Station-Level and Network-Wide SHAP Explanation of CNN Models for Seismic Cycle Monitoring: Evidence from Norcia 2016

Laboratory, theoretical and field data suggest that fault-zone properties should evolve during the seismic cycle as stress rises prior to failure and drops during earthquake rupture. Lab work shows systematic changes in elastic properties during the seismic cycle and that these changes can be used to predict lab earthquakes.  Recent work shows that in some cases these results are also applicable to tectonic faults. Seismic data show a clear distinction between fault zone properties pre and post mainshock, as well as post-seismic time-dependent changes in elastic properties. Here we extend these works by developing tools to distinguish seismic waves pre/post mainshock that are both predictive and physically interpretable. We train a convolutional neural network (CNN) on RGB spectrograms developed from three component seismograms recorded at seismic stations around the M6.5 2016 Norcia earthquake.  Our model can accurately distinguish foreshocks from aftershocks of the sequence. We train and test models on individual stations and also on all stations and subsets of the stations based on source-station geometry of the Norcia fault.  Models trained on the full set of stations achieve >99% accuracy for foreshock/aftershock classification and models trained on individual stations achieve higher accuracy in tests.  For each set we also performed the SHapley Additive exPlanations (SHAP) technique. We find that specific time-frequency signatures in the RGB spectrograms identify each class.  Here we extend that framework to a multi-station setting by training CNN models on spectrograms from several seismic stations surrounding the mainshock. While the multi-station model achieves high classification accuracy (about 97%), SHAP analysis reveals a substantial reorganization of feature importance, including strong station-dependent variability and a reduced contribution from aftershock-related regions.  Even for data from the reference station (NRCA), SHAP patterns differ markedly from those obtained in the single-station case, suggesting that heterogeneous training distributions alter global attribution mechanisms. To disentangle these effects, we additionally train station-wise CNN models, which achieve very high accuracy and produce more stable and physically coherent SHAP explanations. These results indicate that station-specific propagation effects play a key role in model interpretability and that caution is required when applying SHAP to models trained on spatially heterogeneous seismic datasets. The findings motivate future work toward hierarchical, region-aware, or physics-constrained interpretability frameworks.