Consistent land surface temperature data generation from irregularly spaced Landsat imagery

Abstract

Abstract Land surface temperature (LST), derived from satellite thermal infrared (TIR) sensors, is a key variable for characterization of urban heat island, modeling of surface energy balance, estimation of evapotranspiration and soil moisture, and retrieval of air temperature. Among the satellite TIR sensors in operation, Landsat TIR sensor provides the only feasibility for long-term reconstruction of a LST dataset for environmental applications. However, a holistic technique is not currently available to generate spatially and temporally continuous LSTs from Landsat due to its 16-day revisit frequency, impact of atmospheric conditions and the SLC (Scan Line Corrector) -off gap. Previous algorithms had been developed to overcome these limitations, it is still not possible to generate LSTs at any desired date with consistent accuracy and corrections. Therefore, this study aimed to devise an algorithm to reconstruct consistent, daily LSTs at Landsat spatial resolution based solely on Landsat imagery. By selecting Beijing, China, as the study area, a total of 512 images from 1984 to 2011 were downloaded from the USGS online portal and were consistently calibrated to surface reflectance and brightness temperature. The cloud-, cloud shadow-, and snow-contaminated pixels were excluded according to quality flags; and a further screening procedure based on temporal information of Landsat spectral bands 2, 4, and 5 was conducted. Brightness temperatures were converted to LSTs through the single channel algorithm with input of water vapor from the NCEP Reanalysis dataset. Field LSTs were collected from 11 weather stations in Beijing in the year of 2008, 2009, and 2010. The proposed algorithm included four modules: Data filtEr, temporaL segmentation, periodic and trend modeling, and GAussian process (DELTA). Accuracy assessment showed that, compared with the in situ LSTs from weather stations, satellite-derived LSTs inverted through the single channel algorithm had an average accuracy of 2.3 K. Further comparison between LSTs reconstructed from the DELTA algorithm and those collected from weather stations in the year 2008 yielded a mean error of 3.5 K. Twelve LST maps reconstructed from the DELTA in 2000 showed that LSTs of different land covers exhibited similar seasonal patterns and reached their maximal values in June/July. Using LST of every August 15th as an example, the SUHI (surface urban heat island) intensity of Beijing was computed, which ranged from 3.3 K to 5.3 K from 1984 to 2011, with an increase pattern of LST in both rural and urban areas.