Early meteorological in-situ observation is obtained by manual timing recording, and the temporal resolution is low. The above leads to limitations in the quality of historical data produced by land-surface data assimilation systems. Aiming at the problem of low temporal resolution of in-situ observation in the past, we propose a deep learning-based method to improve the resolution of observation time series, Deep Interpolation. DI introduces a training strategy using uncertainty to weight loss functions to solve the "seesaw phenomenon" in multi-task learning and realize collaborative temporal downscaling of multiple observation variables, including temperature, surface pressure, relative humidity, and wind speed. The experimental results in China show that the DI has achieved significant performance improvement compared with the linear interpolation method and ERA5-land, and the average correlation coefficient is as high as 0.934. DI can accurately capture the hourly fluctuation characteristics of various meteorological variables, especially in high latitude, grassland, and cultivated land regions, showing more vital correction ability and more stable model performance. This study provides an effective solution for improving the temporal resolution of historical meteorological observation data. It lays a foundation for further enhancing the accuracy and reliability of climate models and land surface assimilation systems. Improving the temporal resolution of meteorological observation data can better support the research on climate change, the construction of disaster early warning systems, and decision-making in many fields, such as agriculture and energy.

Long-term analysis of climate trends and patterns relies on continuous and high-frequency observation data sets. Still, due to limitations in historical meteorological observation techniques and national policies, most weather stations worldwide can only provide three, four, or eight observations per day, hindering climate change research progress. To solve the problem of low-frequency daily observation in part of global meteorological stations, we propose a time-downscaling model of observation series based on deep learning, Land Surface Observation Simulator-Time Series Version (LOS-T), taking 2m air temperature as an example. LOS-T, combined with multimodal technology and Transformer architecture, effectively merges multiple types of data, including low-frequency observations, ERA5-land, and geographic information, to convert low-frequency observations into hourly high-frequency observations. The model showed significant accuracy improvements by training on millions of meteorological observations worldwide, especially on downscaling the data, which only has three observations per day. The results showed that LOS-T substantially improved over baseline models such as Bilinear and vanilla Transformer on several metrics such as MAE, RMSE, COR, and R2. In addition, case studies have confirmed that LOS-T can effectively utilize ERA5-land’s high-frequency temperature change information to improve the accuracy and robustness of predictions, even when there is a significant deviation between ERA5-land data and Ground Truth. In short, LOS-T provides new ways to refine global meteorological observation data and helps advance climate science.