Downscaling of Landsat and MODIS Land Surface Temperature Over the Heterogeneous Urban Area of Milan

Abstract

Remotely sensed images of land surface temperature (LST) with high spatial resolution are required for various environmental applications. For instance, finer resolutions (FRs) are essential to capture thermal details in urban textures. To meet the requirements of sharper and sharper images, this study carries out a downscaling from coarser spatial resolution LST images to FRs using relationships between LST and spectral indexes (SIs) representative of different land cover types over the heterogeneous area of Milan. Different regressive schemes were applied to downscale LST of Landsat Thematic Mapper (TM) and Terra MODIS images during four summer passages. The regressions were first evaluated on Landsat images aggregated at 960 m resolution and downscaled to 480, 240, and 120 m. For the four Landsat scenes, the best regression models include both vegetation and built-up/soil SIs: the root mean square (rms) error, around 1 K for 480 m and 2 K for 120 m, is clearly below the LST standard deviation of each reference image, assumed as LST spatial variability. Then, contemporary MODIS data were downscaled from 960 m to the above FRs, and the best models include again both vegetation and built-up/soil SIs. The rms error is higher than the correspondent Landsat one (in some cases exceeds 3 K), but always below the LST spatial variability. A compression of the range of LST values for the MODIS-downscaled images was found with respect to the Landsat disaggregated images: this shortcoming in the LST retrieval affects the MODIS downscaling accuracy.