Akademik Veri Yönetim Sistemi
Pure and Applied Geophysics, 2026 (SCI-Expanded, Scopus)
Northwestern Türkiye is a region with high earthquake activity under the complex compressional and extensional tectonic regime of the Western Anatolian block between the Eurasian and African plates, with dominant control from the North Anatolian Fault Zone. In seismically active regions, accurate and timely P- and S-wave arrival picking is crucial for earthquake monitoring and parameter estimation. Although several deep learning algorithms have been proposed for this purpose, models trained in one region often degrade when applied to other regions, where the noise conditions and seismic network characteristics are different. In this study, we present TroyNet, a PhaseNet-derived seismic phase-picking architecture tailored for northwestern Türkiye. TroyNet features a U-Net-like architecture and consists of a dilated residual encoder, minimal skip connectivity, nearest-neighbor upsampling, and a positional-transformer bottleneck to capture distant temporal dependencies. For coping with extremely imbalanced classes, our network incorporates a customized class-weighted loss function during training. Noise-based clipping and SNR-based sample weighting were evaluated in controlled ablations; the final configuration excludes sample weighting. TroyNet was compared with a retrained PhaseNet baseline implemented and evaluated under identical conditions. TroyNet achieved higher P- and S-wave recalls with similar level of phase detection precision. These results demonstrate that incorporating region-specific data characteristics, architectural improvements, and tailored training can improve phase picking in noisy environments.