O'Sullivan et al. (2025) Efficient Likelihood and Machine‐Learning Models for Spatiotemporal Rainfall Estimation and Imputation

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: International Journal of Climatology
• Year: 2025
• Date: 2025-08-15
• Authors: Brian O'Sullivan, Gabrielle Kelly
• DOI: 10.1002/joc.70072

Research Groups

Not specified in the provided text.

Short Summary

The study develops and evaluates a likelihood-based imputation method and a DeepKriging approach to efficiently handle missing values in large spatiotemporal precipitation datasets.

Objective

• To improve computational efficiency and accuracy in imputing missing values within high-dimensional spatiotemporal rainfall data while maintaining the ability to generate global models for interpretability and prediction.

Study Configuration

• Spatial Scale: Irish rainfall network consisting of 474 stations.
• Temporal Scale: Monthly precipitation totals.

Methodology and Data

• Models used:
  • Likelihood-based imputation (Maximum Likelihood Estimation via Expectation-Maximisation algorithm, utilizing block composite likelihood and block Toeplitz decomposition).
  • Spatiotemporal DeepKriging (Deep Neural Network with spatiotemporal basis functions).
  • Baseline regression methods: Elastic-Net Chained Equations (ENCE) and Multiple Imputation by Chained Equations with Direct Use of Regularised Regression (MICE DURR).
• Data sources: Monthly precipitation totals from an Irish rainfall network.

Main Results

• ENCE and MICE DURR showed the highest accuracy during 10-fold cross-validation.
• The accuracy advantage of regression-based methods (ENCE, MICE DURR) diminished when validating against temporal or spatiotemporal gaps.
• The likelihood-based method and DeepKriging provided superior interpretability and the capacity for prediction beyond existing station locations by fitting a global spatiotemporal model.

Contributions

• Proposes a computationally efficient likelihood-based framework for large-scale rainfall data imputation.
• Introduces a DeepKriging approach using neural networks and basis functions to capture spatiotemporal dependencies.
• Demonstrates that global spatiotemporal models offer advantages over local regression-based imputation in terms of spatial prediction and robustness to specific gap types.

Funding

Not specified in the provided text.

Citation

@article{OSullivan2025Efficient,
  author = {O'Sullivan, Brian and Kelly, Gabrielle},
  title = {Efficient Likelihood and Machine‐Learning Models for Spatiotemporal Rainfall Estimation and Imputation},
  journal = {International Journal of Climatology},
  year = {2025},
  doi = {10.1002/joc.70072},
  url = {https://doi.org/10.1002/joc.70072}
}