Indonesia Target-Domain Seismic Waveform Dataset for Training-Free Domain Adaptation in TinyML Earthquake Detection

This dataset contains preprocessed three-component (Z, N, E) seismic waveform windows recorded at Indonesian broadband stations (BMKG–GEOFON network), used as the target-domain benchmark in the paper "Kernel Density Re-Embedding: A Training-Free Domain Adaptation Framework for TinyML Seismic Detection" (Bakti, Setyanto, Muhammad, and Wibowo, IEEE Open Journal of Signal Processing, 2026). Each event contains a paired 7-second (700-sample, 100 Hz) earthquake window and a 7-second pre-arrival ambient-noise window, extracted around the automatically picked P-wave arrival (recursive STA/LTA, 0.5 s/15 s, threshold τ = 2.5), quality-filtered by post-arrival energy continuity and a minimum SNR of 3.0 dB. Waveforms are amplitude-normalized to the peak absolute amplitude of a 9-second reference window [P − 1 s, P + 8 s]. From 40,430 raw candidate events downloaded from BMKG–GEOFON, 5,159 events (12.8% retention) passed all quality filters, yielding a balanced dataset of 10,318 windows split into: - Training split (indonesia_train_data.json): 4,127 event pairs, used to build the KDE reference density matrix and for the Full Retraining comparison baseline.- Blind test split (indonesia_test_data.json): 1,032 event pairs, held out and never seen during KDE re-embedding or retraining, used for all reported blind-test performance metrics. Each JSON file is keyed by a unique event ID (network_station_YYYYMMDD_HHMMSS) and contains normalized Z/N/E earthquake and pre-arrival noise windows plus metadata (network, station, P-arrival time, source file, SNR). The paper's KDE Re-Embedding and ESP32-S3 hardware deployment use only the Z-component fields. Raw seismic waveform data provided by Badan Meteorologi, Klimatologi, dan Geofisika (BMKG) and GEOFON network stations. Processing, quality filtering, and dataset curation performed by the authors for this study.