A study of the improvement in cloud-masking capability of data from a Moderate Resolution Imaging Spectroradiometer (MODIS) relative to data from an Advanced Very High Resolution Radiometer (AVHRR) is performed. MODIS offers significant advances over AVHRR in spatial resolution and spectral information. Three MODIS scenes that present a range of cloudiness, surface type, and illumination conditions are analyzed. AVHRR local area coverage (LAC) and global area coverage (GAC) data were synthesized from the most spectrally comparable MODIS channels. This study explores the benefits to cloud masking offered by MODIS beyond that offered by AVHRR. No global generalization can be inferred from this limited analysis, but this study does attempt to quantify the added benefit of MODIS over AVHRR for three scenes. The sole focus is on the levels of cloud contamination in clear AVHRR pixels; the misclassification of clear pixels as cloudy is not addressed. For the scenes studied, the results of the additional MODIS tests revealed measurable residual cloud contamination in both AVHRR LAC and GAC clear pixels. From this analysis, the contamination of the clear pixels in AVHRR LAC data was between 1% and 3% for the cases studied. The levels of contamination of the clear GAC pixels revealed by MODIS cloud tests ranged from 2% to 4%. MODIS was able to reveal roughly 2% more cloud contamination of clear GAC pixels than was revealed by LAC. This result indicates that the increase in spatial resolution offered by MODIS may be as significant to reducing cloud contamination as is the increase in spectral information. Inclusion of the results of AVHRR spatial uniformity tests applied to MODIS or LAC pixels revealed potentially much more cloud contamination of clear GAC pixels. The larger values of potential cloud contamination revealed by spatial uniformity tests were not apparent in the clear-sky products.
An analysis of the derived SST, land surface temperature (LST), and normalized difference vegetation index (NDVI) fields was conducted to explore the impact of the MODIS-inferred cloud contamination on these products. The results indicated minimal effects on the distribution of SST, LST, and NDVI derived from AVHRR LAC data. Errors in the GAC SST and LST had standard deviations of 0.1 and 0.3 K, respectively. The GAC NDVI error distribution has a standard deviation of 0.03 for all scenes. The GAC error distributions showed little bias, indicating that cloud-masking differences between the AVHRR and MODIS should not introduce a discontinuity in the AVHRR/MODIS/Visible Infrared Imaging Radiometer Suite (VIIRS) SST and NDVI data records.
One of the most important sources of meteorological and climate data are satellite imagers. Imagers are defined as instruments that measure reflected solar and thermal radiation and are distinguished from sounder instruments by having fewer spectral channels and high spatial resolution. With the launch of the Earth Observing System (EOS) satellite Terra and its imager, the Moderate Resolution Imaging Spectroradiometer (MODIS), the remote sensing community enters a new age of multispectral imager data. MODIS provides 36 channels with a global resolution of 1 km and a resolution in a few channels of 250 and 500 m. The instrument noise of the thermal channels of MODIS is specified to be roughly 0.1 K for a 300-K scene, while the instrument noise for the Advanced Very High Resolution Radiometer (AVHRR) for a 300-K scene is 0.17 K. Currently, the operational imager data on polar orbiting satellites, AVHRR, provides five channels (0.63, 0.86, 3.75, 10.8, and 12 μm) with a global resolution of 4 km and limited coverage at a resolution of 1 km. The AVHRR data record currently spans 20 years (1981–2001) and is expected to continue until 2018 as the primary imager on the European Meteorological Operational (METOP) polar orbiting satellite. The METOP AVHRR will provide global 1-km data. The AVHRR instrument on the National Oceanic and Atmospheric Administration (NOAA) operational polar satellites will be replaced by the Visible Infrared Imaging Radiometer Suite (VIIRS) instrument in 2008. The VIIRS instrument will most likely be a multispectral instrument with capabilities comparable to MODIS. Even though not originally designed for quantitative remote sensing, the AVHRR potentially offers the most important source of existing imager data for climate studies.
A crucial component of any imager climate dataset is the cloud mask applied to it. Cloud masking is one of the largest sources of error for some products commonly used in climate change studies such as those regarding surface temperature and the normalized vegetation index. Now that the AVHRR is being replaced by newer imagers with more channels, a question is raised as to how useful the AVHRR data record is compared to the data from the newer instruments. One may also question in what way the increase in spectral information and spatial resolution has affected the cloud-masking ability and subsequent clear-sky products of MODIS. In addition, MODIS offers reduced sensor noise compared to AVHRR. For example, the MODIS noise equivalent temperature difference for the 3.75-, 11-, and 12-μm channels are specified to be 0.05 K compared to 0.12 K for the analogous AVHRR channels.
This paper quantifies the improvement of the multispectral imager, MODIS, over AVHRR in the area of cloud masking. It is important to note that in some conditions, such as the polar night, the spectral information offered by MODIS clearly improves the ability to perform cloud masking. The cloud contamination of interest here is that occurring in regions where the AVHRR and its cloud-masking ability are thought to perform adequately well for climate studies. The scientific questions it addresses can be summarized as follows.
Relative to the MODIS data, what is the level of cloud contamination in the AVHRR data record due to cloud-masking limitations?
What is the impact of cloud contamination on the AVHRR data record?
Does this additional cloud contamination introduce any discontinuity in the AVHRR–MODIS or AVHRR–VIIRS data record?
To answer these questions, an analysis of several MODIS scenes will be performed using a core AVHRR cloud mask supplemented by the additional MODIS tests. A test in this context is defined as a condition placed on the value of channel or channel combinations used that separates clear from cloudy conditions. A test is passed if the presence of cloud is detected. As explained later, the AVHRR data is synthesized from the most similar MODIS channels. The differences between the cloud masks will be used to diagnose the relative level of cloud contamination present in the AVHRR results relative to the MODIS data. This methodology assumes that if a MODIS test detects a cloud, the cloud is really present. The impact of this cloud contamination will be assessed by its impact on the sea surface temperature (SST), the land surface temperature (LST), and the normalized difference vegetation index (NDVI), three of the more successful AVHRR climate records and two products known to be susceptible to cloud contamination. The level of this contamination will answer the question of whether the relative cloud-masking differences will cause discontinuities in the AVHRR–MODIS or AVHRR–VIIRS data records.
2. Cloud masks
A crucial component of any imager data record is the cloud mask applied to it. A vast amount of research activity has been spent developing cloud masks for imager data. Most cloud-detection algorithms are designed to exploit unique spectral and/or spatial signatures of clouds to separate clear from cloudy pixels (Inoue 1987; Yamanouchi et al. 1987; Saunders and Kriebel 1988; Ackerman et al. 1990; Stowe et al. 1999; Baum and Trepte 1999). Less common are techniques that use neural network–type algorithms trained on manually cloud-masked data (Uddstrom et al. 1999; Welch et al. 1989). Neural networks are a proven methodology for cloud masking but are not used by NOAA or the National Aeronautics and Space Administration (NASA) for the routine cloud masking of AVHRR or MODIS data, and therefore a neural network approach is not adopted here. In addition, many algorithms use estimates of clear-sky values to look for the contrast offered by cloudy pixels (Rossow and Garder 1993a,b; Vemury et al. 2001). While the knowledge of clear-sky values has obvious benefits to any cloud mask, the masks in this study do not use them. The goal here is to focus on the relative impact of the additional spectral information by using MODIS. The methodology is to use a core AVHRR cloud-mask algorithm supplemented by tests offered by the enhanced spectral information of MODIS. This methodology focuses on looking for cloud contamination of the clear pixels. The misclassification of clear pixels as cloudy is not addressed in this study. The latter problem does impact cloud analysis of the data record but it does not impact the clear-sky radiance estimates except in the amount of coverage. Regions where clear pixels are misclassified as cloudy can occur for snow/ice or desert surfaces. The cloud mask, however, does try to mitigate these misclassifications and the scenes used in this study avoid these types of conditions.
The core AVHRR cloud-mask algorithm used here is derived from the Clouds from AVHRR Phase I (CLAVR-1) cloud mask algorithm (Stowe et al. 1999). CLAVR has been used to produce a long-term climate quality imager data record (Stowe and Jacobowitz 1997). This paper assumes that CLAVR is a representative AVHRR cloud mask. The regions where AVHRR cloud masks have difficulty have been avoided to reduce concerns about the performance on the AVHRR results. For any one scene, tuning the AVHRR thresholds could improve performance and reduce the impact of additional MODIS spectral information. We chose to use the original CLAVR thresholds that were developed for global application. The CLAVR thresholds were developed for the afternoon AVHRR orbits. The Terra orbit has a descending equator crossing time of 1030 local time which provides similar solar viewing angles as seen by the afternoon NOAA satellites, indicating that differences in viewing geometry should be minimal. Note that this study is not an attempt to validate the NASA/MODIS cloud mask. Several of the cloud tests in the NASA/MODIS mask replicate the AVHRR tests from CLAVR. Since CLAVR has the longer heritage, its tests and thresholds were used. The relative impact of the MODIS spectral information is the crucial issue, not the specific thresholds on any one particular test.
a. The AVHRR cloud mask
CLAVR-1 was used for this study because it has been adopted by NOAA for its operational cloud mask and has been applied to the entire afternoon orbiting AVHRR (NOAA-7, -9, -11, -14) data record as part of the NOAA AVHRR Pathfinder for Atmospheres Project (PATMOS). The seven cloud tests used in CLAVR are listed in Table 1. These tests are analogous to most spectrally based AVHRR cloud masks, such as the Apollo Program (Saunders and Kriebel 1988). The CLAVR cloud-mask philosophy is as follows. Cloud test thresholds are set such that no clear pixel should exceed them. Restoral tests are used to restore pixels to clear after they erroneously pass cloud tests under certain stressing conditions. A crucial part of CLAVR is the use of spatial uniformity tests applied to 2 × 2 pixel arrays. For a pixel to be classified as clear it must pass tight uniformity constraints. The clear thermal/reflectance uniformity test thresholds in CLAVR are 0.5 K/0.3% for ocean pixels and 3.0 K/9% for land pixels.
The version of the CLAVR algorithm used here is slightly modified from the original. In the original CLAVR algorithm, pixels were classified as clear, mixed, or cloudy with a sequential test decision tree. To aid in understanding the results, the mixed category is split into mixed-clear and mixed-cloudy categories. As in CLAVR-1, any pixel array that fails a spatial-uniformity test applied to 2 × 2 pixel arrays of the 0.65-μm reflectance and 11-μm brightness temperature is called mixed. In this modification, the mixed-cloudy pixels have passed at least one cloud test and the mixed-clear pixels have passed no cloud test. Appendix A discusses the modifications made to CLAVR-1, and Table 2 summarizes the cloud-mask codes used in this study. Only clear pixels have historically been used to make clear-sky products from the AVHRR data record by NOAA/NESDIS; mixed-clear pixels are not used because they are assumed to be potentially cloud-contaminated. The use of dynamic thresholds on reflectance and temperature does allow mixed-clear pixels to be reclassified as clear. This procedure was implemented by Vemury et al. (2001) in the CLAVR-3 algorithm.
The choice of the CLAVR cloud-mask algorithm is not meant to limit the results of this paper to a particular algorithm. It was chosen because it is representative of many AVHRR multispectral cloud masks. Its strength is in its proven performance, its relevance to the climate data community through use in the PATMOS program, and its relevance to the operational satellite product user community through its use by NOAA.
b. The MODIS cloud mask
The MODIS instrument has 36 channels with a nominal spatial resolution of 1 km at nadir with some channels having resolutions of 250 m. The cloud mask applied by NASA to the standard MODIS data is discussed by Ackerman et al. (1998). The MODIS cloud mask is similar in philosophy to CLAVR in that a series of cloud tests are applied to each pixel. Even though the MODIS cloud-mask algorithm employs 13 cloud tests on the 1-km data (see Table 4 of Ackerman et al. 1997), Table 3 lists just those four tests that utilize the additional spectral information unique to MODIS. Of the remaining nine MODIS tests, six are replicated by CLAVR tests and three had not yet been implemented in the routine MODIS cloud mask at the time of this study. The MODIS cloud shadow, the heavy aerosol, and the 250-m reflectance tests are also not used in this study. As discussed at the MODIS cloud-mask workshop held at the University of Wisconsin—Madison (UW) Cooperative Institute for Meteorological Satellite Studies (CIMSS), these three tests are still undergoing testing and their validity on a global basis is still being determined. It is assumed that they would have little impact on the scenes chosen here.
The 1.38-μm test is designed to detect high cirrus clouds in the daytime, especially over land surfaces (Gao et al. 1993; Hutchinson and Chow 1996). The 8.55-μm test exploits the spectral variation of cloud absorption through the 8–12-μm window region to separate clouds from clear sky (Ackerman et al. 1990; Strabala et al. 1994). The 6.7-μm channel is a water vapor absorption feature, and the 13.9-μm channel is in a CO2-absorption feature. The strength of the absorption in both these channels is sufficient to remove the surface emission for most atmospheric conditions. The thresholds on these tests attempt to detect high clouds (Ackerman et al. 1997). Two additional tests listed in the MODIS Cloud Mask Algorithm Theoretical Basis Document are applied to the 1-km MODIS data using non-AVHRR channels. These tests involve the 0.94 and the 3.9-μm channels. To date, these last two tests are not implemented in the UW/CIMSS MODIS cloud mask and it is unclear if they will be. It is assumed that the exclusion of these two tests will not alter significantly the results of this study.
The MODIS thresholds on many tests have three values indicating a low, middle, and high sensitivity to cloud. To maximize the impact of the MODIS data, the thresholds used in this study were the high-sensitivity values. The values in Table 3 were provided by Richard Frey of UW/CIMSS and are used in the MODIS cloud-mask data operated by the University of Wisconsin. MODIS is still a new instrument, and its calibration and cloud-mask thresholds will certainly change over time. In addition, the MODIS thresholds used in this study have been verified through previous studies involving the MODIS airborne simulator and other hyperspectral measurements (Ackerman et al. 1998).
During the period of this study, the MODIS calibration was continually updated. Because cloud masking typically has less sensitivity to calibration than other quantitative algorithms, such as aerosol retrievals, there is little risk that MODIS calibration errors will nullify the results of this study. For example, the AVHRR cloud mask used here has been applied successfully during periods when some AVHRR channels were known to be poorly calibrated. Based on the AVHRR experience, sensor noise within the specified levels exhibits little impact on the cloud mask. Therefore, the effects of differing sensor noise from MODIS and AVHRR are also assumed to have no impact on the results of this study.
It is important to note that the official NASA/MODIS cloud mask contains different tests using the AVHRR-like channels including estimates of the surface temperature. This study is not an attempt to validate the MODIS cloud mask. By using CLAVR, which has a long heritage in producing climate-quality imager data records, we have chosen not to redefine the AVHRR thresholds based on the MODIS choices. For any one scene, any one test could be tuned to give optional results. The CLAVR thresholds are derived for global application, and no attempt is made to fine-tune the AVHRR mask for the scenes used here. The relative impact of the MODIS spectral information is the crucial issue here, not the specific thresholds on any one particular test.
The difference in equator crossing times of the Terra and the NOAA polar orbiting satellites prohibits a direct comparison of the data, except in a statistical sense. The local equator crossing time of Terra is 1030 (descending) and the equator crossing time is roughly 1630 (ascending) for NOAA-14 and 0730 (ascending) for NOAA-15. Therefore, only near the poles would the orbits overlap within two hours to allow a direct comparison. Cloud masking at high latitudes is challenging, making this choice unattractive. Since MODIS possesses channels that are similar to the AVHRR channels, the AVHRR data used in this study were synthesized by directly substituting the most similar MODIS channels. AVHRR channels 1, 2, 3b, 4, and 5 were simulated with MODIS channels 1, 2, 20, 31, and 32. AVHRR channel 3a (1.6 μm) was not used because it comprises a small fraction of the AVHRR data record and is not used by the UW/CIMSS MODIS cloud mask. The central wavelengths for the MODIS channels used are 0.65, 0.86, 3.75, 11, and 12 μm. The MODIS data used in this study have a spatial resolution of 1 km. The actual AVHRR data record consists of two datasets at differing spatial resolutions. The local area coverage (LAC) AVHRR data has a nominal nadir resolution of 1.1 km. The AVHRR channels derived directly from the 1-km MODIS data will be used as the LAC substitute. LAC data from AVHRR is available for only 10% of each AVHRR orbit. The directly broadcasted 1-km AVHRR data [high-resolution picture transmission (HRPT)] has been collected by regional sites. While not a global data record, the LAC or HRPT data offer a continuous 1-km dataset over several global regions, such as North America, Europe, and Australia. Global coverage from AVHRR is obtained with global area coverage (GAC) data with a nominal nadir spatial resolution commonly cited as 4 km. GAC data is derived from LAC data by averaging four out of five contiguous LAC values on every third scan line. Figure 1 shows schematically the GAC sampling pattern as applied to a 3 × 5 array of 1-km pixels. Therefore, the GAC samples 4 out of a possible 15 pixels and has an actual resolution of 1 km × 4 km and an apparent resolution of 3 km × 5 km. The GAC data used in this study are derived using this same procedure applied to 1-km MODIS data.
To perform this analysis, three scenes were selected. While three scenes cannot reproduce the wide range of conditions seen globally, these scenes were selected to give varying surface, cloud, and illumination conditions and should be sufficient to begin to address the questions posed in this study. Table 4 lists the location and times of the three scenes. Each scene represents a 5-min orbit segment of the 1-km MODIS data. The calibration of the MODIS channels was performed using the calibration coefficients given with the level 2 data. The conversion from radiance to brightness temperature was performed using the same procedure used by NOAA for the AVHRR thermal channels. The polynomial coefficients were generated by filtering the Planck radiance with the average spectral response functions for each detector in the focal plane array for each channel.
The first scene (scene A) is of the eastern United States, with a large band of clouds stretching from the Northeast to the Southwest. Predominantly clear conditions are found over the Midwest and northern Plains. Ahead of the large cloud band, small convective systems are seen in the southeast United States and the Gulf of Mexico. An image of the 0.86-μm reflectance of scene A is shown in Fig. 2. Scene B is a daytime scene of the Yucatan Peninsula (Fig. 3) characterized with small cumulus over land and a wide range of cloudiness over the ocean. The stripe seen in Fig. 3 was deleted from any subsequent analysis. Also seen over the bright clouds are some saturated pixels, but this has no effect on the analysis. In both scenes A and B, spectral reflection can be seen in the ocean near the middle of Figs. 2 and 3. Scene C is a nighttime scene over the eastern tropical Pacific, with a tropical storm clearly evident. Figure 4 shows the 11-μm brightness temperature image. This scene contains a mix of intense convection, cirrus, and lower-level clouds. Even though these scenes vary in region and meteorology, the results shown later are similar for all three regions.
4. Cloud-mask comparison
In this section, a comparison of the cloud masks is performed. The cloud masks were computed for each scene using the MODIS, LAC, and GAC (AVHRR-type) data. Figures 5–7 show the cloud masks for each scene using the MODIS cloud mask. Because cloud contamination of the clear pixels has potentially the most impact on the utility of the AVHRR data record, cloud contamination will be defined as any detection of cloud in a pixel determined to be clear using an AVHRR cloud mask. In the GAC analysis, a distinction will be made between cloud contamination due to detected clouds and that due to pixel spatial nonuniformity (which is indicative of cloud).
a. LAC cloud-mask comparison
Because the LAC and MODIS cloud-mask datasets in this study have the same resolution and dimensions, the comparison of the cloud-mask results can be done directly. The results for each scene are given in Tables 5–7. Because the LAC data is derived from the MODIS data, the differences seen here are due solely to presence of the additional MODIS tests listed in Table 3. Tables 5–7 are organized so that each row represents the distribution of the MODIS cloud code values, given in Table 2, for each LAC cloud code. The results for the LAC mixed-cloudy and cloudy codes (2 and 3) are not shown because they are the same in this analysis. Once a mixed-cloudy or cloudy decision is made in the LAC cloud mask, it is constrained to be the same in the MODIS cloud mask because it uses the same AVHRR cloud tests. Each row should sum to 100%, and perfect agreement would be represented by a diagonal matrix. The most important results are those that will impact the AVHRR data record, specifically the cloud contamination of the clear pixels.
The two daytime cases show similar results. The comparison with MODIS shows that roughly 4% of the clear LAC pixels are cloud-contaminated. Cloud contamination is defined here as MODIS mask values >1 for clear LAC pixels, which means MODIS shows cloud for pixels defined as clear in LAC. The contamination of the mixed-clear values is about 10%. The detection of higher levels of contamination in the mixed-clear pixels is consistent with the CLAVR assumption that spatial nonuniformity is a signature of cloud contamination. These results confirm this.
Scene C shows an order of magnitude less cloud contamination of the clear LAC pixels (0.2%). This is probably explained by the disappearance of the 1.38-μm test in the MODIS mask and the inherent cirrus detection abilities in nighttime AVHRR cloud tests [i.e., the CLAVR-1 Cirrus Test (CIRT)]. For example, even though the 8.55-μm test is detecting clouds at night, the CIRT tends to detect those same clouds resulting in little impact of the 8.55-μm test. The CIRT is a test involving the brightness temperature difference between the 3.75- and the 12-μm channels. This test responds strongly to the presence of cirrus. The CLAVR Uniform Low Stratus Test (ULST), also applied at night, involves a similar brightness temperature difference and responds strongly to low stratus. As in the daytime case, the mixed-clear cloud contamination is higher (2%) than the clear pixels.
b. Relative impact of MODIS tests
As shown above, comparison with the MODIS cloud mask revealed cloud contamination in the clear LAC data. It is important to understand which MODIS tests were responsible for detecting this contamination. This is a relevant issue because the instrument replacing the AVHRR is now being developed. A prioritization of the benefit of additional channels to the AVHRR will aid this process. The relative percentage of the cloud contamination detected by each MODIS test for each scene is given in Tables 8–10. The values in these three tables are the percentages of pixels passing the particular MODIS test that passed none of the AVHRR cloud tests, and were derived from the clear and mixed-clear LAC cloud-mask values.
For the daytime scenes, the 8.55-μm test is responsible for most (>90%) of the difference between the LAC and MODIS masks. The 1.38-μm test is responsible for much less detection of LAC cloud contamination than the 8.55-μm test. This is true especially for scene B, which has less cirrus than scene A.
For the nighttime scene, all cloud contamination of the LAC clear and mixed-clear pixels was due to the 8.55-μm test. It is important to note that the other MODIS tests are being passed (cloudy) over a significant portion of each scene. In other words, the MODIS tests are working properly but have a high level of redundancy with each other and the AVHRR tests. For example, the 6.7- and 13.9-μm tests do not detect cloud contamination because the pixels that pass those tests have already passed an existing AVHRR test.
c. GAC cloud-mask comparison
In the above analysis, any difference between the LAC and MODIS cloud masks was due solely to the impact of the additional cloud tests offered by MODIS. In this section, the cloud-mask results using GAC data are compared to results of cloud masks using LAC and MODIS data. Figure 8 shows an image of the GAC mask for scene B. The LAC and GAC masks look similar overall but differences exist, especially in the mixed-clear regions and in the general reduction in the amount of clear pixels. Much of this difference is due to tight constraints of the clear uniformity tests and the impact of spatial resolution on the inferred 2 × 2 pixel uniformity. In the GAC and MODIS cloud-mask comparisons, both the spectral and the spatial difference between the datasets are potential sources for differences. To separate out these effects, a comparison between the GAC and LAC cloud masks was performed to isolate any differences due solely to spatial-resolution effects. The differing spatial resolutions of the MODIS and GAC cloud masks prohibits a direct comparison. To compare the 1-km LAC and MODIS cloud masks to the 4-km GAC mask, the distribution of the 1-km cloud-mask values are compared to the cloud-mask value for the GAC pixel in which they are contained. In selecting the 1-km cloud values to compare to the GAC value, the identical sampling employed to construct the GAC channel observations is used. The GAC sampling may, however, introduce its own error if the four pixels chosen are not representative of the entire 3 × 5 pixel array. To explore this sampling error, a super-GAC (SGAC) sampling is also used. The SGAC sampling compares the GAC cloud-mask value to the cloud-mask values from all of the 3 × 5 pixels that fall within the GAC pixel. Using Fig. 1 as a reference, the SGAC sampling allows all 1-km values to be included in this analysis, while the GAC sampling includes the values from the gray pixels only.
The results in this section are presented as tables showing the distribution of the LAC or MODIS cloud codes within each GAC cloud code. Tables 11–13 show the comparison of the GAC and LAC cloud-mask results, and Tables 14–16 show the comparison of the GAC and MODIS cloud-mask results. The values not in parentheses are based on the four LAC or MODIS pixels sampled from the 3 × 5 pixels using the actual GAC sampling. The values in parentheses are derived using all 15 pixels falling within the GAC pixel using the SGAC sampling. The values for each row sum to 100%. For example, for scene A (Table 11), 0.38% of the LAC pixels within the clear GAC pixels were cloudy using the GAC sampling. Using the SGAC sampling, 0.52% of the LAC pixels were cloudy. Comparison of the results using the GAC and SGAC sampling show little sensitivity to the sampling method. This confirms that the GAC sampling itself does not introduce any significant cloud contamination in GAC results. All further discussions in this section will deal with the comparisons of the 1-km cloud-mask values chosen using the actual GAC sampling.
Tables 11–16 give the complete results of the LAC and MODIS versus the GAC cloud masks. In this analysis, the LAC and MODIS masks differ only in that the MODIS mask has more cloud tests. As a result of these additional cloud tests, the percentage of mixed-cloudy and cloudy pixels in the MODIS cloud mask always exceeds the comparable percentages from the LAC cloud mask. The most important measure of cloud contamination is that of the clear (0) GAC pixels. For example, only the clear GAC pixels were used to derive the clear-sky climate data records in the PATMOS dataset. In the most conservative sense, GAC cloud contamination can be defined as any nonclear LAC or MODIS pixels occurring within the GAC pixel. Using this definition, the GAC cloud contamination is computed as 100% minus the first column of the first row of Tables 11–16. For example, the GAC cloud contamination relative to the MODIS cloud mask for scene A is approximately 12%.
Analysis of the distribution of the LAC and MODIS cloud codes within the clear GAC cloud-mask values (first row) in Tables 11–16 show that this cloud contamination is dominated by the presence of mixed-clear LAC or GAC pixels. Mixed-clear pixels have passed no cloud test and only the presence of spatial variability prevents them from being classified as clear. We therefore find it useful to define a second, more conservative, definition of clear GAC cloud contamination, which ignores the presence of mixed-clear LAC or GAC. Using this definition, the clear GAC cloud contamination is computed as the sum of the third and fourth columns of the first row of Tables 11–16. When quantifying the level of GAC cloud contamination, the second definition of cloud contamination will be used because it is a more direct measure of the effect of the additional MODIS channels.
The levels of clear GAC cloud contamination for all scenes using both definitions are given in Table 17. The results show that when mixed-clear LAC or GAC pixels are included as cloud contamination, the levels of contamination are roughly 12% (scene A), 17% (scene B), and 27% (scene C). The levels of contamination revealed by the MODIS comparison are larger than those revealed by the LAC comparison for the daytime scene. For the nighttime scene, the GAC contamination is similar for both the LAC and MODIS comparisons. When mixed-clear LAC or MODIS pixels are excluded from the clear GAC contamination computation, the levels of contamination drop significantly and range from 2% to 4% for the scenes studied here. Using this definition, the contamination relative to the MODIS mask is roughly 2% higher than the contamination relative to the LAC mask for the daytime scenes. This increase in revealed clear GAC contamination by MODIS is due solely to the additional cloud-detecting capability of the additional spectral information used in the MODIS cloud mask. While 2% is a small absolute increase in contamination, it represents well over a 100% relative increase in revealed cloud contamination compared to the LAC results. For reasons described earlier, the MODIS and LAC results are similar for the nighttime cases (scene C).
In summary, the definitive level of clear GAC contamination depends greatly on whether mixed-clear pixels are included as cloud contamination. Inclusion of mixed-clear pixels gives levels of clear GAC contamination of roughly 10%–30%, while exclusion gives levels of clear GAC contamination of roughly 2%–4%. Analysis of the clear-sky products such as the aerosol optical depth has confirmed the benefits of exclusion of mixed-clear pixels from processing. To resolve this apparent conflict, the next sections will compare clear-sky products generated using the LAC, GAC, and MODIS cloud masks. The computed contamination of these clear-sky products will allow for at least a definitive estimate of the effect of this contamination of the clear GAC pixels. The differences in the surface products derived from clear GAC and MODIS pixels will help determine which definition of clear GAC contamination is more representative of the true level of contamination.
5. SST/LST comparison
In the previous section, analysis of the cloud-mask results between the AVHRR (LAC and GAC) and MODIS masks revealed that the addition of the MODIS tests did result in a reduction of the number of clear pixels and indicates that the clear AVHRR data record may suffer from cloud contamination. Relative to the MODIS data, the level of potential contamination of clear pixels in the LAC data varied from 1% to 3%. The level of contamination in the GAC data was found to vary from roughly 10% to 30% or 2% to 4%, depending on the definition of contamination used. In this section, an attempt is made to judge the impact of the potential cloud contamination by evaluation of the derived SST and LST values. The SST record from AVHRR is one of the longest, most utilized, and most important climate records from satellite data. The SST has traditionally been derived from comparison of colocated clear-sky radiances and buoy estimates of the bulk sea surface temperature. In this study, the current regression applied by NOAA to the NOAA-14 AVHRR nighttime data is used. The expression used to derive the SST is the multichannel SST (MCSST) equation (McClain et al. 1985), and the current regression has the form given by (1):
where a = 0.9367, b = 1.1301, c = 0.5979, and μ is the cosine of the satellite viewing zenith angle. While not derived for the MODIS instrument, the AVHRR values are used in this study. Since we are applying this algorithm only to MODIS datasets, any absolute error will not effect these results. Over land, we have used (1) to derive an LST. No attempt is made to correct for surface emissivity differences and this should not impact relative differences in SST or LST computed from MODIS, LAC, or GAC cloud masks. The important aspect of the SST equation for this application is that it contains the correct sensitivity to cloud contamination. Clouds can both raise and lower the values of SST retrievals. Clouds are typically cooler than the surface, and the SSTs derived from opaque clouds are typically lower than the true SST. For nonopaque clouds such as cirrus, the elevation in 11–12-μm brightness temperature difference can act to raise the SST above its true value.
a. LAC SST/LST comparison
A comparison of the LAC- and MODIS-derived SST distributions for all scenes is presented in Fig. 9. For scenes A and B, distributions for the land and ocean are presented separately. The distributions are shown only for the clear pixels. Also shown in each figure are the distributions of the SST for the pixels called clear in the LAC but not clear in the MODIS cloud mask. The two numbers shown next to each entry in the figure legends are the mean and the standard deviations of the distributions. This format will be used for all remaining figures in this paper. The shift in the distribution of nonclear MODIS pixels to colder values of SST and LST indicates that the MODIS cloud tests do tend to reduce the cloud contamination in the SST and LST distribution relative to the LAC values, which supports the assumption that cloud contamination acts to reduce the SST or LST values. Due to the small number of misclassified clear pixels (1–3), effect on the derived SST distribution from the LAC data is very small. The effect of the potential cloud contamination is also minimized because the distribution of SST or LST values for the nonclear MODIS pixels is similar to the clear distribution and varies by only 4 K in the mean. This result clearly indicates that addition of MODIS tests into the AVHRR cloud mask did reduce the amount of cloud contamination. The distribution of the nonclear MODIS pixels show that the effect of cloud contamination levels on misclassified LAC pixels could be on the order of a few degrees, with a mean of about 4 K. The effect of the addition of the MODIS tests on the total distribution of SST derived for each scene was small. The difference between the means for the SST LAC and MODIS distributions was less than 0.1 K, and less than 0.2 K for the LST distributions.
b. GAC SST/LST comparison
The results from the above comparison showed that the addition of the MODIS tests had little effect overall on the derived SST and LST distribution from the LAC cloud mask. A more important record is the SST derived from the AVHRR GAC data because it is the source for the global SST product. To estimate the level of cloud contamination in the GAC SST product, the LAC and MODIS SST or LST values were averaged over the pixels that compose each GAC pixel. In this case, the sampling used in making the GAC data was the SGAC sampling, which uses all 1-km pixels within the GAC pixel in the analysis. Only pixels determined to be clear by the LAC or MODIS cloud mask were used in the average, and only clear GAC pixels with at least one clear LAC or MODIS value were used in this analysis. This results in three GAC-resolution SST or LST fields and allows for pixel-level differences to be computed.
The results of the analysis are shown as histograms of the SST or LST differences (ΔSST or ΔLST) between the GAC and the MODIS values. The histograms were computed using a bin size 0.10 K, and the values are shown as the percentage of the ΔSST values that fell into each bin. Figure 10 shows the results for the ocean pixels in all scenes and Fig. 11 shows the results for the LST analysis. In both figures, only the GAC–MODIS SST or LST difference distributions are shown. The reason for the omission of the GAC–LAC results is that they are almost indistinguishable from the GAC–MODIS results. The similarity in these results again indicates that spatial information may be as helpful as additional spectral information for cloud masking. The effect of cloud contamination should cause negative values of ΔSST and ΔLST. The distributions of neither ΔSST nor ΔLST for each scene show a significant negative bias. Analysis of these fields showed the cause of the differences between the GAC and MODIS SST and LST values to be largely driven by the pixel-scale variation of the MODIS data within the GAC pixel. The fact that the distributions are similar indicates that the pixel-scale variations in the MODIS SST were similar for each scene. This is likely a function of both the surface temperature variations in the scenes and the performance of the MODIS instrument. The predicted performance of the MODIS channels used in computing the SST is 0.05 K. It is likely then that instrument noise is accounting for much of the difference between the GAC and MODIS SST. The larger standard deviation in distributions of ΔLST (0.3 K) than ΔSST (0.1 K) is consistent with the higher expected spatial variability over land than ocean. From these studies, it can be concluded that the effect of cloud contamination in the GAC SST or LST is at a level below the inherent pixel-scale variation of each quantity. It is evident that any potential error in the GAC SST or LST values due to cloud contamination is less than 0.1 K, which is less than the calibration accuracy of both MODIS and AVHRR. The lack of significant bias in the GAC–MODIS SST or LST shows that the effect of the differing cloud-masking capabilities should not prohibit the use of continuous AVHRR–MODIS SST or LST climatologies. This is a crucial question because MODIS is a precursor to the future instruments that will replace the AVHRR as NOAA's operational imager.
6. NDVI results
Another important AVHRR-derived climate record is the NDVI, which is computed as the following ratio using the 0.65- and 0.86-μm reflectances (Gutman et al. 1995):
Refinements of the NDVI have been developed to produce a product less sensitive to viewing geometry and atmospheric effects. As in the case with the above SST results, this analysis will not be sensitive to the absolute magnitude of the NDVI. As long as the identical algorithm is applied to clear pixels from the MODIS, LAC, and GAC cloud masks, the effects of cloud contamination can be isolated. Due to the increase in the chlorophyll reflectance at 0.86 μm relative to 0.65 μm, NDVI tends to be positive for green vegetation. The decrease in atmospheric scattering and increase in water vapor absorption at 0.86 μm tend to cause negative values for pixels composed of unvegetated land or open water. Clouds tend to have similar reflectances at both wavelengths and produce values of NDVI near zero.
a. LAC NDVI comparison
As done in the LAC SST analysis, the LAC NDVI analysis was performed comparing the distribution of NDVI values computed from the clear LAC and clear MODIS pixels over land. The results for the land pixels of scenes A and B are shown in Fig. 12. By comparing the NDVI distribution for the pixels that were clear in LAC but not in the MODIS mask to the clear NDVI distributions, we can see that the effect of the additional MODIS tests is to screen pixels with NDVI lower than the clear values. This is consistent with a reduction in cloud contamination. These results are qualitatively similar to those seen in the SST analysis. The resulting impact of the potential cloud contamination on the derived NDVI distribution is very small, owing to the small number of pixels involved. The difference in the means of the LAC and MODIS NDVI distributions is less than 0.01 for both scenes.
b. GAC NDVI comparison
A similar analysis to the GAC SST study described above was performed on the NDVI. Again, the clear LAC and MODIS pixels within each GAC pixel were averaged to give GAC-resolution fields for comparison. Figure 13 shows the GAC–MODIS and GAC–LAC NDVI difference distributions for scenes A and B. An immediate conclusion from Fig. 13 is that the GAC–MODIS and GAC–LAC results are again indistinguishable as was the case in the SST analysis. This result reiterates that an increase in spatial resolution may offer as much benefit to imager-based cloud-masking capability as an increase in spectral information. The distributions of NDVI difference have a standard deviation of roughly 0.025 and mean value of 0.006. This lack of a significant bias shows that cloud-masking capability differences will not introduce a discontinuity in the AVHRR–MODIS NDVI record.
The addition of the MODIS cloud tests does reveal cloud contamination in the AVHRR LAC and GAC cloud masks. In this sense, cloud contamination is defined as the detection of cloud by MODIS in pixels classified as clear by AVHRR LAC or GAC. Using this definition, comparison of LAC and MODIS cloud masks for the scenes studied showed that roughly 1%–3% of the LAC clear pixels suffer from cloud contamination. Comparison of the GAC and MODIS cloud masks showed the level of detected cloud contamination to vary from 2% to 4% for all scenes. The MODIS comparison revealed roughly twice the amount of GAC cloud contamination as revealed by the LAC comparison for the daytime scenes and comparable levels for the nighttime scene.
The above definition of cloud contamination ignores the detection of mixed-clear MODIS or LAC pixels within clear GAC pixels, because mixed-clear pixels are only potentially cloud contaminated. A less conservative definition of clear GAC contamination that included the effects of mixed-clear MODIS or LAC pixels gave levels of clear GAC contamination of 10%–30% for the scenes studied. The experience with AVHRR GAC data does support the conclusion that mixed-clear pixels are potentially cloud-contaminated and should not be used for clear-sky products. The small level of contamination seen in the clear GAC surface products, however, gives support to the more conservative estimates of clear GAC contamination revealed by the MODIS cloud tests.
The effect of this relative cloud contamination of the AVHRR data on the retrieved distributions of SST, LST, and NDVI were found to be minimal. The mean of the distribution of the SST derived from the clear MODIS and LAC pixels was found to be within 0.01 K for each scene. The same value for the LST was 0.05 K for each scene. The LAC NDVI distributions differed by only 0.01 in the mean compared to the values derived using the MODIS cloud mask.
The distribution of the pixel-based SST or LST differences between the GAC and MODIS data and the GAC and LAC data showed indistinguishable results. Assuming that the MODIS and the LAC offer superior cloud masking relative to the GAC, these differences can be considered errors in the GAC values due to cloud masking. For the ocean pixels, the SST error in the GAC data had a standard deviation of 0.1 K and a mean of 0.005 K. For land pixels, the standard deviation of the GAC LST error was 0.3 K and the mean was 0.005 K. The errors in the GAC NDVI distributions were found to have a standard deviation of 0.025 and a mean of 0.006.
These quantitative results give rise to three qualitative conclusions. MODIS does appear to measurably improve the capability to distinguish clear pixels over the current capability with AVHRR LAC and GAC data. The improvement in cloud masking offered by the increased global spatial resolution of MODIS appears to be as significant as the increased spectral information. The lack of any significant mean error in the GAC SST and NDVI products shows that cloud-masking differences between AVHRR and MODIS do not introduce a discontinuity in these data records for the type of scenes studied in this paper. This is a crucial result for future climatologies based on AVHRR, MODIS, and VIIRS data records.
In future work, this type of analysis will be applied to other applications, such as derived cloud and aerosol properties. The goal of these studies will be the same. They will attempt to quantify the impact of the additional spectral and spatial information offered by MODIS or VIIRS on the accuracy of the products and on the continuity with the AVHRR data record.
This work was supported by the NOAA Integrated Program Office. The thresholds and methodology of the MODIS cloud tests were generously provided by Richard Frey of the Cooperative Institute for Meteorological Satellite Science (CIMSS) at the University of Wisconsin. Crucial reviews of this manuscript were provided by Sharon Nebuda of NASA's Data Assimilation Office and other generous reviewers.
The Experimental CLAVR Cloud Mask (CLAVR-x)
The cloud mask used in this study is a modified version of the CLAVR-1 cloud mask. It is important to note that the modifications do not impact the pixels in the clear cloud code except for the glint-detection procedure described below. This version of the cloud mask is incorporated into the experimental CLAVR (CLAVR-x) processing system. CLAVR-x is a developmental test bed for an operational cloud product from NOAA and contains algorithms for cloud masking, cloud typing, and cloud product generation.
The largest difference between the cloud mask in CLAVR-x and CLAVR-1 is a division of the CLAVR-1 mixed pixels into mixed-clear and mixed-cloudy as described in section 2a. As the results in this study show, the use of a mixed-clear category allowed for additional insight on the effect of the MODIS tests on the cloud-mask comparisons. The spatial-uniformity test used to separate clear and mixed-clear pixels in CLAVR-x are identical to those in CLAVR-1. To separate cloudy from mixed-cloudy pixels, a different set of uniformity tests is used. The goal of the cloud uniformity tests is to recognize that large values of nonuniformity in reflectance, R, or window brightness temperature, T, for pixels that pass cloudy tests are probably indicative of partly cloudy pixels. The obvious problem with this logic is the presence of broken high cloud layer over a uniform lower cloud layer. The spatial uniformity presented by multilayer clouds is typically less than that seen due to partly cloud conditions. The thresholds, while preliminary, are an attempt to detect partly cloudy pixels rather than multilayer clouds. To be consistent with CLAVR-1, partly cloudy pixels are classified as mixed-cloudy. The reflectance cloud spatial-uniformity test used in CLAVR-x is
and the window channel (11 μm) cloud spatial-uniformity test is
The mean, maximum, and minimum values in (A1) and (A2) are computed using 2 × 2 pixel arrays. Every pixel in the 2 × 2 array has therefore the same uniformity values. If any one of the above tests is failed, but the pixel has passed at least one cloud test, it is classified as mixed-cloudy.
Another modification to CLAVR-1 is the processing of pixels within the cone of specular reflection over water surfaces. CLAVR-1 applied a limited set of reflectance based cloud tests where specular reflection was possible (Stowe et al. 1999). In CLAVR-x, a test involving the ratio of the computed solar reflectance in the 3.75-μm channel to the reflectance in the 0.63-μm channel is used to look for the effects of specular reflection within the region where specular reflection is possible. The fact that clouds scatter radiation much more diffusely than the ocean surface and absorb significantly at 3.75 μm causes this ratio to be typically much less than 0.50 for cloudy pixels. Values of this ratio for clear pixels with sun glint are often near unity. This test therefore applies a threshold of 0.75. If the ratio is below 0.75, the reflectance tests are applied. If the ratio is above 0.75, the reflectance cloud-mask tests are not applied because glint may be affecting the reflectance channels. This test is beneficial because the CLAVR-1 daytime infrared tests, the Thermal Gross Cloud Test (TGCT) and the Four Minus Five Test (FMFT), could miss cloud within the region where specular reflection was possible. This missing of cloud in the glint region did not affect the PATMOS results because these CLAVR-1 pixels were classified as restored clear glint and treated as missing. As seen in the cloud-mask images in Figs. 5 and 6, the CLAVR uniformity tests often result in mixed-clear rather than clear pixels in the glint regions. This acts to reduce the potential cloud contamination caused by this modification.
In order to facilitate the addition of the MODIS tests in Table 4, the sequential decision-tree-type of cloud-mask logic used in CLAVR-1 (Stowe et al. 1999) has been simplified. All cloud tests are applied to all pixels. Restoral tests are applied after the cloud tests and act to negate the effects of some tests under certain problematic conditions. This change makes it possible to analyze all cloud tests that were passed for any pixel. This type of analysis was shown in Tables 8–10. In CLAVR-1, once a pixel passes a cloud test it is not generally subjected to any other cloud tests, unless it is considered a stressing condition (restoral). CLAVR-1 adopted a decision-tree approach for computational speed needed to support NOAA operations.
The only possible modification described above that will affect the clear pixels is the extension of the reflectance tests into the glint region using the glint test. The distribution of the clear-sky products do not show any evidence that this modification is causing significant cloud contamination. Because this test is applied to both the AVHRR (LAC and GAC) and MODIS cloud masks, it cannot affect the relative difference, which is the focus of this study.
Adjustment of the 1.38-μm Threshold
This study attempts to use MODIS data to detect cloud contamination in the clear AVHRR data. The MODIS data and cloud test thresholds were taken directly from the MODIS level 2 data from the NASA Goddard Space Flight Center (GSFC) and the UW/CIMSS MODIS team. The current status of MODIS calibration is that all channels used in this study are within specification. One potential calibration concern was noted by the occurrence of slightly negative values of the channel 26 (1.38 μm) reflectance. Figure B1 shows the distribution of the channel 26 reflectance for the clear and mixed-clear pixels over land for scenes A and B, derived using the MODIS cloud mask. From these distributions, there appears to be an offset of −0.5% in the channel 26 reflectance. This small offset is only a concern because the cloud-mask threshold applied to the 1.38-μm reflectance is 3%. Because this study desires to use MODIS to detect as much cloud contamination as possible, the 1.38-μm reflectance threshold was lowered to 2.5%.
Corresponding author address: Andrew K. Heidinger, NOAA/NESDIS Office of Research and Applications, World Weather Building, E/RA1, Room 712, 5200 Auth Road, Camp Springs, MD 20746-4304. Email: email@example.com