First-arrival time detection based on data-driven enhanced Akaike information criterion

First-arrival time picking is a fundamental and critical step in seismic data processing. The conventional Akaike Information Criterion (AIC) method is widely used due to its computational simplicity and lack of fixed threshold requirements. However, its stability and accuracy are limited by the artificially designed features. We propose a data-driven enhanced AIC algorithm (DEAIC) to improve the accuracy of first-arrival time detection. DEAIC utilizes the encoder of a denoising autoencoder as an adaptive feature extractor to map original seismic signals into a latent space. Parallel AIC detection is performed on multi-channel latent representations to obtain candidate first-arrival times. The high-precision first-arrival time can be estimated through density-based clustering and weighted fusion. DEAIC combines the unsupervised learning algorithm with the AIC algorithm to simultaneously leverage the strengths of the data-driven approach and physical knowledge. The experiment results demonstrate that DEAIC exhibits higher precision and stability on the field seismic dataset compared to the conventional AIC methods.