Worldwide validation of an Earth Polychromatic Imaging Camera (EPIC) derived radiation product and comparison with recent reanalyses

Highlights

• An irradiance dataset derived from the EPIC onboard NASA’s Deep Space Climate Observatory is validated.

• This EPIC-derived dataset is produced using machine learning as opposed to the traditional radiative transfer.

• The quality of this dataset is however largely unsatisfactory as compared to traditional datasets.

• The data contains undesirable artefacts, which could be due to deficiencies during training.

• This EPIC-derived dataset is not recommended in its present form for solar resource assessment purposes.

Abstract

A very recent gridded product for the hourly global horizontal irradiance (GHI), derived from the measurements of the Earth Polychromatic Imaging Camera (EPIC) onboard the Deep Space Climate Observatory (DSCOVR) launched by a NOAA/NASA/USAF consortium, is validated at 31 locations worldwide, from January, 2017 to June, 2019. In contrast to those traditional methods that leverage (simplified) radiative transfer, this EPIC-derived product uses machine learning – a random forest model – to map out the connection between satellite-observed variables of various kinds and GHI. Nonetheless, the detailed validation conducted here shows that the quality of this EPIC-derived GHI dataset not only does not outperform those traditional gridded solar radiation datasets, but also contains undesirable artifacts that can be possibly attributed to inadequacies in the machine-learning procedure. For these reasons, it is not recommended to use this EPIC-derived dataset in its current form for solar resource assessment purposes.

Introduction

Gridded modeled surface irradiance from satellite-derived and reanalysis radiation products has become an integral part of modern solar resource assessment and forecasting, among many other applications. Remote-sensed irradiance is derived primarily from data collected by instruments onboard geostationary satellites or polar orbiters. The former continuously observe a large stretch of the earth-atmosphere system from different positions along the equator, but cannot provide useful cloud information at high latitudes. Examples of modeled solar irradiance products derived from such observations include the GOES-based National Solar Radiation Database over the Americas (Sengupta et al., 2018), an irradiance database derived from INSAT-3D over the Indian peninsula (Kamath and Srinivasan, 2020), the Meteosat-based Copernicus Atmosphere Monitoring Service (CAMS) radiation service (based on the Heliosat −4 method) over Europe and Africa (Qu et al., 2017), and the Meteosat-based SARAH database over the same region (Müller et al., 2015). Conversely, polar orbiters observe only a small swath along their orbits but can observe the polar regions. In that category, the main satellite sensors are AVHRR, MODIS, and CERES, and examples of surface irradiance databases developed from their observations include CLARA (Karlsson et al., 2017), CERES-SYN1deg (Rutan et al., 2015, Kato et al., 2018), and a variety of regional or global MODIS-derived products, e.g. Huang et al., 2018, Ryu et al., 2018 and Wang et al. (2020). In recent years, the advent of the deep-space satellite era has been marked by the launch of the DSCOVR mission in 2015, whose Earth Polychromatic Imaging Camera (EPIC) instrument provides observations of various quantities from its locked position at the L1 Lagrange point (L${}_1$ point). A Lagrange point is the equilibrium point of a low-mass object under the influence of two massive orbiting objects, which makes it an excellent location for a satellite. Among all such points, the L${}_1$ point is the closest from Earth and allows the DSCOVR Earth Polychromatic Imaging Camera (EPIC) to view the sunlit Earth’s hemisphere almost entirely. The EPIC observations have started to be used to derive surface irradiance estimates at the global scale as the earth rotates in EPIC’s field of view, Su et al., 2018, Hao et al., 2019a and Hao et al. (2020).

Particularly noticeable is the product described by Hao et al., 2019a, Hao et al., 2020, which offers GHI (also known as downward surface shortwave radiation in other fields) and PAR at the Earth surface, obtained from EPIC observations via a random forest model. Located at $\approx 1.5$ million km from Earth, EPIC benefits from a quasi-complete view of the sunlit side of Earth from sunrise to sunset, and makes observations with different narrow-band filters (Burt and Smith, 2012). These observations provide information about ozone, SO${}_2$, aerosols, clouds, and other quantities at high temporal resolution (Marshak et al., 2018). Considering its wider viewing angle than geostationary weather satellites, EPIC is expected to overcome some of the limitations of the current remote-sensing-based GHI estimations, such as those due to effects of orbital geometry. Additionally, given the fact that the spatial resolutions of some of the synoptic-scale satellite-derived GHI products discussed above – or, even more so, of other products generated by general circulation models (i.e., reanalysis) – are often coarse, EPIC-derived high-resolution surface irradiance products have the potential of becoming an important alternative for demanding applications, such as solar resource assessments, for which high spatiotemporal resolution is critical. Such applications, however, also require high accuracy, which raises the need for validation and a detailed performance assessment at global scale. Currently, two different modeling approaches, described in Su et al. (2018) and Hao et al. (2020) have been proposed to derive surface irradiance from EPIC observations. Only the latter results are publicly available at https://data.pnnl.gov/project/13053, however, so that what follows exclusively relates to that database.

In a recent investigation (Yang and Bright, 2020), a total of eight gridded hourly GHI products, which included six satellite-derived datasets and two global reanalyses, were validated against measurements from 57 BSRN stations (Ohmura et al., 1998, Driemel et al., 2018), over a period of 27 years (1992–2018). That study appears to offer the most comprehensive validation of gridded irradiance products to date. One of its key findings was that products based on satellite observations outperformed the reanalysis estimates, thus confirming previous results of the literature, e.g., Urraca et al., 2018a, Urraca et al., 2018b, Salazar et al., 2020 and Bojanowski et al. (2014). One important advantage of reanalyses, however, is their global nature, which makes them able to produce surface irradiance results at high latitudes, thus circumventing the limitation of databases derived from geosynchronous satellite observations. This feature is highly desirable in the present context of a worldwide dissemination of solar technologies, and justifies specific validation studies under such conditions (Babar et al., 2019). (Although irradiance retrievals based on polar orbiter observations do not have that latitude limitation, they are typically at much coarser spatiotemporal resolution than their geosynchronous counterparts, and are thus seldom used in solar resource assessments, in particular.) Considering the solar industry’s strong demand for reliable irradiance databases, it is obvious that a need exists for the determination of which database(s) might have the lowest bias and highest accuracy over any given area. This demand has also attracted commercial data providers, whose products appear to compete favorably with their public-domain counterparts (Bright, 2019, Yang and Bright, 2020). Consequently, to provide fair and informative comparisons, the overall validity of any new product must be established by comparing its quality and accuracy to that of earlier products, under the same framework with which those earlier products were validated. Under this principle, the present study extends the validation framework used by Yang and Bright (2020), and applies it to the specific EPIC-derived product developed by Hao et al., 2019a, Hao et al., 2020. Stated differently, the data pre-processing method herein used, i.e., the preparation of ground-truth and other competitor products, is identical to that presented in Yang and Bright (2020), which makes the current results immediately comparable. For convenience, this new EPIC product derived with the random forest algorithm of Hao et al., 2019a, Hao et al., 2020 is referred to as EPIC-RF hereafter, for the reason that these authors did not assign a name to their product, and to distinguish it from other existing or future EPIC-derived products, such as the one from Su et al. (2018).

The EPIC-RF tested here has a spatial resolution of $0.1°\times 0.1°$, which is finer than that of all existing reanalyses, but currently extends over only a four-year period from June 2015 to June 2019, which is considerably less than the many decades covered by reanalyses or some satellite-derived products. Considering that the data from June 2015 to December 2016 were used to train the machine-learning models (Hao et al., 2019a), one ought to only validate data from January 2017 to June 2019. This period can be regarded as too short to be of general interest in application demanding long time series, such as solar resource assessments. Nevertheless, if this new database proved to be highly reliable, it would have the potential of becoming a benchmark against which all other databases would be evaluated. Moreover, considering that EPIC-derived surface irradiance products can provide global coverage, just like reanalyses, it is worth comparing the two approaches on an equal footing. As mentioned above, reanalysis-based irradiance products have not been found as accurate as most satellite-derived products. Hence, any new global satellite-derived product, such as the EPIC-RF under scrutiny here, would need to at least outperform existing reanalyses to be of potential interest. This justifies why the present study evaluates the worldwide performance of the EPIC-RF GHI product directly against the two reanalyses that were evaluated by Yang and Bright (2020): MERRA-2 and ERA5.