Improving Cloud Detection for AATSR Land Surface Temperature Retrievals
Sembhi, Harjinder1; Ghent, Darren1; Remedios, John1; Bulgin, Claire2; Merchant, Chris2
1University of Leicester, UNITED KINGDOM; 2University of Edinburgh, UNITED KINGDOM

Automated cloud detection methods are an integral part of any satellite retrieval of land surface temperature (LST) and should accurately and efficiently distinguish cloud- contaminated brightness temperatures from those belonging to the underlying surface. Typically, cloud detection methods use threshold-based brightness temperature differences (BTD) or reflectance ratios (RR) algorithms, although other methods (for example, Bayesian cloud detection) are fast proving to be more effective.

In this study, we apply these techniques to data from the Advanced Along-Track Scanning Radiometer (AATSR) which has a combination of thermal and visible channels. Previous studies and our own have shown that the ESA standard cloud mask can grossly underperform over a variety of different land surface types, often missing out thin cirrus clouds or falsely flagging cities and coastlines. To overcome these cloud detection issues, three potential cloud masking methodologies have been developed with ESA and NCEO support based on a) updated thresholds for cloud flags within the ESA standard cloud mask; b) a probabilistic-simulation cloud mask based on AATSR thermal channels and c) the University of Edinburgh Bayesian cloud mask (based on SST Bayesian thresholds).

We assess the overall performance of the ESA standard cloud mask against these three cloud masks and provide a detailed investigation of which cloud masks fail with particular attention paid to which land types prove to be more difficult to mask. The traditional approach to evaluating cloud mask performance is usually based on daytime scene-by-scene comparisons with respect to an automated or manual classification of the scene. However, our evaluation method specifically uses manual masks developed in the SYNERGY project as well as the quality of LST retrievals as a metric. Our assessment is enhanced by regional evaluation of LST when gridded to level-3 - both in terms of grid cell examination and time series analysis. Furthermore, local assessment is carried out over intra- and inter-annual cycles through comparison of cloud-masked satellite-retrieved LST and in situ measurements. These additional metrics highlight weaknesses in the candidate detection methods that are often missed in scene-by-scene assessments.

With preparations for the upcoming Sea and Land Surface Temperature (SLSTR) instrument on-board Sentinel-3 - the successor to AATSR - to provide operational LST data under the framework of the Global Monitoring for Environment and Security (GMES) service, it is imperative that cloud detection methods are comprehensively evaluated for both day and night retrievals to instil confidence in the retrieved observations of LST. These results are relevant not only to SLSTR but also SLSTR-OLCI synergy.