Quantifying solar radiation at the earth surface with meteorological and satellite data

94 5.3 Comparison of SIS, DSSF and ERA-Interim solar radiation with measured data.. The solar energy reaching the earth surface, referred throughout this thesis as surface solar radiatio

QUANTIFYING SOLAR RADIATION

AT THE EARTH SURFACE WITH METEOROLOGICAL AND SATELLITE DATA

J

֒

edrzej S Bojanowski

Examining committee:

ITC dissertation number 242

ITC, P.O Box 217, 7500 AE Enschede, The Netherlands

ISBN: 978-90-6164-371-5

Cover design by Zosia Dzier˙zawska

Printed by ITC Printing Department

QUANTIFYING SOLAR RADIATION

AT THE EARTH SURFACE WITH METEOROLOGICAL AND SATELLITE DATA

DISSERTATION

to obtainthe degree of doctor at the University of Twente,

on the authority of the rector magnificus,

This dissertation is approved by

Dr ir A Vrieling, assistant promoter

To Zosia,

my family and friends

One could imagine that carrying out a PhD research is but an intellectual challenge,backed by a good deal of hard work, and that what one learns during the PhDproject is specialized knowledge of a relatively narrow discipline This was myconviction, at least, when I decided to start a PhD study funded by the EuropeanCommission’s Joint Research Centre and academically supervised by the Faculty

of Geo-Information Science and Earth Observation (ITC) of the University ofTwente Four years later, as I am writing these acknowledgements, I know thatwriting a PhD is a truly transformative experience; to such an extent that I feel

a different person now For this transformation, I am indebted to all the great,knowledgeable and kind people I was honoured to work with during these last fouryears in the Netherlands, Italy and Switzerland

First and foremost, my deepest appreciation goes to my supervisors from ITC:Andrew Skidmore and Anton Vrieling If someone would ask me to describe perfectsupervision, I would simply explain the way how Andrew and Anton supervised

my PhD project I am proud and grateful for having had the possibility to workwith them, and to learn from them Moreover, I truly hope that I will be able toemploy a similar professionalism while developing my own research career I wouldlook forward to continue working with Andrew and Anton in the future, since itwas not only scientifically exciting, but also immensely satisfying on a personallevel

I owe my gratitude to Andrew for all his enthusiasm and help when, manyyears before I started this PhD thesis, we examined opportunities for carrying out

a PhD at ITC Working under Andrew’s supervision had for long been a dreamstart of a research career for me Consequently I was very happy when eventually

we found an opportunity to realize this During my PhD project, Andrew alwayswatched over me, making sure that my research was going in the right direction;

vii

conscien-The other persons at ITC that I wish to thank are Kees de Bie and ValentijnVenus, for their kind help in defining the research plan of my PhD project, andEsther Hondebrink-Lopez and Loes Colenbrander for their outstanding help withthe formalities during the project, and in the preparation of the final thesis.

I am indebted to Bettina Baruth for giving me the opportunity to join theAGRI4CAST team at the Joint Research Centre Bettina dedicated a lot of hertime to me at the beginning of the PhD project, and helped me in defining theresearch goals that formed the basis of my project I truly appreciate her readiness

to share knowledge and expertise with me I hope that the results of my PhDproject will prove beneficial to the AGRI4CAST team

I would like to thank Marcello Donatelli for the discussions we had; they werenot limited to solar radiation modelling, but touched on a lot of scientific issues

in general Our discussions were extremely inspiring and stimulating for me, andresulted in the joint research paper that is presented as Chapter 2 of this thesis.Working with Marcello was not only enlightening, but also very enjoyable

My work in the Joint Research Centre was made pleasant in large part due to

my excellent colleagues from the AGRI4CAST team A special thanks to GregoryDuveiller, Lorenzo Seguini, and my office-mate Andrea Maiorano, for scientificdiscussions, their support, and patience in listening to my thoughts and complaints

I am grateful to Allard de Wit and Gerbert Roerink for the collaboration at thebeginning of my project The conclusions of the joint paper (Chapter 3) prepared

a solid ground for my further research

I would like to express a special word of thanks to J¨org Trentmann and ChristineTraeger-Chatterjee for introducing me in the CM SAF community Our discussionsduring the CM SAF meetings definitely enriched my thesis and, moreover, assured

me that satellite climatology is a discipline I want to explore further (as I amcurrently doing)

of the examining committee of this thesis Although she was not involved in myPhD project, she provided a great contribution to starting up my research career,without which I would have never achieved what I have

I am grateful for all the support I received from Heike Kunz, Reto St¨ockli andChristof Appenzeller, as well as from all my esteemed colleagues at MeteoSwissduring the last months of my PhD project Finishing the thesis was a tough periodfor me, and I truly appreciate the encouragement they all gave me

Regrettably, I cannot acknowledge by name all my fantastic friends in Italy,because the list would not fit in here Particular appreciation goes out to myfriends from the volleyball and basketball teams: Pallavolo Ispra, Ispra Lakers and

I Trigliceridi They all know that without training several times per week, I wouldnot have found the energy and stamina to proceed with my thesis

I will be forever indebted to my beloved friends in Warsaw, whom I had left when

I decided to study abroad, and who make Warsaw a place that I am determined tocome back to some day Regardless of how long and of how far away from Warsaw

I am, I always feel their support and friendship, and it motivates me to carry onwith what I am doing

Lastly, from the bottom of my heart, I would like to thank my whole familyfor all their love and encouragement: especially both my parents, who raised megiving me work ethic and humility, and supported me all my life with remarkablesteadfastness My mother, whose moral support has been indispensable My father,who drove a truck for thousands of kilometres to help me relocate between Poland,Italy, and Switzerland My brother Micha l, who shared with me his own PhDexperience, and cheered me up when I got lost in my research My grandmotherKonstancja, professor of biochemistry, whose interest and academic advice havebeen much appreciated My aunt Agnieszka, who taught me mathematics and(perhaps more importantly) analytical thinking And most of all my loving wifeZosia, to whom I dedicate this book, who is the best person in the world, and whomakes every place where we live feel like home

1 Introduction 1

1.1 Background 1

1.2 Solar radiation data sources 3

1.2.1 Measuring solar radiation 4

1.2.2 Empirical models 5

1.2.3 Physically-based models 6

1.2.4 Satellite observations 7

1.3 Scope and objectives 8

1.4 Outline of the thesis 12

2 Auto-calibration of solar radiation models 13

2.1 Introduction 13

2.2 Data 15

2.2.1 Data from weather stations 15

2.2.2 Meteosat Second Generation data 15

2.3 Methods 17

2.3.1 Temperature-based solar radiation models 17

2.3.2 Clear-sky transmissivity 19

2.3.3 Auto-calibration procedure 20

2.3.4 Evaluation of the auto-calibrated models 22

2.3.5 Simulation of evapotranspiration 23

2.4 Results 24

2.4.1 Evapotranspiration simulation 27

2.5 Discussion 31

2.6 Conclusions 34

xi

xii Contents

3 Evaluation of MSG-derived solar radiation 37

3.1 Introduction 37

3.2 Datasets 40

3.2.1 CarboEurope flux tower radiation measurements 40

3.2.2 Weather stations 42

3.2.3 MCYFS gridded global radiation estimates 42

3.2.4 ERA-Interim radiation product 43

3.2.5 DSSF derived from MSG by the LSA-SAF 44

3.3 Validation and intercomparisons 46

3.3.1 Validation with observed radiation from CarboEurope 46

3.3.2 Comparison with observed radiation from operational weather stations 48

3.3.3 Validation and intercomparison of DSSF, MCYFS and ECMWF products 52

3.3.4 Trend analysis of DSSF data 54

3.3.5 Impact on simulated crop yields 56

3.4 Discussion 61

3.5 Conclusions 62

4 Calibration of solar radiation models using MSG data 65

4.1 Introduction 65

4.2 Materials 67

4.2.1 Surface solar radiation from MSG 67

4.2.2 Selection of weather stations 68

4.3 Methods 69

4.3.1 Estimating solar radiation from sunshine hours 69

4.3.2 Estimating solar radiation from cloud coverage 70

4.3.3 Estimating solar radiation from air temperature 70

4.3.4 Calibration of the models 71

4.3.5 Spatial interpolation of model coefficients 71

4.3.6 Evaluation of solar radiation model coefficients 72

4.4 Results 73

4.4.1 Evaluation of the models employing Meteosat Second Generation-based calibration 73

4.4.2 Interpolation of model coefficients 77

4.5 Discussion 77

4.6 Conclusions 83

Contents xiii

5 Comparison of solar radiation data sources 85

5.1 Introduction 85

5.2 Solar radiation datasets 90

5.2.1 The CM-SAF’s Meteosat First Generation data 90

5.2.2 The LSA-SAF’s Meteosat Second Generation data 90

5.2.3 ERA-Interim reanalysis 91

5.2.4 Interpolated weather station data 91

5.2.5 Ground measurements used for validation 93

5.3 Methods 94

5.3.1 Comparison with ground measurements 94

5.3.2 Comparison of the solar radiation datasets 95

5.4 Results 98

5.4.1 Comparison with ground measurements 98

5.4.2 Comparison of the solar radiation datasets 104

5.5 Discussion 110

5.5.1 Quality of solar radiation measurements 110

5.5.2 Concatenating satellite-based products 112

5.5.3 Backup solution for the merged satellite-based dataset 114

5.6 Conclusions 116

6 Synthesis 119

6.1 Introduction 119

6.2 Improvements in solar radiation modelling 120

6.3 Accuracy of satellite-derived solar radiation 122

6.4 Merging solar radiation datasets 125

6.5 Key practical outcomes of this thesis 128

6.6 Perspectives 129

A Performance statistics 133

Summary 135

Samenvatting 137

References 139

Curriculum vitae 149

ITC Dissertations Series 151

List of Tables

radiation models for the ground- and auto-calibration methods 25

calibration method 75

Supit-Van Kappel, and Hargreaves models depending on the

calibration method 76

radiation calculated against measured solar radiation 100

solar radiation estimates 107

xv

List of Figures

auto-calibration 26

for ground- and auto-calibration 28

ground-and auto-calibration 29

Campbell model and measured solar radiation 31

model against measured solar radiation 32

radiation at operational weather stations 50

radiation 51

xvii

xviii List of Figures

observed radiation at operational weather stations 52

3.8 Annual average global radiation estimated by three products 54

3.9 Differences in annual average global radiation estimates between MCYFS, DSSF and ECMWF 57

3.10 Potential total above-ground biomass of maize as calculated by MCYFS using standard MCYFS and DSSF global radiation 58

3.11 Water-limited total above-ground biomass of maize as calculated by MCYFS using standard MCYFS and DSSF global radiation 59

4.1 Spatial distribution of differences in RRMSE of the ˚ Angstr¨om-Prescott, Supit-Van Kappel and Hargreaves models depending on the calibration method 74

4.2 Box plot of the MBE and RRMSE of the of ˚Angstr¨om-Prescott, Supit-Van Kappel and Hargreaves models 78

4.3 Interpolated ˚Angstr¨om-Prescott model coefficients 79

4.4 Interpolated Supit-Van Kappel model coefficients 80

4.5 Interpolated Hargreaves model coefficients 81

5.1 Weather stations from the JRC-MARS database 92

5.2 Number of weather stations reporting solar radiation 94

5.3 Comparison of SIS, DSSF and ERA-Interim solar radiation with measured data 99

5.4 RMSE of SIS, DSSF and ERA-Interim solar radiation 101

5.5 MBE and standard deviation of bias error of SIS, DSSF and ERA-Interim solar radiation 102

5.6 Comparison of SIS and measured solar radiation for each Meteosat First Generation mission 103

5.7 The monthly-aggregated MBE of SIS, DSSF and ERA-Interim solar radiation 104

5.8 Annual average (2005) solar radiation from SIS, DSSF, ERA-Interim and JRC-MARS 105

5.9 Standard deviation (2005) of solar radiation from SIS, DSSF, ERA-Interim and JRC-MARS 106

5.10 Kernel density plot of the differences between SIS, DSSF, ERA-Interim and JRC-MARS solar radiation 108

5.11 Spatio-temporal distribution of differences between SIS, DSSF, ERA-Interim and JRC-MARS solar radiation 109

List of Figures xix

5.12 RRMSE and MBE derived from a comparison between SIS, DSSF,ERA-Interim and JRC-MARS solar radiation 1105.13 Per-grid bias removal between SIS and DSSF solar radiation 114

List of Acronyms

Satellites

and Water Management

xxi

xxii List of Acronyms

Short Range Forecasting

List of Symbols

xxiii

xxiv List of Symbols

calculation)

semivari-ogram calculation)

According to the recent earth energy budget estimates by Wild et al (2013), 22

solar radiation reaches the earth’s surface (Fig 1.1) The solar energy reaching

the earth surface, referred throughout this thesis as surface solar radiation, global

radiation or simply solar radiation, depends on the geographic location,

orienta-tion of the surface, time of the day, time of the year, and atmospheric composiorienta-tion

1

2 Introduction

(Boes, 1981) Therefore solar radiation varies significantly in time and space Thediurnal cycle is caused by the Earth’s rotation around its own axis, which changesthe angle at which the solar radiation reaches the surface The Earth’s move on theorbit around Sun and the axial tilt of Earth in relation to the orbital plane causethe seasonal fluctuation of day length and sun elevation angle (Linacre and Geerts,1997) The day length and sun elevation angle at a given location also depend onthe location’s latitude For any location and time, the solar radiation reaching thetop of the atmosphere can be obtained by accounting for these regular (diurnal andseasonal) fluctuations in the earth-sun geometry, and a solar constant (e.g Sell-

ers, 1965) This top-of-atmosphere radiation is also referred to as extra-terrestrial

throughout this thesis neglecting the 11-year cycle of the sun activity (Willson andHudson, 1991) In the absence of an atmosphere, terrain slope and aspect would

be the only factors modifying solar radiation Using a good-quality digital terrainmodel, these terrain effects can be estimated, also accounting for the shadowingand reflections from the nearby surfaces (e.g Kumar and Skidmore, 2000; St¨ockli,2013) However, the atmosphere has a strong effect on the solar radiation received

at the earth surface, thus causing additional variability (e.g Liou, 2002) Clouds

are mostly responsible for the high variability in the atmospheric transmissivity,

since they can reflect significant part of incoming solar radiation Estimating theatmospheric transmissivity and thus the solar radiation reaching the surface ischallenging, because the atmospheric composition (including clouds) has to bemodelled

Location-specific time series of accurate solar radiation data are important formany scientific and application fields, including agriculture, ecology, biodiversity,hydrology, meteorology and climatology Solar radiation is also a critical inputparameter for the design, performance prediction, and monitoring of solar energydevices (Davy and Troccoli, 2012; Monforti et al., 2014) In this thesis the mainconcern, however, is to obtain accurate solar radiation estimates that can be used

to improve results of crop growth models

Crop growth models are useful for several purposes Firstly, they have helped inquantifying the environmental limits to specific crop production at a given location(Aggarwal and Penning de Vries, 1989) Secondly, they have been extensively usedfor forecasting crop yields (e.g Baruth et al., 2007; Boogaard et al., 2002) Finally,running crop growth models with future weather scenarios as input can help inassessing the impact of expected climatic shifts on future crop production (e.g.Tubiello et al., 2000; Wolf et al., 1996) In addition to other weather variables (i.e.air temperature and precipitation), solar radiation is an essential variable required

1.2 Solar radiation data sources 3

Fig 1.1 Schematic diagram of the global mean energy balance of the Earth The bers (in W m −2 ) indicate best estimates for the magnitudes of the globally-averaged solar radiation fluxes together with their uncertainty ranges, representing present day climate conditions at the beginning of the twenty first century from (from Wild et al., 2013).

num-by most crop growth models (Fig 1.2), typically at a daily temporal resolution.Solar radiation has been identified by Nemani et al (2003) as the major constraint

to plant growth in Europe and in other areas worldwide (Fig 1.3), thus justifyingthe purpose in this thesis of improving the accuracy of solar radiation estimates

1.2 Solar radiation data sources

Direct measurements of solar radiation at weather stations are the most accuratesource of solar radiation data, provided that the equipment is well-maintainedand regularly calibrated Various methods have been developed to obtain solarradiation estimates for locations, where it is not directly measured The simplestsolution is to assign measured values from a nearby station (e.g Hunt et al.,1998) or to use spatial interpolation methods (e.g Bechini et al., 2000; Ertekinand Evrendilek, 2007) However, the density of solar radiation measurements isoften not sufficient for reliable interpolation A different approach, which is notdirectly based on measured solar radiation at nearby stations, is to model solarradiation Satellite observations have also provided an alternative means to derive

1.2.1 Measuring solar radiation

Global solar radiation, which consists of direct and diffuse radiation, is typicallymeasured by a pyranometer (Gueymard and Myers, 2008a) Due to the sensitivity

1.2 Solar radiation data sources 5

Fig 1.3 Geographic distribution of potential climatic constraints to plant growth derived from long-term climate statistics (from Nemani et al., 2003).

of the device, the precise measurement of solar radiation requires frequent tion and maintenance (Ohmura et al., 1998) This means that it is difficult andcostly to collect solar radiation time series for a large number of stations with

calibra-a consistently-high mecalibra-asurement calibra-accurcalibra-acy Even though the number of stcalibra-ationsthat measure solar radiation is increasing (Wild et al., 2013), it is small compared

to the number of stations that record air temperature and precipitation The tio between stations observing radiation and those observing air temperature andprecipitation may be as low as 1:500 at the global scale (Thornton and Running,1999; Trnka et al., 2007) The sparse network of stations that directly measuresolar radiation is therefore insufficient to accurately estimate solar radiation forlarge areas using interpolation techniques

ra-1.2.2 Empirical models

Empirical solar radiation models employ relationships found between atmospherictransmissivity and other meteorological variables These variables mainly includesunshine duration (e.g Ertekin and Evrendilek, 2007), cloud cover (Supit and VanKappel, 1998) and air temperature (e.g Abraha and Savage, 2008), but can be alsosupported by other variables such as precipitation and humidity (e.g Weiss andHays, 2004; Wu et al., 2007) In all cases, the input meteorological variable mainlyrelates to the amount of clouds which is the key limiting factor of solar radiationfor reaching the earth surface, as clouds reflect a significant part of incoming

6 Introduction

solar radiation The accuracy of solar radiation models varies depending on themeteorological variable used (e.g Mavromatis, 2008; Trnka et al., 2005; Wu et al.,2007) A very high correlation between daily sunshine duration and daily solar

measures the time per day of direct solar radiation not attenuated by clouds Cloudcover is usually visually assessed by the observer (WMO, 2008) several times perday (e.g every 3 hours) Average daily cloud cover calculated from these discreteobservations is less correlated with solar radiation than sunshine duration (e.g.Mavromatis, 2008; Trnka et al., 2005; Wu et al., 2007) The relation betweensolar radiation and daily air temperature range has been extensively exploitedbecause of the high number of weather stations measuring air temperature (e.g.Bristow and Campbell, 1984; Hargreaves et al., 1985) The difference betweenmaximum and minimum air temperature is a proxy of the cloud amount reflectingincoming solar radiation Relatively high differences between day-time and night-time air temperatures usually occur during the cloud-free conditions In that casethe incoming solar radiation is not filtered by the clouds during the day, whileduring the night the infrared emission from the soil surface is rapidly lost in theatmosphere and the air temperature decreases more quickly Conversely, opaquecloud cover limits the amount of solar radiation reaching the surface during theday (and thus decreases the air temperature), while during the night clouds reflectback the infrared radiation emitted from the surface (and limit the decrease of airtemperature) The occurrence of advection of cold or warm air masses can reducethe accuracy of air temperature-based model estimates, as the relation betweenthe air temperature range and cloud amount is not preserved in that case.One of the key limitations of empirical solar radiation models is their site-dependent coefficients (e.g Abraha and Savage, 2008) These coefficients can beaccurately determined only for weather stations where solar radiation has beenmeasured, although they can be interpolated for other locations (Van Kappel andSupit, 1998) The accuracy of solar radiation models decreases when their empiricalcoefficients cannot be determined at the location where the model is applied

1.2.3 Physically-based models

Physically-based models employ algorithms to model the transfer of solar radiationthrough the atmosphere based on known relationships and laws from physics Solarradiation transferring the atmosphere is absorbed by gases (i.e water vapour, oxy-

1.2 Solar radiation data sources 7

gen, carbon dioxide, methane, nitrous oxide and ozone) and scattered by molecules,aerosols and cloud particles (e.g Liou, 2002) Due to the complexity of the radia-tive transfer and high temporal and spatial variability of atmospheric composition,the physically-based models require numerous input parameters in order to locallytune them Physically-based models generally rely on estimates of atmosphericcomposition derived from numerical weather prediction models (Dee et al., 2011).For instance, in the ERA-Interim reanalysis of the European Centre for Medium-Range Weather Forecast (ECMWF), water vapour information (including clouds)

is produced by the forecast model, whereas for other gases and aerosols, logical information is used (Dee et al., 2011) Application of weather predictionmodels can lead to long-term solar radiation datasets covering large areas, butwith low spatial resolution typical for the weather forecasting systems (such as for

Geostationary satellites are the most important source of satellite-derived solarradiation data These satellites have a fixed position of 35786 kilometres abovethe earth surface (i.e the equator, e.g Jones and Vaughan, 2010) This fixed po-sition allows constant scan of the exposed hemisphere, and as a consequence, toprovide satellite observations at high temporal resolution (i.e 15 minutes) Thespatial resolution of these observations is between 3 and 5 km Their weakness isthe increase in the apparent size of the pixel with latitude, and longitude (Ineichen

et al., 2009) Frequent observations are essential in estimating the amount of solarradiation that reaches the earth surface, which is strongly influenced by the highlyvariable atmosphere Nowadays, satellite observations from geostationary weather

8 Introduction

satellites, including the American Geostationary Operational Environmental lites (GOES), the European Meteosat missions (Schmetz et al., 2003), the Indian

Meteorological Satellite (HIMAWARI) (Thies and Bendix, 2011), are available forthe whole globe except the poles, which are not visible from satellites positionedover the equator

Polar orbiting satellites can observe polar regions since their orbits pass nearlyabove both poles Satellite sensors on a polar orbit of about 800 km above the earthsurface can deliver satellite imagery of high spatial resolution (i.e 250–1000 m),but at low temporal frequency (e.g 1 day) This low temporal resolution limits theusability of a single polar orbiting satellite in the operational estimation of surfacesolar radiation However satellite series such as NOAA/AVHRR, with more thanone satellite in polar orbit, can deliver a full global coverage several times a day(Kogan et al., 2011) The NOAA data has been used by Karlsson et al (2013) tocreate 30-year long time series of global solar radiation However, given the hightemporal variation of solar radiation, the frequent observation of instruments onboard geostationary satellites are expected to deliver solar radiation estimates at

a higher accuracy than the polar orbiting satellites

1.3 Scope and objectives

Despite the wide range of methods that exist for retrieving surface solar radiation,existing solar radiation estimates often do not fulfil the specific requirements ofoperational applications (such as operational crop growth models) These require-ments are described below

Accuracy

The error of solar radiation estimates propagates into modelled crop growth cators Studies have shown that crop growth models have similar accuracy whenusing solar radiation measurements versus using empirical solar radiation modelsthat employ sunshine duration (Pohlert, 2004) The lower accuracy of other so-lar radiation models, which are based on cloud cover or air temperature range,negatively influence the accuracy of modelled crop growth (Trnka et al., 2007).However, these reported accuracies of solar radiation model outcomes are typi-

indi-1.3 Scope and objectives 9

cally related to models calibrated with locally-measured solar radiation values andapplied to the meteorological variables measured at the same location For cropgrowth modelling at regional scales, solar radiation estimates are required at everylocation To achieve this, spatial interpolation of model coefficients and/or mod-elled solar radiation is required, which reduces the accuracy In this context, solarradiation estimated from satellite observations has an essential advantage of beingspatially continuous, thus no interpolation is needed

Spatial coverage and resolution

This thesis aims at improving the accuracy of solar radiation estimates to be used

in regional and continental studies (such as crop growth monitoring) Thereforethe sought solar radiation dataset has to cover large areas Since the crop growthmonitoring system for Europe driven by the European Commission’s Joint Re-search Centre (MARS Crop Yield Forecasting System, MCYFS) is used here as

a case study, the minimum spatial coverage is set to Europe However, the larger(global) extent is of high interest, and thus it is under consideration throughoutthe thesis The target spatial resolution of the solar radiation dataset is 25 km, asused by the MCYFS However the spatial resolution of solar radiation estimatesfrom geostationary satellites, which are used here extensively, is higher (3–5 km).Therefore the potential solar radiation dataset of spatial resolution higher than 25

km is discussed in the thesis

Temporal coverage and resolution

Most of the crop growth models such as the World Food Studies crop growthmodel (WOFOST, van Diepen et al., 1989) and CropSyst (St¨ockle et al., 2003)used for crop monitoring at regional scale operate with a daily time step Thisdefines the required temporal resolution of the input parameters including solarradiation The required temporal coverage is more difficult to identify If the cropgrowth model is used for forecasting of the final yield, the analysis often relies onregression between time-series of historic simulated and reported crop yields Insuch case, a consistent time series of minimum 15 years is required (Baruth et al.,2007)

10 IntroductionAccessibility in near real-time

For operational crop growth monitoring systems, near real-time data on surfacesolar radiation are required In order to monitor in a timely manner the rapidchanges that occur during the growing season as crops respond to the weather,data on input variables should be available with a maximum delay of 1–3 days(Duveiller et al., 2013) Only up-to-date information about crop status allowsfor the issuing of early warning on alarm situations such as severe drought orplant sunburn This near real-time requirement has to be taken into account whendesigning the operational crop growth monitoring systems as it may limit thesophistication of the data processing that can be tolerated (Duveiller et al., 2013).Following the requirements listed above, the overall objective of this thesis is:

To provide and test approaches for improved estimation of daily surface solar radiation to be used for operational crop growth modelling in Europe.

This objective can be split up into three main research questions:

Question 1:How can the accuracy of solar radiation estimated from existingempirical models be improved?

The accuracy of solar radiation estimates from empirical models mainly depends

on the meteorological variable used as input and the effective calibration of themodel Normally, calibration requires measured solar radiation from the same lo-cation where the model is applied However, such measurements are not availablefor many stations The hypothesis in this thesis is that effective calibration can beachieved without the requirement for solar radiation measurements This wouldstrongly enhance the applicability of solar radiation models at regional scales

Question 2: What is the spatial and temporal variability of the error ofsatellite-derived surface solar radiation datasets?

During the last decades satellite observations have been successfully used toestimate surface solar radiation (Section 1.2.4) Even though the performance ofmethods for solar radiation retrieval from satellite data has been analysed in pre-vious studies for a limited number of locations (Geiger et al., 2008; Posselt et al.,2011), not much is known about the spatial and temporal variability of the error

in satellite-derived solar radiation estimates To address this research question,

1.3 Scope and objectives 11

a high number (> 500) of weather stations measuring solar radiation will be usedfor evaluation Even though these stations are not equally distributed in space,they should be sufficient to detect spatial patterns in the solar radiation estimates.The spatial and temporal variability in the accuracy of solar radiation estimates isexpect to have several sources, which are discussed in the following However, a de-tailed quantitative analysis of the influence of each source of error on the accuracy

of solar radiation estimates is not in the scope of this thesis

It is expected that the accuracy of solar radiation estimates will be lower forregions (temporarily) covered by snow, as snow can be erroneously detected ascloud Thus a problem of snow might introduce a seasonal variability in accuracy

of solar radiation data as well as a variability related to latitude and elevation Thetemporal variability of error of satellite-derived surface solar radiation datasets can

be also introduced by differences in sensor characteristics and their degradation,since long-term solar radiation datasets consist of observations from several instru-ments The accuracy of solar radiation estimates can be lower in complex terrain:

a given location can receive less direct radiation due to the shadowing effect ofthe neighbouring surfaces Moreover, the sky view from a given location can belimited by the surrounding terrain, which limits the incoming diffuse radiation

It is also expected that solar radiation estimates from geostationary satellites areless accurate more distant from the nadir view of the sensor, where the size of theimage pixel and the satellite viewing angle are increasing

Question 3: Can currently available satellite-based solar radiation datasets

be merged to create accurate long-term solar radiation time series?

Only recently a greater effort has been put into the creation of satellite-basedsolar radiation datasets and into the facilitation of access to these datasets (Schulz

et al., 2009; Trigo et al., 2011) Depending on the objectives of the data producer,the resulting dataset may be updated in near real-time to enable application ofthe radiation estimates in operational systems or, alternatively, may provide longerhomogeneous historical records (i.e not updated in near real-time) that are ap-plicable for climate studies The comparability of resulting datasets that are builtaround different objectives has not been studied in detail previously, hence little

is known regarding the possibility of merging them Several applications, such asoperational regional crop growth modelling defined beforehand, require solar radi-ation datasets that would combine both features of being homogeneous and beingprolonged in near real-time Several issues are expected to hinder the concatena-

12 Introduction

tion of solar radiation datasets including different spatial resolution and differentsensor characteristics, which introduce inconsistencies in the solar radiation timeseries Moreover, the differences between solar radiation datasets are expected to

be not uniformly distributed in time and space

1.4 Outline of the thesis

The core of this thesis is composed of Chapters 2–5, which are based on four reviewed research papers Below the research goals of each chapter are introducedand the link to the overall objective of the thesis is explained

peer-Chapter 2 aims at improving the applicability of solar radiation models forthe locations, where solar radiation measurements are not available for the modelcalibration The chapter introduces an auto-calibration procedure, which allowsestimating solar radiation from other meteorological variables without prior cali-bration of the solar radiation model Chapter 2 tackles the first research question

(Question 1 ).

Chapter 3 evaluates the solar radiation estimates derived from Meteosat SecondGeneration data against high-quality solar radiation measurements Further, theimpact of remote sensing based solar radiation estimates on modelled crop growth

is assessed Chapter 3 partially contributes to the answer to the second research

question (Question 2 ).

Chapter 4 explores the possibility to improve the accuracy of solar radiationmodels through calibrating them using solar radiation estimates from MeteosatSecond Generation data The proposed method allows creating maps of solar ra-diation model coefficients for the whole of Europe, and thus estimating solar ra-diation from meteorological variables for every location Chapter 4 together with

Chapter 2 answer the first research question (Question 1 ).

Chapter 5 compares solar radiation estimates from different data sources cluding those created with methods described in Chapter 3 and 4) Based on thiscomparison the recommendations for the operational use of solar radiation esti-mates at European scale are given and discussed Chapter 5 aims at answering the

(in-third research question (Question 3 ).

Finally, Chapter 6 gives a summary and critique of the main findings, the tical relevance of the presented methods, as well as the perspectives for furtherdevelopment

evapo-of measured values In addition, prediction evapo-of crop yields requires solar radiationestimates at similar spatial and temporal resolutions as for other weather vari-ables derived from global and regional circulation models This cannot be achievedmerely with ground solar radiation measurements.

Different approaches have been developed to estimate solar radiation, whichcan be grouped into: (1) physically-based radiative transfer models, (2) empiri-cal models, based on the statistical relationship between measured meteorologicalvariables and incoming solar radiation, and, more recently, (3) Bayesian neuralnetwork methods (Tymvios et al., 2005; Yacef et al., 2012) Physically-based ra-diative transfer models generally require a large number of input variables Al-though empirical models are less data-demanding with respect to input variables,the accuracy of these models largely depends on reference solar radiation data,required to calibrate model coefficients Accurate calibration can only be achievedfor locations where solar radiation is measured For locations without solar radi-ation measurements, model coefficients are interpolated (e.g Bechini et al., 2000;Fodor and Mika, 2011; Van Kappel and Supit, 1998; Miller et al., 2008), which

∗ This chapter is based on: Bojanowski, J.S., Donatelli, M., Skidmore, A.K., and Vrieling, A (2013) Auto-calibration method of temperature-based solar radiation models.

Environmental Modelling and Software, 49:118–128.

13

14 Auto-calibration of solar radiation models

can result in larger errors (Bojanowski et al., 2013b) Interpolation of the solarradiation models will be further discussed in Chapter 4

Most empirical models utilise the daily range between minimum and maximumair temperature Clear days show a greater air temperature range: daytime airtemperatures are high because clouds do not absorb incoming solar radiation;night-time air temperatures on the other hand are low because infrared radiation

is emitted from the earth’s surface to the atmosphere and not radiated back byclouds This relationship is however weaker during conditions of advection, whichreduces the performance of air temperature-based empirical models in some regionsfor specific periods during the year Despite this limitation, these empirical modelshave proved effective to accurately estimate solar radiation at several locations(e.g Abraha and Savage, 2008; Bristow and Campbell, 1984; Grant et al., 2004;Hargreaves et al., 1985; Trnka et al., 2005)

Two common empirical solar radiation models based on daily air temperaturerange are the models proposed by Bristow and Campbell (1984) and Hargreaves

et al (1985) The Bristow and Campbell model exploits a saturation-type, nential relationship between daily total solar radiation and daily air temperaturerange In contrast, the Hargreaves model uses a linear relationship between solarradiation and the square root of daily air temperature range These models aretypically calibrated based on measured solar radiation, resulting in site-specific co-efficients Locations measuring solar radiation are relatively sparse over Europe, soapproaches for calibrating the models for weather stations without solar radiationmeasurements would be useful

expo-The auto-calibration procedure, which we present in this chapter, can be usedfor the calibration of air temperature-based solar radiation models without solarradiation measurements The term ’auto-calibration’ is used here to indicate a cal-ibration without reference solar radiation data Our procedure is fundamentallybased on the assumption formulated by Allen (1997) that on clear-sky days themodel should approximate but not over-predict potential solar radiation Potentialradiation can easily be calculated for any location, as it is a function of location,day of the year, and atmospheric composition Thus, the auto-calibration algorithmfirstly identifies cloud-free days, and secondly optimises the model’s empirical co-efficients to meet Allen’s assumption This is done for a specific location using only

a time series of daily air temperature ranges thereby allowing solar radiation to beestimated without prior calibration with measured solar radiation data

The specific objectives of this study are: (1) to introduce a procedure to calibrate empirical solar radiation models, (2) to evaluate the performance of theauto-calibration procedure based on the Bristow and Campbell, and Hargreaves air

auto-2.2 Data

2.2.1 Data from weather stations

Meteorological data were obtained from 126 weather stations These data were tracted from the Joint Research Centre’s Monitoring Agricultural Resources Unit(JRC-MARS) database, which is the input for the MARS Crop Yield ForecastingSystem (Baruth et al., 2007; Boogaard et al., 2002) The stations range from lat-

1677 m They are located in ten countries: France, Germany, Italy, the Netherlands,Poland, Portugal, Spain, Tunisia, Turkey, and the United Kingdom Each station

at least 60% of the days during 2005–2010 All but six stations provided dailywind speed and water vapour pressure means, thus allowing for the estimation ofevapotranspiration

2.2.2 Meteosat Second Generation data

We used Meteosat Second Generation data to create two look-up tables to be usedwithin the auto-calibration procedure: (1) a map of mean annual cloud fractionalcover, and (2) a map of clear-sky atmospheric transmissivity The data used toestimate cloud fractional cover are and for the clear-sky atmospheric transmissivityretrieval are described in the following subsections It should be emphasized thatcurrent satellite data are not required for using the auto-calibration procedure,since, once calculated, two satellite-derived maps (look-up tables) are stored withinthe auto-calibration algorithm Alternative methods to those used here exist toderive both of these maps Clear-sky transmissivity may alternatively be estimated

16 Auto-calibration of solar radiation models

based on atmospheric composition (e.g Fortin et al., 2008), while a map of meancloud fractional cover can be derived from interpolation of values measured atweather stations However, the MSG-derived estimates have the advantage of beingspatially continuous and thus allow deriving the mean cloud fractional cover andclear-sky transmissivity maps without interpolating ground measurements

2.2.2.1 Cloud fractional cover

To obtain a map of mean annual cloud fractional cover we used five years (2007–2011) of the monthly cloud fractional cover product provided by EUMETSAT’sSatellite Application Facility on Climate Monitoring (CM-SAF) The product isderived from 15-minute pixel-based cloud detection from visible and near-infra-redMeteosat Second Generation SEVIRI data (Derrien and LeGl´eau, 2005), developed

by the Satellite Application Facility for the project ’Support to Nowcasting andVery Short Range Forecasting’ (Geiger et al., 2008) The monthly cloud fractionalcover is delivered by the CM-SAF at 15 km × 15 km resolution For each pixel,

we calculate the average of all monthly values to obtain the mean annual cloudfractional cover (in percent)

2.2.2.2 Surface solar radiation

To estimate clear-sky atmospheric transmissivity, we used six years (2005–2010) ofthe down-welling surface short-wave radiation flux (DSSF) daily product derivedfrom the Meteosat Second Generation satellite data, for which the pixel size is

3 km at the equator and approximately 5 km in Central Europe (Geiger et al.,2008) The solar radiation product is generated and made freely available by theLand Surface Analysis Satellite Applications Facility (LSA-SAF) and further pro-cessed by the Flemish Institute for Technological Research (VITO) on behalf ofthe Joint Research Centre’s Monitoring Agriculture Resources (MARS) Unit Thesolar radiation obtained for wavelengths between 0.3 µm and 4.0 µm is considered

in this product In the retrieval scheme used, the down-welling surface short-waveradiation flux is approximated by multiplication of the top-of-atmosphere solarradiation and the effective transmittance of the atmosphere or cloud atmospheresystem (Geiger et al., 2008) The effective transmittance is calculated using twoapproaches depending on whether a given pixel is classified as clear or cloudy Thisclassification is done based on the 15-minute cloud mask, which monthly derivative

is described in Section 2.2.2.1