Evaluation of the AMSU land surface temperature algorithm for skin and shelter-air temperature retrievals

Accurate satellite retrievals of skin and shelter-air temperature (i.e., at 2 m height) over land are important for climate studies, weather prediction models and meteorological applications. Skin temperature measurements on the ground are not routinely available, whereas in-situ measurements of shelter-air temperature are routinely made over the US and other parts of the globe. However, retrievals from satellites offer coverage in time and space that cannot be matched with ground measurements. The AMSU land surface temperature algorithm utilizes the microwave window frequency channels at 23, 31 and 50 GHz. These lower frequency channels achieve maximum penetration through clouds, thus enabling retrievals in near all-weather conditions. In contrast, the land surface temperature retrievals from satellite IR measurements are affected by atmospheric clouds and thus deteriorate during cloudy conditions.

The AMSU land surface temperature algorithm has been derived from radiative transfer model calculations including land surface emissivities, with algorithm coefficients empirically adjusted for skin temperature retrievals. The objective of this paper is to evaluate this algorithm for its ability to also retrieve shelter-air temperatures. Shelter-air temperature observations were obtained from an extended network of in-situ meteorological stations over the CONUS US and matched up with the AMSU measurements and estimated skin temperatures. Match-ups included microwave measurements from the three NOAA satellites, NOAA-15, -16, and -17, descending and ascending passes (approx. 4-hour sampling interval per day) in December and July, 2002. A microwave screening procedure was developed to remove effects of snow-cover, precipitation and surface wetness. Next, the screening algorithm was applied to the entire collocated match-up pairs. Statistical analysis of the screened match-up pairs revealed a quadratic relationship between the shelter-air observed and the AMSU-estimated skin temperatures with a high coefficient of determination, standard error of 2.7 K and mean error of 2.1 K, a dramatic improvement of about 50% compared to errors associated with unscreened match-ups. The ensemble regression coefficients were applied to convert estimated skin temperatures into shelter-air temperatures. Comparisons between observed and estimated shelter-air temperatures found remaining satellite-dependent biases despite improved retrieval errors, due to the diurnal temperature effects. To remove these remaining biases, regression coefficients were derived for each satellite from corresponding match-up pairs. As a result, the sheltered-air temperatures retrieved were bias-free, with further reduced mean error of 2.0 K and standard error of 2.5 K. Work is on-going to inter-compare the AMSU land surface temperature retrievals with those of the GOES satellite. Future work will also include tests of the AMSU land surface temperature algorithm with asymmetry correction in order to eliminate the angular dependent biases, and tests of limb-correction effects.