Deep-learning velocity model building by jointly using seismic first arrivals and early-arrival waveforms

Seismic first arrivals and early-arrival waveform data contain rich velocity structure information and are widely used in velocity model building. Such methods based on first arrivals are stable but have low resolution, while seismic waveform methods have high accuracy but low stability. A joint use of first arrivals and seismic waveforms may improve both the stability and accuracy of velocity model building. However, first arrivals and waveforms are different in physical meanings and data structures, hence it is difficult to combine them directly in conventional traveltime tomography or waveform inversion. In this paper, we propose a deep-learning velocity model building method by jointly using seismic first arrivals and early-arrival waveforms. By transforming the dimensions of the heterogeneous data in feature datasets, the first arrivals and early-arrival waveforms are input into the neural network through different channels for building the same velocity model. The results of numerical tests show that the accuracy of our method is better than that using only the first-arrivals or only the early-arrival waveforms in deep -learning modeling and traveltime tomography, demonstrating the complementing advantage of jointly using first arrivals and early-arrival waveforms in velocity model building. The idea of using heterogeneous data for deep learning velocity model building here could be useful for combining other data in high-precision velocity model building.