A machine-learning approach for the reconstruction of the ground shaking fields in real-time

Simone Francesco Fornasari
Veronica Pazzi
Giovanni Costa

Ground shaking field

Spatial representation of the effects of an earthquake in terms of a ground motion parameter
Post-emergency management
Limited number of seismic stations

Ground shaking maps provide a spatial representation of the effects of an earthquake in terms of ground motion parameters: even though they have been used to a wide spectrum of applications, their main purpose is related to post-emergency management. Traditional algorithms, such as ShakeMap, generate these maps from a limited number of observations and using ground motion prediction equations. One of the limitations of this approach is that it needs information about the magnitude and location of the event: to give an example these parameters are available to the Italian Civil Protection on average 12 minutes after the event. On the other hand the data from the stations are available at the computational center a few seconds after being recorded. We then tried to address this temporal gap of information developing an algorithm able to reconstruct the shaking maps in real-time exploiting the data from the stations alone.

Implemented workflow
(Model architecture adapted from Fukami et al., 2020)

Convolutional Neural Network

Voronoi tessellation

To do so we adopted a machine learning approach: our final model consists of an ensemble of 5 CNNs [NEXT], each one with the same architecture but trained independently on different dataset developed from the INGV ShakeMap archive. The data from the stations are retrieved from a monitoring system, such as Antelope, and preprocessed to extract a ground motion parameter at the station [NEXT]: since CNN cannot handle sparse data, we introduced a Voronoi tessellation of the station data to extend the information recorded at one station to all its nearest neighbours. A map of the active stations is used to introduce the information about the geometry of the network, and a Vs30 map is used as a proxy for the site effects. Each member of the ensemble output a reconstruction of the shaking map which is used to compute the outputs of the ensemble that are an average shaking map and its associated uncertainty.

Output comparison
(2016 M6.5 Norcia earthquake)

Simultaneous events

Robustness to network changes

Possible causes:

Data transmission problems
Addition/Removal of stations
Temporary problems

Real-time capabilities
(2016 M6.5 Norcia earthquake)

Conclusions

The developed method:

Can fill a "temporal gap" in the seismic monitoring
Has results comparable with (resampled) ShakeMap
Has useful feature for real-time applications
Is extensible (parameters and areas)

A paper about this work is ~~under review~~ published at BSSA Thanks for the attention and stay tuned!

Supplementary material

Applied Architecture

Based on Fukami et al. (2020)
Input:

Voronoi tessellation
Station position map
Vs30 map

Architecture:

5 models ensemble
4 layers per model
12 5x5 filters per layer

Training details

Optimizer used: Adam
Training dataset: 90% training and 10% validation (10-fold validation)
500 epochs limit (or early stopping on validation loss)
Loss function: $L = \frac{||\tilde{s}-s||_2}{||s||_2}$
48-sample batches used

Reconstruction uncertainty:

$$ \sigma = \sqrt{\sigma_{a}^2+\sigma_{e}^2}$$ $$\sigma_a = \mu\sqrt{e^{\ln(10)^2\sigma_{G}^2}-1}, \qquad \sigma_e=\sqrt{\frac{1}{m}\sum_{i=0}^m\mu_i^2 - \left(\frac{1}{m}\sum_{i=0}^m\mu_i\right)^2}$$