SeisTask

SeisTask is a semi-synthetic, task-oriented, seismic time-series dataset for event detection. Each waveform contains either a signal or noise. For details about the dataset curation process, see the pre-print Meta-Learning to Address Data Shift in Time Series Classification (https://arxiv.org/abs/2601.09018). The dataset is composed of 102,060 waveforms. Each waveform is 5 seconds long sampled at 100 Hz (500 samples) in 2 dimensions (X and Y). Each waveform either contains a synthetic earthquake signal embedded in noise or is pure noise. The noise profiles are obtained from the Stanford Earthquake Dataset [1]. SeisTask is composed of 243 unique tasks, each containing 420 total waveforms, 210 of them contain signal and 210 are pure noise. It sets up a binary classification problem for studying adaptive learning strategies. The dataset is located in the .csv file, metadata included in the .h5 file. The indices of each are aligned such that the i'th index in the metadata file corresponds to the i'th index of hte data file. An example of load, filtering, and plotting some waveforms is provided. [1] Mousavi, S. M., Sheng, Y., Zhu, W., Beroza G.C., (2019). STanford EArthquake Dataset (STEAD): A Global Data Set of Seismic Signals for AI, IEEE Access, doi:10.1109/ACCESS.2019.2947848