Evaluation of statistical climate reconstruction methods based on pseudoproxy experiments using linear and machine learning methods

Abstract

Abstract. Three different climate field reconstruction (CFR) methods employed to reconstruct North Atlantic-European (NAE) and Northern Hemisphere (NH) summer season temperature over the past millennium from proxy records are tested in the framework of pseudoproxy experiments derived from three climate simulations with Earth System Models. Two of these methods are traditional multivariate linear methods (Principal Components Regression, PCR and Canonical Correlation Analysis, CCA), whereas the third method (Bidirectional Long-Short-Term Memory Neural Network, Bi-LSTM) belongs to the category of machine learning methods. The Bi-LSTM method does not need to assume linear and temporally stable relationships between the underlying proxy network and the targeted climate field, in contrast to PCR and CCA. In addition, Bi-LSTM incorporates information on the serial correlation of the time series. All three methods tested herein achieve reasonable reconstruction performance in both spatial and temporal scale. Generally, the reconstruction skill is higher in regions with denser proxy coverage, but reconstruction skill is also achieved in proxy-free areas due to climate teleconnections. All three CFR methodologies generally tend to more strongly underestimate the target temperature variations as more noise is introduced into the pseudoproxies. The Bi-LSTM method tested in our experiments shows relatively worse reconstruction skills compared to PCR and CCA, yet it brings some encouraging results on capturing extreme cooling climate signals. This indicates that this nonlinear CFR method could be a potential methodology for past climate extremes analysis.

Reconstructed Jing River streamflow from western China: A 399-year perspective for hydrological changes in the Loess Plateau

“…With the successful introduction of machine learning algorithms from artificial intelligence science in dendrochronological studies, reconstruction methods have been extended and improved in recent years Jevšenak and Levanič, 2016). In particular, the nonparametric methods such as Long Short-Term Memory (LSTM) (Nasreen et al, 2021;Zhang et al, 2022), Random Forest (RF) Zhao et al, 2022) and Artificial Neural Network (ANN) (Salehnia & Ahn, 2022;Zhang et al, 2000) algorithms have been applied to overcome the limitations of traditional linear models bring higher accuracy and extensibility. Nevertheless, their application of nonparametric methods is restrictive, and the single algorithm is not universally applicable in every tree ring reconstruction.…”