Machine Learning and Deep Learning for Digitizing Scanned Images of Seismic Reflection Data

We present the use of Machine Learning algorithm of K-Nearest Neighbors (KNN) and Deep Learning algorithm of Artificial Neural Network for converting scanned images of seismic reflection data into digital seismic format. Digital seismic data format provides more flexibility for users to implement seismic processing algorithms like de-noising, enhancement and converting it into standard format, namely SEGY for further analysis such as seismic interpretation and seismic attributes generation. Varieties of seismic image in color density representation with different color maps and conventional wiggle display have been tested. Digitizing color density image consists of three main steps: recognizing Red-Green-Blue (RGB) representation in each pixel, creating a look-up table of RGB amplitude and substitute the RGB with a color that falls within the color-scale's dynamic range. Meanwhile, the approach is slightly different for image with wiggles representation. Several images’ attributes such as gradients and edge gradients are generated for better input uniqueness against known target amplitudes during model establishment. This pre-trained model is then used for predicting seismic amplitudes at specific pixel location in respect to a set of image attributes.The outcome of the conversion shows promising results. A qualitative interpretation for similarity check between input and output in terms of seismic events, horizon, faults, stratigraphy and other geological/geophysical features are used to validate the quality of digitized images.