Open Access
Add to BibSonomyAbstract
Error propagation in Earth’s atmospheric, oceanic, and land surface parameters of the satellite products caused by misclassification of the cloud mask is a critical issue for improving the accuracy of satellite products. Thus, characterizing the accuracy of the cloud mask is important for investigating the influence of the cloud mask on satellite products. In this study, we proposed a method for validating multiwavelength satellite data derived cloud masks using ground-based sky camera (GSC) data. First, a cloud cover algorithm for GSC data has been developed using sky index and bright index. Then, Moderate Resolution Imaging Spectroradiometer (MODIS) satellite data derived cloud masks by two cloud-screening algorithms (i.e., MOD35 and CLAUDIA) were validated using the GSC cloud mask. The results indicate that MOD35 is likely to classify ambiguous pixels as “cloudy,” whereas CLAUDIA is likely to classify them as “clear.” Furthermore, the influence of error propagations caused by misclassification of the MOD35 and CLAUDIA cloud masks on MODIS derived reflectance, brightness temperature, and normalized difference vegetation index (NDVI) in clear and cloudy pixels was investigated using sky camera data. It shows that the influence of the error propagation by the MOD35 cloud mask on the MODIS derived monthly mean reflectance, brightness temperature, and NDVI for clear pixels is significantly smaller than for the CLAUDIA cloud mask; the influence of the error propagation by the CLAUDIA cloud mask on MODIS derived monthly mean cloud products for cloudy pixels is significantly smaller than that by the MOD35 cloud mask.
© 2014 Optical Society of America
1. Introduction
Satellite remote sensing products play an important role in climate change studies and understanding the state of Earth’s environmental system. Generating the cloud mask by using a cloud-screening algorithm is an important step in the development of satellite remote sensing products. The role of the cloud mask is to distinguish between clear and cloudy pixels in satellite images. The reliability of the cloud mask influences the accuracy of satellite products that characterize the properties of clouds, aerosols, trace gases [1], and ground surface parameters [2]. Satellite products, except for those concerned with cloud property products, use cloud free pixels (hereafter referred to as clear pixels) in the generation process. However, their results might be contaminated by cloud if “clear” pixels determined by the cloud mask include partial or thin clouds [3,4]. On the other hand, cloud properties, such as the cloud optical thickness, effective particle radius, and cloud top temperature, are determined using a “cloudy” cloud mask, which excludes clear sky. So, the accuracy of such products might be reduced if some cloudy pixels, such as those covering cloud edges, include some clear sky [5–7]. Therefore, validation of the satellite derived cloud mask is very important to understand the influence of the cloud mask on satellite products.
Validation of the cloud mask is also useful for integration of different remote sensing products and correction of the bias error in the cloud property products derived from different satellite sensors. For example, when analyzing the time-series cloud property products from National Oceanic and Atmospheric Administration (NOAA)/Advanced Very High Resolution Radiometer (AVHRR) and Terra or Aqua/Moderate Resolution Imaging Spectroradiometer (MODIS) equipment, it is necessary to quantitatively investigate the propagation of errors caused by cloud masks from different satellite data in addition to errors caused by cloud property algorithms and the signal-to-noise ratio of the satellite sensor. Furthermore, it is also useful for the users of cloud properties and other remote sensing products to be able to choose a suitable satellite product to provide reference information.
The accuracy of the cloud mask depends not only on the reliable collection of multispectral data from various satellite sensors, but also on the performance of the cloud-screening algorithm. Stowe et al. [8] developed a cloud-screening algorithm called “CLAVR-1” for the NOAA/AVHRR. The CLAVR-1 algorithm classifies the satellite images into clear, mixed, and cloudy categories. Ackerman et al. [9] developed the MODIS cloud-screening algorithm called MOD35 for use in MODIS cloud detection. The MOD35 algorithm consists of many threshold tests with static threshold values. Frey et al. [10] improved the MODIS cloud mask for Collection 5 (MOD35-C5) and confirmed that significant improvements have been made by reprocessing it. Further, Ackerman et al. [9] and Holz et al. [11] evaluated the accuracy of the MOD35-C5 with lidar observation data from ground, aircraft, and satellite-borne instruments. Ishida and Nakajima [12] developed the cloud and aerosol unbiased decision intellectual algorithm (CLAUDIA) for used in multispectral satellite imagers, such as the currently operating Terra and Aqua/MODIS, the Cloud and Aerosol Imager (CAI) sensor onboard the GOSAT Greenhouse Gases Observing Satellite [13], the planned second generation global imager (SGLI) onboard the Global Change Observation Mission–Climate (GCOM–C) satellite [14], and the multispectral imager (MSI) onboard the Earth Clouds, Aerosols and Radiation Explorer (EarthCARE) satellite [4]. CLAUDIA is designed to realize unbiased cloud detection and calculates a clear confidence level (CCL) that indicates the probability of cloud existence using a real quantity varying from 0.0 (cloudy) to 1.0 (clear) for each individual test. Nakajima et al. [2] investigated the cloud detection performance of multispectral imagery from existing and future satellite sensors over different types of ground surfaces using CLAUDIA. They have also optimized the efficient multispectral band satellite sensor used in cloud detection. However, validation of the satellite data derived cloud mask and error propagation from the cloud mask by MOD35 and CLOUDIA on the satellite product have not been investigated sufficiently.
There are several cloud mask validation methods, which use active satellite remote sensing data, high-resolution passive satellite remote sensing data, ground-based sky camera (GSC) data, and ground-based active sensor data. Among them, the advantage of using GSC data is that the spatial and time resolutions are high, and are easy to set up in various places. The disadvantage of this method is that the observation area is narrower than that of a satellite sensor. Ackerman and Cox [15] reported that the total cloud cover using the geostationary satellite data and GSC data was similar for the daytime period. They also suggested the possibility for evaluating the satellite derived cloud mask using GSC data. Cazorla et al. [16] developed a sky imager for cloud cover assessment. Furthermore, Heinle et al. [17] and Ghonima et al. [18] developed an automatic cloud classification method using GSC data. However, the validation method for evaluating the accuracy of the satellite data derived cloud mask with GSC data has not been investigated sufficiently.
In this paper, we proposed a method for validation of the MOD35-C5 and MODIS derived cloud mask by CLAUDIA (MODIS-CLAUDIA) using the sky camera cloud mask (SCCM). For that, the accuracy of the SCCM was investigated first. Then, accuracy assessments of the MOD35-C5 and MODIS-CLAUDIA cloud masks were performed using SCCM data. Furthermore, the influence of the error propagation from MOD35-C5 and MODIS-CLAUDIA on multiwavelength MODIS data derived reflectance, brightness temperature, normalized difference vegetation index (NDVI) for clear pixels, and reflectance and brightness temperature for cloudy pixels was investigated using SCCM.
2. Materials and Methodology
In this study, the GSC data, Terra and Aqua/MODIS Level 1B radiance calibrated data (MOD-L1B), and MOD35-C5 from October 2012 through October 2013 were used for validating the MODIS derived cloud mask. The GSC data were obtained from the sky camera site at the Space Information Center (TSIC) of Tokai University Japan. MOD-L1B and MOD35-C5 data were collected from the MODIS homepage (http://modis.gsfc.nasa.gov). Among them, MOD-L1B data were used for generating the MODIS-CLAUDIA cloud mask.
Figure 1 shows the sky camera system in the TSIC, which was set up by the Japan Aerospace Exploration Agency (JAXA) in October 2012 for validating the GCOM-C/SGLI cloud mask. A Nikon D7000 fisheye camera is used in the sky camera system. In many studies, a shadow band filter is used as a Sun mask in the sky camera system in order to avoid saturation due to solar irradiation [18,19]. In this study, a FUJIFILM neutral density filter with 0.5 mm thickness was used in the sky camera system for reducing the saturation pixels in GSC data. The sun mask algorithm was also used to extract the saturation pixels instead of the shadow band filter.
Fig. 1. Appearance of the ground-based sky camera system in Tokai University Space Information Center, Kumamoto prefecture of Japan (Location: 32° 50’ 10” N, 130° 52’ 11” E).
A. Cloud Cover Algorithm
1. Cloud Cover Algorithm for Sky Camera
Yamashita and Yoshimura [20] developed a threshold-based method using the sky index (SI) and the brightness index (BI) to determine Sun, cloudy, and clear pixels from GSC data. SI indicates the blueness of the sky condition, and BI indicates the sky brightness. These indices are calculated from RGB (red, green, and blue channel) true color composed images as shown below:
where is the digital number (DN) of the blue channel, is the DN of the red channel, and is the DN of the green channel.SI values range continuously from to 1.0. When the sky is clear (blue sky) SI is close to 1.0, and SI shifts to when the sky is cloudy. BI ranges from 0.0 to 1.0; for bright skies, BI is close to 1.0, and BI shifts to 0.0 for dark skies.
To produce the SCCM, BI and SI indices were first generated from collected GSC data by using Eqs. (1) and (2). Then, these indices were used to determine thresholds for Sun, cloudy, and clear pixels. Figure 2 shows the thresholds of the Sun and cloud masks in BI-SI data for one month. Area 1 shows the clear pixels away from the Sun with low brightness, Area 2 shows the clear pixels around the Sun with higher brightness, Area 3 shows the cloudy pixels covered by thin cloud (Cirrus) with high blueness, and Area 4 shows the general cloud with lower blueness. Based on the distribution characteristics of the cloudy and clear pixels in the BI-SI table from daily GSC data over one month, the threshold for the cloud cover algorithm was determined. Four BI-SI threshold values for the cloud mask and one value for the Sun mask were determined first, as shown in Eqs. (3) and (4). Next, an SCCM was generated from the collected GSC data using the thresholds of the Sun and cloud masks. The Sun mask discounts both cloudy and clear pixels. Finally, clear coverage of the GSC data was calculated from the SCCM. Clear coverage ranges from 0.0 to 1.0, indicating the ratio of the number of clear pixels to the total number of sky camera image pixels. When the sky is clear, clear coverage is close to 1.0; for a cloudy sky, clear coverage shifts to 0.0:
where BI_SUN is the threshold for the Sun mask.Fig. 2. Threshold of the Sun and cloud mask in BI and SI plotted data from sky camera observations from 1 October 2012 to 31 October 2012. (Area 1, clear pixels away from the Sun; Area 2, clear pixels around the Sun; Area 3, cloudy pixels covered by thin cloud (cirrus); Area 4, general cloud pixels.)
2. CLAUDIA Cloud-Screening Algorithm
The MODIS-CLAUDIA is generated from the CCL by the CLAUDIA cloud-screening algorithm. The CCL is calculated from threshold tests of the multiwavelength MODIS data. Ishida and Nakajima [12] proposed that threshold tests can be categorized into two groups (Table 1). Due to the wavelength characteristics of the satellite data used in the G1 and G2 groups, the sensitivities for detecting the cloudy and clear pixels of the two groups are different from each other. Threshold tests in Group 1 are sensitive for clear areas but tend to incorrectly identify clear areas as cloudy due to a confusing ground surface. The threshold tests in Group 2 can correctly extract cloudy areas, but tend to incorrectly identify cloudy areas as clear. Considering the categorization of threshold tests, CLAUDIA derives the overall CCL from results of each test by ensuring neutral cloud screening. The representative value of the CCL for Group 1, , is derived from the geometric mean as follows:
where is the CCL of the th threshold test, and is the number of threshold tests in Group 1. Equation (5) means that when CCLs of all tests in are 0, is 0 (cloudy), whereas even if only one test is 1, is 1 (clear). On the other hand, the representative value for Group 2, , is determined from Eq. (6): where is the number of threshold tests in Group 2, which means that when CCLs of all tests in are 1, is 1 (clear), whereas even if only one test is 0, is 0 (cloudy). This implies that the representative value of Group 1 is derived to be “cloud conservative” and that of Group 2 is derived to be “clear conservative.” Finally, the overall CCL () is derived from Eq. (7), which implies that the cloud discrimination of Group 2 has priority:
Table 1. Threshold Tests of CLAUDIA for Ocean and Landb
The definition of CCL is illustrated in Fig. 3. We set two thresholds, upper limit and lower limit. If an observed value for a threshold test is larger (smaller) than the upper limit (lower limit), this pixel is considered to be clear (cloudy) and CCL is defined as 1 (0). If the value is between the upper and lower limit, this pixel is considered to be ambiguous, and the CCL is defined as between 0 and 1, calculating with linear interpolation. In contrast, there are four categories in the MOD35-C5 cloud mask: cloudy, uncertain, probably clear, and confidently clear. Details of the MOD35 algorithm are given by Ackerman et al. [9,21] and Frey et al. [10]. The advantage of MODIS-CLAUDIA is that it is possible to determine thresholds for clear and cloudy pixels from CCL values in a way that depends on the user’s needs. Figure 4 shows the MODIS RGB composite image and MODIS-CLAUDIA cloud mask in an area surrounding Japan. MODIS-CLAUDIA is represented by a real number lying between 0.0 and 1.0 [12]. However, to distinguish between cloudy and clear pixels, it is necessary to determine the threshold. In this study, threshold for completely cloudy (), cloudy (), and clear () pixels was determined for the validations.
Fig. 4. MODIS cloud mask by CLAUDIA and MOD35 cloud-screening algorithm on 25 October 2012. [Red triangle mark in (a), location of the sky camera site].
B. Process for the Validations
Since the spatial resolution of the sky camera data is much finer than MODIS data, the SCCM cloud mask is assumed to be a true representation for validating the MODIS derived MOD35-C5 and MODIS-CLAUDIA cloud masks. GSC data closest to the observation times of MODIS data were collected for the validations. Because the interval between GSC observations is 5 min, the time difference between GSC data and MODIS data is less than 5 min.
For validation of the MODIS cloud masks produced by the MOD35-C5 and CLAUDIA algorithms, SCCM data were compared with MOD35-C5 and MODIS-CLAUDIA at a resolution of pixels in the sky camera site. Furthermore, the influence of the cloud mask on MODIS products was investigated by comparison of the MODIS derived reflectances, NDVI, and brightness temperatures in clear and cloudy pixels from the different cloud masks of SCCM, MOD35-C5, and MODIS-CLAUDIA.
3. Results and Discussion
The influence of the cloud mask on satellite products for clear and cloudy pixels was investigated. Figure 5 shows the SCCM and threshold values for determining the cloudy and clear pixels used in the validations. From the comparison of the RGB images and cloud masks, we can confirm that the cloud mask area and cloud distribution in the RGB image were well matched for completely clear, completely cloudy, and partially cloudy sky conditions. We can also confirm that the Sun mask in completely clear images and the partially thin cloud mask in completely cloudy images are judged correctly, after comparing the RGB composed images and cloud mask data. In other words, the thresholds of the cloud and Sun masks efficiently distinguish clear and cloudy pixels. Thus, in this study, SCCM data are regarded as the true values for validation of MOD35-C5 and MODIS-CLAUDIA.
Fig. 5. Comparisons of the sky camera data, cloud mask, and BI, SI plotted data on 25 October 2012. (White dashed line shows thresholds of the cloud mask; percentage values in the cloud mask show cloud cover in the images.)
A. Comparison of the MODIS Derived Cloud Mask to Sky Camera Data
The clear coverage of the SCCM data obtained by the threshold-based BI-SI method was compared with MOD35-C5 and MODIS-CLAUDIA in pixel. As shown in Fig. 6, MOD35-C5 and MODIS-CLAUDIA cloud masks judge the sky to be completely cloudy or completely clear most often. However, there are some pixels that the SCCM judges to be partially cloudy when MOD35-C5 and MODIS-CLAUDIA indicate either completely cloudy or completely clear. Compared to MOD35-C5, both MODIS-CLAUDIA and the SCCM judge skies as being “clear” much more often than MOD35-C5. Conversely, both MOD35-C5 and the SCCM judge skies as being “cloudy” much more often than MODIS-CLAUDIA.
Fig. 6. Comparison of the SCCM and MODIS derived MOD35 and MOD-CLAUDIA data with pixel in sky camera site from 1 October 2012 to 1 October 2013.
The producer and user accuracies of the “completely cloudy,” “cloudy,” and “clear” classes in MOD35-C5 and MODIS-CLAUDIA with pixel resolution were investigated using the SCCM data (Fig. 7, Table 2). The producer’s accuracy relates to the probability that a reference sample will be correctly mapped and measures the errors of omission: the number of correctly classified samples of a particular category divided by the total number of reference samples for that category. In contrast, the user’s accuracy indicates the probability that a sample from the classified result actually matches that from the reference data and measures the error of commission: the number of correctly classified samples of a particular category divided by the total number of samples being classified as that category. For a “completely cloudy” class of the SCCM, the count number of pixels of the MOD35-C5 that are judged as “cloudy” is larger than that for MODIS-CLAUDIA. For a “clear” class of the SCCM, the count number of pixels that MODIS-CLAUDIA judges as “clear” is larger than that of MOD35-C5 (Fig. 7). As shown in Table 2, the producer accuracy of the “cloudy” class in MOD35-C5 is 98.8%, which is larger than the corresponding accuracies of MODIS-CLAUDIA. The producer accuracy of the “clear” class in MODIS-CLAUDIA is 92.11%, which is larger than those of MOD35-C5. However, user accuracy of the “cloudy” class in MOD35-C5 is smaller than MODIS-CLAUDIA; also the user accuracy of the “clear” class in MODIS-CLAUDIA is smaller then MOD35-C5. In other words, MOD35-C5 tends to judge ambiguous pixels as “cloudy,” and MODIS-CLAUDIA tends to judge them as “clear”.
Table 2. User and Producer Accuracy of MOD35 and MOD-CLAUDIA
B. Influence of the Cloud Mask on Satellite Products
From the validation results in the previous section, we can see that MODIS-CLAUDIA is likely to judge a pixel as “clear” even if partial or thin clouds are included in the observed pixels. On the other hand, MOD35-C5 is likely to judge such pixels as “cloudy.” We call this “clear sky contamination.” For clarifying the influence of such ambiguous cloud masks on satellite products, MODIS multiwavelength data derived reflectance, brightness temperature, and NDVI in clear and cloudy pixels by different cloud masks were investigated.
Figure 8 shows the monthly mean reflectance in the 0.65 and 0.86 μm channels, the brightness temperature in the 11.0 μm channel, and NDVI derived from MODIS multiwavelength data in clear pixels produced by MOD35-C5, MODIS-CLAUDIA, and SCCM cloud masks. The variation of the reflectance in channel 0.65 μm, the brightness temperature, and the NDVI for clear pixels produced by MOD35-C5 is more similar than the MODIS-CLAUDIA to the SCCM results [Figs. 8(a), 8(c), and 8(d)]. Reflectance in channel 0.86 μm [Fig. 8(b)] in clear pixels using MODIS-CLAUDIA and MOD35-C5 is not significantly different from that obtained with the SCCM. Reflectance of the 0.65 μm channel in clear pixels using MODIS-CLAUDIA differs significantly from that obtained using the SCCM from July 2013. Standard deviations of the reflectance of the 0.65 μm channel obtained from MODIS-CLAUDIA also differ from those obtained using the SCCM. The brightness temperature of the 11.0 μm channel [Fig. 8(c)] in clear pixels obtained using MODIS-CLAUDIA, MOD35-C5, and SCCM matches well from October 2012 to March 2013. However, MODIS-CLAUDIA results are significantly lower than those of SCCM from May 2013. The NDVI values from MOD35-C5 and SCCM [Fig. 8(d)] agree well. However, MODIS-CLAUDIA results are lower than those of SCCM from July 2013. In summary, MOD35-C5 results match those of SCCM better than MODIS-CLAUDIA, because pixels having partial or thin clouds are judged to be clear by MODIS-CLAUDIA.
Fig. 8. Monthly mean of the reflectance in 0.65 and 0.86 μm channels, brightness temperature in 11.0 μm channel, and NDVI in clear pixels by different cloud masks.
Figure 9 shows the monthly maximum value composition (MVC) of the reflectance, brightness temperature, and NDVI for clear pixels. The monthly MVC reflectance results for MODIS-CLAUDIA, MOD35-C5, and SCCM are significantly different from each other. However, brightness temperature and NDVI match well. The reason for this is considered to be that if a “clear” pixel is contaminated by thin or partial cloud, reflectance in the pixel increases, while brightness temperature and NDVI decrease. Thus, by using the monthly MVC method it is possible to prevent the propagation of errors from the cloud mask to the brightness temperature and NDVI results.
Fig. 9. Monthly maximum value composition in 0.65 and 0.86 μm channels, brightness temperature in 11.0 μm channel, and NDVI in clear pixels by different cloud masks.
The reflectance channel in the satellite sensor is effective for retrieving clouds’ optical thicknesses. The thermal band channel is effective for monitoring cloud top temperature [22–24]. To investigate the influence of the cloud mask on the cloud property products, monthly mean reflectances and brightness temperatures for cloudy pixels were compared (Fig. 10). In contrast to the clear pixel results shown in Fig. 8, the variation of the reflectance and brightness temperature for cloudy pixels obtained using MODIS-CLAUDIA is more similar to the SCCM results than those obtained using MOD35-C5. Reflectance in the 0.65 μm channel obtained using MOD35-C5 differs significantly from the SCCM results, especially from July 2013. The variation of the reflectance in the 0.86 and 0.65 μm channels is almost the same. Furthermore, the reflectance obtained using MOD35-C5 is significantly less than that of MODIS-CLAUDIA and SCCM. This result is consistent with the result, discussed in Section 3.A, that MOD35-C5 is likely to judge as “cloudy” pixels that include partial or thin cloud. In that case, part of the sunlight in the pixel can reach the ground surface. When the ground surface albedo is lower than the cloud albedo, reflectance in this pixel would be lower than the reflectance in a completely cloudy pixel. The variations of the brightness temperature from MODIS-CLAUDIA and SCCM are almost the same. However, the brightness temperature from MOD35-C5 is larger due to clear sky contamination in cloudy pixels.
Fig. 10. Monthly mean reflectance of the 0.65 and 0.86 μm channels, and brightness temperature (11.0 μm channel) in cloudy pixels by different cloud masks.
4. Conclusions
In this study, we proposed a method to validate MODIS multiwavelength derived MOD35-C5 and MODIS-CLAUDIA cloud masks using sky camera data. The employed method proved that GSC data are efficient to evaluate the accuracy of the satellite derived cloud masks. The error propagation from MODIS-CLAUDIA and MOD35-C5 to satellite derived ground surface and atmospheric parameters was also investigated. Our conclusions are as follows:
- • The user and producer accuracies of the “cloudy” class in MOD35-C5 are larger than those in MODIS-CLAUDIA; for the “clear” class MODIS-CLAUDIA accuracies are larger than those of MOD35-C5. In other words, the MOD35-C5 cloud mask tends to judge ambiguous pixels as “cloudy,” while MODIS-CLAUDIA tends to judge them as “clear.”
- • Errors in the cloud mask significantly influence the satellite derived monthly mean reflectance, brightness temperature, and NDVI for clear pixels. The influence of the error propagation by MOD35-C5 on satellite derived monthly mean reflectance, brightness temperature, and NDVI for clear pixels is significantly smaller than MODIS-CLAUDIA. However, using the monthly MVC of the brightness temperature and NDVI can prevent this effect.
- • Errors in the cloud mask significantly influence the monthly mean reflectance and brightness temperature for cloudy pixels. The influence of the error propagation by MODIS-CLAUDIA on cloud products is significantly smaller than that by MOD35-C5.
According to the results above, it was demonstrated that sky camera data are efficient for validating multiwavelength satellite imager derived cloud masks and investigating the influence of the cloud mask on the satellite product despite fewer channels. The results presented in this study are not only useful for users of the multiwavelength satellite data to select the cloud mask, but are also available for the science team of a satellite mission to validate the accuracy of the cloud mask and remote sensing product in the prelaunch phase of the satellite sensor.
In this study, the cloud mask of the sky camera is assumed to be a true representation for validating the MODIS derived cloud mask. However, it is considered that the validation results include some bias, which is caused by different view geometries from the sky camera and satellite sensor. As a future work, it is important to investigate the bias caused by different view geometries.
This work was supported by the GCOM-C/SGLI and EarthCARE project of the Japan Aerospace Exploration Agency (JAXA), and the Japan Science and Technology Agency (JST), CREST/EMS/TEEDDA. This study was also partly supported by the GOSAT project of the National Institute of Environmental Study, Tsukuba, Japan.
References
1. J. Xu, F. Schreier, P. Vogt, A. Doicu, and T. Trautmann, “A sensitivity study for far infrared balloon-borne limb emission sounding of stratospheric trace gases,” Geosci. Instrum. Methods Data Syst. Discuss. 3, 251–303 (2013).
2. T. Y. Nakajima, T. Tsuchiya, H. Ishida, T. N. Matsui, and H. Shimoda, “Cloud detection performance of spaceborne visible-to-infrared multispectral imagers,” Appl. Opt. 50, 2601–2616 (2011). [CrossRef]
3. T. Y. Nakajima, T. N. Matsui, H. Letu, K. Suzuki, H. Ishida, N. Kikuchi, G. L. Stephens, T. Nakajima, and H. Shimoda, “Cloud sciences using satellite remote sensing, cloud growth model, and radiative transfer,” Proc. SPIE 7859, 785905 (2010).
4. T. Y. Nakajima, T. M. Nagao, H. Letu, H. Ishida, and K. Suzuki, “On the cloud observations in JAXA’s next coming satellite missions,” Proc. SPIE 8523, 852316 (2012).
5. Z. Zhang and S. Platnick, “An assessment of differences between cloud effective particle radius retrievals for marine water clouds from three MODIS spectral bands,” J. Geophys. Res. 116, D20 (2011).
6. T. M. Nagao, K. Suzuki, and T. Y. Nakajima, “Interpretation of multiwavelength-retrieved droplet effective radii for warm water clouds in terms of in-cloud vertical inhomogeneity by using a spectral bin microphysics cloud model,” J. Atmos. Sci. 70, 2376–2392 (2013). [CrossRef]
7. T. M. Nagao, T. Y. Nakajima, H. Letu, and H. Okamoto, “Cloud microphysical properties as seen from spaceborne passive multi-spectral imagers: interpretation in terms of vertical and horizontal inhomogeneity by using modeling and other spaceborne instruments,” Trans. JSASS Space Tech. Japan 12, 1–6 (2014).
8. L. L. Stowe, P. A. Davis, and E. P. McClain, “Scientific basis and initial evaluation of the CLAVR-1 global clear/cloud classification algorithm for the Advanced Very High Resolution Radiometer,” J. Atmos. Ocean. Technol. 16, 656–681 (1999). [CrossRef]
9. S. A. Ackerman, K. I. Strabala, W. P. Menzel, R. A. Frey, C. C. Moeller, and L. E. Gumley, “Discriminating clear sky from clouds with MODIS,” J. Geophys. Res. 103, 32141–32157 (1998). [CrossRef]
10. R. A. Frey, S. A. Ackerman, Y. Liu, K. I. Strabala, H. Zhang, J. R. Key, and X. Wang, “Cloud detection with MODIS. Part I: improvements in the MODIS cloud mask for Collection 5,” J. Atmos. Ocean. Technol. 25, 1057–1072 (2008). [CrossRef]
11. R. E. Holz, S. A. Ackerman, F. W. Nagle, R. Frey, S. Dutcher, R. E. Kuehn, and B. Baum, “Global moderate resolution imaging spectroradiometer (MODIS) cloud detection and height evaluation using CALIOP,” J. Geophys. Res. 113(D8), 27 (2008).
12. H. Ishida and T. Y. Nakajima, “Development of an unbiased cloud detection algorithm for a spaceborne multispectral imager,” J. Geophys. Res. 114(D7), 16 (2009).
13. H. Ishida, T. Y. Nakjima, T. Yokota, N. Kikuchi, and H. Watanabe, “Investigation of GOSAT TANSO-CAI cloud screening ability through an intersatellite comparison,” J. Appl. Meteorol. Climatol. 50, 1571–1586 (2011).
14. H. Letu, T. Y. Nakajima, and T. N. Matsui, “Development of an ice crystal scattering database for the global change observation mission/second generation global imager satellite mission: investigating the refractive index grid system and potential retrieval error,” Appl. Opt. 51, 6172–6178 (2012). [CrossRef]
15. S. A. Ackerman and S. K. Cox, “Comparison of satellite and all-sky camera estimates of cloud cover during GATE,” J. Appl. Meteorol. 20, 581–587 (1981).
16. A. Cazorla, F. J. Olmo, and L. Alados-Arboledas, “Development of a sky imager for cloud cover assessment,” J. Opt. Soc. Am. A 25, 29–39 (2008). [CrossRef]
17. A. Heinle, A. Macke, and A. Srivastav, “Automatic cloud classification of whole sky images,” Atmos. Meas. Tech. Discuss. 3, 269–299 (2010).
18. M. S. Ghonima, B. Urquhart, C. W. Chow, J. E. Shields, A. Cazorla, and J. Kleissl, “A method for cloud detection and opacity classification based on ground based sky imagery,” Atmos. Meas. Tech. Discuss. 5, 4535–4569 (2012).
19. C. W. Chow, B. Urquhart, M. Lave, A. Dominguez, J. Kleissl, J. Shields, and B. Washom, “Intra-hour forecasting with a total sky imager at the UC San Diego solar energy testbed,” Sol. Energy 85, 2881–2893 (2011). [CrossRef]
20. M. Yamashita and M. Yoshimura, “Development of sky conditions observation method using whole sky camera,” J. Jpn. Soc. Photogramm. Remote Sens. 47, 50–59 (2008).
21. S. A. Ackerman, R. E. Holz, R. Frey, E. W. Eloranta, B. C. Maddux, and M. McGill, “Cloud detection with MODIS. Part II: validation,” J. Atmos. Ocean. Technol. 25, 1073–1086 (2008). [CrossRef]
22. T. Y. Nakajima and T. Nakajma, “Wide-area determination of cloud microphysical properties from NOAA AVHRR measurements for FIRE and ASTEX regions,” J. Atmos. Sci. 52, 4043–4059 (1995). [CrossRef]
23. K. Kawamoto, T. Nakajima, and T. Y. Nakajima, “A global determination of cloud microphysics with AVHRR remote sensing,” J. Climate 14, 2054–2068. (2001).
24. H. Iwabuchi and T. Hayasaka, “A multi-spectral non-local method for retrieval of boundary layer cloud properties from optical remote sensing data,” Remote Sens. Environ. 88, 294–308 (2003). [CrossRef]
