Spatiotemporal data imputation plays a crucial role in traffic flow monitoring, air quality detection, and climate forecasting. However, the collected sensor data is often incomplete in the time dimension, and the uneven sparse distribution of sensors can also cause data loss in the spatial dimension, resulting in a lack of fine-grained spatiotemporal datasets. Spatiotemporal data imputation aims to fill missing data based on observed values and generate data for target locations based on context. In existing methods, the autoregressive methods suffer from error accumulation, while simple conditional diffusion models do not fully utilize the spatiotemporal relationship between observed data and missing data. To address these issues, we propose the Refine Diffusion Probability Imputation (RDPI) framework. The spatiotemporal data imputation problem is divided into two stages. In the initial stage, preliminary missing data is generated by deterministic imputation methods such as deep neural networks. In the refine stage, we train a conditional diffusion model that takes both forward and reverse processes into consideration, to bridge the gap between real values and initial values generated in initial stage. RDPI combines the rapid generation capability of deterministic models and integrates the efficient likelihood estimation capability of diffusion models. The experiment results of multiple datasets show that RDPI outperforms existing methods in various missing patterns of real spatiotemporal data, and can effectively generate fine-grained sensor data.