Data Reuse Case Studies

AI And Machine Learning


Estimating Cascading Hazards Chain (Xu el al. 2022)

Causal Bayesian network that depicts causal dependencies among different seismic ground failures, building damage, Damage proxy maps (DPMs), and environmental noises.
Causal Bayesian network that depicts causal dependencies among different seismic ground failures, building damage, Damage proxy maps (DPMs), and environmental noises. Read more.

N refers to N locations/pixels in the target area. The posterior probability of landslide, liquefaction, and building damage at each location are the objectives that the system estimates. Green boxes refer to the variables that have data constraints. Blue circles refer to nodes that are not observed or unknown. θ, ϕ, μ, η are the causal coefficients, which are unknown, that quantify the causal effects of parent nodes to landslide, liquefaction, building damage, and DPMs.


To Susu Xu and her team, the key to developing a ML algorithm to estimate earthquake induced cascading hazards chain is reproducibility. Their research uses satellite data to identify surface ground changes after different earthquakes worldwide. Their challenge was distinguishing the causes of these ground changes: building damage, liquefaction, landslides, or other secondary events that happen during earthquakes. To this end, the team developed an approach that combines statistical cause inference and geospatial hazard models, to differentiate multiple earthquake-induced cascading secondary hazards and building damage from the remote sensing observations, specially InSAR imagery. As ground truth data to validate their estimates of the causal mechanisms of damage they used field research data sources from multiple earthquakes (Hokkaido (2018), Ridgecrest (2019), Central Italy (2016), and Puerto Rico (2020)) in which researchers identified the spatial patterns of building damage, landslides, and liquefaction. Many of these ground truth datasets were gathered by GEER and STEER, the geotechnical and structural extreme events reconnaissance organizations of NHERI, and are hosted in DesignSafe. Xu explained that comparing results to those obtained by others is extremely difficult because researchers use different input data modalities (e.g., optical satellite imagery, SAR imagery, Aerial imagery/Streetview imagery) and models, and not all the models are open source. Therefore, having citable ground truth data that is permanently accessible via DesignSafe allows researchers to validate results consistently across research projects using the ground truth. In fact, the team cites other researchers that use this same ground truth to validate their own algorithms1. Reproducing results is an important issue, and having reliable and findable ground truth data that many can use to validate and compare results is one step along the path to reproducibility.

Publication Reusing Data:

Xu, Susu, et al. Seismic Multi-Hazard and Impact Estimation via Casual Inference from Satellite Imagery.” Nature Communications, vol. 13, no. 1, 1, Dec. 2022, p. 7793. www.nature.com, https://doi.org/10.1038/s41467-022-35418-8.

Dataset Reused:

Brandenberg, S., C. Goulet, P. Wang, C. Nweke, C. Davis, M. Hudson, K. Hudson, S. Ahdi, J. Stewart, (2021) "GEER Field Reconnaissance", in Ridgecrest, CA earthquake sequence, July 4 and 5, 2019. DesignSafe-CI. https://doi.org/10.17603/ds2-vpmv-5b34 v1

Miranda, E., J. Archbold, P. Heresi, A. Messina, I. Rosa, I. Robertson, K. Mosalam, T. Kijewski-Correa, D. Prevatt, D. Roueche. (2020) "StEER - Puerto Rico Earthquake Sequence December 2019 to January 2020: Field Assessment Structural Team (FAST) Early Access Reconnaissance Report (EARR)." DesignSafe-CI. https://doi.org/10.17603/ds2-h0kd-5677 v1

Miranda, E., A. Acosta Vera, L. Aponte, J. Archbold, M. Cortes, A. Du, S. Gunay, W. Hassan, P. Heresi, A. Lamela, A. Messina, S. Miranda, J. Padgett, A. Poulos, G. Scagliotti, A. Tsai, T. Kijewski-Correa, I. Robertson, K. Mosalam, D. Prevatt, D. Roueche. (2020) "StEER - 07 Jan. 2020 Puerto Rico Mw6.4 Earthquake: Preliminary Virtual Reconnaissance Report (PVRR)." DesignSafe-CI. https://doi.org/10.17603/ds2-xfhz-fz88 v1

1 Burrows, K., Milledge, D., Walters, R. J., & Bellugi, D. (2021). Improved rapid landslide detection from integration of empirical models and satellite radar. Nat. Hazards Earth Syst. Sci.Discuss.