Impact of Gaussian Transformation on Cloud Cover Data Assimilation for Historical Weather Reconstruction

Abstract

Abstract Old descriptive diaries are important sources of daily weather conditions before modern instrumental measurements were available. A previous study demonstrated the potential of reconstructing historical weather at a high temporal resolution by assimilating cloud cover converted from descriptive diaries. However, cloud cover often exhibits a non-Gaussian distribution, which violates the basic assumptions of most data assimilation schemes. In this study, we applied a Gaussian transformation (GT) approach to cloud cover data assimilation and conducted observing system simulation experiments (OSSEs) using 20 observation points over Japan. We performed experiments to assimilate cloud cover with large observational errors using the Global Spectral Model (GSM) and a local ensemble transform Kalman filter (LETKF). Without GT, meridional wind and temperature exhibited deteriorations in the lower troposphere compared with the experiment with no observations. In contrast, GT reduced the 2-month root-mean-square errors (RMSEs) by 5%–15% throughout the troposphere for wind, temperature, and specific humidity fields. Significant improvements include zonal wind at 500 hPa and temperature at 850 hPa with 6.4% and 7.3% improvements by GT, respectively, compared with the experiment without GT. We further demonstrate that the additional GT application to the precipitation background field improves precipitation estimation by 12.2%, with pronounced improvements over regions with monthly precipitation of less than 150 mm. We also explored the impact of cloud cover GT on a global scale and confirmed improvements extending from around the observation sites. Our results demonstrate the potential of GT in high-resolution historical weather reconstruction using old descriptive diaries. Significance Statement To reconstruct the historical weather, cloud cover information from old diaries can be used by incorporating high-resolution model simulations. However, cloud cover is not normally distributed and violates an important assumption when combining cloud cover observations with model simulations. Our results demonstrate that transforming the cloud cover distribution into a normal distribution could improve wind speed, temperature, and humidity fields in the model. We demonstrate the critical role of the transformation in a nonnormally distributed variable when combined with models and show the potential of diary-based weather information to reconstruct historical weather.