Modeling aerosols' radiative effects during this biomass-burning event first requires accurate quantification and meticulous characterization of emission from field burning of crop residue. Here, emission intensities of trace gases and particulate matters, specifically including carbon dioxide (CO2), CO, methane (CH4), non-methane organic compounds (NMOCs), nitrogen oxides (NOx), ammonia (NH3), sulfur dioxide (SO2), BC, organic carbon (OC) and particulate matter (PM2.5 and PM10 are particles with aerodynamic diameter less than 2.5 and 10 microns respectively), were estimated based on a bottom-up method. According to the farming season (available at http://zzys.agri.gov.cn) and province-level statistics on crop cultivation (NBSC, 2013), we can deduce that the intensive agricultural fires in June were mainly related to wheat straw burning as a consequence of the extensively spreading cultivation mode of “winter wheat–summer corn/rice” in East China. Burned biomass at province-level was calculated based on statistical data of crop productions, residue-to-production ratios and percentages of crop residues burned in the field. Emissions of various pollutants were derived from the product of burned mass and experiment results on crop-specific combustion efficiencies and pollutant-specific emission factors. The detailed methods and involved datasets are described in our previous work (Huang et al., 2012b).

To determine the locations and time of crop residue fires, MODIS (Moderate Resolution Imaging Spectroradiometer, available at https://ladsweb.nascom.nasa.gov/data/) Thermal Anomalies/Fire Daily L3 Global product (MOD/MYD14A1) combined with burned area product (MCD45A1) were introduced for the purpose of emission spatiotemporal allocations (Giglio et al., 2003; Boschetti et al., 2009). MOD/MYD14A1 provides fire identification by examining the brightness temperature relative to neighboring pixels. MCD45A1 was also incorporated in this work because its bidirectional reflectance model-based change detection approach has been proved to be capable of presenting a more accurate mapping of smaller fragments of burn scars (Roy and Boschetti, 2009). The Global Land Cover (GLC) product with a spatial resolution of 1 km was used in this study to identify the burning of different biomass. Only fire detections that occurred on farmland, i.e., land cover classes defined as “farm” and “mosaic of cropping”, were identified as field burning of crop residue. Emission at province level estimated using the aforementioned method was then allocated equally to each fire spot.

The fire emission estimation developed in this work was compared with the FINN fire emission dataset. Spatially, these two emission inventories generally were consistent with each other because the locations for the fires in both inventories are based on MODIS Thermal Anomalies product (Fig. S1 in the Supplement). Some inconsistencies, such as the density of fire in central Jiangsu, are attributed to the different land cover dataset applied for the identification of underlying biomass type. FINN fire emission estimation used MODIS Collection 5 Land Cover Type data (Wiedinmyer et al., 2011), while we employed Global Land Cover data. This inventory differs slightly from FINN estimation in magnitude. Taking CO emission in the inner model domain for instance, we estimate that 4.5 Tg CO was emitted while FINN gives the value of 7.5 Tg during the first half of June 2012. It might be attributed to different methods to estimate burned biomass. FINN used MODIS Vegetation Continuous Fields to assign the burned mass. The fuel loading of farmland was assumed to be 0.5 kg m-2 (Wiedinmyer et al., 2006). However, in China, crop straw is used in multiple ways that differ regionally, like biofuel, biogas production and animal feed supply, which is highly dependent on crop species. We estimated the emission using a “bottom-up” method by fully considering crop yields, crop-specific straw usage and combustion efficiency.

During this agricultural fire event, the spatial pattern of fire detections in Fig. 2a indicates that open burning of straw mostly concentrated in northern parts of Anhui and Jiangsu province and got extremely severe on 9 and 13 June, as displayed in Fig. 2b. Burning of crop residues dominated local emissions of atmospheric pollutants when compared with corresponding anthropogenic emissions. Taking BC for instance (Fig. 2c and d), emission rate from field burning of crop residues far outweighed that from industry, power plant, residential activity and transportation combined (Li et al., 2015).

The numerical simulations in this study were conducted using WRF-Chem version 3.6.1, which is an online-coupled chemical transport model considering multiple physical and chemical processes, including emission and deposition of pollutants, advection and diffusion, gaseous and aqueous chemical transformation, aerosol chemistry and dynamics (Grell et al., 2011). The model has been widely utilized to investigate aerosol–radiation–cloud interactions and aerosol–boundary-layer feedback (Grell et al., 2011; Zhao et al., 2013; Fan et al., 2015; Huang et al., 2015; Ding et al., 2016; Gao et al., 2016). In the present work, we adopted two nested model domains (Fig. 1a). The coarse domain is centered at 115.0∘ E, 33.0∘ N with a grid resolution of 20 km covered East China and its surrounding areas. The fine resolution of 4 km for the inner one allowed better characterization of small-scale physical processes, especially those linked to convective motions, cloud formation and rainfall onset. There were 31 vertical layers from the ground level to the top pressure of 50 hPa, 20 of which were placed below 4 km to achieve finer vertical resolution within the boundary layer. The initial and boundary conditions of meteorological fields were updated from the 6 h NCEP (National Centers for Environmental Prediction) global final analysis (FNL) data with a 1∘ × 1∘ spatial resolution. To investigate the aerosols' radiative effects on 10 June 2012 when the precipitation was substantially modified, the simulations were conducted for the time period from 20 May to 15 June. The meteorological initializing date for 10 June was 12:00 UTC on 9 June. Each run covered 60 h and the last 48 h modeling results were kept for further analysis. The chemical outputs from the preceding run were used as the initial conditions for the following run. The first 20 days were regarded as the model spin-up period for atmospheric chemistry, so as to better characterize aerosol distributions and minimize the influences of initial conditions and allow the model to reach a state of statistical equilibrium under the applied forcing (Berge et al., 2001; Lo et al., 2008).

Key parameterization options for the WRF-Chem modeling were the Noah land surface scheme to describe the land–atmosphere interactions (Ek et al., 2003), the YSU boundary layer scheme (Hong, 2010) and the RRTMG short- and longwave radiation scheme (Mlawer et al., 1997). The Lin microphysics scheme that accounts for six forms of hydrometer (Lin et al., 1983) together with the Grell cumulus parameterization was applied to reproduce the cloud and precipitation processes (Grell and Devenyi, 2002) for the coarse domain. Cumulus parameterization was switched off for the inner domain. For the numerical representation of atmospheric chemistry, we used the CBMZ (Carbon Bond Mechanism version Z) photochemical mechanism combined with MOSAIC (Model for Simulating Aerosol Interactions and Chemistry) aerosol model (Zaveri and Peters, 1999; Zaveri et al., 2008). Aerosols were assumed to be spherical particles. The size distribution was divided into four discrete size bins defined by their lower and upper dry particle diameters (0.039–0.156, 0.156–0.625, 0.625–2.5 and 2.5–10.0 µm). Aerosols in each size bin were assumed to be internally mixed and their optical properties, including extinction coefficient, single-scattering albedo (SSA) and asymmetry factor, were computed based on Mie theory (Fast et al., 2006) using volume averaged refractive indices (Barnard et al., 2010). Similar model configurations and settings have achieved good performance in our previous simulations over East China (Huang et al., 2015; Ding et al., 2016). Detailed configurations and domain settings are listed in Table 1.

Both natural and anthropogenic emissions were included for the regional WRF-Chem modeling in the present work. Typical anthropogenic emissions were obtained from the Multi-resolution Emission Inventory for China (MEIC) database (Li et al., 2015), in which emissions sources were classified into five main sectors: power plants, residential combustion, industrial processes, on-road mobile sources and agricultural activities. This database covers most of anthropogenic pollutants, such as SO2, NOx, CO, volatile organic compounds (VOCs), PM, BC and OC. NH3 emission over China was derived from Huang et al. (2012a). VOCs emitted from typical anthropogenic activities and aforementioned crop residue burning were speciated into model-ready lumped species using profiles for Carbon Bond Mechanism (Hsu et al., 2006). The biogenic VOC and NO emissions were calculated online by using the Model of Emissions of Gases and Aerosols from Nature (MEGAN) that embedded in WRF-Chem (Guenther et al., 2006). More than 20 biogenic species, including isoprene, monoterpenes (e.g., α-pinene and β-pinene) and sesquiterpenes, were considered and then involved in the photochemistry calculation. In China, crop residues are usually burned in piles, which are characterized by short-lived and small-scale smoldering. Consequently, the plume rise of biomass-burning plumes was not considered in this study, and the straw fire emission was placed in the lowest two levels from the surface to around 50 m in this simulation.

Previous studies have shown that, under highly polluted conditions, the aerosol–radiation interaction (ARI) dominated over the aerosol–cloud interaction (ACI) that is related to aerosols' ability to act as cloud condensation nuclei (CCN) (e.g., Rosenfeld et al., 2008; Fan et al., 2015). We also conducted another numerical experiment which included both ACI and ARI. The ACI-induced radiative perturbations were much less notable than those caused by ARI both at the surface and in the atmosphere (Fig. S2), implying the dominant role of ARI during this kind of synoptic-scale fire event. Since that the focus of this study is on ARI and ACI's effect was not that significant, the prognosed aerosol was disabled to act as CCN or ice nuclei in the simulations and therefore the effects from ACI were not accounted for in the following analysis. Accordingly, wet scavenging of aerosol was disabled too. In order to disentangle aerosols' role in radiative transfer and subsequent effects on cloud and precipitation during this biomass-burning event in mid-June of 2012, we designed three parallel numerical experiments. Domain settings and model configurations for these simulations were exactly the same as mentioned before. The control (CTL) experiment did not include aerosol's effects on either longwave or shortwave radiation transfer. On the contrary, the other two took account of aerosols' perturbations on radiation transfer: ARI-A with anthropogenic emissions (anthropogenic activities refer to power generation, transportation, industrial and residential activities hereafter) and ARI-AB that included both anthropogenic activities and biomass-burning emissions.

As demonstrated by existing studies (Andreae et al., 1988; Huang et al., 2012c; Ding et al., 2013a), air quality dramatically deteriorated and the visibility was impeded during biomass-burning events. We compare the simulated daily averaged PM10 concentration with the corresponding measurements derived from Air Pollution Index (API, available at http://datacenter.mep.gov.cn/) in Fig. 3 (if not mentioned specially, the simulation refers to ARI-AB experiment hereafter). Both observations and simulations manifested the fact that intensive agricultural fires led to the severe pollution in mid-June. Since 9 June, when the detected fire spots became intense and extensive, PM10 concentrations in northern Anhui and northwestern Jiangsu province began to increase, especially for those regions near the fire location. For instance, the observed daily mean PM10 concentrations reached up to around 250 µg m-3 at Fuyang (FY) and Xuzhou (XZ) and even exceeded 400 µg m-3 at Bengbu (BB) on 9 June (the locations of cities mentioned in this article are labeled in Fig. 2). Although the simulated temporal variations agree with observations, model-predicted PM10 concentrations at FY and BB were 196 and 168 µg m-3 respectively. The underestimation might be due to the fact that rapid formation of secondary aerosol like sulfate and secondary organic matters is not so well described in current atmospheric chemical transport models (Capes et al., 2008; Xie et al., 2015). XZ and BB suffered from the second-round fire smoke 2 days later, with a maximum daily mean concentration of 548 µg m-3 observed at BB. Figure 4 illustrates the satellite-retrieved 660 nm aerosol optical depth (AOD) and SSA from MODIS Aerosol Product MOD04_L2 (daily level 2 data produced at the spatial resolution of 10 km, Collection 6) around 11:00 local time (LT) on 9 June when the first round of extensive fire pollution broke out. Their comparisons with ARI-AB modeled spatial distributions of PM2.5 and BC column-integrated mass loadings further confirm model's ability to reproduce atmospheric pollution for this event. The AOD observation shows that high aerosol loadings were concentrating in northeast Anhui and the north-central Jiangsu, shaping a belt of pollution from the fire sites to the downwind areas. A similar pattern was also simulated by the model. The PM2.5 mass loading was found to exceed 200 mg m-2 near BB, NJ and most parts of central Jiangsu. This strap-shaped pollution was particularly obvious in terms of BC column concentrations, which was also consistent with a relatively lower SSA along BB, Yangzhou (YZ) and Taizhou (TZ). While solely including anthropogenic emissions, ARI-A experiment failed to represent the spatial pattern of high AOD in the northern Anhui and Jiangsu and the low SSA value near BB (Fig. 4a, d).

Along with the severe air pollution and poor visibility, anomalies in meteorology occurred on 9–10 June. Ding et al. (2013a) found that, during these 2 days, a sharp decline existed in the observed air temperature in NJ and YZ compared with weather forecast results and NCEP FNL data, but the simulations and observations showed a good agreement when the heavy air pollution was not present before 9 June and after 10 June. At YZ the temperature difference was as high as 5.9 and 9.2 ∘C on 9 and 10 June respectively. Simultaneously, measured solar radiation intensity and sensible heat flux showed very low values on 10 June in comparison with non-episode days. Moreover, local meteorological agency forecasted a convective rainfall to occur in NJ and surrounding areas in the afternoon of 10 June, with the rainfall center passing by NJ around 14:00 LT. However, this forecasted rainfall never happened that day.

On the basis of ground-based measurements, vertical sounding data, remote-sensing images and their comparisons with numerical simulations, we found that agricultural fires worsen regional air quality to a large extent and caused a series of anomalies in temperature and precipitation in the mid-June of 2012. How the biomass-burning plumes influenced the air temperature and precipitation will be the main issue to be addressed in the following discussions.

To better understand aerosols' role in the energy re-allocation on 10 June when precipitation was evidently modified, radiative forcing in the atmosphere and at the ground surface was estimated by differentiating the CTL, ARI-A and ARI-AB simulations (Fig. 5). At the surface, daily mean incident shortwave radiation was weakened by 45.5 W m-2 (averaged over the inner domain) as the extinction of aerosol was quite large with a satellite-observed 660 nm AOD exceeding 2.0 (Fig. 4b). Meanwhile, about 60.4 W m-2 shortwave energy was blocked in the atmosphere over the inner domain due to the fact that absorbing aerosols were accumulated on that day. A positive domain-averaged radiative forcing of +14.9 W m-2 was simulated at the top of the atmosphere on 10 June. Comparatively, radiative perturbations caused by agricultural fires (ARI-AB minus ARI-A) were more substantial than those due to anthropogenic emissions (ARI-A minus CTL) in magnitude, particularly in the atmosphere, as presented in Fig. 5. Spatially, radiative effects due to anthropogenic activities concentrated in the economically developed Yangtze River Delta region while agricultural fires exerted significant impact on radiation balance in northern and central Jiangsu and the northern part of Anhui. Table 2 compares the radiative perturbations caused by anthropogenic activities and biomass-burning emissions over three regions with distinct precipitation changes (marked in Fig. 8). As shown, both of them tended to heat the atmosphere and cool the ground surface. Fire plumes dominated the radiative effect in terms of atmospheric warming. Radiation measurements collected at Hefei (HF) and sensible and latent heat flux recorded at Lishui (in South Nanjing) are compared with the diurnal variations of corresponding simulations in Fig. 6, which supports that significant radiative perturbations took place at NJ and HF. Substantially weakened daytime solar irradiance was observed on 10 June, when the peak value of downwelling shortwave radiation was 618.3 W m-2 at HF and was only 309.7 W m-2 at NJ. Taking aerosol's effect on radiation into account tended to predict lower downwards solar radiation, which was closer to observation for both cities. Reduction in shortwave energy hitting the surface in turn decreased outgoing heat fluxes and therefore simulated sensible and latent heat fluxes at 12:00 LT on 10 June in ARI-AB experiment decreased by 89.3 and 76.1 W m-2, respectively, compared to the CTL experiment.

Overall, the magnitude of the radiative forcing on 10 June was comparable in northern Anhui and central Jiangsu, differing from the distribution pattern of fire-induced air pollution that remarkably concentrated in northern Anhui. As revealed in our previous estimation, among all components of the ambient aerosols, BC is the most important disturber of shortwave radiation transfer at the surface and in the atmosphere as well (Huang et al., 2015; Ding et al., 2016). Although fire emission mostly concentrated in the northern Anhui and resulted in a high BC concentration of 20 µg m-3 there, high-altitude BC was spread much more broadly. At an altitude of 2 km, BC concentration around 5 µg m-3 stretched from northern Anhui to central Jiangsu (Fig. S3). Such distinct distributions between two layers were partly attributed to the stagnant condition near the surface and stronger horizontal transport in the upper level. It is emphasized that upper-level BC has higher absorbing efficiency (Ding et al., 2016). That is why the distributions of both positive radiative forcing in the atmosphere and negative forcing at the surface were generally consistent with BC's spatial pattern in the upper air.

The perturbations in the energy budget and the following re-allocation gave rise to substantial modulation in vertical stratification of air temperature. In comparison with the CTL experiment, ARI-AB experiment predicted an obvious decline in near-surface temperature by considering the effects of aerosol–radiation interaction. Hourly observed 2 m air temperature was compared with corresponding simulations by two experiments during the time period from 8 to 15 June. Model-performance statistics including mean bias, mean error and root mean square error are presented in Table 3. As shown, CTL simulation had a systematic positive bias in 2 m temperature and ARI-AB predicted lower temperature for both areas near fire locations (BB and XZ) and downwind regions (NJ and SY). The decreases in temperature were pronounced in BB and XZ with a large difference of approximately 1.2 ∘C, which notably narrowed the gaps with observations. On 10 June when the fire-induced pollution became intensive, the magnitude of surface cooling was remarkably high near the fire sites. Temperature response in Fig. 5e support this cooling effect. For instance, compared to CTL, simulated near-surface temperature by ARI-A and ARI-AB experiment at XZ was cooled by almost 1.2 and 8.0 ∘C at 20:00 LT on 10 June (Fig. 7b). In addition to the cooling tendency of near-surface temperature, aerosols' radiative effects also increased air temperature at a higher altitude, which were more apparent over the downwind areas (Fig. 5f). According to the comparisons between simulated temperature profiles by the three parallel experiments in Fig. 7, the warming of air temperature was particularly evident around an altitude of 2 km at SY with a maximum of 3.0 ∘C, and the biomass-burning aerosols played a leading role.

The different temperature responses over the source region of fire emission and downwind areas could be partially interpreted by the fact that, near the fire locations, pronounced surface cooling counteracted part of the atmospheric warming, which would otherwise elevate upper-air temperature, through vertical mixing; for the downwind area where the surface was less radiatively cooled, the atmosphere was prone to being warmed. As a result of surface cooling and atmospheric heating, vertical convective motions were weakened, triggering perturbations in pressure and wind fields (Fig. 5e and f). It is obvious that suppressed convection was generally along with the resultant wind convergence around 2 km and surface divergence, which may further play a significant role in water vapor transport, entrainment and also cloud formation.

In addition to the attenuation of solar radiation and the modulation in temperature gradients, precipitation also showed many disparities between CTL and ARI-AB simulations. The satellite observation from Tropical Rainfall Measuring Mission (TRMM) Multisatellite Precipitation Analysis product (3B42), which provides merged-infrared precipitation information at a 0.25 × 0.25∘ spatial resolution and has been demonstrated to perform well in East China (Simpson et al., 1988; Zhao and Yatagai, 2014), was used to evaluate the simulated precipitation. As demonstrated in Fig. 8, ARI-AB experiment agrees better with TRMM observations than CTL experiment in terms of precipitation intensities and also spatial pattern on 10 June. Specifically, CTL and ARI-A simulation suggested a convective rain in Zone 1 (NJ and its adjacent areas) around 14:00 LT (the locations of zones 1–3 are marked in Fig. 8); however the ARI-AB simulation did not show any precipitation then, consistent with the TRMM observations. Besides, ARI-AB displayed enhanced precipitation in northern Jiangsu province. A precipitation with the intensity of 3 and 5 mm h-1 was predicted by ARI-AB experiment in Zone 2 (XZ and its adjacent areas) and Zone 3 (SY and its adjacent areas), which, however, never occurred in the CTL and ARI-A experiments. Concerning temporal variations, 3 h precipitation rates for these three zones derived from TRMM 3B42 retrievals are plotted in Fig. 9. Compared to the CTL and ARI-A experiments, the ARI-AB experiment, which considered radiative effects of aerosol from both anthropogenic and biomass-burning emissions, succeeded in capturing the approximate onset time for all the three regions.

Though the modeling work here characterized cloud and precipitation anomalies during the biomass-burning event, we may also question to what extent the modeling reproduced the relevant processes in the real world. As widely acknowledged, accurate simulation of smoke plume and prediction of clouds are both challenging for regional/global models. One contributor to the uncertainties is the quantification of fire emission. The magnitude was determined by statistical information and laboratory experiment data, the accuracy and representativeness of which may introduce some uncertainties. The spatiotemporal distribution of fire emission was allocated based on MODIS retrievals. Loss of information due to cloud coverage and poor detection efficiency of short-lived or small-scale fires are major limitations (Giglio et al., 2003). Another challenge is quantification of heat release from biomass burning and subsequent effects on local and regional meteorology. Furthermore, much attention has been paid to the vertical distribution of absorbing aerosol, to which the cloud response is highly sensitive (Koch and Del Genio, 2010). The vertical profile of absorbing aerosol in this simulation underwent little constrain due to limited observation at that time. The regional model is hardly capable of precisely presenting turbulent flows and vertical transport, thus introducing uncertainties in three dimensional distributions of BC. It also should be noted that BC is co-emitted with other components such as OC and sulfur dioxide that oxidize to sulfate (Xie et al., 2015). Mixing with other scattering aerosol would considerably amplify the absorbing efficiency of BC. A model's ability to account for the evolution of mixing state and how to quantify its amplification also affect the simulated radiative behaviors. Besides, poorly recognized secondary organic carbon formation processes and its light absorption make it imperative to reassess and redefine the chemical mechanism and optical properties of OC in models (Saleh et al., 2014). The large uncertainty in simulating clouds and further aerosol–cloud interaction is another limitation (e.g., Wang et al., 2011; Tao et al., 2012). To improve the model performance in all these chemical and physical processes, more comprehensive measurements and modeling efforts are needed in the future.