Satellite drift is a historical issue affecting the consistency of those few satellite records capable of being used for studies on climate time scales. Here, the authors address this issue for the Pathfinder Atmospheres Extended (PATMOS-x)/Advanced Very High Resolution Radiometer (AVHRR) cloudiness record, which spans three decades and 11 disparate sensors. A two-harmonic sinusoidal function is fit to a mean diurnal cycle of cloudiness derived over the course of the entire AVHRR record. The authors validate this function against measurements from Geostationary Operational Environmental Satellite (GOES) sensors, finding good agreement, and then test the stability of the diurnal cycle over the course of the AVHRR record. It is found that the diurnal cycle is subject to some interannual variability over land but that the differences are somewhat offset when averaged over an entire day. The fit function is used to generate daily averaged time series of ice, water, and total cloudiness over the tropics, where it is found that the diurnal correction affects the magnitude and even the sign of long-term cloudiness trends. A statistical method is applied to determine the minimum length of time required to detect significant trends, and the authors find that only recently have they begun generating satellite records of sufficient length to detect trends in cloudiness.
The satellite era has provided us with an unprecedented ability to globally monitor the earth. While the applications of this are numerous, from a climatological perspective detection of statistically significant anomalies over this relatively short era pose some unique challenges. This is especially true for cloud, which is a key driver of the earth’s energy budget and potential bellwether for changes in climate. Satellite measurements of clouds are difficult to standardize, not only for use with validating climate models, which typically do not explicitly resolve subgrid cloud processes, but also for intercomparison with other satellites. This is attributable to many factors including variation among cloud-masking algorithms and sensor differences such as viewing angle, pixel footprint size, spectral variation, and calibration (Foster et al. 2011). In addition, the satellite records long enough to identify trends on climate time scales, such as the National Oceanic and Atmospheric Administration (NOAA) polar orbiter series (POES), are also those most prone to historical issues such as lack of onboard calibration systems and satellite drift. Having said this and despite the obstacles, there have been several initiatives aimed at solving these problems resulting in greater flexibility and compatibility of satellite cloud records (Klein and Jakob 1999; Webb et al. 2001; Stubenrauch et al. 2009; Pincus et al. 2012).
Recently work has been done to address the historical issues of the long-lived Advanced Very High Resolution Radiometer (AVHRR) sensor record by leveraging the capabilities of modern satellite sensors (Heidinger et al. 2010, 2012). These multiplatform approaches are proving to be effective tools for providing reliable, consistent, long-term satellite cloud climatologies. The work presented here is focused on addressing the issue of satellite drift, a source of significant uncertainty for the AVHRR, which has been flown on the NOAA POES since November 1978. As a satellite’s time in polar orbit progresses the orbital height gradually decays resulting in a slow drift of local equatorial crossing time. When monitoring phenomena with a diurnal cycle, such as cloud formation and microphysical processes, the changing local time for sensor measurements can cause spurious trends in long-term records. This occurrence is known as aliasing and must be addressed before multiple satellites can be merged into a consistent record. Resolution of a diurnal cloudiness cycle is a common enough pursuit (Bergman and Salby 1996; Cairns 1995; Dai and Trenberth 2004; Gray and Jacobson 1977; Rozendaal et al. 1995; Slingo and Yang 2001; among many others). One reason for this is that the diurnal cloudiness cycle is associated with a well-defined solar forcing cycle. Another is that it affects many ocean and land surface exchange processes as well as convection, boundary layer formation, longwave and shortwave fluxes, and latent heat release. Taken together the diurnal cloudiness cycle is an important validation tool for models or studies concerning the energetics of the surface and atmosphere. These studies provide benchmarks with which to validate our own derivation of the diurnal cloudiness cycle. Previous studies have fitted sinusoidal curves to diurnal cycles of maritime cloud liquid water path (O’Dell et al. 2008; Wood et al. 2002). The O’Dell et al. (2008) study used a sinusoidal fit function to correct for satellite drift during the construction of a microwave-based monthly maritime liquid water path climatology. Here we apply a similar method of constructing a diurnal cycle over the lifetime of the AVHRR imager record to generate a daily averaged global record of cloudiness over all surface types. The datasets and methodologies used in this study are described in section 2. In section 3, the results are presented including a validation of the diurnal cycle against a geostationary satellite as well as its long-term stability and an exploration of the error introduced by inconsistent diurnal sampling over the lifetime of the AVHRR record. A time series of total, ice, and water cloudiness is examined in the tropics and an assessment of our ability to detect statistical trends is performed. Section 4 contains discussion and a summary of the results.
2. Data and method
The Pathfinder Atmospheres Extended (PATMOS-x) is a suite of cloud products developed by NOAA. The PATMOS-x record primarily used in this study is generated from the AVHRR sensors flown on the NOAA polar orbiter satellite series and more recently the European Organization for the Exploitation of Meteorological Satellites (EUMETSAT) Meteorological Operation (MetOp) series, though the PATMOS-x processing framework has been generalized to ingest and produce cloud products from the Moderate Resolution Imaging Spectroradiometer (MODIS) and Geostationary Operational Environmental Satellites (GOES) data as well. This flexibility can be a great advantage when attempting a product validation, as evidenced by the AVHRR–GOES diurnal cycle comparison included in this study. The AVHRR/2 series has five sensors with central wavelengths at 0.63, 0.86, 3.75, 10.8, and 12.0 μm and flew on NOAA-7 through NOAA-14. The AVHRR/3 series has an additional channel with a central wavelength of 1.6 μm and flew on NOAA-15 through NOAA-19 as well as the EUMETSAT MetOp satellites. The AVHRR record begins in 1978 with the first four-channel AVHRR flown on Television and Infrared Observation Satellite-N (TIROS-N; missing the 12.0-μm channel), but in this study we are concerned with the five-channel record, beginning in 1981.
The pixel-level product used here is based on measurements from global area coverage (GAC) data. Each GAC pixel is the average of four 1.1-km AVHRR pixels. The geographic area represented by one GAC pixel is approximately 3 km by 5 km. Specifically we use the level-2b product, which is the level-2 pixel-level data subsampled to a 0.1° equal-angle global grid. Ancillary data include National Centers for Environmental Prediction (NCEP) reanalysis data used for surface temperature, snow cover, and vertical profiles of water vapor and temperature. Surface albedo is derived from MODIS observations (Moody et al. 2008).
The calibration of the AVHRR visible channels is done using AVHRR-MODIS simultaneous nadir overpasses (SNOs), AVHRR-AVHRR SNOs, and stable earth targets (Heidinger et al. 2010). The cloud-masking algorithm uses a naive Bayesian scheme with six classifiers calibrated over multiple surface types (Heidinger et al. 2012). The training data necessary for the naive Bayesian approached is obtained using Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) SNOs. Classifiers include 0.63-μm reflectance, 10.8-μm emissivity, 10.8-μm temperature contrast for adjacent fields of view, 10.8- and 12-μm temperature contrast, and daytime and nighttime 3.75-μm emissivity. The skill of the detection algorithm is based on a number of factors including but not limited to surface type, ancillary data, solar zenith, and sensor zenith angle. Some of these factors cause biases in cloud detection, for which it is important to account. For example, cloud detection uncertainties are consistently higher over polar regions because of cold temperatures and highly reflective surfaces like snow and sea ice. There are also small discrepancies in day/night cloud detection as not all classifiers are available at all times of day (there is no visible reflectance measurement at night and the 3.75-μm classifier is trained separately for night and day). These differences are most apparent near the terminator where the visible scattering effects of a setting sun are pronounced. Considerable effort has been put forth to minimize the day/night effect, and sensitivity studies of the PATMOS-x record suggest the effect is in most cases minimal. Other factors, such as sensor viewing angle, occur in a distributed and ubiquitous manner for polar orbiters so as not to cause a consistent bias (this is not the case for geostationary satellites, which we will address later). Solar viewing angle is linked directly to satellite orbital drift, meaning that cloud detection bias related to solar zenith angle is coupled to actual diurnal variation in cloudiness. The end result of this methodology is to correct for the satellite-observed diurnal cycle, which includes both the actual diurnal change and bias caused by solar zenith angle, though we hope to separate these in future work. A more detailed description of this detection algorithm can be found in Heidinger et al. (2012). The statistics generated on the probability of cloud detection relative to the active sensors on CALIPSO can also be used as an estimate of instantaneous uncertainty.
The cloud phase is derived from the cloud type algorithm, which provides seven distinct classifications: 1) clear, 2) fog, 3) liquid water cloud, 4) supercooled water cloud, 5) opaque ice, 6) cirrus, and 7) multilayer cirrus (Pavolonis et al. 2005). For this study classes 2–4 are considered water cloud and classes 5–7 are considered ice cloud. The initial cloud type classification is based solely on the 10.8-μm brightness temperature to account for the relationship between cloud-top temperature and phase. Additional tests using the near-IR AVHRR channels are applied thereafter.
Currently, NOAA operates three GOES satellites located at 135°, 75°, and 60°W. GOES-11, which forms the bulk of the GOES-West (135°W) data, provides IR channels at 3.9, 6.7, 11, and 12 μm. The GOES data used here have been processed using the suite of PATMOS-x algorithms described above. The most significant differences between PATMOS-x/AVHRR and PATMOS-x/GOES are the spatial and temporal resolution of the measurements: the GOES-11 imager has a spatial resolution of 4 km and temporal resolution of 30 min. In addition the nature of geostationary satellites means the sensor viewing angle is constant for each field of view, in contrast to a polar orbiter where scanning geometry can change from overpass to overpass. Functionally this means that sensor zenith angle dependent biases in cloud detection offset in polar orbiters but remain constant in geostationary satellites.
c. Correction for satellite drift
The description of this methodology is for a general case, meaning it could be applied to any observable cloud property that experiences a sinusoidal diurnal cycle. In this study, we use fractional coverage as the cloud property in question. To correct for satellite drift, we leverage the drifting local measurement time over the course of the entire AVHRR record. Figure 1 shows the local equatorial crossing time for the NOAA polar-orbiting satellite series along with the MetOp series. Since the beginning of the five-channel record in 1981, satellites have made measurements spanning almost the entire 24-h period of a day. We use this coverage to generate a fit function for each cloud property over the course of the day. We use a sinusoidal, two-harmonic, fit function of the form
where a, b, c, d, and e are empirically derived coefficients; F(x) is the cloud property being fitted; and x is the local overpass time for each satellite observation. To account for geographic and seasonal differences fits are derived for each month and 1° × 1° box globally. The fit function is used to calculate hourly values for each cloud property and a daily mean is derived integrating over all satellite overpasses using an equation of the form
where xobs and yobs represent the observed subset of local time and cloud property values for all satellite overpasses on a given day. The right-hand side of Eq. (2) can be described as the observed cloud property interpolated to a specific hour of day (represented by i) and then multiplied by a normalized weighting term designed to give greatest weight to those satellite observations closest to that hour. In this way the fit function derived from Eq. (1) is used to interpolate to each hour of the day while maintaining the primacy of actual observations when calculating the mean daily cloudiness. It was noted during review that a similar daily mean would be reached if, instead of interpolating to each hour of the day, we use Eq. (1) to construct a ratio of observed times to the diurnal mean and then take the average observed values adjusted by said ratio (we will refer to this as the unweighted method). Both methods work equally well in scenarios where the observed diurnal cycle matches the climatological average calculated in Eq. (1). For scenarios where the diurnal cycle does not match the climatological average, which is not uncommon because of synoptic-scale weather patterns, both methods are subject to error because of their dependence on a time-invariant diurnal cycle. However, it is unclear whether these methods differ in such a way as to cause systematic differences in diurnal corrections over time. Specifically, a concern was raised that our weighted method [Eq. (2)] would put undue emphasis on those times of day where observations were available, while the counterargument to this was that the weighting allowed the daily average to be based more on observations than a climatological mean. To ascertain how different these methods performed over time and whether one proved better suited than the other, we implemented them over Brazil for the entire record. The results of this can be seen in Fig. 2, where the weighted versus unweighted correction is plotted against one another. Each point represents a corrected monthly mean and includes water, ice, and total cloudiness. Over 2000 months are represented in Fig. 2. We found that there was no bias between the two methods and that the standard deviation was less than 1/10th of a percent cloudiness, suggesting the methods performance was virtually identical. We have kept the method using Eq. (2) because it provides additional information in the form of the estimated shape of the diurnal cycle for each individual day, but both methods work and we would like to thank the anonymous reviewer that sparked an enlightening discussion on the relative merits of each.
Figure 3 shows the maxima and relative amplitude of the sinusoidal fit functions globally for total cloudiness. Several physical processes identified in previous diurnal cycle studies (Cairns 1995; Dai and Trenberth 2004; Rozendaal et al. 1995; Slingo and Yang 2001) can be seen including the following:
Diurnal maxima over persistent stratocumulus deck regions occur around early morning, consistent with the morning to afternoon decrease in cloudiness corresponding with decreasing inversion bases and warming of the lower troposphere.
Larger continental amplitudes are driven by thermal inertia and convective cloudiness cycles.
The largest amplitudes in Northern Hemisphere generally occur during June, July, and August, and the largest amplitudes in Southern Hemisphere generally occur during December, January, and February.
The absence of distinct cycles is represented by small amplitudes in the Southern Ocean and parts of the intertropical convergence zone (ITCZ) and northern oceans consistent with persistent cloud bands.
a. Comparison against GOES-11
To validate the diurnal cycle we compare against data from a geostationary satellite. To this end a mean diurnal cycle is constructed by averaging the hourly fields of GOES-11 imagery for each month of 2009. Processing the GOES-11 data through PATMOS-x has the added benefit of reducing uncertainty due to algorithm differences. A χ2 goodness-of-fit test is applied to the observed GOES-11 cycle versus the expected AVHRR fit function. A value of χ2 < 35.2 corresponds to the 95% confidence interval (p = 0.05). Figure 4 shows the spatial distribution of χ2 in the North Pacific as well as a histogram. The histogram in Fig. 4 shows that about 90% of the χ2 values fall within this range. The spatial map in Fig. 4 shows increasing χ2 found around areas with strong diurnal cycles such as land and the ITCZ. Some of this difference may be attributed to noise inherent in a single month of GOES-11 observations. To investigate further, two subregions are examined over the entire year. For these regions, the mean observed hourly GOES-11 cloudiness is plotted against the AVHRR record interpolated to each hour using the sinusoidal fit function. Figure 5a shows the first subregion, located in a relatively stable portion of the central Pacific with little diurnal variation. Not unexpectedly, the sinusoidal fit function performs well under these conditions, and AVHRR and GOES-11 track closely for most months, resulting in a cloudiness bias (daily averaged GOES-11 minus AVHRR) of −1.8% with a standard deviation of 2.3%. One notable difference between the two can be seen near dawn, where cloud detection near the terminator becomes somewhat sporadic for GOES-11. The second subregion, which is shown in Fig. 5b, is over Mexico and has a strong convective diurnal cycle. This subregion shows a second harmonic in its cycle, and while the fit function is able to reproduce the angular frequency of this cycle we see variations in amplitude and phase. These differences still fall within range of the GOES-11 standard error, which is large over this subregion as compared to the more stable central Pacific subregion. In addition differences in phase and amplitude often offset when averaged over the entire day, resulting in a cloudiness bias of −1.4% with a standard deviation of 2.5%, similar to that of the central Pacific subregion. Over the entire North Pacific, we find a cloudiness bias of 1.3% with a standard deviation of 5.1%. It should be noted that constant viewing angles inherent to geostationary satellites cause shadowing effects that may affect cloud detection near the limb of sensor measurements, contributing to this bias.
b. Diurnal stability over time
Figure 5 suggests a mean diurnal cycle derived over the AVHRR record is comparable to that derived from a geostationary satellite, but differences in phase over land bring into question the stability of the diurnal cycle over time. Interannual variability in cloudiness is of no consequence, provided the shape of the diurnal cycle remains consistent, but relative shifts within the diurnal cycle (i.e., a change in peak cloudiness time or amplitude) could result in biases in the interpolation term in Eq. (2). To explore the stability of the diurnal cycle the difference between PATMOS-x/AVHRR 0230 and 1430 LT cloudiness is examined over the entire record. Changing difference over time suggests either the amplitude or phase of the diurnal cycle is changing as well. The 0230 and 1430 LT times were selected because the entire AVHRR record maintains overpasses very close to these times, reducing the possibility that sinusoidal fit function would greatly affect the results. Figure 6 plots the results of these calculations for the Northern and Southern Hemispheres over land and ocean. We find that the morning–afternoon difference over the oceans for both hemispheres remains stable over the course of the record for all seasons, remaining within 2% cloudiness. Over land there is more variability, in particular continental winter cloudiness varies by more than 2%. The Southern Hemisphere is slightly more variable than the Northern Hemisphere, but this may be due to sampling due to the relatively small amount of Southern Hemisphere land. This variability explains some of the differences in phase and amplitude seen between the continental PATMOS-x/GOES-11 and PATMOS-x/AVHRR diurnal cycles. None of the linear fits to the time series were found to constitute a significant trend.
c. Diurnal sampling uncertainty
One potential source of variance is the changing frequency of diurnal sampling over the course of the record. Referencing Fig. 1, the PATMOS-x/AVHRR is a single-satellite record (two overpasses daily) until approximately 1992. In contrast, there are as many as five satellites contributing to the PATMOS-x/AVHRR record in 2009. To estimate the effect diurnal sampling has on the variance of the cloud record, we again compare the PATMOS-x/AVHRR record to that of GOES-11 over the North Pacific for January 2009. For each day and for each 1° x 1° scene, we compare the difference between GOES-11 and various numbers of AVHRR sensors, resulting in comparison of approximately 28 000 individual scenes. The selection of AVHRR sensors is not random but rather based on optimizing the coverage of local equatorial crossing time. The four satellites used were NOAA-15, NOAA-18, NOAA-19, and MetOp-A. NOAA-19 was the first satellite removed as it local crossing time coincided closely with that of NOAA-18. Next removed was MetOp-A followed by NOAA-15, leaving a comparison of GOES-11 against NOAA-18. A randomization technique would likely have produced greater differences between GOES-11 and the AVHRR satellites, but the launching of satellites over the AVHRR record was planned to maintain morning and afternoon coverage so we chose to reflect that in this analysis. A small correction term is calculated for each scene to account for differing sensor footprints and viewing angles between the GOES-11 and AVHRR imagers. For each AVHRR overpass, the nearest GOES-11 image is located (within 15 min) and a difference of cloudiness is calculated between the two. The correction term is the average difference for all overpasses each day. Figure 7 shows the normalized distribution of GOES/AVHRR daily differences in cloudiness based on the number of satellite overpasses used for the calculation of daily cloudiness. The GOES-11 daily cloudiness is the average of 48 daily images of the scene, while the AVHRR daily average is the diurnally corrected average detailed in the methodology section. The distributions are roughly Gaussian in shape, with the curve flattening (increased variance) with decreasing number of AVHRR satellite overpasses used in the daily averaging. The standard deviation of the differences for one, two, three, and four satellites as a percent cloudiness is 14%, 8%, 6%, and 4%, respectively. The larger the standard deviation, the greater the variability in the time series. This means that there is greater uncertainty associated with the calculated daily cloudiness early in the AVHRR record when only a single satellite is available. The following section will describe the importance of this when determining uncertainty, a necessary step before looking to see if there are statistically significant trends in the satellite record.
d. Other sources of uncertainty
In the following section, we present climatological cloudiness time series based on the diurnal correction methodology. While the results of the time series are intriguing, to claim with confidence that the trends detected are statistically significant would require a full accounting of several sources of uncertainty. Specifically it would require the aggregation and attribution of synoptic- and larger-scale variability, algorithm uncertainty (including things such as view angle dependence, phase determination, and Bayesian detection probabilities), calibration uncertainty, and diurnal sampling uncertainty. While some of these are known with confidence and others we have touched upon here (i.e., diurnal sampling), the task of aggregating various, oft-correlated uncertainty values is not trivial. As such, we will restrict our conclusions here only to the magnitude of satellite drift bias.
e. Diurnally corrected tropical time series
Figure 8 shows an example of a PATMOS-x time series of water, ice, and total cloudiness over the tropics, separated by ocean and land, with the diurnal correction methodology applied. We have chosen the tropics as an example in part because of the distinct diurnal cycles found over the continents and marine stratocumulus regions and in part because the lack of highly reflective surfaces and cold temperatures found in polar regions reduces the uncertainty of the cloud detection algorithm. The error bars are the standard deviation of daily values calculated for each month. The variance is larger over land than ocean and over ice clouds rather than water clouds. This is due to the relative challenge of cloud detection over heterogeneous land with high albedo and spectral deficiencies in AVHRR, making cirrus detection difficult. In addition, jumps in maritime cloudiness after the El Chicón and Pinatubo volcanic eruptions are likely attributable to increased aerosol loading. While this may represent a real increase in cloudiness due to more available condensation nuclei (Song et al. 1996), at least part of the increase may be due to false detection. It is possible, however, that part of the shift from ice to liquid water cloudiness during these periods is due to stratospheric aerosol disrupting the radiometric contrast between the 10.8- and 12.0-μm channels, thereby affecting the cloud phase algorithm. Linear trends are calculated using the weighted least squares fit method, incorporating the error bars. Significant differences between the corrected and uncorrected time series are seen, especially with regard to trend detection. The uncorrected time series is the daily average of all available satellite overpasses. The corrected PATMOS-x shows almost flat total cloudiness over land with a positive slope of 0.15% ± 0.16% decade−1 cloudiness, while the uncorrected record shows a decrease in cloudiness at −0.28% ± 0.21% decade−1. Conversely, the corrected PATMOS-x record shows a significant increase of ice cloud over the ocean at a rate of 0.59% ± 0.14% decade−1 while the uncorrected record is an order of magnitude less, 0.03% ± 0.10% decade−1. Interestingly, the corrected and uncorrected PATMOS-x records over ocean and land all show sizeable decreases in tropical water clouds ranging between −0.76% and −1.09% decade−1. These results suggest that diurnal correction of the PATMOS-x record is an essential step if one’s goal is to detect long-term trends. This also begs the question of whether statistically significant cloudiness trend detection is possible given the limited length of the satellite record.
f. Trend detection
As stated in section 3d there are multiple sources of uncertainty that must be considered when attempting to detect a statistically significant trend in a climatological record, and this is currently being addressed in a separate study. However, assuming for a moment the trends found in Fig. 8 accurately represent the aggregate sources of uncertainty, we feel it worth noting how long the PATMOS-x record must be for the trends to be considered statistically significant. With this is mind a statistical approach developed in previous studies (Weatherhead et al. 2002; Weatherhead et al. 1998) has been applied to the tropical (30°S–30°N) time series of water, ice, and total cloudiness. For an estimated trend ω, in this case calculated using the weighted least squares fit method and shown in the bottom-left corner of each panel in Fig. 8, we can assume that a real trend at the 95% confidence level is established when |ω/σω| > 2, where σω is the standard deviation of ω. A method of estimating σω can be found in Weatherhead et al. (1998) with an exact derivation found in the appendices. An estimation of the number of years (n*) that it would take to detect a real trend of magnitude |ω| with probability of 0.90 is
where N is the noise of the time series (that portion not explained by ω or seasonal signals), calculated by removing the monthly mean and slope of the linear fit from the time series; ϕ is the autocorrelation of N; and ɛ is the white noise, meaning that portion of N not explained by ϕ. In this equation, it is assumed that ϕ is less than 1. This equation shows the estimate of n* is contingent on the autocorrelation of the noise, the magnitude of the estimated trend and the unexplained variance in the dataset. The values of n* are shown in the top-right corner of each panel in Fig. 8. Values of n* are less than the 30-yr PATMOS-x record for the ice cloud over ocean and land and water cloud over ocean, slightly higher than 30 yr for water cloud over land and total cloud over ocean and significantly higher for total cloudiness over land. This suggests that (i) the transition between liquid and ice phase clouds is likely to be at least if not more important than total cloudiness; (ii) 30 yr appears to be the minimum length required to detect small trends in cloudiness; and (iii) increased variance due to a dearth of diurnal sampling, particularly in the early years of the AVHRR record, should increase the number of years needed to detect significant trends.
Producing satellite cloud records of sufficient quality and consistency to be useful in detecting trends on climate time scales continues to be a challenge. Here, we address one of the primary obstacles of the PATMOS-x/AVHRR record: a slow shift in equatorial crossing time as the numerous contributing polar-orbiting satellites drift over their lifetimes. A correction to the record is applied by fitting a mean diurnal cycle to a sinusoidal, two-harmonic fit function derived over the five-channel AVHRR record (1981–2010). An average daily value of cloudiness is interpolated from the fit function using all available ascending, descending, morning, and afternoon satellite overpasses. The diurnal cycle is validated over the North Pacific using one year of GOES-11 measurements, and good agreement is found between the two. A test to verify the stability of the diurnal cycle over time is applied, and it is found the assumption of diurnal stability over time is reasonable, though continental cloudiness experiences greater interannual variability than maritime cloudiness does and this could affect the goodness of fit of the mean diurnal cycle. Comparison of the corrected and uncorrected time series reveals that the magnitude and even sign of cloudiness slope over time changes, suggesting a diurnal correction is a necessary step to perform for studies interested in detecting trends over climate time scales. A statistical method is applied to estimate the necessary length of time needed to detect a significant trend given the slopes estimated by the least squares fit method. The results suggest that 30 yr is the minimum length of time required, but this is contingent on the magnitude of the trend and the autocorrelation and unexplained variability of the data such as that caused by inconsistent diurnal sampling over the lifetime of the AVHRR record. This implies we are only now generating satellite records of sufficient length to begin making meaningful statements about climate trends. The diurnal correction method introduced here is general in its scope, meaning it may be applied to other cloud properties with measureable diurnal cycles such as optical depth and effective radius, though the shape of the fit would need to account for a daytime-only cycle, as the retrieval of those properties rely on the availability of visible channels. Future work includes further validation of these time series against independent data sources as well as quantification of error sources such as diurnal sampling, instantaneous measurement, and monthly variability. Another future goal is the generation of a suite of diurnally corrected level-3 cloud products that together may form a clearer picture of those cloud feedback mechanisms driving the earth’s energy budget.
We thank the NOAA National Climatic Data Center (NCDC) and their Climate Data Record (CDR) program for financial support of this project. The views, opinions, and findings contained in this report are those of the author(s) and should not be construed as an official National Oceanic and Atmospheric Administration or U.S. government position, policy, or decision.