Downscaling Landsat-8 land surface temperature maps in diverse urban landscapes using multivariate adaptive regression splines and very high resolution auxiliary data

Abstract

We propose a method for spatial downscaling of Landsat-8 derived LST maps from 100(30m) resolution down to 2-4m with the use of the Multiple Adaptive Regression Splines (MARS) models coupled with very high resolution auxiliary data derived from hyperspectral aerial imagery and large-scale topographic maps. We applied the method to four Landsat 8 scenes, two collected in summer and two in winter, for three British towns collectively representing a variety of urban form. We used several spectral indices as well as fractional coverage of water and paved surfaces as LST predictors, and applied a novel method for the correction of temporal mismatch between spectral indices derived from aerial and satellite imagery captured at different dates, allowing for the application of the downscaling method for multiple dates without the need for repeating the aerial survey. Our results suggest that the method performed well for the summer dates, achieving RMSE of 1.40-1.83K prior to and 0.76-1.21K after correction for residuals. We conclude that the MARS models, by addressing the non-linear relationship of LST at coarse and fine spatial resolutions, can be successfully applied to produce high resolution LST maps suitable for studies of urban thermal environment at local scales.

A simple method for determination of fine resolution urban form patterns with distinct thermal properties using class-level landscape metrics

“…1) with population of 229,941, 106,940, and 258,018 (Office for National Statistics (2013) respectively and a temperate oceanic climate according to the Köppen-Geiger climate classification system. The three towns are characterised with contrasting histories: modern-day gardencity, medieval, and industrial, respectively, collectively representing a wide range of urban form patterns (Grafius et al 2016;Zawadzka et al 2019).…”

“…Landsat 8 LST maps for the three towns at original 30(100)m spatial resolution were downscaled to 2(4)m resolution using Multiresolution Adaptive Regression Splines method and ancillary data including spectral indices and green-grey infrastructure footprints, described in detail in Zawadzka et al (2019). The mixed spatial resolution of the coarse LST imagery stems from the fact that Landsat 8 TIR bands are captured at 100 m and are subsequently resampled, using the bilinear convolution method, by data provider (USGS-United States Geological Survey).…”

Downscaling Aster Land Surface Temperature over Urban Areas with Machine Learning-Based Area-To-Point Regression Kriging

“…To further improve the accuracy of downscaling LST, many researchers have proposed new LST downscaling methods by integrating more related land surface parameters into the regression model [9,[23][24][25][26]. The geographically weighted regression (GWR) model has been widely used for LST downscaling by combining various land surface parameters [27].…”

“…More accurate LST data has wide applications in meteorological, hydrological, and ecological research and urban heat island monitoring.have been widely applied in meteorological, hydrological, and ecological research and urban heat island monitoring on both regional and global scales [4][5][6][7][8]. Nevertheless, the spatial resolution of existing LST products from 60 m to 1 km, is insufficient to reveal the local intrinsic interactions between LST and land covers [9,10], especially in complex urban environments. The available Landsat LST products can have a moderate spatial resolution of 30 m, but these LST products are interpolated only by the cubic convolution resampling method without considering the correlation between LST and ancillary parameters, such as normalized difference vegetation index (NDVI), normalized difference building index (NDBI), albedo, and so on.…”

Resolution Enhancement of Remotely Sensed Land Surface Temperature: Current Status and Perspectives

“…The NDVI has a widely acknowledged trapezoid relationship with LST [31] and is the most popular regression kernel used in the current literature. In urban areas and other less vegetated areas, the NDBI can be used to portray the distribution pattern for LST [32]. Several kinds of auxiliary data that could be used to parameterize land use patterns [12], topographic characteristics [13], or energy balance conditions [33] have also been included to increase the accuracy of spatial enhancement.…”