A Scale-Invariance-Based Algorithm Application for Land Surface Temperature Downscaling in Denmark

With an ever-growing recognition of Land Surface Temperature (LST) as a key Essential Climate Variable (ECV), it becomes utmost important to have such a variable at the fine spatial and temporal scales of urban spaces and dynamics. Sentinel-3 provides coarse LST (1 km, daily) based on thermal imagery acquired by its Sea and Land Surface Temperature Radiometer (SLSTR) as well as fine Spectral Directional Reflectances (SDRs, 300 m, every two days) synergically inferred from both SLSTR and the Ocean and Land Colour Instrument (OLCI), which gives the opportunity for using the latter as a predictor in the downscaling of the former. Herein, two scale-invariance-based architectures were developed: a single-timestamp (STS) model, trained with coarse data of the timestamp whose fine target it infers; and a multi-timestamp (MTS) one, trained with multiple timestamps. Note that while several Machine Learning (ML) models besides Linear Regression (LR) were considered for the MTS architecture, only LR was used for the STS one due to the limited amount of available data which the former require for hyperparameter tuning. The models were developed over four Danish Functional Urban Areas (FUAs) using SRD-derived indices and seasonal and geospatial predictors and validated against Landsat data. While Gradient Boosting (GB) achieved the best coarse-scale performance it corresponded to the worst fine-scale performer together with Random Forest (RF), indicating scale invariance breakdown. Tree-based models performed poorly due to extrapolation limitations, whereas Neural Net (NN) and LR proved more robust. After residual correction, single-timestamp LR achieved the best fine-scale performance, making it the most reliable and recommended architecture for operations. The overall results showed that, although ML models may better predict the target at their training scale, their performance may not significantly generalise at others, therefore revealing scale specificity. Furthermore, the results suggested that usage of the more general multi-timestamp architecture instead of the single one may deteriorate performance.