Methodology to obtain accurate sea surface temperature from locally received NOAA-14 data in the Canary-Azores-Gibraltar area *

The sea surface temperature (SST) is an important geophysical parameter essential for quantitative studies of the Earth’s atmosphere and oceans. The SST imagery remotely sensed by the Advanced Very High Resolution Radiometer (AVHRR) aboard the NOAA satellite series has been conveniently used in various fields, for example, oceanography, meteorology, fishery, etc. The need for high accuracy in the measurement of sea surface temperature from space was recognised in the early stages of planning the World Climate Research Program. Absolute SST accuracy better than 1oC is necessary in the study of thermal structures of the sea surface (fronts, isotherms, etc.). With the objective to obtain accurate and automatic temperature maps, a methodology has been developed, that is currently operational in our Remote Sensing Center, including some improvements in the commonly used procedure. These procedures applied to the images from the AVHRR sensor aboard NOAA-14 consist of: (i) A procedure for AVHRR Infrared (IR) channels calibration using non-linear correcting functions based on those described by Steyn-Ross et al. (1992) and Walton et al. (1998). SCI. MAR., 65 (Suppl. 1): 127-137 SCIENTIA MARINA 2001

(iii) A new split-window function (SWF) to perform atmospheric and emissivity correction in CANIGO area.This has been accomplished by using the matchup data set of AVHRR brightness temperature (channel4 and 5) and in situ SST data Coefficients are estimated from regresion analysis using 60 co-located in situ and satellite measurements (rnarcnupsj ana examining Uie formai ciependencies of the variables TI, and T,, , where TI, and TI, are the radiative temperatura at 11 and 12 p respectively, restricting in-sini measurements to be within about 50" of the satellite zenith angle. (iv) An automatic procedure for geomehic correction based on a satellite orbital model and a similarity detectia algorithm.This procedure uses the automatic identifícation of coastal features in a digitised coastline map and in the georeferenced image, achieving accuracies to the order of the pixel.
In the next section the theoretical base and a practicd applicaiion of the procedure have been detailed.That radiance is converted to brightness temperature, that is the rnagnitude of interest, using the inverse Planck function with the appropriate central wavelength nurnbers that are provided by NESDIS.

AWOMATIC IMAGE GEOREFERENCING
The characteristic of the satelhe scanning, which covers a 2800 km wide swath, the spacecraft's speed, altitude and amtude, and the Earth's c w ature and rotation, produce signidcant distortions in the images.It is necessary for irnage referencing to identify the geographic co-ordinates comesponding +A nm Z r n n r r a &val /AL--+ rag~-r;nn\ rrr +r\ 1 w -t -9 iu aii r i h i a ~; r .y-L., \ u u u ~ 1rribiruibui6, v i ~v lurvru u pixel comsponding to given geographic ceordinates (inverse referencing).
A good o v e ~e w of existing methods for geomemc correction of saíeiiiíe data Is gheñ iti Krasnopolsky andBreaker (1994) andRosborough et d. (1994).In the referentes several models of varying comp1exity can be found (Ho and Asern,  1986; Bachmann and Bendix, 1992, Moreno and   Meiia, 1993; lllera er aL, 1996).In general the procedure for the georeferencing of NOAA-AVHRR images is based on the combination of an orbital model and ground control points (GCPs).
The objective is to combine the advantages of both methods to obtam a procedure fdly automatic and ogmatiod which aiiows tiie gwrreferencing of NOAA-AVHRR images wirh high a ~~~~l a c y and without operator intervention.
The orbital moded assumes a circular orbit (Ho and Asem, 1986) and the nominal Kepplerian elements, as weii as the time of the first scanned h e , the longitude of the ascending node and the equator mssing hour.These parameten can be obtained, automatically, from the NOAA-Leve1 li3 data set heder and from the weekly archive of the equator crossing file, respectively.In order to c o m t the errors caused by these simplifications, by a non-zero value for the spacecraft's roll, pitch and yaw agd by

The Sarellite Data
The AVHRR raw data cm NOAA-14 is daily received at our station at the Remote Sensing Centre, Univasity of Las Palmas.Figure 6 shows the sysmmic procedure perfonmed to the images to obtain the matchup database between the in-situ SST measu~ements and the nearlycoincident sateliite observatíons.
With-the objcctive to obtain very high quality data to estimate the atmosphcric model coefficients, the foiiowing procedure was fo~awed:.

=-e---1UtXI
Il*lUrt;.detection algorithm, as described in last section.In order to achieve sub-pixel precision, the sub-scenes were autornatically corrected using ground control points iocated sawegically closer to the in-situ measurement co-ordinates.

Visual inspection of the
4. Cloud detection: the algorithm developed is a mutiband threshold method based on the multitest systern implernented by Saunders and Kriebel (1988).The method was adapted to our system including some modifications for the Canigo geooraphical zone based on experimental data.The w flow diagrams of the day and night cloud detection algorithms are presented in Figure 7 and will be briefly described next.-4,: this test is applied in &y irnages, after the previous ones.In cbannel 2, cenfred at 0.9 pm, the sea reflectivity has a value of less than 6% while for clouds is much higher.In our algorithm the visible threshotd depends on the solar zenith angle and the düEerence T3,7*-T:jiiñi.-Gross cfoud test: it is a simple thexmal infiared threshold test using the brightness temperature calculated from channel 4. In our case the threshold has h e n fixed to S 283X to consider cloud presence.
Coman.-The 2 common tests included ui the algorithrn are: -T~ i , " ñ T i ~ test: tbis test is used to detect cims, whose difference between the ttiermal IR c h e l s is high.Saunders and Kriebel (1988)  split-window algorithm.For our data set we have plotted Th-T,, vemis TI,-T,,, appearing that a quadratic regresion seems to be more appropriate than a linear one, especialiy for higher values of TI,-TI,.
Our algorithm includes the mentioned ideas and it has the following expression: The results are presented in Figure 8 and Tabie 2. The avemge errors are ob~iously, about h e data to obtain the algoríthm, 0°C and the dispersion is only  Database, that includes a fairly large collection of highquality in situ SSTs with a reasonably wide dismbution in space and time.Also the results obtained wouici Liave to ' be compad wiui ouier algorithms that use coefficients dependent on satellite zenith angle.

SST EsOmation Fnnctions
We wish to express our gratitude to the Educational Council of the Canary Islands Government (Contract PI1998/06f3), and to the CICYT-Commis- The st;rfce ter??pe=~y~e (SST) is m i q mtant geophysical parameter essentid for quantitative studies of the Earth's atmosphere and oceans.The SST imagery remotely sensed by the Advanced Very Eigh Resoiution Radiometer (AVEREj aboard &e NOAA satellite series has been conveniently used in various fields, for exarnple, oceanography, meteorology, fishes: etc.The need for high accuracy in the measurement of sea surface temperature from space was recognised in the early stages of planning the World Climate Research Prograrn.Absolute SST accuracy better than 1°C is necessary in the snidy of the-d stn'ctiios nf fhe seir S ~~Z C P (frmts, isotherms, etc.).With the objective to obtain accurate and automatic temperature maps, a rnethodology has been developed, that is cmntly opemtional in our k m ü i e Semhig &ñW, iñdudhg S ú m improvements in the commonly used procedure.These procedures applied to the images from the AVHRR sensor aboard NOAA-14 consist of: (i) A procedure for AVHRR Infrared (IR) channels calibration using non-linear correcting functions base.on those described by Steyn-Ross et d.(1992) and Walton et al. (1998).METHODOLOGY TO OBTAM SEA SURFACE TEMPERATURE FROM NOAA DATA 127 (ii) An improved algonthm for cloud detection based on the method of the multiband thresholds adapted to the CANIGO zone after experimental checking.
I C D UULU uu UIG auwairy UL UIG ~ucusured radiances and the accuracy of the SST retrievai algorithrn that converts the measured radiantes to sea surface temperature.Thus, it is of great irnportímce ulat the aigoriulms UiaÍ invoive i n h d rneasurernents get precise values for the d a n c e .This fact requires a welldehed and consistent calibration pmcedure, understood as the radiometer output &&al vaiues (or count) relative to the actual radian= of the observed scene.In principie, if the radiometer was perfectly stable, any necessary calíbrations could be canied out on ground prior &o the launch, obiaining a precise relation between the counts and the radiance.However, because the characteristics of the m-board sensors and electronics may dríft with time and esvimnmental changes, a dynamic calibration produre is required.Tñe AVHRR IR channeb calibration m flight is performed by the monitoring of the radiometer output when it views two targets whose radiance vaiues are known or can be computed: cold space as a zero-radiance reference and an interna1 calibration blackbody target (ICT) located in the instniment base plate and whose temperanire is measured with a four platinim-resistance thermometers (PRTs).These two measurements are used to define a continuously updated linear mapping between count and radiance.For the NOAA-14 AVHRR sensor, a procedure based on the ones described by Steyn-Ross et al. (1992) and Waiton et al. (1998) is used.This procedure cúrrects the insirumental effects of the AVHRR thermal infrared channels centered at 10.8 pm (channei 4j ano i2 pm (channei 3, whose Eg-Sd-Te detectors exhibit slight but well-defined non-linearities in their response to incoming radiation.It assignes a non-zero radiance to fhe cold space view response as proved by Waiton et al. (1998).
,G , , , , ., , , , ., , , dL-' Y detection algonthm is used.Systematic errors in the Orbital Model , The prob1em of georeferencing consists of finding the relation between the co-ordinates in the image &e, colurnn) and the geographic m r d inates (lat, 10x1).This wiii allow us to geographically locate any pixel in the image (direct reference), or to identify the corresponding image co-ordinates on an AVHRR image, given a set of geograpbic coadinates (inverse reference): This is calculated by rnean of the orbital model foIiowing the procedure described by Ho and Asem (1986) (merse referencing).The basic relations for the orbit, the scanner and the Earth are similar to the ones described by Ho and Asem (1986).However, udike this rnodel that adjusts satellite height (hs) and inciination (r] by means of a GCP and the Bachmann and Bendk (1992) ami 111era et d(1996) models, that calculate the longitude of the ascending node (AE) and the time difference between the quator crossing and the scanned h e (t, ) using one or several GCPs, we introduce nomiaal values (i, h, ; l , , t,).The use of nominai values, especially the sate1lite's height and inclination, longitude of the ascending node and the equator crossing hour, as weii as the simplifications used considering a circular orbiit modeI around a sphencal Earth brings us to the unavoidable errors when the obtained image is superimposed on a digitised coastline.En order to carry out an analysis of the possible errors, we measure on the coastline and calculate, by means of the orbitai model, the coordinates of a group of control poinrs, which are used as a test.Results obtaíned using exclusively the orbitai model to correct the NOAA 14-AVHRR image comsponding to 5 of June 1998, with a size of 1400 columns by 1032 rows are shown in Figure 1.
Figure l(a) shows that the error in the nadir is around 1 pixel instead of zem as a result of thc trans-FG.1.b r s baween the m c a s d coordinates in the digitised coanlinc and aKMe caicukatcd by mean of the ~&tai modcI for a set of GCPs.(a) As a funaion of the columns, and (b) as a functwn of tbe lines.lation due to the nominal value of &.The errors in the rows are only of around 3 pixels, basically due to the variation generami by the sateEte clock cirift and the nominal value of the satellite inclination.The most important misaiignrnent is generated by the errm in rhe sateUite altitude, which is unavoidable when using a cirailar orbit.As can be appreciated, the error hcreases when we move to the edges of the swath.If the nominal sateliite aItitude estimated is higher or lower than the actual for the orbit under consideration, the error will be positive to the east or west ~spectively.Finally, h e errors do not depend on the h e number as can be observed in Figure l(b).Automatic procedure of georefemncing We have developed an algonthm that allows the elimination of the e m analysed in the previous METHODOLOGY TO OBTAIN SEA SURFACE TEMPERATURE FROM NOAA DATA 129 section.This procedure for the automatic NOAA-AVHRR images georeferencing shows four clearly separated steps: l .' h e irnage or subscene georeferencing is performe-by rneans of the orbital model using as input the nominal kepplexian elernents and the n d parameters from NOAA. 2. A sirnilarity algorithm is used which identifies autornatically coastal features in the coastline-digitised map and in the georeferenced image, using a window correlation of 11 pixels search radius.The similarity measurement used, unlike Bordes et al. (1992), is the cross correlation coefficient, where i is the shift in mws, j the sbift in columns, R and G are the reference and gradient windows and pR, pG are the means insiáe the considered windows.This coefficient is a good indicator of the two windows spatial registration.3. The values obtained, afta passing the diíferent local and global consistency tests, are used for a bilinear regression anaiysis similar to the usual GCP method, Lin '=ao+a,.Lin+a, Col Col'=bo+b,.Lin +b, Col (3) 4. Given known mapping functions h e georeferenced image, obtained by the orbitai model, is transformed and resampIed usiog a zero ordar híarpo1ation ttchnique.The global scheme of the procedure is shown in Figures 2. Figure FIG.2.-Global pxcdm for tbt automatic g#>~cfsrci)cing of NOAA-AVHRR images.
~f thp. 5 A W Rw.nsor chaanels.C'hsnnels 1 and 2 asing a linear COIrection in accdance with ihe NOAA and channe1s 3,4 and 5 by means of a non-linear calibration, as > --i L -2 intages: the selected imagcs matching ín time with the in-situ measurements were carefdly analysed by expert oceanographers by means of a high resolution monitor on its different & ~e i s with he objective to disregard those images whose areas closer to the co-ordinates of the in-situ measuremtnts showed: RG. 6. -Systcmatic piocedure to generate tfie matchup data set -Cloud contamination.-SST artefacts, such as warm patches (due to diurna1 heating) or sand clouds.3. Images were geornetrically corrected accord-. .mg tv *U:e G i * & I i i eeuLi Li d aild ,aíTii:*lyj.

Night. -
The 3 night tests included in the aigorithm are very simple due to the use of 6xed thresholds.-T,,,-T,,, test: is vexy useful to detect clouds oi iow anci m d u m aititucie, b a s d on rhe spectrai emissivity variation in the water clouds.Here we make a distinction between the water clouds and the ice particle ciouds because they have different physicd propemes.In our case the threshold has been fued to T,,?-T, 2 5OK to consider cloud presente.An additional test is included to detect clouds in desen areas: T , , + % -Tlzi, 2 02°K -T 9.7p -T ,* test.this --nieht test ---tries ----!o det~tf semitransparent clouds composed of ice particles (cims).The thteshold decided is T 3.,w -T,,w 2 3°K to consider cloud presence.P-----l-.-.l *-,.*.*Le 1 :-:-*LA J-C---U I V U J GIVUU LCSC.U1G LGY Y U i i l L 15 111 UIG UGllltition of the appropriate threshold temperature.Over the sea it is relatively straightforward as the SST varies slowly in space and time.Over land, however, the large day-to-day and ara-to-area variability in surface temperatures due to the different land uses and meteorological conditions makes the definition of a single overall discrimíaation threshold temperature much more difficult.h order to consider cloud presence fhe following test has been used: Day.-The 3 day tests included in the algorithm are: -tie --&-, t ge~eTQ+es a considerable increase in the channel 3 detected radiance.So, if the temperature Merence between diannels 3 and 4 is very high, we can assume riiat is aue to the sungiint.I n e esta~iisneci threshold is T,,, -T,,, > 25°K.

Fig. 8 .
Fig.8.In order to get an aigonthm in the Canigo are* we have exarnined the formai dependencies of the variables TI, md T,,.In both algorithms the T,, coefficient is equal to 1, inducing a bias related to the increment of the atmospheric correction when the SST increases and was initially detected comparing the satellite temperatures and the real measurements w i t h buoys (McCiain et al., 1985).AisoBarton (1995) points out that regional algorithms trend to have a coefficient close to 1 for TI,.McMillin et d.(1984), Coll et al.(1994)  and Fran~ois et aL(1995)  demonsmte that the atmospheric comaion should be provided in terms of a second order pohomy in T,,-T,?.Barton (1995) mmtioned the improvements in the FIG. 9. -Canigo arca SST map resulñng frorn quation (6) for the image of 5 Jum 1998 rakm by NOAA-14 (orbit nurnber 17680).Land and clouds are masked m biack and white rrspoctively.