The Climatology of Australian Aerosol

Airborne particles or aerosols have long been recognised for their major contribution to uncertainty in climate change. In addition, aerosol amounts must be known for accurate atmospheric correction of remotely sensed images, and are required to accurately gauge the available solar resource. However, despite great advances in surface networks and satellite retrievals over recent years, longterm continental-scale aerosol data sets are lacking. Here we present an aerosol assessment over Australia based on combined sun photometer measurements from the Bureau of Meteorology Radiation Network and CSIRO/AeroSpan. The measurements are continental in coverage, comprising 22 stations, and generally decadal in timescale, totalling 207 station-years. Monthly climatologies are given at all stations. Spectral decomposition shows that the time series can be represented as a weighted sum of sinusoids with periods of 12, 6 and 4 months, corresponding to the annual cycle and its second and third harmonics. Their relative amplitudes and phase relationships lead to sawtooth-like waveforms sharply rising to an austral spring peak, with a slower decline often including a secondary peak during the summer. The amplitude and phase of these periodic components show significant regional change across the continent. Fits based on this harmonic analysis are used to separate the periodic and episodic components of the aerosol time series. An exploratory classification of the aerosol types is undertaken based on (a) the relative periodic amplitudes of the Ångström exponent and aerosol optical depth, (b) the relative amplitudes of the 6and 4-month harmonic components of the aerosol optical depth, and (c) the ratio of episodic to periodic variation in aerosol optical depth. It is shown that Australian aerosol can be broadly grouped into three classes: tropical, arid and temperate. Statistically significant decadal trends are found at 4 of the 22 stations. Despite the apparently small associated declining trends in mid-visible aerosol optical depth of between 0.001 and 0.002 per year, these trends are much larger than those projected to occur due to declining emissions of anthropogenic aerosols from the Northern Hemisphere. There is remarkable long-range coherence in the aerosol cycle across the continent, suggesting broadly similar source characteristics, including a possible role for intercontinental transport of biomass burning aerosol.


Introduction
Suspended atmospheric particles or aerosols are extensively studied because of their multiple environmental roles, and because -unlike well-mixed gases -their concentration and properties vary widely over space and time. Aerosols adversely affect human health (particularly in urban areas), and exert a further range of impacts due to their modulation of electromagnetic radiation within the atmosphere. Scattering and absorption by aerosol remains a major source of uncertainty in climate projections 5 (Boucher et al., 2013), while satellite images of the Earth's surface require correction for aerosol effects. In addition, aerosol data are also required for the accurate assessment and forecasting of available solar energy (e.g., Perry and Troccoli, 2015).
Major enhancements in surface networks and satellite retrievals over the last 15 years have radically boosted knowledge of global aerosol. In particular, surface networks including AERONET (Holben et al., 1998), SKYNET (Nakajima et al., 2007;Hashimoto et al., 2012) and GAW/PFR (WMO, 2005) have provided regional aerosol climatologies and supported process 10 studies (e.g., Eck et al., 2008), while several major satellite programs have championed the characterization of aerosols from orbit. Of note in this respect are NASA's flagship EOS A-train instruments MODIS (Levy et al., 2010), MISR (Martonchik et al., 2009) and CALIPSO (Young et al., 2013). In addition, the ATSR series of instruments operated by ESA has made a significant contribution to satellite aerosol measurement from space, particularly the AATSR instrument launched on the ENVISAT platform in March 2002 (Grey and North, 2009;Qin et al., 2015). However, retrieval of aerosol from satellites 15 remains challenging (Mishchenko et al., 2009), and requires ongoing availability of validation data from surface networks.
In addition to these more recent sensors, the 30-year record available from the NOAA/AVHRR series is proving valuable, as shown by the recent examination of the aerosol indirect effect of maritime aerosol on cloud properties (Zhao et al., 2016).
Since aerosol emission is highly correlated with human population and land mass, there has been a striking northernhemisphere bias in aerosol characterisation efforts to date. However, there is a growing need for improved characterisation 20 of southern hemisphere aerosol. In particular, with lower levels of anthropogenic aerosol (AA), the southern hemisphere is on average less subject to 'masking' of the greenhouse warming than the northern hemisphere (Rotstayn et al., 2013). Consequently, low AA regions including Australia experience the greenhouse gas component of warming with less moderation from aerosol radiative forcing, to a degree dependent on improved knowledge of Australian natural aerosol. Rotstayn et al. (2013) sought to quantify the unmasking implied by projected future decreases in AA, and their findings will be further considered 25 below. Furthermore, recently Stephens et al. (2015) showed that the contribution of each hemisphere to the planetary albedo averaged over the last 13 years has been in near-perfect balance, with the increased reflection from northern hemisphere land exactly balanced by increased atmospheric scattering in the southern hemisphere. While the full ramifications of this finding are yet to be fully explored, there is a pressing need to better characterise southern hemisphere aerosol emissions.
Australia is the largest dust source in the southern hemisphere (Tanaka and Chiba, 2006) and is a significant source of smoke 30 aerosol from savanna burning (Mitchell et al., 2013). Hence it is clear that Australia's aerosol emissions make up a significant fraction of the southern hemisphere total. Since aerosol contributes to the planetary albedo both directly, and indirectly through the modulation of cloud properties, the need for improved continental-scale characterisation of Australian aerosol is clear.
Studies of Australian aerosol to date have often been restricted in time and space. While coastal maritime aerosol has been extensively studied (Gras and Ayers, 1983;Gras, 1991), studies of the continental aerosol are more restricted. Thus Scott et al. (1992) characterised the aerosol cycle caused by seasonal biomass burning at Broome in the tropical north west of Australia in the late 1980s, with more extended analysis of regional tropical aerosol carried out by Mitchell et al. (2013); see also references cited therein. Studies of the smoke from Eucalypt fires in southern temperate forests are few in number, but include Mitchell et al. (2006) and Radhi et al. (2006). 5 Australian dust emission has been well-studied from a geomorphological perspective (e.g., Leys et al., 2008;Bullard et al., 2008) but less so via sun photometry, where the measurand -column aerosol optical depth -is the required input for climate and remote sensing studies. However, Mitchell et al. (2010) analysed sun photometer and nephelometer measurements from Tinga Tingana in the Australian arid zone during the Millennium drought (2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010) and found an approximate doubling in both column and near-surface aerosol during the austral summer over the duration of the drought. Feedbacks related to dust 10 aerosol were modelled by Rotstayn et al. (2012), who found that inclusion of interactive dust in the model amplifies the impact of the ENSO cycle on Australian climate.
While satellite retrieval of aerosol over Australia is promising in terms of the extensive areal and temporal coverage available over a sparsely-populated continent, the dense-dark vegetation (DDV) assumption inherent in the MODIS aerosol retrieval algorithm does not apply well over most of Australia (Levy et al., 2010;Gillingham et al., 2012). Problems in other sensors 15 and retrieval methods are discussed by Qin et al. (2015), who show that the dual-view scan geometry of the AATSR instrument offers particular advantages under Australian conditions. However, the ongoing need for validation data from surface networks is universal. They showed that fine-mode anthropogenic aerosol is decreasing over those countries with strict environmental regulations, 20 and vice-versa. Coarse-mode dominant aerosol depends strongly on meteorological conditions subject to climate change, with stations close to regions of increasing desertification showing a significant increase in coarse-mode dominant aerosol.
In this paper we present a continent-wide climatology of Australian aerosol obtained at 22 stations, spanning all mainland states, and comprising a total of 207 station-years. The stations are operated by two agencies, the Australian Bureau of Meteorology, and the Commonwealth Scientific and Industrial Research Organization (CSIRO). The CSIRO stations are affiliated 25 with NASA's worldwide AERONET sun photometer system, as further described below. Section 2 describes the two measurement systems and the data processing applied to each. The result of this step is a time series of multi-spectral monthly mean aerosol optical depth. Section 3 describes the common analysis applied to these time series. Section 4 presents a discussion of the results, with conclusions given in Section 5. The Appendix presents the monthly aerosol climatology at each of the 22 stations in tabular form. 30 ning in 2000, confirm the Bureau of Meteorology methodologies produce aerosol optical depth within the WMO uncertainty requirement.
Both networks utilise a combination of Langley methods to find a reference wavelength calibration and then apply the General Method developed by Forgan (1994) to calibrate the other wavelength channels (excluding the channels affected by water vapour). The General Method assumes the stability of the relative aerosol size distribution -but not the magnitude of the optical depth -throughout a morning or afternoon, and uses the reference wavelength aerosol optical thickness as the predictor for the least squares solution of the wavelength calibration value at the top of the atmosphere. As a Langley-type method involves an interpolation to the top of the atmosphere in real atmospheric conditions the selection of points not impacted 5 by water or ice particles is key to the uncertainty of the top of atmosphere value. The two networks use different sampling strategies and methods to select points for the calibration process. The operational measurements are used to provide the in-situ calibrations.
The Bureau of Meteorology radiometers are co-located at solar and terrestrial irradiance stations that utilise the WMO Global Climate Observing System (or Baseline Surface Radiation Network) solar monitoring methodologies (Ohmura et al.,10 1998). The solar trackers employ active tracking sensors that mimic a 868 nm radiometer but with a wide field of view, and sample at 1 Hz. The standard deviation of the active tracking sensor is recorded every minute with the 0.12 s samples from the SPO2 radiometers. If the active sensor was not available the standard deviation of the direct irradiance from a pyrheliometer is used as an equivalent indicator of variability in the minute prior to the SPO2 measurement. The first pass of the selection process of data to use for SPO2 calibration is to eliminate the near zero SPO2 signals that are less than 5 times the measurement 15 resolution (typically 0.005 V), and those periods when the standard deviation of the active sensor signal is more than 0.005 V.
As the 868 nm active sensor signals in clear sun in typical Australian conditions is within 25% of the solar noon value for aerosol air masses less than 6, use of a fixed standard deviation is sufficient. The remaining data for each wavelength are then subject to the Harrison and Michalsky (1994) algorithm. If sufficient data remain, namely at least 30 points for aerosol air masses less than 6 and greater than 2, the calibration process is applied and the result recorded.

20
The AeroSpan measurements are based on a 'triplet' of the routine measurements obtained 30 s apart, spanning a minute.
The standard deviation among the three measurements comprising the triplet is used for analogous filtering, as described in more detail by Mitchell and Forgan (2003).
For both networks, determinations of the natural logarithm of the top of atmosphere signal (ln V 0 ) at 1 AU earth-sun distance are compiled during the deployment period, which may range from months to over a year. The time series of the ln V 0 is then 25 used to derive the time dependence of the ln V 0 through linear interpolation using a single slope where possible, or a series of linear functions if the calibration is changing in a non-linear manner. For some wavelengths of some radiometers the non-linear changes were rapid and difficult to model; these periods have been eliminated from the data analysis. The calibration functions for each wavelength providing a top of atmosphere value for each day are used in the derivation of optical depth for the entire measurement data set, not just the points selected for the calibration process. 30 Determination of aerosol optical depth requires subtraction of all relevant molecular extinction. For the present 'aerosol' channels (i.e., not including the 812/936 nm water vapour channels), these are nitrogen dioxide and ozone. The correction for 5 Atmos. Chem. Phys. Discuss., doi:10.5194/acp-2016-1021 Manuscript under review for journal Atmos. Chem. Phys. Published: 16 January 2017 c Author(s) 2017. CC-BY 3.0 License.
NO 2 is small (typically <0.002 at 380-440 nm and <0.001 at 500 nm (Mitchell and Forgan, 2003) and was excluded from the present analysis; however ozone absorption is more significant. The total column ozone determined from the Total Ozone Monitoring Satellite (TOMS) measurements was used in this analysis. In particular, TOMS-EP (Earth Probe) was used from For the elimination of cloud-affected aerosol optical depth measurements the two networks use different methods. In 5 AeroSpan the filtering is similar to that described by Smirnov et al. (2000) used by AERONET. This first evaluates the shortterm temporal variation based on triplets, then applies a further test based on the daily standard deviation of the aerosol optical depth. For the Bureau data the method of Alexandrov et al. (2004) is used after the elimination of signals less than 5 times the measurement resolution. This applies a moving temporal window of 15 minutes in span across a time series examining the variance in the initial estimate of the aerosol component of the optical depth.

Calculation of Ångström exponent
The Ångström exponent α captures the spectral slope of the aerosol optical depth, and is defined as where τ is the spectral aerosol optical depth, and α is defined across the wavelength range (λ 1 -λ 2 ). For the AeroSpan stations, the Ångström exponent was defined using the wavelength pair (440-870 nm), in part because these channels have always 15 been present on the CE-318 instruments. For the Bureau stations, the Ångström exponents were calculated using the 500 nm wavelength and the 868 or 870 nm channel when available. For the Bureau data sets prior to 2009, 868 nm was only available from the active sensor, and led to an underestimate of the true optical depth due to the large opening angle. A comparison of the Bureau Ångström exponents using 500-778 and 500-868 was possible when the 868 nm channel was added in the last decade of measurements; no statistically significant difference was found. As a result, in the period before 2009, when the 868 nm 20 was only available from the active sensor at Bureau sites, the 500-778 nm pair was used to calculate the Ångström exponent statistics reported below. A similar rationale supports the comparability of the AeroSpan and Bureau Ångström exponents, even though the basis wavelength pairs differ.
For the majority of the 22 stations, the annual mean aerosol optical depth at the 500 nm wavelength is less than 0.06, and at 868 nm less than 0.03. Hence even with the low uncertainties (∼0.007) for the individual aerosol optical depth measurements, 25 the uncertainties in the individual Ångström exponent measurements can be relatively large (>0.2). The monthly statistics of Ångström slopes based on individual values was compared to the monthly mean wavelength pair gradients derived from the General Method calibrations. The differences in monthly means were typically less than 0.1 except for those months where the mean aerosol optical depth at 500 nm was less than 0.020. The comparison of the Ångström exponent methods of measurement and subsequent statistics is the subject of future work. In this section we introduce the data sets, then define the model fitting procedure used as a basis for the analysis. This allows separation of the observed aerosol times series into a periodic component, with the residual interpreted as an episodic component. This is used as a basis for classification of the aerosol across the continent in terms of (a) the periodic amplitudes found for the Ångström exponent (a measure of aerosol size) and aerosol optical depth; (b) the harmonic content of the periodic com-5 ponent of the aerosol optical depth; and (c) the balance between the episodic and periodic components of the aerosol optical depth.

Monthly time series
Since the object of this work is to define the monthly climatology of Australian aerosol, aerosol time series were preprocessed into monthly means for subsequent analysis. The analysis involved (a) model fitting (as described below) and (b) spectral 10 decomposition using a Fast Fourier Transform (FFT). Due to instrument malfunction and occasional removal for calibration, these time series contain infrequent gaps. The model fitting was carried out using all available data, simply ignoring periods when no data were available. However, the FFT required a complete time series, so interpolation for missing data was required.
This was carried out using a combination of spline fitting, and fits based on Kalman filtering, as provided in the R package.
Testing showed that the interpolation had no significant effect on the Fourier analysis, which was employed only to assess the The right-hand panels show the power spectrum obtained from the FFT. From the latter, the dominance of the annual cycle is clear, with significant contributions at the 6 and 4-month periods, varying between sites. This general pattern was evident 20 at all sites. The increasing complexity of the power spectrum at Wagga Wagga ( Figure 4) arises due to the high incidence of episodic smoke plumes from forest fires in the south eastern Eucalypt forests. Other stations -not shown here -exhibit similar characteristics, supporting the analysis outlined here.

Model fitting
The equation describing the model fit is as follows: where a 0 is the time-invariant offset of the fit to the (dimensionless) aerosol optical depth time series, and a 1 is the linear coefficient with units of inverse time. The following term represents the sum of J sinusoidal components, with amplitude b j and frequency ω j . The phases of the sinusoidal components are encapsulated by t * j , which is the time in months corresponding to the first maximum in the cosine function following the zero-time reference for the j th harmonic component, referred to 30 below as the phase-time. The zero-time reference was chosen as the first December in the time series at each station. This aids the pictorial representation of the phase angles of the harmonic components, with the months of the phase-time corresponding to the hours of a clock face, with December at 12 O'clock (see Figure 6 in Section 4.1). Guided by the identification of the annual cycle and its second and third harmonics in the FFT analysis discussed above, three sinusoidal terms were employed (J=3), with periods of 12, 6 and 4 months corresponding to frequencies ω of 1/12, 1/6 and 1/4 cycles per month, respectively.
An example of the Fourier decomposition of one of the aerosol time series -in this case from Tinga Tingana -is shown in Figure 5. The sinusoidal components are shown in the lower panel, with their respective phase-times marked by arrows.

5
Their sum (together with the offset and linear term further discussed below) results in the fit depicted by the orange line, representing the optimal fit to the data (shown here by the climatology in black) derived using a least-squares optimization provided within the R language. Clearly, the fitting using the three component sinusoid is capable of representing the dual peaks in aerosol optical depth observed at Tinga Tingana and in other in arid zone stations, the peaks occurring in the spring and summer. Although the three-term sinusoidal fit has insufficient high-frequency content to represent the narrow peaks observed 10 at this site, the higher frequency components required are not found to be continent-wide, and the use of this three-component harmonic analysis is sufficient for classification of the various aerosol regimes as further discussed below.
A list of the fitted parameters for all sites is given in Table 2. As well as the parameters defined in Equation 2, the tables lists the square of Pearson's correlation coefficient (R 2 ). The linear trend parameter a 1 and its significance is further discussed below.

Amplitude and phase of the aerosol optical depth of harmonic components
A geographic depiction of the fitted amplitudes and phases at all sites is shown in Figure 6. This clearly shows the increase in amplitude of the periodic components from south to north, and the general coherence of their phase relationship across the continent. However, at Darwin the phase-time of the annual cycle is noticeably earlier (10.3 months) than at all other tropical 20 sites (11.0 months), suggesting an earlier onset of dry season at Darwin than at other sites. Comparison of the fitted functions with those at other tropical stations (Broome, Lake Argyle and Jabiru) indeed shows slightly earlier onset of the rise toward the October peak at Darwin than at the other tropical stations. In addition, the other stations show a broader peak declining later than at Darwin. This earlier decline from the October peak contributes to earlier phase-time at Darwin than at the other sites. Both effects can be understood as the high population density around Darwin relative to the other stations, leading to 25 many fires early in the dry season. By contrast, the burning in the more remote regions is more driven by natural ignition, and hence peaks later when the fuel is drier. The reduced influence of human activity -with fewer late season fires being lit closer to Darwin -supports the persistence of the smoke aerosol later in the dry season at the more remote stations.
A regional phase change is also evident in south-eastern Australia, in the west-east transect Adelaide, Mildura, Wagga Wagga and Canberra. The corresponding phase-times are 10.4, 11.0, 0.1 and 0.6 months respectively, showing a trend toward a later character of the more easterly sites. This is particularly true of Wagga Wagga and Canberra, where major incursions of smoke aerosol are found in the austral summer months. In contrast, the arid zone aerosol has a greater magnitude during the austral spring.
Aerosol sources contributing to the noted rise during spring are heterogeneous, with suggested contributions from fine aeolian dust (Mitchell et al., 2010), a combination of biomass burning smoke, fine dust, and maritime aerosol from long-range transport (Radhi et al., 2010), and exogenous biomass burning aerosol originating in southern Africa or even South America Both Geraldton (t * 12 =11.9 months) and Learmonth (11.7 months) on the Western Australian coast show an aerosol peak near the end of the calendar year, and are thus distinct from both the tropical stations further north, and the temperate stations further south. These stations lie in the significant but poorly-studied north-west pathway for dust transport out of the central arid zone (Karlson et al., 2014), and their aerosol dynamics warrants closer study.

Comparison of harmonic amplitudes of Ångström exponent and aerosol optical depth
The fitting procedure described above was applied to the time series of both aerosol optical depth and Ångström exponent.
For the Ångström exponent, this leads to an analogous set of fitted parameters to those shown in Table 2, although for brevity, these are not shown. However, once the optimised fit for each station is obtained, its peak-to-peak amplitude can be determined for both the aerosol optical depth and Ångström exponent. A scatterplot of the amplitudes found for the Ångström exponent 15 against that for the aerosol optical depth is shown in Figure 7, for all stations where the data record is 5 years or more.
Some separation of different aerosol regimes is possible from Figure 7. In particular, the tropical stations are easily identified by their large aerosol optical depth amplitude (A τ >0.2) and clustering around Ångström exponent amplitude (A AE ∼ 1.0).
A large number of stations lie within 0.03 < A τ <0.1 but with a large range of A AE between 0.4 and 0.9. This rectangular region was divided into 'Temperate' and 'Arid', with the suggested division at A AE ∼0.6. Perhaps surprisingly given its coastal 20 location, this classification places Adelaide in the Arid regime, possibly a result of the transport of dust over Adelaide during the summer months. By contrast, the more easterly stations at Wagga Wagga and Canberra have lower A AE , as does Kalgoorlie, in temperate inland Western Australia.

The three stations Tennant Creek, Learmonth and Rockhampton fall outside the suggested classification rectangles. Tennant
Creek is clearly a tropical station with somewhat reduced A τ , consistent with its location between Darwin and Alice Springs. 25 Learmonth combines the A AE of a tropical station with the A τ of an Arid zone station, while Rockhampton shows a puzzlingly small A τ given its location in coastal tropical Queensland. The aerosol regimes at Learmonth and Rockhampton warrant further study. The vertical arm contains a mixture of arid and temperate stations, increasing toward Wagga Wagga and Canberra, the only stations where A 4:12 exceeds A 6:12 . This is a consequence of the large summer smoke aerosol signal associated with 10 these stations, as already discussed. The 'outlier' of Mildura arises because of the relatively small amplitude of the annual component, rather than excessively high amplitudes at 6 and 4 months.

The balance between episodic and periodic aerosol
The residual or difference between the optimised model fit to the observed time series (equation 2) and the data itself is considered to be the 'episodic' component of the aerosol signal, in the sense that it is determined by unpredictable episodes 15 such as fires or dust storms which depart from the 'periodic' component that govern the parameters of the fitted model. Figure 9 shows the standard deviation of the residual (expressed as a fraction of the peak-to-peak amplitude of its periodic variation) against the peak-to-peak amplitude of the periodic function fitted to the aerosol optical depth at 500 nm. The standard deviation was chosen as it includes extreme values, unlike other measures of variation such as the inter-quartile range. Figure 9 allows separation of (a) the four tropical stations (b) a cluster of 5 stations with episodic variation at ∼20%, and (c) 20 an extended series of stations with episodic variation between 24% up to 50%. The stations with episodic variation above 30% include the smoke-affected Wagga Wagga (46%) and Canberra (49%), and the dust-dominated Tinga Tingana (36%). Geraldton (39%) and Adelaide (32%) are also in this group, although the aerosol sources responsible are less clear. This result suggests that proximity to forested areas prone to irregular bushfires (unlike the savanna burning in the tropical north) engenders the greatest episodic variation found across the sites studied.

Trends in aerosol optical depth
Trends in aerosol optical depth are difficult to discern because of the inherent stochastic variability imposed on the periodic signal. Hence long time series are necessary to discern a small signal amidst much noise. For the present data set, we limit the trend analysis to those time series longer than 10 years. This avoids the pitfall that can appear when a multi-year (but sub-decadal) regional change causes a change in aerosol loading. Such a change caused an increase in dust aerosol over the The model used in the present study does return a linear trend, as the coefficient a 1 listed in Table 2, together with the significance of the trend, expressed as the probability that the trend arose from a random distribution. It is standard practice to declare the trend significant if the probability of it arising randomly is 5% or less. In the present case, three of the stations fulfill this criteria, but we relax it slightly to 6% to enable inclusion of a fourth station. The four stations satisfying both duration of record and significance criteria are listed in Table 3.
The table shows that the trends are negative but small. For Lake Argyle and Broome in Australia's north west (Kimberley) 5 region, the trend expressed as a change in aerosol optical depth at 500 nm is -0.002 per year, with a standard uncertainty of ±0.0008. Assuming the uncertainty is normally distributed, there is a 67% probability that the trend lies between -0.0028 and -0.0012 per year. For Rockhampton and Alice Springs, the best estimate of the trend is -0.0009 per year, with a corresponding range of -0.0012 to -0.0005. In summary, it can be stated that there is evidence for a small decrease in aerosol loading at these four sites, with a stronger (but still small) negative trend over the Kimberley stations.  Table A1.
For sites where the data record is over 5 years in length, the monthly climatology of aerosol optical depth at 500 nm is displayed in Figure 10. The time axis runs from July to June (rather than January to December) so as not to bisect the aerosol peak, which occurs in the austral spring-summer period (September to February). The grouping of stations into regions was

Correlation among aerosol time series
The groupings adopted in Figure 10 are used to examine correlation of the station time series within groups. The resulting values of Pearson's correlation coefficient between each station pair in a group are shown in Tables 4-7. Unsurprisingly, given the tight clustering found in much of the foregoing analysis, the tropical group shows high correlation between stations, with even the distant pair Broome-Jabiru yielding R=0.8 (Table 4). Although the correlation between arid zone station pairs 5 is weaker across the board, the pair Alice Springs-Birdsville showed the highest coefficient in the present study, at R=0.92 (Table 5). The large episodic aerosol component in the bushfire-prone part of the Temperate group reduces expectations of high correlation coefficients, so it is surprising to note the coefficient of R=0.795 between Wagga Wagga and Canberra (Table 6), suggesting regional-scale linkage via correlated smoke episodes. The results seen in Table 7 for the 'Unclassified' stations are yet more surprising, particularly the coefficient of R=0.73 between the pair Learmonth-Rockhampton, located on opposite 10 sides of the continent at a distance of ∼3730 km. The implication is that the mechanisms responsible for aerosol generation and transport have much in common at both regional and trans-continental scales.
A further mechanism underlying this correlation is the inter-continental transport of aerosol plumes from southern Africa and possibly South America considered by Rosen et al. (2000). Although they measured the plumes from only one site, such material is likely to be well-dispersed following long-range transport, and was shown to contribute ∼0.02 to the mid-visible 15 aerosol optical depth over Mildura. This is significant given the aerosol 'offset' of 0.05-0.06 typical of the arid zone stations (Table 2), so the widespread presence of this aerosol source could well boost the correlation among widely separated stations.
Further analysis of the prevalence and uniformity of these layers is required.

Summary and conclusions
Analysis of the aerosol records at 22 stations across the Australian continent has been carried out with dual aims in mind: (a) 20 generation of benchmark aerosol climatologies across the continent, and (b) classification of Australian continental aerosol types based on their measured time series. The latter goal was achieved by Fourier analysis which suggested that most of the seasonal variation could be explained by just three periodic terms, corresponding to the annual cycle and its second and third harmonics. Fitted coefficients arising from a model constructed on this basis support the following principal conclusions.
1. The proposed model explains 50% or more of the variance at 17 of the 22 sites (77%), and hence provides a basis on 25 which to infer aerosol classification and to examine the time series for possible trends.
2. Analysis of the fitted parameters suggests that Australian aerosol can be classified into three groups: Tropical, Arid and Temperate, although there is overlap between the latter two groups on some measures. 4. Arid zone aerosol is seasonally bimodal, with dual peaks in the austral spring and summer (typically September -January). This is evident in the increase in amplitude of the 4-monthly harmonic relative to the annual (12-month) signal.
5. Temperate stations in this study were influenced by smoke aerosol in a manner that increased with proximity to the Eucalypt forests of the Great Dividing Range. This is evinced as an increase in the episodic aerosol component, derived from the residual between the fitted (periodic) model and the measured data.  7. Strong correlation was found both among stations in the same group, and, remarkably, between stations separated by the full width of the continent. It follows that, on a monthly timescale, aerosol sources over Australia are driven by mechanisms that do not vary greatly either on regional or continental scales. There is evidence that inter-continental transport of biomass burning aerosol may play a significant role in this.
Despite the strong coherence in Australian aerosol just noted, we caution that this applies to monthly-mean data. The extent to 15 which such coherence applies on shorter timescales is unclear, although its persistence among tropical stations was explored by Mitchell et al. (2013). Although a monthly climatology is useful for low frequency applications, including solar resource prediction and arguably climate change, it is by no means suitable for applications requiring instantaneous aerosol fields such as the atmospheric correction of remote sensing imagery from satellites. Hence future work will consider the relation between the instantaneous and monthly mean aerosol, and the predictive capacity of the latter. Future work will also be directed to filling in the geographical gaps evident in the present study, particularly the lack of stations over Victoria and Tasmania.   give the AOD offset and linear coefficient respectively, the latter in units of AOD per year. The following column tabulates the significance of the trend, expressed as the probability that the trend arises from a random distribution. Trends were considered significant for P(a1) of     (blue), 6 months (red) and 4 months (green). The arrowed phase-times of the three sinusoids were 11.4, 2.9 and 1.1 months respectively. The phase-times are referenced to the first December in each time series, denoted as month zero in this plot. The observed climatology is shown as the black solid line. The linear trend apparent in the dashed line is not significant at the 5% level (see Table 2).