Spectral Emissivity (SE) Measurement Uncertainties across 2.5–14 μm Derived from a Round-Robin Study Made across International Laboratories

Abstract

Information on spectral emissivity (SE) is vital when retrieving and evaluating land surface temperature (LST) estimates from remotely sensed observations. SE measurements often come from spectral libraries based upon laboratory spectroscopic measurements, with uncertainties typically derived from repeated measurements. To go further, we organised a “round-robin” inter-comparison exercise involving SE measurements of three samples collected at seven different international laboratories. The samples were distilled water, which has a uniformly high spectral emissivity, and two artificial samples (aluminium and gold sheets laminated in polyethylene), with variable emissivities and largely specular and Lambertian characteristics. Large differences were observed between some measurements, with standard deviations over 2.5–14 μm of 0.092, 0.054 and 0.028 emissivity units (15.98%, 7.56% and 2.92%) for the laminated aluminium sheet, laminated gold sheet and distilled water respectively. Wavelength shifts of up to 0.09 μm were evident between spectra from different laboratories for the specular sample, attributed to system design interacting with the angular behaviour of emissivity. We quantified the impact of these SE differences on satellite LST estimation and found that emissivity differences resulted in LSTs differing by at least 3.5 K for each artificial sample and by more than 2.5 K for the distilled water. Our findings suggest that variations between SE measurements derived via laboratory setups may be larger than previously assumed and provide a greater contribution to LST uncertainty than thought. The study highlights the need for the infrared spectroscopy community to work towards standardized and interlaboratory comparable results.

1. Introduction

Spectral emissivity (SE) is an intrinsic material property, defined as the ratio (0–1) of the electromagnetic radiation emitted by an object at a particular wavelength to that emitted at the same wavelength by a perfect blackbody at the same thermodynamic temperature [1]. SE values range from 0 to 1 emissivity units, albeit when averaged across the mid-wave infrared (MWIR) or long-wave infrared (LWIR) atmospheric windows most natural materials have an SE higher than 0.4 and 0.6 respectively. SE information is essential for the derivation of land surface temperature (LST), an Essential Climate Variable (ECV) important to the understanding and modelling of many Earth system processes from local to global scales [2,3].

LST retrieval algorithms primarily use remotely sensed observations of electromagnetic radiation in the long-wave infrared (LWIR; 8–14 μm) part of the thermal infrared atmospheric window. Some LST algorithms do make use of the mid-wave infrared (MWIR; 3–5 μm) atmospheric window, though these are less common because daytime MWIR measurements are a mixture of thermally emitted and solar reflected radiation [3]. Usually the specific emissivity information required for use in LST retrieval algorithms is the SE integrated over the spectral response function of each of the spectral measurement channels considered in the algorithm [4], though for convenience we still refer herein to SE since typically it is by knowing this that the band-integrated spectral integrated emissivity values are determined.

However, errors in emissivity typically result in significant LST biases. For example, for typical Earth surface conditions, SE uncertainties of 0.01 deliver typical uncertainties of around 0.6 K in the retrieved LST [5]. Given this and recent experimental studies into angular and structural emissivity dependence [6,7], accurate knowledge of SE has been identified as one of the greatest challenges to retrieving sufficiently precise LST to support a wider range of applications [8].

Remotely sensed LST algorithms typically require either (i) knowledge of the SE or its spectral integral in advance, as with widely used split-window algorithms (for example [9]) or (ii) estimate SE or its spectral integral as part of the retrieval process, as with the Temperature Emissivity Separation (TES) algorithm [10]. Laboratory measurements of SE, typically made using Fourier transform infrared (FTIR) spectrometer setups, are commonly used in both approaches, either when deriving the split-window coefficients [11], for calibration of satellite or airborne sensors [12] or for ground-truthing of the LST and SE outputs [13,14,15].

Interest in SE measurements has increased in recent years, largely due to advances in thermal remote sensing and a concerted effort to reduce LST retrieval uncertainty following the classification of LST as an ECV. Campaigns such as Fiducial Reference Measurements for validation of Surface Temperature from Satellites (FRM4STS) have focused on such efforts [16,17]. Laboratory measurements of SE are generally considered to be the “truth” in such work, either for measurement of samples that can be transported without modifying the sample and its emissivity or for evaluating the accuracy of field methodologies on appropriate samples [18]. One key advantage of laboratory SE measurements is the highly controlled conditions under which measurements can be made compared to typical field measurement conditions, and potentially the higher spectral resolution often possible with laboratory setups [19]. Laboratory SE measurements are therefore commonly used as reference measurements when comparing them to those derived using field, airborne or satellite observations [20,21,22].

Given the importance of SE information, multiple laboratories have now developed capabilities for determining SE from thermally emitted or reflected infrared radiation measurements of target samples. Whilst the former “thermal emission” approach is used in for example the SLUM (Spectral Library of impervious Urban Materials) library of Kotthaus et al. [23], the method requires that the samples be heated to well above room temperature, which for some materials is not always possible and which can introduce issues with regards to sample temperature homogeneity when the sample is removed from the heat source to be measured. The latter “reflected radiation” approach instead illuminates a room temperature sample with infrared radiation and measures how much of the radiation is reflected, with SE then determined through use of Kirchhoff’s Law [24]. Key advantages of this approach are that no artificial heating of the samples is required, so all types of sample are analysable, and sample temperature inhomogeneity is not an issue. The approach has been widely applied to provide much of the SE data populating the most commonly used online spectral emissivity libraries, such as the ECOSTRESS (ECOsystem Spaceborne Thermal Radiometer Experiment on a Space Station)—formerly ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer)—spectral library [25]. SE data from the ECOSTRESS spectral library are used within the TES algorithm to provide the spectra required to derive some of the algorithm coefficients [26].

However, the quality of laboratory SE measurements is not always apparent, and there are relatively few reflectance standards readily available for use in the MWIR and LWIR spectral regions with which to assess this, unlike in the near-infrared (NIR) or shortwave infrared (SWIR) [27]. SE quality metrics for an individual laboratory’s SE measurements have often been provided as uncertainty values based on repeated measurements of the sample with the same equipment (for example [28]), but comparisons of laboratory SE measurements derived for the same samples but with different equipment and laboratory setups are rare [27,29]. Here we redressed this gap through a “Round-Robin” study involving seven international laboratories all measuring the same set of reference samples whose SE they determined using their own equipment and measurement protocols. The differing SE measurements were intercompared and their inconsistencies explored to understand the impact that any identified differences in SE would have on remotely sensed LST determination.

The lead investigators of this “Round-Robin” study are based at the National Centre for Earth Observation (NCEO) in King’s College London (KCL), and their SE measurement setup shown in Figure 1 uses a very similar set of equipment to that used in the Department of Earth Systems Analysis at the University of Twente (UT-ITC) and detailed in [27]. SE measurements of the target sample are inferred from reflected infrared radiation measurements made by a Bruker VERTEX 70 spectrometer and application of Kirchhoff’s Law, with the sample positioned under a port of a diffuse highly reflective gold-coated integrating sphere and illuminated by intense radiation coming from an external mid-infrared (MIR) source (Figure 1). Hecker et al. [27] describe two measurement approaches to derive SE from these types of reflectance measurements—the substitution and comparative calibration methods that are described in detail below.

The substitution method of SE derivation uses a material of known emissivity (in this case a Labsphere Infragold™ plate) as a reference sample, and this is first placed under the sample port (e.g., of the integrating sphere of Figure 1) and a “reference measurement” made. The reference sample is then replaced by the target sample and the “sample measurement” made. Spectral reflectance is then calculated through the ratio of these two measurements. In the comparative calibration method of SE derivation, the sample and reference samples are mounted simultaneously (e.g., through multiple ports or use of the internal sphere wall as for the reference), their measurements made consecutively (e.g., through use of an internal rotating mirror), with the sample spectral reflectance again calculated through their ratio. Theoretically, the comparative method should provide some benefit since it avoids a known limitation in the substitution method (the so-called “substitution error”), where changes in the total internal sphere reflectance between measurements of the reference and the sample cause underestimation of sample reflectances (and thus overestimation of emissivity) as discussed in [30]. Hardy and Pineo [31] determined that the substitution error could be as much as 25% for low reflectance samples and 12% for samples with medium reflectance. Corrections have been developed [32,33] but even these are known to include errors of up to 1% from approximations in the calculations. However using both the substitution and comparative methods with the setup at UT-ITC, Hecker et al. [27] observed differences between the SEs derived, with those calculated using the substitution method in closer agreement with other spectra, thus questioning the assumption that the comparative method provided improved results. They attributed these differences to variations in the measurement geometries between the reference and sample measurements made using the comparative method. The measurement setup at the KCL laboratory has been designed to attempt to overcome this issue, and design specifications were to have as identical a path length as possible between the sample and reference measurement when using the comparative method. Within the current work we will therefore also assess the relative performance of the KCL setup when performing SE retrieval with these two different measurement approaches.

Our objectives are therefore threefold: (i) investigate the consistency of SE measurements derived from measurements made in different international laboratories through a Round Robin intercomparison study using reference samples, (ii) evaluate the substitution and comparative calibration methods of SE measurement using the setup shown in Figure 1, and (iii) assess the impact that any SE differences and uncertainties stemming from the results of (i) and (ii) have on typical satellite LST retrievals.

2. Materials and Methods

2.1. SE Sample Standards

Two artificial SE samples were used in this study, with one specular and one diffuse to test the setup performances for samples with different scattering properties. Sample 1 was a thin 60 mm × 60 mm aluminium sheet showing specular reflective behaviour and laminated in polyethylene (PE). Sample 2 was a thin 47 mm × 57 mm diffusely reflective gold sheet also laminated in PE. The samples are shown in Figure 2 and were selected primarily for their robustness and their ability to be used in multiple different laboratory setups having different sample holder sizes and different measurement alignments (e.g., side-looking and down-looking instrumentation). Additionally, both have known spectrally varying properties over the 2.5–14 μm spectral region since the metal foils have low emissivity but the PE film has spectral regions of high absorptivity that result in spectral regions of high emissivity [34]. Due to the nature of the materials used in these reference samples, measurements of SE that required heating of the samples are inappropriate and only laboratories where SE measurements are obtained from directional hemispherical reflectance (DHR) setups were considered.

In addition to the two artificial SE standards shown in Figure 2, the laboratories participating in the Round Robin were also requested to make measurements of a sample of distilled water if their setup was permitting. Distilled water is widely available and its SE is well known and available in the ECOSTRESS spectral library, with SE measurements of distilled water featuring in other inter-comparison studies including [27]. Additionally, since distilled water has a low spectral reflectance of only a few percent, the retrieval of accurate SE poses a challenge for the signal-to-noise capability of setups that measure in DHR mode [28], thus providing a good test of a laboratory’s capability.

2.2. Laboratory SE Measurement Setups and Schedule

In total twelve different measurement setups conducted at seven different international research organisations were used to measure the SE of the two artificial sample standards. All setups other than the Agilent 4300 Handheld FTIR at CSIRO (Commonwealth Scientific and Industrial Research Organisation) involved integrating spheres to derive spectrally resolved hemispherical reflectance (𝜌) measurements over the 2.5–14 μm spectral region. The Agilent does not have an integrating sphere but instead uses mirrors to capture diffuse reflectance from a target. Five of the laboratories provided additional SE measurements of distilled water. Included in the distilled water comparison is a distilled water spectrum measured by John Hopkins University from the ECOSTRESS spectral library [25], also considered in [27] and which was found by [28] to be within 0.17% of a theoretical spectrum of distilled water calculated using optical constants of water in the infrared [35]. This spectrum is identifiable as ECOSTRESS Spectral Library (ESL).

The setup and instruments used within each of the seven laboratories vary in design, interferometer type, age, reference standard (typically gold), and general measurement protocols. A detailed description of each setup is provided in Appendix A, Appendix B, Appendix C, Appendix D, Appendix E, Appendix F and Appendix G, with an overview in

Table 1. An overview of the spectrometer setups of the multiple research organisations participating in this Round Robin surface spectral emissivity measurement intercomparison. The setups are presented here in the order they were used to measure the SE of the selected samples, with information about the instrument, reference standard, gases purging from and references in the literature where available. The names in brackets identify how the laboratories will be referred to subsequently.

Research Institution + Country		Instrument Specification	Reference Standard	Purge Gas	References
National Centre for Earth Observation at Department of Geography, King’s College London (KCL)	UK	Bruker VERTEX 70 FTIR spectrometer with external gold-coated integrating sphere, external detector and external water-cooled globar infrared source	Internal Infragold sphere wall (Comparative) Infragold Labsphere target (Substitution)	Dry, CO2 free air	-
Fraunhofer Institute for Solar Energy System ISE, Optosol GmbH (Optosol)	DE	Bruker VERTEX 80 FTIR spectrometer modified with external integrating sphere and detector	-	-	-
Department of Earth Systems Analysis, University of Twente (UT-ITC)	NE	Bruker VERTEX 70v FTIR spectrometer with external gold-coated integrating sphere, external detector and external water-cooled globar infrared source	Infragold standard	Dry, CO2 free air	[19,27,36]
Planetary Spectroscopy Laboratory at the German Aerospace Center (DLR)	DE	Bruker 80v FTIR spectrometer with gold-coated integrating sphere	Infragold standard	-	[37]
National Aeronautics and Space Administration-Jet Propulsion Laboratory (NASA-JPL)	US	Nicolet 6700 FTIR spectrometer with an external Labsphere integrating sphere	Infragold standard	Dry, CO2 free air	[38]
ONERA—The French Aerospace Lab (ONERA)	FR	Bruker Equinox 55 FTIR spectrometer equipped with a Labsphere Infragold-coated integrating sphere	Infragold Labsphere target	Dry, CO2 free air	[39]
Commonwealth Scientific and Industrial Research Organisation (CSIRO)	AUS	(1) Bruker VERTEX 80v spectrometer with Bruker integrating sphere with multiple sample ports (2) Agilent 4300 Handheld FTIR spectrometer	Infragold target Coarse silver target	Unknown if operational None	[40]

. Note that CSIRO were able to use two different instruments in the Round Robin—a Bruker VERTEX 80v with an integrating sphere with multiple possibilities of port placement for the sample and reference standards, and an Agilent 4300 Handheld FTIR.

The Round Robin was organised sequentially, in that the artificial reference SE samples were measured by one laboratory and then sent onto the next, in the order shown in Table 1. Samples of distilled water were provided by the laboratory participants themselves from local supplies. Measurements commenced in September 2017 at KCL and Optosol. Samples were returned to KCL both halfway (Oct 2018) and at the end (Sept 2019) of the exercise to be remeasured on the same setup and with the same methodology to check for absolute changes to the samples. ONERA’s Infragold reference standard was additionally sent to KCL to be measured using the KCL setup to determine the impact of using a different laboratory’s reference standard to derive reflectance.

2.3. Measurement Protocol

To enable easy intercomparison of the final SE data, all participating laboratories were requested to make spectral reflectance measurements across the 2.5–14 μm spectral range (4000–714 cm⁻¹) at spectral resolutions of 4 cm⁻¹, 8 cm⁻¹ and (if capable) 0.5 cm⁻¹, with all other settings kept as their standard protocol. The majority of participant laboratories made their sample reflectance measurements (${\rho }_{\mathrm{s}}$) by comparing the measured radiation reflected from the sample $\left(V_{\mathrm{s}}\right)$ with that reflected from a reference standard $\left(V_{\mathrm{r}}\right)$ of known (and near unity) spectral reflectance across the wavelength range of interest such that:

$${\rho }_{\mathrm{s}}\left(\lambda \right)=\frac{V_{\mathrm{s}}\left(\lambda \right)}{V_{\mathrm{r}}\left(\lambda \right)}{\rho }_{\mathrm{r}}\left(\lambda \right)$$

(1)

where $\lambda$ is wavelength and ${\rho }_{\mathrm{r}}\left(\lambda \right)$ the spectral reflectance of the reference standard, generally provided by manufacturer or from previous calibration. KCL and UT-ITC made additional open port measurements and subtracted these from both the sample and reference measurements during the spectral reflectance calculation in Equation (1) in order to remove background radiation, as detailed in [27]. ONERA, KCL and NASA-JPL made additional adjustments to the measured data to compensate for the substitution error described in Section 1.

Sample emissivity $\left({\epsilon }_{\mathrm{s}}\right)$ was inferred from measured reflectance using Kirchhoff’s law [24]:

$${\epsilon }_{\mathrm{s}}\left(\lambda \right)=1-{\rho }_{\mathrm{s}}\left(\lambda \right)$$

(2)

The number of scans varied between the different methods, ranging from 60 scans for each sample at CSIRO using the Agilent 4300 Handheld FTIR to eight repeat measurements of 512 scans for each sample at UT-ITC. For most setups, a full measurement sequence (reference, sample and any additional measurements) took between 5 and 30 min.

A summary of the method of SE determination used by each participant laboratory is presented in

Table 2. Details of the methods of spectral emissivity (SE) determination used by each participant laboratory, with abbreviations identifying how the individual measurements will be subsequently referred to.

Institute	Measurement Description	Measurement Abbreviation
KCL	(1) Substitution method with sample and reference alternately placed under bottom port; sphere correction and open port subtraction applied	KCL_sub
	(2) Comparative method with sample under bottom port and internal gold wall of the integrating sphere as reference; open port subtraction applied	KCL_comp
Optosol	Comparative method using internal gold wall of the integrating sphere as reference	Optosol
NASA-JPL	Substitution method with sample and reference alternately placed under bottom port; sphere correction applied	NASA-JPL
UT-ITC	Substitution method with sample and reference alternately placed under bottom port; open port subtraction applied	UT-ITC
DLR	Substitution method with sample and reference alternately placed over sample port	DLR
ONERA	Substitution method with sample and reference alternately placed under sample port. A second measurement of each is made with a tilted beam to correct for substitution method error	ONERA
CSIRO	(1) Comparative method with reference under the bottom port of the Bruker VERTEX 80v FTIR spectrometer and sample in the top port	CSIRO_BG-B_S-T
	(2) Comparative method with reference in the top port of the Bruker VERTEX 80v FTIR spectrometer and sample under the bottom port	CSIRO_BG-T_S-B
	(3) Comparative method using the internal wall of the integrating sphere of the Bruker VERTEX 80v FTIR spectrometer as reference and sample under the bottom port	CSIRO_BG-W_S-B
	(4) Comparative method using the internal wall of the integrating sphere of the Bruker VERTEX 80v FTIR spectrometer as reference and sample in the top port	CSIRO_BG-W_S-T
	(5) Substitution method for the Agilent 4300 Handheld FTIR spectrometer	CSIRO_Agilent

, with abbreviations identifying how the individual measurements will be subsequently referred to. Detailed descriptions of the individual measurement protocols are presented in Appendix A, Appendix B, Appendix C, Appendix D, Appendix E, Appendix F and Appendix G.

3. Results

Due to the very strong similarities in the results obtained at the different spectral resolutions used, we report here the 4 cm⁻¹ results only. This resolution was used by all laboratories and the one for which the widest intercomparisons can be made. Conclusions from the other spectral resolution measurements were similar.

3.1. Sample Stability

Figure 3 shows the mean and standard deviation derived for the two artificial samples from three measurements collected at KCL (with identical parameters and setup) at the start, midway through and at the end of the exercise as detailed in Section 2.2. Despite slight physical abrasion observed on Sample 1 at the end of the study, differences between the emissivity measurements made at different times are small for both samples, as shown by the generally low standard deviation ($<0.01$ for most wavelengths, with the peak in standard deviation around 4.3 μm attributed to insufficient purging during one measurement as it appears in the CO₂ region). Variability is slightly greater for the specular sample (Sample 1) than the diffuse sample (Sample 2), with the mean standard deviations across $2.5-14 \mathsf{\mu }\mathrm{m}$ 0.009 and 0.006 respectively for these two samples. However, these levels of variability are still within the ranges observed in the reproducibility tests in similar studies [27]. These data indicate that the absolute SEs of Samples 1 and 2 were stable throughout the experiment, and that any SE differences found between the different laboratory measurements cannot be attributed to changes in the samples over time.

3.2. Comparison between Different Laboratory’s Emissivity Measurements

3.2.1. Absolute Differences between Emissivity Measurements

The SEs of the two artificial samples and distilled water measured using the setups listed in Table 2 are shown in Figure 4 and Figure 5 respectively. From Figure 4, we can see that there are some large SE differences between the measurements of both artificial samples at certain wavelengths. These differences are reduced for distilled water (Figure 5), although there were fewer measurements here as not all laboratories were able to make measurements of this sample. For all three samples however, there appear to be a group of measurements within the LWIR region consisting of some of those from CSIRO with lower emissivities (around 0.07 and 0.05 less than the majority of the measurements for the artificial samples and distilled water respectively). In the case of the artificial samples, there also appear to be a top group with spectra from DLR and NASA-JPL, which are around 0.05 higher than the majority of the measurements, although the distilled water measurements from these two laboratories are in close agreement with most of the others.

Variability is observed to be wavelength dependent and is greater in the MWIR than in the LWIR region for all samples, as shown by the standard deviations for the two regions presented in

Table 3. Mean (μ) and standard deviation (σ) of SE (ε) for each sample averaged over a specified wavelength range, calculated using all spectra. The number in brackets shows the standard deviation as a percentage of the mean. The subscripts a, b and c refer to the wavelength ranges averaged over, where a = 2.5–14 μm, b = 3–5 μm (MWIR region) and c = 8–14 μm (LWIR region).

		Spectral Emissivity (μ ± σ)
Sample 1	${\epsilon }_a$	0.574 ± 0.092 (15.98%)
	${\epsilon }_b$	0.476 ± 0.107 (22.54%)
	${\epsilon }_c$	0.855 ± 0.046 (5.33%)
Sample 2	${\epsilon }_a$	0.713 ± 0.054 (7.56%)
	${\epsilon }_b$	0.659 ± 0.060 (9.14%)
	${\epsilon }_c$	0.877 ± 0.037 (4.19%)
Distilled Water	${\epsilon }_a$	0.962 ± 0.028 (2.92%)
	${\epsilon }_b$	0.954 ± 0.030 (3.18%)
	${\epsilon }_c$	0.970 ± 0.024 (2.52%)

. This is likely due to increased atmospheric absorption in the MWIR—where there are strong absorption bands for CO₂ and H₂O compared to in the LWIR region—and the differences in how each setup compensate for these atmospheric effects (if at all). There seems to have been an issue with the CO₂ purging in the DLR setup when measuring the artificial samples as these results report an increase in emissivity in the CO₂ absorption band (4.3 μm), which is not present in the measurements from the other laboratories as can be seen in Figure 4. This is not apparent in the DLR measurement of distilled water however (Figure 5).

As shown by the standard deviations in Table 3, SE differences were largest for the specular sample (Sample 1), where the standard deviation of the measurements was ±0.092 over the full wavelength range (2.5–14 μm). The maximum observed difference between two measurements (DLR and CSIRO BG-W_S-T, Sample 1) was 0.762 emissivity units, but this occurred around 4.3 μm and was thus within the CO₂ absorption band and seemed likely to be associated with an insufficient atmospheric compensation in the DLR system as discussed earlier. However, DLR and CSIRO consistently produced the highest and lowest emissivities respectively as evident from

Table 4. The mean SE differences for each setup over specified wavelength ranges, calculated as $\left(\epsilon –{\epsilon }_{\mathrm{mean}}\right)$, where the mean emissivity ${\epsilon }_{\mathrm{mean}}$ was calculated using all spectra. As before, the subscripts a, b and c refer to the wavelength ranges averaged over, where a = 2.5–14 μm, b = 3–5 μm (MWIR region) and c = 8–14 μm (LWIR region).

	Sample 1			Sample 2			Distilled Water
	$\Delta {\mathit{\epsilon }}_{\mathit{a}}$	$\Delta {\mathit{\epsilon }}_{\mathit{b}}$	$\Delta {\mathit{\epsilon }}_{\mathit{c}}$	$\Delta {\mathit{\epsilon }}_{\mathit{a}}$	$\Delta {\mathit{\epsilon }}_{\mathit{b}}$	$\Delta {\mathit{\epsilon }}_{\mathit{c}}$	$\Delta {\mathit{\epsilon }}_{\mathit{a}}$	$\Delta {\mathit{\epsilon }}_{\mathit{b}}$	$\Delta {\mathit{\epsilon }}_{\mathit{c}}$
CSIRO_Agilent	−0.113	−0.145	0.031	0.052	0.056	0.036	-	-	-
CSIRO_BG-B_S-T	0.030	0.030	0.002	0.040	0.049	−0.001	-	-	-
CSIRO_BG-T_S-B	−0.070	−0.087	−0.057	−0.085	−0.101	−0.055	−0.049	−0.053	−0.043
CSIRO_BG-W_S-B	−0.090	−0.111	−0.061	−0.091	−0.107	−0.056	−0.039	−0.042	−0.033
CSIRO_BG-W_S-T	−0.107	−0.132	−0.073	−0.035	−0.035	−0.057	-	-	-
DLR	0.183	0.229	0.069	0.088	0.105	0.053	0.013	0.015	0.010
KCL_comp	−0.014	−0.014	−0.004	−0.013	−0.014	0.001	0.016	0.016	0.016
KCL_sub	0.023	0.028	0.017	0.024	0.028	0.020	0.019	0.019	0.017
NASA-JPL	0.089	0.113	0.046	0.015	0.014	0.035	0.019	0.022	0.015
ONERA	−0.018	−0.016	−0.006	−0.022	−0.022	−0.005	-	-	-
Optosol	0.050	0.052	0.015	0.017	0.012	0.010	-	-	-
UT-ITC	0.037	0.052	0.021	0.010	0.016	0.018	0.021	0.023	0.018

, which presents the mean absolute differences of each individual measurement from the mean of all measurements. The measurements made at DLR had the greatest positive bias compared to the mean while all measurements made at CSIRO using the VERTEX 80v FTIR spectrometer (with the exception of that with the reference in the lower port and the sample in the top port, CSIRO BG-B_S-T) had the greatest negative biases compared to the mean.

Fewer participants made SE measurements of distilled water (Figure 5), but there was greater agreement among these than for the artificial samples, with reduced standard deviations in all spectral regions (Table 3). Much of the variation in the distilled water spectra appeared due to noise, given the difficulties of measuring a high emissivity (low reflectance) sample on a DHR setup. The CSIRO measurements of distilled water made using the VERTEX 80v spectrometer have a negative bias (0.05) compared to the other laboratories, which was consistent with the results of both artificial samples from this setup. In contrast, the measurements from the other laboratories (DLR, KCL, NASA-JPL and UT-ITC) were in very close agreement with the spectrum from John Hopkins University available in the ECOSTRESS spectral library (ESL). This can be observed in Figure 6, where the mean and standard deviation of the differences between the spectra of distilled water from the ECOSTRESS spectral library and the measurements of distilled water from DLR, KCL, NASA-JPL and UT-ITC are shown. The spectrum from NASA-JPL over the LWIR spectral region in particular is in close agreement with that from the ECOSTRESS spectral library, reflecting the reduced noise observed in the NASA-JPL distilled water spectrum compared to the other laboratories (Figure 5).

Figure 7 and Figure 8 compare each laboratory’s SE measurements against the mean and standard deviation of all SE measurements for Sample 1 and Sample 2 respectively over the LWIR atmospheric window (8–14 μm). In this spectral region, which was most commonly used for remote sensing of LST, the measurements of the artificial samples from DLR and NASA-JPL were consistently above the mean, with the DLR spectra greater than one standard deviation away from the mean and with shallower absorption features at 10.8 μm and 13 μm for Sample 1 (Figure 7). Results from CSIRO and KCL in this same spectral region show, respectively, that (i) the positions of the sample and reference standards and (ii) the method of calibration to reflectance both impact the absolute emissivity values retrieved, even on a single setup. At KCL, higher SEs are retrieved when using the substitution method of calibration (KCL_sub) than when using the comparative method of calibration (KCL_comp), with the latter the closest of all laboratory measurements to the mean of all measurements over 8–14 μm, as shown in Table 4. At CSIRO, all the SE measurements made using the Bruker VERTEX 80v FTIR spectrometer were derived using the comparative method, with sample and reference simultaneously mounted. However, merely changing the position of the sample and reference targets changed the derived SE of every samples. This can be seen by the differences between the derived SEs with the reference target in the lower port and the sample in the top port (CSIRO BG-B_S-T) and the derived SEs with the reference target as the internal wall and the sample in the top port (CSIRO BG-W, S-T) in Figure 7 and Figure 8. CSIRO’s highest SE values—and those most in agreement with the other laboratory measurements—are recorded for both artificial samples with the reference target in the lower port and the sample in the top port (CSIRO BG-B_S-T), while the lowest emissivities for Sample 1 are found with the reference target as the internal wall and the sample in the top port (CSIRO BG-W, S-T), and for Sample 2 with the reference target as the internal wall and the sample in the bottom port (CSIRO BG-W, S-B). Due to the nature of the distilled water sample, only measurements with the water sample placed under the bottom port were possible at CSIRO with the Bruker VERTEX 80v. Of these, the SEs derived using the internal wall as reference (CSIRO BG-W, S-B) are consistently slightly higher by approximately 0.008 than those measured with the reference target in the top port (CSIRO BG-T, S-B; Figure 5).

3.2.2. Differences in Spectral Shapes between Emissivity Measurements

The spectral shapes of the SE measurements made at the different laboratories are generally consistent for the Lambertian sample (Sample 2) and for the distilled water measurements, despite absolute differences in retrieved SE. However, this is not the case for the specular sample (Sample 1), where wavelength shifts of up to 0.09 μm are evident in the 9.8–11 μm spectral region as shown in Figure 9 for a subset of the laboratories (Optosol, UT-ITC and ONERA). The wavelength outputs from FTIR spectrometers are known to have some variation resulting in wavenumber calibration procedures as in [42,43]. However, shifts of this magnitude are larger than would be expected from this. While they could suggest alignment issues within the FTIR, if that were the case we would see shifts for all samples—generally the wavelengths would be shifted by a constant correction term dependent on the laboratory setup. Given that the shifts are observed for the specular sample only and that they are observed in measurements on the same setup (Figure 10), incorrect calibration of wavelength outputs is therefore determined not to be the cause.

A more likely cause of the shifts is the different incident angles in each method. This is because, for a specular sample, the resonance wavelength will change with the incident angle, assuming a cavity effect due to the thin layer coating. For example, the ONERA measurement sequence includes a tilted beam measurement for the compensation of the sphere substitution error whereas the beam was on the normal (0°) in the Optosol setup. This theory is supported by the fact that similar shifts are observable over that wavelength range between different measurements at CSIRO that were made using the VERTEX 80v spectrometer with different sample and reference positions (Figure 10). Here the measurements made with the sample in the bottom port (which had good agreement with each other) appear to be up to 0.13 μm out of phase with those where the sample was in the top port. No spectral shifts were observed between the KCL measurements of Sample 1 with different permutations (KCL_sub and KCL_comp, Figure 7).

Other potential causes of the wavelength shifts could be changes in the water vapour and CO₂ conditions between the sample and reference measurements, non-uniformity in the PE film structure and thickness for Sample 1, different sample orientations at the time of measurement as in [44], or differences in the spectral data interval as detailed in [29] and caused by different settings in zero-filling factors for example. To evaluate the impact of sample orientation of position of the spectral features, measurements were made of Sample 1 at KCL at different orientations (0–315° in increments of 45°). The locations of the spectral features in this spectral region agreed between measurements at 0°, 90°, 180° and 270° but small wavelength shifts of approximately 0.04 μm were observed between these measurements and the measurements at 45°, 135°, 225° and 315° (which were in agreement with each other). It is likely therefore that the different sample orientations or differences in the illumination angles used within the different measurement setups could therefore at least partly explain the spectral shifts observed.

3.3. Comparison between Different Laboratories’ Reference Standards

To identify whether the cause of the differences could be attributed to different reference standards used in the substitution approach, a comparison of SEs calculated using two different laboratories’ reference standards was conducted, shown in Figure 11. Using ONERA’s reference standard (with absolute reflectance provided by ONERA) within the KCL setup in the substitution mode reduces the differences between the measured emissivities of the artificial samples by between 10 and 50%. However, it does not equalise them, with the KCL measured emissivities—including those derived using the ONERA reference sample—higher than the ONERA-measured emissivities for both samples between 8 and 14 μm (with the exception of the minima around 13 μm). It also does not account for the wavelength shift between certain spectral features of Sample 1 detailed in Section 3.2.2.

3.4. Implications of SE Differences for LST Retrieval

To demonstrate the impact of the SE differences we observed on typical estimates of remotely sensed LST, spectrally integrated surface emissivities $\left({\epsilon }_i\right)$ for the artificial reference samples and distilled water were calculated across each the five thermal infrared (TIR) bands of the spaceborne ASTER instrument, which is commonly used for LST determination [45,46,47]. ASTER bands 10–14 are centred at 8.3 μm, 8.7 μm, 9.1 μm, 10.6 μm and 11.3 μm respectively in regions of high atmospheric transmittance as shown in Figure 12. Band integrated emissivities were calculated using:

$${\epsilon }_i=\frac{{\int }_{{\lambda }_0}^{{\lambda }_1}S{r}_{\mathsf{\lambda }}\left(i\right)\epsilon \left(\lambda \right)d\lambda }{{\int }_{{\lambda }_0}^{{\lambda }_1}S{r}_{\mathsf{\lambda }}\left(i\right)d\lambda }$$

(3)

where $i$ indicates the band number, $\lambda$ wavelength, $\epsilon \left(\lambda \right)$ the measured spectral emissivity and ${\lambda }_0$ and ${\lambda }_1$ the lower and upper bounds of the band and $S{r}_{\mathsf{\lambda }}$ is the spectral response function for each band.

A typical mid-latitude summer situation was simulated, assuming observation of the each of the samples with a land surface brightness temperature (BT) of 300 K, a sky BT of 260 K and a very near-surface remotely sensed observation (to ensure negligible atmospheric transmissivity and path radiance effects, and thus focus on the surface emissivity impacts only). LSTs corresponding to each ASTER band 10–14 were calculated as in [8], such that the LST of band $i$ was equal to:

$$LS{T}_{\mathrm{i}}=B_{\mathrm{i}}^{-1}\left[\frac{1}{{\epsilon }_{\mathrm{i}}}\left(L_{\mathrm{surf},\mathrm{i}}-\left(1-{\epsilon }_{\mathrm{i}}\right)L_{\mathrm{sky},\mathrm{i}}^{\downarrow }\right)\right]$$

(4)

where ${\epsilon }_i$ is the surface emissivity in band $i$ coming from Equation (3), $B_{\mathrm{i}}^{-1}\left(L\right)$ is the inverse Planck function describing the blackbody equivalent temperature T (kelvin) of spectral radiance $L_{\mathrm{i}}$ ($\mathrm{W}·{\mathrm{m}}^{-2}·{\mathrm{sr}}^{-1}·{\mathsf{\mu }\mathrm{m}}^{-1}$) in band $i$, $L_{\mathrm{surf},\mathrm{i}}$ the spectral radiance ($\mathrm{W}·{\mathrm{m}}^{-2}·{\mathrm{sr}}^{-1}·{\mathsf{\mu }\mathrm{m}}^{-1}$) corresponding to the surface BT in band $i$ and $L_{\mathrm{sky},\mathrm{i}}^{\downarrow }$ the spectral radiance ($\mathrm{W}·{\mathrm{m}}^{-2}·{\mathrm{sr}}^{-1}·{\mathsf{\mu }\mathrm{m}}^{-1}$) corresponding to the sky BT in band $i$.

Figure 13 shows the statistical distribution of LSTs calculated using Equation (4) by sample and by ASTER TIR band, with the box showing the interquartile range and whiskers the distribution (excepting outliers). LSTs derived using the convolved artificial sample emissivities range by over 3.5 K in each band, with a maximum difference of 17.8 K (Sample 1, Band 13). These differences greatly exceed both the GCOS target accuracy and currently achievable requirements for LST as an ECV, which are 1 K and 2–3 K respectively [49]. The difference of nearly 20 K for Sample 1 in Band 13 is related to this waveband covering the spectral range containing the observed wavelength shifts in the Sample 1 SE measurements, and being an area of increased atmospheric attenuation and thus stronger downwelling irradiance impact.

The range of the LSTs calculated using the distilled water emissivities is reduced compared to those calculated using the artificial sample emissivities, reflecting the greater agreement between the SE measurements of distilled water. However, the range of the LSTs calculated using distilled water spectra convolved to the ASTER TIR bands is still 2.5 K in all bands, thus still exceeding the GCOS target accuracy requirements.

4. Discussion

Absolute differences observed between the different measurements were larger than anticipated with no clear cause. The derived emissivity does not correspond to different spectrometer types as may be expected: DLR and CSIRO’s measurements were both made on setups based on a Bruker VERTEX 80v FTIR spectrometer, albeit with different spheres, however there were large differences between these measurements for all three samples over the full wavelength range. Furthermore, the measured absolute differences cannot be solely attributed to the use of the reference standard, although results from Section 3.3 indicate that the uncertainty in reference standard calibration is a key factor in the SE uncertainty. The contribution to uncertainty from reference calibration is particularly pertinent given that Labsphere Infragold standards can no longer be bought with NIST traceability calibration certificates. Differences in the absolute reflectance of the reference standards could be due to different coatings but could also result from physical damage or degradation from humidity absorption. To reduce uncertainty regarding the latter, laboratories should ensure regular calibration of their reference standards to monitor for drift.

Results from all three samples considered indicated that three of the four CSIRO measurements made using the VERTEX 80 had a consistent negative bias, with the measurement made with the reference target in the lower port and the sample in the top port (CSIRO BG-B_S-T) the only one in agreement with the others. Given the similarity in the spectral shapes measured by this setup and those measured at other laboratories for all three samples, this could suggest that CSIRO may need to re-characterise their reflectance standards for the other three configurations used with the VERTEX 80v spectrometer. However, these differences may also be due to the different optical path lengths of the sample beam with each permutation, or to directional reflectance effects of the sample at the different incident angles of each permutation. Further investigation is recommended if CSIRO wish to use any of the three biased configurations (e.g., to measure liquid samples, which is not currently possible with the sample in top port configuration). Removing these three measurements from the analysis reduces standard deviations to ±0.089 (14.69%), ±0.038 (5.16%) and ±0.008 (<1%) for Sample 1, Sample 2 and distilled water respectively across the 2.5–14 μm wavelength range. Furthermore, the impacts on LST are reduced considerably without these measurements, with the range of the LSTs calculated using distilled water spectra convolved to the ASTER TIR bands reduced to <0.45 K in all bands.

Differences were also observed between measurements made on the KCL setup with the reference target in different positions. Possible causes for these differences at KCL are (i) differences in the path lengths for each measurement setup that remain unaccounted for as discussed in Hecker et al. [27], (ii) insufficient correction for the substitution error in the substitution method, (iii) differences in reflective properties of the reflectance targets (being flat and curved respectively for the substitution and comparative methods) and (iv) incorrect characterization of the reference target spectral reflectance (the Infragold reference panel and sphere wall for the substitution and comparative methods respectively). In terms of the latter possible cause, it was identified that scaling the provided absolute spectral reflectance of the internal wall of the KCL integrating sphere (which is used as the reference target during the comparative method) by a factor of 0.87 (so that ${\rho }_{\mathrm{r}}=$ 0.84) brought the derived SE very close to that derived using the substitution method (which used the Infragold reference panel as the reference) for all three samples, as shown over the LWIR region in Figure 14. However, the reflectance of the internal wall coating should be much higher than this given the material type and the fact that both this system and the Infragold target used as the reference target for the substitution method are new. Note that it is not physically possible for the quoted spectral reflectance of the Infragold reference panel to be higher by this amount as this would result in emissivities above $1$. Evaluating the performance of the two methods provides mixed results. While measurements derived using the comparative method of calibration were found to be closer to the mean for all samples, KCL’s distilled water emissivity measurements retrieved using the substitution method of calibration were observed to be closer to the ECOSTRESS spectral library spectrum than those derived using the comparative method. This could indicate that the mean may have been negatively skewed by the afore-mentioned bias in the CSIRO measurements. Discounting the three negatively biased measurements from CSIRO from the analysis supports this: KCL’s measurements derived using the substitution method were in good agreement (<0.01) with the recalculated mean over 2.5–14 μm for all three samples while KCL’s measurements derived using the comparative method were in poorer agreement, with differences of up to 0.04 across 2.5–14 μm. Based on these results, further investigation should be conducted to determine the cause of the differences between the measurements from KCL and to determine an optimal approach.

The increased variability in the MWIR than LWIR observed is likely due to the increased atmospheric effects in this region, with the DLR measurements of Samples 1 and 2 clearly impacted in the CO₂ region (Figure 4). An alternate explanation for the reduced variability in the LWIR for both artificial samples could be because this is an area of high emissivity (and thus low reflectance), which Hecker et al. [27] observed to be areas of better agreement in their intercomparison of emissivity spectra from different laboratories.

This latter interpretation could also be why the distilled water measurements (with uniformly high emissivity) had reduced variability compared to the artificial sample SE measurements (which had variable emissivities between 0 and 1). However, it is more likely that the increased variability of the SE measurements of the artificial samples was due to the composition of these samples and their interaction with different setups. Tsilingiris [50] provide a transmittance spectrum for polyethylene (PE), and considering this against the measured spectra of Samples 1 and 2 in Figure 4, it is clear that variability amongst the different measurements was lower in the regions where PE had a low transmittance (3.5 μm, 6.9 μm and 13.8 μm). In all other spectral regions, the PE formed a multilayer system, which is potentially sensitive to directional illumination characteristics. Differences in the incident angles upon the samples within the different measurement setups could therefore at least partly explain the SE variations seen. Given that Sobrino and Cuenca [51] observed that emissivities in field measurements tended to decrease with increases in the observation angle, results from this study indicate that future work should be conducted to explore whether emissivities from DHR setups in the laboratory similarly correlate with the incident angle. Materials with expected directional behaviour should in particular be considered.

The observed spectral shifts between different measurements of the specular sample also raise interesting questions about the impact of incident angles on SE measurements. While this may not be an issue for non-specular samples without coating (and therefore for most natural samples), these discrepancies indicate that further work should be conducted to confirm and investigate the impact of the incident angle on the spectral stability and absolute emissivity. Spectral shifts of the magnitude observed in this study will have more implications when working with data from hyperspectral rather than multispectral thermal imagers. Conversely airborne hyperspectral instruments such as NASA-JPL’s airborne Hyperspectral Thermal Emission Spectrometer (HyTES) sensor have so many narrowband (18 nm) spectral channels between 7.5 and 12 µm that TES and spectral smoothness approaches can be applied, which reduce the need to prescribe emissivity in advance [52,53]. However, such TES approaches often rely on laboratory emissivity spectra to derive empirical relationships used within the algorithm, and so their accuracy is still important even though for each hyperspectral image pixel the emissivity is directly retrieved [10]. Such wavelength shifts may also have implications for use of spectral emissivity features in, e.g., mineral identification studies [54], and also may affect in situ LST measurements given that radiometers commonly used in LST validation studies are affected—such as the Heitronics KT15.85 IIP radiometer with a spectral range of 9.6–11.5 μm [16,55].

Consideration of how the uncertainty of individual SE measurements translates into retrieved LSTs in Section 3.4 indicates the importance of reducing uncertainties in laboratory SE measurements to improve remotely sensed LST estimates. While it must be acknowledged that the artificial samples are not representative of many land surfaces, with most natural surface less reflective in the LWIR, similar samples may be observed in remote sensing of urban areas (e.g., for urban heat island monitoring) or monitoring of plastic pollution in water [7,56,57].

5. Summary and Conclusions

Surface spectral emissivity data collected with twelve different laboratory spectrometer measurement setups made at seven different laboratories were compared over the MWIR and LWIR spectral ranges in a Round-Robin intercomparison exercise. All measurements were based on the principle of illuminating the sample with an intense source of infrared radiation, measuring the reflected signal, and converting this to an emissivity spectrum using Kirchhoff’s Law. Three different samples were used for the exercise. The first two were artificial samples constructed from gold and aluminium sheets each laminated in PE films that had Lambertian and specular characteristics respectively and with widely varying emissivity features across the 2.5–14 μm spectral range. The third sample was distilled water, which has a relatively flat emissivity spectrum close to unity.

Comparing the measurements from the different laboratories we found that the inter-setup variability of the SE measurements was larger than anticipated, with differences in magnitude and spectral shape. Standard deviations of ±0.092 (15.98%) and ±0.054 (7.56%) were identified across the 2.5–14 μm spectral range for Samples 1 and 2 respectively. Repeated measurements using the same measurement setup and protocol at different times confirmed that observed SE differences were not attributable to changes in the sample properties over the course of the study but were rather due to the different setups and measurement procedures used in the various laboratories. Variability was greater in the MWIR rather than LWIR spectral region, likely due to differing efficiencies of atmospheric purging, which impact this region more. SE differences across the LWIR atmospheric window (8–14 μm), which is the most important for the remote sensing of LST, were ± 0.046 and ± 0.037 respectively for Samples 1 and 2, and most of the Sample 2 SE measurements were within 0.02 of the mean. The greater variability for the specular sample (Sample 1) over this region was attributed to spectral shifts, with differences between identification of spectral maxima and minima of up to 0.09 μm between different laboratories and up to 0.13 μm between different positional permutations on one setup were observed in the 9.8–11 μm spectral range in particular. Investigation indicated potential causes of these spectral shifts to be different sample orientations during measurements or differences in the incident angles within the different measurement setups. The latter cause was also identified as a potential cause of the absolute emissivity differences. Further investigation is therefore recommended into the impact of directional effects in laboratory measurements of emissivity (particularly for materials with known directional behaviour) given recent advances into understanding the angular dependence of emissivity for field and satellite measurements [6,58].

Use of different reference standards was found to contribute to the observed SE differences between different laboratory measurements but not to be the sole factor. Nonetheless, the differences observed from use of a different reference standard suggest that uncertainty in the reference standard calibration is a key factor in emissivity uncertainty in laboratory measurements. Regular calibration of the reference standards is recommended to reduce this uncertainty.

SE variability was comparatively lower for distilled water than for the artificial samples, with a mean emissivity of 0.962 ± 0.028 determined over the 2.5–14 μm spectral range. These uncertainties are larger than those observed in other studies but standard deviations were reduced to 0.008 when discounting measurements from one laboratory with a consistent negative bias (likely indicating inaccurate calibration of their reference standard). Other contributions to the SE differences included the method of calibration to reflectance, different incident angles and the placement of the sample and reference standards within the same setup. Regarding the setup at KCL, the primary laboratory used in this study, the measurements indicate that the comparative method of calibration was closer to the mean of all measurements. However consideration of the distilled water suggested instead improved performance using the substitution method and therefore further investigation is recommended.

The impact of the determined spectral emissivity differences on LST retrieval was evaluated by considering a typical mid-latitude summer scene with the surface emissivities set to the sample emissivities convolved to the ASTER TIR bands. With Sample 2 (the diffuse sample) considered more representative of natural surfaces than Sample 1, use of the three samples in this simulation provided LST error estimates over diffuse surfaces, over an extreme case of 100% specular surfaces and over water bodies. Calculated LSTs using the convolved artificial sample emissivities ranged by over 3.5 K in each band, with a maximum difference of 17.8 K (Sample 1, Band 13). The range of the LSTs calculated using distilled water spectra convolved to the ASTER TIR bands (and thus an error estimate for retrieval of LST over water bodies) was lower, but still around 2.5 K. The variability of the artificial samples and the distilled water emissivities measured at different laboratories would therefore result in uncertainties in LST estimates that exceeded the target accuracy requirements for satellite observations of land surface temperature, even without considering contributions from atmospheric effects.

Overall, our findings highlight the need for the infrared spectroscopic community to work towards standardized and interlaboratory comparable results, with regular calibration of reflectance standards and the laboratory setup against SI traceable standards.