2026-10-06T13:05:10Z https://api.figshare.com/v2/oai

oai:figshare.com:article/34041801 2026-10-01T09:37:37Z category_25870 item_type_3 month_year_10_2026

Processed seismic survey data for validating DDPM Jun Su (25155612) Ryoichiro Agata (9325472) Gou Fujie (16801691) Yasuyuki Nakamura (16801685) Seismology and seismic exploration seismic airgun This record contains the processed inputs, model outputs, logs, and test products supporting Su, Agata, Fujie & Nakamura, "Noise reduction for wide-angle OBS record sections with a denoising diffusion probabilistic model" (JGR: Machine Learning and Computation, in revision). Version 2 adds the ensemble-averaged DDPM outputs, the leave-one-pass-out (LOPO) references and outputs, the self-similarity and baseline test products, the direct U-Net checkpoint, and the environment files requested during review. All waveform arrays are stored in .npz files written by the Profiles class and can be read and visualized with the code archived at DOI: 10.5281/zenodo.20922233 (source code, trained DDPM checkpoint model-final.pt, GPL-3.0). An example of reading these files is given in the Jupyter notebook resultview_standalone.ipynb in that record. The raw SEG-Y survey data are distributed by the JAMSTEC Seismic Survey Database (DOI: 10.17596/0002069; cruise KH-24-JE03, line NT24, OBS 1–40, passes NT24-OBS1–5 and NT24-MCS1–2; cruise KR09-06, line A2, OBS 10).<h3 dir="ltr">Training logs and previews</h3><ul><li>converge.log — L1-L2 hybrid loss for the training set at every epoch (Fig. 2c).</li><li>validate.log — L1-L2 hybrid loss for the validation set, evaluated every 10 epochs (Fig. 2c).</li><li>result_images.tar — image previews of all 40 validation record sections (input, DDPM output, diversity-stacked reference).</li></ul><h3 dir="ltr">Single-run DDPM outputs (Figs. 1–7, Supplementary Figs. S1–S6)</h3><ul><li>result-test-move80-allshot.out.npz — 200-m validation set, 8 hold-out stations × 5 passes (40 sections). Index catalog below.</li><li>result-mcs-move80-allshot.out.npz — 50-m Noto test set (NT24-MCS), source-location-matched one-in-four shots, 40 stations.</li><li>result-nwp-move80-allshot.out.npz — NW Pacific test set (KR09-06, OBS 10), resampled to 250 Hz, traces 450–922.</li></ul><h3 dir="ltr">Ensemble-averaged DDPM outputs (Table 1 DDPM column, Supplementary Table S2)</h3>Reconstructions averaged over 20 independent ancestral samples, in the same Profiles format as the single-run files:<ul><li>ensemble.test.avgout.npz — scheme A (200-m validation).</li><li>ensemble.mcs.avgout.npz — scheme B (50-m Noto).</li><li>ensemble.nwp.avgout.npz — scheme C (NW Pacific).</li></ul><h3 dir="ltr">Repeated-sampling (self-similarity) test (Supplementary Section S3, Table S1)</h3><ul><li>self_similarity_test.tar.gz — the 20 per-run rebuild.out snapshots for the validation set, per-run metric tables (masked/unmasked RMS, hybrid loss, AIC and CC delays and misfit ratios), and the pairwise cosine-similarity reports (unmasked, direct-wave masked, masked + saturated at ±1 and ±0.3) as CSV/JSON.</li></ul><h3 dir="ltr">Baseline comparison (Table 1, Fig. 8, Supplementary Table S2)</h3> <ul><li>baseline.tar.gz — rebuild outputs and metric tables for the direct U-Net and the trace-wise DeepDenoiser (pretrained model) on schemes A–C, together with the pooled (Table 1) and station-wise (Table S2) evaluation outputs.</li><li>unet-direct_model-final.pt — checkpoint of the deterministic direct U-Net baseline (same backbone as the DDPM, trained 200 epochs on the same prestack/diversity-stacked pairs). Load with baseline/rebuild_unet_direct.py --checkpoint from the software record.</li></ul><h3 dir="ltr">Index catalog for the validation arrays</h3>Waveform array Profiles[n,:,:] (shot-station, time_sample, trace) in result-test-move80-allshot.out.npz and ensemble.test.avgout.npz:<table><tr><td>n=</td><td>NT24-OBS1</td><td>NT24-OBS2</td><td>NT24-OBS3</td><td>NT24-OBS4</td><td>NT24-OBS5</td></tr><tr><td>OBS 4</td><td>0</td><td>1</td><td>2</td><td>3</td><td>4</td></tr><tr><td>OBS 6</td><td>5</td><td>6</td><td>7</td><td>8</td><td>9</td></tr><tr><td>OBS 9</td><td>10</td><td>11</td><td>12</td><td>13</td><td>14</td></tr><tr><td>OBS 21</td><td>15</td><td>16</td><td>17</td><td>18</td><td>19</td></tr><tr><td>OBS 24</td><td>20</td><td>21</td><td>22</td><td>23</td><td>24</td></tr><tr><td>OBS 26</td><td>25</td><td>26</td><td>27</td><td>28</td><td>29</td></tr><tr><td>OBS 27</td><td>30</td><td>31</td><td>32</td><td>33</td><td>34</td></tr><tr><td>OBS 28</td><td>35</td><td>36</td><td>37</td><td>38</td><td>39</td></tr></table> 2026-10-01T09:37:37Z Dataset Dataset 10.6084/m9.figshare.34041801.v1 https://figshare.com/articles/dataset/Processed_seismic_survey_data_for_validating_DDPM/34041801 GPL 3.0+