Temporal changes of radon (222Rn) concentration in subsurface air are analyzed in comparison with ensuing seismicity on the basis of neural network methods. The data for analysis are represented by five-year series of one-site radon measurements, seismic activity observations and meteorological data. Changes in radon concentration are described quantitatively by windowed estimates of temporal, statistical and complexity features, which serve as material for the training neural network classification method to divide into conditionally “strong” and “weak” daily seismicity following in time. The multilayer perceptron based classification model tuning involves selection of informative features, search for optimal sizes of their estimation windows, and analysis of the seismicity categorization threshold. Two types of classification models were studied, which differed by the method of partitioning into conditionally “strong” and “weak” seismic events – partitioning based on the magnitude of events and intensity in points. It is shown that the best classification is achieved on a limited set of statistical and complexity features and reaches 83% for the neural network model based on intensity in points. We conclude that accurate feature extraction from temporal changes of subsurface radon concentration can give implicit precursors of earthquakes.
Shepelev et al. (Sun,) studied this question.