Introduction

The Satellite Applications Facility on Land Surface Analysis (LSA-SAF) generates a set of products related with land, land-atmosphere interactions, and bio-geophysical parameters. The LSA-SAF products are available on a pixel basis, in the satellite nominal resolution (3 km at nadir in the case of SEVIRI/Meteosat, centred at 0E). Furthermore, an indication of the expected accuracy of each retrieved value is also given, either in the form of quality flags, or as estimated error bars. In the case of Land Surface Temperature (LST), the quality information is still qualitatively. Here we present a statistic characterization of estimates provided by the LST algorithm, which allow the determination of a realistic error bars for each retrieved value. The error statistics of the LST algorithm are based on sensitivity studies, which take into account the expected errors of the main inputs, namely: (i) sensor noise for SEVIRI channels 10.8μm and 12.0μm; (ii) emissivity uncertainties – mostly dependent on the dominant land cover for each pixel; and (iii) error statistics of ECMWF forecasts of total column water vapour (TCWV). LST is estimated for clear-sky pixels, and thus, the impact of undetected clouds is out of scope of the present study. The uncertainties in the input variables are assumed to have a Gaussian distribution, with known standard deviation. In the case of TCWV this is available from ECMWF error statistics, while in the case of surface emissivity the error standard deviation is estimated as a function of the pixel land cover class. The propagation of the different error sources in the LSA-SAF LST algorithm allows the estimation of the respective error standard deviation, as a function of input variables. The results of the sensitivity study show that sensor noise, by itself, is responsible for LST inaccuracies of the order of 0.3C. Uncertainties in surface emissivity have a higher impact for dry atmospheres (reaching values of the order of 0.6C to 0.8C). Although the emissivity impact is smaller for moist conditions, the overall LST error tends to increase with TCWV; it is estimated that LST inaccuracies higher than 2C occur for less than 10% of the retrieved values.