2018/12/26 by Bas Peters, Justin Granek, Peters, Bas +3 · 1 citation
Earth and Planetary Sciences · Engineering · #Seismic Imaging and Inversion Techniques #Reservoir Engineering and Simulation Methods #Hydraulic Fracturing and Reservoir Analysis
paper · pdf · doi:10.48550/arxiv.1812.11092
Detecting a specific horizon in seismic images is a valuable tool for\ngeological interpretation. Because hand-picking the locations of the horizon is\na time-consuming process, automated computational methods were developed\nstarting three decades ago. Older techniques for such picking include\ninterpolation of control points however, in recent years neural networks have\nbeen used for this task. Until now, most networks trained on small patches from\nlarger images. This limits the networks ability to learn from large-scale\ngeologic structures. Moreover, currently available networks and training\nstrategies require label patches that have full and continuous annotations,\nwhich are also time-consuming to generate.\n We propose a projected loss-function for training convolutional networks with\na multi-resolution structure, including variants of the U-net. Our networks\nlearn from a small number of large seismic images without creating patches. The\nprojected loss-function enables training on labels with just a few annotated\npixels and has no issue with the other unknown label pixels. Training uses all\ndata without reserving some for validation. Only the labels are split into\ntraining/testing. Contrary to other work on horizon tracking, we train the\nnetwork to perform non-linear regression, and not classification. As such, we\npropose labels as the convolution of a Gaussian kernel and the known horizon\nlocations that indicate uncertainty in the labels. The network output is the\nprobability of the horizon location. We demonstrate the proposed computational\ningredients on two different datasets, for horizon extrapolation and\ninterpolation. We show that the predictions of our methodology are accurate\neven in areas far from known horizon locations because our learning strategy\nexploits all data in large seismic images.\n