Towards a Deep Learning Approach for Data-Driven Short-Term Spatiotemporal Earthquake Forecasting

The development of novel deep learning-based earthquake monitoring workflows has led to a rapid increase in the availability of earthquake catalogue data. Earthquake catalogues are now being created by deep learning algorithms at significantly reduced processing times compared to catalogues built by human analysts and contain at least a factor of ten more earthquakes. The use of these rich catalogues has been shown to have led to improvements in the predictive power of statistical and physics-based forecasts. Combined with the increasing availability of computational power, which has greatly contributed to the recent breakthrough in the field of artificial intelligence, the use of rich datasets paired with machine learning workflows seems to be a promising approach to uncovering novel insights about earthquake sequences and discovering previously undetected relationships within earthquake catalogues.

Our focus is on employing deep learning architectures to produce high-quality earthquake forecasts. Our hypothesis is that deep neural networks are able to uncover underlying patterns within rich earthquake catalogue datasets and produce accurate forecasts of earthquakes, provided that a representative dataset that accurately reflects the properties of earthquake sequences is used for training. We use earthquake catalogue data from different geographical regions to build a time series of spatiotemporal maps of past seismicity. We then split this time series into training, validation, and test datasets in order to explore the ability of deep neural networks to capture patterns within sequences of seismicity maps and produce short-term spatiotemporal earthquake forecasts.

We assess the performance of the trained deep learning-based forecasting models by using metrics from the machine learning and time-series forecasting domains. We compare the trained models against a null hypothesis, the persistence model, which assumes no change between consecutive time steps and is commonly used as a baseline in various time series forecasting settings. The persistence null hypothesis has been proven to be a very effective model due to the fact that when only background seismicity is observed, there is very little change between consecutive time steps. We also evaluate the relative performance of different deep learning architectures and assess their suitability for dealing with our specific problem. We conclude that deep learning techniques are a promising alternative to disciplinary statistics and physics-based earthquake forecasting methods as, once trained, deep learning models have the potential of producing high-quality short-term earthquake forecasts within seconds. This realisation can influence the future of operational earthquake forecasting and earthquake predictability.