"Spectogram Geomagnetic Dataset"

"Abstract: This dataset contains a curated collection of geomagnetic spectrograms designed for the development and benchmarking of deep learning models in earthquake precursor detection. The data is derived from the national magnetometer network operated by the Meteorological, Climatological and Geophysical Agency (BMKG) of Indonesia, comprising 25 monitoring stations distributed across the archipelago.The dataset encompasses 1,972 spectrogram samples generated from 256 unique earthquake events with magnitudes ranging from $M4.0$ to $M7.0+$ occurring between 2018 and 2025. Raw three-component magnetometer data ($H, D, Z$) sampled at 1 Hz was processed through a 0.001\u20130.5 Hz bandpass filter to isolate the Pc3\u2013Pc5 frequency range. Each sample represents a 6-hour temporal window captured prior to seismic events, transformed into $224 \\times 224$ pixel spectrograms using Short-Time Fourier Transform (STFT) and normalized according to ImageNet statistics.To support multi-task learning, each sample is meticulously labeled for magnitude classification (4 classes: Normal, Moderate, Medium, Large) and azimuth direction (9 classes). The dataset specifically highlights Ultra-Low Frequency (ULF) bands (0.001\u20130.01 Hz), consistent with lithospheric emission theories and the Lithosphere-Atmosphere-Ionosphere Coupling (LAIC) hypothesis.Data Structure:Spectrogram Images: 1,972 normalized STFT images ($224 \\times 224$ pixels) in PNG\/TIFF format.Metadata Table: CSV file containing earthquake event IDs, magnitudes, azimuth directions, station coordinates, and temporal window timestamps.Geomagnetic Indices: Corresponding Kp, Dst, and F10.7 indices for each sample to facilitate magnetospheric noise discrimination.Potential Applications: This dataset is intended for researchers in geophysics, remote sensing, and machine learning to explore resource-efficient architectures, temporal attention modules, and physics-informed loss functions for disaster mitigation systems"