Convnetquake Eval

convnetquake-eval

Convolutional Neural Network for Earthquake Detection and Location — Perol et al. (2017) (arXiv:1702.02073, 2017)

What this evaluates

Evaluates a convolutional neural network's ability to detect seismic events versus background noise and classify their geographic origin using raw waveform data. It probes the model's generalization to unseen temporal periods and non-repeating seismic events.

Datasets

• Oklahoma Seismic Dataset (OGS) — total 834029; splits: train (-1), test (131281); repo https://github.com/tperol/ConvNetQuake

Metrics

• detection accuracy (primary) — range: percent
  • Percentage of windows correctly classified as either earthquake or noise across the entire test set.
• location accuracy — range: percent
  • Percentage of windows predicted as events that are correctly assigned to their true geographic cluster (classes 1-6).

Input / output format

Input: 10-second, 3-channel (HHZ, HHN, HHE) ground velocity waveform windows sampled at 100 Hz, normalized per month by subtracting the mean and dividing by the absolute peak amplitude.

Output: Discrete class label (0 for noise, 1-6 for geographic cluster) or probability distribution over the 7 classes.

Scoring recipe

def compute_metrics(preds, golds):
    correct_det = sum(1 for p, g in zip(preds, golds) if (p == 0) == (g == 0))
    det_acc = correct_det / len(golds)
    detected = [i for i, p in enumerate(preds) if p != 0]
    correct_loc = sum(1 for i in detected if preds[i] == golds[i])
    loc_acc = correct_loc / len(detected) if detected else 0
    return det_acc, loc_acc

Common pitfalls

• Location accuracy is calculated only on windows predicted as events, not on the full test set.
• The test set uses a strict temporal split (July 2014) rather than random shuffling, so results do not generalize to other months without retraining.
• The dataset is heavily imbalanced (~831k noise vs ~2.9k events), requiring noise injection augmentation to prevent overfitting.

Evidence (verbatim from paper)

The detection accuracy is the percentage of windows correctly classified as earthquake or noise. We obtain 74.5 % location accuracy on the test set (see Table[1]).

Citation

@misc{perol2017convnetquake,
  title={Convolutional Neural Network for Earthquake Detection and Location},
  author={Perol et al. (2017)},
  year={2017},
  note={arXiv:1702.02073}
}

• arXiv: 1702.02073