Variable seismic waveforms representation: Weak-supervised learning based seismic horizon picking

Seismic horizon picking via deep learning models have been advanced rapidly and proven popular. However, the prediction result is highly depended on the quality of the train set and the manually interpreted train set is very difficult to obtain. To guarantee a reasonable prediction result, we usually need manually interpret 20%–30% seismic data to generate the train set for the convolutional neural network (CNN) based seismic horizon picking methods, which is very time consuming. Besides, the performance of the deep learning based seismic horizon picking methods are also greatly influenced by the architecture of the selected CNN model. The core theory of using CNN for seismic horizon picking is recognizing different waveform patterns and classifying them with the corresponding labels. In this work, we first develop the variable waveform representation (VWR) algorithm to simulate the waveform features and then produce a large number of seismic traces as the train data. The VWR can deal with the task of the train set generation by only using very few seismic traces. Afterward, we propose a novel CNN model by integrating the holistically-nested module and the SegNet model for solving seismic horizon picking, which is termed as the HSegNet model. It should be noted that the holistically-nested module is able to recognize the holistic image and learn the multiple features, which is efficient for the waveform pattern recognition. To demonstrate the efficiency and robustness of our proposed HSegNet model, we apply it to field data and compare with the widely used U-Net model. The experiment results show that our proposed VWR-HSegNet can achieve the prediction result with 99% test accuracy by only manually interpreting 0.13% seismic traces. Whereas, the U-Net model can only get 91% test accuracy prediction result but using 30% manually interpreted seismic traces as the train set.