Enhancing spatial resolution of Landsat derived land surface temperature: A novel downscaling approach using an extreme learning machine: Enhancing spatial resolution of Landsat derived land surface temperature

The study presents an approach for downscaling Landsat retrieved land surface temperature (LST) from 30 m spatial resolution to 10 m using an extreme learning machine (ELM) algorithm. The LST was retrieved from Landsat thermal data using a split window algorithm. For LST downscaling, seven vegetation indices were calculated from Dove satellite data. ELM was utilized to establish a causal relationship between LST and the downscaling variables. The mean absolute error (MAE) achieved during the training and testing of ELM was 0.48 and 0.51, respectively. Since ELM finds the hyperparameters analytically, it led to reduced training time and better generalization. Assuming the causal relationship between LST and variables to be scale-invariant, the LST was downscaled from 30 to 10 m resolution. The downscaled LST at 10 m was validated using LST values collected by ground survey, and the MAE between the downscaled and validation LST data was found to be 1.3°C. The downscaled results were also spatially compared with the land cover map of the study area. It was found that the maximum effect of LST downscaling was observed over heterogeneous built-up areas, where the LST spatial variability became more pronounced after downscaling, compared to other land covers.