HTML
--> --> -->Ground-based observation and satellite remote sensing are the two most commonly used cloud observation methods and have high spatial coverage and long time series (Lu et al., 2015). Ground-based observation includes human visual observations and ground-based automatic cloud detection (Kazantzidis et al., 2012). Since ground-based automatic cloud detection is restricted by short time series and low spatial coverage, human visual observation is still the most important source of cloud information (Kotarba, 2009; Feister et al., 2010; Huo and Lu, 2012; Lu et al., 2015). Visual observation is conducted at meteorological stations, which are also called Synop stations. This is the most traditional observation approach to obtain long-term cloud fraction data, and offers a relatively dense spatial coverage.
Satellite remote sensing is another important observation approach to obtain cloud fraction data. Compared with Synop observations, satellite data are not influenced by subjective factors. This observation method also provides the opportunity to obtain continuous and spatially uniform observations of cloud conditions (K?stner et al., 2004; Fontana et al., 2013). Nevertheless, the data quality varies with the different characteristics of satellites, such as the spectral, spatial and temporal resolution of the sensors (Fontana et al., 2013). In recent decades, satellite remote sensing has been developing rapidly and is considered to be the most important method of remote sensing in cloud detection. The MODIS instrument, onboard the Aqua and Terra satellites, is a passive imager with 36 spectral channels and a spatial resolution of 250 to 1000 m. Previous studies have shown that MODIS has higher cloud recognition capabilities, as well as better calibration and geometry, compared with other operational sensors (Platnick et al., 2003; Lu et al., 2015). Comparisons between MODIS and other satellites have indicated that the observational quality of MODIS represents an improvement over ISCCP and AVHRR (Heidinger et al., 2002; Kotarba, 2015).
Satellite-derived TCC has been compared with visual surface observations (Meerk?tter et al., 2004; Kotarba, 2009; Fontana et al., 2013; Ma et al., 2014; Lu et al., 2015) and ground-based instruments (Key et al., 2004; An and Wang, 2015) in different regions of the world. The results show good consistency between satellite and surface observations in some regions (K?stner et al., 2004; Meerk?tter et al., 2004), but also that MODIS tends to overestimate the cloud cover when compared with the surface observations in other regions (Kotarba, 2009; Fontana et al., 2013). The satellite-observed TCC is generally higher in winter and lower in summer, as determined from the observations of ISCCP, AVHRR and MODIS (Rossow et al., 1993; K?stner et al., 2004; Kotarba, 2009). (Meerk?tter et al., 2004) pointed out that, in areas with serious haze pollution in the Mediterranean, the satellite-observed cloud cover is much higher than in clean areas. Research in China has shown that the consistency between satellite and visual-surface-observed TCC is probably affected by air pollution and snow cover (Lu et al., 2015). Also, the cloud cover from satellite and surface observations has been reported to show greater deviation over the North China Plain (NCP) compared with other regions (Ma et al., 2014).
The NCP is an area with serious air pollution. Rapid economic growth over the past three decades has resulted in severe atmospheric pollution and frequent haze events (Che et al., 2014; Chen and Wang, 2015; Li, 2016). The aggravated pollution is accompanied by high aerosol loading levels (Qiu and Yang, 2000; Luo et al., 2001; Li et al., 2013; Zhang et al., 2013) and reductions in visibility (Che et al., 2007) and solar radiation (Che et al., 2005; Liang and Xia, 2005; Xia, 2010). In regions with high aerosol optical depth(AOD), the so-called shadowing effect caused by aerosols will lead to a smaller Synop-detected value of cloud fraction compared with the true value (Lu et al., 2015). Another important affect caused by high AOD is that MODIS tends to misjudge aerosol plumes as cloud in regions with heavy aerosol concentrations (Shang et al., 2014; Mao et al., 2015). However, comparisons between satellite and visual surface observations are still rare over areas with high atmospheric pollution like the NCP, particularly over the long term and in recent high-haze years.
In this paper, we present a detailed comparison of MODIS cloud cover data with Synop observations over the NCP and its surrounding regions during the period from December 2002 to November 2013 in daytime, and December 2002 to November 2009 at nighttime. We assess the discrepancies between the two datasets over high haze pollution regions and analyze these discrepancies with respect to cloud with different cloud-top heights (CTHs) and cloud optical thicknesses (COTs). The possible factors (particularly in terms of aerosol) related to the discrepancies between MODIS and Synop data are discussed for different cloud types.
2
2.1. Cloud fraction from MODIS
The satellite-observed TCC was derived from MODIS onboard the Terra and Aqua satellites, which passes over each region of the world twice a day in daytime and at nighttime. For Terra, the overpass time is around 1130 LST (Local Standard Time, UTC+8) during the daytime and 2330 LST at nighttime. For Aqua, meanwhile, the overpass time is around 1330 LST during daytime and 0130 LST at nighttime. The MODIS collection 6 MYD06/MOD06 and MYDATML2/MODATML2 cloud and aerosol products were used, downloaded from the Level 1 and Atmospheric Archive and Distribution System (http://ladsweb.nascom.nasa.gov/). The cloud detection results were recorded into 1-km (at nadir) spatial resolution MODIS cloud mask. According to the cloudiness likelihood of a given pixel, it was labeled as "cloudy", "uncertain——probably cloudy", "probably clear" or "confidently clear". The first two conditions were regarded as cloudy and the latter two as clear when calculating the cloud fraction (Platnick et al., 2003). The cloud mask product was generated into cloud fractions at 5-km resolution by calculating the proportion of cloudy pixels from every 25-pixel cloud mask group (Menzel et al., 2008).
For comparison of satellite and surface observations, the usual approach is to average the satellite-derived cloud fraction or cloud mask data within the field of view (FOV) of the surface observation. Previous studies have found that a FOV with a radius of 30 or 35 km agrees better with the observers' FOV at each Synop station (Minnis et al., 2003; Meerk?tter et al., 2004; Dybbroe et al., 2005; Fontana et al., 2013). In China, studies have found that satellite and surface observations correlate best when using a FOV with a 35-km radius (Lu et al., 2015). At each Synop station, we calculated the average MODIS cloud fraction within the surrounding 35-km radius to obtain the MODIS-observed TCC from Terra and Aqua, separately.
2
2.2. Cloud fraction from surface data
The surface TCC data are visual estimations of cloud cover and cloud type produced by observers at meteorological observation stations, which are sited in open areas away from buildings and trees in order to ensure the FOV is unaffected. The data were provided by the China Meteorological Sharing Service System (CMDSSS, 2016). In total, 121 Synop stations were chosen in the research area. Synop observations were performed at eight times a day at 3-h intervals——at 0200, 0500, 0800, 1100, 1400, 1700, 2000 and 2300 LST. To minimize the effect of the time differences between Synop and MODIS observations, possible approaches include choosing the Synop TCC nearest to the MODIS overpass time (Lu et al., 2015), calculating the average of two time points adjacent to the MODIS overpass time (Fontana et al., 2013), and interpolating the Synop TCC to the MODIS overpass time (Kotarba, 2009). In this study, the Synop TCC at three times nearest the overpass time (0800, 1100 and 1400 LST during daytime and 2000, 2300 and 0200 LST at nighttime for Terra; 1100, 1400 and 1700 LST during daytime and 2300, 0200 and 0500 at nighttime for Aqua) were interpolated to the satellites' overpass times with linear interpolation in order to reduce the errors caused by observational time deviation.In terms of the dark conditions at nighttime seriously influencing the accuracy of visual surface observations (Minnis et al., 2003), the main existing method is to choose observations made at illuminations greater than that from a half-moon at zenith. The illumination of the moonlight from the lunar altitude and phase can be determined by the ephemeris and date (Hahn et al., 1992). The Extended Edited Cloud Report Archive (EECRA) is a dataset compiled based on global surface observation datasets. EECRA offers the relative lunar illuminance and flags denoting sufficient illumination from moonlight, twilight, or sunlight during the period 1971 to 2009 for land-based stations. In this study, 81 stations in or near the research area were chosen. For each Synop station to be compared, the nearest EECRA station was identified and their illuminations considered to be approximately equal.
Synop observations of cloud types divide the cloud at three levels into 10 types, separately. For the sake of analysis of cloud with different forms, we redivided clouds into 10 categories following the classification method defined by the International Meteorological Organization. The 10 cloud types were: cumulus cloud (Cu), cumulonimbus cloud (Cb), stratocumulus cloud (Sc) stratocumulus cloud (St), nimbostratus cloud (Ns), altostratus cloud (As), altocumulus cloud (Ac), cirrus cloud (Ci), cirrostratus cloud (Cs), and cirrocumulus cloud (Cc).
2
2.3. Auxiliary data sets
For the analysis of the factors influencing observations, five auxiliary datasets of CTH, COT, AOD at 550nm, satellite view zenith angle (VZA), and snow cover, were used. All were derived from Terra and Aqua MODIS Collection 6 data products. The AOD data were derived from the Deep Blue (DB) and Dark Target (DT) combined algorithm, and only the highest quality flag (QF=3) AOD data were used. The DT algorithm was developed to detect AOD over dark surfaces such as vegetation and ocean (Remer et al., 2005; Levy et al., 2007a, 2007b). In contrast, the DB algorithm can retrieve AOD over bright surfaces such as desert and snow (Hsu et al., 2004; Bilal and Nichol, 2015). The DT/DB algorithm is a "best of" AOD product with a wide coverage and high precision (Green et al., 2009; Levy et al., 2013; Bilal and Nichol, 2015). The snow cover data were derived from the MODIS snow and sea ice products MOD10/MYD10, which provide the snow cover and ice cap at a 0.05° resolution (Hall et al., 2006). All these auxiliary data were averaged within the same FOV, like the TCC.


3.1. Climatology of TCC from Aqua MODIS and Synop observations
In order to realize the overall distribution of MODIS- and Synop-observed TCC, we first calculated the climatic field as well as the temporal variation of the TCC. As shown in Fig. 1, the cloud fraction showed distinct seasonal changes. The TCC observed by MODIS was generally greater than that from the Synop observations. The latter showed the lowest TCC in winter and highest in summer, yet the MODIS value was high both in summer and winter, and relatively low in spring and winter. The TCC observed by the two methods showed best consistency in summer and greatest deviation in winter. Analysis of the TCC climatic field is shown in Fig. 2. In general, the TCC of the southern part was higher than the northern part, which was roughly the same for MODIS and Synop observations. Meanwhile, it is notable that in winter the MODIS-observed TCC in the northern part was much larger than the Synop observation during daytime, while at nighttime both the MODIS- and Synop-derived TCC showed low values. In the southern part, the MODIS-observed TCC was high both in daytime and at nighttime, while the Synop observation was relatively low.2
3.2. Comparison between TCC from Terra and Aqua MODIS
We conducted a detailed 11-year (December 2002-November 2013) comparison between the MODIS-derived TCC from the Terra and Aqua satellites. Figure 3 compares the monthly averaged TCC observed by Aqua and Terra for all stations. The correlation coefficient (R) between the TCC derived from Terra MODIS and Aqua MODIS was 0.77 for daytime and 0.72 for nighttime, suggesting that Terra MODIS and Aqua MODIS were highly coherent. The Aqua MODIS observation results were slightly larger than those of Terra MODIS for both daytime and nighttime. This may be affected by the satellites' different overpass times.2
3.3. Comparison between daily MODIS and Synop observations
A comparison between the MODIS and Synop TCC was conducted daily during the period December 2002-November 2013, and the statistical results are shown in Fig. 4 and Table 1. The positive differences between the MODIS and Synop observations were significantly more than the negative ones, and 55% of all differences during daytime and 50% of all differences at nighttime ranged from 0% to 20% (Fig. 4), indicating that the MODIS-observed data were generally greater than the Synop data. The mean difference (D ms) between the MODIS- and Synop-observed data was 13.95% for Terra and 15.25% for Aqua (Table 1).Table 1 explicitly shows that the deviation at nighttime was greater than that during daytime. The D ms at nighttime was 2% to 3% higher than that during daytime, and the RMSE at nighttime was 3% to 4% higher than during daytime. The R during daytime was 0.69 and 0.67 for Terra and Aqua, respectively, which was higher than the R at nighttime (0.65 and 0.64). This may be affected by the lack of a visible channel, which would reduce the accuracy of the MODIS observation (Kotarba, 2009).
2
3.4. Seasonal variations between MODIS and Synop TCC
The difference between MODIS and Synop TCC also varied with season (Table 2). As shown in Table 2, the deviation between the two datasets was greatest in winter. In winter, the D ms and RMSE were the largest among the four seasons; the D ms reached 29.53% and 31.07% and the RMSE 46.42% and 47.59% for Terra and Aqua, respectively. Meanwhile, the R in winter was smallest among the four seasons, being only 0.56 and 0.55 for Terra and Aqua, respectively. The R was similar to a comparison of MODIS and Synop TCC in Poland; however, the D ms was much higher than that in Poland, which was 7.28% in January 2004 (Kotarba, 2009).In contrast, the difference between the MODIS and Synop TCC was smallest and most consistent in summer. Both the D ms (2.31%-7.99%) and RMSE (24.86%-33.72%) were much smaller than in the other three seasons, and the R was relatively high (0.65-0.78). The mean D ms during daytime and at nighttime in our study regions was 4.46% for Terra and 6.07% for Aqua, which is comparable to the 4.38% in Poland in July 2004 (Kotarba, 2009). Previous research in China found similar results. (Ma et al., 2014) found that the D ms calculated by full-year data was 15.09% in North China, while the D ms decreased to 5.29% after removal of the winter data. (Lu et al., 2015) found that in the China area the correlation between the two observation results was highest in summer (0.736) and lowest in winter (0.667).
The deviation between MODIS and Synop TCC in spring and autumn was between that of summer and winter, and did not show any great difference. The high R, ranging from 0.69 to 0.75, suggested good consistency between MODIS and Synop TCC. Table 2 shows that in all seasons the D ms and RMSE during daytime were much smaller than at nighttime and the R during daytime was much higher than that at nighttime, indicating that the TCC observed during daytime was much better than that at nighttime.
2
3.5. Relationship between the cloud fraction deviation and CTH/COT
Considering different cloud types may influence both MODIS and Synop observations and further influence the D ms, two physical characteristics——CTH and COT, which are important parameters to distinguish different cloud types——were chosen to discuss their relationship with D ms. Because of the lack of COT observations at night, both discussions focus on the data during daytime only.

Figure 5a shows the average CTH under different D ms levels. It can be clearly seen from the figure that when the D ms was less than zero the average CTH was at a relatively high value. In the area that the D ms was near zero the average CTH was at a peak. Meanwhile, when the D ms was greater than 0.2 the CTH showed a sharp decrease with an increase in the D ms. When the D ms was close to 1 the average CTH was near 1 km.
Figure 5b facilitates further discussion on the D ms distribution for clouds with different CTH. The figure shows that under conditions with lower CTH the distribution frequency of bigger D ms was much higher, and D ms values greater than 0.2 mainly appeared when CTH was less than 2 km. With an increase in CTH the proportion of bigger D ms values reduced rapidly. This result shows that MODIS more easily detects cloud with low CTH, which Synop observations were otherwise unable to detect.
The other characteristic, COT, is discussed based on the results in Fig. 6. Figure 6a shows that the average COT with D ms near zero was obviously higher. In contrast, the average COT with deviation larger than 0.2 was generally low in value. The distribution of D ms under different COT is presented in Fig. 6b. As can be seen in the figure, large deviation mainly occurred under conditions with low COT. When COT was greater than 20, nearly all deviations were smaller than 0.2.Meanwhile, when COT was smaller than 12, the frequency of deviations greater than 0.2 was nearly half.
It can be inferred from the analysis above that MODIS tends to detect cloud with low CTH and small COT that is otherwise undetected by Synop observations, meaning there may be cases that Cu and Sc clouds are detected by MODIS but undetected or underestimated by Synop observations. Given that previous research has proven that MODIS tends to judge the layer of aerosols at low altitude as cloud (Shang et al., 2014; Mao et al., 2015), it is possible that MODIS in the present study judged the aerosol layer as cloud, leading to the high D ms. Another possibility is that the surface FOV was larger for high cloud; surface observations can see high cloud in a larger radius than low cloud, which increases the surface-observed high cloud fraction.
2
3.6. Possible reasons for the difference between MODIS and Synop
To explore the possible factors influencing the consistency and deviation between MODIS and Synop TCC observations, we analyzed the relationship between the deviation with AOD, VZA, and snow cover.


The averaged D ms in different VZA intervals (Fig. 7a) showed that the deviation of TCC increased with an increase in VZA. Under conditions of VZA <30°, the D ms basically maintained at a low level (<7%); whereas, at VZA >30°, the D ms increased systematically with VZA. The averaged D ms at the largest VZA interval (23.45%) was 19.68% greater than the D ms at the smallest VZA interval (3.77%). The regression result also showed that larger VZA would lead to larger MODIS observations. This is consistent with previous studies (Maddux et al., 2010; An and Wang, 2015). The larger VZA would decrease the clear space between clouds, especially for thick clouds like convective clouds, on account of the vertical sides of the clouds would be viewed by the satellite. This effect was especially obvious for convective clouds or broken clouds. Another possible reason was that pixels with larger VZA have larger size and longer observation path lengths, which may increase the satellite-observed TCC.
Analysis of the relationships between cloud types and VZA (Figs. 7b-d) showed that high VZA would lead to the D ms of most categories of clouds being higher when the VZA was higher. Besides, observations of broken clouds were more likely to be affected by the VZA. Cu, Cb and Ac showed significant increasing trend as the VZA became larger. In contrast, cloud covering the whole sky had a relative stable observation result. The trends of Sc, St, Ns and As were not as obvious as the other types of cloud. It is worth noting that the D ms of most cloud types was positive, while that of Ns and As was near zero, possibly because both MODIS and Synop observations were near to 1 under these conditions; plus, Cs was negative in every VZA, which was possibly because MODIS had a relative weak detection ability for thin ice cloud, as proven by (Holz et al., 2008).
Figure 8 shows the spatial distribution of Aqua MODIS AOD and the averaged D ms between MODIS and Synop at each Synop station in the four seasons. Because of the lack of AOD observations at nighttime, only observations during daytime were analyzed. The D ms was averaged from the D ms of Terra and Aqua MODIS. In all seasons, the distribution of the averaged D ms was consistent with the distribution of AOD. Stations with low D ms values mainly distributed in the northwestern area and Shandong's coastal area, which were the low AOD value areas. In contrast, the D ms in central and western Shandong, central and eastern Henan, as well as southern Hebei, were generally higher than in other areas.
Note that in the Liaoning area the D ms was slightly larger than in areas with the same AOD value (Fig. 8). This phenomenon became quite obvious in winter (Fig. 8d). In winter, the D ms in Liaoning was even larger than that in the border regions of Shandong, Hebei and Henan, where AODs were largest. This might be influenced by snow cover and the low solar height angle due to Liaoning being located at high latitudes.
To further investigate the impact of AOD on the difference between satellite and Synop TCC observations for each cloud type, the D ms values at different AOD intervals were calculated (Fig. 9). Figure 9a shows that the D ms was greater at high AODs. In all seasons, the D ms tended to increase generally with an increase in AOD. In summer, the D ms increased monotonically with increasing AOD over the entire AOD range. In spring, autumn and winter, the D ms increased with AOD values at AOD <1.5, whereas the D ms showed no remarkable change and even dropped slightly with increasing AOD at AOD >1.5. That may have been caused by a small amount of high AODs (Fig. 9) or the environment was not so different to satellite and visual surface observations at AOD >1.5.
Analysis of different cloud types (Figs. 9b-d) showed that the D ms of most cloud types increased with AOD. The most obvious were Cu, Ac, Ci and Cs, which did not cover the whole sky. Misjudging the aerosol layer as cloud by MODIS may be the reason behind this phenomenon.
To further investigate the influence of snow cover, the relationship between snow cover and the D ms in winter is shown in Fig. 10. The distribution of D ms values showed consistency with the distribution of snow cover. In winter, the main areas with high D ms values appeared in the provinces of Liaoning and Shandong. The above analysis shows that high AODs in Shandong induced high D ms values; however, Liaoning had much lower AOD values. The large snow coverage in Liaoning may have resulted in the higher D ms.

Analysis of the effect of cloud characteristics on the observational deviation found that CTH and COT had an obvious influence on D ms. Cloud with low CTH was more likely to cause a higher MODIS observational result and lower Synop observational result, while this frequency reduced significantly when the CTH was lower than 4 km. Another point is that observations with significant deviations mainly occurred when COT was less than 12.
Analysis showed that a large VZA would lead to a larger MODIS-observed TCC, and this effect was more obvious for clouds occurring in clumps than cloud covering the whole sky. Besides, thin clouds like Cs would lead to a negative D ms, and a high VZA value would improve the MODIS detection. Similar results were seen for the effect of AOD. The spatial distribution of the difference between MODIS and Synop matched well with the AOD distribution, and the difference increased with an increase in AOD. The difference in the NCP and its surrounding regions was higher than that in Poland, Europe (Kotarba, 2009), suggesting that high pollution may induce a greater MODIS TCC. In addition, high snow coverage may affect MODIS observations, thus resulting in a high difference in northern areas.