A multi-layer perceptron approach to downscaling geostationary land surface temperature in urban areas

Machine learning is a subfield of artificial intelligence, rooted into statistics. It is a flexible and interdisciplinary tool that can be used for solving several types of problems. This work focused on the use of multi-layer perceptron (MLP) to solve a regression type problem: downscale land surface temperature (LST) from the SEVIRI sensor onboard Meteosat Second Generation series of satellites to a 750 m spatial grid. The choice of an MLP in preference of other machine learning algorithms was motivated by the intention of developing the simplest algorithm that provides acceptable results. Data from VIIRS onboard Suomi National Polar-Orbiting Partnership was used to compute the target LST. The resulting model was trained on 2019-2022 data and its performance accessed on 2023 data. To compare downscaled LST at different hours than the ones retrieved by SNPP, data from three additional sensors were used: VIIRS onboard both Joint Polar Satellite System 1 and 2 and the SLSTR onboard Sentinel 3A.

The comparison between the target and the models’ results are optimistic given that its performance on new data is similar to its training performance, i.e. no overfitting. The distribution of model temperature values follows the one of target temperature values.

The new downscaled dataset is then used to analyze the diurnal behavior of the surface urban heat island with an improved spatial resolution than the one a geostationary is able to provide, for the city of Madrid during cold/heat waves.