Explainable AI for microseismic event detection

TL;DR

The paper tackles the interpretability challenge of deep learning-based microseismic event detection by applying Grad-CAM and SHAP to PhaseNet, revealing that the model attends to physically meaningful P- and S-wave features and that SHAP attributions align with seismic polarization. It introduces a SHAP-gated inference scheme that fuses the model's output with an explanation-based metric, achieving a higher F1 score (0.98) than the baseline (0.97) on a 9,000-window test set and demonstrating improved noise robustness. The work demonstrates that explainability techniques can both validate model decisions against geophysical knowledge and directly improve detector performance, offering a pathway toward explainable-by-design seismic AI systems. This approach provides a template for integrating XAI into critical geoscience applications and may generalize to other detectors and architectures, such as Transformer-based models.

Abstract

Deep neural networks like PhaseNet show high accuracy in detecting microseismic events, but their black-box nature is a concern in critical applications. We apply explainable AI (XAI) techniques, such as Gradient-weighted Class Activation Mapping (Grad-CAM) and Shapley Additive Explanations (SHAP), to interpret the PhaseNet model's decisions and improve its reliability. Grad-CAM highlights that the network's attention aligns with P- and S-wave arrivals. SHAP values quantify feature contributions, confirming that vertical-component amplitudes drive P-phase picks while horizontal components dominate S-phase picks, consistent with geophysical principles. Leveraging these insights, we introduce a SHAP-gated inference scheme that combines the model's output with an explanation-based metric to reduce errors. On a test set of 9,000 waveforms, the SHAP-gated model achieved an F1-score of 0.98 (precision 0.99, recall 0.97), outperforming the baseline PhaseNet (F1-score 0.97) and demonstrating enhanced robustness to noise. These results show that XAI can not only interpret deep learning models but also directly enhance their performance, providing a template for building trust in automated seismic detectors.

Figures (7)

• Figure 1: Schematic illustration of the network architecture. The input consists of 30-s three-component seismograms sampled at 100 Hz, yielding an input dimension of 3 × 3001. The network outputs three probability sequences of equal length, corresponding to P-pick, S-pick, and noise classes. Blue rectangles indicate neural network layers. Arrows denote operations, as summarized in the lower right corner. The model comprises four stages dedicated to down-sampling and four stages for up-sampling. Down-sampling is carried out using 1-D convolutions with a kernel size of 7 and a stride of 4. In contrast, up-sampling is performed through deconvolutions, which serve to restore the sequence length from the previous stage. Skip connections combine feature maps from each down-sampling stage with the corresponding up-sampling stage (indicated by dashed rectangles), thereby aiding in convergence. The last layer utilizes a softmax activation function to produce class probabilities (adapted from zhu2019phasenet).
• Figure 2: Grad-CAM visualizations of PhaseNet attention on the vertical (Z) component for three representative cases: (a) a high-SNR event with clear P- and S-wave arrivals, (b) a low-SNR event, and (c) a noise-only window. The heatmaps show Grad-CAM attribution intensity (red shading) overlaid on the waveform amplitude. Darker shades correspond to stronger model attention. For high-SNR events, attention is sharply concentrated around the P-wave onset with secondary activation near the S arrival; for low-SNR events, the attention remains aligned with the arrivals but becomes broader and less intense; for noise, attention is diffuse and unstructured, indicating no consistent phase-like focus.
• Figure 3: Zoomed-in Grad-CAM views of the vertical (Z) component for (a) a high-SNR and (b) a low-SNR event, corresponding to Figure \ref{['fig:gradcam3']}a–b. The heatmaps emphasize the specific areas where the model exhibits the greatest emphasis in distinguishing between P- and S-waves arrivals. The high-SNR example illustrates a distinct and clearly defined activation centered on the arrival of the P component, while, in the low-SNR scenario, there is a broader and less pronounced area of focus, which suggests a decrease in model confidence when faced with noisy conditions.
• Figure 4: Distributions of absolute SHAP values ($|\phi|$) for 5,000 signal and 5,000 noise windows. Panels (a,b) show P-class attributions, and panels (c,d) show S-class attributions, separated by component (E, N, Z). For signal windows, vertical (Z) contributions slightly dominate the P-class, while horizontal components (E and N) dominate the S-class. Noise windows show SHAP magnitudes that are low across all components, which means that there is no clear evidence for either phase class. These distributions complement the summary statistics in Table \ref{['tab:shap_summary']}.
• Figure 5: Violin plots of Shapley value distributions for 5,000 signal and 5,000 noise windows. (a) P-class SHAP distributions for signal windows; (b) P-class for noise; (c) S-class for signal; (d) S-class for noise. Signal windows show strong and coherent SHAP concentrations: the vertical (Z) component dominates P-phase detections, while horizontals (E and N) dominate S-phase detections. Noise windows exhibit uniformly low SHAP values across all components. Summary statistics are provided in Table \ref{['tab:shap_summary']}.