Enhanced First-Break Picking Using Hybrid Convolutional Neural Network and Recurrent Neural Networks

First-break (FB) picking plays an important role in many applications of seismic study. Different machine-learning-based methods have been proposed to solve FB picking problem. But little has been done to exploit convolutional neural networks (CNNs) and recurrent neural networks (RNNs) in hybrid form for FB picking classification. Moreover, the input samples’ requirement for different hybrid models is also underinvestigated. In machine learning, FB picking problems can be formulated as regression or classification. However, solving it using regression does not yield better results. Therefore, in this study, FB picking is formulated as a binary classification problem, and the state-of-the-art CNN and RNN are combined to form hybrid models to enhance FB arrival picking. Eight hybrid models and three standalone models are investigated using different numbers of features and input samples. Experimental results show that hybrid models have superior performance over the baseline CNN model. The maximum accuracy obtained is 95.94%, and the
$F1$
-score is 95.92% for CNN bidirectional long short-term memory (CNNBiLSTM). Applying convolutional operation after recurrent operation in the hybrid model exhibits robust performance for fewer (nine) input samples. Specifically, gated recurrent unit CNN (GRUCNN) is robust for a lesser (nine) number of input samples, and CNNLSTM shows better accuracy for ten input samples of amplitude in FB arrival classification. In addition, GRU shows higher performance scores compared with the baseline model. Furthermore, our findings recommend using a higher number of features with fewer samples or a higher number of samples with fewer features for efficient FB picking classification.