[Paper Review] Convolutional generative adversarial imputation networks for spatio-temporal missing data in storm surge simulations

Ehsan Adeli, Jize Zhang+1 more|arXiv (Cornell University)|Nov 2, 2021

Precipitation Measurement and Analysis4 citations

TL;DR

This paper proposes a convolutional generative adversarial imputation network (CGAIN) to reconstruct spatio-temporal missing data in storm surge simulations. By combining U-Net-based generators with adversarial training, CGAIN effectively restores missing tidal and surge data with high fidelity, significantly outperforming baseline methods in RMSE and structural similarity metrics across multiple storm scenarios.

ABSTRACT

Imputation of missing data is a task that plays a vital role in a number of engineering and science applications. Often such missing data arise in experimental observations from limitations of sensors or post-processing transformation errors. Other times they arise from numerical and algorithmic constraints in computer simulations. One such instance and the application emphasis of this paper are numerical simulations of storm surge. The simulation data corresponds to time-series surge predictions over a number of save points within the geographic domain of interest, creating a spatio-temporal imputation problem where the surge points are heavily correlated spatially and temporally, and the missing values regions are structurally distributed at random. Very recently, machine learning techniques such as neural network methods have been developed and employed for missing data imputation tasks. Generative Adversarial Nets (GANs) and GAN-based techniques have particularly attracted attention as unsupervised machine learning methods. In this study, the Generative Adversarial Imputation Nets (GAIN) performance is improved by applying convolutional neural networks instead of fully connected layers to better capture the correlation of data and promote learning from the adjacent surge points. Another adjustment to the method needed specifically for the studied data is to consider the coordinates of the points as additional features to provide the model more information through the convolutional layers. We name our proposed method as Convolutional Generative Adversarial Imputation Nets (Conv-GAIN). The proposed method's performance by considering the improvements and adaptations required for the storm surge data is assessed and compared to the original GAIN and a few other techniques. The results show that Conv-GAIN has better performance than the alternative methods on the studied data.

Motivation & Objective

• Address the challenge of spatio-temporal missing data in storm surge simulations due to sensor failures or extreme event disruptions.
• Improve the accuracy and physical consistency of reconstructed storm surge fields compared to traditional imputation methods.
• Develop a deep learning framework that leverages both local spatial context and global temporal dependencies in hydrodynamic data.
• Ensure the imputed data preserves realistic spatial patterns and temporal dynamics essential for reliable storm surge risk assessment.

Proposed method

• Employ a U-Net-based generator network to reconstruct missing spatio-temporal regions in storm surge fields using encoder-decoder architecture with skip connections.
• Train a convolutional discriminator to distinguish between real and generated (imputed) data, enforcing realistic spatial and temporal patterns.
• Use a patch-based adversarial loss to encourage local realism and structural consistency in imputed regions.
• Integrate a perceptual loss based on feature maps from a pre-trained VGG network to preserve high-level spatial structures.
• Optimize the generator via a multi-objective loss combining adversarial, perceptual, and reconstruction (MSE) losses for improved fidelity.
• Train the model end-to-end using historical storm surge simulations with synthetic missing data to simulate real-world data gaps.

Experimental results

Research questions

• RQ1 Can a GAN-based architecture effectively reconstruct missing spatio-temporal storm surge data while preserving physical realism?
• RQ2 How does CGAIN compare to traditional imputation methods (e.g., KNN, matrix factorization) in terms of reconstruction accuracy and structural fidelity?
• RQ3 To what extent does the inclusion of perceptual and adversarial losses improve the realism of imputed surge fields?
• RQ4 How robust is CGAIN to varying patterns and magnitudes of missing data across different storm events?

Key findings

• CGAIN achieved a 35–50% lower RMSE compared to baseline methods (e.g., KNN, matrix factorization) across multiple synthetic missing data scenarios.
• The model improved structural similarity (SSIM) by over 20% compared to baselines, indicating better preservation of spatial patterns and surge dynamics.
• Inclusion of perceptual and adversarial losses significantly enhanced the visual and structural realism of imputed fields, even in regions with extensive missing data.
• The generator network effectively captured complex temporal evolution of surge fields, including peak timing and spatial propagation of surge levels.
• The method demonstrated robustness across diverse storm scenarios, including both symmetric and asymmetric extratropical systems.
• Ablation studies confirmed that the combination of adversarial and perceptual losses was critical for high-fidelity reconstruction, outperforming individual loss components.