Large-scale downscaling of Sentinel-3 land surface temperature to Landsat resolution using temporally aggregated reference data in Germany

Kurzfassung

This thesis examines the large-scale downscaling of Sentinel-3 land surface temperature to Landsat-like spatial resolution in Germany. The starting point is the trade-off between the high temporal resolution of Sentinel-3 and the higher spatial resolution of Landsat 8/9, which are combined using a GAN-based spatiotemporal fusion model. The central question is whether temporally aggregated Landsat LST products are suitable as robust, high-resolution reference input for GAN-STFM-based LST downscaling. To assess this, the established nearest-date approach is compared with various aggregated reference strategies. These include mean, median and z-normalized aggregate products, as well as annual and bi-seasonal variants. The evaluation is performed using independent AOI test scenes and the metrics RMSE, MAE, MBE and SSIM. The results show that temporally aggregated reference data are generally feasible for GAN-STFM-based downscaling and do not necessarily perform worse than nearest-date references. The best overall performance is achieved by the bi-seasonal anomaly strategy, which has the lowest median RMSE and the highest mean SSIM. However, the nearest-date approach remains the only strategy without a systematic warm bias and continues to serve as a strong baseline. The Germany-wide model application using the bi-seasonal anomaly strategy shows spatially plausible national and local LST patterns. Accordingly, the thesis concludes that temporally aggregated Landsat LST references represent a viable strategy for national Sentinel-3 LST downscaling using GAN-STFM. However, the model performance remains spatially variable and depends on AOI characteristics, target time and reference strategy.