Enhanced Video Inpainting: A Deep Learning Approach for Historical Weather Reconstruction

AbstractWe investigate the applicability of deep learning (DL) methods for reconstructing daily weather data. Inspired by video inpainting, we propose a novel method, WeRec3D, which utilizes a three‐dimensional convolutional neural network. Our approach was developed iteratively by evaluating seven modeling improvement techniques. The resulting method reduces the validation error by 67% compared to a two‐dimensional baseline, decreasing the error from RMSE = 0.4620 and MAE = 0.311 to RMSE = 0.1527 and MAE = 0.1093. Additionally, we demonstrate the impact of the spatial distribution of observations on reconstruction accuracy and propose a potential integration with the analogue resampling method. WeRec3D is trained and validated in a self‐supervised manner using ERA5's surface temperature and pressure data over Europe. On a hold‐out set from 1950 to 1954, the validation results in an MAE of 1.11°C and 199 Pa. As a case study, we reconstruct the 1807 heat wave and validate it using a leave‐one‐out method in space. Compared to the original data, the reconstructed time series exhibit a correlation of at least 0.91, with a maximum normalized RMSE and standard deviation delta of 0.58 and