Resolving the spatial and temporal dynamics of gross primary productivity (GPP) of terrestrial ecosystems across different scales remains a challenge. Remote sensing is regarded as the solution to upscale point observations conducted at the ecosystem level, using the eddy covariance (EC) technique, to the landscape and global levels. In addition to traditional vegetation indices, the photochemical reflectance index (PRI) and the emission of solar-induced chlorophyll fluorescence (SIF), now measurable from space, provide a new range of opportunities to monitor the global carbon cycle using remote sensing. However, the scale mismatch between EC observations and the much coarser satellite-derived data complicate the integration of the two sources of data. The solution is to establish a network of in situ spectral measurements that can act as a bridge between EC measurements and remote-sensing data. In situ spectral measurements have already been conducted for many years at EC sites, but using variable instrumentation, setups, and measurement standards. In Europe in particular, in situ spectral measurements remain highly heterogeneous. The goal of EUROSPEC Cost Action ES0930 was to promote the development of common measuring protocols and new instruments towards establishing best practices and standardization of these measurements. In this review we describe the background and main tradeoffs of in situ spectral measurements, review the main results of EUROSPEC Cost Action, and discuss the future challenges and opportunities of in situ spectral measurements for improved estimation of local and global estimates of GPP over terrestrial ecosystems.
Accurate quantification of carbon fluxes across space and time is of primary
importance to climate scientists, land use managers, and policymakers
(Beer et al., 2010; Ciais et al., 2014; Joos et al., 2001). Carbon budgets can be estimated with high
accuracy at the ecosystem level (e.g. Clement et al., 2012;
Grace et al., 2006; Zanotelli et al., 2015), but global estimates of gross
primary productivity (GPP) and carbon balance in terrestrial ecosystems
still have high levels of uncertainty (Alton,
2013; Balzarolo et al., 2014; Beer et al., 2010; Enting et al., 2012; Jung
et al., 2011; Keenan et al., 2012; Piao et al., 2013). The primary method
used to measure the net flux of carbon dioxide (CO
Given that most of the factors affecting carbon fluxes have strong spatial
and temporal components it is difficult to envisage upscaling without the
use of remote-sensing data, the only means to provide regular and spatially
continuous observations of the Earth surface. One of the most widely applied
approaches to assimilate remotely sensed data is to estimate GPP through a
light use efficiency (LUE) model (Monteith, 1972; Reichstein et al., 2014; Ruimy et al., 1994):
In ecosystems dominated by evergreen species, the seasonal variation in GPP
can be strongly controlled by LUE in addition to, or instead of, fAPAR (e.g.
Garbulsky et al., 2008; Gamon, 2015). The LUE term is usually estimated as the product of the potential
maximum LUE (
Importantly, LUE generates optical signatures that can be measured with
optical remote-sensing instruments mounted on airborne or satellite
platforms. These signatures are the photochemical reflectance index (PRI),
and the emission of solar-induced chlorophyll
Noting that the relationship between NDVI and fAPAR tends to saturate at high canopy densities (Myneni and Williams, 1994; Olofsson and Eklundh, 2007), other approaches have also been used to estimate vegetation carbon uptake. For example, the Enhanced Vegetation Index (EVI) (Huete et al., 2002) efficiently describes the seasonal variability in GPP across both dense and sparse vegetation canopies (Schubert et al., 2010, 2012; Sims et al., 2006; Sjöström et al., 2011; Xiao et al., 2004a, b, 2010). More recently, the plant phenology index (PPI) (Jin and Eklundh, 2014) has been shown to be linearly related to green leaf area index (LAI), and better related to seasonal GPP variations than NDVI and EVI of coarse-resolution MODIS data at northern latitudes. This illustrates the value of investigating the relationship between carbon uptake and spectral information in flux footprint areas beyond the LUE model depicted in Eq. (1).
Integrating satellite and EC data into large-scale carbon models is not straightforward. The spatial mismatch between EC measurements and coarser grid-cell information in models and most satellite-derived remote-sensing data adds significant uncertainty (Chen et al., 2012; Oren et al., 2006). The large viewing angle of many satellite products, e.g. MODIS, results in ill-defined and variable footprint areas leading to additional geometric uncertainties (e.g. Tan et al., 2006). Furthermore, airborne and space-borne data need to be corrected for atmospheric absorption and scattering effects (Karpouzli et al., 2003; Richter, 2011), a process that again can add further uncertainty (Drolet et al., 2005; Hilker et al., 2009). All these physical limitations could be substantially reduced by including in situ long-term spectral measurements to the network of EC flux sites (Gamon et al., 2010; Hilker et al., 2009).
Coordinated in situ spectral measurements require a network of stable sensors that
follow the same measurement standards, calibration protocols, and have
traceable technical specifications to allow across-site comparisons.
Following the example of the EC Fluxnet community
(
The goal of EUROSPEC COST was to promote the development of common measuring protocols and new instruments for in situ spectral measurements, bringing together scientists and industries in order to increase the reliability, value and cost-efficiency of such measurements. This was done so that field-installed spectral sensors could be used as a “bridge” between the EC and optical remote-sensing communities.
The action was divided in four working groups (WG): WG1, network and state-of-the-art characterization. The goal was to characterize the variability of spectral measurements and methods being used across flux sites in Europe; WG2, intercomparison and standardization of instruments. The goal was to characterize the sources of variability between sensors, methods and protocols; WG3, new instruments. The goal was to promote the development of new instruments that better match sensor design, specifications, cost and purpose. And WG4, upscaling methods. The goal was to evaluate challenges and tools to upscale point observations to the footprint area and beyond.
The main objective of this review is to contextualize and synthesize the accomplishments made during EUROSPEC and to identify a number of challenges and opportunities for the near future. We describe the background of in situ long-term spectral measurements and their main tradeoffs, followed by presenting the main results of each EUROSPEC WG and by a final discussion on future challenges and opportunities of these measurements.
Remote-sensing measurements can be collected from platforms that may operate at variable distance from the Earth's surface: from satellites for regional–global extent measurements, to field spectrometers mounted on top of towers for close-range in situ measurements. In between these two scales are airborne platforms including piloted and unpiloted aircraft, kites and blimps that can measure at multiple scales depending on height. In EUROSPEC we focused on long-term in situ optical measurements conducted from EC towers.
There are a number of important differences between close-range in situ measurements and the traditional remote sensing from aircraft or satellites. In situ measurements, sometimes referred to as proximal sensing, are conducted at short distances and are to a large extent free from the atmospheric absorption and scattering effects that affect traditional remotely sensed data (Cheng et al., 2006; Meroni et al., 2009; Thenkabail et al., 2002). In situ measurements can be used to track the spectral properties of individual biological elements (leaves, shoots, plants, homogeneous canopies) while traditional remote sensing tends to measure at coarser scales where multiple species, soil and non-vegetated areas may contribute to the measured signals. Most importantly, in situ measurements can provide data at high-temporal resolution, something that cannot be accomplished with traditional remote sensing. All these characteristics make in situ measurements ideal to study and disentangle the link between optical signals and carbon flux dynamics, as well as for calibrating and validating satellite data and atmospheric correction algorithms (Brook and Ben-Dor, 2015; Czapla-Myers et al., 2015; Hilker et al., 2009).
In situ spectral measurements involve the measurement of the down-welling (incoming) and up-welling (both reflected and emitted) radiation fluxes from the Earth surface. These measurements can be conducted with variable setup and approaches, and the optimal solution will depend on the purpose, characteristics of the site and amount of resources available.
Measurements of down-welling and up-welling radiation can be carried out either in sequence (when a single sensor or spectrometer is used), or simultaneously (when two separate sensors or spectrometers are used) (Fig. 1). These are also addressed as single beam/field-of-view (SFOV) or dual beam/field of view (DFOV) configurations, respectively, and have their own advantages and disadvantages (see Table 1). A SFOV system is generally configured with a single sensor (or spectroradiometer) and will be generally cheaper to set up than a DFOV using two sensors. Having a single sensor means also that there is no need to inter-calibrate the sensor pair. However, long-term and unattended measurements with an SFOV system face the challenge of automating a single sensor/spectrometer to measure both down-welling and up-welling radiation. This automation usually involves moving parts (e.g Meroni et al., 2011; Sakowska et al., 2015), which may become a problem for long-term field operation under certain environments and entail a time delay between up-welling and down-welling measurements, which in turn may generate noisy data under cloudy conditions. Similarly, DFOV systems also have associated advantages and disadvantages. Because radiometric measurements are temperature-sensitive (Saber et al., 2011), DFOVs based on two spectrometers are particularly sensitive to temperature. A practical solution is to keep the two sensors at constant temperature, e.g. by housing them in a temperature-controlled enclosure (e.g. Drolet et al., 2014), but this might not be always possible due to power limitations. Also, regular intercalibration of the two sensors will be essential in DFOV measurements (see e.g. Anderson et al., 2006; Gamon et al., 2015; Jin and Eklundh, 2015). Additionally, long-term measurements with a DFOV may be constrained by aging-dependent degradation of the two sensor heads (see Sect. 3.2.4). These limitations were partly overcome with new DFOV systems developed during EUROSPEC that include a single spectrometer (see Sect. 3.3). The advantage of a DFOV system is that it guarantees quasi-simultaneous measurements of down-welling and up-welling radiation, each within a few hundred milliseconds of the other and may be easier to automate because it does not require moving parts to shift from up-welling to down-welling measurements. Importantly, systems such as the Piccolo and SIF-Sys (developed during EUROSPEC) share the benefits from both SFOV and DFOV systems as they include a single spectrometer but make use of bifurcated fibre optics to sample two fields of view (Fig. 1).
Main instrument configurations adopted for in situ spectral measurements.
Advantages and disadvantages of different approaches and configurations for in situ spectral measurements.
Spectral information can be acquired at different spectral resolution, which depends on the sampling intervals (discrete or continuous) and the width of the spectral bands (Fig. 2). Accordingly, spectral measurements can be classified into multispectral or hyperspectral. Multispectral sensors measure a limited number of spectral bands, from two to five bands found for example in the SKYE SKR-1800 or 1860 sensor series (Skye Instruments Ltd, UK) the Decagon-SRS series (Decagon devices Inc, WA, USA) or the Cimel five-band sensors (Cimel Electronique, FR), or up to 16 bands found in the Cropscan MSR16R (Cropscan Inc., MN, USA) (Balzarolo et al., 2011; Sakowska et al., 2014). The bandwidth of these sensors (in terms of full width at half the maximum response, FWHM) is at the order of 10 nm or greater, and the sampling across a specific spectral range is typically discrete (Fig. 2). These sensors are typically manufactured using optical filters, light emitting diodes (LEDs) and photodiode detectors (Norton, 2010; Ryu et al., 2010). These sensors are characterized by relatively low cost (from a few hundred to a few thousand euros/dollars), ease of maintenance, weather-proof design, and low power consumption. Hence, they are useful and affordable instruments for deployment at flux tower sites for extended periods of time. In addition, their relatively low cost allows mounting of several sensors in different positions to study spatial heterogeneity. Multispectral sensors can be deployed to measure a number of vegetation indices (e.g. NDVI or PRI) to track and study vegetation phenology and seasonality. They can also be used to produce satellite calibration and validation data, provided that their spectral configuration can be related to that of the spaceborne sensor.
Different sampling strategies: discrete vs. continuous. Spectral resolution denoted by bandwidth and sampling intervals. The narrower the bandwidth and the shorter the interval between bands the higher will be the resolution at which spectral features can be resolved. Note that sensor responses are here represented as Gaussian just for simplicity reasons, but typical sensor responses may present different shapes.
In contrast, hyperspectral sensors (more often addressed as spectrometers, or spectroradiometers when they are radiometrically calibrated) can measure hundreds of spectral bands, often 250 or more, with bandwidths usually less than 10 nm full width at half maximum (FWHM) and sampling intervals from less than 1 to 10 nm depending on configuration (Fig. 2). The obvious advantage of hyperspectral sensors is that they can resolve more detailed features of the vegetation (Milton et al., 2009) and serve to estimate parameters that require higher spectral resolution, such as the emission of chlorophyll fluorescence (Meroni et al., 2009). Moreover, since hyperspectral information can be resampled to coarser spectral resolutions, data from hyperspectral systems can be flexibly convoluted to match spectral bands of different remote sensors (Olsson et al., 2011) increasing its value as a source of satellite calibration and validation data. In addition, hyperspectral data can be used to mine new spectral band combinations to match different ecosystem variables (e.g. Balzarolo et al., 2015; Heiskanen et al., 2013; Inoue et al., 2008; le Maire et al., 2008; Milton et al., 2009; Tagesson et al., 2015; Wang et al., 2011; Yao et al., 2010). Milton and coworkers (2009) presented an extensive review of how hyperspectral proximal sensing, or field spectroscopy, has developed and listed the most commonly available field spectrometers. Spectrometers are very complex opto-electro-mechanical instruments and tend to be expensive, from a few thousand euros/dollars for optical benches measuring in the VNIR, to several tenths of thousands for field instruments measuring in the VNIR and the SWIR. The common limitation of all these spectrometers is that they are not designed for unattended or long-term field operation. Accordingly, users need to build their own weatherproof housing, power supply, automatic datalogging, and control units (see next section). As a result, the overall cost of these user-made systems is difficult to quantify because in addition to off-the-shelf components they involve plenty of in-house skilled technician hours. Field spectrometers are also more susceptible to physical damage (due to their inherent complexity), and are more difficult and expensive to automate for continuous or periodic logging applications. In addition, these systems tend to be considerably larger and heavier than their multispectral counterparts, presenting a structural challenge to their deployment on flux towers. Despite these limitations, the number of such measurements is rapidly increasing (Drolet et al., 2014; Huber et al., 2014; Pacheco-Labrador and Martín, 2015; Rossini et al., 2012; Sakowska et al., 2015).
Reflectance factors relate the radiant flux reflected by a target surface to the radiant flux incident on it, and they can be measured using different instrument configurations (see Schaepman-Strub et al. (2006) for a full mathematical explanation of the different factors and terms). Three main instrument configurations have been applied to in situ field measurements to quantify incoming and reflected radiation and estimate reflectance factors: bi-conical, hemispherical-conical and the bi-hemispherical configurations (Fig. 1).
Hemispherical-conical measurements use a foreoptic diffuser assembly, designed to have a cosine response at changing solar zenith angle to estimate down-welling irradiance, and a conical foreoptic for upwelling measurements which can be installed at nadir or off-nadir. The hemispherical-conical configuration lends itself to both multispectral and hyperspectral measurements from flux towers.
Bi-conical measurements rely on a diffuse white reference panel, typically of Spectralon® (Labsphere Inc., NH, USA), reflecting down-welling solar radiant flux, normally viewed from nadir through a fixed angularly limited (conical) field-of-view foreoptic, to provide the reference measurement. The potential limitation of using a reference panel is that it needs to be kept clean and stable over time which may become a challenge in the field due to particle deposition (but see Sakowska et al. (2015) in Sect. 3). In practice, because both direct and diffuse light contribute to the up-welling signal reflected from the reference panel when measuring under field conditions, field data obtained in a bi-conical instrument configuration can be used to derive hemispherical-conical reflectance factors (HCRF) (Schaepman-Strub et al., 2006).
Bi-hemispherical measurements use a foreoptic diffuser to capture both down-welling and up-welling irradiance. Bi-hemispherical measurements require a nadir-view installation and have the great advantage of enabling the sampling of a wider area. The main limitation of this configuration is that while the hemispherical-conical measurements can be taken by observing the canopy at nadir or off nadir, all viewing directions (both nadir and off-nadir) contribute to the bi-hemispherical measurements (Meroni et al., 2011). For this reason, bi-hemispherical measurements tend to be more sensitive to variations in illumination geometry compared to hemispherical-conical measurements collected with a nadir view particularly for large illumination zenith angles.
When research questions become global, as in the case of global carbon cycle monitoring, networking becomes a key methodological element to ensure consistent implementation of measuring protocols, data sharing and management. One of the objectives of EUROSPEC was to contribute to build up a distributed European spectral sampling network to foster data sharing in order to better understand relationships between optical responses of vegetation and carbon cycle.
In situ long-term spectral measurements at flux towers are still accomplished with
instruments that are of variable design and performance, use different
configurations, are installed with contrasting geometries, and are conducted
with different calibration or quality assessment regimes. In EUROSPEC, we
conducted a detailed review of proximal sensing measurements based on the
responses to a questionnaire obtained from groups working at 40 flux tower
sites in Europe including two sites from Africa and Australia
(Balzarolo et al., 2011). In situ measurements included
SFOV and DFOV systems, bi-hemispherical or hemispherical-conical
configurations, and both multispectral and hyperspectral sensors. The study
portrayed a lack of consensus on what are the most suitable proximal
sensing systems and methods to support EC measurements. No standards were
being applied in terms of system performance (e.g. non-linearity in
response, signal-to-noise ratios, and cosine response of down-welling
radiant flux foreoptics); measurement geometries (e.g. hemispherical vs.
conical and their combinations); different foreoptic field-of-views;
installation geometry (e.g. nadir or off-nadir; height of sensor above
target surface), or calibration regimes (e.g. regularly calibrated by
manufacturer, calibrated in situ or even not regularly calibrated)
(Balzarolo et al., 2011). The lack of regular
calibrations was presented as a fundamental limitation to overcome in order
to produce high-quality data, reliability of time series analysis and to
enable inter-comparison of results between network sites, e.g. Integrated
Carbon Observation System (ICOS) sites (
Balzarolo et al. (2011) suggested also that two possible levels of instrumentation could be considered. The first, termed the Basic Standard, would include only multispectral broadband sensors to estimate selected vegetation indices. The second, termed Advanced Standard, would augment these multispectral sensors sites with hyperspectral sensor systems. The question remains as to what specific instruments and sensors would be more appropriate in each case. Anderson et al. (2013) conducted a field intercomparison experiment to assess the reproducibility of measurements collected by different sensors used at flux tower sites. The analysis showed that lower-cost spectroradiometer systems performed similarly to more costly models and suggested that cost-effective and accurate measurements in the PAR range can also be acquired using lower-cost instrumentation. Similar conclusions were obtained by Harris et al. (2014) when they compared the performance of lower-cost multispectral sensors with a reference spectroradiometer to estimate the photochemical reflectance index (PRI). Another conclusion of these studies was the importance of characterizing sensor properties to allow inter-comparison of results between sensors and sites (see next section). Further long-term field instrument intercomparisons will be needed before final conclusions can be drawn from these studies.
EUROSPEC managed to establish an active network including scientists from 28 countries from Europe and beyond. This network remains active under a new COST Action (OPTIMISE-ES1309) and under the umbrella of SpecNet. Together we hope to continue promoting the standardization and implementation of optical measurements across flux sites.
Limited consideration had been given to the comparability of spectral measurement protocols and systems before EUROSPEC (but see Castro-Esau et al., 2006; Pfitzner et al., 2011). Because the use of the same type of sensor in all sites is neither realistic, and perhaps nor desirable as it would undermine the development of new sensors, we need to examine the factors that influence the variability in the data collected with different sensors. In EUROSPEC we dealt with a number of these factors: linearity of spectrometer response, impact of cosine diffusers and reference panels, effect of sensor FOV, and temporal stability of measurements and calibrations.
Linearity refers to the linear relationship between the signal generated by a radiation sensor and the impinging light power. Any dependence of this relationship on additional factors leads to a systematic error in the measurements which needs to be characterized and corrected. Pacheco-Labrador et al. (2014) and Pacheco-Labrador and Martín (2014) assessed the linearity of one of the commercial field portable spectroradiometers, the Unispec-DC (PP Systems) currently used in unattended systems at EC sites (Hilker et al., 2010; Pacheco-Labrador and Martín, 2015), finding that both the grey level measured and also the integration time had an effect on linearity. They showed that non-linearity could be a significant problem in hyperspectral proximal sensing, especially for in situ and long-term unattended measurements. The impact of grey level-dependent non-linearities may be significant when estimating narrow band indices, such as the photochemical reflectance index (PRI), and therefore cannot be left uncharacterized and uncorrected (Pacheco-Labrador and Martin, 2014). The impact of non-linearity can be minimized avoiding the most non-linear region of the dynamic range. In turn, non-linearities related with the integration time affected also the characterization of other instrumental artifacts (Pacheco-Labrador et al., 2014). This dependence, previously reported in cameras (Ferrero et al., 2006) but not in field spectroradiometers, is significant when the integration time is close to the readout time of the sensor (i.e. a photodiode continues to collect photons during the time when the signal is being processed (the readout time), producing an extra signal that is added to that obtained during the integration time). Despite that integration-time-dependent non-linearities have been characterized only in the Unispec-DC (Pacheco-Labrador and Martín, 2015), it would be recommended to avoid integration times close to the instrument readout time, unless the integration-time-dependent non-linearity has been characterized.
Irradiance (i.e. downwelling radiant flux) can be measured with cosine-corrected foreoptics pointed vertically up, or with a diffuse white reference panel (see Fig. 1). Importantly, the materials, calibration status, and method selected to measure irradiance may have an impact on the result. Biggs et al. (1971) highlight the need for a properly designed foreoptic to avoid spectral variations caused by changing sun azimuth and zenith angles. Malthus and MacLellan (2010) demonstrated that the material selected for the foreoptic diffuser can significantly affect the spectra with angular/wavelength dependencies and a poor cosine response. Similarly, they reported that the performance of cosine diffusers in the short wave infrared (SWIR) tends to be very poor with high signal attenuation above 1400 nm (Malthus and Mac Lellan, 2010), and suggested that a diffuse reference panel will provide a better cosine response than cosine-corrected foreoptics. Importantly, because reference panels present and angular-dependent time degradation at such wavelengths, especially when used in the field (e.g. Anderson et al., 2002; Georgiev and Butler, 2007) (Fig. 3), recalibration and regular maintenance is essential. Recalibration requires a dedicated laboratory facility as demonstrated by Georgiev et al. (2011) for the SWIR region of the solar spectrum. In turn, Labsphere, the manufacturers of Spectralon® provide guidance on how to clean reference panels. Consequent care, careful cleaning and recalibration of reference panels are essential to minimise error propagation and uncertainties when conducting spectral measurements.
Spectralon® panel reflectance before (dotted line) and after cleaning (solid line). The panel had been `lightly' and carefully used in the field for one season. Data provided by C. MacLellan, NERC/NCEO Field Spectroscopy Facility, GeoScience, University of Edinburgh.
Selection of reference panel material is also very important. Manufacturers of reference panels for spectroscopy such as LabSphere (Spectralon®) or SphereOptics (Zenith polymer®) use a sintered fluoropolymer manufactured to have a very high reflectance, possibly in excess of 96 % and approximate a Lambertian reflectance across the 400 to 2500 nm spectral region. Alternatively, PTFE sheets (i.e. Teflon) can be purchased at lower cost. However, PTFE sheets have lower reflectance, approximately 80 %, and have higher specular reflection. Also, because PTFE sheets are not manufactured to be used as “references” there may be variability between individual sheets and wavelength-dependent reflectances may be unknown. Overall, PTFE sheets are not recommended as field spectroscopy reference standards.
Similarly, the material and design of cosine receptors affect the estimation
of hemispherical reference factors (Malthus and MacLellan, 2010)
and consequently, the indices derived from them. Therefore, significant and
unquantified uncertainties will be introduced when comparing data from sites
that used different cosine receptors or sites characterized by a different
range of solar zenith angles (SZA). For example, for SZA greater than
60
Field spectroscopists normally assume that the Earth surface sampled by a non-imaging spectrometer with a limited FOV foreoptic is spatially delimited by the solid angle specified by the manufacturer, and that the response across the surface delimited by the FOV is the same for all points inside the given FOV (Castro-Esau et al., 2006; Ferrier et al., 2009; Murphy et al., 2005; Nichol and Grace, 2010). In practice, the spectral response within FOV of a field spectrometer is not constant (i.e. certain areas within the FOV contribute more to the signal than others) (Mac Arthur et al., 2012; Eklundh et al., 2011) and this can be determined by the viewing angle and the instrument's Directional Response Function (DRF) (CIE, 1987) which can be characterized. The DRF will be affected by both the internal design of the spectrometer (e.g. open path or fibre optic transfer to individual detectors) and the foreoptics used. When measurements of heterogeneous surfaces were simulated using the measured DRFs, significant differences were found between simulated reflectance factors and those expected from the manufacturers' specifications (Mac Arthur et al., 2012). Even when less optically complex spectrometers, measuring only across the VNIR region are considered, the Earth surface sampled is not necessarily that inferred from the manufacturers' specified FOV-included solid angle (Caras et al., 2011). The manufacturers of some spectrometers now offer optical elements within their foreoptic mounts to defocus the foreoptics and thereby homogenize the light received (e.g. the ASD FS pistol grip “scrambler”), or have improved the optical components used to minimize chromatic aberrations and heterogeneities and again, homogenize the light received prior to it being distributed to the detectors (e.g. SVC HR-2014i spectrometers). Therefore, the spectrometers' response should be more closely represented by a Gaussian or Cauchy response, albeit centre-weighted, with all areas within the FOV represented in the integrated measurement. These limitations affect the estimation of reflectance factors measured from heterogeneous Earth surfaces (Mac Arthur et al., 2013) because the sample area is ill defined and unknown but systematic sampling errors appear. In contrast, multispectral field sensors normally comprise of individual foreoptics/detector assemblies for each spectral band and subsequently have less complex optical paths than their hyperspectral counterparts, and each sensor can be more reasonably assumed to have a centre-weighted and Cauchy response, though this response is also affected by the viewing angle of the instrument (Eklundh et al., 2011). For a more detailed discussion of the FOV and DRF of field spectrometers and multispectral sensors we refer readers to Mac Arthur et al. (2012) and Eklundh et al. (2011), respectively.
The temporal stability of the measurements and the calibrations are essential factors to be considered when conducting long-term in situ spectral measurements. Factors such as diurnal or seasonal fluctuations in temperature, gradual particle deposition onto optical parts (e.g cosine diffusers or reference panels), or any other processes causing a temporal drift in the functioning will interact with the measured signals and calibrations. In turn, the impact of these factors will depend on the signals we are measuring and the instrumentation we use. For example, because the impact of these factors may be wavelength-dependent it may interfere with the estimation of reflectance indices. Similarly, in DFOV systems constructed around two sensors, the differential impact of these factors in each sensor may also introduce significant errors. Unfortunately, the quantitative characterization of these sources of variability, and the establishment of a set of recommendations, remains a key question after EUROSPEC and clearly requires further attention. We briefly introduce the topic and present some indicative data that we hope will help the reader to understand the importance of temporal stability.
Stability issues can be grouped around two points: the temporal stability of the calibration or cross calibration of a sensor pair, and thermal stability of the measurements.
Temporal stability of the calibration/cross calibration. Sensor
calibration against a source of known spectral and radiometric properties is
needed to derive radiometric units and control for spectral shifts in sensor
response. Similarly, cross calibration of two sensors (e.g. Gamon et al.,
2015; Jin and Eklundh, 2015) is essential for deriving reflectance factors
using two different sensors (e.g. DFOV systems) and to control for
between-sensor variability. Importantly, particle deposition, component
aging, or partial damage of sensor components such as optical fibres may
cause a change in these calibrations which we need to detect, quantify and
correct for. For example, the temporal degradation of the white reference
panel becomes a critical issue in systems such as the ASD-WhiteRef (see
Sect. 3.3), which thanks to the system design was found to be insignificant
(maximum of 2 % differences at 400 nm) over the measuring period
(Sakowska et al., 2015). In the absence of additional information, the
general recommendation is to start with an intensive
calibration/cross-calibration scheme and adjust the frequency later on when
the stability of the calibrations for the specific field conditions is known.
Key questions that the user should consider are the following: what is the temporal drift in calibration for
the specific sensors and measuring conditions? What is the impact of this
drift on the resulting signal/indices? What is the optimal calibration/cross-calibration
frequency? Thermal stability. Changes in temperature may have an impact on both the
intensity and the spectral information of the measured signal. Accordingly,
characterizing the temperature stability of a spectral system and its impact
on the signal we seek to measure is a critical step when designing and
deploying in situ spectral measurements. For example, the radiometric response to
temperature in silicon diodes is more pronounced in the NIR compared to the
visible. Saber et al. (2011) characterized the change in the temperature response of a spectrometer relative to that
obtained at the calibration temperature (20
Conducting in situ long-term spectral measurements in the field is not straightforward. In addition to a number of logistic and infrastructural requirements, long-term field measurements require instrumentation specially designed and conceived for the task. One of the goals of EUROSPEC was to identify the main requirements of such sensors and to promote the development of new dedicated instrumentation. As part of these activities we organized a Science-Industry Interaction Meeting where EUROSPEC scientists got together with representatives of the “spectrometry” industry sector. A number of general requirements for field optical sensors were identified and are summarized in Box 1. In addition, industry representatives raised the issue of how to cover the non-recurring engineering costs associated with instrument development. The possibility of establishing partnerships and seeking funding for joint collaborative projects between science and industry was suggested, particularly to produce prototypes for new instruments.
Four different hyperspectral systems were identified during EUROSPEC for continuous proximal sensing from EC towers in Europe (Fig. 4):
A
temperature-controlled spectrometer system for continuous and unattended
measurements of canopy spectral radiance and reflectance (UNIEDI System)
developed by the University of Edinburgh (Drolet et al., 2014) that has been
operating at the FluxNet Hyytiälä site ( The Multiplexer Radiometer Irradiometer (MRI) developed by the Remote
Sensing of Environmental Dynamics Laboratory, Dipartimento di Scienze
dell'Ambiente e del Territorio e di Scienze della Terra, Università degli
Studi Milano-Bicocca (Italy) and deployed for relatively short periods (weeks
to months) in the context of different projects (Bresciani et al., 2013;
Cogliati et al., 2015). The HyperSpectral Irradiometer (HSI) also
developed by the previous group which has operated in the field from 2009 to
2011 (Meroni et al., 2011; Rossini et al., 2012, 2014). The AMSPEC-MED
system, a version of the automated, multi-angular spectroradiometer system
AMSPEC II (Hilker et al., 2010) modified by the Environmental Remote Sensing
and Spectroscopy Laboratory (SpecLab), Spanish National Remote Sensing (CSIC)
and the Centro de Estudios Ambientales del Mediterráneo (CEAM) in Spain.
This system has been operating at Las Majadas Fluxnet site in Spain
(
General requirements for in situ long-term optical sensors.
Hyperspectral systems in use before and during EUROSPEC. The UNIEDI
system
The first three systems are based on commercially available spectrometers from Ocean Optics, relatively low cost and compact optical benches housed in temperature-controlled environments and operated by dedicated software. The main difference between each of these systems lies in their design.
The UNIEDI system (Fig. 4a) has a hemispherical-conical configuration and is
a DFOV system that uses a pair of spectrometers (Ocean Optics, USB2000
The MRI system also has a hemispherical-conical configuration, but it is an
SFOV system with a single spectrometer. A commercially available optical
multiplexer is used to switch the input to the spectrometer from down-welling
to up-welling radiant flux. Irradiance can be measured through a fibre
connected to either a cosine-corrected diffuser or an up-looking integrating
sphere foreoptic. Up-welling radiance is measured through a bare optical
fibre with an FOV of 25
The AMSPEC-MED system is based on a commercial Unispec dual-channel VIS-NIR spectroradiometer (PP-Systems, Amesbury, MA, USA) equipped with a motor-driven pan-tilt unit that allows measuring up-welling radiance in a range of zenithal and azimuthal angles. Similar to the UNIEDI, the system is a DFOV system and, therefore, cross calibration between spectrometers is performed regularly using a Spectralon® panel. Because the system is operated with solar panels, temperature control is not possible due to power restrictions. Instead, temperature sensitivity of each of the spectrometers and its impact on the resulting hemispherical-conical reflectance factors was characterized in the laboratory and used in signal post-processing (Pacheco-Labrador and Martín, 2015). Note that power constraints are not system-dependent but rather site-specific, depending on power availability and site temperature range.
In an attempt to address some of the limitation of the systems reviewed above and based on discussions between groups during EUROSPEC, three new approaches were developed (Fig. 5):
The Piccolo system, developed by the UK Natural Environment Research
Council (NERC) Field Spectroscopy Facility (FSF) Geoscience, University of
Edinburgh, is based on a DFOV hemispherical-conical configuration with a
cosine-corrected foreoptic to capture down-welling radiant flux and a
configurable up-welling channel to capture up-welling radiant flux. The
up-welling foreoptic can either be fitted with a view angle limited
foreoptic or with another cosine-corrected receptor to enable a
bi-hemispherical measurement approach to be adopted (Fig. 1). The novelty of
this system is the use of low-weight components for decreased weight, and
the use of bifurcated fibre optic with electronic shutters for decreased
time delay between up and down-welling measurements
(Mac Arthur et al., 2014). In
addition, as both light inputs can be closed at the same time, the systems'
dark current (inherent electrical noise) can be recorded and used in post-processing. The Piccolo system is currently undergoing service life cycle
testing and will be field trialled in a number of flux towers in the near
future. In addition, the low weight and DFOV mode of this system makes it
compatible with unmanned aerial vehicle (UAV) applications, opening a new
range of research possibilities. A similar configuration has been adopted
in the SIF-Sys (Burkart et al., 2015) developed by the
Forschungszentrum Jülich GmbH. The system hosts a low cost and small
size spectrometer (STS-VIS, Ocean Optics, Inc., Dunedin, US) and uses also a
bifurcated optical fibre with optical shutters to split the optical signal
between two channels: one channel pointing to a white reference panel to
measure the down-welling radiant flux and the down-looking channel measuring
the radiant flux up-welling from the vegetation. SIF-Sys is specifically
intended to measure SIF and, for this reason, it is equipped with an LED
emitting at the wavelength of SIF (at 760 nm). The LED is placed in the
instrument down-looking FOV and it is used as a reference to assess the
uncertainty of passive SIF retrieval in field conditions. SIF-Sys has been
tested in dedicated field experiments and will be installed at flux towers
for long-term and unattended data collection in the near future. The
ASD-White Ref system (Sakowska et al., 2015) is an automated system designed
for continuous acquisition of measurements using an ASD FieldSpec
spectroradiometer. The WhiteRef system was developed by the Forests and
Biogeochemical Cycles Research Group, Sustainable Agro-Ecosystems and
Bioresources Department, Research and Innovation Centre–Fondazione Edmund
Mach, San Michele all'Adige, together with the Institute of
Biometeorology–National Research Council, Florence in Italy, and the
contribution of NERC Field Spectroscopy Facility, School of Geosciences,
University of Edinburgh, and has been deployed in a grassland site in the
Viote del Monte Bondone in northern Italy. The main advantage of this system
is the possibility to scan in the VNIR and SWIR regions (350 to 2500 nm)
using a popular and commercially available spectrometer. The system is SFOV
and measures in a hemispherical-conical configuration with an FOV of
25
Hyperspectral systems developed during EUROSPEC. The Piccolo system
In situ spectral measurements are essential for the successful upscaling of optical and flux data across space and time. In particular, the temporal match between in situ spectral measurements and flux data facilitates the characterization, modelling and validation of their linkage. Spatial and temporal scales are tightly connected with each other and neither temporal or spatial upscaling can be fully accomplished without giving attention to the other. Considering that the temporal link between optical and flux data can be covered with in situ spectral measurements the main question is probably that of upscaling these signals across space from the footprint to the landscape level (Fig. 6).
Upscaling of optical and flux data across space and time. While long-term in situ spectral measurements help us establish a link between optical and flux data across time, new tools like UAVs are still needed to facilitate the spatial upscaling from the footprint to the landscape level.
In the process of integrating remote-sensing data with flux measurements an assumption is commonly made: the match between flux footprint and image pixel (e.g. Beer et al., 2010; Tramontana et al., 2015). The same assumption can be used between flux data and in situ spectral measurements. However, a number of factors related to footprint variability, pixel heterogeneity, the BRDF properties of the surface, and the geometry of the measurements can momentarily or systematically decouple optical and flux data adding noise or bias to their relationship.
Despite efforts to orientate the FOV of in situ spectral measurements to cover the dominant footprint of EC measurements (e.g. using footprint modelling techniques), the flux footprint will still differ from that of optical measurements most of the time due to footprint variability. Most flux sites are located in places with homogeneous vegetation where footprint variability is not expected to decouple flux and optical data. For example, accurate modelling of the flux footprint did not improve the predictive power of optical data to estimate GPP in a Mediterranean savanna (Pacheco-Labrador et al., 2015) or in a subalpine grassland (Vescovo et al., 2015). However, the mismatch can be relevant in sites with heterogeneous vegetation like agricultural land, ecotones, or sites with adjacent patches of vegetation. In these sites, characterization of the area of interest and footprint modelling will be critical for the successful implementation of data-driven models, e.g. the light use efficiency model introduced in Eq. (1). For example, when estimating GPP in an agricultural area using MODIS data and a footprint model, Gelybó et al. (2013) were able to reduce the RMSE by 28 % compared to non-footprint weighted values.
Dealing with the effect of optical vs. flux footprint mismatch is challenging from a point of view of tower-based measurements. One of the conclusions from EUROSPEC was that new tools are needed to characterize these scale issues more precisely. One of them is the use of small and relatively affordable UAVs or remotely piloted aircrafts (RPAs) on which lightweight spectrometers, both multi- and hyperspectral, and cameras can be deployed. Hyperspectral imaging systems onboard of aircraft or unmanned aerial vehicles (UAV) can for example provide high spatial resolution imagery enabling the identification of pure species pixels within the flux footprint (Zarco-Tejada et al., 2013a). The flexibility, maneuverability, and capacity to view the same target from different heights allows for the impact of footprint variability to be studied and in situ spectral measurements with coarser satellite or airborne data to be bridged, facilitating their interpretation and un-mixing (Fig. 6). For example, the availability of pure pixels can be used to investigate the effect of aggregating different species or land-cover classes on the resulting hyperspectral signal (Zarco-Tejada et al., 2013b). As reported in Gamon et al. (2015) or Whitehead K. and Hugenholtz (2014), the cost effectiveness of UAV platforms makes them a valid solution to address footprint variability. Two UAV-based statistical sampling approaches are possible to systematically address footprint variability: (i) with no previous knowledge a regular grid might be recommended, whereas (ii) if the spatial patterns of vegetation are already known, a stratified sampling for different vegetation types might be more efficient. Overall, the systematic optical sampling of the footprint/pixel area can serve to characterize the different sources of error when upscaling from in situ spectral measurements to the satellite pixel level. These topics have just started to be addressed as low cost UAVs and proper instrumentation are becoming available. The technology is relatively under-explored in the context of flux scaling studies, but there are a growing number of papers that comment on the utility of UAVs for fine-scale sensing of landscape ecology and vegetation parameters (e.g. Dandois and Ellis, 2013).
Data from in situ and remote-sensing measurements are also affected by the structure of the canopy and the geometry of the observation and illumination per se (Jones and Vaughan, 2010). The reason is that photons hitting a surface are preferentially scattered (or reflected) in given directions depending on the properties of the surface. This can be characterized by the BRDF of the surface (Nicodemus et al., 1977; Roberts, 2001). In other words, if we measured an ideal plant canopy with constant fAPAR and LUE, our sensors would still register diurnal and seasonal variations in vegetation parameters due to variations in solar elevation and azimuth. This is particularly relevant when comparing seasonal time series of optical data which may have been acquired with significantly different sun elevations (e.g. in boreal latitudes). Accordingly, knowledge on the BRDF properties of the surface under examination becomes essential to correct for these geometry effects.
The BRDF can be quantified and investigated by mounting sensors on pan-tilt heads (e.g. Huber et al., 2014), by deploying a number of (low-cost) sensors with different fixed off-nadir positions, or by using the UAV systems discussed above. Hilker et al. (2007, 2010) presented a hyperspectral system capable of quantifying and measuring these effects, the AMSPEC and AMSPEC II systems (Automated Multi-angular Spectro-radiometer for Estimation of Canopy reflectance). The AMSPEC system (see previous chapter) is a DFOV system that samples hemispherical-conical reflectance factors at different observation angles from the canopy surrounding the tower. Multiangular measurements are used to retrieve the BRDF and can be used to normalize observations to the same viewing and illumination conditions. Data acquired by AMSPEC system over forest stands in North America showed how optical indices, such as the PRI, can be influenced by view angle and shadow fractions (Hilker et al., 2008b). Moreover, retrieved BRDF estimates allow mimicking off-nadir observations of remote sensors and provide a top-of-canopy reference for atmospheric corrections (Hilker et al., 2009).
Quantifying and modelling the spatiotemporal dynamics of the carbon cycle
remains a key goal in climate change and global biogeochemistry research.
Global questions call for global initiatives to provide sensible data at the
global scale. Flux tower networks such as FluxNet (
EC measurements provide good temporal resolution of carbon fluxes at the ecosystem level but they are limited by spatial resolution and coverage. In contrast, remote-sensing data provide good to moderate spatial resolution and coverage but are limited by temporal resolution. The complementarity and synergy between these two sources of data is clear but their integration remains a challenge due to scale mismatch. We need a Rosetta stone to help us translate and link the information from these two sources of data: something that can be done only via in situ spectral measurements. On one hand, in situ spectral measurements can provide the same optical indices than satellites, serving as a landmark to interpret, calibrate and validate remotely sensed data products (i.e. we can establish a link between satellite data and ground optical data). On the other hand, because data from in situ spectral measurements have comparable temporal resolution and relatively similar biological footprint to that of EC measurements, they can be used to develop quantitative models that associate the two signals.
Overall, there is a clear need to establish a global network of sites with standardized and coordinated in situ spectral measurements to facilitate the integration of remotely sensed data and EC data towards improving the global monitoring of the carbon cycle. In addition, such network is also needed to calibrate and validate satellite data products, and to resolve and avoid problems that appear when inferring ecosystem properties directly from satellite data, such as the “spurious amazon green-up” (Morton et al., 2014; Soudani and François, 2014); or the controversy around the remote sensing of foliar-nitrogen (Knyazikhin et al., 2013; Townsend et al., 2013).
The EUROSPEC Cost Action was a starting point for the organization of the
European community of scientists working with in situ spectral
measurements. We identified many areas that still need further work and
perhaps the main conclusion of EUROSPEC was to realize that we need more
projects such as EUROSPEC. As a continuation, a new COST Action (ES1309)
“Innovative optical tools for proximal sensing of ecophysiological
processes” (OPTIMISE) (
Despite that regional level networking projects such as the EUROSPEC and
OPTIMISE COST Actions, AusCover (
The following challenges and opportunities were identified during EUROSPEC:
Need to compile information on best practices for in situ spectral measurements.
Information on what to purchase, how to install, maintain, calibrate,
analyse, and store the data from in situ spectral measurements is to some extent
available from a number of studies conducted as part of EUROSPEC or by other
groups (see e.g. Anderson
et al., 2011; Balzarolo et al., 2011; Gamon et al., 2015; Harris et al.,
2014; Jin and Eklundh, 2015). These types of studies will most likely
continue to appear in the near future. However, a major up-to-date synthesis
effort is urgently needed to provide a comprehensive treatise on such
measurements. This would facilitate the different phases of decision-making
by site PIs and promote standardization within relevant networks such as
ICOS and FLUXNET. Quantifying and dealing with uncertainty. Measurement uncertainty is
instrument- and environment-specific. Accordingly, characterization of
sensor performance and quantification of measurement uncertainty is crucial
to producing accurate data (Anderson
et al., 2011; Castro-Esau et al., 2006; Jung et al., 2012). Anderson et al. (2011) have
demonstrated that laboratory-derived measurement uncertainties do not
present a useful means of quantifying all uncertainties in field
spectroscopy. Laboratory measurements can serve to define features such as
signal-to-noise ratio, noise equivalent radiance and linearity, but these
uncertainties are added to by complexities of the hemispherical illumination
environment experienced in the field. Clearly, the optimal way to
characterize measurement uncertainty is to do so under the conditions that
typify the measurement scenario. Protocols for systematic measurement
uncertainty characterizations in the field should be adopted in the future. Need for characterization and calibration. Networks of research sites
engaged in optical sampling should follow an instrument characterisation and
calibration scheme to ensure direct result inter-comparison (Anderson et
al., 2013; Balzarolo et al., 2011). Optical sensors could for instance be
characterized and calibrated against a common standard in a central
laboratory prior to field deployment then tested annually to monitor change
or degradation. In addition, portable calibration/verification standards
could be rotated periodically around sites to conduct validation
measurements across space and time. Cross calibration of sensors in DFOV
systems is also critical and should be accomplished regularly. Calibration
frequency will depend on signal drift rate which may be instrument- and
climate-dependent. Accordingly, it would seem logical to characterize and
adjust calibration demands to each site and instrument, for example by
calibrating at high frequencies during the first measuring season and
adjusting later on depending on signal drift rate. Need for a “smart” data repository and information access portal.
Spectral data are time intensive to collect, but their analysis is even more
time consuming. In turn, most spectral data collections remain poorly
documented which greatly reduces their use for data sharing, if not
nullifying it. There is an urgent need for a spectral information system that
(a) establishes a data pool that can hold spectral data collected from
various instruments, providing them in an easily accessible and generic form,
and (b) includes metadata that is standardised to a degree that allows data
selections to answer new science questions. Such databases have been
developed (Bojinski et al., 2003; Ferwerda et al., 2006; Hueni and Tuohy,
2006), but their adoption by the spectroscopy community has been slow.
Currently, there are only a few available and persisting spectral information
systems, the most prominent one being the open-source SPECCHIO (Hueni et al.,
2009). SPECCHIO has seen many upgrades over time with a large contribution by
the Australian National Data Service (Hueni et al., 2012) and support from
EUROSPEC. The challenges for the future are numerous, but most pressing
appears the issue of automated data quality and metadata standards. SPECCHIO
is currently being further developed under the new COST Action OPTIMISE. Upscaling issues: BRDF and footprint analysis. Scaling up in situ spectral
measurements to those acquired from airborne and satellite platforms and
linking this optical data to that of EC measurements remains an issue (Ju
et al., 2005; Simic et al., 2004; Wu and Li, 2009). BRDF effects, footprint
variability, and scale mismatch are factors that constrain our capacity to
link and upscale remotely sensed and EC data. The rapid advances in UAV
technology have opened new opportunities to deal with these challenges.
Micro-hyperspectral field spectrometers and imaging sensors can now be
mounted on UAVs. These measurements can serve to retrieve the BRDF of
challenging Earth surfaces, such as forest canopies, to measure footprint
optical variability, or to sample the same target at different heights
facilitating the treatment of the scale issue. Deployment of specialized
instruments on board UAVs with a view to collecting narrowband multispectral
or hyperspectral data will constitute a step change in scientific
understanding of the connection between spectral data and multiple ecosystem
processes. Investigating this potential is again one of the goals of the new
COST Action OPTIMISE. Permanent platform for communication and information dissemination. The
need for and the potential of a permanent channel for cross-talk between
research communities as well as between scientists and industry stakeholders
was identified. Such a channel could be for example a moderated mailing list
(e.g. Fluxnet type) or making use of other social media. This platform would
provide the opportunity to share know-how and best-practices between users
helping to promote standardization. In addition, it would also promote
collaboration between research groups as well as between scientists and
industry stakeholders, which in turn might foster the development of new
instruments.
We acknowledge the support from COST, Action ES0903/EUROSPEC. In addition we acknowledge the following funding sources: Academy of Finland (Project Number 12720412, 1138884, and 1288039) to A. Porcar-Castell; NERC Field Spectroscopy Facility, GeoSciences, University of Edinburgh to A. Mac Arthur; FCT/MEC, Portugal, (BPD/SFRH/78998/2011) to S. Cerasoli; BIOSPEC/MEC (CGL2008-02301/CLI) and FLUXPEC (CGL2012-34383) to M. P. Martín and J. Pacheco-Labrador Methusalem Program, Flemish Government, Belgium to M. Balzarolo. Edited by: G. Wohlfahrt