Multi-Source Geophysical Parameter Dataset of Earthquake Cases in Sichuan-Yunnan Region for Deep Learning and Its Application

The Sichuan-Yunnan region is characterized by intense neotectonic movements and seismic activities. Over the past two decades, a large amount of geophysical observation data have been accumulated. Among them, earthquakes with magnitudes ≥4.5 are particularly concerned due to the significant damage they cause. Deep learning technology, based on the principle of data-driven, can mine the implicit features among data, such as the correlation between geophysical parameter characteristics and their variations and the occurrence of moderate to strong earthquakes. However, while seismic event-based single-sample seismic wave detection datasets are abundant, geophysical background datasets are currently relatively scarce.Based on this, in this article it uses the dataset of 798 earthquakes with magnitudes ≥4.5 that occurred in the Sichuan-Yunnan region over the past 20 years. Historical earthquake catalogs, gravity, faults, crustal velocity, crustal thickness, Moho depth, lithology, and groundwater with strong correlation to earthquake occurrence within a certain spatial range centered on the earthquake source were collected. Through data processing methods such as calculation, cleaning, and normalization, an annotated dataset was created. In order to ensure the balance of positive and negative samples, it also selected geophysical data from the same region with an equal number of non-significant earthquakes (magnitude 3 and below, which have a significant energy difference from earthquakes above magnitude 4.5) as positive examples, and created a labeled negative example dataset. Examples, negative examples, and data composition were elaborated based on accuracy. The evaluation indicators such as recall rate were used to analyze the performance of the dataset in four classic learning models, all of which achieved an accuracy of about 80%. Finally, the quality of the dataset was verified through transfer learning in other regions, which was not lower than the accuracy of the dataset testset. These indicate that the constructed dataset has good quality, applicability, and generalization. This article can provide reference for the construction of other deep learning seismology datasets.