Missing value imputation of long-term GNSS monitoring time series using a graph-conditioned diffusion model

GNSS monitoring time series often suffer from missing values caused by receiver failures, equipment maintenance, power outages, and other uncontrollable factors, which degrade data quality and affect subsequent analysis. Existing imputation methods usually focus on short-term correlations and therefore cannot effectively handle GNSS data with large temporal spans and continuous missing points, while also overlooking the dynamically evolving dependencies among variables. To address these issues, we propose a Diffusion-based Satellite Data Imputation framework, termed DISA, which is built upon denoising diffusion probabilistic models and integrates a GNN-based condition extraction module to capture evolving variable dependencies, a Selective-UNet denoising backbone to model long-term dependencies, and an iterative replacement strategy to improve imputation accuracy. Experimental results on real-world datasets demonstra-te that DISA outperforms existing baseline methods, especially in scenarios with long time spans and continuous missing data.