The LMDz-INCA model couples a general circulation model developed at the Laboratoire de Météorologie Dynamique (LMD; Hourdin et al., 2006) and a global chemistry and aerosol model INteractions between Chemistry and Aerosols (INCA; Folberth et al., 2006; Hauglustaine et al., 2004). A more recent description of LMDz-INCA is presented in Szopa et al. (2013). To simulate CH4 and CO2 concentrations, we run a standard version of the model with a horizontal resolution of 2.5∘ in longitude (i.e., 144 model grids) and 1.27∘ in latitude (i.e., 142 model grids) (hereafter this version is abbreviated as “STs”) and a zoomed version with the same number of grid boxes, but a resolution of ∼0.66∘ in longitude and ∼0.51∘ in latitude in a region of 50–130∘ E and 0–55∘ N centered over India and China (hereafter this version is abbreviated as “ZAs”) (Fig. 1; see also Wang et al., 2014, 2016). It means that, in terms of the surface area, a grid cell from STs roughly contains 9 grid-cells from ZAs within the zoomed region. Both model versions are run with 19 and 39 sigma-pressure layers, thus rendering four combinations of horizontal and vertical resolutions (i.e., ST19, ZA19, ST39, ZA39). Vertical diffusion and deep convection are parameterized following the schemes of Louis (1979) and Tiedtke (1989), respectively. The simulated horizontal wind vectors (u and v) are nudged towards the 6-hourly European Center for Medium Range Weather Forecast (ECMWF) reanalysis dataset (ERA-I) in order to simulate the observed large-scale advection (Hourdin and Issartel, 2000).

The atmospheric concentrations of hydroxyl radicals (OH), the main sink of atmospheric CH4, are produced from a simulation at a horizontal resolution of 3.75∘ in longitude (i.e., 96 model grids) and 1.9∘ in latitude (i.e., 95 model grids) with the full INCA tropospheric photochemistry scheme (Folberth et al., 2006; Hauglustaine et al., 2004, 2014). The OH fields are climatological monthly data and are regridded to the standard and zoomed model grids, respectively. It should be noted that the spatiotemporal distributions of the OH concentrations have large uncertainties and vary greatly among different chemical transport models; therefore, the choice of the OH fields may affect the evaluation for CH4 (especially in terms of the annual gradients between stations and the seasonal cycles). In this study, as we focus more on the improvement of performance gained from refinement of the model resolution rather than model–observation misfits and model bias in CH4 growth rates, the influences of OH variations on model improvement are assumed to be very small given that the OH fields for both ZAs and STs are regridded from a lower model resolution and thus don't show much difference between the two model versions.

The CH4 and CO2 concentrations are simulated over the period 2000–2013 with both STs and ZAs. The first 6 years (2000–2005) of the simulations are considered as model spin-up; thus, we only compare the simulated CH4 and CO2 concentrations with observations during 2006–2013. The initial CH4 concentration field is defined based on the optimized initial state from a CH4 inversion that assimilates observations from 50+ global background stations over the period 2006–2012 (Locatelli, 2014; Locatelli et al., 2015b). The optimized initial CH4 concentration field for the year 2006 is rescaled to the levels of the year 2000 and used as the initial state in our simulations. The time step of model outputs is hourly.

The prescribed CH4 and CO2 surface fluxes used as model inputs are presented in Table 1. We simulate the CH4 concentration fields using a combination of the following datasets: (1) the interannually varying anthropogenic emissions obtained from the Emission Database for Global Atmospheric Research (EDGAR) v4.2 FT2010 product (http://edgar.jrc.ec.europa.eu, last access: 21 October 2016), including emissions from rice cultivation with the seasonal variations based on Matthews et al. (1991) imposed to the original yearly data; (2) climatological wetland emissions based on the scheme developed by Kaplan et al. (2006); (3) interannually and seasonally varying biomass burning emissions from Global Fire Emissions Database (GFED) v4.1 product (Randerson et al., 2012; Van Der Werf et al., 2017; http://www.globalfiredata.org/, last access: 3 May 2017); (4) climatological termite emissions (Sanderson, 1996); (5) climatological ocean emissions (Lambert and Schmidt, 1993); and (6) climatological soil uptake (Ridgwell et al., 1999). Note that for anthropogenic emissions from sectors other than rice cultivation, the seasonal variations are much smaller, and a monthly sector-specific dataset is currently not available for the whole study period. Therefore we do not consider seasonal variations in CH4 emissions from those sectors. Based on these emission fields, the global CH4 emissions in 2010 are 543 TgCH4yr-1 and 191 TgCH4yr-1 over the zoomed region. For the years over which CH4 anthropogenic emissions were not available from the data sources when the simulations were performed (namely, the years 2011–2013), we use emissions for the year 2010.

The prescribed CO2 fluxes used to simulate the concentration fields are based on the following datasets: (1) three variants (hourly, daily and monthly means) of interannually varying fossil fuel emissions produced by the Institut für Energiewirtschaft und Rationelle Energieanwendung (IER), Universität Stuttgart, on the basis of EDGARv4.2 product (hereafter IER-EDGAR, http://carbones.ier.uni-stuttgart.de/wms/index.html, last access: 14 December 2014) (Pregger et al., 2007); (2) interannually and seasonally varying biomass burning emission from GFEDv4.1 (Randerson et al., 2012; Van Der Werf et al., 2017; http://www.globalfiredata.org/, last access: 3 March 2017); (3) interannually and hourly varying terrestrial biospheric fluxes produced from outputs of the Organizing Carbon and Hydrology in Dynamic EcosystEms (ORCHIDEE) model; and (4) interannually and seasonally varying air–sea CO2 gas exchange maps developed by NOAA's Pacific Marine Environmental Laboratory (PMEL) and Atlantic Oceanographic and Meteorological Laboratory (AOML) groups (Park et al., 2010). Here ORCHIDEE runs with the trunk version r1882 (source code available at http://forge.ipsl.jussieu.fr/orchidee/wiki/SourceCode/ORCHIDEE, last access: 3 June 2018, with the revision number of r1882), using the same simulation protocol as the SG3 simulation in MsTMIP project (Huntzinger et al., 2013). The climate forcing data are obtained from CRUNCEP v5.3.2, while the yearly land use maps, soil map and other forcing data (e.g., monthly CO2 concentrations) are as described in Wei et al. (2014). The sums of global net CO2 surface fluxes in 2010 are 6.9 PgCyr-1 and 3.9 PgCyr-1 over the zoomed region. For the CO2 fossil fuel emissions, the IER-EDGAR product is only available until 2009. To generate the emission maps for the years 2010–2013, we scale the emission spatial distribution in 2009 using the global totals for these years based on the EDGARv4.2FT2010 datasets. The detailed information for each surface flux is listed in Table 1.

As inversions use concentration gradients to optimize surface fluxes, it is important to have a metric based upon cross-site gradients. We take Hanle in India (HLE – 78.96∘ N, 32.78∘ E; 4517 ma.s.l., Fig. 1, Table 2) as a reference and calculate the mean annual gradients by subtracting CH4 (CO2) at HLE from those of other stations. HLE is a remote station in the free troposphere within SEA and is located far from any important source or sink areas for both CH4 and CO2. These characteristics make HLE an appropriate reference to calculate the gradients between stations. Concentration gradients to HLE are calculated for both observations and model simulations using the corresponding smoothed curves fitted with the CCGvu routine (see Sect. 2.3). The ability of ZAs and STs to represent the observed CH4 (CO2) annual gradients across all the available stations is quantified by the mean bias (MB, Eq. 1) and the root-mean-square deviation (RMSE, Eq. 2). In Eqs. (1) and (2), mi and oi indicate respectively the modeled and observed CH4 (CO2) mean annual gradient relative to HLE for a station i. MB=∑i=1Nmi-oiNRMSE=∑i=1Nmi-oi2N

Two metrics of the model ability to reproduce the observed CH4 (CO2) seasonal cycle are considered: the phase and the amplitude. For each station, the seasonal phase is evaluated by the Pearson correlation between the observed and simulated harmonics extracted from the original time series, whereas the seasonal cycle amplitude is evaluated by the ratio of the modeled to the observed seasonal peak-to-peak amplitudes based on the harmonics (Am/Ao).

For each station, the performance of ZAs and STs in representing the phase (timing) of the synoptic variability is evaluated by the Pearson correlation coefficient between the modeled and observed synoptic deviations (residuals) around the corresponding smoothed fitting curve (see Sect. 2.3), whereas the performance for the amplitude of the synoptic variability is quantified by the ratio of SDs of the residual concentration variability between the model and observations (i.e., normalized standard deviation, NSD, Eq. 3). Further, the overall ability of a model to represent the synoptic variability of CH4 (CO2) at a station is quantified by the RMSE (Eq. 4), a metric that can be represented with the Pearson correlation and the NSD in a Taylor diagram (Taylor, 2001). In Eqs. (3) and (4), mj (oj) indicates the modeled (observed) synoptic event j, whereas m‾ (o‾) indicates the arithmetic mean of all the modeled (observed) synoptic events over the study period. Note that for the flask measurements, j corresponds to the time when a flask sample was taken, whereas for the continuous measurements, j corresponds to the early morning (00:00–03:00 LST, for mountain stations located higher than 1000 m a.s.l.) or afternoon (12:00–15:00 LST, for other stations) period of each sampling day. NSD=∑j=1Nmj-m‾2N∑j=1Noj-o‾2NRMSE=∑j=1Nmj-oj2N

For each station, the model's ability to reproduce the mean CH4 (CO2) diurnal cycle phase in a month is evaluated by the correlation of the hourly mean composite modeled and observed values, whereas model performance on the diurnal cycle amplitude is evaluated by the ratio of the modeled to the observed peak-to-peak amplitudes (Am/Ao). For each station, daily means are subtracted from the raw data to remove any influence of interannual, seasonal or even synoptic variations.

The annual mean gradient between a station and the HLE reference station relates to the time integral of transport of sources or sinks within the regional footprint area of the station on top of the background gradient caused by remote sources. For CH4, Fig. 2a and b shows the scatterplot of the simulated and observed mean annual gradients to HLE for all stations. In general, all the four model versions capture the observed CH4 gradients with reference to HLE, and the simulated gradients roughly distribute around the identity line (Fig. 2a and b). Compared to standard versions, the zoom versions (ZAs) better represent the CH4 gradients for stations within the zoomed region (closed circles in Fig. 2a and b), with RMSE decreasing by 20 and 16 % for 19- and 39-layer models (Fig. 2a and b and Table S1a). Note that increasing vertical resolution does not impact the overall model performance much, but the combination with the zoomed grid (i.e., ZA39) may inflate the model–observation misfits at a few stations with strong sources nearby (e.g., TAP and UUM in Table S2a). The better performance of ZAs within the zoomed region is also found for different seasons (Fig. S3). Outside the zoomed region (open circles in Fig. 2a and b), the performance of ZAs does not significantly deteriorate despite the coarser resolution.

When looking into the model performance for different station types, ZAs generally better capture the gradients at coastal and continental stations within the zoomed region, given the substantial reduction of RMSE compared to STs (Table S1). For example, significant model improvement is found at Shangdianzi (SDZ – 40.65∘ N, 117.12∘ E; 293 ma.s.l.) and Pondicherry (PON – 12.01∘ N, 79.86∘ E; 30 ma.s.l.) (Fig. 2a and b), with each having an average bias reduction of 28.1 (73.0 %) and 30.3 (94.7 %) ppb respectively compared to STs for the 39-layer model (Table S2). This improvement mainly results from reduction in representation error with higher model horizontal resolutions in the zoomed region through better description of surface fluxes and/or transport around the stations. Particularly, given the presence of large CH4 emission hotspots within the zoomed region (Fig. S4), ZAs makes the simulated CH4 fields more heterogeneous around emission hotspots (e.g., North China in Fig. S5), having the potential to better represent stations nearby on an annual basis if the surface fluxes are prescribed with sufficient accuracy.

However, finer resolutions may enhance model–data misfits due to inaccurate meteorological forcings and/or surface flux maps. For example, for the coastal station Tae-ahn Peninsula (TAP – 36.73∘ N, 126.13∘ E; 21 ma.s.l.) with significant emission sources nearby (Fig. S6), both ZAs and STs overestimate the observed CH4 gradients by >+15 ppb, and ZA39 perform even worse than other versions (Table S2). The poor model performance at TAP suggests that the prescribed emission sources are probably overestimated within the station's footprint area (also see the marine station GSN, Fig. S6), and higher model resolutions (whether in horizontal or in vertical) tend to inflate the model–observation misfits in this case. In addition, as stated in several previous studies (Geels et al., 2007; Law et al., 2008; Patra et al., 2008), for a station located in a complex terrain (e.g., coastal or mountain sites), the selection of an appropriate grid point and/or model level to represent an observation is challenging. In this study we sample the grid point and model level nearest to the location of the station, which may not be the best representation of the data sampling selection strategy (e.g., marine sector at coastal stations) and could contribute to the model–observation misfits.

Both ZAs and STs can generally capture the CO2 annual gradients between stations, although not as well as for CH4 (Fig. 2c and d). In contrast with CH4, ZAs do not significantly improve representation of CO2 gradients for stations within the zoomed region, with the mean bias and RMSE close to those of STs (Table S1b). At a few stations (e.g., TAP, Fig. S8), ZAs even degrade model performance (Table S2b), possibly related to misrepresentation of CO2 sources in the prescribed surface fluxes and transport effects. Again increasing model vertical resolution does not impact the overall model performance much.

With finer horizontal resolution, the model improvement to represent the annual gradients is more apparent for CH4 than for CO2. One of the reasons may point towards the quality of CO2 surface fluxes, especially natural ones. They are spatially more diffuse than those of CH4 and temporally more variable in response to weather changes (Parazoo et al., 2008; Wang et al., 2007). Therefore, the regional variations of net ecosystem exchange (NEE) not captured by the terrestrial ecosystem model (e.g., ORCHIDEE in this paper) may explain the worse model performance on the CO2 annual gradients compared to CH4 and less apparent model improvement. Further, the spatial resolution of the prescribed surface flux may also account for the difference in model improvement between CO2 and CH4 (e.g., the spatial resolution of anthropogenic emissions is 1∘ for CO2 and 0.1∘ for CH4). Therefore, with the current setup of surface fluxes (Table 1), ZAs are more likely to resolve the spatial heterogeneity of CH4 fields, and its improvement over STs is more apparent than that for CO2.

The model performance for the seasonal cycle depends on the quality of seasonal surface fluxes, atmospheric transport, and chemistry (for CH4 only). For CH4, both ZAs and STs capture the seasonal phases at most stations within the zoomed region very well (Fig. 3a), and model resolutions (in both horizontal and vertical) do not significantly impact the simulated timing of seasonal maximum and minimum. The seasonal phases at Plateau Assy (KZM – 43.25∘ N, 77.87∘ E; 2524 ma.s.l.), Waliguan (WLG – 36.28∘ N, 100.90∘ E; 3890 ma.s.l.) and Ulaan Uul (UUM – 44.45∘ N, 111.10∘ E; 1012 ma.s.l.) are not well represented, which is probably related to unresolved seasonally varying sources around these stations. The sensitivity test simulations prescribed with wetland emissions from ORCHIDEE outputs show much better model–observation agreement in seasonal phases (Fig. S9). For stations outside the zoomed region, the performance of ZAs is not degraded despite the coarser horizontal resolutions (Fig. S10).

With respect to the seasonal amplitude, the performance of STs and ZAs shows a significant difference at stations influenced by large emission sources. For example, the seasonal amplitudes of AMY and TAP are strongly overestimated by STs (Am/Ao=2.99 and Am/Ao=5.11 for the 39-layer model; Fig. 3a), while ZAs substantially decrease the simulated amplitudes at these two stations with improved model–observation agreement (Am/Ao=2.24 and Am/Ao=2.80 for the 39-layer model; Fig. 3a). However, at SDZ the seasonal amplitude is even more exaggerated by ZAs, especially when higher vertical resolution is applied (Am/Ao=1.70 and Am/Ao=2.03 for ST39 and ZA39; Fig. 3a). The two contrasting cases suggest that increasing horizontal resolution does not necessarily better represent the CH4 seasonal cycle, and model improvement or degradation depends on other factors such as accuracy of the temporal and spatial variations of prescribed fluxes, OH fields and meteorological forcings. In addition, as it is found for annual CH4 gradients, we note that the simulated seasonal amplitudes at stations in East Asia (AMY, TAP, GSN and SDZ) are consistently higher than the observed ones (Fig. 3a), implying that the prescribed CH4 emissions are probably overestimated in this region.

The CO2 seasonal cycle mainly represents the seasonal cycle of NEE from ORCHIDEE convoluted with atmospheric transport. Figure 3b illustrates that both ZAs and STs capture the CO2 seasonal phases at most stations well, and a high correlation (Pearson correlation R>0.8) between the simulated and observed CO2 harmonics is found for 14 out of 20 stations within the zoomed region. However, the simulated onset of CO2 uptake in spring or timing of the seasonal minima tend to be earlier than observations. This shift in phase can be as large as >1 month for several stations (e.g., HLE, JIN and PON in Fig. 3b), yet cannot be reduced by solely refining model resolutions. At BKT in western Indonesia, the shape of the CO2 seasonality is not well captured (R=0.27 and R=0.30 for ST39 and ZA39; Fig. 3b). Given that representation of the CH4 seasonal phase at BKT is very good (R=0.97 for ST39 and ZA39; Fig. 3a), the unsatisfactory model performance for CO2 suggests inaccurate seasonal variations in the prescribed surface fluxes such as NEE and/or fire emissions. As for CH4, the performance of ZAs is not degraded outside the zoomed region despite the coarser horizontal resolutions (Fig. S11).

With respect to the CO2 seasonal amplitude, 10 out of 20 stations within the zoomed region are underestimated by more than 20 %, most of which are mountain and continental stations (Fig. 3b). The underestimation of CO2 seasonal amplitudes at these stations is probably due to the underestimated carbon uptake in northern midlatitudes by ORCHIDEE, which is the case for most land surface models currently available (Peng et al., 2015). Another reason may be related to the misrepresentation of the CO2 seasonal rectifier effect (Denning et al., 1995), which means that the covariance between carbon exchange (through photosynthesis and respiration) and vertical mixing may not be well captured in our simulations even with finer model resolutions.

The day-to-day variability of CH4 and CO2 residuals are influenced by the regional distribution of fluxes and atmospheric transport at the synoptic scale. For CH4, as shown in Fig. 4a, both STs and ZAs capture the phases of synoptic variability at most stations within the zoomed region fairly well, with 15 out of 18 stations showing model–observation correlation r>0.3. Increasing horizontal resolution can more or less impact model performance, yet the direction of change is station dependent. In general, ZAs improve correlation in phases for most marine and coastal stations compared to STs (e.g., CRI and HAT; Fig. 4a), while degradation in model performance is mostly found for mountain and continental stations (e.g., KZM and SDZ; Fig. 4a). With increased horizontal resolution, better characterization of the phases would require accurate representation of short-term variability in both meteorological forcings and emission sources at fine scales. This presents great challenges on data quality of boundary conditions, especially for mountain stations located in complex terrains or continental stations surrounded by highly heterogeneous yet uncertain emission sources.

Regarding the amplitudes of CH4 synoptic variability, 12 out of 18 stations have NSDs within the range of 0.6–1.5, and ZAs generally give higher NSD values than STs for most of these stations (Fig. 4b). For stations with NSDs >1.5, ZAs tend to simulate smaller amplitudes and slightly improve model performance (e.g., GSN, HLE and SDZ; Fig. 4b). One exception is UUM. Given the presence of a wrong emission hotspot near the station in the EDGARv4.2FT2010 dataset (Fig. S6), ZAs greatly inflate the model–observation misfits (Fig. S13). The sensitivity test simulations prescribed with an improved data version EDGARv4.3.2 show much better agreement with observations, although the simulated amplitudes are still too high (Fig. S13). In addition, it is interesting to note that stations in East Asia generally have NSDs >1.5 (e.g., GSN, TAP, SDZ and UUM; Fig. 4b), again suggesting overestimation of the prescribed CH4 emissions in this region.

For CO2, as shown in Fig. 4c and d, 12 out of 20 stations within the zoomed region have model–observation correlation r>0.3, whereas 14 out of 20 stations have NSDs within the range of 0.5–1.5. With finer model resolution, significant model improvement (whether regarding phases or amplitudes of CO2 synoptic variability) is mostly found at marine, coastal and continental stations (e.g., AMY, DSI and SDZ; Fig. 4c and d); for mountain stations, on the contrary, phase correlation is not improved and representation of amplitudes is even degraded (e.g., HLE, LLN and WLG; Fig. 4c and d). As mentioned above for CH4 synoptic variability, the model degradation at mountain stations may arise from errors in mesoscale meteorology and regional distribution of sources or sinks over complex terrains, probably as well as unresolved vertical processes.

When we examine model performance for CO2 vs. CH4 by stations, there are stations at which phases of synoptic variability are satisfactorily captured for CH4 but not for CO2 (e.g., BKT, PBL, PON; Fig. 4a and c). At PON, a tropical station on the southeast coast of India, the simulated CO2 synoptic variability is even out of phase with observations all year around and during different seasons (Fig. S14; Table S3). The poor model performance should be largely attributed to the imperfect prescribed CO2 surface fluxes. As noted by several previous studies (e.g., Patra et al., 2008), CO2 fluxes with sufficient accuracy and resolution are indispensable for realistic simulation of CO2 synoptic variability. In this study, the daily to hourly NEE variability does not seem to be well represented in ORCHIDEE, especially in the tropics. Further, for stations influenced by large fire emissions (e.g., BKT), using the monthly averaged biomass burning emissions may not be able to realistically simulate CO2 synoptic variability due to episodic biomass burning events. In addition, the prescribed CO2 ocean fluxes have a rather coarse spatial resolution (4∘×5∘), which may additionally account for the poor model performance, especially for marine and coastal stations.

The diurnal cycles of trace gases are mainly controlled by the covariations between local surface fluxes and atmospheric transport. To illustrate model performance on diurnal cycles, we take a few stations with continuous measurements as examples. For CH4, as shown in Fig. 5a, the mean diurnal cycles can be reasonably well represented at the marine/coastal stations GSN and PON for the specific study periods (also see Table S4), although monthly fluxes are used to prescribe the models. Compared to STs, the diurnal cycles simulated by ZAs agree much better with observations (Fig. 5a), which is possibly due to more realistic representation of coastal topography, land–sea breeze, and/or source distribution at finer grids. However, there are also periods during which the CH4 diurnal cycles are not satisfactorily represented by both model versions or model performance is degraded with higher horizontal and/or vertical resolutions (Table S4). The model–observation mismatch may be due to the following reasons. First, the prescribed monthly surface fluxes are probably not adequate to resolve the short-term variability at stations strongly influenced by local and regional sources, especially during the seasons when emissions from wetlands and rice paddies are active and temporally variable with temperature and moisture. Second, the sub-grid scale parameterizations in the current model we used are not able to realistically simulate the diurnal cycles of boundary layer mixing. Recently new physical parameterizations have been implemented in LMDz to better simulate vertical diffusion and mesoscale mixing by thermal plumes in the boundary layer (Hourdin et al., 2002; Rio et al., 2008), which can significantly improve simulation of the daily peak values during nighttime and thus diurnal cycles of tracer concentrations (Locatelli et al., 2015a).

Representation of the CH4 diurnal cycle at mountain stations can be even more complicated, given that the mesoscale atmospheric transports such as mountain-valley circulations and terrain-induced up-down slope circulations cannot be resolved in global transport models (Griffiths et al., 2014; Pérez-Landa et al., 2007; Pillai et al., 2011). At BKT, a mountain station located on an altitude of 869 ma.s.l., the CH4 diurnal cycle is not reasonably represented when model outputs are sampled at the levels corresponding to this altitude (level 3 and level 4 for 19-layer and 39-layer models). The simulated CH4 diurnal cycles sampled at a lower model level (level 2 for both 19-layer and 39-layer models) agree much better with the observed ones (Fig. 5a). This suggests that the current model in use is not able to resolve mesoscale circulations in complex terrains, even with the zoomed grids (∼50 km over the focal area) and 39 model layers.

For CO2, as shown in Fig. 5b, the simulated diurnal cycles at GSN and PON correlate fairly well with the observed ones for their specific study periods (also see Table S5). The amplitudes of diurnal cycles are greatly underestimated, although this can be more or less improved with finer horizontal resolutions (Fig. 5b). As for CH4, the model–observation discrepancies mainly result from underestimated NEE diurnal cycles from ORCHIDEE and/or unresolved processes in the planetary boundary layer. Particularly, neither ZAs nor STs are able to adequately capture the CO2 diurnal rectifier effect (Denning et al., 1996). For stations strongly influenced by local fossil fuel emissions, underestimation of the amplitudes may be additionally attributed to fine-scale sources not resolved at current horizontal resolutions. This is the case for PON, a coastal station 8 km north of the city of Pondicherry in India with a population of around 750 000 (Lin et al., 2015), where the amplitudes of diurnal cycles are underestimated for both CO2 and CH4 (Fig. 5a and b). Again at BKT, as noted for CH4, a better model–observation agreement is found for the CO2 diurnal cycle when model outputs are sampled at the surface layer rather than the one corresponding to the station altitude (Fig. 5b). Note that even the simulated diurnal cycles at the surface level are smaller compared to the observed ones by ∼50 %, suggesting that the diurnal variations of both NEE fluxes and terrain-induced circulations are probably not satisfactorily represented in the current simulations.