A Multi-Layer Perceptron Approach to Downscaling Geostationary Land Surface Temperature in Urban Areas

Remote sensing of land surface temperature (LST) is a fundamental variable in analyzing temperature variability in urban areas. Geostationary sensors provide sufficient observations throughout the day for a diurnal analysis of temperature, however, lack the spatial resolution needed for highly heterogeneous areas such as cities. Polar orbiting sensors have the advantage of a higher spatial resolution, enabling a better characterization of the surface while only providing one to two observations per day. This work aims at using a multi-layer perceptron-based method to downscale geostationary-derived LST based on a polar-orbit-derived one. The model is trained on a pixel-by-pixel basis, which reduces the complexity of the model while requiring fewer auxiliary data to characterize the surface conditions. Results show that the model is able to successfully downscale LST for the city of Madrid, from approximately 4.5 km to 750 m. Performance metrics between training and validation datasets show no overfitting. The model was applied to a different time period and compared to data derived from three additional sensors, which were not used in any stage of the training process, yielding a R2 of 0.99, root mean square errors between 1.45 and 1.58 and mean absolute errors ranging from 1.07 to 1.15. The downscaled LST is shown to improve the representation of both the temporal variability and spatial heterogeneity of temperature, when compared to geostationary- and polar-orbit-derived LST individually. The resulting downscaled data take advantage of the high observation frequency of geostationary data, combined with the spatial resolution of polar orbiting sensors and may be of added value for the study of diurnal and seasonal patterns of LST in urban environments.