This is a technical supplement to the function nitpix, which describes an algorithm for estimating sea surface temperature from Advanced Very High Resolution Radiometer (AVHRR) data.
This section describes the methods employed to process the level 1 archive tapes into digital sea surface temperature imagery. The example below is for a particular project (C4S - the Central California Coastal Circulation Study).
The first processing step extracts all AVHRR spectral bands of interest, in a geographic domain of interest. For C4S the domain of interest was centered at 36.0 N, 121.5 W, extending over about 5 degrees of latitude and 6 degrees of longitude. This domain was selected to cover the coastal region from just west of Los Angeles to just north of San Francisco, and offshore for several hundred kilometers. The first processing step avin automatically determines the records, which must be extracted from the level 1 tape to cover this domain, and places the results into a TeraScan dataset. It is also at this stage that earth location of this dataset is computed using satellite orbit elements supplied on a routine basis by the U.S. Department of Defense. Earth location at this point is typically good to within a few kilometers, i.e. a few picture elements (pixels) of the AVHRR. At the same time, ancillary data are also extracted from the level 1 tape, which is used for radiometrically calibrating the AVHRR data, that is, converting the raw digital counts to per cent albedo (for the case of channels 1 and 2), and to blackbody equivalent brightness temperature in degrees Celsius (for channels 3, 4, and 5).
Only those channels that will be employed for sea surface temperature estimation are extracted from the level 1 tape at this first stage. As noted earlier, the daytime set of channels is 2, 4, and 5, while at night channel 3 is substituted for channel 2. This choice is modified for the case of NOAA-6 NOAA-8 and NOAA-10, where the AVHRR instrument did not include a channel 5. It is also modified to exclude channel 3 except for those periods where that channel was sufficiently noise-free to be of use. Channel 3 was only employed as a result during the first few months of 1984 with NOAA-7, and the first 7 months of 1985 with NOAA-9.
The second stage of processing involves a visual display of one or more channels of the extracted AVHRR data on an interactive color monitor, a peripheral device attached to the TeraScan system. The location of several readily identifiable land points, such as Point Conception, Monterey, Point Arenas, are permanently stored on the system, and displayed as a plus mark (+) on the monitor, at the location according to the earth reference data computed from the Defense Department orbit elements, spacecraft clock time, and the pointing angle of the AVHRR scanner. As noted above, these marks typically fall within a few kilometers, or pixels, of the actual land point viewable on the color display.
Interactive routines were developed to bring the reference marks into exact alignment with the actual land points. This is done by first adjusting the spacecraft clock time, which is known to be inaccurate by +/- 0.5 seconds, sometimes as much as 1 second. A one second clock error translates to a 6.6 km location error along the track traveled by the satellite. Following time correction, another set of corrections is made for inaccuracies in the attitude control system of the spacecraft. The system for maintaining the pitch, roll, and yaw is supposedly good to 0.2 degrees, which translates to about two pixels in the AVHRR. In practice, pitch variations are indistinguishable from clock error variations, and yaw errors appear to be small in most cases. Thus, small corrections for clock and roll error are normally sufficient to align the reference marks. On the TeraScan system this procedure only takes a few seconds.
The third stage of processing draws upon the ancillary calibration data to produce a dataset of radiometrically calibrated data in per cent albedo and degrees Celsius. This processing avcal is done in accordance with methods described in Kidwell [1985]. Prior to scanning the earth below, the AVHRR first views cold space. Following the earth scan, the instrument views the inside of its housing, within which are imbedded four platinum resistance thermometers. From this and prelaunch data, one may readily convert the raw count data to engineering units, that is radiance. The radiances may alternately be expressed as a blackbody equivalent temperature.
It is important to keep in mind that ocean viewing temperatures obtained at this point are not sea surface temperatures, but rather brightness temperatures as seen by the AVHRR viewing through the earth's atmosphere. Brightness temperatures are typically colder than actual sea surface temperature for a variety of reasons, as discussed in Bernstein [1982]. Most of the factors that come into play, namely clouds, water vapor, aerosols, departure of the sea surface from being a true blackbody, all tend to lower the brightness temperature by an amount that may range from one to several degrees Celsius or more, depending upon various atmospheric condition.
The fourth stage of processing attempts to eliminate cloud and other atmosphere-induced effects, to produce a dataset of actual sea surface temperature in those areas that are sufficiently cloud-free. The dominant errors and approaches to correct for errors are discussed in numerous papers. A selection of references includes Deschamps and Phulpin [1980], Bernstein [1982], and McClain et al. [1985]. The approach applied in this stage of processing the C4S data set most closely resembles that described in McClain et al. [1985].
In this approach, the data is subjected to a variety of tests to identify and eliminate cloudy areas. The blackbody temperatures of the infrared channels for the cloud-free areas are then combined as a linear weighted sum, where the weights have been predetermined through modeling and data studies to give the best estimate of sea surface temperature.
The tests are sequential; only those pixels that pass all tests are judged to be cloud-free. Failure to pass any test ends the testing sequence for that pixel, which is then assigned the value bad_value to identify it as being cloudy. The procedure then moves on to apply the same sequence of tests to the next pixel, and so on through the entire domain. The sequence of tests are somewhat different for day and night passes. Daytime passes will be discussed first.
The first test of a daytime pass is for view angle. The AVHRR scanner sweeps the earth to yield zenith viewing angles that range from 72 degrees at the swath edges to zero at nadir directly below the satellite. Experience has shown that data from zenith angles greater than 60 degrees yields poor results. The first test thus eliminates any pixels with zenith view angles greater than 60 degrees.
For the next set of tests, the pixel in question is imbedded at the center of a 3 x 3 array of its neighboring pixels. This unit array is examined in several ways, with the objective of determining if any spatially very small clouds, smaller than the AVHRR's pixel size, are present. Even a small cloud occupying only one tenth the area of the pixel, and say 10 degrees \uo\dC cooler than the sea surface temperature, would produce a brightness temperature 1 \uo\dC cooler than that of a cloud-free scene. To detect these so-called sub-resolution clouds, all nine elements of the unit array are examined, and the difference in channel 4 temperature between the warmest and colder elements computed. If this difference exceeds 0.45 \uo\dC, the unit array is judged to be cloud-contaminated, the pixel at the array's center is set to zero, and the 3 x 3 array is moved over to be centered at the next adjacent pixel. Clearly, cloud-free areas with sufficiently sharp frontal changes in sea surface temperature will also fail this test. Increasing the difference threshold would alleviate this problem, but at the expense of passing through more cloud-contaminated pixels. The choice of 0.45 \uo\dC was arrived at after trying other values, and represents a necessary compromise. The screened out frontal regions are generally quite narrow, and actually help by outlining and drawing attention to the frontal areas.
If the unit array channel 4 temperature min-max difference is less than 0.45 \uo\dC, the array is then subjected to a similar difference test on its channel 2 albedos. If the min-max difference exceeds 0.25% albedo, the array is again judged to be cloud-contaminated, with the array re- centered over the next pixel.
If the unit array passes both these channel 2 and channel 4 local difference tests, then the pixel at its center is assumed to be free of small clouds, as well as free of cloud-cover that would have cloud tops at varying altitudes and hence varying temperatures. Thus, these tests are very effective against most convective cloud situations. These tests fail however to screen out stratus cloud that uniformly covers all elements of the unit array. For daytime the uniform cloud test is quite simple: the average channel 2 value is computed for the array. Cloud-free arrays will have low albedo, and stratus-filled arrays very high albedo, typically around 15% or higher. The actual threshold for rejecting the array is however set at 5%. This value rejects all cloudy situations, and passes all cloud-free situations where the AVHRR is not viewing into the sunglint off the sea surface. In the sunglint areas it becomes difficult or impossible to distinguish presence or absence of cloud using channel 2. Setting the threshold at 5% eliminates the sunglint contaminated areas, as well as those covered by uniform clouds.
Daytime pixels passing the above tests are determined to be cloud-free. The logical next step normally would be to estimate sea surface temperature at the unit array's center pixel as:
[Eq. 1] SST(ctr) = a*T4(ctr) + b*T5(ctr) + c
where T4(ctr) and T5(ctr) are the center pixel's channel 4 and 5 brightness temperatures in degrees Celsius, and a, b, and c are the previously mentioned weighting coefficients. These coefficients are listed in McClain et al. [1985] for both NOAA-7 and NOAA-9. However, a and b have values of about 3.6 and -2.6. Even with effective cloud-screening, T4 and T5 of the center pixel will have uncertainties of about 0.1 \uo\dC. These weights will thus amplify this noise uncertainty by the square root of the sum of the squares of the weights, a factor of 4.4. This will result in an SST that appears quite noisy. To counteract this effect, Eq. 1 is regrouped to form:
[Eq. 2] SST(ctr) = a'*T4(ctr) + b'*{T4(array)-T5(array)} + c'
where T4(array) and T5(array) are now the 3 x 3 array average brightness temperatures. Clearly, b'=b, while a'=a-b. This reformulation takes into account the fact that the SST is very nearly equal to T4, and the second term in Eq. 2 is a small correction for atmospheric effects, primarily water vapor. Such atmospheric effects normally have large spatial scales, so averaging over the array will not change their value compared to their value for the center pixel alone. The array averaging of T4 and T5 reduce the noise uncertainty of individual pixel values by the square root of the number of pixels in the array, namely a factor of 3. This counteracts the noise amplification factor b for this term of Eq. 2, which as noted above is about 2.6. Application of this approach yields SST imagery of excellent quality and with noise characteristics nearly identical to the original channel 4 brightness temperature data.
The actual values of a', b', and c' employed for the C4S processing are derived from those in McClain et al. [1985], and are as follows:
a' b' c'
NOAA-7 1.0346 2.5779 -0.61
NOAA-9 0.9864 2.6705 +0.52
For night data, the sequence of tests begins as it did with the daytime data, with the view angle test following by the channel 4 min-max test. Since channel 2 is not available at night, channel 3 is critical to the screening of uniform stratus. For night data, the difference temperature diff34 = T3 - T4 is an excellent discriminator for distinguishing cloud surfaces from the cloud-free sea surface. This results from a difference in emissivity properties between the sea surface and clouds at the two different channel wavelengths. For channel 4, both cloud and sea surface emissivity is very nearly unity. For channel 3, the sea surface emissivity is also near unity, but cloud emissivity is much less, typically around 0.75. As a result, diff34 will be significantly negative for cloud surfaces. For sea surface viewing, diff34 is related to total amount of water vapor in the atmospheric path through which the sea surface is being viewed. For very moist conditions diff34 is significantly positive, while for moderately moist conditions diff34 is weakly positive. Under dry conditions diff34 is weakly negative, and for the very driest atmospheric conditions, diff34 can attain a value of about -1.0.
There does appear to be a slight overlap that can occur in which diff34 cannot be used to distinguish between cloud and sea surface viewing, when the atmosphere is extremely dry. This situation occurs in only an extremely small number of cases, however. The following figure schematically represents the manner in which diff34 may be used:
|--*--|
cloud.................
..dry........mod.........moist.. ocean
--+---------+---------+---------+---------+-----
-2 -1 0 1 2
T3 - T4 (\uo\dC)
In this figure, the overlap only occurs for a narrow range of diff34. However, as the uncertainty or noise grows in the AVHRR, especially in the noisy channel 3 problem noted earlier, the error bars grow around the actual estimated value of diff34, in this figure plotted at a value of -0.5, with error bars of +/-0.3 \uo\dC. As the noise grows and these bounds increase, it becomes increasingly difficult to separate cloud from clear conditions at night, unless the atmosphere has sufficient total water vapor. Yet the stratus found along the California coast is typically associated with subsiding dry air conditions, tending to produce situations where the clear sea surface diff34 will normally be in the range around zero or slightly negative. Thus, use of channel 3 for nighttime uniform stratus cloud detection requires a fairly noise- free channel 3.
In practice, when the channel 3 noise level has grown to exceed 0.7 \uo\dC rms, even 3 x 3 unit array averaging has not helped to reduce it significantly due to the fact that the noise signature tends to extend over more than 3 pixels. For the C4S data, only during the first three months of the NOAA-7 1984 data, and during the 7 months of 1985 NOAA-9 data, were the channel 3 noise levels sufficiently low to permit nighttime stratus screening. The NOAA-6 and NOAA-8 channel 3 noise levels also remained unacceptably high during their respective periods of data availability in 1984. As a result, nearly all of the nighttime data in 1984 cannot be screened for stratus clouds. Examples of such situations will be discussed in a following section describing specific images in the C4S set.
Finally, for the case of NOAA-6, NOAA-8 and NOAA-10, only channels 2 and 4 are available during daytime, and channel 4 alone at night. No multi-channel equation such as Eq. 2 can be used in these cases. For this situation, the following approach was employed. There is a rough tendency for the second term in Eq. 2, which is proportional to total precipitable water in the atmosphere, to be proportional to the sea surface temperature. The difference between the actual SST and T4 is thus found to be about 1 \uo\dC when T4 is around 10 \uo\dC, and about 2 \uo\dC when T4 is around 20 \uo\dC. For NOAA-6 and 8, the conversion of T4 to SST was simply to multiply T4 by a factor of 1.1, clearly a crude method, but at least invertible.
Once the data, whether day or night, has passed all screens, the result is a TeraScan dataset, bad_value in the (presumably) cloudy areas, and set to the computed SST elsewhere. The geometric distortion and earth location - line/sample relation is different from one overpass to the next. This is corrected by establishing a fixed grid, which is then populated by resampling the TeraScan dataset on a pixel by pixel basis. The simplest nearest neighbor algorithm is used. The result is a geometrically fixed set of final output arrays.
References
Bernstein, R. L., Sea surface temperature estimation using the NOAA-6 advanced very high resolution radiometer, J. Geophys. Res. 87, 9455-9465, 1982.
Chelton, D. B., R. L. Bernstein, A. Bratkovich, and P. Kosro, The Central Coastal Circulation Study, Trans. Am. Geophys. Un., 68, 12-13, 1987.
Chelton, D. B., A. Bratkovich, R. Bernstein, and P. Kosro, Poleward California during the spring and summer of 1981 and 1984, J. Geophys. Res., 10, 620, 1988.
Deschamps, P. Y., and T. Phulpin, Atmospheric correction of infrared of sea surface temperature using channels at 3.7, 11 and 12 um, Boundary-Layer Meteorology, 18, 131-143, 1980.
Atmospheric Administration, Satellite Data Services Division, Washington, D.C.
McClain, E. P., W. G. Pichel, and C. C. Walton, Comparative performance AVHRR)based multichannel sea surface temperatures, J. Geophys. Res., 90, 11,587-11,601, 1985.
None.
avcal, avin, avfix, avfilt3, etx, fastreg, master, nitpix.
None.
Last Update: $Date: 1999/05/10 21:18:27 $