This paper investigates how hyperspectral reflectance (between 350 and 1800 nm) can be used to infer ecosystem properties for a semi-arid savanna
grassland in West Africa using a unique in situ-based multi-angular data set
of hemispherical conical reflectance factor (HCRF) measurements.
Relationships between seasonal dynamics in hyperspectral HCRF and ecosystem
properties (biomass, gross primary productivity (GPP), light use efficiency
(LUE), and fraction of photosynthetically active radiation absorbed by
vegetation (FAPAR)) were analysed. HCRF data (
Hyperspectral measurements of the Earth's surface provide relevant information for many ecological applications. An important tool for spatial extrapolation of ecosystem functions is to study how spectral properties are related to in situ measured ecosystem properties. These relationships found the basis for upscaling using Earth observation (EO) data. Continuous in situ measurements of hyperspectral reflectance in combination with ecosystem properties are thereby essential for improving our understanding of the functioning of the observed ecosystems. Strong relationships have, for example, been found between information in the reflectance spectrum and ecosystem properties such as leaf area index (LAI), fraction of photosynthetically active radiation (PAR) absorbed by the vegetation (FAPAR), light use efficiency (LUE), biomass, vegetation primary productivity, vegetation water content, and nitrogen and chlorophyll content (e.g. Thenkabail et al., 2012; Tagesson et al., 2009; Gower et al., 1999; Sjöström et al., 2009; Sims and Gamon, 2003). In situ observations of spectral reflectance are also important for parameterisation and validation of canopy reflectance models, as well as space- and airborne products (Coburn and Peddle, 2006).
Very few sites across the world exist with an instrumental setup designed for
multi-angular continuous hyperspectral measurements. Leuning et al. (2006)
present a system mounted in a 70 m tower above an evergreen eucalyptus
forest in New South Wales, Australia, which measures spectral hemispherical
conical reflectance factors (HCRFs) Different reflectance
terminologies have been used to inform on spectral measurements in the field
by the remote sensing community leading to suggestions to the proper use of
the terminology (Martonchik et al., 2000). All field spectro-radiometers
measure HCRF (hemispherical conical reflectance) if the field of view (FOV)
of the sensor is larger than 3
There are many methods for analysing relationships between hyperspectral reflectance and ecosystem properties, such as multivariate methods, derivative techniques, and radiative transfer modelling (Bowyer and Danson, 2004; Ceccato et al., 2002; Danson et al., 1992; Roberto et al., 2012). Still, due to its simplicity, the combination of reflectance into vegetation indices is the major method for upscaling using EO data. By far, the most commonly applied vegetation indices are those based on band ratios, e.g. the normalised difference vegetation index (NDVI), which is calculated by dividing the difference in the near-infrared (NIR) and red wavelength bands by the sum of the NIR and red bands (Tucker, 1979; Rouse et al., 1974). The NIR radiance is strongly scattered by the air–water interfaces between the cells, whereas red radiance is absorbed by chlorophyll and its accessory pigments (Gates et al., 1965). The normalisation with the sum in the denominator is a means to reduce the effects of solar zenith angle, sensor viewing geometry, and atmospheric errors as well as enhancing the signal of the observed target (e.g. Qi et al., 1994; Inoue et al., 2008).
Wavelength specific spectral reflectance is known to be related to leaf characteristics such as chlorophyll concentration, dry matter content, internal structure parameters, and equivalent water thickness (Ceccato et al., 2002). Hyperspectral reflectance data can be combined into a matrix of normalised difference spectral indices (NDSIs), following the NDVI rationing approach. Correlating the NDSI with ecosystem properties provides a way to gain an improved empirically based understanding of the relationship between information in the reflectance spectrum with ground surface properties (e.g. Inoue et al., 2008). Several studies have analysed relationships between hyperspectral HCRF, NDSI, and ecosystem properties (e.g. Thenkabail et al., 2000; Cho et al., 2007; Psomas et al., 2011; Inoue et al., 2008; Gamon et al., 1992; Feret et al., 2008; Thenkabail et al., 2012). Still, it is extremely important to examine these relationships for different ecosystems across the Earth and investigate their applicability for different environmental conditions and under different effects of biotic and abiotic stresses.
Information about the instrumental setup for the measured environmental variables. HCRF is hemispherical conical reflectance factor, GPP is gross primary productivity, LUE is light use efficiency, and FAPAR is fraction of photosynthetically active radiation absorbed by the vegetation. Min and Max are minimum and maximum values measured, respectively, DW is dry weight, C is carbon, and MJ is megajoules. The year started is the first year with measurements. Time is in UTC. For more information about the instrumental setup, see Tagesson et al. (2015b).
A strong correlation between an NDSI and an ecosystem property does not necessarily indicate that the NDSI is a good indicator of vegetation conditions to be applied to EO systems. Visible, NIR, and shortwave infrared (SWIR) have different sensitivity to variations in solar zenith angles, stand structure, health status of the vegetation, vegetation and soil water content, direct/diffuse radiation ratio, and sensor viewing geometry. The influence of sun–sensor geometry on the reflected signal has been studied using radiative transfer models and airborne (e.g. AirMISR) as well as satellite-based data from instruments such as CHRIS-PROBA, MISR, or POLDER (Huber et al., 2010; Maignan et al., 2004; Javier García-Haro et al., 2006; Jacquemoud et al., 2009; Verhoef and Bach, 2007; Laurent et al., 2011). However, effects of variable sun angles and sensor viewing geometries are not well documented in situ for different plant functional types of natural ecosystems except for some individual controlled experiments (Hilker et al., 2008; Sandmeier et al., 1998; Schopfer et al., 2008). Improved knowledge regarding the influence from sun–sensor variability on different NDSI combinations is thereby essential for validating the applicability of an NDSI for EO upscaling purposes.
The Dahra field site in Senegal, West Africa, was established in 2002 as an in situ research site to improve our knowledge regarding properties of semi-arid savanna ecosystems and their responses to climatic and environmental changes (Tagesson et al., 2015b). A strong focus of this instrumental setup is to gain insight into the relationships between ground surface reflectance and savanna ecosystem properties for EO upscaling purposes. This paper presents a unique in situ data set of seasonal dynamics in hyperspectral HCRF and demonstrates how it can be used to describe the seasonal dynamics in ecosystem properties of semi-arid savanna ecosystems. The objectives are threefold: (i) to quantify the relationship between seasonal dynamics of in situ hyperspectral HCRF between 350 and 1800 nm and ecosystem properties (biomass, gross primary productivity (GPP), LUE, and FAPAR), (ii) to quantify the relationship between NDSI with different wavelength combinations (350 to 1800 nm) and the measured ecosystem properties, and (iii) to analyse and quantify effects of variable sun angles and sensor viewing geometries on different NDSI combinations.
All measurements used for the present study were conducted at the Dahra field
site in the Sahelian ecoclimatic zone north-east of the town Dahra in the
semi-arid central part of Senegal (15
Map and photos of the Dahra field site and measured areas. The map
shows the location of Dahra within the Sahel (orange area).
A range of meteorological variables have been measured from a tower at the
Dahra field site for more than 10 years: air temperature (
The total above-ground green biomass (g m
Net ecosystem exchange of CO
A possible source of error in a comparison between EC-based variables and
spectral HCRF is the difference in footprint/instantaneous field of view
(IFOV) between the sensors. The IFOV of the spectroradiometer setup contains
only soil and herbaceous vegetation. The footprint of the EC tower was
estimated using a model based on measurement height, surface roughness, and
atmospheric stability (Hsieh et al., 2000). The median point of maximum
contribution is at 69 m, and the median 70 % cumulative flux distance is
at 388 m from the tower. The footprint of the EC tower contains semi-arid
savanna grassland with
The daytime NEE was partitioned into GPP and ecosystem respiration using the
Mitscherlich light-response function against PAR
Gaps in GPP less than or equal to 3 days were filled using three different methods: (i) gaps shorter than 2 h were filled using linear interpolation, (ii) daytime gaps were filled by using the light-response function for the 7-day moving windows, and (iii) remaining gaps were filled by using mean diurnal variation 7-day moving windows (Falge et al., 2001). A linear regression model was fitted between daytime GPP and APAR for each 7-day moving window to estimate LUE, where LUE is the slope of the line.
Ground surface HCRF spectra were measured every 15 min between sunrise and
sunset from 15 July 2011 until 31 December 2012 using two FieldSpec 3
spectrometers with fibre optic cables (Table 1) (ASD Inc., Colorado, USA).
The spectroradiometers cover the spectral range from 350 to 1800 nm and have
a FOV of 25
Each sensor measurement starts with an optimisation to adjust the sensitivity of the detectors according to the specific illumination conditions at the time of measurement. The optimisation is followed by a dark-current measurement to account for the noise generated by the thermal electrons within the ASD instruments that flow even when no photons are entering the device. The measurement sequence starts with a full-sky-irradiance measurement, followed by measurements of the seven angles of the land surface and finalised by a second full-sky-irradiance measurement. Thirty scans are averaged to one measurement to improve the signal-to-noise ratio for each measurement (optimisation, dark current, full-sky irradiance, and each of the seven target measurements). The full measurement sequence takes less than 1 min. The two ASD instruments are calibrated against each other before and after each rainy season. Poor-quality measurements caused by unfavourable weather conditions, changing illumination conditions, and irregular technical issues were filtered by comparing full-sky solar irradiance before and after the target measurements (Huber et al., 2014). The spectral HCRF was derived by estimating the ratio between the ground surface radiance and full-sky irradiance. For a complete description/illustration of the spectroradiometer setup, the measurement sequence, and the quality control, see Huber et al. (2014).
NDSI using all possible combinations of two separate wavelengths were
calculated as
The effects of variable solar zenith angles on different NDSI combinations were studied with nadir HCRF measurements. In order to capture the seasonal dynamics, data were taken over 15 days during four periods: (1) the dry season in 2012 (day of year (DOY) 71–85), (2) the fast growth period in 2011 (start of the rainy season) (DOY 200–214), (3) the peak of the growing season in 2011 (DOY 237–251), and (4) the senescent period in 2011 (the end of the rainy season) (DOY 278–293). Only days with full data coverage were used in order not to include bias in the results from days with incomplete data sets. The median HCRF of the 15 days was calculated for each wavelength for every 15 min between 08:00 and 18:00 (UTC). These HCRF values were combined into NDSI with different wavelength combinations. Finally, daily mean and standard deviation for all wavelength combinations were calculated. Diurnal variability in the NDSI was assessed with the coefficient of variation (COV), which is the ratio between the standard deviation and the mean. The COV gives an indication of effects related to variable solar zenith angles.
To capture directional effects in the NDSI related to the variable view
zenith angles (15, 30, and 45
We examined the relationship between predictor variables (daily median
hyperspectral HCRF, and NDSI from nadir observations) and response variables
(biomass, GPP, LUE, and FAPAR). A comparison between fitted linear and
exponential regression models indicated no improvement by fitting exponential
regression models; we hence choose to use linear regression analysis
(Supplement). Possible errors (random sampling errors, aerosols, dust or
water on the sensor heads, electrical senor noise, filtering and gap-filling
errors, errors in correction factors, sensor drift, and instrumentation
errors) can be present in predictor and response variables. We thereby used a
reduced major axis linear regression to account for errors in both the
predictor and response variables when fitting the regression lines. In order
to estimate the robustness of the empirical relationships, we used a
bootstrap simulation methodology, where the data sets were copied 200 times
(Richter et al., 2012). The runs generated 200 sets of slopes, intercepts,
and coefficients of determination (
A filter was created for the analysis between NDSI and ecosystem properties; all NDSI combinations with a COV higher than 0.066 in any of the four periods (dry season, fast growth period, peak of the growing season, and senescent period) and all NDSI combinations with ANIF values higher than 1.2 and lower than 0.8 in any of the four periods were filtered. The ANIF thresholds of 1.2 and 0.8 and the COV threshold of 0.066 were used since values then vary less than 20 % due to effects of variable sun–sensor geometry. NDSI including the water absorption band (1300–1500 nm) was also removed as it is strongly sensitive to atmospheric water content and is less suitable for spatial extrapolation of ecosystem properties using air/spaceborne sensors (Asner, 1998). Finally, NDSI combinations including wavelengths between 350 and 390 nm were removed owing to low signal-to-noise ratio in the ASD sensors (Thenkabail et al., 2004).
Time series of the measured variables:
The coefficient of variation (COV), i.e. the ratio between daily
standard deviation and the daily mean (measurements taken between 08:00 and
18:00 (UTC)), for different normalised difference spectral index (NDSI)
wavelength (
Daily average air temperature at 2 m height ranged between 18.4 and
37.8
The range in HCRF is large across the spectral space, and would hide the seasonal dynamics in hyperspectral HCRF if directly shown. Therefore, to clearly illustrate these seasonal dynamics, the ratio between daily median nadir HCRF and the average HCRF for the entire measurement period was calculated for each wavelength (350–1800 nm). This gives a fraction of how the HCRF for each wavelength varies over the measurement period in relation to the average of the entire period (Fig. 2d). In the visible (VIS) part of the spectrum (350–700 nm) there was a stronger absorption during the second half of the rainy season and at the beginning of the dry season than during the main part of the dry season and the start of the rainy season. There was stronger NIR absorption (700–1300 nm) at the end of the rainy season and the beginning of the dry season, whereas the absorption decreased along with the dry season. Strong seasonal variation was observed in the water absorption region around 1400 nm following the succession of rainy and dry seasons. HCRF in the SWIR (1400–1800 nm) generally followed the seasonal dynamics of the visible part of the spectrum.
The strongest effects of solar zenith angles and variable viewing geometry on
NDSI were observed at the peak of the growing season 2011 (Figs. 3 and 4, and
S1–S5 in the Supplement). In the main section of the paper, we hence choose to
present the figures from this period; figures from remaining periods are
located in the Supplement. The most pronounced effects of solar
zenith angles were observed for NDSI combining SWIR and NIR wavelengths, and
with VIS wavelengths between 550 and 700 nm (
The anisotropy factor (ANIF) for different normalised difference
spectral index (NDSI) wavelength (
Median correlation coefficient (
HCRF values for all wavelengths except the water absorption band at 1100 nm
were strongly correlated with biomass (Fig. 5a). The strongest correlation was
found at
Wavelengths of the hemispherical conical reflectance factors (HCRF)
(
The relationship between GPP and nadir-measured hyperspectral HCRF is
inverted as compared to other correlation coefficient lines (Fig. 5b), since
GPP is defined as a withdrawal of CO
Coefficient of determination (
The least-squares linear regressions with the strongest relationships
between the normalised difference spectral index (NDSI) and
LUE was negatively correlated with HCRF in the blue and red spectral ranges
and in the water absorption band at 1100 nm, and it was positively correlated
in the NIR wavelengths (Fig. 5c). HCRF at 761 nm yielded the strongest
positive correlation (
FAPAR was negatively correlated with nadir-measured HCRF for most wavelengths
(Fig. 5d); the higher the FAPAR, the higher the absorption, and thereby the lower
the HCRF. The strongest correlation was found at a blue wavelength
Effects of solar zenith angles and sensor viewing geometry were similar
(Figs. 3 and 4), since they affect HCRF measurements in a similar way (Kimes,
1983). In dense and erectophile canopies, HCRF increases with sensor viewing
and solar zenith angles, because a larger fraction of the upper vegetation
canopy is viewed/illuminated, whereas the shadowed lower part of the canopy
contributes less to the measured signal as shown previously by several
studies (Huete et al., 1992, 2014; Jin et al., 2002; Kimes, 1983). However,
the radiative transfer within a green canopy is complex, and differs across
the spectral region (Huber et al., 2014). Less radiation is available for
scattering of high-absorbance spectral ranges (such as the VIS wavelengths),
and this tends to increase the contrast between shadowed and illuminated
areas for these wavelengths, whereas in the NIR and SWIR ranges, more
radiation is scattered and transmitted, which thereby decreases the
difference between shadowed and illuminated areas within the canopy (Kimes,
1983; Hapke et al., 1996). A recognised advantage of NDSI calculations is
that errors/biases that are similar in both wavelengths included in the index
are suppressed by the normalisation. However, for a given situation where
errors/biases are different for the wavelengths used, such as effects
generated by sun–sensor geometry, errors/biases will affect the value of
the index. This was also the case at the Dahra field site: NDSI values were
strongly affected at wavelength combinations with large differences in
effects of variable solar zenith angles (Fig. 6 in Huber et al., 2014) and
variable view zenith angles (Fig. 4 in Tagesson et al., 2015b). This effect
is especially pronounced in the case of low index values (closer to 0),
whereas larger index values (closer to 1 and
A strong diurnal dynamic does not necessarily mean a poor NDSI. For example, the photochemical reflectance index (PRI) was created for assessing diurnal dynamics in the xanthophyll cycle activity (Gamon et al., 1992). Stomatal closure due to high temperatures could also influence diurnal dynamics of vegetation properties (Lasslop et al., 2010), and hence the diurnal dynamics of NDSI. However, diurnal variation in reflectance caused by diurnal variability in vegetation status is assumed minor in relation to the diurnal variability caused by changes in solar zenith angles. Additionally, in our study we are interested in relationships in seasonal dynamics between ecosystem properties and NDSI; diurnal variation can thereby interfere and introduce uncertainty into such relationships.
The importance of directional effects for the applicability of normalised difference spectral indices has been pointed out as an issue in numerous papers (e.g. Holben and Fraser, 1984; van Leeuwen et al., 1999; Cihlar et al., 1994; Fensholt et al., 2010b; Gao et al., 2002). This study confirms these challenges for NIR/SWIR-based indices, but the results also indicate several wavelength combinations from which these effects are less severe and potentially applicable to EO data without disturbance from viewing/illumination geometry for this type of vegetation. Multi-angular HCRF data provide additional information of, for example, canopy structure, photosynthetic efficiency, and capacity (Bicheron and Leroy, 2000; Asner, 1998; Pisek et al., 2013; Huber et al., 2010), and this unique in situ-based, multi-angular, high-temporal-resolution data set may thus be used for future research of canopy radiative transfer and BRDF (bidirectional reflectance distribution function) modelling (Jacquemoud et al., 2009; Bicheron and Leroy, 2000). The multi-angular data set is also highly valuable for evaluation and validation of satellite-based products, where the separation of view angle and atmospheric effects can only be done using radiative transfer models (Holben and Fraser, 1984).
The strong correlation between biomass and most of the spectrum indicates the strong effects of phenology on the seasonal dynamics in the HCRF spectra (Fig. 5a). Variability in VIS (350–700 nm) HCRF for vegetated areas is strongly related to changes in leaf pigments (Asner, 1998), and this can also be seen in Fig. 2d since absorption was much stronger during the rainy (growing) season than during the dry season. Previous studies have generally shown positive relationships between NIR HCRF and biomass since a large fraction of NIR radiation is reflected in green healthy vegetation to avoid overheating (e.g. Hansen and Schjoerring, 2003; Asner, 1998). Here, a strong negative relationship between NIR HCRF and dry weight biomass is generally observed (Fig. 5a), indicating stronger NIR absorption with increased biomass. However, a strong positive NIR HCRF correlation with vegetation water content was seen (figure not shown). A possible explanation could be that the sampled biomass at the end of the rainy season contained some senescent vegetation, and a correlation against vegetation water content is hence closer to green healthy vegetation. This relationship is, however, interesting and should be studied further to better understand the respective importance of canopy water and leaf internal cellular structure for the NIR HCRF of herbaceous vegetation characterised by erectophile leaf angle distribution in semi-arid regions. We found the strongest correlation for biomass with a SWIR wavelength, thereby confirming the studies by Lee (2004) and Psomas et al. (2011) in that SWIR wavelengths are good predictors of LAI or biomass.
The NDVI is known to saturate at high biomass because the absorption of red
light at
The maximum absorption in the red wavelengths generally occurs at 682 nm as
this is the peak absorption for chlorophyll
Both LUE and GPP were most strongly correlated with HCRF at 761 nm, which is
the oxygen A band within the NIR wavelengths. HCRF at 761 nm is commonly
used for estimating solar-induced chlorophyll fluorescence due to radiation
emitted by the chlorophyll, and it has been suggested as a direct measure of
health status of the vegetation (Meroni et al., 2009). Earth observation data
for estimating fluorescence should have very high spectral resolution
(
The strongest wavelength combinations for estimating LUE for this semi-arid
ecosystem was NDSI[688, 435]. The 688 nm wavelength is just at the base of
the red-edge region, again indicating the importance of this spectral region
for estimating photosynthetic activity. The wavelength at 435 nm is at the
centre of the blue range characterised by chlorophyll utilisation, and
strongly related to chlorophyll
FAPAR is an estimate of radiation absorption in the photosynthetically active
spectrum and thereby strongly negatively correlated with most parts of the
spectrum (Fig. 5d). FAPAR remained high during the dry season because of
standing dry biomass that was slowly degrading over the dry season (Fig. 2g).
The seasonal dynamics in FAPAR is thereby strongly related to senescence of
the vegetation, which explains why FAPAR was most strongly correlated with blue
wavelengths (
Very limited multi-angular hyperspectral in situ data exist, even though they have been, and will continue to be, extremely valuable for an improved understanding of the interaction between ground surface properties and radiative transfer. In this study, we have presented a unique in situ data set of multi-angular, high temporal resolution hyperspectral HCRF (350–1800 nm) and demonstrated the applicability of hyperspectral data for estimating ground surface properties of semi-arid savanna ecosystems using NDSI. The study was conducted in spatially homogeneous savanna grassland, suggesting that the results should be commonly applicable for this biome type. However, attention should be paid to site-specific details that could affect the indices, such as species composition, soil type, biotic and abiotic stresses, and stand structure. Additionally, the biophysical mechanisms behind the NDSIs are not well understood at the moment, and further studies are needed to examine the applicability of these indices to larger regions and other ecosystems. Due to it being a two-band ratio approach, NDSI does not take full advantage of exploring the rich information given by the hyperspectral HCRF measurements. In the future, this hyperspectral HCRF data set could be fully explored using, for example, derivative techniques; multivariate methods; and creation, parameterisation, and evaluation of BRDF and radiative transfer models.
Even though several other methods exists which fully exploit the information in the hyperspectral spectrum, results of the present study still indicate the strength of normalised difference indices for extrapolating seasonal dynamics in properties of savanna ecosystems. A number of wavelength spectra that are highly correlated with seasonal dynamics in properties of semi-arid savanna ecosystems have been identified. The relationships between NDSI and ecosystem properties were better determined than, or at the same level as, results of previous studies exploring relationships between hyperspectral reflectance and ecosystem properties (Kumar, 2007; Cho et al., 2007; Mutanga and Skidmore, 2004; Psomas et al., 2011; Ide et al., 2010). By also studying the impact from varying viewing and illumination geometry, the feasibility and applicability of using indices for upscaling to EO data were evaluated. As such, the results presented here offer insights for assessment of ecosystem properties using EO data, and this information could be used for designing future sensors for observation of ecosystem properties of the Earth.
This paper was written within the frame of the project entitled Earth
Observation based Vegetation productivity and Land Degradation Trends in
Global Drylands. The project was funded by the Danish Council for
Independent Research (DFF) Sapere Aude programme. The site is maintained by
the Centre de Recherches Zootechniques de Dahra, Institut Sénégalais
de Recherches Agricoles (ISRA).
Edited by: M. Rossini