Detection of possible hydrological precursor anomalies using long short-term memory: A case study of the 1996 Lijiang earthquake

Historical data on underground fluid can be valuable for detecting potential hydrological precursors. Although different analysis methods have been employed, nonlinear dynamic fluctuations and the lack of auxiliary observations makes hydrological precursors difficult to identify. In this study, we investigated hydrological time series of data from six monitoring sites to identify possible precursors to the Lijiang earthquake. Two long shortterm memory (LSTM) models were developed to simulate and forecast hydrological variations, with and without the use of auxiliary data. The results indicate that both models are useful and valid in forecasting short-term hydrological variations. The LSTM model developed using only historical data can provide good accuracy and validity in simulating and forecasting short-term hydrological variations when no auxiliary meteorological observations are available. The results show that observations made during seismically inactive periods can be better modeled than those made during active periods. Furthermore, the exponentially weighted moving average control chart was employed as a detector to identify anomalous signals. The results show that there was a false precursor in the time series from the six monitoring points (wells and springs). Hydrological anomalies at four of the wells occurred 1-6 months before the Lijiang earthquake and can be attributed to changes in tectonic stress. However, the anomalies at one of the other wells may be unrelated to this earthquake. An anomalous signal observed at a spring could not be related to earthquakes and may have been caused by other factors. This study shows that LSTM models are useful in detecting possible hydrological precursor anomalies.