Data Quality Aware Deep Learning for Reliable Seismic Event Detection

Reliable seismic event detection from continuous waveform data remains fundamentally challenged by heterogeneous data quality across sensor networks, where variations in instrumentation performance, environmental noise, and recording conditions degrade detection accuracy and prediction calibration. Conventional deep learning approaches assume uniform input quality, leading to unstable confidence estimates and reduced performance when faced with degraded signals characteristic of operational monitoring. This paper presents a quality aware deep learning framework that integrates waveform reliability into both model architecture and training dynamics through dual pathway integration. Each three-component seismic trace is assigned a scalar quality score Q \in [0 1] computed from three interpretable indicators like signal-to-noise ratio, glitch fraction, and clipping fraction which is concatenated with CNN extracted latent features to provide quality conditioning while simultaneously weighting the training loss to prioritize high fidelity examples during optimization. Implemented as a compact one-dimensional CNN with three convolutional blocks and global average pooling, the framework is evaluated on waveforms from the Southern California Seismic Network spanning the full quality spectrum Q ε [0.322, 1.0]. The quality aware model achieves F1=0.949, average precision AP = 0.989, and ROC-AUC= 0.989. Quality stratified evaluation reveals robust performance across degradation levels, ranging from AP = 0.592 on the lowest quality decile (Qε[0.322, 0.497]) to AP>0.999 on high-quality data (Q>0.955), demonstrating the framework’s ability to maintain detection capability under challenging conditions where signal reliability is compromised. Calibration analysis demonstrates well aligned predicted probabilities with expected calibration error (ECE)<0.02, ensuring reliable confidence estimates for operational decision-making. The proposed framework contributes a transparent, reproducible approach for quality aware learning that extends beyond seismology to any sensor-based monitoring domain challenged by heterogeneous data conditions.