ESA CLIM4cities - Land Surface Temperature Downscaled from Sentinel-3 SLSTR Level-2 LST products - Copenhagen, Aarhus, Aalborg and Odense, 2020-2023

Overview

The present dataset provides land surface temperatures (LST) of 300 m resolution for the Danish functional urban areas of Copenhagen, Aarhus, Aalborg, and Odense between 2020 and 2023 (including) as well as other items. There are six main items issued by the dataset, as described below.

Data Products

• lst_sen3_downscaled.zip - zip folder with fine LST data obtained by applying the scale-invariance-based downscaling model downscaler.joblib on coarse Sentinel-3 LST;
• lst_sen3_interpolated.zip - zip folder with fine LST data obtained by applying cubic spline interpolation on coarse Sentinel-3 LST (and further filling the propagated NaN values with the result obtained with bilinear interpolation);
• lst_lansat.zip - zip folder with validating fine Landsat 8/9 LST data (with original 30 m resolution);

Model

• downscaler.joblib - the downscaling model;

Validation Results

• score_downscaler.json - the coarse and fine validation scores obtained by the downscaling model (using coarse Sentinel-3, coarse and fine Landsat 8/9 LST data as ground truths, respectively);
• score_interpolation.json - the coarse and fine validation scores obtained by a mean model and the above-mentioned interpolation model, respectively (using coarse Sentinel-3, coarse and fine Landsat 8/9 LST data as ground truths, respectively);

Methodology

The downscaled data was obtained from coarse Sentinel-3 SLSTR Level-2 LST products, with a resolution of 1000 m, georeferenced onto a regular grid and upsampled using the scale-invariance-based downscaling model applied on such data as well as pure spatial data and spectral indexes (derived from georeferenced Sentinel-3 SYN Level-2 products, with a resolution of 300 m). Such predictors corresponded to coastal distance, imperviousness density (IMD), tree cover density (TCD), fractional vegetation cover (FVC) and normalised difference water index (NDWI). The spatial predictors, which were originally even finer than the ones of the SYN products, were reprojected to the grid of the latter. A total of 112 late-morning timestamps, for which the cloud cover fraction in the area of interest did not surpass 5 %, were processed. Landsat 8/9's matched fine LST data (of 30 m resolution) from 7 timestamps, which were further georeferenced and reprojected to Sentinel-3's coarse and fine grids, were used for validation of the fine predictions. The obtained downscaling model may be regarded as a refinement of the well-known TsHARP and DisTrad architectures by considering a larger number of predictors. This downscaling model is actually comprised by several sub-models, one per timestamp: training and inference of each one is done by firstly reprojecting a copy of the fine predictors to LST's coarse grid, training a base linear regression model with the coarse predictors and target, applying the base model on the fine predictors (therefore, assuming scale-invariance) and correcting the fine prediction by adding to it the finely interpolated residual of the coarse prediction.

Related Work

An article and a PyPI python package have resulted from the CLIM4cities project. 

Funding

The dataset was produced within the CLIM4cities project funded by the European Space Agency (ESA Contract No. 4000143628/24/I-DT, AI Trustworthy Applications for Climate) and by the Recovery and Resilience Plan Investment RE-C05-i02: Interface Mission - CoLAB, certified by the National Innovation Agency (Project N.º 01/C05-i02/2022).