A latent diffusion framework for spatiotemporal enhancement of geostationary satellite imagery

Abstract

Geostationary satellite archives are affected by a persistent trade-off between spatial resolution, temporal sampling rate, and long-term sensor continuity. We propose a latent diffusion framework for simultaneous spatial and temporal enhancement of geostationary satellite imagery, whereas most existing approaches focus on these tasks separately. The model performs temporal interpolation and spatial super-resolution within a single conditional generative pipeline, operating in the latent space of an autoencoder. The study also provides a systematic evaluation of key design choices. Experiments on GOES-16 imagery show that reconstruction quality strongly depends on the sampling configuration. DDIM provides the best overall trade-off between distortion-based fidelity, perceptual quality, and temporal consistency, outperforming PLMS across interpolation regimes. Conditioning-space interpolation achieves the strongest perceptual performance while preserving competitive PSNR values, whereas post-sampling latent interpolation improves temporal smoothness at the cost of reduced high-frequency detail. Experiments on GOES-11 imagery suggest that the framework can generate coherent enhanced sequences under domain shift, supporting its use as an exploratory tool for legacy archives, although merely as a qualitative evaluation, without any calibrated physical validation. The results position latent diffusion as a practical approach for enhancing geostationary satellite time series, balancing perceptual realism, temporal coherence, structural fidelity, and color accuracy.