反射震測法常用於探測地下構造並建構地下速度模型，在反射震測資料處理流程中，最關鍵的步驟之一為速度分析。速度分析的品質好壞將直接關係到最終地下構造的成像品質。傳統的速度分析方法為先計算出波線資料的速度譜(velocity semblance spectrum)後，經由人工判讀點選速度譜相對聚焦處以得知各個地層相對應的反射波走時及地層速度。然而，反射震測測線往往動輒數十至數百公里長，若完全經由人工分析將曠日費時。此外，目前國際學界雖已積極利用深度學習方法開發人工智慧方式挑選震測地層速度，但尚無完全克服在受雜訊影響下開發出可信賴之自動化速度分析方法。本研究旨在借助深度學習方法，開發快速有效的自動化速度分析方法，以期在不影響最終成像品質的狀況下，加速反射震測資料處理流程，以利研究者能更快速地進行後續震測地層分析。
本研究將反射震測速度分析問題視為圖像辨識中的語意分割(semantic segmentation)問題。輸入資料為速度譜，運用加入注意力門(Attention Gates, AGs)的U-Net神經網路模型對速度譜中的各個資料點進行分類。另外，因無法得知真實地下構造，野外資料皆為無標註資料(unlabeled data)。因此，本研究使用有限差分數值模擬之震測資料(synthetic seismic data)進行訓練，並在部分資料中加入隨機雜訊以貼近真實情況。經訓練後於測試資料集可得精確率(precision)、召回率(recall)及F1值分別為0.947、0.590及0.727，經過訓練後的網路可辨別真實地層訊號與複反射訊號造成的速度譜振幅高點。於1995年TAICRUST計畫所搜集之EW9509-1測線實測結果可發現，本研究提出之模型可於沉積盆地（沖繩海槽及弧前盆地）之淺層地層有良好的速度預測結果，而於缺乏反射面之火山島弧區域(琉球島弧)及深部地層有較差的預測結果，此可能與本研究模型未考慮波浪雜訊及模型最大速度僅止於每秒五公里有關，未來可進一步改善以更適合現地震測資料使用。

The seismic reflection method is often used to detect underground structures and construct subterranean velocity models. In the workflow of reflection seismic data processing, one of the most critical steps is velocity analysis. The quality of velocity analysis will directly affect the imaging quality of the final underground structure. The traditional velocity analysis method is to calculate the velocity semblance spectrum from original seismic data, manually interpret and pick the relative local semblance spectrum amplitude high and derives the corresponding reflected wave travel time and velocity of each horizon. However, due to tens to hundreds of kilometers long seismic profiles, manual picking velocity is time-consuming and has a less spatial resolution. This study aims to develop a fast and effective automatic velocity analysis method by using deep learning techniques, producing a high-quality velocity model and accelerating the reflection seismic data processing workflow.
In this study, we considered velocity analysis problem as a semantic segmentation problem. The input data is the velocity semblance spectrum, and each data point in the input is classified using the U-Net architecture with Attention Gates (AGs). In addition, due to underground structures being unknown, the field seismic data are unlabeled. Thus the training data of this study is composed of synthetic seismic data and data with random noise from finite difference simulation. After training, results show precision of 0.947, recall of 0.590 and F1 score of 0.727, respectively. Our model can recognize velocity semblance spectrum amplitude highs, picking accurate velocity and avoiding water column multiples. In addition to a long-offset multichannel seismic profile EW9509-1 collected by TAICRUST project in 1995, our model behaves well in sedimentary basins but behaves poorly in arc volcanism area and deep sedimentary basins. This is probably because our training dataset did not consider deep (limited to 5 km/s in velocity) or complex crustal structures and did not include water column noise generated from tide and artificial sources, which can be improved in the future.

［1］G. J. Crutchley and H. Kopp, Reflection and Refraction Seismic Methods, In: Submarine Geomorphology, A. Micallef, S. Krastel and A. Savini (Eds.), Cham: Springer International Publishing, 2018.
［2］Ö. Yilmaz, Seismic Data Analysis: Processing, Inversion, and Interpretation of Seismic Data, Tulsa, OK, USA: Society Exploration Geophysicists, 2001.
［3］海洋學門資料庫：〈海洋學門資料庫─震測資料查詢〉，取自https://www.odb.ntu.edu.tw/seisdb/test/。
［4］Y. K. Chen, "Automatic velocity analysis using high-resolution hyperbolic Radon transform," Geophysics, Vol. 83, No. 4, Jul 2018, pp. A53-A57.
［5］G. Fabien-Ouellet and R. Sarkar, "Seismic velocity estimation: A deep recurrent neural-network approach," Geophysics, Vol. 85, No. 1, Jan 2020, pp. U21-U29.
［6］W. L. Wang, G. A. McMechan, J. W. Ma and F. Xie, "Automatic velocity picking from semblances with a new deep-learning regression strategy: Comparison with a classification approach," Geophysics, Vol. 86, No. 2, Mar 2021, pp. U1-U13.
［7］O. Ronneberger, P. Fischer and T. Brox, "U-Net: Convolutional Networks for Biomedical Image Segmentation," Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, Springer International Publishing, 2015, pp. 234-241.
［8］R. S. Ferreira, D. A. B. Oliveira, D. G. Semin and S. Zaytsev, "Automatic Velocity Analysis Using a Hybrid Regression Approach With Convolutional Neural Networks," IEEE Transactions on Geoscience and Remote Sensing, Vol. 59, No. 5, May 2021, pp. 4464-4470.
［9］F. Chollet, "Xception: Deep Learning With Depthwise Separable Convolutions," Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jul 2017.
［10］I. Goodfellow, Y. Bengio and A. Courville, Deep Learning, MIT Press, 2016.
［11］W. S. McCulloch and W. Pitts, "A logical calculus of the ideas immanent in nervous activity," The Bulletin of Mathematical Biophysics, Vol. 5, No. 4, Dec 1943, pp. 115-133.
［12］F. Rosenblatt, "The perceptron: a probabilistic model for information storage and organization in the brain," Psychological Review, Vol. 65, No. 6, Nov 1958, pp. 386-408.
［13］M. T. Hagan, H. B. Demuth and M. Beale, Neural network design, Boston: PWS Publishing Co., 1996.
［14］B. T. Polyak, "Some methods of speeding up the convergence of iteration methods," USSR Computational Mathematics and Mathematical Physics, Vol. 4, No. 5, Jan 1964, pp. 1-17.
［15］J. Duchi, E. Hazan and Y. Singer, "Adaptive Subgradient Methods for Online Learning and Stochastic Optimization," Journal of Machine Learning Research, Vol. 12, Jul 2011, pp. 2121-2159.
［16］G. Hinton, "Neural Networks for Machine Learning-Lecture 6a Overview of mini-batch gradient descent," 2012, http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf.
［17］D. P. Kingma and J. Ba, "Adam: A method for stochastic optimization," arXiv preprint arXiv:1412.6980, 2014.
［18］Y. LeCun, L. Bottou, Y. Bengio and P. Haffner, "Gradient-based learning applied to document recognition," Proceedings of the IEEE, Vol. 86, No. 11, Nov 1998, pp. 2278-2324.
［19］S. Minaee, Y. Boykov, F. Porikli, A. Plaza, N. Kehtarnavaz and D. Terzopoulos, "Image Segmentation Using Deep Learning: A Survey," IEEE Trans Pattern Anal Mach Intell, Vol. 44, No. 7, Jul 2022, pp. 3523-3542.
［20］T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár and C. L. Zitnick, "Microsoft COCO: Common Objects in Context," Computer Vision – ECCV 2014, Springer International Publishing, 2014, pp. 740-755.
［21］O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Misawa, K. Mori, S. McDonagh, N. Y. Hammerla, B. Kainz, B. Glocker and D. Rueckert, "Attention U-Net: Learning Where to Look for the Pancreas," arXiv, 2018.
［22］M. J. Park and M. D. Sacchi, "Automatic velocity analysis using convolutional neural network and transfer learning," Geophysics, Vol. 85, No. 1, Jan 2020, pp. V33-V43.
［23］F. Luporini, M. Louboutin, M. Lange, N. Kukreja, P. Witte, J. Huckelheim, C. Yount, P. H. J. Kelly, F. J. Herrmann and G. J. Gorman, "Architecture and Performance of Devito, a System for Automated Stencil Computation," Acm Transactions on Mathematical Software, Vol. 46, No. 1, Apr 2020, pp. 1-28.
［24］M. Louboutin, M. Lange, F. Luporini, N. Kukreja, P. A. Witte, F. J. Herrmann, P. Velesko and G. J. Gorman, "Devito (v3.1.0): an embedded domain-specific language for finite differences and geophysical exploration," Geoscientific Model Development, Vol. 12, No. 3, Mar 2019, pp. 1165-1187.
［25］A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai and S. Chintala, "PyTorch: An Imperative Style, High-Performance Deep Learning Library," Vol. 32, 2019.