In southern UK, many growers are involved in outdoor vegetable production, however soft fruit production is most commonly polytunnel-based. The confined space created by tunnels defines phytoclimates that can be optimized for instance to extend the length of the growing season through temperature and venting regulation and by enabling a precision irrigation approach.

Tunnels consequently offer a powerful control of the plant's growing environment.

In the case of soft fruits, their rate of development is both temperature and radiation dependent. The spectral part of the solar radiation responsible for the growth of plants, algae, and of certain microorganisms is called the Photosynthetically Active Radiation (PAR), and typically lies in the wavelength range [400, 700] nm. During early summer 2018, May–July, a long period of hotter and sunnier than average growing conditions led to a very focused peak in production, resulting in a significant reduction in price and a marked increase in wastage. The management of crops through the ventilation or the shadowing of tunnels offers some potential to regulate the production schedule and therefore to enhance economic performance. An accurate knowledge of climatological and real-time PAR is valuable here. The relative scarcity of PAR measurements performed at terrestrial stations motivated researchers to seek alternatives by estimating PAR from satellite imagery (PAR-satellite methods). The purpose of this study is to compare five methods to estimate PAR against the measurements performed at three stations located in the area of soft fruit cultivation. Results will be discussed taking into account the constraints and challenges in accessing reliable long-term and real-time PAR datasets that are needed for this type of application.

This study has four main advantages: first, the PAR-satellite methods anterior to 2010 have never been confronted to each other in UK. Second, this analysis includes the most recent outcomes in terms of models for the delivery of spectral radiation values (2015 onwards). Third, it enables growers to immediately access long-term, regularly distributed and worldwide databases of PAR radiation values, which is essential in particular when the size of their exploitation is too small to support the costs of installation and maintenance of a complete meteorological station. Finally, the fourth strength is that the recent and reliable methods also cover water areas, opening new opportunities for other application domains such as the monitoring of the phytoplankton in the ocean.

In-situ measurements and quality checks are presented in Sect. 2. Section 3 describes the five methods explored in this analysis to derive PAR-satellite methods, separated into two groups. All methods are applied to two satellite-derived datasets of the solar broadband irradiance received at the surface (SSI) in all-sky conditions: HelioClim-3 version 5 (HC3v5) and CAMS Radiation Service (Copernicus Atmosphere Monitoring Service, abbreviated CAMS-Rad in the rest of the paper). The two datasets are briefly presented in Sect. 3.1. Section 3.2 and 3.3 respectively introduce group 1 and group 2 PAR-satellite methods. Section 4 is dedicated to the validation protocol. Results are presented in Sect. 5. Section 6 proposes an interpretation of the results and explore the cause of the underestimation observed for all methods at one of the three sites. Section 7 provides a few perspectives on potential improvements of the service and to widen the work carried out in the framework of these activities. Section 8 concludes the present document.

The three UK stations are located as depicted on Fig. 1:

Aberystwyth University, located on the west coast of Wales, provides six years of relevant PAR measurements between 2012 and 2017;

The quality of the measurements was assessed following the approach of Opálková et al. (2018). This quality check corresponds to an adaptation of the quality control of Korany et al. (2016) that applies to broadband measurements. This QC consists of checking that each measurement of PPFD at ground level and for a given instant t falls into the range [0.03, 1] ⋅ PPFDToA, where PPFDToA represents the same quantity but at the top of atmosphere. PPFDToA was computed from the solar spectral irradiance at the top of atmosphere defined by Gueymard (2004) by summing up the irradiances within the band [400, 700] nm.

The Aberystwyth University and Abbotts Hall sites provide measurements of good quality. The quality check has only discarded a few isolated data. The quality check applied on Cartmel Sands measurements discards a few slots, but it seems that this quality check procedure is not perfectly suited or sufficient to take rid of all issues at this site. This aspect is discussed in Sect. 6.2.

At least for Aberystwyth University and Abbotts Hall, all data points shadowed or flagged as suspicious were rejected, leading to trustful data sets ready to serve as references.

Numerous geostationary satellites have been operating for decades now, providing a significantly long record from which fields of SSI have been estimated. Consequently, several companies and governmental agencies are now providing access to different data sets of SSI worldwide. Given this opportunity, numerous methods exist to exploit the SSI to derive spectral radiation values. We applied the five methods to the HC3v5 and CAMS-Rad SSI, yielding an ensemble of results to compare with the surface PAR measurements. The two SSI data sets are described in the following sub-section. The two following sub-sections describe the five methods which have been separated into two groups. The first group comprises three methods that apply a constant coefficient directly to the SSI to derive the surface PAR in all weather conditions and are presented in Sect. 3.2. The second group comprises two methods which combine an accurate estimation of the spectral radiation in cloud-free conditions with a cloud extinction factor and are described in the Sect. 3.3.

Time series of 30 min means of SSI were generated from 15 min native values of HC3v5 and CAMS-Rad. At each station, the five PAR methods were applied to each SSI data set, leading to a total of 10 time series to assess per station.

The selected validation protocol is the usual one; time and space coincident estimates are compared to the corresponding measurements. Night values are discarded. Correlation coefficients are computed. The sets of differences (estimatesmeasurements) are computed and summarized by their bias, and the root mean square error (RMSE). Relative values are computed by dividing the bias or the RMSE by the mean of the measurements and are given in percent. 2-D histograms between measurements and estimates are also drawn.

Results at Aberystwyth University, Abbotts Hall and Cartmel Sands are respectively given in Tables 2–4. Each column of the tables represents each PAR-satellite method; the first three columns correspond to the methods in group 1, and the two final columns to group 2. The two first lines respectively report the number of samples used for validation and the mean of the ground-based measurements. Then for each SSI, the next three lines correspond respectively to the bias and its relative value, the RMSE and its relative value, and the correlation coefficient.

The number of coincident samples available for validation at Aberystwyth University (line 1 of Table 2) is the largest of the three sites with 49 882 samples. The best performing method for HC3v5 is Discretized-Kato, with a bias of -4 µmol m-2 s-1 (-1 % in relative value), a RMSE of 130 µmol m-2 s-1 (28 %) and a correlation coefficient of approx. 0.96 (Table 2). Similar results are obtained with the method Weighted-Kato. The range of variation for the bias of the group 1 methods is large, with values ranging from -24 µmol m-2 s-1 (5 %) for Jacovides up to 61 µmol m-2 s-1 (13 %) for Szeicz. Correlation coefficients are identical for all methods. The RMSE exhibits more or less the same values for all methods and ranges from 131 µmol m-2 s-1 (28 %) for Udo and Aro up to 153 µmol m-2 s-1 (33 %) for Szeicz.

Results obtained for CAMS-Rad exhibit a weaker performance throughout than for HC3v5 for all methods, except for the bias of Jacovides (-4 µmol m-2 s-1, -1 %). The absolute values of the bias and the RMSEs are greater for CAMS-Rad than for HC3v5 and the correlation coefficients are smaller. All methods exhibit similar performances, though the Szeicz method offers slightly weaker performances than the others.

The results at Abbotts Hall are fairly similar to those at Aberystwyth University. The number of available samples is 15 365, with a mean of the measurements of 486 µmol m-2 s-. The best performing methods are Weighted-Kato and Discretized-Kato combined with HC3v5: the bias is respectively 0 and 2 µmol m-2 s-1, leading to a relative bias of 0 % in both cases. The RMSE is 142 µmol m-2 s-1 (29 %) for both methods. The correlation coefficients are the same for the five methods: 0.944–0.945 with HC3v5 and 0.918–0.919 for CAMS-Rad. As at Aberystwyth University, the results obtained for CAMS-Rad exhibit a weaker performance throughout than for HC3v5.

Cartmel Sands exhibits similar results in terms of RMSE and correlation coefficients. The level of performance of all the methods based on CAMS-Rad is also lower than based on HC3v5. Nevertheless, the situation at Cartmel Sands is very different compared to the two other sites in terms of bias. Results obtained for the biases are as if all biases for either Aberystwyth University or Abbotts Hall had lost 5%. Szeicz method demonstrated an over estimation at the two other sites (bias of 13 % to 18 %), is now performing good with a bias of for 29 µmol m-2 s-1, 6 % HC3v5 and 18 µmol m-2 s-1, 4 % for CAMS-Rad. The smaller bias in absolute value is given by the Udo and Aro method for HC3v5 (-18 µmol m-2 s-1, -4 %) and Szeicz for CAMS-Rad (18 µmol m-2 s-1, 4 %).

As an example, 2-D histograms of PAR computed with Discretized-Kato combined with HC3v5 are presented in Figs. 3–5 corresponding to the three sites. The number of samples in each bin is coded in colour, increasing from dark blue, representing 1 sample, to dark red, representing a few tens of samples for this bin. For this method, the points are fairly aligned along the 1:1 line at the three sites (low bias) with a limited scattering of the points (low RMSE). The underestimation observed at Cartmel Sands is explained by the green cloud of points around 400 µmol m-2 s-1 just below the 1:1 line.

The 2-D histograms of the other methods exhibit similar shapes and scattering, in line with the fact that all RMSE values lies in the range [30, 35] % and correlation coefficients always exceeds 0.91. Only the slopes of the clouds of points differ from one method to another. These results confirm that the temporal variability of the measurements is well reproduced by all the PAR-satellite methods, despite a necessary post-processing to adjust bias for some of the presented methods.

Concerning the Weighted_Kato and the Discretized_Kato method, authors are currently working on a spectral version of the clear-sky index Kclear adjusted for the PAR range. Preliminary results demonstrated that the hypothesis that clouds are not spectrally dependant led to an under-estimation of approximately 10 % in the resulting PAR values. This will be the topic of another publication.

Another mean of potential improvement is to provide a better model of the albedo in the PAR range by exploring data sets of bidirectional reflectance distribution functions (BRDF) available from the MODIS mission to estimate the albedo in the PAR range. This could be based on the monthly average maps of BRDF parameters built by Blanc et al. (2014) available in the broadband range as well as in ten spectral bands.

Collaboration has already been set up to extend this validation activity across Europe, mainly in France, Spain, and Czech Republic with more than 15 sites in total. PAR data have been collected for a site in Kenya.

Discussions have also been initiated with users working in different domains, such as the monitoring of algal blooms along sea shores or in lakes, and for the monitoring of the phytoplankton. They have observed that phytoplankton of the northern hemisphere in the Atlantic Ocean is migrating further north due to global warming and climate change. They are investigating the opportunity to exploit PAR-satellite methods in their monitoring tool.

This paper reports on a comparison of five methods to derive PAR from two SSI datasets derived from satellite imagery against the measurements performed at three UK stations. A great attention has been paid to the quality of the measurements prior to the comparison. An in-depth analysis of the measurements available for the Cartmel Sands station demonstrated that this dataset should not be used. The variability of the measurements in time is well reproduced by all methods for this area. Overall, methods based on HC3v5 exhibit better results. Those using the CAMS-Rad data set globally overestimates irradiances as it misses cloudy situations, which is in agreement with the literature. Group 2 methods, and in particular Discretized-Kato combined with the cloud extinction of the SSI HC3v5, outperform methods of group 1 confirming the importance of successfully modelling the spectral distribution of the solar radiation in cloud-free conditions, with a bias equal to zero or slightly negative, a RMSE of 140 µmol m-2 s-1 (approx. 30 %) relative to the mean of the measurements) and a correlation coefficient of 0.95. This methods turns to be satisfactory to meet the diverse challenges of the specific application of assessment of the PAR potential and the real-time monitoring at more than sixty soft fruit farms relying upon tunnel production.

The validation carried out in this analysis will be extended to other sites in Europe and in Kenya in a forthcoming effort. To help in understanding the merits and drawbacks of the different PAR-satellite methods, they will be confronted to other methods, such as the alternative proposed by EUMETSAT which plans to enlarge the panel of the CM-SAF products with monthly PAR data sets as part of the next version of SARAH (version 3) developed and maintained by the DWD.

This work belongs to a wider project which aims at providing irradiances in any spectral range to meet the numerous requirements from users in various applications, for instance in solar energy, human health, agro-meteorology and e-cosmetics.