Evaluation of conditional diffusion for cross-sensor land surface temperature downscaling across African cities

Fine-resolution land surface temperature (LST) is needed to characterize urban thermal patterns, but satellite observations typically trade spatial detail for temporal coverage. This study evaluates whether conditional diffusion can reconstruct ECOSTRESS-scale LST from coarse-resolution METEOSAT observations and fine-resolution environmental descriptors. Paired ECOSTRESS and METEOSAT scenes from six African cities were used for model development and retrospective evaluation. The 212 development-city evaluation scenes from 2022 were excluded from training but contributed to validation and checkpoint selection, while Johannesburg was designated as the geographic holdout and did not contribute to training or normalization. Each model predicts the residual between ECOSTRESS and aligned METEOSAT LST and adds this correction to the observed coarse thermal field. Conditional diffusion using 50-step DDIM sampling was compared with SRCNN, a deterministic U-Net, and the uncorrected METEOSAT input using radiometric, structural, multiscale, and spatial-consistency diagnostics. Predictions were reconstructed without ECOSTRESS-derived clipping. Across the 212 scenes, DDIM-50 with a fixed 10-member ensemble reduced mean RMSE from 5.143 K for METEOSAT to 4.936 K and median RMSE from 3.838 to 3.482 K. The paired mean contrast is −0.207 K (95% interval −0.400 to −0.011); after excluding two scenes flagged by a reference-only quality-control rule, the paired contrast retains the same direction. Ensemble averaging improved pixel-wise stability but reduced part of the fine-scale variability. These results show that the no-clipping comparison supports a small average radiometric improvement from the fixed-member diffusion product while making the trade-off between stable temperature reconstruction and preservation of local spatial detail explicit.