Downscaling land surface temperatures at regional scales with random forest regression

Environmental monitoring with satellite data is facilitated by frequent observations at a fine spatial scale. As land surface temperature (LST) is one environmental key variable, we implemented a random forest (RF) regression approach to increase the spatial resolution of LST maps from ~1km, routinely available in daily repetition from the Moderate Resolution Imaging Spectroradiometer (MODIS), to ~250m. LST was downscaled based on its relationship to topographic variables derived from digital elevation data of the Shuttle Radar Topography Mission (SRTM), land cover data (MODIS product MCD12Q1), and surface reflectances in the visible red and near-infrared, which both are provided with the MODIS/Terra daily product MOD09GQ at ~250m resolution. The approach was tested for a complex landscape in the Eastern Mediterranean, the Jordan River Region, with LST fields from aggregated Landsat-7 ETM+ and MODIS (MODIS/Terra LST product MOD11A1) data; as reference at the finer scale, we used Landsat-7 derived LST data. For the ideal-case scenario with both degraded and reference values from the same sensor (Landsat-7 ETM+), root mean square errors (RMSE) of downscaled LSTs ranged from 1.02K to 1.43K for six different acquisition dates. When compared to the widely-adopted and in parallel applied TsHARP sharpening method, that is based on the relationship between NDVI and LST, downscaling accuracy with RF improved up to 19%. Applying the RF approach to MODIS LST products yielded RMSEs from 1.41K to 1.92K, whereas the TsHARP method and also a uniform disaggregation by resampling provided only slightly worse results. For the real MODIS LST product, downscaling with RF was affected by lower thermal contrasts in the image data that hindered an adequate training to reproduce temperature variations at the finer scale of ~250m. In this context, we assume the LST product of the Visible Infrared Imaging Radiometer Suite (VIIRS) instrument (as a successor of MODIS onboard the SUOMI NPP platform) to be a better candidate for downscaling, as it provides a spatial resolution of ~750m. One advantage of the RF approach is that predictor datasets can easily be adapted to data availability. In an extended RF approach with all Landsat spectral bands, downscaling results for formerly aggregated Landsat data improved distinctly and now ranged from 0.98K to 1.33K. This approach is also promising for the downscaling of real MODIS or VIIRS LST data as it may be combined with already available reflective data fusion models that are able to blend Landsat data and sensor data with a coarse spatial resolution (given spectral bandwidths corresponding to Landsat) to generate temporally dense synthetic Landsat time series.