Reconstructing historical climate fields with deep learning

Area of Science:

• Climate Science
• Artificial Intelligence
• Data Science

Background:

• Historical climate records are often incomplete due to missing measurements.
• Existing methods for reconstructing climate data have limitations.

Purpose of the Study:

• To develop and evaluate a novel deep learning approach for reconstructing historical climate fields.
• To address the challenge of sparse and missing data in climate records.

Main Methods:

• Utilized a deep learning model based on Fourier convolutions.
• Trained the model on numerical climate model output.
• Applied the model to reconstruct large, irregular areas of missing climate data.

Main Results:

• Successfully reconstructed historical climate fields with high realism.
• Accurately reproduced significant climate events (e.g., El Niño/La Niña) with minimal input.
• Demonstrated superior performance compared to kriging and other machine learning methods.
• Showcased generalization to higher resolutions and unseen data patterns.

Conclusions:

• The deep learning approach offers a powerful new tool for climate data reconstruction.
• This method enhances our ability to study past climate events and fill data gaps.
• The model's flexibility allows application across various climate fields and data scenarios.