We use airborne in situ measurements as well as ground- and space-based remote sensing observations of tropospheric water vapour isotopes from the project MUSICA (MUlti-platform remote Sensing of Isotopologues for investigating the Cycle of Atmospheric water). This combination of multi-platform observations allows a comprehensive model validation, since the different isotope measurement methods are sensitive to different vertical, horizontal, and temporal scales. Dyroff et al. (2015) and Schneider et al. (2016) gave a detailed overview of the campaign and the measurements.

The limited-area model COSMOiso (Pfahl et al., 2012) is an isotope-enabled version of the non-hydrostatic numerical weather and climate prediction model COSMO (Steppeler et al., 2003). The isotope implementation incorporates two parallel, purely diagnostic water cycles for the heavy water isotopes H218O and HD16O, which experience exactly the same physical processes as H216O except for fractionation during phase changes. The prognostic isotope multilayer soil model TERRAiso couples the isotope content of the soil to COSMOiso (Dütsch, 2016; Christner et al., 2018), thereby accounting for fractionation during soil evaporation. Plant transpiration is treated as non-fractionating and transmits the isotope signals of the contributing soil layers according to the vegetation's rooting depths.

The COSMOiso simulation used in this study extends from 1 July to 31 August 2013. This time period includes the summer 2013 MUSICA campaign (see Sect. 2.1). The model domain was chosen such that it covers large parts of the North Atlantic, West Africa, and parts of the Mediterranean (Fig. 1a). With this large domain we intend to encompass all relevant source regions of moisture for the SHL, which can be convectively lifted into the Saharan air layer. The simulation is performed with explicit convection at a horizontal grid spacing of 0.125∘ (in rotated coordinates, corresponding to approximately 14 km) and with 60 hybrid levels in the vertical. Marsham et al. (2013b), Pearson et al. (2014), and Pante and Knippertz (2019) showed that explicit convection leads to a more realistic representation of the West African Monsoon already with a relatively coarse model resolution on the order of 10 km. Initial and lateral boundary conditions are provided every 6 h by the isotope-enabled global climate model ECHAM5-wiso (Werner et al., 2011) at a spectral resolution of T106 (corresponding to a horizontal grid spacing of approximately 1∘) and on 31 vertical levels. In order to keep the long simulation close to reality, horizontal winds in COSMOiso at 850 hPa and above are spectrally nudged towards ECHAM5-wiso. The nudging is performed at every time step (60 s) and operates at zonal and meridional wavenumbers of five and less. All other COSMOiso fields run freely in the model domain. The spectral nudging technique at small wavenumbers forces the large-scale flow in the limited-area model towards the large-scale flow in the global climate model without directly affecting small-scale weather features (von Storch et al., 2000; Schubert-Frisius et al., 2017). The ECHAM5-wiso, in turn, was nudged towards the ERA-Interim reanalysis dataset from the European Centre for Medium Range Weather Forecasts (Dee et al., 2011). The ECHAM5-wiso nudging also includes temperature and surface pressure.

To identify the SHL location from the COSMOiso simulation output, we use the methodology proposed by Lavaysse et al. (2009). The heat low detection criterion relies on the cumulative probability distribution of low-level atmospheric thickness between 700 and 925 hPa at 06:00 UTC over West Africa. A high percentile of this variable is used as a threshold to define the area of the SHL. Unlike Lavaysse et al. (2009), we always average the thickness values over 2 d before computing the cumulative probability distribution in order to have a smoother transition of the SHL location between 2 consecutive days. Moreover, in contrast to Lavaysse et al. (2009), we use the 80th percentile instead of the 90th percentile since the COSMOiso domain is smaller than the area used by Lavaysse et al. (2009), which extends about 20∘ further east and 5∘ further south. Comparing the identified SHLs in COSMOiso and ERA-Interim, the latter calculated from the 90th percentile in the larger domain, confirms that our chosen threshold leads to a realistic identification in terms of location and spatial extent. The SHL masks used in this study only contain grid points over continental Africa; i.e. points over the adjacent subtropical North Atlantic are not considered. Finally, only SHL features with a circumference larger than 1000 km are considered in order to avoid a fragmented structure of the SHL, which could occur due to the high spatial resolution of COSMOiso. The top of the SHL is defined by the planetary boundary layer height as diagnosed in COSMOiso with a bulk Richardson number criterion.

For the interpretation of observed and simulated water vapour isotope signals, we use kinematic backward trajectories computed from three-dimensional hourly COSMOiso wind fields with the Lagrangian analysis tool LAGRANTO (Wernli and Davies, 1997; Sprenger and Wernli, 2015). The trajectories start every hour and from every 20 hPa between 300 and 900 hPa above Tenerife (28.30∘ N, 16.48∘ W). They are calculated 10 d backward in time or until they leave the model domain. Different COSMOiso variables such as δD, qv, qi, qc, rain water qr, snow water qs, surface precipitation, and boundary layer height are interpolated along each trajectory and stored together with the trajectory position every hour. In addition, we attribute each trajectory to one of four predominant transport pathways, which we define as follows for this study: (1) North African air originating from the convective boundary layer of the SHL (hereafter referred to as SHL BL); (2) North African air coming from the free troposphere above the SHL (SHL FT); (3) North African air from outside the SHL region (typically descending from tropical upper levels south of the SHL; TRP AFR); and (4) North Atlantic air (typically subsiding from the upper-level extratropical North Atlantic; NA). We attribute a trajectory to the SHL BL regime if the air parcel was at least at one time step in the convective boundary layer of the SHL during its travel; to the SHL FT regime if the air parcel was at least at one time step above but never in the boundary layer of the SHL during its travel over the SHL; to the TRP AFR regime if the air parcel was at least once above continental Africa but never in or above the SHL; and to the NA transport pathway if the air parcel was never above continental Africa. We will investigate whether these four transport pathways have a distinct isotopic signature and lead to contrasting atmospheric conditions on timescales of hours to days in the Canary Islands region.

We validate the different transport pathways in COSMOiso by comparing trajectories calculated with winds from COSMOiso and 6-hourly ERA-Interim fields. The ERA-Interim trajectories start every hour from the same starting positions as the COSMOiso trajectories and run 10 d backward in time. The validation criterion is based on comparing the origin of the COSMOiso trajectories with the origin of the ERA-Interim trajectories, which are regarded as “truth”. We distinguish between trajectories from continental Africa and trajectories from the North Atlantic. For each arrival time and vertical level of the trajectory starting profile, trajectories arriving within a 6 h interval centred at the considered arrival time are compared. A 6 h time window for the comparison is chosen to achieve a more robust validation. If at least four of the seven COSMOiso trajectories in the considered time window and at the specific vertical level agree with the corresponding ERA-Interim trajectories about the origin (continental Africa vs. North Atlantic), the transport in COSMOiso is considered “reliable” at the respective trajectory arrival time and vertical level.

For the comparison between COSMOiso and ground-based FTIR remote sensing observations, we regard a specific observation time as reliable if at least 75 % of the levels between 400–700 hPa (representative of the 4.9 km FTIR retrieval level) are associated with a reliable transport according to the aforementioned evaluation criterion.

Figure 3 shows the comparison of COSMOisoqv and δD in water vapour with the aircraft observations. The flights were characterised by a fast ascent up to 7 km altitude followed by a slow step-wise descent. They covered an area of about 50 km in the zonal and 25 km in the meridional direction over the ocean south of Tenerife (Fig. 1b). Each of the seven flights was performed in the morning during 1 to 2.5 h, yielding a total flight time of about 14 h. The vertical airborne profiles reveal large variability in humidity and isotopic composition. Specifically, the measured qv and δD values range from 0.08 to 17 gkg-1 and from -590 ‰ to -50 ‰ along the flight paths with fast variations. COSMOiso reasonably captures this observed variability that is mostly associated with the vertical gradient in qv and δD. While modelled and observed qv and δD values agree well in the lower troposphere (mean absolute differences at 0–2 km a.s.l.: Δqv(0–2 km)=2.19gkg-1, ΔδD(0–2 km)=38‰), differences occur for both variables in the middle and upper troposphere (Δqv(2–7 km)=0.89gkg-1, ΔδD(2–7 km)=71‰) and are largest around 6 km a.s.l. (Δqv(>6km)=0.57gkg-1, ΔδD(>6km)=114‰). At this altitude, qv and δD tend to be too high in COSMOiso compared to the measurements. These differences can be partly explained by synoptic-scale uncertainties in COSMOiso associated with the very heterogeneous meteorological conditions that occurred during the flights (see Dyroff et al., 2015). Considering the synoptic-scale variability of qv and δD in COSMOiso in the Canary Islands region by sampling minimum and maximum values in a horizontal 5∘×5∘ box around each data point along the flight path (red and blue shading in Fig. 3) reveals that the differences are mostly not significant, except for 30 July. On this day, aircraft measurements show very strong spatial gradients in qv and δD around 5.5 km altitude, which are not represented in COSMOiso in the considered box. In general, however, COSMOiso is able to capture the observed vertical humidity and isotope profiles in the middle to upper troposphere, albeit with some horizontal displacements.

The strong horizontal gradients in qv and δD at 500 hPa (Fig. 4a and b) highlight the very heterogenous conditions in the Canary Islands region (black box in Fig. 4) at 12:00 UTC 21 July 2013, which is representative of all flight days apart from 30 July. Slight spatial shifts of these gradients can lead to a rather different isotope and humidity signal in the area of the flight tracks (red box in Fig. 4). The comparison of qv at 500 hPa between COSMOiso and ERA-Interim (Fig. 4c) confirms that COSMOiso overall correctly simulates the large-scale distribution of the free-tropospheric moisture. Still, we can observe a north-westward shift of the frontal zone towards the Canaries in COSMOiso compared to ERA-Interim, where the front is located farther to the south-east along the African west coast. Due to the low spatial resolution of ERA-Interim, however, it is not clear how accurately ERA-Interim represents this frontal situation. Figure 4e and f reveal why the quantified synoptic uncertainty in COSMOiso does not reduce the differences between the modelled and observed qv and δD values on 30 July 2013. On this day, COSMOiso is far too moist in the Canary Islands region compared to ERA-Interim and consequently also too enriched in heavy isotopes. (In the following, for reasons of readability, we will only use the term “enriched” when referring to “enriched in heavy isotopes” and will proceed analogously for “depleted”.) Analysis of backward trajectories calculated from COSMOiso and ERA-Interim fields (see Fig. A1) reveal that the overestimated qv and δD values in the model most likely result from a moistening of the air parcels by deep convection over tropical West Africa about 5 d prior to arrival over the Canaries and again near the African coast 2 d prior to arrival. In addition, we have to consider that the very sharp vertical humidity gradients measured by the ISOWAT spectrometer (Fig. 9 in Dyroff et al., 2015) cannot be fully resolved by COSMOiso with a vertical resolution of about 400 m in the middle troposphere. The lower spatial and temporal resolution of the COSMOiso output fields compared to the airborne in situ measurements in general constitutes a source of uncertainty for the validation of modelled qv and δD.

Figure 5a displays the aircraft observations and the corresponding COSMOiso data in the {qv,δD} space. Both distributions have a very similar slope but the COSMOiso distribution covers a much narrower δD range and tends to be too enriched compared to the airborne data. Positive δD differences are present throughout the atmospheric column but most pronounced at low qv and δD values (Fig. 5a and b). The qv range covered by COSMOiso is comparable to the aircraft data and there is no clear tendency towards positive or negative differences from aircraft observations, except for the humid bias in COSMOiso at low qv (Fig. 5c). The representativeness of this model–aircraft comparison, however, is limited, since the data are not fully independent but comprise correlated observations from vertical qv and δD profiles that were performed on 7 specific campaign days.

In this section, we first compare satellite-based IASI observations with post-processed COSMOiso data that have the same vertical characteristics as the IASI water vapour isotope product (Sect. 2.1.3). The datasets consist of 18 063 IASI and 141 703 COSMOiso samples that were retrieved at 4.2 km altitude in a 10∘×10∘ box centred around Tenerife (Fig. 1a) in July and August 2013. This large number of observational and model data enables a statistically robust model validation. Figure 5d shows the {qv,δD}-pair distributions of COSMOiso and IASI. The distributions overlap to a great extent and cover ranges of qv from 0.4 gkg-1 to 5 gkg-1 and of δD from -360 ‰ to -130 ‰. The slope of the COSMOiso {qv,δD} distribution is somewhat less steep than the slope of the IASI distribution, yielding a positive qv model bias for high δD values (δD>-180‰) and a negative qv model bias for low δD values (δD<-320‰) compared to IASI. Comparing the COSMOiso and IASI δD distributions confirms the high consistency between the model and observations (Fig. 5e). Only a slightly negative model bias can be observed for higher δD values (δD>-260‰). Similarly, the comparison of the COSMOiso and IASI qv distributions underlines the good agreement in the middle troposphere in spite of small differences in the range of the highest and lowest qv values. Attributing the source of the qv and δD differences in COSMOiso is difficult owing to the complex characteristic of the remote sensing retrievals and the post-processed COSMOiso data. Synoptic-scale uncertainties in the model, as discussed in the context of the aircraft observations in Sect. 3.1, constitute a minor source of the observed qv and δD differences due to the large data sampling area.

The {qv,δD}-pair distributions of the ground-based FTIR remote sensing observations (Fig. 5g; type 2 product; see Sect. 2.1.2) and the corresponding COSMOiso data multiplied with the FTIR kernels (see Sect. 2.1.2) are similar to the distributions shown for the COSMOiso–IASI comparison (Fig. 5d). They cover a similar {qv,δD} range and in addition, the COSMOiso distribution also shows a slightly tilted slope compared to FTIR towards higher qv values for high-δD air and lower qv values for low-δD air. Likewise, the comparison of the FTIR (type 2 product) and COSMOiso δD distributions indicates a tendency towards δD values that are too low in the model compared to the observations (Fig. 5h). These differences, however, mainly occur on days with a less reliable air parcel transport in the model (see Sect. 2.4). Uncertainties in the transport of air parcels are often found during transitions from the North Atlantic flow regime to an African flow regime and vice versa and are related to heterogeneous atmospheric conditions (as discussed in Sect. 3.1), which imply that subtle changes in the initialisation of the trajectories (time, pressure) can lead to a completely different transport pathway. Finally, comparing modelled and observed (type 1 product; see Sect. 2.1.2) qv does not show a clear shift in COSMOiso (Fig. 5i), but in contrast to the comparison against space-based IASI data, the statistical representativeness of the ground-based remote sensing observations is limited. The 383 FTIR observations contain only a few independent samples since there are multiple observations per day in general (varying between 1 and 25 data points). Still, the high consistency between the comparisons with two different remote sensing datasets is encouraging and forms a sound basis for evaluating variations in qv and δD in COSMOiso vs. FTIR observations on short timescales. As a side note we mention that the inconsistency of the model comparison with aircraft in situ data vs. with remote sensing data might point to a positive δD bias of the remote sensing observations. However, we use the final MUSICA NDACC (FTIR) data product in this study, which is calibrated to the in situ aircraft profiles and therefore bias corrected (Schneider et al., 2016). In addition, comparing the different observational datasets is difficult since the in situ and remote sensing techniques sample the atmosphere differently in terms of spatial and temporal resolution.

We briefly come back to the setup of our COSMOiso simulation with explicitly resolved convection. Figure B1 in the appendix shows an alternative simulation with parameterised convection, which leads to larger and mostly positive model biases in comparison with airborne, ground-, and space-based observations. This is in agreement with previous studies, which have already reported positive mid-tropospheric δD biases in model simulations with parameterised convection due to an overestimated vertical moisture transport (e.g. Werner et al., 2011; Risi et al., 2012a, b; Christner et al., 2018).

Figure 6 shows the time series of the COSMOiso and FTIR qv (type 1 product) and δD (type 2 product) data. There is a pronounced day-to-day variability with qv ranging from 0.45 to 4.3 gkg-1 and δD from -318 ‰ to -128 ‰. Conditions that are moist and enriched in heavy isotopes change with dry and depleted conditions on a remarkably short timescale (e.g. from 18 to 19 July or from 29 to 30 July). Comparing COSMOiso to the FTIR data demonstrates that our simulation accurately reproduces these observed temporal variations in mid-tropospheric qv and δD despite occasional deviations. These differences can occur due to various factors. In the first place, we observe that the deviations are largest on days with a less reliable air parcel transport (see Sect. 2.4 and discussion below in the context of Fig. 7). Synoptic-scale uncertainties in COSMOiso, as already mentioned in Sect. 3.1, may thus partly explain the observed differences. Moreover, model uncertainties in the representation of complex, sub-grid-scale processes in the vicinity of the mountain slopes of Tenerife could also contribute to differences between the model and observations. Finally, there are also uncertainties associated with the remote sensing retrievals and the COSMOiso post-processing. Overall, however, we can conclude that COSMOiso realistically represents the variability in mid-tropospheric qv and δD including fast day-to-day variations.

In Fig. 7, we present the qv and δD composition in the vertical column between 300 and 900 hPa above Tenerife as simulated by COSMOiso for the period of July–August 2013. Substantial variability appears in the qv and δD signals (Fig. 7a and b). In the middle troposphere, short intervals (1–5 d) of dry, low-δD air (δD<-260‰) alternate with longer periods (3–11 d) of humid, high-δD air (δD>-220‰). Transitions between episodes often occur within a few hours and are characterised by sharp qv and δD gradients (e.g. on 19 July and 17 August 2013; see also Fig. 4a–c). The enriched periods themselves also show fast, albeit less pronounced, variations in δD on timescales of hours to days. These contrasting humidity and isotope conditions are closely linked to the four predominant transport pathways (see Sect. 2.4). While the dry and low-δD air primarily comes from the North Atlantic (NA, blue shading in Fig. 7c), the humid and high-δD air mainly originates from West Africa, where we distinguish between air originating from the convective boundary layer of the SHL (SHL BL, red shading in Fig. 7c), the free troposphere above the SHL (SHL FT, purple shading in Fig. 7c), and tropical Africa south of the SHL (TRP AFR, green shading in Fig. 7c). Most pronounced temporal variations in δD occur during transitions from an African transport regime to the NA regime and vice versa. For example, on 19 July and 17 August a strong depletion in heavy isotopes in mid-tropospheric water vapour from about -160 ‰ to -320 ‰ occurs within a few hours, simultaneously with a transition from an African to the NA transport regime. The fast δD variations during enriched periods can be partly explained by changes in the African transport pathways, where higher δD signals tend to appear during the SHL BL transport regime compared to lower δD signals during the TRP AFR regime (e.g. 20–23 August). Not surprisingly, air parcels from tropical Africa (TRP AFR) are more often affected by surface precipitation (black dots in Fig. 7c) and hydrometeors (not shown) than air parcels from the SHL region (SHL BL/FT). The formation of cloud water and precipitation directly influences the water vapour isotopic composition of an air parcel, since heavy isotopes preferentially condense, whereby the remaining ambient water vapour becomes depleted in heavy isotopes. As a result, air parcels from TRP AFR are associated with lower δD values due to condensation in moist convective systems in the Sahel compared to air parcels from the SHL region, where dry convection dominates. For the further analysis of physical and dynamical processes associated with the four contrasting transport pathways, we consider the air layer between 500 and 700 hPa (indicated by the black horizontal lines in Fig. 7). This layer is particularly interesting from a dynamical viewpoint as the large-scale circulation dominates the humidity and isotope composition at these altitudes, whereas local mixing processes between the boundary layer and the free troposphere are less relevant than at lower levels. In addition, the remote sensing observations, against which we validated COSMOiso, have their highest sensitivity around 600 hPa.

The statistical analysis of the mid-tropospheric δD signal highlights that each of the four transport pathways has a distinct isotopic signature (Fig. 8). The NA transport pathway is the most depleted (median δD=-255‰) and most variable (interquartile range IQR=92‰) regime, whereas the SHL BL pathway represents the most enriched (median δD=-122‰) and most homogeneous (IQR=21‰) category (Fig. 8a). Air parcels of the SHL FT (median δD=-148‰, IQR=47‰) and TRP AFR (median δD=-175‰, IQR=55‰) transport regimes are less enriched in heavy isotopes than air parcels of the SHL BL regime but more enriched than NA (Fig. 8a). In general, the three African transport pathways distinctly differ from the NA pathway with respect to δD as IQRs of the distributions do not overlap (Fig. 8a). The {qv,δD}-pair distributions of the African and NA regimes differ in a similar manner (Fig. 8b). The three African transport pathways also show some differences between their {qv,δD}-pair distributions. There is a clear contrast in the δD range of the SHL FT and TRP AFR distributions for the 60 % contour, whereas the qv ranges largely overlap (Fig. 8b). Hence, these two transport pathways have a similar qv that apparently results from different physical processes. While the SHL FT pathway is primarily influenced by dry convective mixing without fractionation, the TRP AFR pathway is mainly affected by moist convection and thus by microphysical processes in addition to mixing. Thanks to the characteristic isotopic signature, a discrimination between these two fundamentally different transport pathways is possible. This emphasises the added value of water vapour isotopes for investigating physical processes and transport pathways that affect the subtropical tropospheric humidity. Figure 8b further highlights the homogeneous character of the SHL BL category as well as its particularly high δD and qv values. By contrast, the {qv, δD}-pair distribution of the NA category reflects the dryness and depletion in heavy isotopes of this transport pathway.

This analysis confirms the current state of knowledge about the contrasting moisture transport conditions over the eastern subtropical North Atlantic, resulting from an alternation of humid, isotopically enriched air primarily coming from Africa and of dry, depleted air mainly originating from the upper-level extratropical North Atlantic (González et al., 2016; Lacour et al., 2017). In addition, our work shows that North African air masses affected by the SHL (pathways SHL BL and SHL FT) and air masses originating from the Sahel region further south (pathway TRP AFR) are associated with a distinct isotope signature. The combination of high-resolution numerical isotope modelling with multi-platform isotope observations, which represents an expansion of the previous observation-oriented studies by González et al. (2016) and Lacour et al. (2017), offered the possibility to directly link the observed synoptic timescale variability of specific humidity and isotope composition to the origin of moisture. In particular, it allows for studying the isotopic composition along backward trajectories from the Canary Islands region and thereby disentangling the governing physical processes that affect the subtropical free-tropospheric moisture budget (see Sect. 4.2).

Detailed analysis of isotopic signals along backward trajectories provides valuable insights into the process history of air parcels. Figure 9 summarises the median 10 d Lagrangian evolution of pressure p, δD, and qv of the four different transport pathways identified in this study. The median SHL BL pathway involves weak subsidence from about 550 to 640 hPa (Δp4-10d=90hPa(6d)-1) 4–10 d before arrival, followed by a lifting to 600 hPa (Δp1-4d=-30hPa(3d)-1) 1–4 d prior to arrival and a subsequent descent to 620 hPa (Δp0-1d=20hPa(1d)-1) during the last 2 d before arrival over the Canary Islands. The initial subsidence goes along with an increase in δD (ΔδD4-10d=50‰(6d)-1) and qv (Δqv,4-10d=1.8gkg-1(6d)-1). The subsequent lifting occurs without changes in the median δD and qv. The SHL BL pathway is further characterised by substantial variability in p and δD 4–10 d before arrival, which considerably reduces in the last 3 d before arrival. Both are related to SHL dynamics. Convective mixing in the SHL of dry, low-δD air from the upper-level extratropics with moist, high-δD air from low levels effectively homogenises the isotopic composition, producing a well-mixed and relatively homogeneous air mass that subsequently travels towards the Canary Islands. The homogeneous, enriched composition of the SHL BL air is a unique characteristic of this transport pathway (Fig. 8a) and a clear imprint of the dry convective mixing in the SHL. Figure 10a and b show a representative example of the SHL BL pathway. Air parcels with a low δD descend from the upper-level extratropical North Atlantic around the Atlas Mountains into the SHL (red squares in Fig. 10a), where they experience considerable enrichment in heavy isotopes and moistening due to dry convective mixing with high-δD air from lower levels. The moisture of these low-level air parcels presumably originates from evaporation over the eastern Mediterranean and the Black Sea, as identified with extended ERA-Interim backward trajectories (Fig. 11). Two principal moisture transport pathways dominate the low-level SHL inflow at the time when the air parcels in Fig. 10a and b are located in the SHL (Fig. 11). The first pathway, the so-called south-westerly monsoon flow, brings very humid air from the tropical Atlantic into the southern part of the SHL. The second pathway, the so-called north-easterly Harmattan flow, transports air from north-eastern Europe into the central and northern part of the SHL. These low-level air parcels experience a considerable moisture uptake during their transport across the eastern Mediterranean, with minor contributions from the Black Sea, before they mix with upper-level extratropical air in the SHL. The well-mixed, isotopically homogeneous air then rapidly moves as a coherent air mass over the adjacent North Atlantic directly towards Tenerife (Fig. 10a and b). This direct and fast transport of air towards the Canary Islands is typical for the SHL BL pathway. It essentially conserves the isotopic imprint of the dry convective mixing in the SHL, which is reflected by the constant δD signal along the trajectories during the last 3 d before arrival (Fig. 9).

Although originating from a similar region, air parcels of the SHL FT category follow a different pathway (Fig. 9). Typically, air parcels with a very low initial δD (median δD=-390‰) from the upper troposphere experience a dramatic median increase in δD (ΔδD0-10d=243‰(10d)-1) and a strong moistening (Δqv,0-10d=2.2gkg-1(10d)-1) along with a strong subsidence (Δp0-10d=245hPa(10d)-1) on their 10 d travel towards the Canary Islands. This indicates that at least in our simulation, there is vigorous exchange across the inversion on top of the SHL. This transport pathway exhibits the largest median changes in humidity and isotopic composition among the four identified transport pathways. The pronounced enrichment in heavy isotopes usually occurs in the free troposphere above the SHL due to mixing in of high-δD air that is injected from the boundary layer of the SHL into the free troposphere. Figure 10c and d illustrate how upper-level air parcels get enriched while anticyclonically rotating above the SHL (purple squares in Fig. 10c) before they move off the African coast and slowly approach the Canary Islands from the south. Note that these trajectories move slower than the ones in the SHL BL category and that they therefore reside over West Africa for most of the 10 d period.

The TRP AFR pathway reflects the typical transport of air parcels that descend from upper-level tropical Africa towards the Canary Islands (Fig. 9). The strong median subsidence (Δp0-10d=214hPa(10d)-1) over the Sahel and tropical West Africa goes along with a moderate increase in δD (ΔδD0-10d=109‰(10d)-1) and qv (Δqv,0-10d=1.6gkg-1(10d)-1). In summer, during the West African Monsoon, the Sahel, and therefore air parcels of the TRP AFR pathway, are frequently affected by mesoscale convective systems. This implies that moist convective mixing regularly influences air parcels of the TRP AFR pathway, in contrast to the SHL BL and SHL FT pathways where dry convective mixing is dominant. The exemplary trajectories presented in Fig. 10e and f show the interplay between subsiding air parcels from the upper-level Sahel and ascending air parcels from the low-level tropical Atlantic and Africa. The low-level air parcels are lifted to higher altitudes thereby experiencing rainout (blue squares in Fig. 10e) and depletion in heavy isotopes. The subsiding air parcels, in turn, get enriched during their subsidence over the Sahel due to mixing with air from lower altitudes. However, the increase in δD due to this effect is only minor, since the rising lower-level air experiences depletion in heavy isotopes due to condensation in the convective system. The moist convective mixing over the Sahel leads to a homogenisation of the different airstreams, which subsequently travel over the adjacent tropical Atlantic towards the Canary Islands (Fig. 10e and f). Because of the strong variability in the occurrence of mesoscale convective systems over West Africa, the TRP AFR pathway shows high variability (Fig. 9).

Finally, the NA pathway presented in Fig. 9 summarises the history of dry air parcels strongly depleted in heavy isotopes that originate from the upper-level extratropical North Atlantic. This transport pathway leads to the lowest δD and qv values observed in the Canary Islands region. The humidity and isotopic composition show a slight to moderate change (Δqv,0-10d=0.6gkg-1(10d)-1, ΔδD0-10d=184‰(10d)-1) during the comparatively fast descent (Δp0-10d=295hPa(10d)-1) of the air parcels with low initial δD values (median δD=-493‰) in an environment with regular precipitation events but without deep convective systems. A representative example of the NA pathway is given in Fig. 10g and h. The descending air parcels with a very low δD are either directly advected from the upper-level extratropical North Atlantic towards the Canary Islands or, in a few cases, first lifted from the lower-level extratropical North Atlantic to higher altitudes thereby experiencing rainout (blue squares in Fig. 10g) and depletion in heavy isotopes.

Figure 12 summarises the four different transport pathways by showing the spatial distribution of the air parcels' location 3 d before arrival over Tenerife. Air parcels of the SHL BL pathway are almost exclusively located over the Sahara (Mauretania, Mali, Algeria) before they rapidly move over the adjacent subtropical North Atlantic towards the Canary Islands (Fig. 12a). The direct transport of air parcels from the West African coast towards the Canaries is unique for the SHL BL pathway and effectively conserves the isotopic imprint of the dry convective mixing of low-level moisture of diverse origin advected into the SHL. The SHL FT pathway, in contrast, has the highest air parcel location frequency over the subtropical North Atlantic south of the Canary Islands and over Mauretania (Fig. 12b). These air parcels typically reside several days over the subtropical North Atlantic before reaching the Canary Islands from the south. They potentially experience mixing with enriched air over the North Atlantic, which is reflected by the increase in δD in water vapour during the last 3 d before arrival in Fig. 9. Air parcels of the TRP AFR pathway are predominantly located over the (sub)tropical North Atlantic 3 d prior to arrival in the Canary Islands region (Fig. 12c). In addition, they are also found along the West African west coast and over the Sahel. The broad spread in the air parcels' location underlines the large variability of the TRP AFR regime in terms of pathways of individual trajectories and processes that occur along them. The NA pathway also exhibits substantial variability in the air parcels' location over the North Atlantic (Fig. 12d). This variability is presumably closely related to the synoptic variability in the extratropical North Atlantic storm track region and explains the broad δD distribution of this category (Fig. 8). The peaks in Fig. 12d imply that air parcels either subside directly from the upper-level extratropical North Atlantic towards the Canary Islands or first travel across the North Atlantic, which often goes along with precipitation, before descending into the subtropics to the west of the Canary Islands region.

This section focuses on investigating the large-scale flow conditions associated with the four transport pathways. Previous work has shown that even during the warm season the investigation region is frequently affected by positively tilted upper-level intrusions of high potential vorticity (PV) from higher latitudes, usually associated with Rossby wave breaking over the North Atlantic (Fröhlich and Knippertz, 2008; Papin et al., 2020). Ahead of these troughs, moist mid-level air can be transported northwards around the western flank of the anticyclone overlaying the low-level SHL and lead to precipitation in north-western Africa (Knippertz et al., 2003; Knippertz, 2003). Upper-level troughs and ridges together with African easterly waves determine synoptic-scale fluctuations of the SHL and its associated mid- to upper-level anticyclone (Lavaysse et al., 2010a) as well as of dust export from Africa to the Atlantic (Cuevas et al., 2017).

For each of the four transport pathways, we compute ERA-Interim composites of the geopotential height z at 600 hPa and anomalies from the July and August climatological mean in the period 1979–2018. The 6-hourly geopotential height fields in July and August 2013 are assigned to a transport pathway according to the origin of air parcels arriving at 600 hPa above Tenerife in COSMOiso at the corresponding time step. The composites depict the synoptic-scale conditions 3 d prior to the arrival of the respective air parcels. For reasons of consistency, we only consider time steps when COSMOiso and ERA-Interim trajectories both originate either from Africa or from the North Atlantic.

The analysis reveals distinct geopotential height anomaly patterns in the extratropical storm track region for the SHL BL, TRP AFR, and NA pathways (Fig. 13). The most frequent (32 %) SHL BL transport regime shows a large-scale wave over the North Atlantic with positive z anomalies east of the Canadian east coast and over Europe (Fig. 13a). The latter has a north-east–south-west tilt and expands to the Canary Islands region. In combination with negative anomalies over most of the Sahara, the centre of the North African mid-level anticyclone shifts north-westward to the Moroccan Atlas Mountains such that Saharan air can reach the Canary Islands flowing along its southern flank. Cuevas et al. (2017) referred to this as a high NAFDI (North African Dipole Intensity) index situation.

The relatively rare (16 %) TRP AFR regime also shows a large-scale wave over the North Atlantic with positive z anomalies over the Canadian east coast and over Europe (Fig. 13b). In contrast to the SHL BL pathway, however, the eastern anomaly is confined to continental Europe. Instead, a negative z anomaly with two centres prevails over the eastern extratropical North Atlantic: one associated with an upper-level trough near 45∘ N and one south-west of the Canary Islands region, indicative of an upper-level cutoff and PV intrusion. The North African mid-level anticyclone is centred over north-western Algeria, close to its climatological mean position but with a slightly enhanced amplitude. This dynamical environment enables a direct transport of air parcels from tropical West Africa along the south-western flank of the mid-level anticyclone similar to the situation described in Knippertz et al. (2003) and Knippertz (2003). The enhanced zonal gradient of z south of the Canary Islands then favours intensified northward transport towards the islands. Cuevas et al. (2017) referred to this as a low NAFDI index situation. For the SHL FT pathway, the geopotential height anomaly pattern (not shown) is similar to that of TRP AFR, albeit with less pronounced anomalies in the extratropical storm track region and a weaker trough over the central North Atlantic.

By contrast, the NA transport regime (26 %) is characterised by a pronounced positive z anomaly over the extratropical North Atlantic near the British Isles associated with a ridge to the west and north-west of the Iberian Peninsula as well as a negative anomaly over the Mediterranean Sea (Fig. 13c). These anomalies go along with a clear south-westward shift of the North African mid-level anticyclone over the adjacent subtropical North Atlantic, leading to overall more zonal flow. As a result, African air parcels that are transported westward with the African easterly jet are prevented from reaching the Canary Islands region. The ridge over the extratropical North Atlantic induces subsidence of upper-level air parcels southwards to the Canary Islands.

In summary, this analysis is in agreement with previous work that extratropical Rossby wave dynamics strongly influences the mid-tropospheric transport in the study region and thereby the moisture and isotope variability. Specifically for the Canary Islands, the position and zonal extent of the anticyclone above the SHL determines the influx of air from tropical Africa, the Sahara, or the North Atlantic.

We complete this study by analysing the relevance of the four predominant transport pathways from a climatological perspective. To this end, we compute 10 d ERA-Interim backward trajectories for the period 1979–2018 started every 6 h and from every 20 hPa between 500 and 700 hPa above Tenerife. The trajectories are then grouped into the four different transport pathways (see Sect. 2.4).

The monthly occurrence frequencies of the different transport pathways emphasise the dominant role of the SHL BL and SHL FT categories in July and August (summarised as SHL in Fig. 14). With a median occurrence frequency of about 50 % in these 2 months, SHL pathway clearly exceeds the frequencies of the TRP AFR (∼15 %) and NA (∼33 %) pathways. The rapid increase in the median occurrence frequency of the SHL pathway from 13 % in June to 50 % in July and the gradual decrease in September and October to 6 % are closely related to the SHL onset at the end of June (20 June±9d in the climatological mean; Lavaysse et al., 2009) and its decay in September (17 September±7d). This is in agreement with Lacour et al. (2017), who showed that the seasonality in the transport of air from the SHL region is linked to the SHL activity. The enhanced transport of mid-tropospheric air from the SHL region in summer develops at the expense of the NA transport regime, which clearly dominates during the rest of the year. With the onset of the SHL at the end of June, the median occurrence frequency of the NA pathway drops from 78 % in June to 36 % and 30 % in July and August, respectively, before reaching again values of around 80 % in October. The occurrence frequency of the TRP AFR pathway is less affected by the seasonal evolution of the SHL and varies less in the course of the year (monthly values between 5 % and 22 %).

From October to May, the NA transport regime clearly dominates (85 % occurrence frequency on average) and alternates with the less frequent TRP AFR regime (13 % on average). Note that the TRP AFR pathway comprises by definition North African air from outside the SHL (see Sect. 2.4). This implies that in the absence of the SHL, the TRP AFR pathway also includes air that originates from the Sahara, unlike in summer, when the SHL is fully established and the TRP AFR transport typically represents air from tropical upper levels south of the SHL (see Sect. 4.2).

The interannual variability of the monthly occurrence frequencies of the SHL, TRP AFR, and NA transport pathways is moderate (interquartile ranges in Fig. 14) and amounts to 19 %, 9 %, and 18 %, respectively, in July and August (Table 1). In July and August 2013 (studied in detail in this paper), the transport of mid-tropospheric air from the SHL region is more frequent (58 %) compared to the 1979–2018 climatology (49 %). Specifically, the SHL BL and SHL FT pathways both occur about 5 % more often (34 % and 24 %) than in the climatological mean (29 % and 20 %). The NA pathway, by contrast, is underrepresented in July and August 2013 (24 %) compared to the climatology (35 %). The occurrence of the TRP AFR pathway (18 %) roughly corresponds to the climatological mean (16 %), see also Table 1. We analysed the relationship between teleconnection indices

Data downloaded from the NOAA Physical Sciences Laboratory: https://psl.noaa.gov/data/climateindices/list/ (last access: 20 August 2021).

The occurrence frequencies of the four transport pathways in COSMOiso are approximately equal to ERA-Interim in July and August 2013, with differences not exceeding 2 % (Table 1). Accordingly, the SHL BL and SHL FT pathways appear more often in the time period considered in this study than climatologically, while the NA transport occurs less frequently. However, as the frequencies of the four transport pathways lie within the climatological interquartile ranges, the results of this study are representative of the contrasting atmospheric conditions that prevail in the Canary Islands region in summer.