Using NASA Satellite Data Sources and Geometric Deep Learning to Uncover Hidden Patterns in COVID-19 Clinical Severity

Ignacio Segovia-Dominguez, Huikyo Lee, Zhiwei Zhen, Yuzhou Chen, Michael J. Garay, Daniel J. Crichton, Rishabh Wagh, Yulia R. GelarXiv 2021 · arXiv

Abstract

This study introduces NASAdat, a benchmark dataset integrating NASA satellite-derived atmospheric data with county-level COVID-19 health records. By leveraging Geometric Deep Learning (GDL), specifically Spatio-Temporal Graph Neural Networks, the researchers demonstrate that atmospheric factors like Aerosol Optical Depth (AOD) significantly enhance the predictive accuracy of COVID-19 clinical severity forecasts across the United States. This work bridges the gap between climate science and epidemiology, offering a scalable framework for environmental health modeling.

Environmental factors such as air quality and temperature are known to influence respiratory health, yet integrating these into epidemiological models remains challenging. Two primary hurdles exist: the complexity of satellite data formats, which are often inaccessible to non-specialists, and the irregular spatial lattices of geographic data (e.g., interconnected county borders) that traditional grid-based deep learning models struggle to represent. This research addresses these gaps by providing a streamlined dataset and utilizing graph-based AI to uncover hidden links between atmospheric conditions and COVID-19 hospitalizations.

Methodology: Geometric Deep Learning on Irregular Lattices

The authors employ Geometric Deep Learning (GDL), specifically Graph Neural Networks (GNNs), to model the United States as a graph. In this representation, each node is a county and edges represent shared borders. Unlike standard models, GDL is designed for non-Euclidean spaces, making it ideal for capturing how disease spread and weather patterns flow across irregular geographic boundaries.

Construction of spatial-temporal graphs and climatological features for the NASAdat framework.

The study compared two classes of models: standard Recurrent Neural Networks (RNNs), such as LSTMs, and Spatio-Temporal GNNs, including DCRNN (Diffusion Convolutional Recurrent Neural Network) and GconvLSTM. The core experiment utilized "Granger causality" logic to determine if adding atmospheric variables produced more accurate forecasts than using hospitalization data alone.

Datasets: Introducing NASAdat

The study introduces NASAdat, a unique benchmark dataset designed for "plug-and-play" use by researchers. It integrates several key components:

  • Aerosol Optical Depth (AOD): Derived from NASA’s MODIS instrument, serving as a proxy for PM2.5 air pollution.
  • Surface Air Temperature & Relative Humidity: Derived from the AIRS instrument on the Aqua satellite.
  • COVID-19 Hospitalization Records: County-level data used to measure clinical severity.
Global distribution of atmospheric variables: (a) MODIS AOD, (b) AIRS temperature, and (c) AIRS relative humidity.

All data is mapped to FIPS codes, facilitating easy integration with other socio-economic or health datasets.

Key Results and Insights

The DCRNN model achieved the highest performance, demonstrating that GDL models outperform traditional LSTMs by "understanding" the spatial dependencies between neighboring counties.

The Aerosol Connection

The most striking finding was the predictive power of Aerosol Optical Depth (AOD). In many regions, high aerosol levels—often linked to poor air quality—were a strong leading indicator for increased COVID-19 severity. In states like Pennsylvania and Texas, the improvements gained by including atmospheric factors were statistically significant ($p < 0.05$).

Regional Sensitivity

A "model voting" approach revealed high consensus in industrial and high-traffic corridors, such as Houston, TX, and Philadelphia, PA, suggesting that air quality is a critical risk factor in these areas. The models also indicated that Pennsylvania and Texas showed higher sensitivity to poor air quality than California, likely due to differences in landscape heterogeneity and local pollution sources.

Discussion and Future Work

While the study highlights the potential of satellite-informed AI, limitations remain. Currently, data is averaged at the county level; finer resolution would provide deeper insights but presents technical challenges in data fusion. Furthermore, the study's U.S.-centric focus and the potential correlation between poor air quality and socio-economic disadvantage necessitate careful future work to ensure model fairness. The research team plans to expand NASAdat globally and explore more complex GDL architectures to further refine clinical forecasting.

Summary updated

Abstract

As multiple adverse events in 2021 illustrated, virtually all aspects of our societal functioning -- from water and food security to energy supply to healthcare -- more than ever depend on the dynamics of environmental factors. Nevertheless, the social dimensions of weather and climate are noticeably less explored by the machine learning community, largely, due to the lack of reliable and easy access to use data. Here we present a unique not yet broadly available NASA's satellite dataset on aerosol optical depth (AOD), temperature and relative humidity and discuss the utility of these new data for COVID-19 biosurveillance. In particular, using the geometric deep learning models for semi-supervised classification on a county-level basis over the contiguous United States, we investigate the pressing societal question whether atmospheric variables have considerable impact on COVID-19 clinical severity.
BibTeX
@misc{segoviadominguez2021using,
  author = {Ignacio Segovia-Dominguez and Huikyo Lee and Zhiwei Zhen and Yuzhou Chen and Michael J. Garay and Daniel J. Crichton and Rishabh Wagh and Yulia R. Gel},
  title = {{Using NASA Satellite Data Sources and Geometric Deep Learning to Uncover Hidden Patterns in COVID-19 Clinical Severity}},
  year = {2021},
  eprint = {2110.10849},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2110.10849},
}
Figure