Kathmandu is located in a broad basin at the foothills of the central Himalayas, the valley floor at an average altitude of 1300 m a.s.l. The mountains surrounding the valley have peaks ranging from 2000 to 2800 m a.s.l. Neighboring valleys to the west, north, northeast, and south have substantially lower elevations. The meteorology of the Kathmandu Valley is influenced by large-scale features, western disturbances, and the south Asian summer monsoon, as well as local mountain–valley circulations. As reported by Panday and Prinn (2009) and Panday et al. (2009), during the dry season the diurnal cycle of air pollutants (CO, O3, PM10) is strongly affected by local meteorology connected to the evolution of the convective planetary boundary layer (PBL) and thermal wind flows along the flanks of the mountains surrounding the valley.

The Paknajol site is located (27∘43′4′′ N, 85∘18′32′′ E; 1380 m a.s.l.) near the edge of Kathmandu's tourist district of Thamel. The sampling site is on the terrace (about 25 m a.g.l.) of the Ev-K2-CNR representative office. This is the highest building in the block; thus it has a 360∘ free horizon of at least 300 m. The instruments are located in an air-conditioned room, in order to maintain the correct operating conditions. A UPS – which signifies an uninterruptible power supply – guarantees the continuous measurements in case of (frequent) blackout events; there may be power cuts of up to 18 h a day, especially during the winter months. The sampling heads are placed on the roof just outside this room. The measurement activities, including aerosol and trace gas measurements, were started in February 2013.

The following instruments are used for continuous measurements.

A UV-absorption analyzer (TEI 49i, Thermo Environmental) is used to collect surface O3 measurements. These are linked to the World Meteorological Organization (WMO)/Global Atmosphere Watch (GAW) reference scale (Standard Reference Photometer; SRP 15; see Klausen et al., 2003) hosted at the GAW World Calibration Centre (WCC) at EMPA (Switzerland), via direct comparison with the CNR-ISAC laboratory standard hosted at the Mt. Cimone WMO/GAW global station (Italy). The experimental setup is similar to that described in Cristofanelli et al. (2010).

In order to describe the synoptic-scale atmospheric circulation scenarios over the Kathmandu Valley and the surrounding region, isentropic 5-day back trajectories have been used, computed by the HYSPLIT model (Draxler and Hess, 1998) every 6 h (at 05:45, 11:45, 17:45, and 23:45 NST). With the aim of minimizing the effect of the complex topography and to provide a description of the large-scale circulation in the free troposphere, calculations were initialized at 600 hPa. The model calculations are based on the Global Data Assimilation System (GDAS) meteorological field produced by NCEP Reanalysis data, with a horizontal resolution of 1∘×1∘. In order to aggregate the back trajectories of common origin, and better characterize the synoptic-scale circulation occurring at Paknajol, a cluster aggregation technique (Draxler, 1999) has been applied to the back trajectories. Basically, at each step of the process, the appropriate number of clusters was identified, based on the variations of several statistical parameters; by maximizing between-group variance and minimizing within-group variance, this methodology might identify similar air-mass back trajectories and aggregate them.

To understand how photochemistry and dynamics affect the variation of O3 mixing ratios and to comprehend the origin of elevated O3 levels in the afternoon and evening hours during the pre-monsoon period, we used a recurrent neural network model. These models allow us to simulate the nonlinear relationship between O3 and meteorological parameters that are proxies of photochemistry and dynamics (Lönnblad et al., 1992; Elman, 1990). Considering the strong role of meteorological conditions and photochemistry on the variations of O3 mixing ratios (Pudasainee et al., 2006; Di Carlo et al., 2007) and the fact that meteorological effects usually last for more than 1 day, the more complex architecture of the neural network that uses the recurrent approach takes into account the multi-day effect of meteorology, as well as diurnal boundary layer cycles (Biancofiore et al., 2015). The model uses the observed pressure, temperature, RH, solar radiation, wind velocity and direction, and BC concentrations as input to simulate the O3 mixing ratio (Biancofiore et al., 2015). The inclusion of a sub-group of these proxies allows us to distinguish between the role of dynamics and that of photochemistry in the observed variations of O3 mixing ratio.

The Paknajol area is strongly influenced by local traffic and urban emissions, as it is located near the edge of Kathmandu's tourist center, and near a major thoroughfare. Meteorological observations at the sampling site help in better describing the seasonal and diurnal variability of the air pollutants and SLCF/P in the Kathmandu Valley.

With the aim of identifying the regional transition of the monsoon seasonal regimes, we considered meteorological observations carried out at the Nepal Climate Observatory-Pyramid (NCO-P) station, located at 5079 m a.s.l. near Mt. Everest in the Himalayas. As shown by Bonasoni et al. (2010), the variability of meteorological parameters (i.e., RH and meridional wind component) observed at NCO-P can be used to derive the onset and withdrawal dates of the different seasons on the south side of the Himalayan range (where NCO-P is located). Moreover, as described in the annual report of the India Meteorological Department (IMD, 2014), the seasonal advance of the south Asian monsoon cycle did not differ too much between the location of NCO-P and Kathmandu. Table 1 reports the start and withdrawal dates of each season (pre-monsoon, monsoon, post-monsoon, winter) for the period considered in this study.

Figure 1 shows the variability of the meteorological parameters measured at Paknajol from February 2013 to the end of January 2014. Hourly atmospheric temperature (T, Fig. 1a) values never exceeded 29.5 ∘C, while minima never dropped below 3.5 ∘C. Over the whole measurement period, T had an average value of 18.7±5.6 ∘C (hereinafter, average values are indicated as average ± 1 standard deviation). T was characterized by an evident diurnal cycle, with values peaking in the middle of the day and a minimum in the early morning. Atmospheric pressure (P, Fig. 1b, average value: 865.3±4.1 hPa) showed its minimum values during the summer season, which is characterized by the presence of the monsoon trough over Nepal, accompanied by frequent and intense showers, reaching up to 47 mmh-1 (Fig. 1d). P is characterized by a semi-diurnal cycle, with two minima (at 04:00 and 16:00) and two maxima (at 10:00 and 22:00), with average amplitudes ranging from 1.2 hPa (pre-monsoon) to 3.5 hPa (post-monsoon). RH values (Fig. 1c, average value: 67.1±17.0 %) were high during all of the measurement period, rarely decreasing below 20 % (50 % during the summer monsoon season); it has to be noted that during winter, RH values swing from very high to very low, thus presenting the widest diurnal cycle among all of the seasons. Saturation conditions (RH equal to 95 % or higher) were mainly reached during the most intense rainfalls. In agreement with rainfall reported in Panday and Prinn (2009), about 90 % of annual rainfall was observed during June–August. Figure 1e and f show wind speed and direction, respectively. The sampling site was characterized by low wind speeds, with the majority of winds from the W–NW sector and a secondary contribution from the W–SW sector (see Fig. S1, Supplement). As shown in Panday and Prinn (2009) and Panday et al. (2009), nights were characterized by low wind speeds (maximum speed: 4 ms-1) coming from several directions, mainly explained by katabatic winds descending from the mountain slopes at the edge of the Kathmandu Valley rim; however, during the afternoon, stronger winds (reaching up to 6.5 ms-1) occurred at the measurement site, which was swept by westerly/northwesterly winds which entered through the western passes.

The hourly average (along with daily averages) time series for O3, BC, and particle number concentration are shown in Fig. 2. Figure 3 shows the diurnal variability of these pollutants across the seasons, while seasonal average values are presented in Table 2.

Similarly to other polluted cities, the rush hours and PBL dynamics result in the distinct morning and evening peaks: an increase in traffic activities and congestion, an increase in emissions from cooling/heating activities (LPG, kerosene, and firewood), as well as a decrease in the PBL. The primary emission indicators, i.e., BC and aerosol particle number, reveal such activities. Industries, especially brick kilns, and open garbage burning also contribute to poor air quality in the Kathmandu Valley.

The average value of BC (Fig. 2a) over the whole measurement period was 11.6±10.7 µgm-3. The highest BC concentrations were observed during pre-monsoon and winter seasons (Table 2), with daily values often exceeding 20 µgm-3; while the lowest values occurred during the monsoon season (the lowest daily value recorded was 2.5 µgm-3). These levels are slightly higher than what is reported in a previous study by Sharma et al. (2012) at Pulchowk Campus, in which they reported an average BC of 8.4±5.1 µgm-3, over a year-long study period spanning between May 2009 and April 2010. Another study by Shrestha et al. (2010) reported far lower values of elemental carbon concentration (1.7±0.6 µgm-3) for an urban site 30 km southeast (downwind) of the Kathmandu Valley, during the 2009 pre-monsoon season. The highest seasonal values, observed during winter/pre-monsoon, can be attributed to several factors: an increase in emissions from domestic heating, the use of small but numerous gensets during extended hours with power cuts, the operation of over 100 brick kilns in the valley, refuse burning, as well as lower PBL and lower wet deposition of pollutants in winter months compared to the summer months with intense heat and rainfall. The average diurnal variation in BC concentrations in the different seasons is shown in Fig. 3a. The typical diurnal variation for BC, as also shown in Sharma et al. (2012), reflects the BC profile for an urban site, presenting two daily maxima, with a prominent peak in the morning (between 07:00 and 08:00), and a second one in the evening (between 20:00 and 21:00), as well as two minima at night (between 01:00 and 02:00) and in the afternoon (between 14:00 and 15:00). These two daily peaks reveal the start and build-up of emissions due to local anthropogenic activities, such as traffic and cooking activities. Moreover, also a meteorological component cannot be ignored: this is due to the presence of katabatic winds that lead to the uplift of surface polluted air masses during the night. The following build-up of the morning mixed layer favors the downward mixing of pollutants back to the bottom of the valley (Panday and Prinn, 2009; Panday et al., 2009). This diurnal cycle was observed in all four seasons but the peak values were much higher in winter and pre-monsoon seasons (morning peaks: 41.4 and 33.3 µgm-3, respectively) compared to the post-monsoon and monsoon seasons (12.9 and 11.4 µgm-3, respectively).

Surface ozone (O3) had an average value of 27.0±21.3 nmolmol-1 (1 nmolmol-1 is equivalent to 1 ppb) over the whole measurement period (Fig. 2b). The highest O3 was observed during the pre-monsoon season, while the lowest values were reached during the winter season (Table 2). This spring peak is a feature widely present in south Asia and the Himalayas (see e.g., Cristofanelli et al., 2010; Agrawal et al., 2008). Pudasainee et al. (2006), using measurements made at Lalitpur, an adjacent city to the Kathmandu municipality, suggested that variations of solar radiation, ambient temperature, and precursors (such as NOx and VOCs) can together explain 93 % of the variation in measured ground level O3 at Kathmandu. Following Chevalier et al. (2007), with the aim of attributing the fraction of O3, BC, accumulation, and coarse particles' variability related to day-to-day and diurnal-scale processes, we calculated the ratio of daily/hourly standard deviations. The obtained values (0.54 for O3, 0.59 for BC, 0.81 for accumulation, and 0.71 for coarse particles) indicated that both diurnal and day-to-day variations are important to explain O3, BC, and particle number variations at Paknajol. O3 diurnal variation is shown in Fig. 3b: a peak in O3 mixing ratios characterized the central part of the day (between 11:00 and 13:00), while a minimum was observed in the morning (between 05:00 and 06:00). This diurnal variation is typical for polluted urban sites (Jacobson, 2002) and can be explained in terms of local O3 photochemistry production and removal processes as well as PBL dynamic and vertical air-mass mixing, as discussed in Sect. 3.4.

Particle number concentrations of accumulation (0.28 µm ≤ Dp < 1 µm) and coarse (1 µm ≤ Dp < 10 µm) particles are reported in Fig. 2 (panels c and d, respectively). Unfortunately, due to instrumental failures, no measurements were available after 27 July 2013; only two seasons were covered, i.e., pre-monsoon and monsoon. The average values over the available time period were 505±372 cm-3 for the accumulation particles and 3.3±2.4 cm-3 for the coarse particles. Accumulation and coarse particle concentrations were high during the pre-monsoon season and far lower during the monsoon (Table 2). The average seasonal diurnal cycles (Fig. 3c for accumulation and Fig. 3d for coarse) were somewhat similar to that of BC, presenting two daily peaks, in correspondence to the start of working activities and traffic rush hours, thus indicating common anthropogenic emission sources and similar meteorological influences. The similar behavior between accumulation and coarse particle number concentrations suggests likely common origins, indicating that the main fraction of coarse particles is linked to the resuspension of road dust or ash from local combustion and not to mineral dust transport from desert areas. During the pre-monsoon, the morning peak was higher (1054 cm-3 for accumulation and 7.4 cm-3 for coarse) than the one recorded in the evening (629 and 4.3 cm-3, respectively). The same was true for the accumulation mode even during the monsoon season, although the difference between the two peaks showed smaller amplitude. For coarse particles, however, the evening peak appeared to be higher than the morning peak during the wet season. This can be explained by considering the wet conditions which usually characterized Kathmandu during nighttime in this season; moreover, it has to be noted that this phenomenon may be combined with aerosol drying from the dilution system of the OPC that is not sufficient. Most of the rain occurs during the nighttime, and the wet surface in the early morning prevents emission of roadside dust and soil. As the day evolves, moisture is more efficiently evaporated, leaving dry dust and soil to be resuspended by traffic or winds, thus leading to the appearance of a larger evening peak for coarse particle number.

Figure 4 shows the seasonal box-and-whisker plot for PM10 and PM1 recorded at Paknajol. Prior to 1 April 2013, a PM2.5 sampling head was installed in place of the PM1 one. By considering the whole sampling period, PM10 had an average value of 169±113 µgm-3, which is comparable to the value found by Giri et al. (2006), 133.7±70.3 µgm-3, computed over the period 2003–2005 for the Thamel measurement site (not far from Paknajol), or to the values found in Aryal et al. (2008), which range from 170 to 230 µgm-3 (annual averages) for two busy traffic area stations in Kathmandu. Our value appears slightly higher than those of Giri et al. (2006). This is in line with the increasing urbanization and the growth of the total vehicle fleet which occurred in the Kathmandu Valley. The maximum seasonal average of PM10 was found during winter, while minima occurred during monsoon and post-monsoon seasons (Table 2). PM2.5 presented an average value of 195±83 µgm-3 over its short time period (17 days), while PM1 had an average value of 48±42 µgm-3, with the maximum values during the pre-monsoon and winter seasons and significantly lower values during monsoon and post-monsoon (Table 2). Over the whole measurement period, the ratio PM1 / PM10 was 0.29±0.10, indicating a large contribution of coarse particles to the total aerosol mass. This aerosol mass concentration ratio, whose values were the highest during the pre-monsoon (0.39±0.09) and lowest during winter (0.21±0.05), is similar to the ratios observed for arid sites (Shahsavani et al., 2012; Lundgren et al., 1996), for sites affected both by dust storms originating in Asia (Claiborn et al., 2000), strong African dust outbreak episodes (Alastuey et al., 2005), and dusty roads (Colbeck et al., 2011). Similar ratios were observed also in other large municipalities in south Asia, such as Bilaspur (0.24, Deshmukh et al., 2010) or Raipur (0.28, Deshmukh et al., 2013) in India, or Nanjing (0.34, Wang et al., 2003) in China. In the European cities this ratio is generally higher than in Asia.

We highlighted the wind sector which mostly contributed to the occurrence of high SLCF/P values at the measurement site, as presented in Fig. 5. Here, the angular distribution of the pollutants averaged over WD intervals of 10∘ (green lines) is shown. Also reported in the figure are the distributions of the frequency of wind directions (blue) and the relative abundance of the pollutants (red), weighted by the wind directions, computed according to Gilge et al. (2010). These analyses refer to the whole investigation period and no significant differences were observed, neither by categorizing data as a function of the different seasons, nor by the time of day. WD behavior has already been presented in Sect. 3.1; BC and aerosol particle number (both accumulation and coarse) average values did not show any dependence as a function of wind direction. This is conceivable considering that Paknajol is located in the middle of several pollution sources. O3 angular mean values (green line) showed enhanced values from the W–NW sector (35 nmolmol-1 on average). This leads to a small distortion of the O3 contribution away from the distribution of the wind directions (peaking at 270∘). 48 % of the total O3 recorded at Paknajol station was enclosed in the 240–320∘ wind sector, which perfectly matches the direction from a mountain pass from where, according to Panday and Prinn (2009), air masses can be transported during daytime towards Kathmandu due to thermal transport, indicating the arrival of regional polluted air masses.

By looking at the diurnal variations presented in Fig. 3, the first peak in BC and aerosol particles can be explained in terms of increased emission (traffic and cooking activity) under atmospheric stable conditions and low PBL height or with an additional contribution of down-mixing as the nighttime stable boundary layer breaks up (Panday and Prinn, 2009). Dilution within the higher PBL, arrival of cleaner air from west of the Kathmandu Valley, and decrease of emissions can explain the daily minimum in aerosol and BC observed from 11:00 to 17:00. Conversely, the peak in O3 can be explained in terms of enhanced photochemical production (with respect to nighttime or early morning), as well as in terms of downward vertical mixing of polluted regional air masses from the free troposphere or the nighttime residual layer. When the PBL height starts to decline due to the diurnal decrease of solar radiation and soil heating, along with the increased emissions of evening traffic and cooking activities, a secondary peak in BC aerosol is observed from 18:00 to 22:00. Titration with NO, dry deposition, and less efficient vertical mixing within a more stable PBL lead to the decrease of O3 which finally results in the nighttime minimum, when BC and aerosol particles also present the lowest concentrations due to the decrease of traffic and domestic emissions. Moreover, since measurements were taken on the roof of a tall building, in the presence of a stable nighttime atmosphere, it may be difficult to capture near-surface pollution. These behaviors led to a negative correlation between hourly BC and O3 (Table 3), which was almost constant over all of the considered seasons. The O3 decrease after the noon peak was faster during winter and post-monsoon seasons, while it was more gradual during pre-monsoon and monsoon. Moreover, during the pre-monsoon season, a “bump” in O3 mixing ratios (>50 nmolmol-1) was observed during the afternoon (between 11:00 and 17:00; Fig. 3). The simultaneous decreases of BC and aerosol particle concentrations support a strong role of downward vertical mixing in enhancing O3 and decreasing primary pollutants (BC and aerosol particles). The important role of dynamics in influencing SLCF/P variability is confirmed by the negative (positive) correlation between wind speed and BC (O3). The correlation coefficients (r) are higher by considering daily average values (Table 3), supporting the role of day-to-day meteorology in influencing the SLCF/P.

BC showed significantly higher hourly correlation with accumulation and coarse particles (0.86 and 0.87, respectively), which was lower during the wet season (0.66), strongly supporting common sources and processes that influence their variability (i.e., traffic sources and PBL dynamics). The lower correlation can be explained in terms of different hygroscopicities of BC with respect to other aerosol particles (see Marinoni et al., 2010), which can lead to a lower scavenging efficiency of BC, with respect to other inorganic and organic species, that has been proved especially for non-aged BC (Cozic et al., 2007). Due to the lack of data, no information about the variation of the correlation coefficients computed between accumulation and coarse particles could be given other than during pre-monsoon and monsoon seasons. The high correlation coefficient between BC and accumulation (and coarse) particles could however indicate that BC can be used as an indicator of primary pollution, even when measurements by the OPC are lacking.

O3 showed high correlation with solar radiation (0.71 for hourly and 0.56 for daily values) and temperature, which is considered as a proxy for seasons (0.51 and 0.32); this is somewhat expected for an urban site like Kathmandu, where photochemistry and PBL dynamics (indirectly driven by solar radiation and temperature behavior) play an important role in determining O3 variability (Pudasainee et al., 2006). The correlation with solar radiation exhibited some variability during the year, giving the lowest values (0.59 for hourly values and 0.06 for daily values) during the pre-monsoon season, possibly supporting the enhanced role of atmospheric transport and dynamics in influencing O3 with respect to photochemistry. Apparently, this agrees only in part with the results shown in Pudasainee et al. (2006), in which the authors argued that the “flat peak” in O3 concentrations during the pre-monsoon is mainly due to abundance of solar radiation and higher temperature (justified by high correlation coefficient values).

In order to distinguish the chemical effects from the boundary layer dynamics, we also computed correlation coefficients limiting the data to convective hours only (i.e., between 11:00 and 17:00, according to the wind speed and solar radiation diurnal variations). The slightly weaker correlation between BC and accumulation particle number and, conversely, the increase in correlation between O3 and accumulation particle number, may indicate the role of other processes (e.g., secondary aerosol production) occurring in the air masses which characterize this specific time span (Table S1, Supplement). In particular, we suppose that aged air masses rich in secondary pollutants (i.e., O3 and aerosol) can be transported to the measurement site in the afternoon mixed layer.

Here, we argue that mixing processes with upper residual O3 layers can explain this behavior. Sensitivity tests with a recurrent neural network model, using different subgroups of proxies, have been carried out and the results are shown in Fig. 6, where the observed and different simulated average diurnal O3 mixing ratios are compared. The simulation that included all the proxies reproduced the observed O3 mixing ratios for all hours of the day quite well, whereas a simulation that included only wind speed (a good proxy of atmospheric dynamics) reproduced the afternoon (after 15:00) and evening levels of O3 with accuracy, missing the main O3 peak before noon completely. In contrast, by using both wind speed and solar radiation as input parameters, the model reproduced the peak before noon and the high levels of afternoon–evening O3 well. Putting together the results of these two simulations, we can conclude that the high level of O3 during the afternoon is mainly due to dynamics (vertical intrusion from upper atmospheric layers and/or horizontal advection), for the following two reasons. (i) In the model, the wind speed used as input is enough to reproduce the afternoon concentrations of O3 and (ii) the inclusion of solar radiation does not improve the agreement between measured and modeled O3 during the afternoon, but substantially enhances the agreement between measurements and simulations before noon, when photochemistry, as expected, plays a larger role. The photochemistry contribution varied as a function of the hour of the day, ranging from 6 to 34 %.