Enhancing seismic signal quality assessment with CNNs: A focus on effective data augmentation techniques

• Three CNN models are developed for seismic signal quality assessment. • Domain-specific augmentations outperform generic ones for seismic data. • Time stretching and horizontal flipping significantly boost model generalization. • ResNet-Variant achieves 99.61% accuracy and AUC of 0.9999. • The method enhances automation and reliability in seismic data pipelines. The rapid expansion of seismic data from global networks demands reliable, scalable, and automated methods for evaluating signal quality, an essential step in ensuring accurate seismic analyses. In this work, we present a convolutional neural network (CNN) framework designed for fully automatic seismic signal quality assessment, utilizing data from the Engineering Strong-Motion (ESM) Database. We explore three custom CNN architectures–VGG-Variant, ResNet-Variant, and InceptionNet-Variant–and conduct a systematic examination of data augmentation techniques to enhance model generalization. Using the VGG-Variant as a baseline, we identify time stretching, time shifting, and horizontal flipping as particularly effective augmentations, while rotation, Gaussian noise addition, and frequency masking degrade performance. Applying the most beneficial techniques to all three architectures, the ResNet-Variant achieves state-of-the-art results, with a 99.61 % accuracy and an AUC of 0.9999. These findings highlight the critical influence of domain-tailored data augmentation on CNN-based seismic signal classification. By enabling automated, high-fidelity assessments, our approach reduces the burden of manual inspection and augments decision-making capabilities in seismology and related Earth science disciplines.