IMPACT OF SCENE DEPENDENCE ON AVHRR ALBEDO MODELS

Narrowband observations from NOAA's Advanced Very High Resolution Radi ometer (AVHRR) are used operationally by NOAA to estimate the earth's broadband planetary albedo. Since May of 1988, these broadband albedo estimates have been derived using the two-channel (visible and near-in frared), scene-independent regression model of Wydick et al. The occur rence of relatively large regional bias errors using this model has le d to a study of scene-dependent models that substantially reduce these errors. Three classes of scene stratification are considered: 1) by s urface geography type alone (SFC); and 2) and 3) by surface geography type in combination with cloud amount category (SCN) and normalized di fference albedo index (NDAI) using AVHRR channels 1 and 2. These and t he Wydick model are applied to independent AVHRR global data of 2 July 1985 (hereafter July) and 9 January 1986 (hereafter January). Using E RBE (Earth Radiation Budget Experiment) data as a reference, errors in reflected;nux are computed for each day. The total AVHRR-ERBE shortwa ve flux difference is separated into two terms. One is due to inaccura cy in the calibration of the AVHRR reflectances (calibration error). T he second is the error due to all other sources of the AVHRR-ERBE flux difference. It is referred to as the measurement error. Spatial sampl ing differences (sampling error) and limitations in the mathematical f orm and specification of the AVHRR regression model equations (model e rror) are probably the two primary components of the measurement error . When calibration error vanishes (due to the implementation of calibr ation corrections) and sampling differences are small (i.e., for globa l and zonal averaging), only the model error remains. The Wydick model yields high positive;global bias errors of 22 and 37 W m(-2) for July and January, respectively. In contrast, errors of +/-5 W m(-2) are ob tained with the scene-dependent models (i.e., SCN). When no calibratio n adjustments to the AVHRR data are performed, as in operational proce ssing, the Wydick model produces bias errors of -6.8 and 1.5 W m(-2) f or July and January, respectively. These low bias errors may be mislea ding though as they result from the near cancellation of large model a nd calibration error components. The cancellation is not effective at all latitudes, so the Wydick model tends to generate large north-south error gradients. These latitudinal errors are largely removed by all of the scene-dependent models.