Medium resolution optical satellite data have become an important source of information about the characteristics and dynamics of the land biosphere at regional and global scales. The unique strengths of such data sets afforded by daily satellite coverage and the sensitivity to ecosystem processes have been demonstrated with data from the Advanced Very High Resolution Radiometer (AVHRR) onboard NOAA satellites over more than a decade. These successes have led to the construction of more sophisticated medium resolution optical sensors such as  SeaWIFS onboard the NASA SeaStar satellite launched on August 1, 1997,  ATSR-2 onboard the  ESA ERS-2 satellite launched April 21, 1995, VEGETATION onboard SPOT 4 launched on March 4, 1998,  MODIS and  MISR on the first NASA Earth Observing System platform TERRA launched on December 18, 1999,  GLI on the  NASDA ADEOS-II satellite to be launched in November, 2000, and AATSR and MERIS on the  ESA ENVISAT satellite to be launched some time in June, 2001.

At regional scales, clouds are always present in the images. Compositing procedures are therefore used to create nominally cloud-free images over periods of several days during which the surface conditions may be considered static. High performance computer systems have been designed to process large volumes of data in this manner. In Canada, the GEOcoding and COMPositing System (GEOCOMP Robertson et al., 1992) has been operating at the Manitoba Remote Sensing Centre since 1992 and has been generating seasonal data sets of 10-day composites for the Canada landmass starting in 1993.

Although the resulting composite images are mostly cloud-free, they nevertheless contain some cloud-contaminated pixels. Full pixel clouds are easily detected while pixels containing small clouds might appear cloud-free. Other sources of noise are the angular variations in reflectance as a function of the sun-target-sensor geometry known as the bidirectional reflectance distribution function (BRDF) which varies with surface cover type and season as the vegetation phenology changes; atmospheric effects (absorption and scattering by gases and aerosols); and surface emissivity in the thermal infrared channels. Thus, the Atmosphere, Bidirectional and Contamination Corrections of CCRS (ABC3) software was designed and used to correct the GEOCOMP 10-day composite data for these artifacts. Subsequently, additional procedures were developed and used to generate derived data products that describe the seasonal development of vegetation across Canada or that support the solar radiation budget.

The basic data set consists of 20 10-day composites of NOAA-11 and NOAA-14 AVHRR data between April 11 and October 31 for each year from 1993 to 1999. The composites were prepared using the GEOCOMP system. Briefly, GEOCOMP performs sensor calibration for the five AVHRR bands, registration of the satellite data to ground control points using high resolution image chips, and spatial resampling using a modified Kaiser 16-point kernel. The registered images are input to the compositing process and a compositing criterion (maximum TOA NDVI value) is employed to select the most cloud-free or uncontaminated pixel. No view zenith angle restriction was applied such that all pixels from -69 degrees (backscatter) to +69 degrees (forescatter) were permitted in the composite process. For each 10-day composite, GEOCOMP produced 10 composite channels of data: Top of Atmosphere (TOA) radiance for AVHRR channels 1-5, NDVI, and the four pseudo-bands: view zenith angle (VZA), solar zenith angle (SZA), relative azimuth angle (RA) between VZA and SZA, and the date of the selected pixel. All data were stored in 16-bit resolution.

Since 1999, all AVHRR data processing is performed at the Manitoba Remote Sensing Centre by the new GEOCOMP-n system built by PCI (Adair et al., 2000). A scene id channel has been added to assist data reprocessing by the ABC3 software. Also, starting in 2000 the growing season has been extended to 21 composite periods with the addition of a 10-day composite for the period of April 1-10.

Step 1: Re-calibration of AVHRR data

Because of the sensor degradation after launch, it is necessary to apply time-dependent calibration gains and offsets to derive the TOA radiance for AVHRR channels 1 and 2. CCRS recommended Piece-Wise Linear (PWL) calibration coefficients were applied in GEOCOMP for NOAA 11 (Cihlar and Teillet, 1995) and NOAA 14. However, as typically happens for historical data sets, a definitive set of PWL calibration coefficients was subsequently derived for NOAA 14 which were used in the re-calibration of AVHRR data for channels 1 and 2.

Step 2: Computation of top-of-the-atmosphere reflectance

Top-of-atmosphere (TOA) reflectance for channels 1 and 2 was calculated from the corrected radiance, L*(new), with the formula given by Teillet (1992).

Step 3: Atmospheric correction of AVHRR channels 1 and 2

The SMAC algorithm (Simplified Method for Atmospheric Correction; Rahman and Dedieu, 1994) was employed to convert TOA reflectance to surface reflectance. The processing was carried out using precipitable water vapour content maps based on 6-hourly data. The total atmospheric columnar ozone content was based on multi-year monthly average values. The atmospheric pressure as required for the Rayleigh optical depth correction was derived from a seasonal average barometric pressure map interpolated with a 1-km resolution DEM. These ancillary data were derived from the NCAR Reanalyses Data (Kalney et al., 1996). A constant value of 0.06 was used for the aerosol optical depth at 550 nm based on AEROCAN data for clear days (Fedosejevs et al., 2000). This is based on the assumption that only pixels from the clearest days appear in a 10-day composite. Nominal values were employed for all other atmospheric parameter inputs. The atmospheric corrections were computed on a pixel-by-pixel basis using GEOCOMP viewing geometry pseudo-bands. No pixel mask was applied during atmospheric correction.

Step 4: Identification of contaminated pixels

A new procedure was developed to identify the contaminated pixels (Cihlar, 1996). The procedure, dubbed CECANT (Cloud Elimination from Composites using Albedo and NDVI Trend) is based on the high sensitivity of NDVI to the presence of clouds, haze, smoke and snow. Three features of the annual surface reflectance trend are used: the high contrast between the albedo (represented by AVHRR channel 1) of land, especially when fully covered by green vegetation, and clouds or snow/ice; the average NDVI value (expected value for that pixel and compositing period); and the monotonic trend in NDVI. Four thresholds are required in CECANT to identify partially contaminated pixels. Using these thresholds, a cloud mask was prepared for each composite period.

Step 5: Computation of surface reflectance for AVHRR channel 3

AVHRR channel 3 TOA radiance contains a thermal emissive component and a solar reflective component during the day. Channel 3 contains noisy data, but with some sophisticated processing the surface reflectance (while low) can be retrieved. Based on an empirical relationship with brightness temperature in channel 5, the thermal emissive component can be removed leaving us with an estimate of the surface reflectance in channel 3.

Step 6: Computation of surface temperatures

The goal of the processing of the thermal channels was to obtain surface temperature. The retrieval of precise surface temperature from AVHRR composite images is complex, especially because of the atmospheric attenuation and surface emissivity effects. The modified split window method of Coll et al. (1994) was used which accounts for both atmospheric and surface emissivity effects. Coefficients estimating atmospheric effects were derived by McColl et al. (1994). For soil and vegetation emissivities in channels 4 and 5, a combination of sources was used. The non-linear relationship between the surface emissivity and the NDVI (Van de Griend and Owe, 1993) was used to estimate emissivity for channel 4 as well as the difference in emissivities between channels 4 and 5.

Step 7: Corrections for bidirectional reflectance effects in channels 1, 2 and 3

The BRDF model of Roujean, et al. (1992) was modified to characterize the seasonal BRDF effect for each cover type. The modification of Roujean's BRDF model (Latifovic, 1999) introduced NDVI and the hot spot effect. The CCRS 1995 Land Cover Map of Canada with a pixel size of 1 km (Beaubien et al., 1999; Cihlar, 1996) was used to determine the land cover type. Only cloud-free pixels were included in the derivation of the model coefficients, and no corrections for snow- or ice-covered areas were made. The resulting BRDF model coefficients were used to compute the normalized surface reflectance for channels 1, 2 and 3 for VZA=0 and SZA=45 degrees.

Step 8: Replacement of contaminated pixels for normalized surface reflectance for AVHRR channels 1,2 and 3, NDVI, and surface temperature

Two cases were recognized, pixels contaminated a) during or b) at the start and end of the growing season. For pixels during the growing season, it was assumed that no rapid changes would occur between adjacent 10-day periods and the missing values could therefore be interpolated using the seasonal trajectory for that pixel and data layer (channel). A different problem occurs at the start and end of the growing season where pixels were often missing because of snow or cloud cover. Taking advantage of the cyclical behaviour of reflected solar radiation measured by AVHRR at northern latitudes, it was assumed that the annual trajectory for individual channels as well as for NDVI could be approximated by a second degree polynomial. The polynomial was fitted to the data as a function of time for all clear-sky composite periods, starting with the first clear-sky composite period after August 1 to model the behaviour of NDVI near the end of the growing season. After determining the best fit coefficients, the new values were calculated using the polynomial coefficients to replace contaminated pixels in each channel prior to the first clear pixel or after the last clear pixel. The values for channels 1 and 2 were weighted equally to match the NDVI values after replacement. For the surface temperature, only pixels contaminated during the growing season were replaced; this was done to preserve the information on the length of the growing season.

Step 9: Smoothing of NDVI seasonal profile

For every pixel in composite periods 3 to 18 (in a 20 composite season), a 5-period smoothing was applied where the maximum and minmum NDVI are ignored and the NDVI value for the middle period is replaced by the mean of the NDVI for the 3 remaining composite periods.

There are samples of the corrected image data as well as derived image products.