RNNS For The Analysis Of Geological Long-Term Time Series

Abstract

Geological processes model huge flows of temporal data that are indispensable to the understanding of Earth's dynamical systems, natural hazards of predicting and resources exploration. Geological time series related data involve many of the complex time nonlinear dependency not easily accommodated by traditional statistical methods, as well as lots of long-term time series. In this paper, we consider the use of Recurrent Neural Networks (RNNs) for the analysis of geological time series and their usefulness in terms of their ability to capture temporal dependencies and make assumptions on what can happen in the future regarding geological events (for example). Different RNN architectures incorporating the Long Short-Term Memory (LSTM) network and the Gated Recurrent Unit (GRU) networks were implemented and trained using seismic activity, sedimentation and mineralogical datasets. The results show superior performance of R N N-based models in comparison to the conventional autoregressive and moving average methods with regards to the forecasting accuracy and pattern recognition. Practical implications are the problem of the higher prediction of hazard, the estimation of resources and climate reconstruction. The paper also highlights some limitations such as sensitivity of missing data, computational intensity and required large, labelled data. Future work will focus on developing hybrid models using RNNs and attention mechanisms and combining it with remote sensing data to improve predictive performance.