The imputation of environmental data is crucial due to the frequent incompleteness caused by sensor failures and high operational costs. Traditional statistical methods struggle with complex spatiotemporal correlations, while recent deep learning approaches face challenges like error accumulation and instability, especially with sparse, non-stationary data. We propose a conditional diffusion model (CDEDI), employing cross-attention mechanisms, self-attention with simple imputation embeddings, and global layer standardization. CDEDI effectively captures spatiotemporal dependencies and maintains stable, accurate performance, addressing limitations of existing deep models in atmospheric and aquatic data imputation.