As we celebrate the fiftieth anniversary of NASA’s Apollo missions, images of Earth simulated with the ECMWF Integrated Forecasting System (IFS) are visually compared with pictures collected during space missions of the past five decades, in particular from the Apollo missions (1968–72). The numerical weather reforecasts use the latest version of the IFS and are initialized from (re)analysis data, which provide our current best representation of the atmospheric state for any given date back to the 1950s. Visible images of our planet are produced from the IFS with a simple simulator whose main inputs are the solar fluxes at the top of the atmosphere. First, a comparison to recent imagery from deep space illustrates the high level of performance of the IFS on recent dates. Then, the validation of the IFS against photographs taken by Apollo 11 and 17 both in-flight and from the lunar surface exhibits a significant level of agreement, despite the absence or very limited number of satellite observations available. This short study confirms that the combination of high-quality initial conditions with a modern numerical weather prediction model can yield reasonably accurate reforecasts of global meteorological conditions, especially cloud systems, for dates as far back as the late 1960s.
As part of the celebrations of the fiftieth anniversary of the National Aeronautics and Space Administration’s (NASA) Apollo missions (1968–72), several iconic photographs of Earth taken from the moon back then have been recently shown in the media around the world. Besides their outstanding natural beauty, these pictures can also be considered as pioneering global observations of our planet’s atmosphere, at a time when a handful of meteorological satellites were taking their first faltering steps. Over the past decades, huge progress has occurred on the observation side, from the images provided by the first geostationary satellite ATS-1 in 1966 (Suomi and Parent 1968) to today’s high-spatial- and high-temporal-resolution images available from the latest generation of geostationary satellites (Schmit et al. 2017; Bessho et al. 2016; Schmetz et al. 2002). At the same time and even more amazingly, our ability to predict the many components of Earth’s atmosphere and their interactions has undergone tremendous progress since the 1970s. For example, the European Centre for Medium-Range Weather Forecasts (ECMWF) recently celebrated the fortieth anniversary of its first operational forecast (August 1979) as well as the dramatic improvements in numerical weather prediction (NWP) achieved ever since, through the “quiet revolution” described in Bauer et al. (2015). It therefore seems natural to wonder how a modern NWP system would perform on those bygone times, especially when it comes to simulating cloud patterns on the global scale. In an attempt to shed light on that question, this article proposes a purely visual assessment of weather reforecasts using the independent source of validation provided by pictures taken by space missions spanning the past fifty years, including Apollo 11 and 17. Obtaining realistic weather predictions requires not only a good forecast model but also accurate initial conditions [i.e., (re)analyses]. For recent dates, operational analyses can be directly used as a starting point for weather forecasts. Indeed, these analyses are expected to deliver our current best description of the three-dimensional atmospheric state because they are obtained using the latest version of the forecasting system to assimilate a large number of in situ and satellite observations (around 30 million every day at present). In practice, analyses are produced by running a data assimilation system that merges in a statistically optimal way information from as many observations as possible with information from the model itself. At ECMWF for instance, this is currently done using the four-dimensional variational data assimilation (4D-Var; Courtier et al. 1994) method. For older dates in the NWP era, the quality of the operational analyses produced at the time is likely to be significantly lower than what might be achieved with a more modern forecasting system, due to 1) their coarse resolution, 2) inferior forecasting system, and 3) limited number of assimilated observations (in no particular order). This explains why successive reanalysis projects have been conducted over recent decades, with the goal of generating new atmospheric analyses that would benefit from increased resolution and recent upgrades of the forecasting system, with the ability to assimilate more observations, especially from newly recovered or reprocessed datasets. Recent examples of such reanalysis datasets, which can span the past fifty years or more, include ECMWF’s ERA51 (Hersbach et al. 2020) and ERA-Interim (Dee et al. 2011), MERRA-2 (Gelaro et al. 2017) from the NASA Global Modeling and Assimilation Office (GMAO), and JRA-55 (Kobayashi et al. 2015) produced by the Japan Meteorological Agency (JMA). However, it is worth stressing that despite the constant improvements in the performance of NWP systems, the quality of the produced reanalyses will inevitably suffer from the gradual decrease in the availability of satellite observations as one travels back to the 1960s, especially over oceans.
In this short study, global numerical weather reforecasts were run with ECMWF’s Integrated Forecasting System (IFS) initialized from ECMWF’s best (re)analyses for each selected date. Running reforecasts was necessary not only to ensure the timely comparison of model with the selected photographs, but also to benefit from the latest version of the forecast model as well as from the highest resolution currently affordable.
After a brief description of the data and methodology, various examples of the comparison of increasingly older pictures of Earth from space with corresponding simulated versions from the IFS will be presented and discussed.
Data and methodology
The numerical weather reforecasts shown in this article were run using the latest operational version of the IFS (version 46r1; see www.ecmwf.int/en/publications/ifs-documentation) at the current global resolution of 9 km (unless stated otherwise) and with 137 vertical levels to obtain the most detailed representation of clouds possible. For recent years (post-2015), ECMWF’s operational analyses (9-km resolution) were deemed to provide optimal initial conditions for the reforecasts. For earlier years, global reanalyses either from ERA5 for dates after 1972 or from the older ERA-40 (Uppala et al. 2005) for earlier dates were used for forecast initialization. For a given date in either period, the corresponding reanalysis is expected to provide the best description of global atmospheric conditions we can hope for. It is worth noting that ERA5 (ERA-40) reanalyses were produced at a horizontal resolution of 31 km (125 km), on 137 (60) levels in the vertical and using the IFS version from the year 2016 (2001). Besides, let us keep in mind that more observations have usually been assimilated in ERA5 than in ERA-40 for any date common to both reanalysis datasets.
A program was also specially developed to permit the computation of simulated “visible” images of Earth that could then be compared with images from space missions. This homemade visible image simulator primarily relies on hourly averaged top-of-the-atmosphere (TOA) solar fluxes (default outputs from the IFS), which makes it computationally inexpensive and easy to run offline. High values of the TOA reflected solar flux are typically associated with clouds or snow-covered regions, which will appear in bright white on simulated images. Over clear-sky and snow-free regions, the lower the surface albedo, the darker the appearance. A detailed description of the simulator and its limitations is given in appendix A. More elaborate visible image simulators do exist, which require the specification of optical properties of the surface, clouds and possibly aerosols as well as three-dimensional input data for various atmospheric variables. Examples of such simulators would include Radiative Transfer for TOVS (RTTOV) (Saunders et al. 2018) either using the (slow) discrete ordinates method (Chandrasekhar 1960; Stamnes et al. 1988) or the faster lookup table–based Method for Fast Satellite Image Synthesis (MFASIS) software (Scheck et al. 2016); Oxford–Rutherford Appleton Laboratory (RAL) Optimal Retrieval of Aerosol and Cloud (ORAC) estimation scheme (McGarragh et al. 2018), 6SV2 radiative transfer code (Vermote et al. 1997; Kotchenova and Vermote 2007). However, given their level of complexity and input requirements and hence their higher computational cost, this type of simulator was not considered in the present study whose main purpose was to perform a purely visual comparison of simulated and observed images.
On the observation side, appendix B provides information about the provenance of the NASA images. It is worth stressing that one of the main challenges in this work was to obtain enough reliable information about the exact time and viewing geometry of each picture. The latter would include the distance of the spacecraft to Earth, the geographical coordinates of the subspacecraft point and the tilt angle of our planet in the viewing plane. When such information was not available, a trial-and-error approach was used to match the field of views of the simulated and observed images. In this respect, one advantage of visible over infrared imagery is that the location of the terminator (i.e., the night/day limit) combined with the position of continents can help to confirm the time of an image. Finally, one should note that for recent dates, images from deep-space missions were selected because they provide an unusual and totally independent source of validation for the model as well as a remote view of our planet that is comparable to that of the Apollo photographs taken from the moon.
Simulations from DSCOVR to Apollo
The current level of performance of operational global forecasts will be illustrated in Fig. 1 by considering a view of our planet captured by NASA’s Deep Space Climate Observatory (DSCOVR) Earth Polychromatic Imaging Camera (EPIC) from a distance of 1.5 million km. One advantage of this satellite is that it can provide a continuous viewing of Earth’s sunlit disk with a resolution of around 10 km. The selected DSCOVR/EPIC image shown in Fig. 1a is valid at 2111 UTC 2 September 2015 and is centered on the Pacific Ocean, with the American continent visible on the right-hand side. Corresponding simulated visible images obtained from IFS reforecasts at 21- and 93-h ranges and initialized from ECMWF’s operational analyses are displayed in Figs. 1b and 1c.
The comparison of Figs. 1a and 1b underlines the ability of the 21-h-range reforecast to produce cloud systems that visually agree quite well with the observations. This is particularly true in the midlatitudes of both hemispheres, where clouds associated with frontal regions and cold air outbreaks are well represented in the model, in terms of both location and shape. Even in the tropics where forecast quality is often lower, many cloud patterns appear to be rather realistic, such as the three tropical cyclones cruising over the Pacific. Even more remarkably, the 93-h-range reforecast in Fig. 1c yields a simulated image that looks very similar to its 21-h-long counterpart. This means that the ability of the model to predict clouds in the present case remains very good up to four days ahead (at least). The main differences between Figs. 1b and 1c are found in tropical convective regions, as expected.
Of course, some discrepancies between model and observations can be identified, which could be due either to the simplicity of the visible image simulator or to genuine model imperfections. For instance, low-level clouds in cold air outbreaks and in anticyclones tend to appear brighter in the simulated images, especially over the South Pacific Ocean. Also, the frontal cloud system stretching across the southern mid-Pacific looks slightly broader in the model (Figs. 1b,c) than in the DSCOVR/EPIC image (Fig. 1a), probably as a result of the hourly averaging of solar fluxes in the IFS. Finally, the proper shape and brightness of convective clouds in equatorial regions is harder to predict (especially inside the intertropical convergence zone), but this is a well-known deficiency common to all global forecast models (Dias et al. 2018).
Despite those differences, Fig. 1 illustrates the ability of the latest version of the IFS to produce realistic horizontal distributions of clouds for the selected recent date, especially in the extratropics.
We will now compare the model to a series of four images taken 29 years ago by NASA’s Galileo spacecraft during its first Earth flyby on 11 and 12 December 1990 (Fig. 2a). The distance of the space probe to our planet ranged from 2 to 2.7 million km (a jetliner flying nonstop would cover that distance in about four months!). The corresponding four simulated visible images obtained from a reforecast initialized from the ERA5 reanalysis at 0000 UTC 11 December 1990 are displayed in Fig. 2b. From top to bottom, the four views are centered over South America, the Pacific Ocean, the Indian Ocean/Australia, and Africa, respectively. Forecast ranges are 14, 20, 27, and 32 h, respectively.
Despite the older date considered here, the visual agreement between the simulated and Galileo images is again quite striking. Most of the complex cloud structures found in storms over the Southern Ocean on the space probe’s images are properly depicted in the reforecast. Even in the tropics, the main cloud systems over South America (top row) and Africa (bottom row) are rather accurately represented. This resemblance reflects not only the good performance of the forecast model but also the quality of the ERA5 global reanalysis for the year 1990, with a reasonably large number of assimilated spaceborne observations (mainly from NOAA polar-orbiting satellites).
This is as far away from Earth as this paper will take us. However, it is worth noting that a successful qualitative validation of the model could still be performed on a 2008 image taken by NASA’s Deep Impact mission from a distance of 50 million km, despite the reduced resolution and blurring of the image (not shown here for these reasons).
Apollo 17 (1972). “Blue Marble.”
At 1039 UTC 7 December 1972, the crew of NASA’s Apollo 17 mission captured the iconic “Blue Marble” image (Fig. 3a) while the spacecraft was coasting on its way to the moon, 29,000 km away from Earth. This picture offers a rather unusual and beautiful view of our planet focusing on Africa and Antarctica.
Figure 3b shows the corresponding simulated visible image from an 11-h reforecast started from ERA5 reanalysis data valid at 0000 UTC 7 December 1972. The comparison of Figs. 3a and 3b indicates that most cloud systems south of 30°S are represented fairly realistically. However, a few obvious mismatches can be identified in the extratropics: a mesoscale cloud structure is present just to the east of South Africa in Fig. 3b but not in Fig. 3a; similarly, a well-defined relatively small-scale vortex is found in the Southern Ocean in the model but not on NASA’s image. In the tropics, the level of agreement between model and observation is usually lower, with for instance a clear underestimation of cloud occurrences (or at least of their albedo) by the model over the South Atlantic Ocean. Over central Africa, convective activity is present in both model and observation, but it is not necessarily simulated at the right locations. However, the tropical cyclone observed over India is well captured in the simulation (top-right corner of Fig. 3).
To go beyond the previous 11-h forecast range, Fig. 3c displays a 59-h reforecast initialized from ERA5 and valid at the time of the “Blue Marble” picture. Cloud patterns in the southern midlatitudes as well as the Indian tropical cyclone still agree reasonably well with the Apollo 17 photograph (Fig. 3a); however, compared to the 11-h-range simulation (Fig. 3b), further discrepancies start to appear in the location and shape of individual fronts. Beyond the 3-day range (not shown), the quality of the forecast degrades quite substantially. This behavior strongly contrasts with the remarkable similarity between the 21- and 93-h forecasts found in Figs. 1b and 1c for the recent case of DSCOVR; this is another illustration of the crucial importance of the quality of initial conditions for the performance of medium-range forecasts.
Compared with the 1990 Galileo case [see “Galileo (1990)” section], the somewhat degraded performance of the reforecasts for the 47-yr Apollo 17 case can be explained by the fact that a single polar-orbiting meteorological satellite [NOAA-2 Vertical Temperature Profile Radiometer 2 (VTPR-2)] was assimilated in the ERA5 reanalysis. On the other hand, Fig. 3d, which displays an 11-h reforecast started from the older ERA-40 reanalysis (no satellite data assimilated), exhibits a worse match to the observations (Fig. 3a) than Fig. 3b. This clearly highlights the superiority of ERA5 over ERA-40, due to the larger number of observations assimilated (especially the extra VTPR-2 satellite) but also thanks to the better modeling/data assimilation system and higher horizontal and vertical resolutions of ERA5.
One noteworthy artifact in the simulated images in Figs. 3b–d is the presence of large-scale “cloud streets” over tropical oceans. These patterns originate from the coarser resolution of the reanalyses used to initialize the 9-km reforecasts and they tend to persist in the forecast over oceans wherever calm weather conditions prevail (i.e., in the absence of strong small-scale physical forcings). On the other hand, over land and meteorologically active ocean regions, smaller-scale field patterns can develop right from the beginning of the forecast.
Among the many photographs taken from the lunar surface by the Apollo 17 astronauts, a few of them featured Earth in their background. As an example, Fig. 4a shows a lunar landscape with a large rock in the foreground and the tilted quarter of our planet in the dark sky above it. This picture was taken at 0216 UTC 13 December 1972 and the distance to Earth is around 406,000 km. Despite the small size of Earth’s disk, a comparison with a 27-h reforecast initialized from the ERA5 reanalysis was attempted. Figures 4b and 4c display a magnified view of our planet from NASA’s original image and the corresponding simulation, respectively. The field of view is centered over the Pacific Ocean, with the Arctic on the top right and Antarctica on the bottom left of the sunlit disk. Australia is just emerging on the top left.
Despite its relatively blurry aspect, most cloud features seen on the Apollo 17 photograph seem to have an equivalent in the simulated image, both in the midlatitudes and in the tropics. The fact that some discrepancies can be found in the detail of these cloud structures is not surprising, given the lack of satellite observations used in the 1972 reanalysis. However, from a different perspective, this particular case is interesting because it highlights the fact that even a tiny image of Earth extracted from a 47-yr old photograph can still provide a reasonable source of qualitative evaluation for numerical weather prediction.
Apollo 11 (1969).
Just a few hours before Neil Armstrong set foot on the moon, an iconic picture showing Earth rising above the lunar surface was taken from the Apollo 11 command module in orbit, at around 0500 UTC 20 July 1969 (Fig. 5a). On this stunning photograph, our planet’s polar axis appears almost horizontal and the sunlit area covers the west Pacific, with Australia visible on the top left and the north pole on the right. Figure 5b displays the corresponding visible image simulated from a 29-h 29-km-resolution reforecast initialized from the older ERA-40 reanalysis.
Even though no satellite observations were assimilated in ERA-40 for that day, it is remarkable that many cloud systems seen in Fig. 5a have an equivalent in the simulated image in Fig. 5b. This is particularly true in the extratropics, with for instance the reasonable representation of frontal clouds over the North Pacific and northern Asia (right-hand side of Earth’s disk) and of low-level and midlevel clouds associated with the large anticyclonic vortex centered over New Zealand (bottom left). Over the tropical Pacific (central region of the disk), the level of agreement is not as good. This is easily understandable as a result of 1) the lack of observations over oceans in the ERA-40 reanalyses for 1969 and 2) the long-standing challenge of accurately forecasting convective activity, particularly inside the intertropical convergence zone.
State-of-the-art numerical weather prediction systems such as the IFS can deliver remarkably accurate cloud forecasts on the global scale for recent dates. Even more amazingly, rather realistic cloud simulations can also be obtained for much older cases, provided the same modern forecasting system is initialized using high-quality reanalyses, such as ERA5. A significant contributor to the improvement of atmospheric and surface (re)analyses (and subsequent medium-range forecasts) is our ability to extract information from satellite observations, whose developments owe so much to the space race of the 1950s and 1960s. In other respects, the comparison of simulated and Apollo images illustrates that old photographs from space missions can be used as an unconventional source of validation, despite their remoteness in both space and time. Actually, in the pre-1970 era, such photographs of our planet would constitute our only available source of validation for global cloud predictions, especially over oceans and uninhabited regions.
A possible follow-up to this study could include rerunning the Apollo 11 case using ERA5 reanalysis data, once the latter become available for 1969. It might also be interesting to investigate the potential benefits of using a more complex visible image simulator.
In summary, besides celebrating the huge technological progress and achievements in space exploration since World War II, this short study has also paid tribute to the giant leap in our ability not only to predict the weather into the future, but also in the context of reforecasts of past weather for dates as early as the late 1960s. This rather amazing capability has been made possible through continuous improvements in numerical weather modeling, data assimilation, high-performance computing, but also in the handling (and recovery) of myriad in situ and satellite observations.
I would particularly like to thank my ECMWF colleagues Hans Hersbach, Adrian Simmons, Irina Sandu, Robin Hogan, Richard Forbes, Mark Fielding, Margarita Choulga, Simon Lang, and Alan Geer for their technical help or suggestions during this short study. The three anonymous reviewers as well as BAMS Editor Timothy Schmit should be acknowledged for their comments on how to improve the original manuscript. I am also very grateful to NASA for granting access to their images from the DSCOVR, Galileo, and Apollo missions.
Appendix A: The Visible Image Simulator
The simple visible image simulator developed during this study relies on two main input fields produced by the IFS: 1) the incident solar flux and 2) the net solar flux , both valid at the TOA; is defined as the difference between and the TOA reflected solar flux . By default, and are both archived from the IFS as hourly averages.
The brightness B of each pixel in a simulated image is expressed as
where denotes the maximum value of the TOA reflected flux over the entire image. Exponent γ achieves the gamma correction, which accounts for the nonlinear perception of light by a given camera or imager (set between 0.25 and 0.45 in this study).
As described below, the specification of the color hue for each type of clear-sky scene (land, ocean, vegetation, or ice) is based on the solar albedo α calculated as
A few ancillary fields from the IFS are needed to increase the realism of simulated images: land–sea mask, total cloud cover, snow depth, vegetation cover, and leaf area index (LAI) for both high and low vegetation. The model land–sea mask is used to distinguish between land and ocean areas. Maximum threshold values of 0.24 and 0.10 for the TOA solar albedo are used to identify cloud-free land and ocean, respectively. In addition, cloud-free high-albedo desert surfaces are identified wherever total cloud cover is lower than 0.01 and snow depth is equal to zero. Significantly vegetated regions are identified wherever either the high-vegetation fractional cover is above 0.8 and the high-vegetation LAI is greater than 2 or the low-vegetation fractional cover is above 0.75 and the low-vegetation LAI is greater than 1.5. Different red–green–blue (RGB) color palettes are defined for clear-sky scenes, that is when α is lower than a certain threshold α1. Their respective dominant hue depends on the type of surface: ocher for bare soil and deserts (α1 = 0.24), blue for oceans (α1 = 0.10), and green for vegetation (α1 = 0.14). On the other hand, setting the three RGB color values to 1 for α above 0.35 means that clouds and clear-sky ice-covered surfaces are displayed using a grayscale whose brightness is determined from Eq. (A1). Figure A1 summarizes the shape of the RGB curves for (Fig. A1a) standard scenes and (Fig. A1b) the special case of clear-sky desert areas, as a function of the solar albedo α. In practice, the definition of the color palettes should be adjusted to better match the hues of the observed image, which are camera/imager dependent.
Since the image simulator was designed to work on a few existing 2D output fields from the IFS, it suffers from various imperfections. The hourly averaging applied to the solar fluxes simulated by the IFS may introduce small discrepancies when comparing to the snapshot-style images available from space. For instance, this difference may contribute to a slight broadening of narrow cloudy frontal regions or explain small misplacements of some cloud features in the simulated images. Furthermore, solar fluxes in the IFS are computed over the spectral window ranging from 0.2 to 12.0 µm, which may not exactly match the spectral response function of the camera or satellite instrument. However, even though the IFS spectral window extends beyond the limits of the visible spectrum (0.38 to 0.74 µm), the contribution from visible wavelengths dominates in the solar fluxes computed in the IFS. Hence the term “visible” used to describe the image simulator throughout this paper. An additional simplification in the simulator comes from the assumption that the upwelling shortwave radiation is isotropic. In reality, reflectance can vary with scattering angles, depending on the physical properties of scattering particles, such as effective radius and shape (Scheck et al. 2016). Another possible source of inaccuracy lies in the underestimation of the Rayleigh scattering of radiation in the direction of the camera/imager and originating from points located near the limb or near the terminator (i.e., in case of slantwise propagation of radiation over longer distances through the atmosphere). The 3D effects of cloud shape and shadows on radiation are also neglected. Finally, the possible effect of sun glint (reflection from water surfaces) is not accounted for.
However, one should emphasize that these simplifications do not prevent the successful visual comparison of the simulated images with observed ones, as illustrated throughout this article.
Appendix B: Information about the NASA Images
The source of the NASA images shown in this article can be found in Table B1. Note that some of these images were cropped prior to publication to optimize display size.
The latest reanalysis produced by ECMWF for the Copernicus Climate Change Service.