GKSS 2001/23 Derivation of the Photosynthetially Available Radiation from METEOSAT data K. Shiller Derivation of the Photosynthetially Available Radiation from METEOSAT data Kathrin Shiller Abstrat Two dierent models, a Physial Model and a Neural Net, are used for the derivation of the Photosynthetially Available Radiation (PAR) from METEOSAT data in the German Bight. Both models are onstruted for the alulation of PAR in the German Bight in terms of an easily aessible time series of PAR elds; advantages and disadvantages of both models are disussed. A software pakage was generated in IDL for the realisation on the omputer, its struture is motivated and physi bakground informations are given. With slight modiations all programs an be applied for the alulation of PAR for arbitary regions (but not over land). A zip le ontaining all programs and lasses used, the tehnial html doumentation and this report an be obtained from http://gfesun1.gkss.de/software/meteosat2par. Ableitung der zur Photosynthese zur Verfugung stehenden Strahlung aus METEOSAT Daten Abstrat Zur Ableitung der zur Photosynthese zur Verfugung stehenden Strahlung (PAR) in der Deutshen Buht aus METEOSAT Daten werden zwei vershiedene Modelle, ein Physikalishes Modell und ein Neuronales Netz, benutzt. Beide Modelle sind darauf ausgelegt, PAR in der Deutshen Buht in Form einer einfah zuganglihen Zeitreihe von PAR-Feldern zu berehnen; Vor- und Nahteile beider Modelle werden diskutiert. Zur Realisierung auf dem Computer wurde ein Programmpaket in IDL erstellt, dessen Struktur motiviert wird und physikalishe Hintergrunde erlautert werden. Alle Programme sind so angelegt, da sie sih durh geringfugige Modikationen eignen, PAR auh fur andere Gebiete (aber niht uber Land) zu berehnen. Ein zip File aller benutzter Programme, Klassen, der tehnishen html Dokumentation und dieses Berihtes kann bei http://gfesun1.gkss.de/software/meteosat2par bezogen werden. 2 Contents 1 Introdution 7 2 METEOSAT data Proessing 7 2.1 The METEOSAT System . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 The Two Models for Deriving PAR 3.1 Physial Foundations . . . . . . . . . . . 3.2 The Physial Model . . . . . . . . . . . . 3.2.1 The Heliosat Method . . . . . . . 3.2.2 Software . . . . . . . . . . . . . . 3.2.3 Disussion of the Physial Model 3.3 The Neural Net . . . . . . . . . . . . . . 3.3.1 The Neural Net ffbp1.0 . . . . . 3.3.2 Disussion of the Neural Net . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 9 12 12 15 15 18 23 25 25 29 4 Analysis of the results for the German Bight 30 5 Summary 32 1 Introdution The present work ame into being as a part of the `ENVOC: A New Perspetive of the Oean' projet. Within this projet the work pakage 3:4 is onerned with the development of a model whih determines the primary prodution using a ombination of remotely sensed data from MERIS, ASAR, AATSR or AVHRR and METEOSAT. An important input variable for eah Primary Prodution model is the Photosynthetially Available Radiation (PAR). The derivation of PAR in the `German Bight' from METEOSAT data was the goal of the work presented here. Sine the METEOSAT satellite views about 42% of the earth surfae the rst subtask was to generate a program whih allows to extrat only a subset of METEOSAT data for a speied `region of interest' (roi).This program and some related auxillary programs will be desribed in setion 2 together with a brief introdution to METEOSAT and the data format `OpenMTP'. The next step was to onsider how PAR ould be derived from these `roi-les'. Two dierent methods have been implemented, the usage of a so alled 'Physial Model' as well as a Neural Net parametrization of PAR. Setion 3 starts with some physial foundations valid for both models. The Physial Model, its realization and its limitations will be introdued, followed by the desription of the Neural Network parametrization together with a brief introdution to Neural Nets. For both models a quality estimation of the model used will be inluded. Afterwards the results will be disussed with fous on the `German Bight', the region of interest in the present work. For this purpose data of the day- and month means of PAR in the roi will be used. All programs for this work were written in `IDL' (Version 5:4) in objet-oriented manner. All programs and their purpose will be illustrated in the respetive setion. This will inlude motivations of the lass denitions as well as a desription of some assoiated methods. A tehnial desription of all programs is provided online[1℄. 2 METEOSAT data Proessing This setion is onerned with the extration of a subset of METEOSAT data for a speied `region of interest' (roi). To do so, one rst of all has to deal with the struture of METEOSAT image produts i.e. the `OpenMTP format'. This is done in the next setion inluding an introdution to the METEOSAT servies. Afterwards the software used for the extration and some auxiliary programs will be disussed in some detail. 7 2.1 The METEOSAT System Image produts ontain the basi image data aquired by the Meteosat satellites. These data are obtained every 30 minutes in three spetral wavebands. Sine the satellites are in geostationary orbit, the overage is approximately hemispherial and entered on the sub-satellite longitude rossing with the equator. The main satellite is loated over the Greenwih meridian. The imagery itself onsists of retangular arrays of 8-bit-image pixels. The satellite is spin-stabilised, and the data are aquired at a rate of one image line per satellite rotation. Eah time the satellite rotates, the radiation detetors for the various hannels pan aross the earth 'horizontally' (i.e. eastwards), aquiring one line of data. Between eah rotation, a stepping tilt mirror in the amera optis adjusts the position so that the next line is oset 'vertially' (i.e. northwards) from the last. A total image is aquired every 25 minutes. A ve minute retrae and stand-by period prepares the satellite for the start of the next image, so that the image interval is 30 minutes. The interval is known as `slot', and there are 48 slots in eah day of operations. Slots are numbered from midnight so that slot 1 overs data aquired from midnight to 00:30 UTC. Emissions in the following three spetral wavebands are deteted by Meteosat: VIS: wavelengths in the range 0.5 to 0.9 mirons - showing reeted light in the visible part of the spetrum, IR: wavelengths in the range 10.5 to 12.5 mirons - showing emitted radiation in the thermal infrared part, WV: wavelengths in the range 5.7 to 7.1 mirons - showing radiation in the water vapour absorption bands. Data are aquired at two resulutions. In the Water Vapour (WV) and Infrared (IR) hannels a omplete image onsists of 2500 x 2500 pixels. This gives a spatial resolution of about 5 km at the subsatellite point. In the Visible (VIS) band, data are aquired at twie this resolution. Images are available in dierent data formats. All software desribed in this doument does apply for data in the `OpenMTP le format' only. The mahine level representation of bits and bytes used in the OpenMTP format follows the standard used by UNIX/open systems arhiteture mahines. The open system representation uses the IEEE standard for real number representation and ASCII enoding for harater data. All OpenMTP les onsist of a variable number of logial reords of variable length. This struture ontains three distint omponents: Reord 1, ASCII le header. This ontains general information about the le in ASCII format. Reord 2, binary le header. Contains extensive binary format information on the produt and its alibration. 8 Reords 3-N, image line data. The number of lines N of data depends on the han- nel (see above). Eah image line is (NP+32) bytes in length, where NP is the number of pixels per line. Besides the image line data the produt informations are of great use as one will see in the following setions. All informations presented here were taken from publiations of the `European Organisation for the Exploitation of Meteorologial Satellites EUMETSAT' [2℄[3℄. 2.2 Software Before reading the image line data for the region of interest (roi) for a speied time intervall one needs to hek whether the METEOSAT data are omplete and ontain full disk retied data (due to the fat that a great amount of METEOSAT les has to be proessed one has to make sure that they are uniform and leading to the same roi). To do so, the auxiliary program `list bad files' is provided. The program heks whether the list of les is omplete (i.e. 48 slots per day, 3 types of data (VIS, IR, WV) per slot) by reading all the le names whih will give information about the date, slot and type. In order to hek if all les ontain retied and full disk data the program reads the ASCII-header of all les, in whih general informations are stored. The ASCII-header provides the reord `Desription' (taking one of the values `Image subarea' or `Full disk image') and the reord `ProessingPerf' (proessing performed on the image `Raw Data' or `Retied Data') for this task. If one ore more les are missing or ontaining `bad data' (i.e. `Image subarea' or `Raw data' ) the program will list the date, slot and type of these les in a le named `list bad files.txt'. The main program `OpenMTP2roi' reads the image lines belonging to the roi and writes their roi parts together with an ASCII header (see below) into `roi-les'. For this purpose a lass `OpenMTP file' was dened with several methods to all. In order to determine the image lines of interest the user needs to speify a grid (i.e. a (south-west) starting point (latitude and longitude) a step-width (in latitudinal and longitudinal diretion) and the number of steps in (latitude and longitude). The so speied grid (interpreted as ellipsoidal oordinates) is then onverted to a artesian grid and then to pixel-oordinates, whih allow to adress roi. Before produing the roi-les, i.e. running the main program, one should make sure that the speied grid really leads to an image of the speied region (it was found that the routines onverting geographial- to pixel oordinates work quite well when areas around 0 latitude were hoosen and that errors of up to three lines or pixels oured when the speied region is of latitude 50 degree). For this purpose the auxiliary program `offset' is provided. With this program one an shift the resulting roi to adjust it to the desired position. The framed subarea will be monitored and so an be easily ompared (by viewing) with a geographial map of roi. When it oinides with roi the so found oset parameters an be used afterwards in the main-program. 9 The objet oriented manner in whih the programs where written is espeially useful sine the user has not to deal with large and omplex programs.1 Calulations whih are ommon to all les are done only one in the main program (i.e. the rst time one alls the appropiate method) whih makes it very fast. Suh methods are the `OpenMTP file::grid' method whih alulates the artesian grid speied by the users geographial grid input (see above) and the `OpenMTP file:: grid onvert' method whih alulates the orresponding pixel-oordinate (this method atually runs three times, beause of the dierent image sizes of `.VIS' and the other types (`.IR' and `.WV') leading to three sets of pixel-oordinates). list bad files.pro OpenMTP-files offset.pro OpenMTP2roi.pro roi-les Fig. 1: filmhen.pro The gure shows the data types and the programs designed for eah type in a shemati way. Data types are harterized with the dis symbol, programs are framed. After alling these methods for the dened objet the main program will then extrat the pixels assoiated with the roi in a le (named after date, slot and type of the original input OpenMTP le). These roi-les also ontain an ASCII-header where information ommon to all roi-les (speied grid, size of ASCII-header and size of roi-data) and some astronomial values needed in order to alulate PAR are stored. Besides these informations this header should ontain the alibration informations (but later on it was found that the alibration informations were not present in the OpenMTP-les binary header as they should, so month mean values of the alibration informations had to be expliitly inluded in the programs whih needs them). The benet of the ASCII header will beome lear when disussing the derivation of PAR from these roi-les. A list of ontents of this header is given in the table below. For the astronomial alulations a lass `roi al' was dened with methods assoiated with the region, date and time when data were taken but not with the roi data themselves. Suh methods are: alulating sunrise- and sunset-time (in UTC) 1 At this point it is worth mentioning that the above programs aept '.z' and '.gz' OpenMTP les, so one should never unzip them. 10 Table 1: Contents of the roi ASCII header `Region of Interest' Spetral Content of (OpenMTP) image data (`VISS + VISN (visible south +north) data') Re1Size Size of ASCII header in bytes, always set to `800' Re2Size Size of image in bytes RetMethod Name of retiation method used (OpenMTP) NumberOfLines Number of lines in the roi image (i.e. steps in Latitude) NumberOfPixels Number of pixels in the roi image (i.e. steps in Longitude) Longitude Longitude of south-west starting point of roi Latitude Latitude of south-west starting point of roi Delta Lon step width in longitudinally diretion in roi Delta Lat step width in latitudinally diretion in roi Earth Sun Dist earth sun distane (in AU) os sun zenith osine of sun zenith angle in roi sun azimut rad sun azimuth angle (in rad) in roi diff angle rad azimuth dierene sun-satellite (in rad) in roi os sun sat osine of angle between sun and satellite as seen from roi os sat zenith osine of satellite azimuth angle CalibraCoeff absolute alibration oeÆient (W=m2 ST R) SpaeCounts spae ount (dimensionless) CopyRight `Copyright () GKSS/GFE All rights reserved' Desription SpetralCont (`roi al::sunrise sunset') , alulating the sun position vetor (`roi al:: sun position') i.e. Earth-Sun distane, zenith- and azimuth angle, and alulating the azimuth dierene from sun and satellite (`roi al::azidiff sun sat'). The method for alulating sunrise- and sunset-times is used in the main program for speifying the number of les that should be onverted to roi-les in terms of `slots before (or after) sunrise (sunset)' , e.g. for the onversion to PAR only those les with sunligth present are of interest. The so reated roi-les an now be used to alulate the PAR for the region of interest. To get an idea how a series of suh roi-les looks like the auxiliary program `filmhen' is provided, displaying a sequene of the speied les. A summary of the software disussed in this paragraph an be seen in gure 1. 11 3 The Two Models for Deriving PAR In this setion the two dierent models -the so alled `Physial Model' as well as the Neural Net parametrization- for the derivation of PAR from METEOSAT data will be introdued. First the physial foundations valid for both models will be layed. Afterwards the two models themselves and their omputational realization will be desribed. For eah of the models a brief disussion of its sope and its error budget will be given. 3.1 Physial Foundations The amount of `Photosynthetially Available Radiation' (PAR) is dened by Z 700nm E0 (~x; )d [photons s 1 m 2 ℄; PAR 400nm h where E0 (~x; ), the spetral total irradiane, is the total radiant power per square meter at wavelength oursing through point ~x owing to photons traveling in all diretions. The term E0 =h in the above denition gives the number of photons generating E0 . That means for photosynthesis it is the number of available photons rather than their total energy that is relevant to the hemial transformation. This is beause a photon of, say, wavelength 400nm, if absorbed by hlorophyll, indues the same hemial hange as does a less energeti photon of wavelength 600nm. Note that PAR is by denition a broadband quantity, there is no `spetral PAR'. However, over a wide variety of surfae harateristis, the onversion fator for energy to photons varies by only 10% about a onstant value (this approximation is valid when not taking into aount the spetral dependane of the total irradiane). This simpliation is also made in both models desribed in this setion, i.e. the global irradiane will be derived from METEOSAT data and is onverted via a onstant fator to PAR. The global irradiane reahing the earth's surfae is on one hand depending on the urrent position of the sun relative to the earth in spaetime and on the other hand inuened by sattering and absorption proesses in the earth's atmosphere. Before going into details about the atual applied models these two fators should be explained briey. Figure 2 (taken from [4℄) shows a elestial sphere with the earth at the enter and the sun revolving around the earth. In the elestial sphere, the elestial poles are the points at whih the earth's polar axis uts the elestial sphere (mutatis mutandis for the equator). The intersetion of the plane of the earth's equator with the plane of the sun's revolution, the elipti, makes a xed angle of 23:5Æ . However, the angle between a line joining the enters of the sun and the earth's equatorial plane hanges every instant and is alled the delination angle. The delination angle and the knowledge of the geographial position on earth are suÆient to dene the sun zenith and azimuth angles, whih inuene the amount of 12 North Pole of Celestial Plane Plane of Celestial Equator Autumnal Equinox Delination N Angle Earth S Apparent Path of Sun on the Elipti Plane Summer Solstie 23 5Æ : Sun Vernal Equinox South Pole of Celestial Plane Fig. 2: 90Æ The gure shows the elestial sphere with the apparent path of sun and the sun's delination angle. solar radiation reahing the top of the atmosphere (toa) of the earth. The zenith angle is the angle between the loal zenith and the line joining the observer and the sun. The solar azimuth is the angle in the loal oordinate system between the plane of the observer's meridian and the plane of a great irle passing through the zenith and the sun. Besides these two angles the amount of radiation reahing the toa of the earth is inversely proportional to the square of its distane from the sun. Sine the earth revolves around the sun in an elliptial orbit with the sun at one of the foi the earth-sun distane is, as the delination angle, a funtion of the day in the year only2 . When solar radiation enters the earth's atmosphere, a part of the inident energy is removed by sattering and a part by absorption (see gure 3, taken from [4℄). Both inuene the terrestrial spetrum by onsiderably modifying the spetral energy passing through the atmosphere. The sattered radiation is alled diuse radiation. A portion of this diuse radiation goes bak to spae and a portion reahes the ground. The radiation arriving on the ground diretly from the solar disk is alled diret radiation (and is obviously proportional to the osine of the sun zenith angle). The diuse omponent onsists of several parts. Besides the sattering by air moleules and aerosols the interation of diret solar radiation with the louds results in reeted diuse radiation. Further, a portion of the diret and the diuse radiation reahing the earth after the rst pass through the atmosphere is reeted bak to the 2 The mean earth-sun distane is alled one astronomial unit (AU). The minimum distane is about 0:98 AU, and the maximum about 1:02 AU. The earth is at its losest point to the sun (perihelion) on approximately January the third and at its farthest point (aphelion) on approximately forth of July. The earth is at its mean distane from the sun at 4th of April and at 5th of Otober. 13 Solar Radiation reeted bak to Spae Limit of Earth's Atmosphere Thin Clouds Aerosols Clouds Air Moleules Diret Anisotropi Diuse Radiation arriving on a horizontal Surfae Fig. 3: The gure shows the various proesses inuening the amount of radiation arriving on the ground. sky, whih ontributes to the multiply reeted irradiane. This multiple reeted irradiane (not shown in gure 3) will depend srongly on the reetane properties of the loud system. When the diretional intensity of the diuse irradiane is not uniform over the sky hemisphere, it is alled anisotropi diuse radiation. Loal Vertial Sun Satellite Sun Zenith Angle Dierene of Azimuth Angles Fig. 4: The gure shows the position of the sun and a satellite in the loal oordinate system of an observer on the earth's surfae Before summarizing the minimum of input variables for eah model for deriving the global irradiane from METEOSAT data one has to take into aount the satellite itself. Sine the `observer', the satellite, is plaed in a geostationary orbit it has a xed `viewing' angle onto the earth, i.e. it is never (or only one per day) inline with diret omponent of the irradiane reahing a spei region on the earth. Therefore the amount of irradiane deteted in the satellites radiometer (i.e. the reeted irradiane) is a funtion of the angle between the satellite and the sun as seen from the ground (see 14 gure 4). (For the alulation of this angle the earth's urvature was not taken into aount, sine our roi is suÆient small enough to do so.) The above onsiderations are leading to the following set of input variables (besides the satellite data themselves): date and time when data were taken, loation of the region of interest on earth (longitude and latitude), earth-sun distane for the time the data were taken, zenith angle of the sun in the loal (roi) oordinate system for the time the data were taken, zenith angle of the satellite in the loal oordinate system, angle between the satellite and the sun as seen in the loal oordinate system for the time the data were taken. The physial informations in this setion were taken from[4℄ and [5℄. 3.2 The Physial Model In this paragraph the physial model used for derivation of PAR is introdued. First the model itself will be desribed briey. Afterwards the realization of this model on the omputer together with some auxiliary programs will be visualized. At the end there will be a disussion of the results and a quality estimation. 3.2.1 The Heliosat Method The use of satellite-based methods for the retrieval of surfae solar irradiane relates the earth radiane at satellite altitude to the radiative properties of the system and to the top of the atmosphere solar irradiane. The satellite radiometer measures within a given solid angle the radiation sattered bak to spae by the system `earth-atmosphere'. This quantity is a measure of the planetary albedo and is reversely related to the atmospheri transmission of solar radiation. Sine surfae solar irradiane is strongly determined by the transmission harateristis of the atmosphere, there is a strong omplementarity between global irradiane and the satellite signal. The use of remotely sensed radiane data from satellites is well-suited to the task of surfae solar irradiane estimation for two reasons. First, most of the solar radiation reahing the earth's surfae originates from visible to near-infraread wavelength ( 15 0:4 1:0m). Energy at wavelengths longer than 1:0m is almost totally absorbed by even thinnest louds, and energy at wavelengths shorter than 0:4m is largely lost due to moleular sattering and absorption by ozone. Seond, the louds are the main modulator of the surfae solar irradiane, and they often an be observed easily from spae. A high (low) value of net solar ux at the surfae is onsistently aompanied by a low (high) value of loud optial thikness, and therefore a low (high) value of reeted solar ux at satellite altitude. Meteosat Visible Counts Normalization Reetivity Ground Reetivity Cloud Reetivity Cloud Index Clear Sky Index Clear Sky Model (Dumortier/Page) Global Irradiane Clearness Index Diuse Fration Model (Skartveit et al.) Diuse Fration Diuse Irradiane Fig. 5: The gure shows the basi steps for the estimation of global irradiane from METEOSAT data. Diuse irradiane is alulated by applying a statistial model. The various methods (e.g. [6℄, [7℄) presented in the literature mainly dier in the desription of the relationship between atmospheri transmission and outgoing radiane as seen by the satellite. Physial models diretly desribe the radiative proesses in the atmosphere by means of radiative transfer alulations while empirial methods are based on statistial relationships between satellite and ground measurements. However, 16 most operational methods, inluding the one used here, for the estmation of surfae irradiane atually inlude elements of the other respetive onept. The algorithms for retrieving global irradiane from satellite data where taken from the paper by A. Hammer et al. [8℄ and is referred to as the Heliosat Method. It desribes how to alulate the global irradiane from METEOSAT ounts in VIS. The method is basially driven by the strong omplementary between the planetary albedo reorded by the satellite's radiometer and the surfae radiant ux. The planetary albedo inreases with inreasing atmospheri turbidity and loud over. For a orret estimation of the hange in reetivity for a given image element the inuene of the inoming radiation on the reeted radiation has to be onsidered. Therefore a normalization with respet to the zenith angle of the sun is apllied. This relative reetivity % (C C0 )dist2 %= Irrext os(#) (where C0 is a registration oset, dist is the earth-sun distane in AU, Irrext is the extraterrestrial solar spetral irradiane at mean earth-sun distane for the spetral range of the visible METEOSAT detetor and # is the sun zenith angle) thus gives a measure for the planetary albedo. With the relative reetivities estimated in this manner a loud index n is derived for eah pixel as a measure of loud over: n= % %g % %g where %g and % are the relative reetivities in the lear sky and overast ase, respetively. To establish the before mentioned relationship between atmospheri transmission and planetary albedo, a quantity haraterizing the transmittane has to be introdued. In the paper [8℄ the lear sky index k is used, whih is a funtion of the loud index only and is dened by the ratio of surfae global irradiane Ig to surfae global irradiane under lear sky onditions Ilear (the latter one has to be inferred from a lear sky model): k = = Ig Ilear ; 8 >> < >>: 2:0667 1:2 1 n 3:6667n + 1:6667n2 0:05 : n 0:2 : 0:2 < n 0:8 : 0:8 < n 1:1 : 1:1 < n; so the surfae global irradiane an be alulated, if Ilear is known. The lear sky model for alulating the global irradiane under loudless skies Ilear is desribed as a sum of diret and diuse irradiane. For the diret omponent of lear sky global irradiane a physial model is used [9℄, whereas for the diuse omponent 17 an empirial models is used [10℄. For both models the relevant parameters, besides the earlier mentioned sun parameters Irrext ; dist; # are: a side spei Turbidity TL (see table in [8℄), the relative airmass m = (os(#) + 0:50572(96:07995ir the rayleigh optial thikness %R (m) = (6:6296 + 1:7513m 0:1202m2 + 0:0065m3 #) 1:6364 ) 1, 0:00013m4) 1 . With these parameters the global irradiane under loudless skies Ilear an be alulated as follows Ilear = Idir os(#) + Idiff ; Idir = Irrext dist 2 exp( 0:8662TL %R (m)m); Idiff = Irrext dist 2 (0:0065 + os(#)( 0:045 + 0:0646TL) os2 (#)( 0:014 + 0:0327TL)): In summary the global irradiane at the surfae an then be determined from the lear sky index hareterizing the atmospheri transmittane and the lear sky irradiane. While the latter one is modeled with a site spei turbidity [11℄, the lear sky index is derived via the loud index from the normalized satellite ounts (the relative reetivities). A sheme of the proedure is given in gure 5. The model presented here is limited due to the fat that METEOSAT ounts represent data integrated over quite a large spetral range. That means by denition that the physial model parametrization an only lead to a mean representation of all proesses atually depending on the inoming wavelength. Other fators determining the method's unertainty are inuenes of louds, aerosols and surfae reetion whih again are only haraterized by mean inuenes. 3.2.2 Software After the desription of the physial model used its realization on the omputer will now be disussed. Most of the underlying input parameters in this model are well-dened physial values; only the relative reetivities in the lear sky and overast ase have do be determined seperately from VIS ounts. In order to proess the roi-les the lass `roi file' was dened. It provides several methods inluding reading les (`roi file::file start', `roi file::header info' and `roi file::read roi') and alulating the global irradiane. For alulating the global irradiane the physial model was splitted into two parts, i.e. one method to determine the relative reetivities (`roi file::reflet bytewise') and one method (the atual physial model) whih uses the determined relative reetivities `roi file:: phys model bytewise'. This was done beause one rst has to determine the minimum 18 Fig. 6: The gure shows an example of the output of the `reflet param' auxiliary program. (the above-quoted lear sky ase) and maximum (overast ase) value of the relative reetivities whih then serve as parameters for the atual model. In order to determine these two values the `reflet param' program was used. It alulates the relative reetivities for all pixels whih are `above sea' from a number of roi-les (one should make sure that the program uses roi-les from days with louds as well as from those with lots of sunshine in order to get orret results) and plots them afterwards. From this plot one an then determine the minimum and maximum value and enter them in the `roi file::phys model bytewise' method. Figure 6 shows the output of this auxiliary program for roi-les of Marh and July 1994. Norderney, List data DW Data linear.m jebo read.m time series data Fig. 7: Helgoland data time series i. data The gure shows the onversion of the supplied measured data into data whih an be easily aest with `time series( i)' methods. As mentioned above (see `Physial Foundations') the physial model is deriving the global irradiane only. In order to determine PAR from global irradiane a onversion 19 fator has to be found. For this purpose one obviously needs a set of data of global irradiane and PAR measured at the same site and time. These data were taken from measurements (every ten minutes) for the year 2000 on Heligoland. The data an also be used to test the models used and for the training of the Neural Net (see below). For the year 1994 data from the `Deutsher Wetterdienst' of measurements of the global irradiane (every hour) on Norderney and List were taken for the testing and training. Fig. 8: The gure shows an example of the output of the `glob2par' auxiliary program for measurements on Heligoland of the year 2000. roi-les physmodel testing.pro make mask.pro time series data time series i. data Fig. 9: physmodel testing.pro glob2par.pro The gure shows the auxiliary programs disussed in this setion together with the data they work on. Data types are hareterized with the dis symbol, programs are framed. The measured data were rst onverted into a time series with the help of two MATLAB auxiliary programs `DW data linear' and `jebo read'. The rst one was used for the `Deutsher Wetterdienst' data. Due to the fat that these data were nearly 20 omplete the missing values were derived by linear interpolation. For the latter one the interpolation was not always possible (it was only done when not more than ve sequent measurements were missing). Figure 7 shows the onversion of the data shematially. Due to the diversity of the two sets of data two distint lasses had to be introdued for reading out the data of the time series for a speied time, the lass `time series' for Norderney and List data and the lass `time series i' (`i' for inomplete) for the Heligoland data. For both of these lasses methods were provided for reading data (not only salar type but also vetor- or grid-like) from the time series les (whih have to have a ertain struture disribed in the header of the `time series::read' (or `time series i::read') method) and for returning the value of `data' for a speied time (i.e. the methods `time series::value' and `time series i::value') by interpolation. Fig. 10: The gure shows an example of the output of the `physmodel testing' auxiliary program for the Marh and June 1994 data. The determination of the above onversion fator was done with the auxiliary program `glob2par' whih uses the measured data from Heligoland of PAR and global irradiane. Within the program a linear orrelation of the two data sets is assumed so that a onstant onversion fator is obtained by simple linear regression (assuming a regression line through the origin). The output of the program are two les, one for 21 the regression line and values for the slope and its standard deviation. The other plot gives the deviation of the values from this tted line. Figure 8 shows these two output les, giving a onversion fator of 2:197(0:001)[ mol W s ℄. Before the atual alulation of PAR one should test whether the relative reetivities entered in the `roi file::phys model bytewise' method are well suited for deriving the global irradiane for the region of interest. For this purpose the auxiliary program `physmodel testing' is provided. It ompares the measured data of global irradiane with the alulated ones. The program alulates the global irradine for ve points (over sea) near the measuring station, looks up the orresponding measured value and reates three les, one ontaining a satterplot of the aulated values vs the measured ones; leading to a regression line, one ontaining a plot of the deviations and one of the relative errors. Figure 10 shows the output for the 1994 data (Marh and June). A summary of all auxiliary programs will be given in gure 9. If the output of the `physmodel testing' program was satisfatory (i.e. a mean deviation lose to zero, a slope lose to one and a suÆient small standard deviation) one an then run the main program alled `physmodel gridded'. The main program alulates PAR (from the output of the `roi file::phys model bytewise' method via the onstant onversion fator) for all points of the roi-les and writes the values into a time series (inomplete, grid like type). Using this program is therefor espeially useful for a series of roi-les. The so generated gridded time series an be easily aessed with the `time series i' methods. Fig. 11: The gure shows an example of the output of the `PAR means' program for the month means of Marh 1994 (left) and the day means of Marh 4th 1994 (PAR is in mol m2 s ). For both the region of interest is the `German Bight' with the land overed by a mask. These methods were used to reate another auxiliary program alled `lm gridded ts'. It displays a `movie' of the PAR time series for an interval speied by the user who an also adjust the speed and diretion of this movie interatively. The reated time series of PAR was also used for alulations of day- and month mean values of PAR in the roi. This is useful if the time series overs a range of a month 22 roi-les physmodel gridded.pro time series roi PAR film gridded ts.pro PAR means.pro day-, month means pit Fig. 12: The gure shows the data types and the main programs for deviation of PAR with the physial model in a shematially way. Data types are harterized with the magneti dis symbol, programs are framed. or so and is done with the `PAR means' program. The program will reate a number of output les, one of them ontaining a oloured plot of the deviations of the month mean values of PAR in roi from the total month mean value (red orresponds to values of PAR greater then the total month mean value, blue orresponds to a smaller one) and the other ones (as many as there are days in this month) are the orrespondend plots of the deviations of the day mean values in roi from the total day mean values. (Note that the month mean was alulated as the mean of the day means, so it is atual the mean day value and oinides with the month mean values if all days have the same number of slots with sunshine.) For all of these plots only points in roi `above sea' an be taken taken into aount sine the physial model used is valid only above sea. This is reahed by using a `land-sea mask' reated with the `make mask' program from a roi le with nearly no louds present in roi. Figure 11 shows an example of the output of the `PAR means' program. A summary of all programs diretly related to PAR is given in gure 12. 3.2.3 Disussion of the Physial Model The Physial Model presented has the advantage that it is appliable to any region overed by water; when hanging the region of interest the model by onstrution produe results with the same quality. This is due to the fat that the model is based 23 on mean representations of the various variables the global irradiane depends on. On the other hand the Physial Model has the disadvantage of having large errors (see the satterplot (regression line) and the plot of the deviations in gure 10). The origin of the errors should now be disussed in more detail. In the satterplot a large amount of points is lose to the origin. These points are assoiated with sunrise or sunset, i.e. with a small amount of global irradiane. On sunrise and -set the angle of the inoming irradiane is lose to , the path through the atmosphere is of maximum length and the earth's urvature has non-negletible inuene on the path length. Within the model the urvature is not taken into aount, so it is not surprising that most of these points have overestimated global irradiane. In the plot of the relative errors in gure 10 these points are assoiated with the large positve deviations from 200 to 400 %. Anyway, these points are not orrelated with the asymmetry of the deviations in the above gure, i.e. the width of the distribution is bigger for underestimating then for overestimating the global irradiane. In order to nd the origin of that the Physial Model was tested again, but this time with only one day with sunshine (1.6.1994) and afterwards with only one day with an overast sky (28.3.1994) (for this purpose an auxiliary program `disussion phys' was written whih is just a modiation of the `physmodel testing' program). The testing was done by plotting a satterplot of the model returned values vs the ground measured ones (again the data from the `Deutsher Wetterdienst where used) together with the regression line from gure 10. Additionally the values for the two sites (Norderney and List) are haretarized with two dierent symbols. The output an be seen in gure 13. Fig. 13: The gure shows the testing of the Physial Model for an overasted day (left) and one with lear sky (right). The `x' symbolises points from List the `4' points from Norderney. In the plot for the lear sky ase it is remarkable that nearly the same amount of alulated points is plaed above as is plaed below the regression line, i.e. this plot leads to a symmetri distribution of the orresponding derivations. On the other hand the plot for the overast ase shows an asymmetry: a large 24 amount of points is plaed below the line, i.e. the global irradiane is underestimated very often. This might be explained by multiple bakreetane in the thin loud ase: then the diret omponent of the global irradiane an still pass through these louds, a part of it will be reeted onto the louds and from there bak to the surfae leading to bigger then expeted value of irradiane on the surfae. These proesses are not taken into aount in the model and may be the reason for the disussed asymmetry. In the same plot one an also see that most of the points orresponding to overestimated global irradiane originate from the same measuring site, i.e. above the line there are more `rosses' (orresponding to List). This might be due to the fat that in the overast sky the global irradiane is strongly depending on the site: when hanging the position only by a few kilometers the value might vary drastially. Sine for eah measuring site ve points lose to this site where hoosen for the test of the model this ould explain the overestimating for only one site in a few ases and again the disussed asymmetry. In summary the Physial Models performane is, as expeted, better for the lear sky ase than for the overast one due to the omplexity of the proesses in this ase. The advantage is the broad appliability of the model. 3.3 The Neural Net Another way of determining global irradiae from METEOSAT data is the usage of a Neural Net (NN) parametrization whih will be illustrated in this setion. The NN `learns' from a training sample (ontaining a number of input variables as well as the desired output, i.e. the global irradiane). The so trained NN an then be used for the derivation of global irradiane from the input variables. The onversion to PAR is done via the onstant onversion vator (see above). First of all the NN used will be introdued. 3.3.1 The Neural Net ffbp1.0 Before desribing the Neural Network (NN) and its training briey, some general informations about NN should be given. A NN is an interonneted assembly of simple proessing elements, neurons, whose funtionality is loosely based on the animal neurons. The proseing ability of the network is stored in inter-neuron onnetion strength, or weights, obtained by a proess of adaption to, or learning from, a set of training patterns. Properties ommon to all NN are: NN's provide aurate, fast and onvenient mathematial (statistial) models. 25 NN's have apabilities for informational modeling of dependenies on more than one variable. NN's retrieve aurate algorithms. NN's are a natural tool for multi parameter retrievals. NN's an be fast forward models for diret assimilation. NN's are robust in the presene of noise: small hanges in an input signal will not drastially aet a neurons output. Trained NN's an deal with `unseen' patterns and generalize from the training set. The program used in the present work `ffbp1.0' is a C-program by H.Shiller [12℄ for training of feedforward bakpropagation NN, a multiple non-linear regression method, running on UNIX omputers. Feedforward bakpropagation NN are the most frequently used NN in remote sensing. `Feedworward' means that input values in the rst neuron layer will be only propagated `forward', there are no bak onnetions whereas `bakpropagations' means that during the learning of the NN the desired output is ompared with the NN output, the errors will be traed bak and the weights will be readjusted. The learning funtion implemented is a gradient desent algorithm with momentum term. The program allows for starting with a small net whih an be enlarged later if neessary. In the publiation `Feedforward-Bakpropagation Neural Net Program `ffbp1.0' [12℄ the usage of this program is desribed in detail. Therefore in this doumention there will be no desription of this program but the result of the training proess will be presented. For the training of the NN one needs to reate a training le in whih in eah line a number of input variables is written as well as the desired output. The ouput data where taken from measurments taken in Norderney and List of the `Deutsher Wetterdienst' of 1994 on an hourly basis. Sine METEOSAT data are taken every half an hour one has to have a program whih an interpolate from the `Deutshe Wetterdienst' data for arbitary times. For this purpose again the lass `time series' was used with one method for reading in the `raw' data (`time series::read') and a method `time series::value' whih returns the interpolated value (see above, setion `Software') for any given time (i.e. the time at whih the satellite sanned roi). Input variables for the NN are: the alibrated VIS, IR and WV ounts (i.e. radiane in Watts per square meter and steradian) the azimuth dierene (in rad) between Sun and Satellite the osine of the sun zenith angle divided by the square of the Sun Earth distane (in AU) 26 The last two values an be alulated from the parameters of the roi Asii header. In priniple the alibration oeÆients for the three hannels an be taken from the OpenMTP les binary header, but for the VIS hannel they are only available sine 1995. Details about where the atual alibration oeÆients were taken from an be found in the tehnial desription [1℄. In dierene to the `Physial Model' all three hannels available from METEOSAT data serve as input variables. This is done beause some of the meteorologial properties an be `seen' on the IR and WV images. Suh properties inlude loud motion, temperature, upper tropospheri humidity, lear sky radiane and loud top heights. Due to the fat that these properties are modelled by their mean inuene in the `Physial Model' only, it is expeted that the NN works better and over a wider range of spreading of the input variables. To reate the training les (i.e. one le ontaining the input and ouput variables as olumns ending with `.patt' and one le ontaining the names of all these variables ending with `.dsp') the program `NN training' was used. From the roi-le it hooses ve points lose to Nordernay and List whih are for sure `above sea'. A sheme of the training proess is given in gure 14. roi-les time series data NN training.pro training-les plot nn training.pro ffbp1.0 Fig. 14: The gure shows the data types and the programs needed for the training of the NN in a shemati way. The auxiliary program plot nn training an only be run after the training of the NN. Data types are harterized with the dis symbol, programs are framed. From this training sample about 90% were used for training the NN and the remaining 10% for testing the NN-output. It was found that the training led to best results when two hidden layers with 15 neurons in the rst and 3 neurons in the seond layer for the NN were hoosen. Figur 15 shows the performane of the trained NN for the test sample of the marh and june 1994 data, in whih the NN-output was ompared 27 with the `Deutsher Wetterdienst' measurements. Sine the NN program ffbp1.0 only provides these plots for the test sample the auxiliary program plot nn training (see gure 14) was used to obtain the plots for both, the testing and the training sample, for easier omparing the output of the physmodel testing output (see gure 10) with NN output. The values of errors of the so trained NN are: trainings sample has a total sum of squares of errors: 106.758442, ratio avg.train/avg.test= 0.864312, average of residues: training 106.758442/13453/5= 0.007936 test 20.140849/2317/1= 0.008693. Fig. 15: The gure shows the performane of the trained NN. In the upper left part the regression line of NN-output vs desired output is shown. The upper rigth part shows a plot of the orresponding deviations. Below a plot of the relative errors is given. The so trained NN was used to reate an easily aessable time series of PAR elds. The NN is used in this ontext to extraploate pointlike measurements to enlarge the area. 28 3.3.2 Disussion of the Neural Net By omparing the performane of the Physial Model (see gure 10) and of the NN (gure 15) it is obviously that the NN output has a muh smaller width in the distribution of the deviations from the tted regression line. Also this distribution does not show the asymmetry present in the Physial Model. This is due to the fat that the various proesses inuening the ouput in the overast ase an be taken muh better into aount in a pure statistial model suh as a NN. This statement an be prooven when running the NN for only one day with lear sky and for one day with overast sky (for better omparison the same days as in the disussion of the Physial Model where used). For this purpose the auxiliary program `disussion nn' was written whih is a modiation of the `plot nn training' program. Again, the NN output was plotted vs the measured values as a satterplot together with the regression line from gure 15. For the NN the output shows no asymmetry (i.e. muh more point above or below the regression line) neither in the lear sky ase nor in the overast ase (see gure 16). The better performane for the lear sky ase might be explained due to the strong site dependeny in the overast ase. In the plot the asymmetry of one site (as for List in the testing of the Physial Model performane) annot be seen. It is ommom to both models that the overast ase leads to higher deviations than the lear sky ase. Fig. 16: The gure shows the testing of the Neural Net for an overasted day (left) and one with lear sky (right). The `x' symbolises points form List and the `4' points from Norderney. The disadvatage of the NN is that it an not be applied to any other region overed with water beause the NN has only `learned' the optial properties for this speial region of interest. Anyway, when supplied with ground measurements for any other site another NN an be trained easily in the same manner as disribed above. In summary one an state that the use of a NN for derivation of PAR should be preferred to any Physial Model sine by onstrution a NN an take the various proesses determining PAR on a surfae muh better into aount then a non-statistial 29 model relying on averaged relations. 4 Analysis of the results for the German Bight In the previous setions two dierent models for deriving PAR (via a onstant onversion fator from the derived global irradiane) from METEOSAT data in a region of interest (roi), here the `German Bight', have been introdued. Now the output of these models, the Physial Model and the Neural Net (NN), will be disussed with fous on the roi. For these purpose pitures of the month mean of Marh and June 1994 were alulated with both of the models. Figure 17 shows the output for the Physial Model and gure 18 for the NN. In the gures the alulated mean value is always set to `white', values of greater PAR are `red' and lower PAR is `blue'. Eah olour sale is dened by the largest derivation from the mean value, i.e. in most of the ases only the maximum (or minimum) of the olour sale will be atually used in the orresponding piture. The mask used in the pitures not only overs the land but also small islands and the tidelands between (this is why the mask does not look like the land on a map, but was nesessary beause both of the models only work above sea). Fig. 17: The gure shows the output of the `PAR means' program for the month means of Marh 1994 (left) and of June 1994 (PAR is in mol m2 s ) alulated with the Physial Model. For both the region of interest is the `German Bight' with the land overed by a mask. As expeted the ouput looks similar for both models used. The alulated PAR values for eah month are of the same order for both models and are a bit lower for the NN output. In all ases the derivation from the mean value is of the same order (or bigger) then the orresponding standard derivation when testing the model. Nevertheless the error in the above pitures is muh less then these values due to orrelations from one pixel to the neighbouring ones (i.e. the possibility of an error in alulating PAR for a region with similar values of PAR is muh less ompared with the error on just one single pixel in this region). 30 Fig. 18: The gure shows the same month mean values as gure 17 but this time alulated with the NN. Fig. 19: The gure shows the day means (alulated with the NN) for an overasted day (left, 28.3.1994) and one day with (nearly) lear sky (right, 1.6.1994). Again PAR is in mol m2 s . Another noteworthy fat is the struture present in all of the above pitures: lose to the oast the month mean values of PAR are less then above sea, i.e. in the month mean it is brighter above sea. At a rst glane this may look like an artifat. To proof that this is not the ase gure 19 shows the day means of one day with nearly no louds present and one day with an overast sky (for this purpose the same days as in the disussions of the models were used). One an see that in both of the pitures the above mentioned struture is not visible but that the value of PAR is nearly the same in the whole region. The struture might be explained due to blushing on the oast when old and humid airmasses from above the sea and warmer, dryer ones from above the land `meet' in oastal regions. Anyway, it is obvious that for deriving any parameters in roi whih depend on PAR (as the primary prodution in roi) one should not use a mean value of PAR for the whole region. 31 5 Summary The derivation of the Photosynthetially Available Radiation (PAR) in the `German Bight' from METEOSAT data sueeded. From the data a subset orresponding to this region of interest (roi) was extrated. For the atual derivation of PAR from ounts in the roi-les two dierent models have been used, a so alled Physial Model and a Neural Net (NN) implementation. Both the models alulate the global irradiane whih is afterwards onverted via a onstant fator to PAR. To test the quality of the models their output was ompared with ground measurements of the global irradiane (Norderney, List). From this it turned out that the NN is better suited for the task for several reasons. It rst of all has a good performane independent of the weather onditions, whereas the Physial Model underestimates PAR for an overast sky. Seond the Physial Model uses only the VIS ounts and desribes the inuene of louds and other properties modifying the diuse omponent in an overast ase only via their mean inuene whereas the NN extrats these informations from the IR and WV ounts whih also serve as input values for the NN. The major advantage of the Physial Model is the fat that it is appliable to any region overed by water, by onstrution it will produe results of the same quality. If one wants to use the NN in other regions it has to be `taught' for this region to aount for possible other optial properties in this region. The work presented here is a part of the workpakage 3.4. of the `ENVOC' projet. This workpakage is onerned with the determination of the Primary Prodution in the `German Bight', where PAR serves as an important imput variable. With respet to this projet the most important results are: the value of PAR within this region utuate so that the use of a mean value for PAR is not reommended. Therefore the values of PAR for all points in the region are written into a time series whih an be easily aessed for arbitary times in the range overed by the time series. Seond a robust method for the derivation of PAR has been found: the use of the NN is reommended. 32 Referenes [1℄ K. Shiller http://gfesun1.gkss.de/software/meteosat2par Tehnial Desription available [2℄ The Meteosat System - Satellites, Ground Segment, Missions, Global Coordination EUM TD 05 Revision 4 (2000) [3℄ The Meteosat Arhive - Format Guide No.1 - Basi Imagery, OpenMTP Format EUM FG 1 Revision 2.1 (2000) [4℄ M.Iqbal An Introdution to Solar Radiation Aademi Press, 1983 [5℄ C.D.Mobley Ligth and Water - Radiative Transfer in Natural Waters Aademi Press, 1994 [6℄ R.Frouin et al. A Simple Analytial Formula to Compute Clear Sky Total and Photosynthetially Available Solar Irradiane at the Oean Surfae Journal of Geophysial Researh, Vol.94, No.C7, 1989, 9731{9742 [7℄ W.B.Rossow, R.A.Shier Advanes in Understanding Clouds from ISCCP Bulletin of the Amerian Meterologial Soiety, Vol.80, No.11, 1999, 2261{2287 [8℄ A.Hammer, D.Heinemann, A.Westerhellweg et al. Daylight and Solar Irradiane Data Derived from Satellite Observations - the Satellight Projet Dept. of Energy and Semiondutor Researh, Faulty of Physis, University of Oldenburg, D-26111 Oldenburg, Germany [9℄ J.Page Algorithms for the Satellight programme Tehnial Report, 1996 [10℄ A.Skartveit, J.A.Olseth, M.E.Tuft An Hourly Diuse Fration Model with Corre- tion for Variability and Surfae Albedo Solar Energy 63, 173{183 [11℄ M.Fontoynont et al. Satellight: A WWW Server whih provides high quality day- light and solar radiation data for Western and Central Europe Pro. 9th Conferene on Satellite Meteorology and Oeanography, Paris, 25{29 May 1998, 747{750 [12℄ H.Shiller Feedforward-Bakpropagation Neural Net Program ffbp1.0 GKSS 2000/37 ISSN 0344-9629 33