[Paper Review] Spatio-Temporal Super-Resolution Data Assimilation (SRDA) Utilizing Deep Neural Networks with Domain Generalization

Yuki Yasuda, Ryo Onishi|arXiv (Cornell University)|Dec 7, 2022

Meteorological Phenomena and Simulations4 citations

TL;DR

This paper proposes a physics-informed deep learning framework, Spatio-Temporal Super-Resolution Data Assimilation (4D-SRDA), that combines low-resolution physics-based simulations with a neural network to infer high-resolution atmospheric and oceanic states. By integrating domain-generalized data augmentation via SR-mixup, the method achieves robust, computationally efficient super-resolution inference even under distribution shifts, outperforming baseline methods in accuracy and speed with minimal computational cost.

ABSTRACT

Deep learning has recently gained attention in the atmospheric and oceanic sciences for its potential to improve the accuracy of numerical simulations or to reduce computational costs. Super-resolution is one such technique for high-resolution inference from low-resolution data. This paper proposes a new scheme, called four-dimensional super-resolution data assimilation (4D-SRDA). This framework calculates the time evolution of a system from low-resolution simulations using a physics-based model, while a trained neural network simultaneously performs data assimilation and spatio-temporal super-resolution. The use of low-resolution simulations without ensemble members reduces the computational cost of obtaining inferences at high spatio-temporal resolution. In 4D-SRDA, physics-based simulations and neural-network inferences are performed alternately, possibly causing a domain shift, i.e., a statistical difference between the training and test data, especially in offline training. Domain shifts can reduce the accuracy of inference. To mitigate this risk, we developed super-resolution mixup (SR-mixup)--a data augmentation method for domain generalization. SR-mixup creates a linear combination of randomly sampled inputs, resulting in synthetic data with a different distribution from the original data. The proposed methods were validated using an idealized barotropic ocean jet with supervised learning. The results suggest that the combination of 4D-SRDA and SR-mixup is effective for robust inference cycles. This study highlights the potential of super-resolution and domain-generalization techniques, in the field of data assimilation, especially for the integration of physics-based and data-driven models.

Motivation & Objective

• To develop a computationally efficient method for high-resolution inference in atmospheric and oceanic simulations using low-resolution physics-based models.
• To address domain shift issues in data assimilation caused by statistical discrepancies between training and inference data.
• To enhance robustness of neural network inference in hybrid physics-informed and data-driven modeling through domain generalization.
• To reduce computational cost compared to traditional ensemble-based data assimilation while maintaining high-resolution accuracy.
• To enable practical deployment on large-scale 3D Earth system data by exploring patch-based training strategies.

Proposed method

• The framework alternates between physics-based simulations and neural network inference to generate high-resolution spatio-temporal states from low-resolution inputs.
• A deep neural network is trained end-to-end to perform super-resolution and data assimilation simultaneously, using supervised learning with high-resolution reference data.
• SR-mixup is introduced as a data augmentation technique that creates synthetic training samples by linearly combining random input pairs, improving domain generalization.
• The neural network architecture combines convolutional layers, pixel shuffle upsampling, and transformer-based attention blocks for effective feature extraction and resolution enhancement.
• The method avoids ensemble simulations, significantly reducing computational cost compared to traditional ensemble Kalman filters (EnKFs).
• Training is performed on full 2D flow fields in the current implementation, with future work targeting patch-based training for scalability to 3D systems.

Experimental results

Research questions

• RQ1 Can a deep neural network effectively perform spatio-temporal super-resolution and data assimilation in a physics-informed framework without requiring ensemble simulations?
• RQ2 How does domain shift between training and inference data affect the robustness of neural network inference in hybrid physics-data models?
• RQ3 To what extent does SR-mixup improve generalization and reduce error under distribution shift in Earth system modeling?
• RQ4 How does the computational cost of 4D-SRDA compare to traditional ensemble-based data assimilation methods like EnKF?
• RQ5 Can the proposed framework be extended to 3D Earth system data using patch-based training strategies?

Key findings

• The 4D-SRDA framework achieved high-resolution inference with significantly lower computational cost than ensemble Kalman filters (EnKFs), while maintaining accuracy comparable to HR EnKF under sparse observation conditions.
• When observations were sparse, the error of ST-SRDA became comparable to that of HR EnKF, indicating near-optimal performance under limited observational constraints.
• The wall time of ST-SRDA was substantially shorter than that of EnKFs, demonstrating its computational efficiency.
• Without SR-mixup, the error increased substantially, confirming that domain generalization via SR-mixup is essential for robust inference in the presence of distribution shifts.
• The method demonstrated consistent performance across different observation densities and was robust to noise and statistical discrepancies between training and test data.
• The framework provides a scalable and efficient alternative to ensemble-based methods, with potential for extension to 3D atmospheric and oceanic models through patch-based training.