In this work, the Greenhouse Gases Observing Satellite (GOSAT) Thermal and Near-infrared Sensor for Carbon Observation–Cloud and Aerosol Imager (TANSO-CAI) cloud screening results, which are necessary for the retrieval of carbon dioxide (CO2) and methane (CH4) gas amounts from GOSAT TANSO–Fourier Transform Spectrometer (FTS) observations, are compared with results from Aqua/Moderate Resolution Imaging Spectroradiometer (MODIS) in four seasons. A large number of pixels, acquired from both satellites with nearly coincident locations and times, are extracted for statistical comparisons. The same cloud screening algorithm was applied to both satellite datasets to focus on the performance of the individual satellite sensors, without concern for differences in algorithms. The comparisons suggest that CAI is capable of discriminating between clear and cloudy areas over water without sun glint and also may be capable of identifying thin cirrus clouds, which are generally difficult to detect without thermal infrared or near-infrared bands. On the other hand, cloud screening over land by CAI resulted in greater cloudy discrimination than that by MODIS, whereas detection of thin cirrus clouds tended to be more difficult over land than water, resulting in incorrect identification of thin cirrus as clear. The amount of missed thin cirrus had a seasonal variation, with the maximum occurring in summer. The cloudy tendency of CAI over half vegetation is caused by lack of an effective threshold test that can be applied to MODIS. The statistical results of the comparison clarified the important points to consider when using the results of CAI cloud screening.
Observation of greenhouse gases in the atmosphere is important to the investigation of the carbon cycle mechanism and the prediction of global climate change (e.g., Keeling et al. 1995). However, it is difficult to observe the global variation of greenhouse gases because direct sampling of gases, especially in the upper atmosphere, requires great effort and cost. Remote sensing with instruments on board satellites is expected to enable global monitoring of not only the spatial distributions of gases but also their temporal change (e.g., Rayner and O’Brien 2001). However, the development of sensors that achieve the accuracy required for investigation of the carbon cycle has been a challenging task in remote sensing research (Christi and Stephens 2004).
The Japan Aerospace Exploration Agency (JAXA), Ministry of the Environment (MOE), and National Institute for Environmental Studies (NIES) in Japan have developed the Greenhouse Gases Observing Satellite (GOSAT), which was launched in January 2009. This satellite aims at precise measurement of CO2 and CH4 amounts (Kuze et al. 2009). The first preliminary results have been presented in Yokota et al. (2009). GOSAT is a sun-synchronous satellite with a ground-track repeat cycle of about 3 days. It descends across the equator at about 1300 local time at each location. GOSAT has two instruments, namely, the Fourier Transform Spectrometer (FTS) as the main sensor, and the Cloud and Aerosol Imager (CAI), which is a narrow-band, multichannel passive imager that takes measurements at wavelengths from the ultraviolet to the near infrared. The FTS is designed to measure high-resolution spectra of reflected solar radiance in the near-infrared region and thermal emission from the earth’s surface and the atmosphere to derive molecular absorption of radiation, which is related to gas amounts in the atmosphere and their vertical profiles. However, clouds and aerosols, as well as gases in the atmosphere, influence the radiation that reaches the satellites. It is, therefore, necessary to exclude areas contaminated by clouds for the retrieval of gas amounts. Furthermore, precise estimation of gaseous absorption requires correction of the effects of scattering and absorption by aerosols in the atmosphere (e.g., O’Brien and Rayner 2002). Accordingly, CAI is designed to observe clouds and aerosols simultaneously with FTS measurements.
CAI is a push-broom imager with four bands and measures solar radiances reflected by the earth’s surface or clouds (Kuze et al. 2009). The wavelength combinations of CAI are designed to enable cloud screening, as well as determination of aerosol types, and the optical thickness and effective radius of aerosol and cloud particles. The center wavelengths of bands 1, 2, 3, and 4 are 0.38-, 0.674-, 0.87-, and 1.6-μm, respectively. Bands 1, 2, and 3 have a spatial swath width of about 1000 km and spatial resolution of 0.5 km at nadir, whereas band 4 has a swath width of about 750 km and spatial resolution of 1.5 km at nadir. Table 1 lists the specifications of the CAI bands.
Cloud screening from satellite data requires an algorithm for automated data processing. The development of cloud screening algorithms for satellite observation is a fundamental and important field of study for remote sensing, and many types of algorithms for multichannel imagers, such as the MOD35 algorithm for Moderate Resolution Imaging Spectroradiometer (MODIS) data (Ackerman et al. 1998, 2006), have been developed. Currently, NIES applies the Cloud and Aerosol Unbiased Decision Intellectual Algorithm (CLAUDIA) to the operational procedure for the CAI cloud flag product (Ishida and Nakajima 2009). CLAUDIA consists of several simple threshold tests, based on the differences of radiative properties between clouds and the earth’s surfaces. The cloud flag product not only is used to extract clear-sky areas for processing of FTS data, but also is released to the public and made available through the Internet for remote sensing research.
However, it is necessary to exercise care when using cloud screening results by CAI, because CAI has only four bands, which in some cases may be not sufficient to discriminate clouds over certain surface types. In particular, the lack of wavelengths in the thermal infrared region may result in failure to detect thin high clouds over bright surfaces. Therefore, it is significant to verify cloud screening with CAI for various cloud and surface conditions. In general, cloud screening results can be verified through comparison with data obtained by other methods. This study carries out comparisons with the cloud screening results from Aqua/MODIS data. The use of Aqua/MODIS data for the comparison has several advantages. Aqua/MODIS has 36 channels that cover wavelengths from the visible to thermal infrared regions, thus enabling more correct cloud screening compared with CAI. Simultaneous overlapping of the orbit of Aqua with CAI occurs only intermittently and is limited mainly to the Northern Hemisphere, as shown in Fig. 1, because Aqua crosses the equator during the daytime ascending from south to north. However, a large amount of data coincident in time and location can be collected for the comparison because of the wide swath of Aqua/MODIS. In contrast, the number of ground-based observation sites is limited.
In general, a complication of comparing data from two independent sensors and using two distinct cloud screening algorithms includes not only the differences in the satellite specifications but also the differences in the algorithms. However, CLAUDIA is a versatile algorithm and can be applied to all multichannel imagers. In this research, we applied CLAUDIA to both CAI and MODIS satellite data, and thereby considered only the differences in the wavelength combinations and the sensor specifications.
In section 2, we explain the cloud screening algorithm and its application to CAI and Aqua/MODIS data. Examples of CAI cloud screening results are given in section 3. In section 4 we describe the statistical analysis for determining the accuracy of cloud screening by CAI, and we investigate the characteristics of cloud screening result through several case studies for typical surface conditions. In section 5, problems with the CAI cloud screening revealed through the comparison are discussed. We summarize the paper and describe our conclusions in section 6.
In this section, we briefly describe CLAUDIA and its application to CAI data. Details of CLAUDIA are described in Ishida and Nakajima (2009).
a. Cloud screening using CLAUDIA
The term “threshold test” refers to a comparison of a value derived from satellite measurements (e.g., the reflectance in a channel) with the threshold value that determines the boundary between cloudy and clear-sky areas. A multichannel imager is usually able to perform various types of threshold tests. CLAUDIA consists of many pixel-by-pixel threshold tests to improve the accuracy of cloud screening, because a threshold test that is effective for a certain cloud or surface type may not be appropriate for others. It is, however, unavoidable that ambiguous areas remain even if many threshold tests are applied, because in nature the optical thickness of clouds continuously changes and the boundary between cloud and clear sky can be vague. Some algorithms are designed to bias the discrimination by identifying ambiguous pixels as cloudy (referred to as “clear conservative”) or clear (“cloud conservative”) under the fail-safe concept. On the other hand, CLAUDIA is designed to aim at neutral cloud screening, which means that it induces bias toward neither clear nor cloudy.
To realize neutral cloud screening, CLAUDIA includes the concept of clear confidence level (CCL) and categorization of threshold tests according to their characteristics. The CCL, which indicates the certainty of clear or cloud discrimination, has been introduced in MOD35 (Ackerman et al. 1998) to deal with ambiguous areas of MODIS observations. The CCL is defined as follows: Two threshold values—the “upper limit” and “lower limit”—are defined for a threshold test, rather than only one threshold value as is typical. If an observed value for the threshold test, which has a larger value for clear-sky scenes than that for cloudy scenes, is larger (smaller) than the upper limit (lower limit), the CCL is considered to be 1 (0). If the observed value is between the upper and lower limit, the CCL is calculated by linear interpolation to have a value between 0 and 1, which is indicative of ambiguous areas. The definition of the CCL is illustrated in Fig. 2.
The overall CCL must be comprehensively estimated from integrating the results of all threshold tests. Before the integration, we categorize threshold tests into two groups. The first group (group 1) contains threshold tests that tend to incorrectly identify clear-sky areas as cloudy. The second group (group 2) contains threshold tests that tend to incorrectly identify cloudy areas as clear. The representative CCL for group 1 G1 is calculated to be cloud conservative:
where Fk is the CCL of the kth threshold test of group 1. Equation (1) implies that even if only one threshold test takes the CCL of 1 then G1 = 1 (clear), whereas G1 = 0 (cloudy) only if all the Fk are 0. On the other hand, the representative value of group 2, G2, is given by
where Fk is the CCL of the kth threshold test of group 2. Equation (2) implies that the representative value of group 2 is considered to be “clear conservative.” Our algorithm finally computes the overall CCL, Q, from the geometric mean of the representative values for the two groups as follows:
Therefore, if the CCL of either group is 0 (i.e., cloudy), the overall CCL results in 0. The quantity Q is the output product of CLAUDIA as a result of cloud screening, that is, Q = 1 (0) indicates clear (cloudy) sky, and 0 < Q < 1 indicates ambiguity. Thus, CLAUDIA is able to carry out more neutral cloud screening than either clear or cloudy conservative.
As can be expected from the discussion above, the threshold tests in CLAUDIA are arranged in parallel, not in cascade; that is, all applicable threshold tests are performed independently, without regard for the results from the other threshold tests. CLAUDIA can be applied to any imager, containing various compositions of wavelengths, by adding or omitting possible or impossible threshold tests.
b. Application of CLAUDIA to CAI
Only a few threshold tests can be performed for CAI because it has only four bands. The reflectance of band 3 (0.87 μm) is efficient over water, whereas the reflectance of band 2 (0.68 μm) must be applied over land, because leaves of plants have a large reflectance in the near-infrared region. For more efficient cloud screening using reflectance data, we introduce “minimum albedo” maps, which are constructed from merging the minimum reflectance of each band for a month before the date of data acquisition, and apply it as the threshold values. This scheme is based on the assumption that the condition within a certain area is clear at least one time per month, and the minimum value must correspond to the reflectance of the surface. However, the minimum albedo maps sometimes contain reflectance larger than that of the surfaces not only because of persistent clouds but also sun glint. Figure 3 shows an example of the minimum albedo maps for bands 2 and 3, which are applied to the threshold tests for reflectance over land and water, respectively. Because the CAI swath of an orbit track, especially of band 4, scarcely overlaps with that of the next or previous track, the large reflectance areas due to sun glint cannot be eliminated, remaining in the minimum albedo maps. To deal with sun glint, the threshold value of reflectance for water regions is adjusted depending on the cone angle (θc) given by
where θsol, θsat, and φ are the solar zenith angle, satellite zenith angle, and relative azimuthal angle, respectively.
The ratio of the reflectance in band 2 to that in band 3 [R(band 2)/R(band 3)] can be used to detect optically thick clouds, because the reflectance of cloud in the solar radiation region is almost independent of wavelength if the absorption by air molecules is negligible. However, the threshold test with this reflectance ratio tends to mistake sun-glint regions and bright deserts for clouds. The ratio of the reflectance in band 3 to that in band 4 [R(band 3)/R(band 4)] is effective for detecting clouds over bare soil surfaces, because the reflectance of it in the near-infrared region generally increases with increasing wavelength (e.g., Irish 2000; Chikhaoui et al. 2005; Guerschman et al. 2009). However, R(band 3)/R(band 4) is not sensitive to clouds over deep forest or half vegetation.
The normalized difference vegetation index (NDVI) is based on the large reflectance of leaves in the near-infrared region and their small reflectance in the visible region. NDVI can be applied not only to estimate vegetation density (e.g., Zeng et al. 2000) but also to detect clouds over deep forests. We define the NDVI value for CAI data as follows:
where R denotes the reflectance.
All the threshold tests with CAI—the reflectance of band 2 or band 3, R(band 2)/R(band 3), NDVI, and R(band 3)/R(band 4)—are categorized into group 1; that is, these tests tend to misidentify clear-sky areas as cloudy. Therefore, the representative value of the CCL for group 1, derived from Eq. (1), is also the overall clear confidence level Q. It is expected that the incorrect identification of clear pixels as cloudy by a threshold test for CAI is adjusted by the other tests, resulting in neutral cloud screening. However, if the combination of threshold tests for CAI is insufficient, results of CAI cloud screening will include the bias to cloudy. On the other hand, if the combination of threshold tests is redundant, cloud screening results will have the bias to clear. The estimation of the bias of CAI cloud screening is an objective of this study.
The combination of the threshold tests and their threshold values are changed according to the surface type, water, land, and the polar region (defined in CLAUDIA as the areas with latitude higher than 66.6° or lower than −66.6°). The threshold test with R(band 3)/R(band 4) is not effective at detecting clouds over water. In the polar region where snow-covered areas are widespread, the threshold tests with the reflectance ratio are not applied, because these often misidentify snow-covered surfaces as clouds. The upper and lower limits of all threshold tests are determined by visual inspection, which means comparison of measured values for the threshold tests with the red–green–blue (RGB) images of MODIS and CAI. Table 2 lists the upper and lower limits that are applied to currently operational cloud screening for each surface type. For the determination of water or land, we referred to the U.S. Geological Survey (USGS) 1-km land–water tag file.
c. Application of CLAUDIA to Aqua/MODIS
In this work, CLAUDIA is also applied to MODIS level 1B 1-km radiance data. The details of the application to MODIS data are described in Ishida and Nakajima (2009). The threshold tests that can be applied to MODIS data and the group categorization are listed in Table 3, excluding the polar region. This paper does not include comparisons in the polar region, in which coincident data of CAI and Aqua/MODIS are scarce. The combination of threshold tests for MODIS includes all the tests used for CAI, but with slightly different center wavelengths for each channel. On the other hand, threshold tests of reflectance R(1.38 μm), the ratio of R(0.55 μm) to R(1.24 μm) [R(0.55 μm)/R(1.24 μm)], the ratio of R(0.905 μm) to R(0.935 μm), and the thermal infrared radiance test, which cannot be applied to CAI data, were applied to the MODIS radiances. The investigation of cloud screening results (Ishida and Nakajima 2009) suggests that R(0.55 μm)/R(1.24 μm) is necessary for cloud screening over half-vegetation areas. The threshold tests with thermal channels are also needed to detect thin cirrus clouds, which often have low reflectance. All the threshold tests for Aqua/MODIS are categorized into group 1 or group 2, except to the thermal brightness test Tb(11.0 μm), which is used as the “restoral test.” Regardless of the results of the other tests, a pixel with Tb(11.0 μm) > 297.5 K is identified as clear. This test is effective, especially in the determination of clear areas over bare soil. Because of MODIS detector problems, the 1.64-μm channel sometimes contains negative radiances for some pixels in the L1B 1-km data. The threshold test with R(0.87 μm)/R(1.64 μm), which corresponds to R(band 3)/R(band 4) of the CAI threshold test, is not carried out for such a pixel.
3. Examples of CAI cloud screening
In general, intensities of electrical signals measured with a detector on a satellite are converted to radiances by applying conversion coefficients. Currently, the first conversion coefficients determined from prelaunch on-ground calibration experiments are applied to CAI L1B data, which are distributed to general users. However, Kikuchi et al. (2009) pointed out that the current radiance data are different from the true values because of an increase in dark current, which is an unexpected output when no radiant energy enters the detector. To correct the CAI radiances, a vicarious calibration experiment has been carried out as follows (Ishida et al. 2011): First, the optical thickness and the effective radius of clouds and aerosols were estimated by applying a retrieval algorithm for cloud properties (Nakajima and Nakajima 1995) or for aerosol properties (Higurashi and Nakajima 1999, 2002) to MODIS data. Second, assuming that the MODIS data are true, the alternative conversion coefficients are set such that the cloud and aerosol properties retrieved from the CAI radiance agree with those from the MODIS data at the same location and time. All the cloud screening results in this paper are estimated from the recalibrated CAI data. However, Kikuchi et al. (2009) also indicated that the corrected dark-current signal is still problematic and that the conversion coefficients must be improved through further vicarious calibrations.
Before the comparison, we give some examples of CAI cloud screening in this section and compare them with RGB composite images to demonstrate that cloud screening of CAI data with CLAUDIA is roughly adequate by visual inspection. Figure 4 shows an example of cloud screening over ocean. Figure 4a shows the spatial distribution of the overall CCL obtained from CAI data, and Fig. 4b shows the true-color RGB composite image obtained from MODIS data for the same area at nearly the same time. In this example, almost all areas with optically thick cloud and clear areas determined visually from the RGB image have the overall CCLs of 0 and 1, respectively, while the margins of thick clouds as well as optically thin clouds often have the overall CCL of between 0 and 1. This example indicates that the cloud screening over water by CAI is suitable and capable of extracting not only clear and cloudy areas but also ambiguous areas, in accordance with the expected results by CLAUDIA.
Figure 5 shows an example over land, which consists of bright sand desert and stony desert (Mayaux et al. 2004). The cloudy areas that can be seen in the RGB image almost all have the overall CCLs of 0 or near 0. A large part of the clear areas expected from the RGB image have the overall CCL of 1, but several clear regions expected from the RGB image in the northern and western part of this image area have the overall CCL less than 1. As mentioned later, some types of land tend to be cloudy or ambiguous areas compared with MODIS. A detailed discussion about the causes of the discrepancy is described in section 5.
4. Results of comparison
a. Matching of CAI and Aqua/MODIS data
We extracted the CAI and Aqua/MODIS pixels that are coincident within 0.005° of latitude and longitude, and within a time interval of 5 min, considering these pixels as the same location and time. When this procedure is used, slight discrepancy in cloudy areas between CAI and Aqua/MODIS may exist because clouds are expected to move during the 5-min time interval. Owing to the differing sun–earth–satellite geometry between CAI and Aqua/MODIS, there may be discrepancy between the latitude and longitude of high clouds, especially when the satellite zenith angle is very oblique. To avoid this discrepancy, we selected the pixels for which the Aqua/MODIS satellite zenith angle was less than 45°. The pixels recorded by CAI when the satellite zenith angle was larger than 30° were excluded because of the limited swath of band 4. In addition, pixels that have a cone angle less than 36° for both CAI and Aqua/MODIS were excluded to avoid the sun-glint areas. Since the number of pixels was large, we took a pixel from every third column in the current row, then skipped two rows, and repeated this every third column sampling.
b. Comprehensive comparison
In this section, we discuss the comprehensive comparison of the cloud screening results, showing the statistics of the data collected for a month in each season, November in 2009, and February, May, and August in 2010. In February, only data after the eighth were used because NIES carried out a version update for the CAI L1B data at this time. The CLAUDIA algorithm allows the user to define the value of the overall CCL [Q in Eq. (3)] used to distinguish clear from cloudy pixels. To extract not only cloudy (or clear) pixels, but also ambiguous ones, the value for the boundary should be near 1 (or 0). In the results reported here, we allowed the ambiguous pixels with large (or small) Q to be considered clear (or cloudy). Pixels are grouped into three ranges according to the overall clear confidence level: 0 ≤ Q < 0.1 (cloudy), 0.1 ≤ Q ≤ 0.9 (ambiguous), and 0.9 < Q ≤ 1 (clear). A statistical analysis of the classifications was then performed.
Figure 6 shows the correspondence of Q derived from CAI and Aqua/MODIS for the pixels over water without sun glint for both sensors (i.e., cone angle less than 36°). Through all the seasons, good agreement was found among more than 80% of the compared pixels in the cloud screening results. In November, February, May, and August, about 86%, 91%, 94%, and 91%, respectively, of the cloudy pixels recorded by CAI were in agreement with the classification by MODIS, whereas about 84%, 83%, 87%, and 90%, respectively, of the clear pixels recorded by CAI were in agreement with the MODIS results. On the other hand, for each month less than 3% of the compared pixels have cloud classification completely opposite to the results recorded by Aqua/MODIS. In November, February, May, and August, about 2%, 2%, less than 1%, and 2% of the cloudy pixels recorded by CAI were oppositely identified as clear, respectively, and in all the months about 3% of the clear pixels recorded by CAI were oppositely identified as cloudy. However, the agreement of ambiguous pixels is not good compared with clear or cloudy areas. About 53%, 50%, 60%, and 54% of the ambiguous pixels recorded by CAI in November, February, May, and August were also identified as ambiguous in the Aqua/MODIS dataset, respectively. Through all of the seasons, from 22% to 28% of the ambiguous pixels recorded by CAI were identified as cloudy in the Aqua/MODIS dataset, and from 17% to 23% of the ambiguous pixels recorded by CAI were identified as clear. Furthermore, from 18% to 26% of the ambiguous pixels recorded by Aqua/MODIS were identified as cloudy or clear in the CAI dataset. The histograms of the reflectance at 0.87 μm for each range of Q in November 2009 over water are illustrated in Fig. 7. More than 90% of the clear pixels (0.9 < Q ≤ 1) had the reflectance less than 0.1. The ambiguous pixels (0.1 ≤ Q ≤ 0.9) had the peak of histogram at the range of reflectance 0.05–0.1, while the distribution for the cloud pixels (0 ≤ Q < 0.1) widened.
Figure 8 is the same as Fig. 6 but for the pixels over land. Good agreement was found among more than 61% of the compared pixels, whereas less than 3% of the clear pixels were categorized completely opposite. In November, February, May, and August, about 61%, 66%, 68%, and 67% of the cloudy pixels recorded by CAI were coincident to the results by MODIS, respectively. For the pixels determined cloudy by CAI, the smaller number of pixels in agreement is caused by the larger number of ambiguous pixels by Aqua/MODIS. About 37%, 32%, 26%, and 23% of the cloudy pixels recorded by CAI in November, February, May, and August were ambiguous pixels in the Aqua/MODIS dataset, respectively. Furthermore, about 52%, 45%, 68%, and 74% of the ambiguous pixels recorded by CAI were clear in the Aqua/MODIS dataset, respectively. These results suggest that discrimination by CAI tends to have a cloudy tendency, which means that clear areas over some land surfaces tend to be incorrectly identified as cloudy, compared with the cloud screening results by MODIS. This is consistent with the visual inspection, as mentioned in section 3. On the other hand, about 75%, 89%, 89%, and 89% of the clear pixels recorded by CAI were in agreement with the MODIS results and about 22%, 17%, 8%, and 9% of the clear pixels recorded by CAI were ambiguous in the Aqua/MODIS dataset. This implies that the difficulty of correct detection of clear areas over land by CAI is less than that of cloudy areas.
c. Case studies
In this section, we present individual cases where a discrepancy was found between CAI and MODIS in the cloud screening results.
Figure 9 shows the difference in Q between Aqua/MODIS and CAI for a broken cloud field that consists of roll-shaped, low-level small clouds. The conspicuous discrepancy of the cloud screening result, both the cloudy and clear tendency of CAI, occurs at cloud margins, where ambiguous (0.1 ≤ Q ≤ 0.9) pixels often exist. In this area, about 65% of the clear (0.9 < Q ≤ 1) pixels in the CAI dataset were identified as another category in the Aqua/MODIS dataset, and about 74% of the clear pixels in the Aqua/MODIS dataset were identified as another category in the CAI dataset. In addition, about 50% of the ambiguous pixels in the CAI dataset were identified as cloudy in the Aqua/MODIS dataset, and vice versa. Figure 10 illustrates the correspondence of the difference in Q and the difference in the reflectance at 0.87 μm. When the difference in Q is large, the difference in reflectance is also large. This suggests that the parallax effect, which can cause a large difference in the reflectance, is likely to be a cause of the difference in Q. These comparison results suggest that the differences in cloud screening did not arise from any bias toward clear or cloudy, but rather were caused by the discrepancy between cloud location for CAI and Aqua/MODIS, which becomes more apparent in such a field with small clouds.
It is generally difficult for a sensor such as CAI without infrared channels to discriminate between thin cirrus clouds and the earth’s surfaces. On the other hand, the channels in the near-infrared region with water vapor absorption (Gao et al. 1993) and in the thermal infrared region enable Aqua/MODIS to detect thin cirrus clouds. We extracted the pixels expected to contain thin cirrus, and compared Q values. Here, the following two conditions are used to determine pixels containing thin cirrus clouds over water. First, both of the reflectance in the near infrared (band 3 of CAI and channel 2 of Aqua/MODIS) must be smaller than 0.2. Second, the split window of Aqua/MODIS (difference in the brightness temperature of 10.8–12.0 μm), which is not applied to the threshold tests, must be larger than 2.5 K. Figure 11a shows the comparison results for the pixels over water, identified as cloudy by MODIS and expected to contain thin cirrus clouds during November 2009 and February 2010. It should be noted that results in May and August are not shown because the sun-glint area in these months is widespread, providing too few pixels for a robust comparison. In November and February, about 15% and 12% of the pixels expected to be cirrus were also identified as cloudy in the CAI dataset, respectively. Large amounts of cirrus pixels over water are categorized as ambiguous. However, the number of pixels with completely opposite discrimination (i.e., clear) by CAI is less than 9%. Figure 11b shows the result of the same comparison but for land surfaces. Here, pixels were extracted where the reflectance in the visible region (band 2 of CAI and channel 1 of Aqua/MODIS) was smaller than 0.25. As over water, the same split-window threshold of 2.5 K was used over land. In November, February, May, and August, about 17%, 18%, 0%, and 0% of the cirrus pixels recorded by MODIS are also determined as cloudy in the CAI dataset, respectively. Nonnegligible numbers of pixels identified as cirrus by MODIS were incorrectly classified as clear by CAI. In particular, the comparison in August resulted in identifying about 67% of cirrus pixels as clear.
Figure 12a shows the difference in Q at 1235 UTC 5 November 2009 (a part of the area in Fig. 5), as an example of cloud screening over sand desert and stony desert. Aqua/MODIS seems to correctly discriminate clear areas expected from the composite RGB image. However, the statistical result shown in Fig. 13 reveals that about 34% of the clear pixels recorded by Aqua/MODIS were identified as ambiguous in the CAI dataset. Figure 12b illustrates the distribution of R(band 3)/R(band 4). Comparing Fig. 12a with Fig. 12b reveals that the areas with a cloudy tendency in CAI cloud screening mainly correspond to larger value of R(band 3)/R(band 4).
For an example of cloud screening over half vegetation, a RGB composite image at 1233 UTC 14 November 2009, and the difference in Q are shown in Figs. 14a,b. The area in this image consists of mixed patches of some types of grassland and desert (Mayaux et al. 2004). The difference shows that a large part of the clear region expected from the RGB image has been misclassified as cloudy. Figure 15 illustrates the statistics of the comparison result of the area in Fig. 14. About 4% and 70% of the clear pixels recorded by Aqua/MODIS were identified as cloudy and ambiguous in the CAI dataset, respectively. The main reason for the cloudy tendency in CAI results is the lack of a threshold test with R(0.55 μm)/R(1.24 μm), the results for which are depicted in Fig. 14c for MODIS.
In general, cloud screening over water without sun glint is straightforward even for an imager that has fewer channels, compared with other surface types such as ocean with sun glint or land. The results of the comparison between CAI to Aqua/MODIS cloud screening indicate good agreement over water, although CAI identified a few pixels over water as the opposite condition to the results by Aqua/MODIS. Furthermore, several ambiguous areas in the CAI data are expected to be actually clear or cloudy. Several reasons for this discrepancy can be assumed: movement of clouds, the differences in sun–earth–satellite geometry (e.g., Maddux and Ackerman 2010), differences of the longitude and latitude estimation method for the pixels, and the existence of cirrus. Failed detection of cirrus may increase the underestimation of the cloud amount by CAI. As mentioned above, CAI completely missed at least 7% of thin cirrus clouds over water, while about 80% of the cirrus pixels were categorized as ambiguous. This suggests that CAI may be able to discriminate cirrus clouds from water surfaces. If a user considers ambiguous pixels as also “cloudy,” contamination of thin cirrus into clear areas can be partly avoided. On the other hand, CAI completely missed large amount of thin cirrus over land, especially in the month of August. Because land is usually brighter than water, contrast of reflected solar radiance between thin cirrus clouds and the surface over land tends to be less than that over water. This result suggests that cirrus cloud detection from the FTS measurements will be necessary prior to the retrieval of CO2 and CH4 gas amounts. Indeed, in the NIES operational analysis, the retrieval procedure for gaseous concentration includes an independent algorithm for detection of cirrus from the FTS data, in addition to the CAI CLAUDIA cloud screening.
The comparison reveals the seasonal variation of the amount of missed thin cirrus. The number of thin cirrus that CAI falsely identified as clear has a strong maximum in August and a minimum in February. Several probable reasons for this variation, to be explored in future work, may be related to seasonal changes in vegetation, changes in snow cover, or cloud type differences in each season.
The comparison indicates that CAI cloud screening over land sometimes has a cloudy tendency compared with MODIS, implying that CLAUDIA cannot eliminate the bias that each threshold test includes. In bare soil areas, the cloudy tendency is mainly caused by the increase of R(band 3)/R(band 4), as illustrated in Fig. 12. A likely explanation is the growth of vegetation. In Fig. 5, the western part, where R(band 3)/R(band 4) is large and Q is small, tends to have slightly larger NDVI and slightly smaller reflectance at 0.68 μm, than those of the areas with smaller R(band 3)/R(band 4). This appearance may correspond to a growth of vegetation. However, other areas with large R(band 3)/R(band 4) have smaller NDVI. Another likely explanation is a difference of soil type. Referring to Tsvetsinskaya et al. (2002), the area in Fig. 5 consists of various soil types. A slightly larger R(band 3)/R(band 4) occurs in the area of dunes and lithosols, which also produces large reflectances in band 2. On the other hand, the areas of yermosols tend to have smaller R(band 3)/R(band 4). However, other MODIS data in Africa indicate that this condition is not always appropriate to other areas. More investigation is needed to reveal the causes of the R(band 3)/R(band 4) increase. It should be noted that the increase of R(band 3)/R(band 4) [or R(0.87 μm)/R(1.64 μm) for MODIS] must influence both the CAI and MODIS cloud screening, but only the CAI results in a cloudy tendency. The comparison of R(band 3)/R(band 4) of CAI with R(0.87 μm)/R(1.64 μm) of MODIS, shown in Fig. 16, indicates that R(band 3)/R(band 4) of CAI tends to be slightly larger than that of Aqua/MODIS. As mentioned in section 3, the error in the radiance due to the conversion coefficients (Kikuchi et al. 2009) becomes more apparent in the calculation of reflectance ratio, which is very sensitive to changes in radiance. When applying the same threshold values, larger R(band 3)/R(band 4) of CAI than that of MODIS causes the cloudy tendency of CAI. In addition, the other tests in the MODIS cloud screening can compensate the cloudy tendency caused by the increase of R(0.87 μm)/R(1.64 μm). The error in R(band 3)/R(band 4) may be eliminated through further calibration of the CAI sensor, which will improve the cloud screening over sand desert regions.
Half-vegetation areas also result in the cloud tendency of CAI over land. In the case study mentioned in section 4, the half-vegetation areas tended to have R(band 3)/R(band 4) values about 0.9 or larger and NDVI values less than 0.3. It seems that it is difficult for CAI to discriminate these surfaces from the presence of clouds. In contrast, MODIS is able to discriminate between clear and cloudy areas with the help of an efficient threshold test using R(0.55 μm)/R(1.24 μm).
Here we deduce the influence of the tendency of CAI cloud screening on the estimation of global cloud fraction. We first consider that a pixel with 0 ≤ Q < 0.1 is cloudy and 0.1 < Q ≤ 1 is clear, for both CAI and MODIS. This consideration is expected to reduce the overestimation of cloudy areas by CAI. Then we estimate the ratio of the amount of “cloudy” pixels by CAI to “clear” pixels by MODIS. In November, February, May, and August, the ratios are 0.22, 0.33, 0.15, and 0.1, respectively. This means that CAI is expected to overestimate the cloudy areas over land by 10%–33%. Next, we categorize all pixels containing thin cirrus with 0 ≤ Q < 0.9 as cloudy and those with 0.9 < Q ≤ 1 as clear, for both CAI and MODIS. This consideration is expected to reduce the underestimation of thin cirrus areas by CAI. The ratios of “clear” pixels by CAI to “thin cirrus” pixels by MODIS in November, February, May, and August are 0.26, 0.22, 0.36, and 0.58, respectively. This means that CAI is expected to underestimate the thin cirrus areas over land by 22%–58%.
6. Summary and conclusions
We carried out a comparison of GOSAT/CAI cloud screening with Aqua/MODIS. Although the orbits of the two satellites only intermittently cross each other, the swaths of CAI and MODIS are so wide that a large amount of compared data can be collected. We applied the cloud screening algorithm CLAUDIA to CAI level 1B data and Aqua/MODIS level 1B data to obtain cloud screening results for the comparison. CLAUDIA estimates the clear confidence level, which is able to quantitatively represent the condition (clear, cloudy, or ambiguous) of a pixel. We derived the clear confidence level from CAI and Aqua/MODIS data collected in November 2009 and February, May, and August 2010 to investigate seasonal characteristics. Application of the same algorithm to two satellites is considered to allow verification of the performance of each sensor without concern for algorithm differences. We extracted the pixels at nearly the same location and time from the CAI and Aqua/MODIS data and compared the overall clear confidence levels. Although the comparison between the two satellites includes errors due to the discrepancy in the cloud location, the comparison enabled statistics of the cloud screening performance of CAI to be derived. The following conclusions were drawn from this work:
The statistical comparisons of all data in all seasons proved that CAI is generally capable of discriminating between clear and cloudy areas over water without sun glint, except for thin cirrus clouds.
Cloud screening by CAI over bare soil showed a cloudy tendency. This seems to be due to the sometimes increased values of R(band 3)/R(band 4), which is usually effective for discriminating clouds from bare soil. Cloud screening by CAI over half vegetation also had a cloudy tendency, in contrast to Aqua/MODIS. This is likely due to a lack of the threshold test using the reflectance ratio between 0.55 and 1.24 μm, which is efficient over such surfaces. These results suggest that the use of the CAI cloud flag data over land requires appropriate attention because of the overestimation of cloudy areas.
The comparison using only pixels containing thin cirrus clouds over water suggested that CAI was not able to completely discriminate thin cirrus. However, CAI may be able to detect cirrus clouds if ambiguous pixels are also categorized as “cloud.” On the other hand, the misidentified areas of thin cirrus clouds over land were larger than those over water. The amount of missed cirrus has a strong seasonal variation, with a maximum in August.
The present work does not include comparisons in apparent sun-glint regions. Cloud screening over water with sun glint is difficult, even for MODIS. Comparison of data over snow-covered surfaces was also not performed in this study because of a scarcity in coincident observations between CAI and MODIS. In this study, the variation of the satellite zenith angle is neglected. The statistical comparison using a large number of matched pixels can reduce the dependence of the cloud screening results on the satellite zenith angle. However, a detailed comparison may require consideration of the satellite zenith angle (e.g., Maddux and Ackerman 2010) and bidirectional reflectance distribution function (BRDF), which may influence cloud screening, especially over vegetated areas. To improve the validation of the CAI cloud screening results, other satellite data and methods are needed. We plan to carry out this future work and to improve the cloud flag data by selecting more appropriate threshold values for the threshold tests.
This research is supported by the GOSAT Project of the National Institute for Environmental Studies, Tsukuba, Japan (2006–2010).
Current affiliation: Graduate School of Science and Engineering, Yamaguchi University, Ube, Japan.