RT Journal Article SR Electronic(1) A1 Xu, Yinpo A1 Yin, Cheng A1 Pan, Yingjie A1 Ni, Yudong A1 Zou, Xuefeng A1 Yang, TianfuYR 2021 T1 First‐break automatic picking technology based on semantic segmentation JF Geophysical Prospecting, VO 69 IS 6 SP 1181 OP 1207 DO https://doi.org/10.1111/1365-2478.13088 PB European Association of Geoscientists & Engineers, SN 1365-2478, AB ABSTRACT With the wide application of the high‐density and high‐productivity acquisition technology in the complex areas of oil fields, the first‐break picking of massive low signal‐to‐noise data is a great challenging job. Conventional first‐break automatic picking methods (Akaike information criterion method, energy ratio method, correlation method and boundary detection method) require a lot of manual adjustments due to their poor anti‐noise performance. A lot of adjustments affect the accuracy and efficiency of picking. First‐break picking takes up about one‐third of the whole processing cycle, which restricts petroleum exploration and development progress severely. In order to overcome the above‐mentioned shortcoming, this paper proposes the first‐break automatic picking technology based on semantic segmentation. Firstly, design the time window for primary wave and pick a certain quantity of first breaks from newly acquired data in different zones of the exploration area using the commonly used Akaike information criterion method and interactive adjustments; and then perform pre‐processing on the data within the time window to extract multiple first‐break attribute features and perform feature enhancement, to obtain multi‐dimensional features data blocks, at the same time, label the first breaks. Secondly, u‐shaped architecture network‐like encoding and decoding network is used to implement end‐to‐end feature learning from primary wave attribute data to first‐break labels. The encoding and decoding process of the encoding and decoding network is used to fuse the extraction and feature positioning of primary wave attribute features. At the same time, normalize each layer and use the rectified linear unit function as a non‐linear factor to improve the generalization and sensitivity of network model to low signal‐to‐noise primary waves. Finally, an optimized deep network model is used to predict the first breaks of the data to improve the accuracy and efficiency of the first‐break picking. This method innovatively fuses the multiple conventional automatic picking methods (Akaike information criterion method, energy ratio method, correlation method and boundary detection method) to extract multiple attribute features of primary wave, and improves the accuracy of the training network model to the first‐break detection using the improved UNet‐like encoding and decoding network. The feasibility of the new method is proved by model data. A comparative test is conducted between the new method and the Akaike information criterion method with the actual data, which verifies that the method in this paper has a higher picking accuracy and stable first‐break processing capability for the data with low signal to noise, our method shows a significant advantage when applied to low signal‐to‐noise seismic records from high‐productivity acquisition and can meet the demands of the accuracy and efficiency for near‐surface model building and static calculation of massive data., UL https://www.earthdoc.org/content/journals/10.1111/1365-2478.13088