AdaGeo-SR: Geo-semantic-guided super-resolution of sentinel-2 imagery with AlphaEarth foundation priors

Remote sensing image super-resolution is an important approach for enhancing the spatial detail of Earth observation data. In super-resolving Sentinel-2 multispectral imagery from 10 m to 2.5 m, conventional single-image super-resolution (SISR) methods are largely constrained by purely pixel-driven mechanisms. As a result, they often produce structural discontinuities, blurred edges, and chromatic artifacts when reconstructing complex human-made objects such as narrow roads and farmland boundaries during cross-scale detail reconstruction. To address these limitations, this study proposes a geo-semantic-guided remote sensing super-resolution framework, termed AdaGeo-SR. The framework incorporates dense spatial feature maps extracted from AlphaEarth Foundation (AEF) as geo-semantic priors, providing additional geo-semantic and topological cues for the reconstruction process. To better leverage high-level semantic features, we further design a degradation-simulation-based feature balancing mechanism. By introducing specific degradation conditions during training, this mechanism encourages the network to reduce its over-reliance on the original pixel cues, thereby enhancing its use of external geo-semantic priors. Experiments on the OpenSR-Test and WorldStrat benchmarks show that AdaGeo-SR effectively suppresses structural distortions and hallucination artifacts while maintaining reasonable spectral consistency, leading to improved local geometric coherence. In addition, downstream road extraction experiments using SAM_Road on the NAIP test set further demonstrate that the super-resolved images produced by the proposed framework help reduce mask fragmentation and topological discontinuities. These results highlight the potential of the proposed framework for downstream structure-sensitive tasks and support the generation of structurally more reliable 2.5 m imagery for practical remote sensing applications.