Deep CNN-based seismic channel detection by leveraging synthetic fluvial geometric properties

Manual mapping of channels and related fluvial features can be a challenge when their expressions are subtle, intertwined, and discontinuous. A deep convolutional neural network (CNN)-based pattern recognition workflow is presented to automate pixel‑level labeling of channels and related features. The ideas are demonstrated on a 3D seismic volume from north-central Oklahoma. The core of the workflow is a simple U‑Net (convolutional encoder-decoder network) trained to identify meandering shapes in coherency attribute images by utilizing geometric properties from synthetically generated fluvial terrains. The novelty of the paper is generating a large training dataset (~4000 images) locally by: a) extracting a representative channel, b) leveraging synthetically generated river geometries, c) stretching and squeezing the representative channel to match geometric properties, and d) embedding them into various background textures (noise) extracted from the same volume. The U-Net achieved high training accuracy by being exposed to a variety of signal-noise combinations. Generalization is further ensured by blind testing on previously unseen images. In both image sets –with and without exposed components – the U-Net satisfactorily delineated features of interest like distributaries, low‑sinuosity incised channels, and crevasse‑splay elements. Demonstrated on coherency, the workflow is easily extendable to other attributes and opens a way of reducing the burden of manual digitization in mapping complex fluvial systems. Providing the ability to automatically map these features in a matter of seconds, significantly reducing time consumption and subjectivity involved with this process.