Cloud removal in optical remote sensing typically relies on multi-temporal or multi-source data and large-scale supervised training, which limits its applicability in data-scarce scenarios. Although the deep image prior (DIP), an internal learning regime, enables training-free reconstruction from a single degraded image, it lacks strong physical or semantic constraints and often suffers from over-smoothing in large cloud-covered regions. Geospatial embeddings encode temporal trajectories of surface conditions through various Earth observations, but their potential to assist cloud removal in optical images has not yet been explored. This study first combined the ready-to-use Satellite Embedding dataset, generated by the AlphaEarth Foundations model and released by Google, with the DIP to reveal the potential of this dataset for cloud removal and sub-cloud information reconstruction in optical remote sensing. The results showed that the reduced-dimensional embedding data, obtained by compressing the original 64-dimensional representation to 4 dimensions through an autoencoder, served as an informative reference that effectively facilitated cloud removal and cloud area reconstruction. Spectral consistency and structure reconstruction performance were further improved as the range and shape priors derived from the original embeddings were incorporated into the DIP structure. The shape prior enforced boundary continuity, while the range prior constrained pixel intensities within physically plausible embedding-defined limits. Across two datasets, the proposed configuration with two priors achieved up to ∼0.15 improvement in SSIM, ∼1.4 dB gain in PSNR, and ∼0.019 reduction in RMSE, over cloud-contaminated regions compared to the baseline embedding-guided DIP. These findings demonstrated the potential of embedding-based representations and their derived priors to enhance physical consistency and structural fidelity in cloud removal for optical imagery, greatly reducing the burden of data preparation and improving processing efficiency.
Li et al. (Sun,) studied this question.