🤖 AI Summary
This study addresses the domain gap between synthetic training data and real-world remote sensing imagery in cross-sensor super-resolution. It presents the first systematic evaluation of diffusion models on real Sentinel-2 and PlanetScope images, uncovering critical challenges in synthetic-to-real domain transfer. Leveraging a large-scale geometrically and temporally aligned dataset, the work integrates five state-of-the-art diffusion architectures with self-supervised learning and introduces LPIPS-Sat, a novel perceptual metric tailored for quantifying the domain discrepancy in satellite imagery. The findings reveal that models trained on synthetic data suffer significant performance degradation when applied to real scenes, while those trained directly on real data encounter optimization difficulties and struggle to generalize across the physical and radiometric diversity inherent in actual satellite observations, thereby highlighting fundamental limitations of current approaches.
📝 Abstract
Demand for high-resolution satellite imagery has increased interest in super-resolution (SR) to bridge the spatial resolution gap between freely available missions such as Sentinel-2 and commercial systems like PlanetScope. Because no sensor provides true paired low- and high-resolution observations, SR models are usually trained on synthetically degraded data, creating a domain gap on real cross-sensor imagery. In this work, we provide the first systematic study of how this synthetic-to-real mismatch affects the performance of modern diffusion-based SR models. Using a large, geometrically and temporally aligned dataset of Sentinel-2 and PlanetScope imagery, we evaluate five state-of-the-art diffusion architectures under controlled experimental settings. We also introduce LPIPS-Sat, a domain-adapted perceptual metric based on Sentinel-2 self-supervised features. Our results show two persistent challenges: synthetically trained models degrade sharply on real pairs, while models trained on real cross-sensor data exhibit optimisation difficulties and struggle to adapt to the physical and radiometric diversity. These findings highlight a key limitation of current SR and motivate methods that disentangle super-resolution from domain adaptation.