The HyTES instrument is a Dyson optical configuration with a compact hyperspectral grating spectrometer acquiring data in 256 spectral bands in the TIR range from 7.5 to 12 µm (Fig. 1), and a Quantum Well Infrared Photodetector (QWIP) (Gunapala et al., 2006). This is the first integration of the QWIP with a spectrometer imaging system for Earth science studies that require well-calibrated data. A major advantage of the instrument is its very compact design due to its small form factor and low power requirement (1 kW) when compared to what the aircraft can support (4 kW) (Johnson et al., 2012). A vacuum chamber is used to keep the focal plane system cold using two mechanical cryocoolers (Fig. 1). The chamber has also been proven to support airborne operation of other VSWIR instruments, while maintaining rigidity of its inner precision and optical components. A full description of the HyTES instrument including instrument performance, calibration, and validation is provided by Hook et al. (2016).

HyTES is currently configured to fly on the Twin Otter aircraft and Fig. 1 shows the aircraft and the HyTES instrument looking nadir in flight. For Twin Otter flights, the instrument is calibrated before and after each flight (including any intermediate stops), and the nominal operation for data processing from L0 to L1 is to average the pre- and post-flight calibrations. That being said, a day to day comparison between calibrations in 2014 showed that a single calibration in fact could be substituted with only minor errors for the whole week's campaign.

For detection of trace gases, flights are usually conducted at an altitude of 1 km above ground level (a.g.l.) to minimize atmospheric attenuation between ground and sensor. The HyTES pixel size at 1 km a.g.l. is ∼ 2 m. Figure 2 shows an example of a HyTES data hypercube for a flight over Death Valley, California. Radiances in the vertical slice have been atmospherically corrected for the atmospheric transmission and path radiance using an in-scene atmospheric correction approach.

Using Kirchhoff's law, we can write the hemispherical–directional reflectance as a function of directional emissivity (ρλ=1-ϵλ), and express the at-sensor radiance for a clear-sky pixel with no gas plume attenuation (“off-plume”), (Lλoff) as follows: Lλoffθ=Lλgndτλatm+Lλ↑, where τλatm is the atmospheric transmittance, Lλ↑ is the atmospheric path radiance, and Lλgnd is the total land leaving radiance: Lλgndθ=Lλsurf+ρλLλ↓=ϵλBλTs+1-ϵλLλ↓, where λ is the wavelength; θ the observation angle; Lλsurf the Earth-emitted radiance; ϵλ the spectral surface emissivity; Ts the surface temperature; Lλ↓ the downwelling sky irradiance; τλatm the atmospheric transmittance; Lλ↑ the atmospheric path radiance; BλTs the Planck function defined at temperature, Ts.

The radiance measured by a sensor for a pixel centered on a gaseous plume (“on-plume”) introduces an additional plume thermal emission term, Lλp, and a plume transmissivity term, τλp to account for the additional attenuation of the surface radiance: Lλonθ=Lλgndτλatmτλp+Lλ↑+Lλpτλatm, where Lλp=ϵλpBλTp is the gas plume emission term, Tp is the plume temperature, and ϵλp the plume emissivity. An illustration depicting these components for an observation over a gaseous plume is shown in Fig. 3. These terms can be simplified using some physics-based assumptions. The weak plume transmissivity, τλp, is given by Beer's law: τλp=e-nobλ, where no is the gas column density and bλ is the gas absorbance spectra (“plume signature”), usually extracted from the HITRAN database for the relevant gas constituent. If we assume the gas plume is optically thin and plume absorbance, (nobλ), is small, we can approximate Eq. (4) with a Taylor expansion: τλp≈1-nobλ. The Beer–Lambert law can then be used to write the transmittance as a function of the gas plume effective emissivity: ϵλp=1-τλp≈nobλ. Substituting Eq. (6) into Eq. (3) and rearranging terms yields an equation describing the total at-sensor radiance for an observation centered on a gaseous plume pixel: Lλonθ=[Lλgndτλatm+Lλ↑]+nobλτλatm[BλTp-Lλgnd]. The first term on the right-hand-side of Eq. (7) describes the off-plume radiance, Lλoff (background, or “clutter”), while the second term consists of the plume signature, b, multiplied by the plume strength, which includes the column density, no, multiplied by the thermal contrast term, [BλTp-Lλgnd]. Examination of this term indicates that the detection of gaseous plumes in the TIR requires a finite thermal contrast between the surface and the plume, otherwise the plume strength term approaches zero. To solve Eq. (7), knowledge of the surface temperature of the background, the temperature of the gas plume, the surface emissivity of the background, and the atmospheric terms τλatm and Lλ↑ is needed. The atmospheric terms are estimated using an atmospheric correction technique described in the next section, and are usually fairly constant across an image at the scale of a few kilometers depending on variability in atmospheric water vapor.

The spectral radiance in Eq. (1) will include atmospheric emission, scattering, and absorption by the Earth's atmospheric constituents. In order to isolate the land-leaving surface radiance and separate the surface temperature and spectral emissivity terms, these atmospheric effects need to be removed from the observation. For on-plume pixel observations, the atmospheric compensation isolates the land-leaving radiance contribution in addition to reducing the wavelength dependence of the plume strength, which is the difference between radiance emitted by plume and ground as expressed in Eq. (7). The success of the atmospheric correction depends on the accurate characterization of the atmospheric state that is input into the radiative transfer model (RTM) e.g., MODTRAN (Berk et al., 2005). Independent atmospheric profiles of temperature, water vapor, and other gas constituents (e.g., ozone) are input to the RTM to obtain the atmospheric transmittance, path radiance, and sky irradiance terms. Once the residual effects of the atmosphere have been removed from the observed radiance the surface properties can be obtained.

For multispectral data, where the bands are typically not strongly affected by the atmosphere the RTM approach works satisfactorily, but for hyperspectral data the RTM approach is more challenging when bands are situated in strong atmospheric absorption features and if output model data from the RTM are not accurately spectrally registered with the observed data, then the solution may be unstable. This instability primarily arises because (1) methods used in the RTM to interpolate hyperspectral absorbances introduce error, (2) the sensor's spectral responses functions are not precisely defined, and (3) band-to-band registration issues result in model error. In these cases even small misregistrations between the observed and modeled data near strong absorption lines will amplify instead of reduce the effects of atmospheric attenuation, making correction of the radiance spectrum very difficult. To address these issues, an in-scene atmospheric correction (ISAC) approach was developed for the SEBASS airborne hyperspectral sensor (Young et al., 2002). The main advantage of the ISAC method is that atmospheric correction is accomplished using the hyperspectral data itself without the need for external atmospheric profiles or an RTM. In addition, the issue of spectral band misregistrations is eliminated. An example of this is shown in Fig. 4 where surface brightness temperature spectra are shown from HyTES after atmospheric correction using the RTM approach with MODTRAN (gray line), and the ISAC approach (black line). With a successful atmospheric correction we expect a nearly constant temperature across all bands, which is achievable with the ISAC approach but not with MODTRAN below 8 µm and above 11.5 µm because of band-to-band misregistrations between HyTES data and MODTRAN in the presence of higher water vapor absorption regions.

ISAC relies on finding gray bodies in a given scene with emissivity close to 1 across all bands, ϵλ∼1 (e.g., water, dense vegetation, ice, snow). Then, the observed radiance in Eq. (1) can be written as a linear function with an independent variable, BλTs, and with slope τλatm and y intercept Lλ↓ as follows: Lλoff=BλTsτλatm+Lλ↑. Theoretically, the atmospheric parameters τλatm and Lλ↑ can then be found by simple linear regression by plotting Lλoff vs. BλTs for all pixels on a scene for a given wavelength. We found that using the maximum brightness temperature “most hits” method as proposed by Young et al. (2002) resulted in pixels consisting of different types of soils in agricultural environments, with emissivities < 0.95 often being included in the fitting procedure. This was verified by comparing these pixels with emissivity information from the ASTER Global Emissivity Database (ASTER GED) at ∼ 100 m spatial resolution (Hulley et al., 2015). Misclassification was usually worse over scenes with high temperatures, where bare soils exhibit near-blackbody-like behavior and are confused with true gray bodies such as dense vegetation. These non-gray-body pixels violate the intrinsic assumptions of the ISAC method, leading to errors in the fitting procedure.

To address this issue we developed a spectral variance approach in which the spectral variance in observed radiance was calculated for each pixel and only those pixels with low variance (e.g., a threshold set at less than 8 W m-2) were assumed to be gray-body pixels suitable for use in the fitting procedure. Using this approach resulted in a very good match with gray bodies classified according to the ASTER GED emissivities. The spectral variance approach is a good assumption for low-altitude flights (1 km a.g.l.) in which observed radiance is still representative of underlying surface spectral features, and also because the emissivity spectra of gray-body surfaces such as vegetation, snow, ice, and water are pseudo-invariant in the 8–12 µm range.

A large fraction of the HyTES target sites including those over a few key methane hotspots (e.g., Kern River oil field) were flown over bare regions containing very few gray-body pixels (e.g., vegetation, water) and an alternative ISAC approach had to be developed. In this approach, termed the ISAC-ASTER method, emissivity information from the five ASTER GED TIR bands from 8 to 12 µm were used directly in the ISAC fitting procedure instead of relying on the blackbody assumption. ASTER GED emissivities at 100 m spatial resolution were first geolocated and interpolated onto the HyTES scene and then a principal component (PC) regression approach (Borbas et al., 2007) was used to extend the 5 ASTER band emissivities to the 256 HyTES bands from 7.4 to 12 µm (Hulley et al., 2014).

Starting with a data cube, L of hyperspectral thermal infrared data, contains an image of N columns by n rows, where the columns denote the number of pixels in a given image, and n denotes the number of spectral channels. The goal is to find a wavelength-dependent spectral signature, b, which is assumed to be linearly superimposed on the background signal or clutter. This can be expressed by the following equation: r=∝b+c, where r is the total radiance and can be modeled as a linear combination of signal, ∝b, where ∝ is the strength of a plume signature, b, and c is a noise term that contains both sensor noise and scene clutter. The plume signature b is usually expressed in terms of absorbance, and is typically extracted from the HITRAN database and convolved to the sensor's spectral response. Figure 5 shows an example of CH4 and H2O normalized absorbance spectra in the 7.5–8.1 µm range convolved to the HyTES spectral response. The scene clutter contains radiance contributions from the ground and atmosphere, and is defined as noise with cross-spectral correlations. These spectral cross correlations can be written in terms of a covariance matrix, K: K=〈ccT〉=1NLLT. Given the covariance of the background clutter, K, we can then find the optimum filter vector, q, as follows: q=K-1bbTK-1b, where q is normalized such that the variance, qTKq=1. This ensures that in the absence of the signal, the matched filter image will have a variance of 1. The final clutter matched filter (CMF) image, is calculated by applying q to the original data cube of radiance: CMF=qTL. In order to minimize the effects of striping and other noise in the data, the CMF is calculated in a matrix column-wise fashion (along-track) for a given data swath. The CMF result for each column is then demeaned by subtracting the sample mean from each observation and dividing by the standard deviation using all pixels on the scene. This results in a mean CMF of zero and standard deviation of 1 for each column of data. The final CMF will produce an image in which the intensity correlates with the desired plume signature as defined by b. Values that are classified as outliers in the final CMF are strong evidence for the presence of the desired signature, and their significance quantified by number of sigmas of the distribution; however this metric is only valid if the matched filter distribution is Gaussian (Funk et al., 2001).

We can further define a dimensionless quantity called the signal clutter ratio (SCR), which is computed by applying the signal filter vector in Eq. (11) to the target plume signature, b: SCR=qTb. The SCR can be used as a metric for evaluating the strength of the desired target signal above background clutter, or the radiance emitted by other targets in the field of view. Usually the optimally derived CMF in Eq. (12) will maximize the SCR values derived in Eq. (13). SCR values are normalized from [0 1] and values closer to 1 indicate higher confidence in the presence of the desired gas target pixels in the image data.

The CMF detection algorithm for HyTES is optimized to detect only the strongest CH4 sources using a five-step process. The algorithm is designed to minimize false positives while enhancing plume structure around the strongest sources using a plume dilation algorithm. This algorithm is used to provide qualitative information to help attribute emissions to specific source types and source locations. The CMF can also be tuned to detect more diffuse CH4 enhancements that could be the result of advection further downwind from a specific source. For example, ground surveys have shown that some of the highest concentrations are found downwind at significant distances (hundreds of meters) from the original source (Leifer, 2014). However, lowering the CMF threshold comes at the cost of increasing the likelihood of false positives in the final image.

Once a binary image of the strongest plume pixels is generated from thresholding the CMF result, a dilation algorithm is used to enhance the structure and edges of the plume (Broadwater et al., 2008). The binary image is first dilated within a 2-by-2 pixel neighborhood and then multiplied by the original CMF detection image. This results in an image with modified CMF values in the neighborhood immediately around the original plume pixels. A slightly lower detection threshold is then applied to the new detection image, resulting in a binary image that is again dilated within a 2-by-2 pixel neighborhood. This process repeats until a minimum detection threshold is reached based on the initial threshold set. After each iteration, a contiguity test is applied that removes any pixels with fewer than two neighbors. The result is an adaptive plume-growing algorithm that finds the gas plume edges immediately surrounding the strongest gas plume pixels, while simultaneously reducing any false positives and noise.

A number of different configurations and thresholds were tested, which resulted in a final set of steps that both optimized the presence of the strongest gas plume pixels and simultaneously reduced any false positives and noise. The results of the three primary steps are demonstrated in Fig. 6, which shows a sequence of two methane plumes detected over the Kern River oil field (top panels) and Four Corners (bottom panels).

While HyTES has the ability to detect multiple trace gases, much of this work focused on improving understanding of atmospheric methane given its high importance both for scientists and decision-makers as a key climate-forcing greenhouse gas and ozone precursor. The atmospheric growth rate of methane and controlling emission sources remains highly uncertain at regional to global scales (Dlugokencky et al., 2009; Kirschke et al., 2013; Miller et al., 2014; Rigby et al., 2008). Future changes in surface temperatures and precipitation have the potential to dramatically alter natural methane fluxes from large Arctic reservoirs (Damm et al., 2010; Kort et al., 2012) and tropical wetlands (Dlugokencky et al., 2009), while transformational changes in anthropogenic emissions from fossil fuel production threaten to further increase atmospheric methane abundance (Larsen et al., 2015). Examples of anthropogenic sources of methane include the natural gas and oil supply chains (production, storage, transmission, distribution, consumption), agricultural activities (enteric fermentation, manure management, rice cultivation), landfills, coal mining, stationary and mobile combustion, and wastewater treatment (Thomas and Zachariah, 2012). This work focuses on anthropogenic point source emitters rather than more diffuse area sources, given that the former are both uncertain and more readily detectable with TIR imaging spectroscopy.

Detection of methane from infrared measurements is possible due to the absorption from strong rotational-vibrational transitions (ν4) in the 7.3–8 µm range that have sufficient separation from the strong water vapor band centered at 6.3 µm (see Fig. 5). Hyperspectral satellite sensors like the Infrared Atmospheric Sounding Interferometer (IASI) (Aires et al., 2002), the Tropospheric Emission Spectrometer (TES) (Beer, 2006), and the Atmospheric Infrared Sounder (AIRS) (Tobin et al., 2006) are able to take advantage of these absorption characteristics of methane, however are limited by their coarse spatial resolutions (10 km or more) and insensitivity to near-surface concentrations due to sensor saturation issues. Airborne hyperspectral TIR sensors such as HyTES and others detailed in Table 1 have the imaging capability of detecting methane emission sources at the scale of a few meters, allowing improved characterization of individual point sources towards better understanding of their distribution.

The following section demonstrates a few examples of the capability of HyTES to detect multiple chemical gas species. The ability to distinguish between different trace gas signatures within a single plume consisting of several contiguous pixels is a key advantage of TIR hyperspectral data. The different chemical species that will be shown include NH3, SO2, H2S, and CH4 and their distinctive features in the infrared domain from 7.5 to 12 µm are shown in Fig. 5.