Reconstructed precipitation isotopes in China during the past six decades based on machine learning

Shengjie Wang, Yanqiong Xiao, Yuqing Qian, Hongyang Li, Cunwei Che, Xiaofan Zhu, Mingjun Zhang,
Reconstructed precipitation isotopes in China during the past six decades based on machine learning,
Journal of Hydrology,
Volume 662, Part B,
2025,
134003,
ISSN 0022-1694,
https://doi.org/10.1016/j.jhydrol.2025.134003.
(https://www.sciencedirect.com/science/article/pii/S0022169425013411)
Abstract: The stable isotopes of hydrogen and oxygen in precipitation are efficient tools for understanding the hydrological processes. However, the precipitation isotope observations are discontinuous, which restricts the analysis of long-term trends and spatial patterns of water isotopes. During the past decades, China has been a hot spot for precipitation isotope measurement, which provides a data basis for reconstructing the continuous time series of water isotopes. Using the machine learning methods, here we developed a monthly time series product of stable hydrogen and oxygen isotopes in precipitation from 1961 to 2020 with a spatial resolution of 0.5° (latitude) by 0.5° (longitude), i.e., C-Isoscape PreTS. Approximately 7000 months of precipitation isotopes at more than 200 sampling stations in China were compiled as the training and validation sets in machine learning. Among the three machine learning methods used (XGBoost, MLP, and SVM), the XGBoost method shows great potential to predict the precipitation isotopes, as indicated by the 5-fold cross-validation (R2 = 0.73 and 0.71, for hydrogen and oxygen isotopes). The new nationwide meteoric water line with long-term representativeness is determined as δ2H = 7.7δ18O + 4.0. Compared to several existing global products based on geostatistics and isotope-enabled climate models, the C-Isoscape PreTS product has a good performance across China. This machine learning-based precipitation isotope product may offer valuable insights for studies interested in long-term hydrological and isotopic trends as well as the spatial incoherence of isotope-related hydrometeorological phenomena.
Keywords: Stable isotope; Precipitation; Machine learning; China; Time series