Detection of bad data images in long-term MODIS land surface temperature image time series using statistical outlier detection methods

Abstract

Land surface temperature (LST) data are available for several years at a high acquisition rate from sensors such as a moderate resolution imaging spectroradiometer (MODIS). Advances in remote sensing technology have exploited this large volume of LST image time series (LITS) data for a clearer understanding of land surface processes. However, remotely sensed LST observations from spaceborne platforms are likely to be influenced by severe weather conditions (e.g., clouds, haze, heavy aerosols, heavy rainfall and flooding, hailstorms, and snowfall). Presence of bad data images in LITS can reduce the credibility of results from LITS data analysis. An algorithm to detect bad data images in a long-term MODIS LITS is proposed. Bad data in MODIS LITS were filtered at various stages. Initially, poor quality pixels were removed from LST maps based on quality control information provided in corresponding quality assurance maps. In further stages, thin cloud-contaminated LST pixels were removed that were left undetected by MODIS official cloud screening procedure. Finally, to remove bad data LST maps, median-based statistical outlier detection method was employed, and outliers were detected in LST pixel time series (LPTS) vectors generated at consistent normalized difference vegetation index (NDVI) pixel (CNP) locations. CNPs were selected by employing Mann–Kendall test statistics on annually aggregated MODIS NDVI time series vectors. Decisions were made to identify bad LST maps based on number of outlier LST values appeared in LPTS vectors at CNP locations for each LST map. The proposed algorithm was applied over parts of Gondwana coalfields situated in Jharkhand and West Bengal states of India.