1 Introduction

Clouds are visible clusters of water droplets, ice crystals, or their combination in the atmosphere, and have a strong effect on the incoming solar and outgoing thermal infrared (TIR) radiation. Knowledge of cloud optical, macrophysical, and microphysical properties is essential for accurate determination of the atmospheric energy budget. The cloud top height (CTH) is a key attribute of clouds, which plays an important role in the atmospheric circulation, cloud structure analysis, and radiative forcing. Satellite remote sensing is well-suited for obtaining the high temporal and spatial resolution cloud information such as the CTH on both regional and global scales through the use of various instruments and algorithms.

FY-4A, situated over the equator around 104.7°E, is the first of China’s next-generation geostationary meteorological satellite series (Yang et al., 2017). It was launched on 11 December 2016, and has been operated by the National Satellite Meteorological Center of the China Meteorological Administration since 1 May 2018 (Zhang Z. Q. et al., 2016). The function and performance of the FY-4 series have been greatly improved over those of FY-2 (China’s current operational geostationary meteorological satellite). Therefore, FY-4A would be a benefit to improve meteorological satellite observations. FY-4A is equipped with an Advanced Geostationary Radiation Imager (AGRI), Geostationary Interferometric Infrared Sounder (GIIRS), and Lightning Measurement Imager (LMI; Zhang P. et al., 2016). The AGRI onboard FY-4A has 14 spectral bands in the visible, near infrared, and TIR regions. It is designed to perform the full-disk detection in 15 min and special area detection in 3 min with a nominal nadir resolution of 1 km for the 0.64-µm visible bands or 4 km for the other bands. Operational cloud products are generated by using a variety of techniques and algorithms based on consistent and continuous measurements of the visible (VIS) through infrared (IR) spectrum.

Here we attempt to evaluate and analyze the CTH retrievals derived from FY-4A single field-of-view (FOV) by comparing them with retrievals from Himawari-8, CloudSat, Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO), and Moderate Resolution Imaging Spectroradiometer (MODIS) CTH operational products. Section 2 presents an overview of the CTH estimation algorithms of these satellites. In Section 3, CTHs are compared among the different satellites as functions of the atmospheric layers, cloud phase, and cloud optical thickness (COT). In addition, two typical cases involving variable cloud conditions are used to illustrate the characteristics of CTH products. Section 4 consists of our conclusions as well as suggestions for future work.

2 CTH retrieval algorithm 2.1 FY-4A cloud retrieval algorithm

FY-4A science product algorithms are partly based on algorithms of the FY-2/Visible Infrared Spin Scanning Radiometer (VISSR) and Geostationary Operational Environmental Satellites (GOES-R) CTH (www.GOES-R.gov). The FY-4A algorithm (Min et al., 2017) combines two window channel observations (centered at 10.8 and 12.0 µm) with a single absorption channel (centered at 13.5 µm) to allow for the estimation of cloud height without making major assumptions about cloud microphysics. By applying the CO₂ slicing method (Menzel et al., 2008) to the 13.5-µm observations and the split-window approach (Heidinger and Pavolonis, 2009) to the 10.8- and 12-µm observations, FY-4A can benefit from both the sensitivity to CTH offered by a CO₂ channel and the sensitivity to cloud microphysics offered by window channels. An optimal estimate (OE) approach (Li et al., 2001) is also used in the FY-4A CTH retrieval algorithm according to the framework presented below (Fig. 1). OE can generate the automatic estimation of retrieval errors and reduce those errors by iteration.

2.2 Himawari-8, CloudSat, CALIPSO, and MODIS overview

Himawari-8, situated over the equator around 140.7°E, is the first of Japan’s next-generation geostationary meteorological satellites. It was launched on 7 October 2014, and has been operational since 7 July 2015. The Advanced Himawari-8 Imager (AHI) onboard Himawari-8 is a visible–infrared radiometer like the Advanced Baseline Imager (ABI;Schmit et al., 2005, 2017) launched as part of the U.S.’s new generation of GOES-R. It can collect the imagery with 16 spectral channels ranging from 0.47 to 14 μm (Bessho et al., 2016). Table 1 lists the central wavelength and bandwidth of the FY-4A and Himawari-8 channels used in the CTH retrieval algorithm. The cloud retrieval algorithm of Himawari-8 is like that of the Integrated Cloud Analysis System (ICAS) algorithm, which was developed by Iwabuchi et al. (2016) to retrieve cloud properties from MODIS. ICAS also used the CO₂ slicing/split-window method together with the OE algorithm. Applying ICAS to the observations made by Himawari-8, Iwabuchi et al. (2018) demonstrated that CTHs can be estimated accurately. CTH retrieval errors are generally small in single-layer clouds but large in multi-layer clouds. The presence of a temperature inversion near the cloud tops may also introduce uncertainties for low-level clouds, particularly in the case of stratocumulus clouds that dominate the marine environments. The Japan Meteorological Agency provides L2 CLP (cloud parameters), which include the CTH, cloud top temperature, COT, and cloud type. Data used in this paper have a spatial resolution of 5 km × 5 km.

CloudSat, which is part of the NASA’s Earth Science System Pathfinder (ESSP), is equipped with a 94-GHz nadir-looking cloud profile radar (CPR). CPR is designed to detect the vertical structure of clouds by measuring the backscattered power as a function of distance from radar (Weisz et al., 2007). The CloudSat Data Processing Center (CDPC) provides standard and auxiliary products. The 1B-CPR product records the backscattered power as a function of distance from radar. Backscattered power is sampled every 240 m and classified into 125 range bins, for a total window of 30 km. CloudSat level 2 data products are produced by combining the data from an auxiliary MODIS product, water vapor fields from ECMWF, and data from CALIPSO (Mace and Zhang, 2014). The 2B-GEOPROF in version Epoch06 R04 provides the radar reflectivity data, CPR cloud mask information, and MODIS cloud flags (Delanoë and Hogan, 2008). The Cloud–Aerosol Lidar with Orthogonal Polarization (CALIOP) onCALIPSO is sensitive to thin cirrus, tenuous cloud tops, and aerosols, and thus helping to describe the cloud profile more completely (Holz et al., 2008). The CALIPSO L2 01-km-Cloud-Layer product (Version 4.10) can describe cloud layers at the high spatial resolution: 60 m vertically and 1 km horizontally (Winker et al., 2009). It should be noted that the combined 2B-GEOPROF-Lidar cloud product derived from CPR and satellite-based Lidar (CALIOP/CALIPSO) has not been used in this paper because the overlap with CALIPSO was recently degraded due to a reduction information flight maneuver. The 2B-GEOPROF product of CloudSat and the 01km-Cloud-Layer product of CALIPSO are used in the comparison.

MODIS is a visible–infrared radiation imager deployed onboard theTerra and Aqua satellites. The instruments capture data in 36 spectral bands ranging in wavelength from 0.4 to 14.4 µm at varying spatial resolutions (2 bands at 250 m, 5 bands at 500 m, and 29 bands at 1 km). More details about the MODIS Collection6 L2 cloud product can be referred to in Platnick et al. (2017). The MODIS L2 MYD06 product derived from the Aqua satellite is used for comparison in Section 3, together with the FY-4A L2 CTH, CloudSat 2B-GEOPROF, CALIPSO L2 01km-Cloud-Layer, and Himawari-8 L2 CLP products.

3 Comparison and evaluation 3.1 Validation

For validation, FY-4A CTH retrievals were compared with data from the operational products derived from the active CPR sensor onboard CloudSat, active CALIOP sensor onboard CALIPSO, passive AHI sensor onboard Himawari-8, and passive MODIS sensor onboard Aqua. We used the FY-4A full disk data at a 4-km horizontal resolution. The Himawari-8 data have a similar detection area and observation time, but the resolution is 5 km. The CloudSat data used in this study were obtained from the 2B-GEOPROF CPR cloud mask product, which provides vertical profiles of the cloud mask at resolutions of 1.1 km horizontally and 240 m vertically. Data used in this comparison cover a three-month period from August to October 2017.

Accurate temporal and spatial matches are essential for reducing uncertainties when comparing cloud products. In this paper, we incorporated the CloudSat, CALIPSO, and MODIS data by locating the observation times that were most closely matched to those of FY-4A and Himawari-8, with the maximum time difference of less than five minutes.

The varying observation view angles may introduce uncertainties since CloudSat and CALIPSO data are derived from a nadir-looking perspective. Therefore, a parallax correction is applied to FY-4A and Himawari-8 CTH retrievals based on the nearest-neighbor interpolation approach. Moreover, the view zenith angles of AGRI and AHI are limited to less than 60°. Despite of these caveats, the operational cloud products of Himawari-8, CloudSat, CALIPSO, and MODIS CTH retrievals have excellent performance reputations and are widely applied in atmospheric research (Baum et al., 2012; Iwabuchi et al., 2018). FY-4A CTH retrievals can be accurately evaluated by comparing them with these satellites’ products.

3.2 Comparison of CTH

Table 2 lists the number of the pixels, mean bias error (MBE), mean absolute error (MAE), standard deviation (STDE), and correlation coefficient of CTH products obtained from different satellites. All scene types are considered in the comparison. The input radiance data in FY-4A and Himawari-8 CTH retrievals are completely TIR, meaning that day and night are the same. We also list the proportion of land and water because it may be a factor in the results. MBE and MAE are calculated as in the equations below):

${\rm{MBE}} = \frac{1}{N}\sum\limits_{i = 1}^N {\left({x(i) - y(i)} \right)}, $

(1)

${\rm{MAE}} = \frac{1}{N}\sum\limits_{i = 1}^N {\left| {x(i) - y(i)} \right|} ,$

(2)

where x represents FY-4A or Himawari-8 data and y represents CloudSat or CALIPSO data.

Generally, CTH retrievals derived from FY-4A and Himawari-8 are lower than those from CloudSat, CALIPSO, and MODIS, with negative MBE values. Since the instruments of FY-4A, Himawari-8, and MODIS are all visible–infrared radiation imagers, the results show better consistency if we regard MODIS as the truth. The performance ofFY-4A CTH retrievals is similar to that of Himawari-8 retrievals no matter which satellite product is regarded as the truth. Meanwhile, a comparison of FY-4A and Himawari-8 shows that they are in good agreement, although the mean value of FY-4A CTH retrievals is slightly smaller than that of Himawari-8 retrievals with a negative MBE of –0.31 km.

Although the surface type may be a factor in the CTH retrieval, the proportion of land and water surfaces in Table 2 is similar in each comparison. For clouds over the water surface, FY-4A compares well with CloudSat (e.g., with a correlation coefficient of 0.84); the results are better than those obtained for clouds over land. The same tendency can be found if we regard CALIPSO and MODIS data as the truth, and FY-4A tends to obtain better results for clouds over the water surface than over the land surface.

Previous research has shown that in situations involving multi-layer clouds, particularly when an optically thin cloud overlies a lower cloud, the CTH values are located between the upper and lower clouds and are dependent on the position and characteristics of the upper-layer cloud (Watts et al., 2011). To analyze the influence of multi-layer clouds on CTH retrievals in detail, the frequency distributions of CTH differences for single-layer and multi-layer clouds are shown in Fig. 2. CloudSat and CALIPSO are regarded as the truth in this figure.

In Fig. 2, the red, blue, and green lines represent all clouds, single-layer clouds, and multi-layer clouds, respectively. Figure 2 also shows the MBE in different cases. When the CloudSat CTH retrievals are regarded as the truth, Figs. 2a, b show that the frequency distributions of the CTH difference for single-layer clouds are narrower than those for multi-layer clouds. The presence of multi-layer clouds causes a significant change in the frequency distribution of the CTH difference, although the ratio of the number of these clouds to the total number is relatively small. In the case of multi-layer clouds, the CTH difference tends to be negative (MBE: –1.93 km for FY-4A and –2.43 km for Himawari-8). These results show that FY-4A and Himawari-8 perform relatively poorly when retrieving CTHs for multi-layer clouds, and the retrieve obviously smaller CTHs than CloudSat does. Previous research has also shown that it is difficult to handle multi-layer cloud cases by using the TIR measurements and CO₂-slicing technique (Menzel et al., 2008). In Fig. 2c, we see that the frequency distribution of the CTH difference between FY-4A and Himawari-8 is very narrow, indicating that the bias is smaller. It should be noted that MBE in the case of multi-layer clouds has a positive value of 0.36 km, while it has a negative value of –0.38 km for the single-layer clouds.

Regarding CALIPSO CTH retrievals as the truth, Figs. 2d–f show similar tendencies. MBEs in Figs. 2d, e are obviously underestimated for multi-layer clouds. Figure 2f also agrees with Fig. 2c. As MBE suggests, the mean value of FY-4A CTH retrievals is bigger than that of Himawari-8 for multi-layer clouds but smaller for single-layer clouds. In addition, the STDE and correlation coefficients of FY-4A CTH retrievals for multi-layer clouds are 2.69 km and 0.64, while the corresponding values for Himawari-8 are 3.04 km and 0.54. These results indicate that FY-4A performs slightly better for multi-layer clouds than Himawari-8 if CloudSat and CALIPSO data are treated as the truth.

Based on the different phases of cloud particles, samples can be divided into liquid water clouds, ice crystal clouds, and mixed clouds. Ice clouds have radiation characteristics in the TIR band, differing from those of water clouds. These differences significantly affect the retrieval of cloud altitude (Sourdeval et al., 2013). Figure 3 shows eight scatter diagrams of CTH comparisons for ice and water clouds. CloudSat and CALIPSO data are regarded as the truth separately. Only the single-layer clouds are included to preclude the impact of multiple layers.

When compared with CloudSat as shown in Fig. 3a₁–a₄, most pixels of the ice clouds are below the diagonal line, indicating that CTH retrievals of FY-4A and Himawari-8 are generally smaller than those of CloudSat. However, MBE suggests that there is a tendency for FY-4A as well as Himawari-8 to overestimate CTHs for water clouds, especially those below 3 km.

We regard CALIPSO data as the truth in Fig. 3b₁–b₄. Here, FY-4A and Himawari-8 agree more poorly with CALIPSO than with CloudSat for ice clouds, with greater MAE and STDE values and smaller correlation coefficients. This effect may arise from the sensitivity of CALIPSO to optically thin cirrus clouds (Heidinger et al., 2010), which may be hard work for a radiance imager (Holz et al., 2008). If the cloud tops in a CALIPSO profile consist of liquid water, FY-4A also tends to overestimate the CTHs. The overestimation of Himawari-8 is more obvious with a greater MBE. According to the analysis of MODIS Collection 5 by Baum et al. (2012), temperature inversions near water cloud tops give rise to this CTH overestimation. Iwabuchi et al. (2018) studied CTH retrievals for Himawari-8 and found that the TIR method agrees well with CloudSat data when the temperature decreases at a rate of about 6.5 K km^–1 near the cloud top. When the temperature decreases at a much smaller rate or increases, CTHs derived by using the TIR method are significantly overestimated. According to Fig. 3b, the overestimation mainly occurs for water clouds below approximately 4 km. However, Himawari-8 tends to produce results closer to CALIPSO CTHs than FY-4A, with a significantly smaller MAE if all water clouds are included.

Figure 4 shows vertical variations in the CTH difference for FY-4A and Himawari-8, depicted in blue and orange, respectively, compared with CALIPSO. The diamonds represent MBE and error bars represent STDE. Below approximately 3 km, the CTH difference is positive, meaning that the mean values of CTH retrievals derived from FY-4A and Himawari-8 are higher than those derived from CALIPSO. This result indicates that the overestimation mainly appears below 3 km. Above 3 km, CTH values become negative, indicating that the CTH retrievals of FY-4A and Himawari-8 are lower than those of CALIPSO.

Figure 5 shows that CTH difference varies with COT. The COT is obtained from the Himawari-8 L2 CLP product (more details on this product can be found at www.jma.go.jp). Here, ice and water clouds are examined separately. COT is an important indicator of cloud radiation characteristic. Optically thick clouds have better IR sensitivity for CTP than transparent clouds due to the greater difference between clear and cloudy skies (Ackerman et al., 1998; Weisz et al., 2007; Sourdeval et al., 2013). As COT increases, MAE gradually decreases up to a COT value of around 10 before holding stead regardless of the cloud phase. In ice clouds, MAE shows more variability than the water case, although it exhibits the same tendency of decreasing with the increasing COT.

3.3 Case study

To reveal the strengths and weaknesses of FY-4A CTH retrieval in detail, we analyzed two typical cases containing single-layer clouds, multi-layer clouds, and various cloud types. In Fig. 6, the CTH retrievals derived from FY-4A and Himawari-8 at 0800–0815 UTC 7 September 2017, are presented in the left two plots with the CloudSat track overlaid in black; in the right plot with the CloudSat CPR cloud mask (orbit 60439, 0740–0835 UTC 7 September 2017) displayed in gray, while FY-4A and Himawari-8 data are displayed as red and green circles, respectively. The left-hand plots show that FY-4A and Himawari-8 CTH retrievals are similar under most cloud conditions. In certain areas, however (e.g., near 32°S and 93°E), FY-4A retrieved lower CTH values than Himawari-8.

Using the CloudSat 2B-CLDCLASS (cloud classification) product from CDPC, we can deduce that there are various cloud conditions, including the single-layer clouds (north of 34°S) and multi-layer clouds (such as those found near the areas B and D in Fig. 6c). In addition, there are deep convective clouds with some cirrus near the area E (e.g., between 35°S and 32°S), stratus clouds with some stratified clouds near the areas A and C (between 42°S and 41°S, and between 40°S and 39°S, respectively), and altostratus near the area B (between 41°S and 40°S).

Due to differences in orbits and instrument characteristics, the horizontal resolutions of FY-4A and Himawari-8 are both coarser than those of CloudSat. Therefore, FY-4A and Himawari-8 CTH determinations tend to retrieve an overall average value, whereas CloudSat can capture small horizontal features (such as those near the area A between 42°S and 41°S). On the other hand, FY-4A and Himawari-8 can capture continuous measurements of about one-third of the earth’s surface, while CloudSat can only measure the data along its track.

Near the areas B and D, where multi-layer clouds exist, FY-4A has difficulty obtaining correct CTH retrievals, and derives values that range between the upper and lower cloud layers. CTHs in these areas retrieved by Himawari-8 are similar, though slightly higher. Where there are only single-layer clouds (such as near the areas A and E), the bias is much smaller than that in the areas with multi-layer clouds.

Figure 5 shows that MAEs of FY-4A as well as Himawari-8 CTHs are smaller at high COT values. As shown in the area E, between 35°S and 32°S, FY-4A and Himawari-8 CTHs both agree very well with CloudSat values. North of 38°S, however, the bias is large for the retrieval of optically thin clouds, and it could be that there are thin cirrus clouds that are not being captured by CloudSat. In the single-layer cloud case, FY-4A CTH retrievals, particularly when CTH values are between 1 and 3 km (e.g., near the areas C and F), are generally lower than those captured by Himawari-8.

CALIPSO data are applied to the next case in Fig. 7. The top left image shows the FY-4A CTH south of Australia at 0500–0515 UTC 5 October 2017. The bottom-left image shows data collected by Himawari-8 at the same time, and black lines in the two images represent the CALIPSO track starting from 0509 UTC 5 October 2017. In general, FY-4A and Himawari-8 agree with each other. However, red boxes in the left two images show that Himawari-8 CTH retrievals are bigger than those of FY-4A in a certain area.

For clouds near the area B in the right image, an obvious overestimation in Himawari-8 exists. CALIPSO could capture information from optically thin clouds near the area A, with a vertical resolution of 60 m. However, these same clouds were overlooked by FY-4A as well as Himawari-8.

In the two-layer area C around 34°S, we obtain CTH values between the two layer clouds. Red pixels mostly have higher values than the green ones, indicating that FY-4A CTH retrievals are closer to CALIPSO than those to Himawari-8 for multi-layer clouds.

As the previous statistics suggest, CTH retrievals derived from FY-4A are generally smaller than values derived for single-layer clouds above 3 km by CloudSat and CALIPSO. Pixels around the area D show that Himawari-8 CTHs are slightly higher than those of FY-4A, although they both underestimate the CTH values compared with CALIPSO.

Studies of the above two typical cases show consistent statistics. FY-4A compares well with Himawari-8, although there are differences in certain cloud scenes such as multi-layer and low-level liquid water clouds.

4 Conclusions

FY-4A CTH retrievals can be accurately evaluated by comparison with data retrieved from Himawari-8, CloudSat, CALIPSO, and MODIS operational products. Cloud layers, cloud phases, and COT were investigated as factors affecting the CTH retrievals, and two typical cases were presented and analyzed. The comparison demonstrates the following strengths and weaknesses of FY-4A CTH retrievals.

The performance of FY-4A CTH retrieval using AGRI measurements was most similar to that of Himawari-8, rather than other satellites’ CTH products. Differences between FY-4A and Himawari-8 in terms of instrument characteristics, observation times, satellite view angles, and algorithms in detail may introduce uncertainties. Compared with CloudSat, CALIPSO, and MODIS, FY-4A obtained better CTH retrievals for clouds over water than over land. In atmospheric conditions featuring only single-layer clouds, both FY-4A and Himawari-8 retrieve reasonable CTH values with small biases and large correlation coefficients. However, if multi-layer clouds exist, CTHs obtained by using TIR measurements are underestimated, with values between the upper and lower cloud layers varying depending on the thickness of the uppermost layer. In liquid water clouds, particularly for lower level water clouds with the height between 1 and 3 km, the temperature inversion near the cloud tops is a major source of errors in the CTH retrieval, which may give rise to the CTH overestimation. Mean statistics show that the overestimated CTHs by FY-4A, which are affected by the temperature inversion, are slightly lower than those retrieved by Himawari-8. As the COT increases, MAE of FY-4A CTH retrievals gradually decreases.

Two typical cases with clouds of various layers, types, optical thicknesses, and physical phases were studied. Comparison of FY-4A and Himawari-8 CTH retrievals showed similar results, as they have analogous instruments and methods. In some middle-layer cloud cases, the FY-4A CTH values tended to be lower than those from Himawari-8. The CPR onboard CloudSat and CALIOP onboard CALIPSO, with their high spatial resolutions, have an advantage over FY-4A when retrieving the CTH for clouds with small horizontal extents. Therefore, in certain broken cloud cases, neither FY-4A nor Himawari-8 can accurately capture the horizontal cloud features obtained by the CPR cloud mask. Generally, FY-4A CTH retrievals perform well when there are only single-layer clouds.

Overall, FY-4A and Himawari-8 perform similarly in the CTH retrieval with respect to CloudSat, CALIPSO, and MODIS. The presence of multi-layer clouds and temperature inversions near cloud tops introduces major errors. An improved algorithm and better technology are necessary to improve the results. Better prior estimation in the OE algorithm is needed, and synergistic use of observations from high spatial resolution imagers like AGRI and high spectral resolution sounders (Menzel et al., 2018) such as GIIRS onboard the same satellite (Li et al., 2004 , 2005), should be our future goal.

Acknowledgments. This project is supported by the National Satellite Meteorological Center (NSMC) and China Meteorological Administration (CMA). The CloudSat and Himawari-8 data are provided by the CloudSat Data Processing Center (CDPC) and Meteorological Satellite Center of Japan Meteorological Agency (JMA) respectively.