Stead Distance Prediction Eval

stead-distance-prediction-eval

Exploring Challenges in Deep Learning of Single-Station Ground Motion Records — Çağlar et al. (2024) (arXiv:2403.07569, 2024)

What this evaluates

This benchmark evaluates whether deep learning models can accurately predict the epicentral distance of an earthquake from single-station ground motion waveforms. It specifically probes whether models learn intrinsic seismic features or merely exploit highly correlated auxiliary signals like P/S wave arrival times.

Datasets

• Stanford Earthquake Dataset (STEAD) — total ?; splits: train (-1), test (-1)

Metrics

• Mean Absolute Error (MAE) (primary) — range: other
  • L1 loss computed as the average absolute difference between predicted epicentral distance and true epicentral distance. The true distance is calculated using the haversine formula in kilometers.

Input / output format

Input: Three-channel ground motion time-series signals, optionally concatenated with P/S wave arrival time information.

Output: A single scalar value representing the predicted epicentral distance in kilometers.

Scoring recipe

def compute_mae(predictions, true_distances_km):
    # true_distances_km are precomputed via haversine formula
    mae = np.mean(np.abs(predictions - true_distances_km))
    return mae

Common pitfalls

• Models may achieve low error by simply learning the P-S arrival time difference rather than extracting features from the waveform itself.
• Epicentral distance must be computed using the haversine formula on spherical coordinates, not Euclidean distance.

Evidence (verbatim from paper)

For distance prediction, mean absolute error (L1 loss) is chosen and directly calculated using the haversine distance in kilometers between the epicenter and the recording station.

Citation

@misc{caglar2024groundmotion,
  title={Exploring Challenges in Deep Learning of Single-Station Ground Motion Records},
  author={Çağlar et al. (2024)},
  year={2024},
  note={arXiv:2403.07569}
}

• arXiv: 2403.07569