Night-Time Ground Hyperspectral Imaging for Urban-Scale Remote Sensing of Ambient PM. I. Aerosol Optical Thickness Acquisition

Abstract

Aerosol loadings in vertical atmospheric columns are commonly measured by satellite-borne or ground instruments that remotely sense the spectral extinction through the optical path. However, correlations of these integrated measurements with ground-level particulate matter concentrations are highly influenced by local meteorology, seasonality, and the surface albedo. Moreover, as most measurements are based on solar radiation, they are limited to daytime. To account for these limitations, we study the feasibility of using a ground hyperspectral camera for acquiring images of artificial light sources through horizontal urban-scale open paths during the night, in order to retrieve the apparent spectral aerosol optical thickness. Laboratory-scale measurements demonstrated a linear response of the camera and set the spectral operational range. A procedure for night-time imaging of illuminating targets through ambient open paths has been developed to enable a consistent selection of pixels for analysis, providing measurable apparent aerosol optical thickness. We demonstrate the validity of this procedure by field acquisition of hyperspectral signatures through different arid and rural open paths in the Negev desert, Israel. An open path of 180 m provided a test case for imaging during clear and stable ambient conditions, from which an inherent measurement error of ∼4% was estimated. Imaging through a very long open path indicated an uppermost open path limit of about 4 km, resulting from a significant attenuation of the sensor's spectral response. Imaging and aerosol optical thickness retrieval under common environmental conditions through urban-scale open paths of about 1 km in Haifa is also demonstrated.

Copyright 2012 American Association for Aerosol Research

1. INTRODUCTION

Airborne particulate matter (PM) attributes have a key role when assessing its environmental, ecological, and public health effects. The size of ambient PM spans over multiple scales and is usually characterized by a bi- or a tri-modal lognormal size distribution, including an ultrafine mode with count median diameter (CMD) <0.1 μm, a fine mode with particle diameters between 0.1 and 2.5 μm, and a coarse mode with particle diameters ranging between 2.5 and 10 μm (Friedlander 2000). Ambient aerosols originate from different sources and by different mechanisms. Hence, they experience varied particle size distributions (PSD), chemical composition, and morphology (Kokhanovsky 2008). Since exposure to PM with CMD⩽2.5 μm has been associated with adverse health effects (Pope et al. 2002; Laden et al. 2006), different methods for evaluating size- and compositional-resolved PM concentrations were proposed (Mcmurry 2000). Given that airborne PM is a dynamic phenomenon with a rapid spatiotemporal variation, remote sensing (RS) is a promising approach for retrieving size-resolved PM concentrations along ambient air columns.

Both satellite and ground RS of aerosols are usually performed through vertical or oblique atmospheric columns, providing the aerosol optical thickness (τ _a ) at different visible-NIR wavelengths (0.4–2.2 μm) as a measure of the integrated aerosol concentrations along the open path (Kokhanovsky et al. 2007). This wavelength span is the most sensitive spectral range for retrieval of the fine aerosol particles (Bohren and Huffman 1983). When the particles’ emission (due to their temperature) is negligible and when atmospheric scattering into the open path can also be neglected, the relationship between the aerosol spectral optical thickness (i.e., the wavelength dependent extinction through an open path of length L) and the aerosol extinction coefficient, β(λ) is

where λ is the wavelength, I ₀ is a reference radiance, and I_L is the residual radiance at the end of the open path. The macroscopic representation of Equation (1) is the Beer–Lambert law

The extinction coefficient, β(λ), is a Fredhölm integral of the extinction cross-section, C_e (μm²), and the number size distribution of the aerosol, n(d_p ) (cm⁻³ μm⁻¹), with the extinction cross-section being a function of the wavelength, the particle diameter, d_p , and its refractive index, m(λ)

Retrieval of size-resolved PM concentrations from optical thickness by means of an inversion of the Fredhölm integral is an ill posed problem (Twomey 1977) that can be improved using constraints obtained from hyperspectral (HS) data. Moreover, HS imaging spectroscopy supports multivariate statistical analysis (correlations, least squares regression methods, etc.) that improves the signal to noise ratio and the unmixing of signatures (Goetz 2009). It has been noted (Alakian et al. 2009) that the spectral response of ambient aerosols at the visible-NIR wavebands (0.4–2.5 μm) shows only moderate variation and, therefore, many features of the hyperspectra in this range (i.e., possible spectral variables) are expected to be redundant. However, different ambient PM concentrations and composition may generate partially overlapping signatures (Etzion et al. 2010), which call for a strategy of nonlinear curve fitting and favors multivariable analysis. Moreover, optimally, to obtain the best retrieval the spectral range for analysis should be dynamic and change according to the characteristics of the true environmental (i.e., aerosol) conditions.

Many studies in the past 30 years aimed at retrieving bi- or tri-modal atmospheric aerosol PSD from spectral optical thicknesses in the UV-visible and the NIR ranges, acquired by satellite and airborne mounted instruments (Heintzenberg et al. 1981; Ferri et al. 1995; Wang et al. 1996; Franssens 2001; Kocifaj and Horvath 2005; Kuzmanoski et al. 2007) or by ground-based radiometers (Dubovik et al. 2002; Wang et al. 2002). In general, correlations between spectral optical thickness and ground-level fine PM concentrations are highly dependent on site-specific attributes, seasonality, and meteorological conditions such as the relative humidity (RH) and mixing layer height (Schäfer et al. 2008). In particular, Evgenieva et al. (2009) used LIDAR measurements to show extensive diurnal changes in the vertical distribution of the aerosol concentration over an urban area, in relation to the mixing layer height. Consequently, aerosol optical thickness products from satellite sensors are biased toward the specific daily revisit time as well as toward clear sky and dark surface scenes (Levy et al. 2009). Hence, these products do not provide representative aerosol products for every scene, in particular urban or bright land scenes. Ground RS sites of the AErosol RObotic NETwork (AERONET), which provide data for validation of the satellite products (Holben et al. 1998), are sparsely distributed and are not suitable for monitoring spatial changes in air quality at an urban-scale resolution. In addition, most of the currently used aerosol sensors (excluding LIDAR) rely on solar radiation and are, therefore, restricted to daytime. In contrast, ambient PM_2.5 exhibits peak concentrations not only during daytime but also during the evening (Degaetano and Doherty 2004; Freiman et al. 2006; Zhao et al. 2009). These considerations suggest that for the purpose of public health applications, fence line surveying, etc., complementary ground RS methods are required. In particular, environmental health applications call for methods which will provide ground-level size-resolved PM concentrations at the urban scale, and be able to track the diurnal patterns.

Several studies addressed the potential of ground spectroscopy along horizontal open paths by applying active RS and analyzing the reflection spectra of the system's known light source after it transverses the open path back and forth. Such systems have a dedicated reflector at the periphery of the (relatively short) open path (Paganini et al. 2001; Müller et al. 2005; Varma et al. 2009). In contrast, here we portray the first steps toward the development of a novel passive RS approach for ground imaging spectroscopy procedure with high temporal resolution, applicable for night-time (i.e., dark scenes) urban monitoring of fine PM. Unlike the active RS approach, the new approach is based on measuring emission spectra of common illuminating sources in a nocturnal scene after it traverses once the open path. Moreover, applying HS imaging rather than spectroscopy offers the advantage of simultaneous acquisition of signals from many sources, whose emitted light travels along different open paths of varying length and is captured by different pixels in an image. As a first step in the development of the method, we present here a procedure for analyzing emission signals from one source, i.e., one target in the image. The experimental section is followed by a results section that emphasizes the potential and the limitations of the proposed procedure. An actual field example illustrates changes in the spectral response as a result of varying ambient PM concentrations. Here, we only demonstrate the methodology for acquiring the HS signature and deriving the apparent aerosol optical thickness from it. The retrieval of PM from the apparent spectral aerosol optical thickness will be reported in a subsequent article.

2. EXPERIMENTAL METHODS

A field setup for night-time HS imaging of light source emission signatures in the visible-NIR range, in contrast to standard daytime reflectance spectra acquisition, was developed. Procedures to overcome difficulties specific for night imaging, including heterogeneous intensities of the sources and lack of distinct contours in the overall scene, are presented. A fully controlled analogous laboratory setup was used for radiometric calibration of the sensor's response to a reference radiation source. The spatial and temporal variability of the pixel intensities across the beam area was examined under clear atmospheric conditions at both a short-range open path (intermediate length) and a full-scale (very long) open path. These field campaigns were used to assess the stability of the reference signal and the variability of the target pixels, and for developing an algorithm for selecting pixels of stable signal and high signal-to-noise ratio. Results of these steps were used for calculating the experimental error propagation in the apparent τ_a(λ), and for estimating, based on Mie scattering, the sensitivity of the procedure to ambient aerosol loadings.

2.1. Instrumentation

Hyperspectral imaging was exercised by a Sagnac-like interferometer (Applied Spectral Imaging Ltd., Migdal HaEmek, Israel) that utilizes a VDS Vosskühler Cool-1300Q camera. The camera consists of a Peltier-cooled (−20°C) progressive scan CCD sensor (Sony Exview HAD technology) that fits an exposure range of 2 ms to 1000 s. The sensor is sensitive to radiation between 300 and 1100 nm (180 channels) and produces spectral cubes of 1280 × 1024 pixels with 12-bit radiometric resolution. The acquisition is based on an interferometer optics and a fast Fourier transform algorithm that generates signals with spectral resolution of 1–9 nm full width at half maximum (FWHM). The camera was equipped with a Nikon NIKKOR telephoto lens of 105 mm focal length, f, and 23° 20′ field of view (FOV). The output spectra are in digital numbers (DN).

A spectroradiometer (Spectral Vista Co. GER 2600, Poughkeepsie, NY, USA) was used in the laboratory experiments to characterize the source radiance and to calibrate the digitized signal of the camera to radiance units. Radiance spectra were measured between 325 and 1050 nm (1.2 nm FWHM, 512 channels) and between 1050 and 2500 nm (8–12 nm FWHM, 128 channels) using an aperture of 3° FOV. An aerosol mini spectrometer (Grimm 1.108, Ainring, Germany) was used to simultaneously measure the ambient aerosol number concentrations in 15 size channels between 0.3 and 20 μm. During the short-range field measurements, sampling with the aerosol spectrometer was done near the HS camera and further away from it, along the open path. Due to technical as well as topographical constraints, sampling with the aerosol spectrometer during the long-range field measurements was done only near the HS camera. Data from an onsite portable meteorological station (temperature, RH, and wind speed and direction) were also collected. Both the aerosol spectrometer and the meteorological station recorded 1 min average readings. A global positioning system (Garmin GPSmap60CSx, Southampton, UK) and a total imaging and scanning station (Topcon Imaging Station IS-9003, Livermore, CA, USA) were used to measure the locations of the camera and the target in the field campaigns. A halogen source of 500 W was used as the illuminating target in both the laboratory and field setups, since it produces a broad and continuous emission in the visible and NIR range. The halogen spectral radiance demonstrated typical black body emission spectra between 350 and 2500 nm with a prominent peak at 1260 nm, which corresponds to temperature of 2300 K. The use of a dedicated source allowed for a simple scaling between the field and the laboratory conditions, and a convenient monitoring of the source radiance for testing its long-term stability between imaging sessions.

2.2. Imaging Procedure

Three setups with open paths of different lengths were used to image the halogen source: a 2 m laboratory scale, a 180 m short distance field scale, and a 4.4 km long distance field scale. The exposure conditions were adapted for imaging artificial light source emissions, i.e., opted to minimize the sensor saturation while eliminating modulations caused by the electricity network frequency (50 Hz). Accordingly, during the full-scale imaging, the exposure time and the aperture size of the lens were set to 50 ms per interferometer step and f/16, respectively, ensuing overall acquisition time of 44 s per image (including file construction). Exposure time and aperture size in the short-range field experiments were fixed at 20 ms and f/22, respectively, to minimize the CCD saturation. Images in the laboratory were acquired at the full-scale exposure conditions via a procedure that is detailed below. Equal acquisition time was set for the radiometric measurements, with 64 scans per spectrum.

2.2.1. Laboratory Setup

The reflectance spectra of the halogen lamp were imaged under controlled laboratory conditions, providing data for radiometric calibration and for first estimates of the method sensitivity (by testing the signal repeatability and stability). In the laboratory setup, the halogen source was placed between the sensor and a Spectralon target—a flat standard of known reflectivity (0.25 × 0.25 m²)—facing the target (Figure 1). The target reflected the halogen radiance toward the sensor, which was set 215 cm away from it. A range of signal intensities was obtained by varying the distance of the source from the reflecting standard. Specifically, five source-target distances were sampled for sensitivity analysis and for the radiometric calibration. The acquisition of reflectance spectra in the laboratory setup imitated acquisition of emission spectra in the field setup while avoiding pixel saturation in the spectral image. The reflecting standard was imaged through a 40 cm specially designed transparent chamber (Figure 1) that was continuously purged with dry filtered air. The chamber was later filled with known size-resolved aerosol concentrations to simulate typical ambient optical paths. The entire system was set inside a dark hood with all the components aligned to a common plane using laser pointers. Alternating measurements by the HS camera and the spectroradiometer were applied for each setup.

FIG. 1 Side view (top) and top view (bottom) of the laboratory setup for spectral imaging and radiometric measurements (dimensions in cm). A: the HS sensor (either the HS camera or the spectroradiometer); B: a transparent aerosol chamber; C: halogen source; and D: a reflecting standard.

2.2.2. Field Setup

Field measurements were used to develop a RS procedure for acquiring spectral signatures through urban-scale open paths. Before studying the relationships between ambient PM concentrations and the emission spectra attenuation, the effect of the imaging geometry (i.e., the sensor-target location and orientation) on the spectral signature was studied. Field night scenes were imaged in the Negev desert near Sede Boker (30° 53′ N, 34° 47′ E), about 40 km south to Beer Sheva, Israel. This sparsely populated arid area was chosen to enable imaging through different yet comparable open paths, and to minimize the variations in scattering coefficients of the aerosols in the open path. Specifically, the arid site conditions preclude possible increase in aerosol scattering due to hygroscopic growth and water vapor condensation (Day and Malm 2001) as the temperature drops during night-times. Short open path (180 m) field imaging of emission spectra was conducted along the edge of the Zin valley near Sede Boker (Figure 2). Long open path (4.4 km) field imaging was practiced from Sede Boker (30° 51′ N, 34° 47′ E, 468 m above sea level [a.s.l.]) toward Tsiporim Junction (30° 50′ N, 34° 44′ E, 504 m a.s.l.).

FIG. 2 Aerial photographs of the field experiments. (a) The short (180 m) open path stretched between the sensor location (A) and the target location (B) along the Zin valley. (b) The long (4.4 km) open path stretched between the sensor at the Ben-Gurion Heritage Institute (C) and the target location near the Tsiporim junction (D). (Color figure available online.)

Quasi-simultaneous measurements by the aerosol spectrometer were carried out at the target (halogen source) and the sensor (HS camera) locations in an alternating fashion, due to availability of only one device. Meteorological parameters were measured during the imaging sessions by a meteorological station in Sede Boker. Atmospheric aerosol loadings (daytime records) were obtained from the AERONET site in Sede Boker.

2.2.3. Radiometric Calibration

Radiometric spectra were acquired by the spectroradiometer and fitted to 169 wavelength channels (326–1102 nm) of the HS camera by a cubic spline interpolation. A least squares regression was performed for each wavelength channel based on five averaged spectroradiometer radiance intensities and their corresponding averaged HS camera output (DN). Five repetitions per data point and about 110,000 pixels per image were used for calculating the averages. The different radiance intensities were generated by placing the halogen source at five different distances from the reflecting surface.

2.2.4. Selection of Pixels for Analysis

Artificial light sources such as commercial halogen spotlights and street lights are characterized by an extremely high emission at the center of the beam that sharply decays toward its edges. Therefore, a procedure for pixel selection was developed to avoid “dead” (oversaturated) pixels that normally reside at the center of the beam, and facilitate selection of pixels of considerable intensity. The procedure involves three consecutive steps: (a) an unsupervised classification of the pixels by a k-means algorithm (k = 10, 5% threshold, five iterations), (b) sub-selection of pixels with the highest intensity by a nearest neighbors (kNN) classification (eight neighbors, one standard deviation constraint), and (c) removal of outliers that either show low values of cross-correlation or have RMS values above the CCD noise level in the spectral range of 300–400 nm. Counterpart spectral analysis was programmed in MATLAB (version R2008b, MathWorks). Details on the k-means and kNN algorithms can be found in Seber (1984) and Brereton (2003), respectively. A consistent physical segment of the target was selected throughout the imaging to avoid spectral discrepancy by the heterogeneity and anisotropy of the halogen beam. For this purpose, all the images were referenced image-to-image to ensure consistent pixel selection. Referencing of night scenes was obtained by setting small light indicators at fixed locations within the imaged scene. ENVI software (version 4.2, Research Systems Inc.) was used for image referencing and for initial classification steps.

3. RESULTS

3.1. Radiometric Calibration and Sensitivity Analysis

Linear regression coefficients (R ²>0.85), namely the intercept b ₀(λ) and the slope b ₁(λ), formed a lookup table and revealed a fairly constant and stable response of the CCD to radiance between 500 and 900 nm (Figure 3). The coefficient of variance (CV) of b ₁ ranged between 3 and 4.4%, and the nominal standard deviation of b ₀ was 70–370 × 10⁻¹⁰ W m⁻² nm⁻¹ sr⁻¹, about 4.7–6.7% of the halogen radiance. Beyond the 500–900 nm spectral range, the variance increases because of the rapid decay of halogen radiance below 500 nm and the attenuation in the instrument line shape (ILS), i.e., the sensor sensitivity, above 900 nm. Hence, the 500–900 nm spectral range was set as the operational range of the HS camera. It should be noted that due to the limited number of true measurements, standard deviations were estimated based on 2000 bootstrap samples after verifying a close to normal distribution (Efron 1979).

FIG. 3 Radiometric calibration coefficients derived from the least squares regression (average ± standard deviation). Top: the slope b ₁(λ); bottom: the intercept b ₀(λ). (Color figure available online.)

The lookup table was applied to 10 cross-validation sets and the root mean squared error (RMSE) was 5.2–7.8% of the radiance in the 500–900 nm operational range. Following Tailor's (1997) general error propagation scheme and considering the standard deviation of the DN signals and of the calibration coefficients, the propagation of the radiometric calibration error in Equation (1) results in a spectral error of Δτ _a (λ)=0.121−0.187 in the 500–900 nm operational wavelength range (Figure 4). These results provide a first estimate of the HS camera sensitivity to spectral attenuation due to ambient PM. Specifically, low visibility episodes that are characterized by aerosol optical thickness >0.15 should be detected although the HS imaging system is less sensitive in the 400–460 nm range than the spectroradiometer (Utrillas et al. 2000). Moreover, in view of these results, the DN output of the HS camera can be directly used in the Beer–Lambert Equation (2) to retrieve the aerosol optical thickness in the operational spectral window. It is noteworthy, however, that the radiometric calibration was performed for a specific set of exposure conditions, which were carefully selected to fit the full-scale imaging experiment (Section 3.3). It is appreciated that different field setups may require different exposure conditions and, therefore, some corrections to the calibration coefficients.

FIG. 4 Nominal error of the acquired aerosol spectral optical thickness, Δτ _a , due to the radiometric calibration errors.

3.2. Short-Range Field Measurements

The short open path (180 m) night scene experiment setup was performed by placing the target halogen source and marginal referencing spotlights at the Ben-Gurion Heritage Institute (Figure 5). Under these conditions, the apparent cross-section of the halogen beam was about 800 pixels, in comparison with about 20 pixels when the target was imaged through a long open path of 4.4 km. Yet, the short open path images were characterized by saturation at the center of the emission target, resulting in “blind” pixels. The saturated core (black pixels in Figure 6) was identified by a k-means classification and removed from the initial subset for further analysis. The final pixel subset, after a kNN classification and an outlier removal algorithm (Section 2.2.4), contained about a third of the initial set and was used for assessing the repeatability and consistency of the time-series imaging.

FIG. 5 (a) Night-time and (b) daytime scenery of the short (180 m) open path field imaging at the Ben-Gurion Heritage Institute. The center of the halogen source (black core) is marked by a cross in the night-time scene. The four peripheral registration spotlights are arranged in a quadrilateral shape around it. (Color figure available online.)

FIG. 6 Classification of pixels for analysis (ring arrangement). (a) Selected pixels after the k-means classification step (k=10, 5% threshold, 5 iterations, 313 pixels). (b) Sub-selection of pixels with high intensity by the kNN procedure (k=8, SD=1, 231 pixels). (c) Final selection of the subset of pixels used for analysis after outliers removal (106 pixels). (Color figure available online.)

FIG. 7 Time-average (black line) ± one standard deviation (gray lines) of the halogen emission spectra during the short (180 m) open path field imaging in Sede Boker, the Negev desert, on 20 May, 2009. For comparison, a representative halogen reflectance spectrum from the calibration experiment is also shown (gray dotted line).

FIG. 8 Average (dark solid line) and minimal and maximal (light line) size-resolved PM concentrations measured by an aerosol spectrometer during the short (180 m) open path field imaging in Sede-Boker on 20 May, 2009.

A time series of images was acquired on 20 May, 2009, over a short interval of 20 min between 20:30 and 20:50, to minimize ambient variability, with all the images referenced to the first acquired (reference) image. At each time point, an average spectral signal of the final subset of pixels that survived the above analysis was calculated. The average emission signal of each image was repeatable and had a CV smaller than 4% in the 500–900 nm range, similar to the CV of the spectra in the laboratory experiments. Moreover, whereas the emission signal of the halogen source is similar to the reflectance signature acquired in the calibration setup (Figure 7), it is somewhat narrower and shifted to longer wavelengths. This most likely reflects sampling of a specific spatial segment in the non-homogeneous and non-isotropic emission signature of the light source compared with the homogeneous reflectance signature (since each pixel in the reflectance signature averages the radiance it obtains from all the emission pixels). The aerosol spectrometer measured during the imaging session particle number concentrations of 36.3–39.7 cm⁻³ and 0.79–0.93 cm⁻³ in the size ranges 0.3–2 μm and 2–20 μm, respectively (Figure 8). Calculations of Mie scattering by homogeneous spheres (Bohren and Huffman 1983) with a spectral refractive index of dust (WMO 1986) indicated that such concentrations contribute τ _a (λ)<0.02 (500⩽λ⩽900) along the 180 m open path. Since the RH and the temperature varied only marginally during data acquisition (RH: 45.3–52.6%, T: 21.6–23°C), the short open path field experiment represents imaging through relatively clear and stable atmospheric conditions (wind velocity at 3.5 m above ground level [a.g.l.] ranged between 0.85 and 2.75 m s⁻¹). In particular, variation of the PSD and of the corresponding scattering coefficients as a result of aerosol hygroscopic growth are insignificant for (Day et al. 2000; Day and Malm 2001).

TABLE 1 Meteorological and PM data recorded between 19:00 and 19:30 at the Neve Shaanan air quality monitoring station, Haifa, during the three 2010 selected events (3 May, 20 June, and 18 August). The station is located in close proximity to the open path through which the HS imaging was performed

	3 May, 2010	20 June, 2010	18 August, 2010
PM2.5 (μg m^-3)	13.4–15.3	35.7–47.5	31.0–31.7
PM10 (μg m^-3)	27.3–34.6	112.7–203.3	63.5–67.1
RH (%)	82.6–83.5	49.8–54.1	88.4–88.6
Temperature (°C)	18.4–18.5	27.3–27.5	28.3–28.4
Wind velocity (m/s)	2.51–2.57	0.58–0.65	1.72–2.02

3.3 Full-Scale Field Measurement

A time series of images of the halogen source emission spectra through a long open path (4.4 km) was acquired on May 21, 2009, between 20:10 and 20:50. In comparison with the short open path experiment, a higher particle number concentration in the 0.3–2 μm size range (121.8–146.7 cm⁻³) was measured by the aerosol spectrometer during the long open path experiment. Based on Mie scattering calculations, the corresponding aerosol optical thickness varied between 0.25 and 0.4 for 900 nm and 500 nm, respectively. The consistency of PM_2.5 concentrations during the imaging suggests that the temporal variation of τ _a in the 500–900 nm operational range did not exceed 0.05, i.e., was within the inherent noise of the sensor. The RH (63.2–69.1%) and air velocity at 3.5 m a.g.l. (2.78–5.5 m s⁻¹) were higher than in the short open path experiment, whereas the ambient temperature was somewhat lower (20–21°C). Moreover, the temporal variation of the ambient conditions during the long open path measurements did not seem to introduce considerable changes in the target spectral signatures.

Validation of the consistency of the acquisition and of the pixel selection was performed following the procedure applied for analyzing the short open path time-series. The substantial light extinction that was evident in the full-scale open path measurements circumvented “blind” pixels but, on the other hand, provided fewer and less homogeneous pixels suitable for further analysis. Regaining the sensitivity that was lost with the increase in the open path length is possible by increasing the exposure time, but at the cost of obtaining, again, saturated pixels in the center of the beam, as demonstrated for the 180 m open path.

Two distinct pixel groups were distinguished among the high emission pixels that were selected for analysis: pixels with clear heterogeneous signatures and pixels with relatively homogeneous signatures (Figure 9). These groups of pixels were associated with different beam segments and exhibited spectral signatures of comparable intensities, including attenuation between 680 and 820 nm, which is typical of absorbance by water vapor and oxygen (Schermaul et al. 2001; Petty 2006). It should be noted that according to Alakian et al. (2009), the spectral signature is unaffected by atmospheric gaseous compounds outside their respective absorption bands. Figure 9 depicts the average signals of both classes (total of 21 pixels). Like in the short open path measurements, a CV of about 5% was obtained in the 600–850 nm spectral range. The narrowing of the operational range in the full-scale open path experiment results from the much lower emission intensity experienced by the HS sensor, and implies on a distance constraint of the method.

FIG. 9 Time-series imaging of the halogen emission through the long (4.4 km) open path in Sede Boker on the evening of 21 May, 2009. Inset: pixels with a relatively high emission used for analysis; 1: pixels with a heterogeneous signature; 2: pixels with a homogeneous signature. Averaged signals of pixels from both classes at 20:10 (black solid line), 20:30 (dark gray line), and 20:50 (light gray line). (Color figure available online.)

Further, night imaging was carried out at the Technion, Haifa, Israel (32° 47′ N, 35° 03′ E), where halogen emission imaging was exercised through an open path of 1 km under different ambient conditions. Due to space limitations, a complete account of this urban setup and the corresponding signal analysis will be discussed in a subsequent article; however, it is noteworthy that the HS sensor responded to changes in PM concentrations, as demonstrated below.

Emission signals were acquired during selected evenings between May 2010 and August 2010. Table 1 presents meteorological and PM data that were recorded at a nearby air quality monitoring station during the three selected events (3 May, 20 June, and 18 August). PM_2.5 and PM₁₀ concentrations, as well as the ambient temperature, were significantly higher during the June and August events than in May, whereas the RH in June was much lower than in May and August. All three events were characterized by low wind speed and stable meteorological and PM conditions. The HS imaging (f/22 aperture size, 40 ms) yielded a segment of 26 pixels of relatively homogeneous signatures, amenable for further processing. Considering the stable environmental conditions, we compared the event-averaged signals in the 600–850 nm spectral range and found that the signatures in June and August were attenuated relative to the signature in May, as expected based on the doubling of PM_2.5 between these events. Figure 10 illustrates the spectral attenuation in the June and August emission signatures (I ₂) relative to the May emission signature (I ₁). The spectral signatures in both June and August show a declining trend above 650 nm. Based on Mie scattering, this trend indicates a dominant <0.5 μm PM size mode, like many aerosols commonly found in the east Mediterranean (Dubovik et al. 2002). Nonetheless, the different signatures in June and August suggest that the spectral attenuation in the two events resulted from aerosols with different size distributions. This conclusion is supported by the highly different PM₁₀ concentrations and RH levels in the two events but does not exclude additional effects of different chemical composition, internal structure, and shape. Further analysis of these results will be discussed in a subsequent article that will address PM retrieval from τ _a (λ).

FIG. 10 Attenuation of the halogen emission (I ₂) through a 1 km open path in Haifa on the evening of 20 June, 2010 (gray line) and on the evening of 18 August, 2010 (black line) relative to halogen emission on 3 May, 2010 (I ₁).

4. DISCUSSION

Utilizing artificial light sources, as demonstrated here for a standard halogen source, needs some specific and careful considerations to account for the intensity variability, the small target size, the multiple illuminating sources that oftentimes exist in the imaged scene, etc. Specifically, we assessed the capabilities and limitations of the source–sensor interactions and of the geometrical effects of the imaging setup under controlled (lab), stable (desert), and typical urban (Haifa) ambient conditions. The response of the HS sensor was found to be linear with the radiance, and its sensitivity in the setup studied declined rapidly beyond the spectral range of 500–900 nm. The latter mainly resulted from the diminishing halogen emission spectra at lower wavelengths, and from the attenuated CCD response above 900 nm. As a result, measurements with the current system are restricted to this range. Error propagation analysis of the radiometric calibration revealed that PM variations that correspond to changes of 0.12–0.18 in the aerosol optical thickness should be discerned. Indeed, the HS sensor demonstrated a clear response at high ambient PM loadings over 1 km urban open path. Nonetheless, the inherent noise (∼4%) found in both the laboratory and the short open path field measurements must be appreciated when retrieving ambient PM concentrations. Future improvements are expected to enhance the sensor's sensitivity.

Night-time imaging based on artificial light sources is a challenging task, since the emission is neither homogeneous nor isotropic. Hence, it is highly important to accurately reference the images and to analyze the same physical segment of the light beam. A reliable procedure for selecting the pixels of highest intensity while avoiding saturation effects was developed. Night scenes lack natural contours for image referencing, therefore, a dedicated arrangement of point-size illuminating markers is required. Nonetheless, if street lights are used for such a purpose their spatial arrangement may make this need redundant (Etzion et al. 2009). Another referencing strategy we studied was to image the exact scene during the day and use it as a base for the night imaging. Clearly, this technique is more suitable for urban scenes, and has been successfully applied by us in Haifa, Israel, for imaging different illuminating targets through long open paths in both residential and industrial surroundings.

The long open path imaging in the Negev desert revealed that due to large attenuation of the halogen source emission, an open path of ∼4 km is the uppermost limit for our setup. Furthermore, due to FOV considerations, imaging the halogen source from 180 m resulted in about 25 times more pixels suitable for analysis than when imaging through the 4.4 km open path. Hence, the possibility to reduce the inherent sensor noise by averaging spectra of different pixels is very limited for very long open paths. We are currently exploring ways to improve the SNR, e.g., by changing either the lens or the source configuration. Specifically, applying better telescopic lens and zoom lens may provide higher focal lengths and changeable FOV, whereas imaging a source of higher intensity and using special diffusing element may increase the pool of pixels for analysis. Moreover, even with the current system's 4 km open path limit, such a span is actually very useful for urban-scale measurements, since it can capture neighborhood scale spatiotemporal variations, with the spatial variations captured by azimuthal rotation of the system toward different emission sources. This suburban spatial resolution perfectly matches the requirements of modern environmental epidemiology studies, which are looking for better ways for assessing the exposure variability within the city.

5. CONCLUSIONS

Urban-scale ground RS is a promising approach for monitoring fine ambient PM at high spatiotemporal resolution. Current RS methods for estimating the aerosol spectral optical thickness focus on vertical column observations of the entire atmosphere during daytime, and provide estimates of τ _a (λ) that are not easily related to ground-level PM concentrations in vast regions worldwide. However, by applying the same method, i.e., analysis of spectral scattering and attenuation of radiation in the visible-NIR range, through quasi-horizontal atmospheric layers near the ground, size-resolved aerosol concentrations and some physicochemical characteristics of these particles (by means of their refractive indices) can be deduced. As a first step in this direction, we present here a portable HS camera system that operates at the visible-NIR spectral range and can potentially be used for spatiotemporal monitoring of fine PM concentrations through urban-scale ambient open paths.

This work describes the general scheme and the required steps for acquiring the spectral signature of the apparent aerosol thickness in the visible-NIR range of an urban-scale open path, based on irradiance of a ground halogen target. In contrast to PM retrievals that are based on reflectance spectra of the sun irradiance, we focus on night-time (dark scene) imaging. The aerosol spectral optical thickness that can be obtained from the ground HS imaging system, as demonstrated in this work, provides the input for retrieving fine PM characteristics by means of a novel algorithm, which will be described in detail in a subsequent article.

Ground imaging spectroscopy shows many advantages that have not been studied or utilized to date, with a special potential for identifying episodes of high PM loadings, such as dust events, biomass smoke plumes, and heavy air pollution episodes.

Acknowledgments

Y.E. was supported by the Eshkol Fund of the Israel Ministry of Science and Technology. We wish to express our thanks to Midreshet Sede Boker that facilitated the measurements, Mr. David Klepach from The Jacob Blaustein Institute for Desert Research for providing the meteorological data, and Wolfgang Schwanghart for the total station measurements and for assisting in the night scene setups. The study was partially supported by the Technion Center of Excellence in Exposure Science and Environmental Health (TCEEH).