Optical Remote Sensing

Remote sensing in the optical part of the electromagnetic spectrum (400-2500 nm) involves measuring the combined reflectance of solar radiation from both the Earth's surface and atmosphere. There are several specific definitions of reflectance referenced in Table 13, but in general, reflectance involves the diffuse scattering of light by a geometrically complex surface (Hapke 1993). Reflectance varies according to the degree of collimation, the incident irradiance, and the collimation of the detector. Collimation refers to the degree of angular diffusion of the the incident light or the size of the angular field of view of the detector. For instance, direct beam solar energy is considered highly collimated whereas the diffuse sky radiance is uncollimated; a narrow-angle field of view detector can be considered collimated while a hemispherical sensor such as a pyranometer is an uncollimated detector (Diner et al. 1999).

The percentage of incident solar radiation reflected back by an object is called albedo. Surface objects with a high albedo reflect more solar energy and appear as bright objects in a visible/near-infrared image. Objects with a low albedo reflect only a small portion of incident solar radiation and appear as dark objects on a visible/near-infrared image. Figure 5 summarizes how variations of reflectance depend upon the degree of collimation (Hapke 1993).

Figure 5. Variations of Reflectance

AVHRR Channels 1 and 2 measure the upwelling radiance at the Top-of-Atmosphere (TOA) emanating from both the Earth surface and atmospheric scattering. TOA radiance is converted to a TOA BRF using the the exo-atmospheric solar irradiance, Earth-Sun distance, solar zenith angle, and spectral response functions of the AVHRR channels. The surface albedo is obtained from the TOA BRF after applying an atmospheric correction and a conversion from directional reflectance to hemispherical reflectance. Figure 6 shows a summary of the solar and satellite angles used in processing the AVHRR Polar Pathfinder products.

Figure 6. Summary of Solar and Satellite Angles Used in Processing AVHRR Polar Pathfinder Products

Channel 3A on the NOAA 16 AVHRR/3 instrument allows improved discrimination between snow and clouds by using the 1.6 µm wavelength. At 1.6 µm, snow has very low reflectance, while the reflectance of clouds remains high. Refer to Figure 7. Therefore, both cirrus and optically thick clouds can be directly classified and distinguished from snow at the 1.6 µm wavelength (Warren 1982).

Figure 7. Satellite Channel Wavelengths in Micrometers (µm), and Typical Reflectance Spectra for Snow and Clouds

Thermal Remote Sensing

All objects radiate energy according to their blackbody temperature. A blackbody is a hypothetical object that absorbs all incoming thermal energy, but with none of that energy reflected or transmitted. Since no object can ideally absorb 100 percent of incident energy, a blackbody serves as a comparative measure of thermal emission. As the temperature of an object increases, the total amount of the emitted energy also increases, and the wavelength of that energy becomes shorter, as Wien's displacement law describes. According to the Stefan-Boltzman law, objects with a higher temperature give off more thermal energy per unit area than objects with a lower temperature. Remote sensing in the thermal infrared spectral region (approximately 7 µm to 14 µm) involves measuring the radiance of objects. More specifically, thermal remote sensing measures the differences in the ability of objects to absorb shortwave energy and emit it back as longwave energy. Researchers are primarily interested in measuring the actual temperature of objects, rather than radiance. Radiance is simply a measure of the emitted energy of an object, while temperature is a measure of the kinetic (thermal) energy of an object. AVHRR channels 3-5 measure TOA brightness temperature. Refer to the NOAA Polar Orbiter Data User's Guide for details on how radiance values are converted to surface temperature, and the Derivation Techniques and Algorithms section of this document for algorithms used in the AVHRR Polar Pathfinder Twice-Daily 5 km EASE-Grid Composites.

Clear sky surface skin temperature, albedo, and cloud masking are all derived from the Cloud and Surface Parameter Retrieval (CASPR) system.

1. Channels 1 and 2
   Calibration of the shortwave visible and near-infrared channels of the AVHRR (channels 1 and 2) is originally done according to Rao and Chen (1994), which corrects for sensor drift over time. Channels 1 and 2 are normalized for mean Earth-Sun distance.

1. Channels 3-5
   The calibration of the thermal Channels 3-5 applies non-linear functions to the satellite sensor counts to obtain brightness temperatures as described in (Walton 1998). See the Calibration section for more information. The Version 3 data uses data collected from the NOAA-16 AVHRR/3 instrument, which contains the dual Channel 3 for better cloud detection over snow- and ice-covered surfaces. 
   
   For the NOAA 16 AVHRR/3 instrument, Channel 3A collects data during the day at 1.58-1.64 microns, and the values are reported in units of reflectance (0-1000). The Channel 3A reflectance is obtained by first using the Channel 4 emission temperature to estimate the thermal emission component to the total Channel 3 radiance. This thermal emission component is subsequently removed from the total radiance, leaving only the solar reflectance component (Setvak and Doswell 1991).
    
2. Clear Sky Surface Skin Temperature
   Two algorithms are employed to convert brightness temperature to clear sky surface skin temperature: one for high-latitude ocean and snow-covered land, and another for snow-free land (Key 1999). For retrieval of clear sky skin temperature, a simple regression model is used to correct for atmospheric attenuation. Refer to Figure 10.
   
   
   
   Figure 10. AVHRR Polar Pathfinder Product Generation 
   
   
   
   For high-latitude ocean and snow-covered land, the following equation applies: T_S = a + bT₄ + c(T₄ - T₅) + d[(T₄ - T₅)( sec{q}- 1)]
   
   Where:
   T_S = surface temperature
   T₄ = brightness temperature in channel 4
   T₅ = brightness temperature in channel 5
   q = sensor scan angle
   a, b, c, and d = regression coefficients
   
   To determine the empirical relationship, radiosonde data from drifting ice and land stations in the Arctic and Antarctic were used with a radiative transfer model to simulate sensor brightness temperatures. Surface temperature retrieval methods for sea ice and snow, snow-free land, and open ocean are described in detail in Key et al. 1997. The coefficients are a function of temperature range, hemisphere, and the NOAA satellite that acquired the data. Algorithms are provided in the Surface Temperature Source Code Document.
   
   Bare land is treated separately with:
   
   T_S = a + bT₄ + cT₅ + e*emiss4 + f*emiss5
   
   Where:
   emiss4 = surface emissitivity in channel 4 (0.985)
   emiss5 = surface emissitivity in channel 5 (0.975)

1. Clear Sky Surface Broadband Albedo
   Refer to Figure 6. The retrieval of surface albedo involves four steps:
2. normalize channels 1 and 2 with respect to solar zenith angle
3. convert channels 1 and 2 narrowband reflectance to a broadband reflectance
4. correct the TOA broadband reflectance for anisotropy
5. convert the TOA broadband reflectance to a clear sky surface broadband albedo.

The general methodology in Steps 2 through 4 was used by Csiszar and Gutman (1999) for global land studies. The albedo provided here is a directional-hemispherical, apparent albedo, where apparent albedo is measured by up- and down-looking radiometers in the field.

Step 1: Normalize Channels 1 and 2 with respect to the solar zenith angle:

P_1,toa= C₁ / 100 * cos(zen)

P_2,toa= C₂ / 100 * cos(zen)


Where:
C₁ = percent reflectance for channel 1

C₂ = percent reflectance for channel 2

zen = solar zenith angle

p_1,toa = channel 1 reflectance

p_2,toa = channel 2 reflectance


Step 2: Convert the narrowband reflectances in Channels 1 and 2 to a broadband reflectance. The narrow-to-broadband conversion takes the form:

p_toa = a + bp_1,toa + cp_2,toa


Where:
p_1,toa = channel 1 reflectance

p_2,toa = channel 2 reflectance

p_toa = broadband TOA reflectance

a, b, c = regression coefficients (Refer to Table 16)

To develop the regression relationship, the radiative transfer model Streamer is used to simulate the TOA reflectances over a broad range of viewing and illumination angles, atmospheric conditions, and surface types and albedos (Key and Schweiger 1998). Separate sets of coefficients are determined for different surface types. Refer to the Atmospheric and Anisotropic Reflectance Correction Source Code Document for details.

Step 3: Correct for the dependence of the sun-satellite-surface geometry on reflectance. This is done with data presented in Suttles et al. (1988) who used the Earth Radiation Budget Experiment (ERBE) and Geostationary Operational Environmental Satellite (GOES) data to determine TOA Anisotropic Reflectance Factors (ARFs) for the broad shortwave band over various surfaces. To convert the directional reflectance to albedo, the ERBE/GOES ARFs are used:

a_toa = p_toa / f

Where:

p_toa = reflectance observed at the sensor simulated by Streamer in Step 1

f = anisotropic reflectance factor a_toa = TOA albedo. The f factor is derived from a tri-linear interpolation of two tables: rmatrx and albmn.

rmatrx: Anisotropic reflectance values based on ERBE and GOES. Dimension is (3,10,7,8) with three scene types, ten solar zenith bins, seven viewing zenith bins, and eight relative azimuth bins.

albmn: Normalizing factors for albmn. Dimension is (3,10) with three scene types and ten solar zenith bins.

Preliminary results show that under certain circumstances, the reflectance of mixed pixels of open ocean and ice could have incorrect values. A Channel 1 value less than 0.0 indicates open ocean, and a value greater than 0.3 is pure ice. Values between these limits indicate pixels with mixed surface types. The final value is a weighted average of the two. Algorithms are provided in the Atmospheric and Anisotropic Reflectance Correction Source Code Document.

Step 4: Finally, the apparent clear sky surface broadband albedo is estimated with a regression relationship of the form:

surface_albedo = (a_toa - a) / b

Where:

a and b are a function of water vapor, aerosol amount, and solar zenith angle. The coefficients were determined with Streamer for a variety of surface and atmospheric conditions.

For open ocean, a simpler approach is taken:

surface_albedo = a + b*a_toa + c*cos(zen) + d*pw + e*aertau

Where:

a, b, c, d, e = coefficients determined empirically using modeled albedos
pw = precipitable water
zen = solar zenith angle
aertau = atmospheric aerosols

Details are provided in the Atmospheric and Anisotropic Reflectance Correction Source Code Document. 

For both cases, the aerosol optical depth is set to 0.06. Also, for both cases, the water vapor is estimated from Channels 4 and 5 using the formula:

PW = exp[b0 + b1(T4-T5) + b2(T5)] cos(theta)

Where:

pw = precipitable water
theta = scan angle
b0, b1, b2 = coefficients determined over a range of surface temperatures and water vapor amounts using AVHRR radiances modeled with LOWTRAN-7. Arctic rawinsonde data were employed.
b0 = -10.4974 
b1 = 0.751008 
b2 = 0.0453005

Details are provided in the Retrieval of Precipitable Water Source Code Document. The calculated surface albedo is an apparent albedo, one that is measured with radiometers and which varies with changes in atmospheric conditions, particularly for bright surfaces.

5. Surface Type Mask
   
   Refer to Figure 11 for a flowchart that summarizes the steps in generating the surface type mask. 
   
   
   
   Figure 11. Surface Type from Passive Microwave Data 
   
   
   SSM/I brightness temperatures, regridded from the daily-averaged SSM/I Daily Polar Gridded Brightness Temperatures data set to the 25 km EASE-Grid, were first partitioned into land, including ice sheets or ocean. Over oceans, the NASA Team Sea Ice Algorithm was used to generate first-year and multi-year ice concentrations. For the purposes of the cloud detection procedure, all ocean pixels with an estimated ice concentration greater than zero percent were flagged as sea ice. In some cases, atmospheric effects and wind roughening of the ocean cause false ice concentration retrievals over open ocean. Such areas will appear as incorrectly mapped areas of sea ice in the mask; however, for cloud detection, preliminary work suggests that it is important to avoid excluding areas with less than 15 percent true ice cover. Areas that consist of at least 50 percent multi-year ice are assigned the multi-year flag. The multiyear ice estimate is subject to considerable error and uncertainty, particularly during periods when surface melt is likely. At such times, characteristic multiyear ice emission properties are partially or completely hidden, and melt-freeze metamorphosis or other changes in the snow cover on sea ice may cause first-year ice to appear as a fraction of multi-year ice. Since thresholds were applied to the output from ice concentration, ice type, and snow depth algorithms, slight variations in the values yield a noisy appearance in the resulting binary-flag images, particularly in spring and summer due to algorithm limitations.
   
   Over land, snow cover is detected with an SSM/I algorithm described in Goodison (1989). A gradient value (19 GHz horizontal polarization minus 37 GHz horizontal polarization) greater than zero indicates snow, and the pixel is assigned a snow surface type. No special thresholds are applied, nor are special cases such as melting snow addressed in this mask. The SSM/I data have different coverage areas, and thus the regridding process of the SSM/I leaves some areas without information over certain regions. The Near-Real-Time SSM/I EASE-Grid Daily Global Ice Concentration and Snow Extent and Northern Hemisphere EASE-Grid Weekly Snow Cover and Sea Ice Extent products from NSIDC were used to fill in the missing regions wherever possible.
   
   The resulting 5 km mask is over-sampled from the 25 km SSM/I data. Because coastal areas have a jagged appearance, a 5 km land mask is applied to the surface mask, and the nearest appropriate surface type mask is used.
   
   The SSM/I-based mask is not intended to serve as a substitute for surface information from standard products such as the formal SSM/I Pathfinder snow or sea ice concentration data sets. It should only be used as an indication of the general surface conditions. Users interested in detailed snow and sea ice parameters should consult the standard products listed above. Table 17 lists the cell value for various surface types. 
    

   Table 17. Cell Values for Surface Types

   Cell Value	Description
   10	Open water
   20-29	Predominately first-year ice
   30-39	Predominately multiyear ice
   40	Bare land
   50	Snow-covered land
   60	Ice sheet

Ranges for first-year and multiyear ice types indicate the predominant ice type and the total ice concentration in tens of percent. For example, a value of 24 means that first-year ice predominates and that the total ice concentration is between 41 and 50 percent.

6. Cloud Masks
   
   Three cloud masks are provided since each method has advantages under certain conditions. One is based on multi channel, multiday techniques in the CASPR algorithms; another is obtained by using a long time series of Channel 4; and a third is a modified method of replacing the Channel 4 statistics required in the CASPR algorithm with the Channel 4 time series from the second method. Cloud masks are created separately for the morning and afternoon composites. Users can choose which combinations of cloud masks to use; thus, providing some control over the confidence in the presence of clouds. This may also provide some insight into the types of clouds as well since each algorithm has somewhat different sensitivities.

The above tests are done in a clear-conservative mode, meaning that the threshold is set so that values are more likely to be clear. Next, the following three steps are done:

The only distinction between surface types is land and ocean. Land may be snow-free or snow-covered. Ocean can be open water or sea ice. Additionally, the nighttime procedure does not incorporate any surface type distinction. Comparisons with SHEBA surface observations of cloud amount (synoptic observations every three hours) yield a bias of 0.1 and a root mean square error (RMSE) of 0.3.

1. CASPR Method
   
   CASPR has two modes for cloud masking. The first operates on data from a single day using a combination of spectral tests and thresholding and can be adjusted for greater or less sensitivity to clouds. The second method also uses the results from the single day cloud masking in a cloud conservative mode, but comparisons are also done between consecutive days to ensure clear pixels. A time series of mean values for each channel is generated, and final cloud masking is performed by thresholding the data with the mean values. The multiday method is used with Versions 1 and 2 data (bit 0), while single-day and multiday methods are both used with Version 3 data (bits 0 and 1, respectively). The following discussion describes the details of the CASPR algorithm used to produce the Version 1 and Version 2 bit 0 cloud mask. The algorithm was somewhat modified for Version 3, and is described in CASPR User’s Guide, Part II: Reference.
   
   Cloud detection methods in CASPR are optimized for polar conditions, where cloud screening is complicated by the low contrast between clouds and surface in reflected and thermal bands, and by the lack of solar illumination at night and for part of the year. The basic detection procedure involves a combination of split-window and single channel thresholds and temporal change detection using estimates of clear sky surface temperature and albedo determined separately as a function of surface type. SSM/I data are used to partition the surface into open water, ice, snow-free land, and snow-covered land with supplementation by a high-resolution coastline and an AVHRR-based surface type within 50 km of a coast. This allows tighter optimization of the thresholds since the general surface type is known.
   
   The cloud masking procedure consists of thresholding operations based on modeled sensor radiances. AVHRR radiances are simulated using the Streamer algorithm for a wide variety of surface and atmospheric conditions, and values that approximately divide clear from cloudy scenes are determined. The daytime and nighttime algorithms have very different logic. In the presence of sunlight (daytime), spectral cloud tests work relatively well because false detections labeling clear pixels as cloudy are infrequent. The daytime algorithm uses only spectral tests and is thus conceptually simple: initialize the cloud mask to clear, then apply the cloud tests to label cloudy pixels.
   
   In the absence of sunlight, the procedure is different because there is less spectral information available, and because AVHRR Channel 3 tends to be noisy at low temperatures, which is common during the polar night. Under most conditions, optically thick clouds simply cannot be distinguished from clear sky with the three thermal channels alone. The nighttime algorithm first applies the spectral cloud tests, resulting in some false detections for noisy pixels and for optically thick clouds. The spectral test thresholds are set to be clear-conservative to increase the likelihood that those pixels labeled as clear are in fact clear. Next, a clear sky channel 4 temperature image is constructed by passing a mean filter over the clear pixels, eliminating much of the noise and false detections. Another threshold test (based on the clear sky map) is applied to the image, such that temperature deviations greater than the threshold from the clear sky mean signify that a pixel is cloudy. In this final test, it is possible for the cloud to be either warmer or colder than the surface, especially during the winter.
   
   Most of the cloud test concepts are found in the Air Force SERCAA procedures (Gustafson et al. 1994). Some appear in the NOAA CLAVR algorithm (Stowe et al. 1991), while most were developed and/or used elsewhere. In the discussion below, warm cloud refers to clouds with higher temperatures than the surface, while cold clouds have temperatures lower than the surface. The cloud tests are as follows:
   < >Cirrus test (11 µm - 12 µm, day and night): A pixel is cloudy if BTD45 is greater than CCT_TABLE, where BTD45 is channel four minus channel five brightness temperature difference, and CCT_TABLE is a table of thresholds from (Saunders and Kriebel 1988) extended to include lower temperatures (range: 230-310 kelvin). The thresholds are the expected maximum BTD45 for clear sky.Warm clouds test (11 µm - 12 µm, day and night): For warm clouds BTD45 may be negative, so an additional test is required (Yamanouchi et al. 1987). A pixel is considered to be cloudy if BTD45 is less than WARM45. This test is most applicable at night in the presence of a temperature inversion.Water cloud test (3.7 µm, day only) - A pixel is considered cloudy if REF3 is greater than REF3_THRESH and REF1 is greater than REF1_THRESH, where REF1 and REF3 are channels one and three reflectances, and REF1_THRESH and REF3_THRESH are threshold values.Low stratus and thin cirrus tests (3.7 µm - 11 µm, night only): A pixel is cloudy if BTD34 is less than or equal to LSTTCI_34LO, or BTD34 is greater than or equal to LSTTCI_34HI; where BTD34 is the brightness temperature difference between channels three and four.Examine the time change in spectral characteristics between pairs of days. A pixel is clear if the spectral tests say it is clear, and if it varies little on a daily basis. The Channel 4 brightness temperature and the adjusted brightness temperature difference (4 minus 5) are used in the time tests when any part of the image is dark. The Channel 4 brightness temperature, Channel 1 reflectance, and Channel 3 reflectance are used during the day. 
    
2. Clear sky statistics are compiled for each pixel using all clear values within a distance of the nine pixels and over five days. Means and standard deviations of the Channels 1 and 3 reflectances (day only) and the Channel 4 temperature (day and night) are computed. If a cell occurs along a coastline, only pixels with the same surface type (LAND or NOTLAND) as the center pixel are used. Pixels in the clear sky composite that did not have means and standard deviations assigned, due to the lack of clear values, are filled by bilinear interpolation.

1. The clear sky composite statistics are used in a threshold operation to label or relabel all pixels as clear or cloudy. A pixel can be labeled as cloudy if its brightness temperature is higher than its mean clear value plus some number of standard deviations; or if its reflectance in Channel 1 or 3 is higher than its mean clear reflectance plus some number of standard deviations. Similarly, a pixel can be labeled as cloudy if its brightness temperature is lower than its mean clear value minus some number of standard deviations. If any part of the image is dark, the nighttime final thresholding is calculated for the entire image, though both the daytime and nighttime spectral cloud tests are used in the initial labeling. Refer to Step 1.
2. Time Series of Channel 4 Data 
   
   Note: This method is used for Versions 1 and 2 data only. The second cloud masking method makes use of long time series of Channel 4 data. Beginning with at least 120 days of composited Channel 4 brightness temperatures averaged to 25 km pixels, a combination of median and maximum filtering is applied to 20-day increments of the brightness temperatures for each pixel. The filtering selects the most likely clear-sky temperature for each pixel, and interpolates time to fill in the estimated temperature during cloudy periods. This is assumed to be close to a clear-sky time series of channel four brightness temperatures at each pixel location. A cloud mask is generated by thresholding the actual channel 4 with the time series. Separate time series are created for the morning and afternoon composites, since solar heating may change the mean temperatures.

1. Modified CASPR Cloud Masking
   
   Note: The third cloud mask is done exactly the same as the first method using the CASPR routines, except that the Channel 4 statistical mean is replaced with the Channel 4 values from the time series algorithm described above. Because a cloud can be persistent, the longer time series may provide a better mean value. This is especially critical during the periods of low or no solar illumination when the visible spectral tests are not available for the cloud masking.
8. Time Mask
   An acquisition time is provided for each pixel. The time is stored in an 8-bit (1-byte) integer grid as hour (UTC), scaled by ten. Thus, the acquisition time is known to a precision of ± 0.05 hours (± three minutes) for each pixel.
8. Reduced-Resolution Products
   
   To assist users in selecting 5 km data, or for applications that require a lower data volume product, reduced-resolution data are available for the five AVHRR channels, clear sky surface broadband albedo and skin temperature, average albedo and temperature, valid fraction file, solar zenith angle, satellite elevation angle, sun-satellite relative azimuth angle, surface type mask, cloud mask, cloud fraction files, and Universal Coordinated Time (UTC) of acquisition. These products are subsampled to a grid resolution of 25 km by 25 km. Please see the AVHRR Polar Pathfinder Twice-Daily 25 km EASE-Grid Composites documentation for details and ordering.
9. Discussion of Clear Sky Surface Broadband Albedo and Clear Sky Surface Skin Temperature Algorithms
   
   These algorithms are appropriate for clear-sky conditions only. Nonetheless, the algorithms have been applied to all pixels, not just to pixels flagged as clear-sky. The reasoning for this is three-fold: First, the cloud detection algorithm continues to evolve and improve, so different versions of the cloud mask can be provided to users without negating the value of the surface products for clear-sky pixels. Second, users are encouraged to apply their own cloud detection procedures (including manual interpretation), which can be tuned to maximize performance for specific regions and case studies. These custom cloud masks can then be applied to the AVHRR Polar Pathfinder 5 km clear sky surface skin temperature and surface albedo product, eliminating the need for users to implement their own surface product algorithms. Third, some users may wish to investigate the usefulness of the surface products under marginal cloud conditions that may be detected by AVHRR as being cloudy, but which may actually consist of very thin cloud or diamond dust precipitation. Since the CASPR surface algorithms do not work properly for cloudy pixels, the AVHRR Polar Pathfinder 5 km clear sky skin temperature and albedo grids contain inaccurate values. As noted above, users need to apply the cloud mask provided in this data set or use their own cloud detection method.
   
   Evaluation of the CASPR surface products and the cloud detection approach is ongoing, and has identified several shortcomings of the existing products. The CASPR algorithms are optimized for performance over sea ice. Tests indicate that performance of the standard algorithm over ice sheets warrants improvement. The accuracy of the model simulations that form the basis of the CASPR albedo algorithm decreases with low solar zenith angles, such as the sun being low on the horizon. Thus, albedos calculated in spring and autumn are likely to be less accurate than those for periods with higher sun angles. Also, parameterizations used to address bidirectional reflectance are not equally representative of all snow conditions for a full range of sun and satellite viewing angles.
   
   The performance of the clear sky surface skin temperature algorithms was evaluated using surface observations from an ice camp in the Beaufort Sea north of Alaska during March and April, 1992, and from an annual cycle of data from a land site near Barrow, Alaska (Key et al. 1997). The bias and RMSE at the ice camp were -0.30 degrees and 0.61 degrees, respectively. At the land site, the bias and RMSE were -0.97 degrees and 4.70 degrees, respectively.
   
   The surface type used to calculate clear sky surface broadband albedo is the same as that used to calculate clear sky surface skin temperature. Values for albedo are neglected where the solar zenith angle is greater than or equal to 85 degrees. An annual cycle of data from Barrow, Alaska was used to validate the clear sky surface broadband albedo (Key et al. 1997). The bias and RMSE are 0 percent and 17 percent, respectively, at the land site. However, errors are expected to be smaller over ocean, sea ice, and snow.

When the AVHRR Polar Pathfinder program began, calibrations were based on the NOAA Polar Orbiter Data (POD) User's Guide (Kidwell 1995). Since then, at least four different publications have presented different methods of calibrating AVHRR data. Even with extensive pre-launch testing and calibrations, satellite sensors change over time, and improved methods are developed. In some cases, particularly for the visible channels, the sensor characteristics are not fully predictable, and the satellite must be in orbit before changes can be detected to make the proper corrections. However, the calibrations for this data set follows the guide lines in two publications: NOAA Polar Orbiter Data User's Guide: Section 3.3 Calibration of AVHRR Data was used for NOAA 7 through NOAA 14 satellite data, and NOAA KLM User's Guide: Section 7.1 AVHRR was used for the NOAA 16 satellite data.

Visible and Near Infrared Channels

The POD calibration method for Channels 1 (visible) and 2 (near-infrared) used pre-launch values. The primary use of these channels has traditionally been for vegetation studies over land. Investigations revealed that the two channels drifted from the initial launch conditions. A set of time-varying coefficients was subsequently developed to correct for the drifting sensors. In 1999, the calibration for NOAA-14 changed and new time-varying equations were recommended (Rao and Chen 1999). See the Derivation Techniques and Algorithms section of this document for more information.

Thermal Channels

At the start of the AVHRR Polar Pathfinder program, calibration of the thermal channels followed the guidelines in the POD. The methodology is as follows:

• compute a linear slope and intercept based on the instrument count of space and internal blackbody
• apply the slope and intercept to calculate scene radiance
• convert the radiance to brightness temperature using the Plank functions
• apply a non-linear correction to the temperatures using a look-up table of corrections based on scene and blackbody temperature

This method was optimized for a narrow range of AVHRR sea surface temperature products. The Walton method describes an effective method for correcting AVHRR thermal channels 3, 4, and 5 to accommodate the wide range of temperatures in the polar regions. The NOAA/NASA Land Pathfinder group also selected this method, and it is used for the AVHRR Polar Pathfinder 5km Data Set.

The Walton method applies a non-linear correction to the radiance, which is then converted to brightness temperature using a quadratic function with coefficients derived from pre-flight calibrations. The result, while not optimized for the narrower range needed for sea surface temperature measurements, appears to be an improvement for polar regions.

The Walton calibration method was implemented in March 1999 for thermal channels from all NOAA AVHRR satellites. Correction tables were generated for previously processed data to closely match the new calibration techniques. These corrections appear to be compatible to within 0.1 degrees (Walton et al. 1998).