Atmospheric particulates can directly affect Earth's energy balance through the scattering and absorption of light , as well as serve as cloud condensation nuclei (CCN), leading to changes in cloud properties and precipitation patterns . The most recent IPCC assessment identifies both aerosols' direct and indirect effects as the two largest uncertainties of all anthropogenic radiative forcing components . Space-based remote sensing and model simulations yield aerosol data with global coverage but the assumptions and validation of these techniques require additional information. A wide range of methods exist to provide these additional constraints, with two of the most accurate approaches being airborne in situ measurements and ground-based remote sensing . The latter approach, specifically in the form of a global array of sun photometers known as the Aerosol Robotic Network (AERONET), has been used extensively to characterize aerosols and validate space-based observations. While this method is remarkably accurate, it inevitably contains limitations that many in situ measurement techniques do not share. For example, the parameters provided by the AERONET retrieval are limited at low optical depths and the technique cannot provide the vertically resolved aerosol properties that can be obtained by airborne in situ instrumentation.

The limited number of angles sampled by typical passive satellite sensors, along with the need to correct for surface reflectance, means that inversions designed to retrieve aerosol properties from these measurements require significant assumptions about the aerosol in question. These assumptions frequently take the form of a set of aerosol types (desert dust, biomass burning (BB), urban emissions, etc.) with predefined characteristics that can be used to estimate the optical properties of the aerosol. While current algorithms are adequate to retrieve aerosol optical depth (AOD) and a few other parameters, the results from a wide range of studies have suggested that there may still be room for significant improvements in the aerosol properties assumed in space-based remote sensing retrievals. For example, localized tests using modified, more locally appropriate aerosol models have shown significant improvements in comparisons with AERONET-derived AOD as well as the ability to increase spatial resolution with little cost to retrieval accuracy .

Models used to calculate aerosol forcing and to estimate climate change also rely on assumptions regarding aerosol optical properties. In fact, comparison of nine widely used aerosol forcing models found that the greatest diversity in model estimates of forcing were not in the representation of aerosol loading by the models but in the forcing efficiency, the forcing per unit of loading . The forcing efficiency is affected by wide-ranging values of aerosol optical properties found in the models . A revisit of this model comparison, now involving 16 models, published 7 years later found no narrowing of model diversity in estimates of aerosol radiative effects and forcing and significant diversity when analyzing individual aerosol components . Again, the reason was traced to a significant range of values for factors such as forcing efficiency that stem from lack of constraints in basic aerosol intrinsic properties .

The aerosol characteristics used in passive remote sensing algorithms and climate modeling have been primarily based on inversions of AERONET sky radiance measurements that produce values of total column ambient aerosol properties . Constructing aerosol models from these inversion data requires calculating statistics of the quantities for different groupings of the data corresponding to different aerosol types or classifications. Classifications can be identified using a priori knowledge of dominant aerosol types in different locations , or by using advanced statistical methods such as cluster analysis , recently utilizing Mahalanobis distances . These techniques have also been applied to other high-quality aerosol remote sensing datasets such as High Spectral Resolution Lidar (HSRL) or Multiangle Imaging SpectroRadiometer (MISR) to classify aerosol into dominant types and to derive aerosol models for each type . The advantage of using remote sensing datasets to construct aerosol models for remote sensing or climate applications is to have a set of aerosol models producing a radiance at the top of the atmosphere consistent with the radiance a satellite would measure or that affects the planetary energy balance. The disadvantage of using such datasets to construct aerosol models is that detailed particle information is lost due to ambiguities concerning humidification and height of the particles, and such models cannot be easily linked to particle composition.

In situ measurements of aerosol's optical properties have the potential to provide additional information, like aerosol vertical dependence and particle hygroscopicity, while maintaining an optical consistency comparable to remote sensors. In practice, though, the majority of in situ optical measurements have been limited to parameters, like total scattering, absorption and extinction, which lack the angular information required for a comprehensive aerosol model. Polar nephelometers have occasionally been used to measure the angular dependence of light scattered by ambient aerosols but measurements by these instruments above ground level have been extremely limited. The first published measurements from an airborne polar nephelometer were made in the early 1970s . These measurements were quite advanced for the time but the results presented by the authors are limited to only two phase function averages and lack the context of modern aerosol science. The next polar nephelometer, designed specifically for the measurement of aerosols, to be flown aboard an aircraft was the Polarized Imaging Nephelometer (PI-Neph), which has been active since the fall of 2012 . This instrument has sampled extensively during recent field campaigns and, through the use of modern technology, is able to make significantly more advanced measurements than early polar nephelometers. In addition to the two aerosol instruments, a third polar nephelometer originally designed for the measurement of cirrus clouds has been flown extensively since the 1990s. Recently, this instrument has been modified to allow for measurements of some aerosol particles as well as ice crystals, specifically those with diameters larger than 1 µm and sufficiently higher number concentrations. The results of these measurements are presented by , who performed principal component analysis (PCA) on the light-scattering data to explore the properties of aerosols found in volcanic degassing plumes.

This work explores recent airborne measurements made by the PI-Neph and constitutes the first comprehensive analysis of angular-dependent light-scattering measurements made in situ from aboard an aircraft on common atmospheric aerosol types. The analysis focuses on measurements made during the Studies of Emissions, Atmospheric Composition, Clouds and Climate Coupling by Regional Surveys (SEAC4RS) and the Deep Convection Clouds and Chemistry (DC3) field campaigns. The light-scattering data include both phase function (F11) and linear degree of polarization (-F12/F11) measurements ranging from 4 to 174∘ in scattering angle. These measurements are separated into 2390 different averaging periods for which stable, high-quality data were available. A classification strategy is then developed for the SEAC4RS data to estimate the dominant aerosol type for each case, making use of ancillary trace gas measurements, aerodynamic size distributions and aerosol composition measurements. The data from the DC3 campaign, which has significantly different objectives from SEAC4RS, are classified according to the region where the sample was taken. PCA is then applied to the PI-Neph averages to confirm the validity of the ancillary data classification scheme in a light-scattering context and to explore whether there is enough information in the PI-Neph measurements to classify aerosol types from light scattering alone.

The dataset used in this work is built from measurements made aboard the DC-8 aircraft during the DC3 and SEAC4RS field campaigns, with a focus on measurements made by the PI-Neph airborne polar nephelometer. Time averages are performed on the raw PI-Neph data and the resulting cases are grouped into one of eight predefined aerosol categories according to a novel aerosol classification scheme. This classification scheme makes use of measurements that are independent of aerosol light scattering, including particle composition, aerodynamic size distribution and gas concentrations.

An aerosol classification scheme was developed to estimate the dominate source of each aerosol by focusing on the air mass associated with each PI-Neph average, using ancillary data that include measurements of gases, aerosol composition and physical properties as well as aircraft location. Aerosol optical properties were intentionally omitted from all classification metrics to ensure independence between the classification scheme and the scattering features measured by the PI-Neph. As convective systems have the potential to significantly influence aerosol properties different classification schemes were applied to the DC3 dataset (near convective systems) and the SEAC4RS dataset (generally far from convective systems). The SEAC4RS data were subdivided into five categories corresponding to dust, biogenic, urban, BB emissions and unclassified samples. This classification utilized measurements of particle chemical composition from the PALMS instrument, the aerodynamic size distribution of particles generally associated with the coarse mode and a range of trace gases. The scheme developed to categorize the DC3 data was based on aircraft location relative to three storm domains outlined in the DC3 science objectives as well as the presence of convective systems over the course of the corresponding flight. Both the DC3 and SEAC4RS classification schemes only allow one aerosol type to be assigned to a given PI-Neph sample.

The SEAC4RS dust classification requires that the PALMS instrument identify at least 15 % of the measured particles as mineral dust. It was found that the PALMS algorithm would often classify a significant fraction of particles as mineral dust, even when particle size distribution measurements showed no significant coarse mode. In order to exclude these cases and align our dust classification with more traditional dust aerosol we imposed a set of requirements on the coarse mode of the aerodynamic size distributions. Specifically, the volume concentration measured by the APS (which is insensitive to particles below 0.5 µm) must exceed 2 µm3cm-3 and have an effective radius greater than 0.75 µm. This constraint on the aerodynamic particle size distribution removed several cases in which no obvious source of dust could be identified.

If the dust category was not selected gas tracers and aircraft altitude were used to screen for the remaining three fine-mode-dominated types. used a threshold of 250 pptv acetonitrile – or 250 pptb of carbon monoxide if acetonitrile data are unavailable – as an indicator of BB emissions. We have modified this metric to also include cases with acetonitrile values as low as 190 pptv but only if the sum of the volume mixing ratios of isoprene and monoterpenes is less than 40 % that of acetonitrile. Since isoprene and monoterpenes are well correlated with biogenic emissions this condition permits the inclusion of cases with lower BB concentrations, while still avoiding false positives that can potentially be triggered by strong biogenic emissions of acetonitrile. Accordingly, isoprene and monoterpenes are used as gas tracers for the biogenic category, with a biogenic classification occurring when their combined concentrations exceed 2 ppbv. When the previous three categories are not triggered, the aircraft is within or close to the mixing layer (altitude below 3 km) and NO2 concentrations are greater than 1 ppbv the urban category is selected. A marine aerosol classification occurring whenever the aircraft was directly above large bodies of water was also examined, but the scattering intensity during almost all corresponding periods was below the PI-Neph's lower limit of detection. The few remaining marine cases, as well as all cases that failed to trigger any other classification, are identified as “unclassified”. A decision tree specifying the requirements for each SEAC4RS category is shown in Fig. .

The DC3 campaign had significantly different objectives – namely the study of convective systems – and correspondingly the SEAC4RS classification algorithm does not map well to the DC3 dataset. This fact motivated the decision to classify the DC3 samples by study region (CO, AL or TX/OK) as opposed to gas and composition data. A PI-Neph sample was associated with a given study region if the aircraft coordinates were within the corresponding domain and the corresponding flight path was designed to target active storms in the region. Additionally, in order to restrict the assigned cases to storm inflow measurements, the classification was only applied if the observation was made below 6 km. This constraint produced much more homogeneous aerosol properties for each storm domain by eliminating the cases with highly variable scattering properties found in the higher-altitude outflow aerosols.

Table  shows the number of cases the ancillary data classification scheme assigned to each category, as well as the number of unique flights containing at least one of the corresponding cases. Of the SEAC4RS cases, 70 % received a classification (other than unclassified) while 55 % of the DC3 cases were classified. In both campaigns, the majority of the unclassified cases correspond to high-altitude transit legs that are generally associated with relatively clean air masses. All categories have cases originating from multiple flight days, increasing the likelihood that a given category average is representative of the typical aerosol properties found in that type.

Figure  shows the geographic locations of all classified PI-Neph samples. The extent of the three DC3 study regions can be seen in the spread of the red, beige and maroon circles corresponding to the CO, TX/OK and AL storm categories, respectively. In the summer months, biogenic emissions often dominate the southeastern United States while the western portion of the country is frequently influenced by wildfire smoke. While the SEAC4RS categories are not determined by location, clear patterns emerge that are in strong agreement with these physical expectations.

The California Rim Fire was one of the dominant sources of BB emissions during the SEAC4RS deployment. The fire began on 17 August in Stanislaus National Forest, California, and continued to burn until after the end of the SEAC4RS deployment . Before the fire was fully extinguished its total burn area had grown to 104 000 ha, making it the third largest fire in California's history . The arch of BB cases that are shown in the northwestern portion of Fig.  and traverses from California into southern Canada represents samples dominated by Rim Fire smoke. Emissions from this fire made up 58 % of all classified BB cases. The overwhelming majority of the remaining cases came from smaller wildfires within the United States, primarily from three fires located in Wyoming, Colorado and Kansas . While many agricultural fires were sampled in SEAC4RS these measurements were almost always discarded by the averaging algorithm described in Sect.  due to the very short duration, and high variability, of the corresponding measurement.

The 15 points that met the requirements of the dust classification are shown in yellow in Fig. . This type was only observed in early August over Louisiana and the northern Gulf of Mexico. These cases likely correspond to a transported Saharan air layer that was present over this region at the start of the campaign. There is strong evidence, based on aerosol concentration and composition, that this air mass was relatively pristine and had not mixed significantly with continental air .

In August and September biogenic emissions are ubiquitous in the southeastern United States. The classification scheme presented here conveyed this fact well, with most of the cases in this region falling under the biogenic category. The second most prevalent category over the southeastern United States is the urban type. This classification corresponds well to city centers like Houston and Dallas, Texas, whose emissions were frequently sampled by the DC-8. Additionally, a large strip of urban cases can be seen around the Ohio River valley, an area with a very high concentration of fossil-fuel-based power plants. It is shown in Sect.  that the optical properties of aerosols associated with cities are quite different from the industrial emissions of the Ohio River valley. The possibility of dividing the urban category into two subgroups was explored, but the already limited number of cases made this division impractical. It is likely that other datasets, with a larger number of samples corresponding to urban and industrial emissions, can be more easily understood by dividing the urban classification described here into two separated subcategories that are separated by SO2 concentrations, for example.

A robust averaging procedure was applied to all sample averages of F̃11 and -F12/F11 data of a given aerosol type to obtain curves that are typical of each category. Figure  shows the results of this averaging for all three DC3 storm domains as well as the dust category and the average of all three fine-mode-dominated aerosols (biogenic, urban and BB) from SEAC4RS.

A progression in both F̃11 and -F12/F11 averages is evident as the DC3 storm domain transitions from AL to TX/OK to CO. The increase in forward-scattering peak through this sequence suggests an increased scattering contribution from particles in the larger size ranges that direct the bulk of their scattered energy into the forward angles . Models have shown that large convective systems can agitate surface dust, drawing these particles up into the atmosphere and acting as a significant source of dust aerosol . The elevated forward-scattering peaks among the three DC3 categories are likely driven by increasingly arid surface features, leading to an increased availability of this relatively large dust aerosol . Variations in typical storm wind speeds may have also contributed to variations in the quantity of dust that was suspended over a given region.

The same progression is evident in the backscattering angles of the DC3 storm categories, with the CO storms having the largest fraction of the total scattered light that is directed into the scattering angles larger than 90∘. Mie computations were performed in order to identify the size, shape and complex refractive index changes that are potentially driving this progression. In size distribution and refractive index regimes typical of ambient, fine-mode aerosol (rv≈150 µm; σ=0.38; m=1.5+0.01i) the backscattering region of the phase function is very sensitive to the diameter of the particles, suggesting significantly smaller fine-mode particles in the CO inflow than in the AL inflow. A peak in the fine-mode particle size distribution that is shifted toward smaller diameters also produces larger values of -F12/F11 at side-scattering angles, although this effect may be partially moderated by differences in refractive index.

The average F̃11 data, corresponding to the SEAC4RS fine-mode-dominated categories, shows very weak forward scattering and thus suggests relatively few coarse-mode particles. Mie simulations and inversions of the angular-scattering measurements suggest that the reduced -F12/F11 maximum observed in this fine-mode-dominated data is likely driven primarily by fine-mode particles with slightly larger diameters than those found in the three DC3 categories. The features of the dust-scattering matrix elements, specifically the presence of an extremely strong forward-scattering peak, are typical of an aerosol whose scattering properties are dominated by coarse-mode particles . While the typical integrating scattering coefficient for dust was comparable to the other aerosol types, the strong forward-scattering peak significantly limits the amount of light scattered into other angles. The combination of this low absolute scattering intensity and systematic instrument noise resulting from stray light may produce significant biases in the dust F̃11 and -F12/F11 measurement averages at angles above 90∘. It should also be noted that the true forward-scattering peak of the ambient aerosol may be even larger than the values reported by the PI-Neph, whose sample is subject to inlet cutoff effects, which disproportionately effect the largest particles.

The averages of the three SEAC4RS fine-mode cases are examined individually in Fig. . Visually the averages of the three types produce very similar angular-scattering patterns, especially the biogenic and urban averages. The -F12/F11 peak was slightly larger on average in the BB cases, with this feature most clearly separating the BB aerosols from the other two types. Additionally, small differences in the shape of F̃11 can also be observed in the BB averages, where the relative contribution from the forward and backward-scattering peaks are suppressed relative to the other two types.

The variability within a given type's scattering data (not shown) was the highest in the case of the samples associated with urban emissions. Further examination of this variability showed two distinct subgroups, with the conditions around the Ohio River valley differing significantly from conditions near urban centers. The starkest difference between these two subgroups occurred in the -F12/F11 maxima, with significantly higher peaks occurring in measurements made near the Ohio River valley.

It is evident from the results of the previous section that there are differences in the averaged scattering data that agree well with the physical expectations of each aerosol type. While this result is encouraging, the averages alone do not tell us if these differences are characteristic of the majority of samples or are driven by a few relatively extreme cases. The regularity of the geographic patterns observed in Fig.  does suggest a consistent physical basis for the classification scheme in the majority of SEAC4RS cases but it says nothing about the DC3 classification in which sample location is already the primary classification metric. Additionally, as none of the properties used by the SEAC4RS classification scheme are directly related to aerosol optical measurements, it is possible that the geographic distributions observed capture patterns in the air mass properties that are not reflected in the optical features of the corresponding aerosol populations. In order to confidently say that the majority of cases have aerosol optical properties that are clearly characteristic of the corresponding type we must examine PI-Neph measurements on the level of individual cases. Unfortunately, the subtle differences between many of the scattering measurements and the high dimensionality of the dataset complicates a direct analysis of the relevant features. In the following section, this analysis is simplified by reducing the dimensionality of the PI-Neph measurements with PCA. This approach leads to a clear picture of the type-driven clustering of cases that occurs in scattering element space as well as permits easy identification of the features that are characteristic of each aerosol category.

In this section, PCA is performed on all PI-Neph measurement averages to simplify the scattering data and more easily explore its relationship with the classification categories. Intuitively, PCA transforms the data to a new coordinate system in such a way that the greatest variance lies along the first coordinate, the next largest variance lies along the second coordinate and so forth . This process allows most of the dataset's variance to be captured in the first few principal components and in turn the dimensionality of the measurement can be significantly reduced while still maintaining the bulk of the original information content. Mathematically, the basis vectors of this new coordinate system are the eigenvectors of the data's covariance matrix. In this work the orthonormal basis vectors describing this new coordinate system (i.e., the normalized eigenvectors of the covariance matrix) are referred to as “loadings”, and the basis vector coefficients required to represent each data point are referred to as “scores”. It is important to note that PCA is an unsupervised technique and the results are therefore independent of any hypothesis regarding the data, including the ancillary data classification scheme.

PCA was performed on all 532 nm PI-Neph averages from the combined SEAC4RS and DC3 datasets (2334 samples) simultaneously. While the data from the two campaigns were merged, the unpolarized and polarized measurements were kept separate in the final analysis (i.e., the PCA routine was run twice, once for F̃11 and again for the -F12/F11 dataset). Individual F̃11 measurements can often span several orders of magnitude between the forward-scattering peak and side-scattering angles. To prevent the first few principal component loadings from being dominated by the large absolute variations in forward-scattering intensity, the analysis was performed on the natural logarithm of the F̃11 values. This transformation produces a set of principal components in which the first component, for example, explains the largest possible relative variance in data . No transformation was applied to the -F12/F11 measurements. The angular range of the final inputs to the PCA routine was 5 to 170∘ in the case of both the F̃11 and the -F12/F11 datasets. This range corresponds to the angles in which data were present during all measurement periods over both campaigns. Data points at which instrument noise produced nonphysical (i.e., F11(θ)<0 or |F12(θ)/F11(θ)|>1) values were excluded from the analysis.

The decision to treat the intensity and polarization information separately when performing the PCA was based on two factors. The first stems from the fact that most modern measurements of the optical properties of atmospheric aerosol are polarization insensitive. Isolating the polarization information permits conclusions that are more applicable to polarization-insensitive instrumentation while simultaneously helping to illuminate the potential benefits of adding polarization capabilities to future instrumentation. The second factor results from the fact that PI-Neph data contain sources of systematic noise that are strongly correlated over time and scattering angle but are very weakly correlated between F̃11 and -F12/F11. By separating these datasets, the variability in the data corresponding to these systematic artifacts can be more effectively isolated, allowing the remaining components to more accurately capture the physical variation among the samples. This hypothesis is supported by the fact that some of the loadings closely matched correlations in angular errors that are known to result from certain instrumental artifacts. Additionally, a significantly reduced separation of aerosol types in PCA score space was observed when PCA was performed on the intensity and polarized measurements simultaneously. This observation was consistent regardless of the relative weights applied to the F̃11 and -F12/F11 variances.

In this work PI-Neph measurements from the DC3 and SEAC4RS field experiments were sub-selected and averaged over periods corresponding to stable, high-quality data. An optically independent aerosol typing scheme, making use of ancillary data, was developed and the resulting 2390 cases were separated into seven classified categories, as well as an eighth unclassified category corresponding to cases that did not meet any of the classification criteria. SEAC4RS measurements were separated into biogenic, BB, urban and dust types, based on composition measurements from the PALMS instrument, APS data and the concentrations of various gas tracers. The geographic distribution of the resulting classification was in strong agreement with expectations, suggesting a strong physical basis for the classification metrics. The DC3 dataset was divided into periods corresponding to convective system inflow over one of three storm domains located in Colorado, Texas or Oklahoma and Alabama. There were 1307 cases assigned to one of these seven categories and the remaining 1083 cases, not meeting any of the other classification criteria, were labeled as unclassified.

Phase function and -F12/F11 data are averaged over each aerosol type to obtain scattering patterns characteristic of each classification. The dust category produced a significantly stronger forward-scattering peak than all other types, likely as a result of a comparably large number of coarse-mode particles. The next strongest forward peak was found in the CO storms, followed by the storms in TX/OK and then AL. These scattering patterns were also likely influenced significantly by coarse-mode particles, with the convective systems likely lofting more large particles as the local environment became increasingly arid. There were very small differences in the scattering patterns of the SEAC4RS fine-mode-dominated aerosols. The largest differences between these types was observed in the peak values of -F12/F11, occurring around 90∘, with the BB cases being more strongly polarizing than the other two types.

In order to more easily explore the scattering measurements, as well as further validate the ancillary data aerosol typing scheme, PCA was applied to the PI-Neph measurements and the results were examined as a function of the ancillary data classification. The first few principal components of the F11 data and the first principal component of the -F12/F11 data showed very strong relationships with aerosol type. Two schemes were developed to divide the PCA score space into regions representing the different classification categories. The first of these schemes was based on the Mahalanobis distances between points and the center of the cluster corresponding to each ancillary data classification category. The second scheme simply used a plane to divide the PCA score space into two regions predicting positive and negative classifications of a given type. In both schemes, individual scattering measurements were assigned to the correct type with very high recall, further supporting the validity of the ancillary data aerosol typing scheme, and demonstrating that the PI-Neph data alone are capable of identifying major aerosol types. This conclusion held even in cases in which differences in single-angle scattering intensities between the bulk averages of two different aerosol types was very small, often much less than the error in an individual measurement. For example, the dividing plane prediction technique was able to distinguish BB cases from other fine-mode-dominated aerosols with greater than 90 % recall and TNR, despite the two categories having extremely similar average scattering patterns. The characteristics producing this clear separation between similar aerosol types were subtle and often relied on the relationships among many angles simultaneously. This fact emphasizes the value of multi-angle scattering measurements, as well as PCA's ability to reveal the underlying patterns in these datasets.