Towards a deep learning approach for short-term data-driven spatiotemporal seismicity rate forecasting
Open Access
- 25 November 2025
- journal article
- research article
- Published by Springer Nature in Earth, Planets and Space
- Vol. 77 (1) , 185
Abstract
Recent advances in earthquake monitoring have led to the development of methods for the automatic generation of high-resolution catalogues. These catalogues are created at considerably reduced processing times and contain significantly larger volumes of data concerning seismic activity compared to standard catalogues created by human analysts. Disciplinary statistics and physics-based earthquake forecasting models have shown improved performance when rich catalogues are used. The use of high-resolution catalogues paired with machine learning algorithms, which have recently evolved due to the rise in the availability of data and computational power, is therefore a promising approach to uncovering underlying patterns and hidden laws within earthquake sequences. This study focuses on the development of short-term data-driven spatiotemporal seismicity forecasting models with the help of deep learning and tests the hypothesis that deep neural networks can uncover complex patterns within earthquake catalogues. The performance of the forecasting models is assessed using metrics from the data science and earthquake forecasting communities. The results show that deep learning algorithms are a promising solution for generating short-term seismicity forecasts, provided that they are trained on a representative dataset that accurately captures the properties of earthquake sequences. Comparisons of machine learning-based forecasting models with an epidemic-type aftershock sequence benchmark show that both types of models outperform the persistence null hypothesis commonly used as a benchmark in forecasting the behaviour of other types of non-linear systems. Machine learning forecasting models achieve similar performance to that of an epidemic-type aftershock sequence benchmark on the Southern California and Italy test datasets at significantly reduced processing times - a major advantage in applications to short-term operational earthquake forecasting. Graphical AbstractKeywords
Funding Information
- H2020 Marie Sklodowska-Curie Actions (955515)
This publication has 59 references indexed in Scilit:
- Utilizing deep learning for near real-time rainfall forecasting based on Radar dataPhysics and Chemistry of the Earth, Parts A/B/C, 2024
- Forecast evaluation for data scientists: common pitfalls and best practicesData Mining and Knowledge Discovery, 2022
- A deep learning approach to solar radio flux forecastingActa Astronautica, 2022
- SmaAt-UNet: Precipitation nowcasting using a small attention-UNet architecturePattern Recognition Letters, 2021
- CBAM: Convolutional Block Attention Module2018
- Solar irradiation nowcasting by stochastic persistence: A new parsimonious, simple and efficient forecasting toolRenewable and Sustainable Energy Reviews, 2018
- Net load forecasts for solar-integrated operational grid feedersSolar Energy, 2017
- A 27 day persistence model of near‐Earth solar wind conditions: A long lead‐time forecast and a benchmark for dynamical modelsSpace Weather, 2013
- The Collaboratory for the Study of Earthquake Predictability perspective on computational earthquake scienceConcurrency and Computation: Practice and Experience, 2010
- Evaluating Forecasting Methods2001