Automated picking of seismic first arrivals using single- to
multi-domain self-trained network
Over the last decade, the evolution of survey acquisition design
composed of numerous receivers and shot points has exponentially
increased the seismic data volume. As a result, the demand for efficient
algorithms for detecting seismic first arrivals has been growing.
Machine learning methods, especially convolutional neural networks, have
shown great potential in previous studies. However, most existing
methods lack generalization ability or be sensitive to noise. This paper
proposes a new picking strategy based on the single-to multi-domain
selftrained network to achieve automated picking for the first arrivals.
Our self-trained network learns from a single domain to multiple domains
while rejecting the wrong predicted picking and assigning the
pseudo-first-arrivals-times to the unlabeled dataset. The model inherits
the advantage from both single-and multi-domain picking strategies; it
has high generalizability and noise robustness. In addition, the model
is designed for a limited amount of manually annotated data to reduce
human labor and total seismic processing time. In experiments on an open
hard rock seismic multi-survey dataset, our picking method outperforms
the benchmark on most evaluation metrics. Remarkably, our approach
achieves approximately 10 % accuracy improvement against Lalor site
data with a sampling rate that differs from training data. The
experimental results demonstrate our method’s high generalizability and
robustness for unseen datasets.