Towards a Spatio-Temporally Complete Urban Land Surface Temperature Record: A Multi-Sensor Benchmark and Rank-Based Baseline Evaluation for Thermal Downscaling over Barcelona.

Introduction: Urban heat exposure mapping requires high-resolution land surface temperature (LST), yet no single sensor provides it completely. Landsat thermal is acquired at ~100 m and is only resampled to a 30 m grid, so it does not resolve 30 m structure; revisit is 16 days, and it is cloud-limited. Sentinel-3 SLSTR observes daily at 1 km, ECOSTRESS at 70 m irregularly. Standardised downscaling benchmarks are lacking.
Methods: We introduce a multi-sensor benchmark for Barcelona's metropolitan core: 137 co-registered Sentinel-3/Landsat pairs (2016-2024), split temporally into three disjoint sets: 55 training (2016-2021), 21 validation (2022), and 61 held-out test pairs (2023-2024). Because exposure depends on spatial ranking rather than magnitude, the protocol is rank-based: Spearman correlation, hottest-10% intersection-over-union, hotspot F1, and per-Local-Climate-Zone (LCZ) consistency.
Results: Three baselines set the bar on the 61 held-out pairs. Baseline 1 is a negative control, not a proposed method: bilinear resampling of 1 km Sentinel-3 brightness temperature onto a 30 m grid. As intended, it fails (Spearman -0.06, hotspot F1 0.00, RMSE 14.5 K): coarse brightness temperature carries no exploitable fine-scale structure; resampling is not downscaling. Baseline 2 assigns each pixel its training-year LCZ class mean Landsat LST, ignoring Sentinel-3 (Spearman 0.60). Baseline 3, a per-pixel training-years Landsat climatology corrected by that day's Sentinel-3 anomaly, performs best (Spearman 0.74, hottest-10% IoU 0.31, hotspot F1 0.24, RMSE 4.7 K).
Conclusions: A high-resolution climatological prior is indispensable: learned models must beat it to add value. Its residual is temporal rather than a static covariate: progress requires temporal dynamics. Benchmark, protocol, and baselines give a reproducible foundation for spatio-temporal deep learning, evaluated on the 61 held-out Landsat pairs. On three test dates with usable ECOSTRESS coverage, Baseline 3 is consistent with an independent 70 m sensor (Spearman 0.83, RMSE 4.3 K) and is close to ECOSTRESS against Landsat (0.91); n=3 is too small to rank methods.