing of Environment, 46, 45–59. BUERMANN, W., DONG, J., ZENG, X., MYNENI, R. and DICKINSON, R., 2001, Evaluation of the utility of satellite-based vegetation leaf area index data for climate simulations. Journal of Climate, 14, 3536–3550. CIHLAR, J. and HOWARTH, J., 1994, Detection and removal of cloud contamination from AVHRR composite images. IEEE Transactions on Geoscience and Remote Sensing, 32, 583–589. CIHLAR, J., MANAK, D. and D’IORIO, M., 1994, Evaluation of compositing algorithms for AVHRR data over land. IEEE Transactions on Geoscience and Remote Sensing, 32, 427–437.

2592

R. Fensholt and I. Sandholt

DEERING, D.W., 1978, Rangeland reflectance characteristics measured by aircraft and spacecraft sensors. PhD dissertation, Texas A&M University, College Station, Texas, USA. DEFRIES, R.S. and TOWNSHEND, J.R.G., 1994, NDVI-derived land cover classifications at a global scale. International Journal of Remote Sensing, 15, 3567–3586. DIALLO, O., DIOUF, A., HANAN, N.P., NDIAYE, A. and PREVOST, Y., 1991, AVHRR monitoring of savannah primary production in Senegal, West Africa. International Journal of Remote Sensing, 12, 1259–1279. DYBKJAER, G., NIELSEN, T.T., RASMUSSEN, K., RASMUSSEN, M.S. and TOURE´, A., 2003, Documentation and evaluation of a NOAA data set covering Senegal. Danish Journal of Geography, 103, 125–133. FENSHOLT, R., 2004, Earth observation of vegetation status in the Sahelian and Sudanian West Africa: comparison of TERRA MODIS and NOAA AVHRR satellite data. International Journal of Remote Sensing, 25, 1641–1659. FLUXNET, 2003, http://public.ornl.gov/fluxnet/modissubsetsites.cfm (accessed February 2003). GALVAO, L.S., VITORELLO, I. and FILHO, R.A., 1999, Effects of band positioning and bandwidth on NDVI measurements of Tropical Savannas. Remote Sensing of Environment, 67, 181–193. GALVAO, L.S., VITORELLO, I. and PIZARRO, M.A., 2000, An adequate band positioning to enhance NDVI contrasts among green vegetation, senescent biomass, and tropical soils. International Journal of Remote Sensing, 21, 1953–1960. GAO, X., 1993, A simple bidirectional-reflectance model applied to a tallgrass canopy. Remote Sensing of Environment, 45, 209–224. GAO, X., HUETE, A.R., NI, W. and MIURA, T., 2000, Optical–biophysical relationships of vegetation spectra without background contamination. Remote Sensing of Environment, 74, 609–620. GITELSON, A. and KAUFMAN, Y., 1998, MODIS NDVI optimization to fit the AVHRR data series—spectral considerations. Remote Sensing of Environment, 66, 343–350. GITELSON, A.A. and MERZLYAK, M.N., 1996, Signature analysis of leaf reflectance spectra: algorithm development for remote sensing of chlorophyll. Journal of Plant Physiology, 148, 495–500. GOEL, N., 1989, Inversion of canopy reflectance models for estimation of biophysical parameters from reflectance data. In Theory and Applications of Optical Remote Sensing, G. Asrar (Ed.), pp. 205–250 (New York: Wiley Interscience). GOWARD, S.N. and HUEMMRICH, K.F., 1992, Vegetation canopy PAR absorptance and the normalized difference vegetation index: an assessment using the SAIL model. Remote Sensing of Environment, 39, 119–140. GOWARD, S.N., MARKHAM, B., DYE, D.G., DULANEY, W. and YANG, J., 1991, Normalized Difference Vegetation Index measurements from the Advanced Very High Resolution Radiometer. Remote Sensing of Environment, 35, 257–277. HANAN, N.P., PREVOST, Y., DIOUF, A. and DIALLO, O., 1991, Assessment of desertification around deep wells in the Sahel using satellite imagery. Journal of Applied Ecology, 28, 173–186. HANSEN, L., 2001, Chips Development team (http://www.geogr.ku.dk/chips/index.htm). HEIDINGER, A.K., CAO, C. and SULLIVAN, J., 2002, Using Moderate Resolution Imaging Spectrometer (MODIS) to calibrate advanced very high resolution radiometer reflectance channels. Journal of Geophysical Research, 107, 4702–4711. HENDRICKSON, B.L. and DURKIN, J.W., 1986, Growing period and drought early warning in Africa using satellite data. International Journal of Remote Sensing, 7, 1583–1608. HIELKEMA, J.U., PRINCE, S.D. and ASTLE, W.L., 1986, Rainfall and vegetation monitoring in the Savanna Zone of the Democratic Republic of Sudan using the NOAA Advanced Very High Resolution Radiometer. International Journal of Remote Sensing, 7, 1499–1513.

Evaluation of MODIS and AVHRR VIs with in situ measurements

2593

HOLBEN, B.N., 1986, Characterization of maximum value composites from temporal AVHRR data. International Journal of Remote Sensing, 7, 1417–1434. HUETE, A.R., 1988, A Soil-Adjusted Vegetation Index (SAVI). Remote Sensing of Environment, 25, 295–309. HUETE, A.R., HUA, G., QI, J., CHEHBOUNI, A. and VAN LEEUWEN, W.J.D., 1992, Normalization of multidirectional red and NIR reflectance with the SAVI. Remote Sensing of Environment, 41, 143–154. HUETE, A.R., JUSTICE, C.O. and VAN LEEUWEN, W.J.D., 1999, MODIS Vegetation Index (MOD 13). Version 3. Algorithm Theoretical Basis Document. http://modis.gsfc.nasa.gov/data/atbd/land_atbd.html. HUETE, A., DIDAN, K., MIURA, T., RODRIGUEZ, E.P., GAO, X. and FERREIRA, L.G., 2002, Overview of the radiometric and biophysical performance of the MODIS vegetation indices. Remote Sensing of Environment, 83, 195–213. HUNT, E.R. and ROCK, B.N., 1989, Detection of changes in leaf water content using near and middle-infrared reflectances. Remote Sensing of Environment, 30, 43–54. JUSTICE, C., VERMOTE, E., TOWNSHEND, J.R.G., DEFRIES, R., ROY, D.P., HALL, D.K., SALOMONSEN, V.V., PRIVETTE, J., RIGGS, G., STRAHLER, A., LUCHT, W., MYNENI, R., KNJAZIHHIN, Y., RUNNING, S., NEMANI, R., WAN, Z., HUETE, A., VAN LEEUWEN, W., WOLFE, R., GIGLIO, L., MULLER, J.-P., LEWIS, P. and BARNSLEY, M., 1998, The Moderate Resolution Imaging Spectroradiometer (MODIS): Land remote sensing for global change research. IEEE Transactions on Geoscience and Remote Sensing, 36, 1228–1249. KAUFMAN, Y.J. and TANRE´, D., 1992, Atmospherically resistant vegetation index (ARVI) for EOS–MODIS. IEEE Transactions on Geoscience and Remote Sensing, 30, 261–270. LANDSAT USER’s HANDBOOK, 2002, http://ltpwww.gsfc.nasa.gov/IAS/handbook/handbook_htmls/chapter11/chapter11.html (accessed February 2002). LEPRIEUR, C., VERSTRAETE, M.M. and PINTY, B., 1994, Evaluation of the performance of various vegetation indices to retrieve vegetation cover from AVHRR data. Remote Sensing Reviews, 10, 265–284. LIND, M. and FENSHOLT, R., 1999, The spatio-temporal relationship between rainfall and vegetation development in Burkina Faso. Danish Journal of Geography, Special Issue, 2, 43–55. LOVELAND, T.R., MERCHANT, J.W., OHLEN, D.O. and BROWN, J.F., 1991, Development of a land-cover characteristics database for the conterminous U.S. Photogrammetric Engineering and Remote Sensing, 57, 1453–1463. MALINGREAU, J.P., TUCKER, C.J. and LAPORTE, N., 1989, AVHRR for monitoring global tropical deforestation. International Journal of Remote Sensing, 10, 855–867. MALO, A.R. and NICHOLSON, S.E., 1990, A study of rainfall and vegetation dynamics in the African Sahel using normalized difference vegetation index. Journal of Arid Environments, 19, 1–24. MODIS Reprojection Tool, distribution page, http://edc.usgs.gov/programs/sddm/modisdist/ index.shtml (accessed 1 June 2003). MYNENI, R.B., TUCKER, C.J., ASRAR, G. and KEELING, C.D., 1998, Interannual variations in satellite-sensed vegetation index data from 1981 to 1991. Journal of Geophysical Research—Atmospheres, 103, 6145–6160. NEMANI, R.R., KEELING, C.D., HASHIMOTO, H., JOLLY, W.M., PIPER, S.C., TUCKER, C.J., MYNENI, R.B. and RUNNING, S.W., 2003, Climate-driven increases in global terrestrial net primary production from 1982 to 1999. Science, 300, 1560–1563. OLSSON, L., RASMUSSEN, M.S., RASMUSSEN, K., WARREN, A. and EKLUNDH, L., Recovery of vegetation in the Sahel—trends, patterns, hypotheses and implications. In prep. PRINCE, S.D., 1991, A model of regional primary production for use with coarse resolution satellite data. International Journal of Remote Sensing, 12, 1313–1330. PRINCE, S.D. and GOWARD, S.N., 1995, Global primary production: a remote sensing approach. Journal of Biogeography, 22, 815–835.

2594

Evaluation of MODIS and AVHRR VIs with in situ measurements

PRINCE, S.D., ENGMAN, E.T., SELLERS, P., KERR, Y.H., GOUTORBE, J.P., LEBEL, T., TINGA, A., BESSEMOULIN, P., BROUWER, J., DOLMAN, A.J., GASH, J.H.C., HOEPFFNER, M., KABAT, P., MONTENY, B., SAID, F., SELLERS, P. and WALLACE, J., 1995, Geographical, biological and remote sensing aspects of the Hydrologic Atmospheric Pilot Experiment in the Sahel (HAPEX-Sahel). Remote Sensing of Environment, 51, 215–234. QI, J., KERR, Y. and CHEHBOUNI, A., 1994, External factor consideration in vegetation index development. Proceedings on Sixth International Symposium on Physical Measurements and Signatures in Remote Sensing, 17–21 January, Val d’Ise`re, France, pp. 723–730. RAO, C.R.N. and CHEN, J., 1995, Inter-satellite calibration linkages for the visible and near-infrared channels of the Advanced Very High Resolution Radiometer on the NOAA-7, -9 and -11 spacecraft. International Journal of Remote Sensing, 16, 1931–1942. RAO, C.R.N. and CHEN, J., 1996, Post-launch calibration of the visible and near-infrared channels of the Advanced Very High Resolution Radiometer on the NOAA-14 spacecraft. International Journal of Remote Sensing, 17, 2743–2747. RIDDER, N., STROOSNIJDER, L., CISSE, A.M. and VAN KELULEN, H., 1983, Productivity of Sahelian rangelands, a study of the soils, the vegetation and the exploitation of the natural resources. PPS Coursebook, vol. 1, Agricultural University, Wageningen, Netherlands. RUIMY, A., SAUGIER, B. and DEDIER, G., 1994, Methodology for the estimation of terrestrial net primary production from remotely sensed data. Journal of Geophysical Research, 99, 5263–5283. RUNNING, S.W. and NEMANI, R.R., 1988, Relating seasonal patterns of the AVHRR vegetation index to simulated photosynthesis and transpiration of forest in different climates. Remote Sensing of Environment, 24, 347–367. SELLERS, P.J., 1985, Canopy reflectance, photosynthesis and transpiration. International Journal of Remote Sensing, 6, 1335–1372. TRISHCHENKO, A.P., CIHLAR, J. and LI, Z., 2002, Effects of spectral response function on surface reflectance and NDVI measured with moderate resolution satellite sensors. Remote Sensing of Environment, 81, 1–18. TUCKER, C.J., 1979, Red and photographic infrared linear combinations for monitoring vegetation. Remote Sensing of Environment, 8, 127–150. TUCKER, C.J., 1980, Remote sensing of leaf water content in the near infrared. Remote Sensing of Environment, 10, 23–32. TUCKER, C.J. and SELLERS, P.J., 1986, Satellite remote sensing of primary productivity. International Journal of Remote Sensing, 7, 1395–1416. TUCKER, C.J., DREGNE, H.E. and NEWCOMB, W.W., 1991, Expansion and contraction of the Saharan desert from 1980 to 1990. Science, 253, 299–301. VAN LEEUWEN, W.J.D., HUETE, A.R. and LAING, T.W., 1999, MODIS vegetation index compositing approach: a prototype with AVHRR data. Remote Sensing of Environment, 69, 264–280. VERMOTE, E.F. and VERMEULEN, A., 1999, Atmospheric correction algorithm: special reflectances (MOD09). Algorithm Theoretical Background Document version 4.0, available on line at http://modarch.gsfc.nasa.gov/MODIS/ATBD/atbd_mod08.pdf. VERMOTE, E.F., EL SALEOUS, N.Z. and JUSTICE, C.O., 2002, Atmospheric correction of MODIS data in the visible to middle infrared: first results. Remote Sensing of Environment, 83, 97–111. WILLIAMS, A.P. and HUNT, E.R. Jr, 2002, Estimation of leafy spurge cover from hyperspectral imagery using mixture tuned matched filtering. Remote Sensing of Environment, 82, 446–456.