Reconstructing Historical Climate Data using Deep Learning

Understanding past climate conditions is essential for addressing future climate challenges. However, observational climate datasets often contain missing values, especially in older records, leading to incomplete and inaccurate analyses. Interpolation methods like kriging are commonly employed to address this issue by filling data gaps. Nevertheless, these approaches often fail to effectively reconstruct complex climatic patterns [1, 2].

This study leverages the transformative power of deep learning to accurately reconstruct two observational datasets. The first dataset is an intermediate product of HadEX3 [3], which contains gridded extreme indices over land regions, such as the TX90p index, corresponding to the percentage of days where daily maximum temperature is above the 90th percentile. The second dataset is the Full data GPCC product [4], containing global precipitation fields at monthly frequency. To reconstruct these two datasets with high accuracy, we employ and compare three deep learning approaches: a U-Net with partial convolutional layers, a diffusion model and a graph neural network. In all cases, models are trained on CMIP6 climate model data, evaluated on unseen CMIP6 and ERA5 data and compared to Kriging. The best-performing models are then applied to the observational datasets, providing new insights into historical climate conditions to inform more effective climate adaptation strategies. The reconstructed datasets are being prepared for the community in the framework of the H2020 CLINT project [5] and the Horizon Europe EXPECT project [6].

[1] Kadow C. et al., Nat. Geosci., 13, 408-413 (2020)
[2] Plésiat É. et al., Nat. Commun., 15, 9191 (2024)
[3] Dunn R.J.H. et al., J. Geophys. Res. Atmos., 125, 1 (2020)
[4] Schneider, U. et al., DOI: 10.5676/DWD_GPCC/FD_M_V2022_100 (2022)
[5] https://climateintelligence.eu/
[6] https://expect-project.eu/