Abstract
Reliable observational climate datasets are essential for understanding past variability, detecting long-term trends, and assessing future climate risks. However, many historical climate records suffer from substantial spatial and temporal gaps, particularly before the mid-20th century when observational networks were sparse. Traditional interpolation techniques such as inverse distance weighting (IDW) or kriging are commonly used to fill missing values, yet they often produce overly smooth fields.
Recent advances in artificial intelligence (AI) have demonstrated strong potential for reconstructing missing geophysical data by learning spatial patterns from large datasets. In particular, neural network (NN)–based approaches have been used to reconstruct missing values in observational datasets by using transfer learning, i.e., by applying models trained on physically consistent climate model simulation data. Such approaches go beyond conventional statistical interpolation techniques in reconstructing climate fields such as temperature anomalies (Kadow et al., 2020).
Supported by EXPECT, several DL approaches have been developed and tested at DKRZ to reconstruct two historical observational datasets: HadEX-CAM, which contains temperature-based extreme indices over Europe, and GPCC, which provides global precipitation observations.
