Restoring Clipped Legacy Data with Machine Learning Legacy seismic data, especially when stored in 8-bit format, often suffers from severe clipping. This loss of dynamic range m...
Restoring Clipped Legacy Data with Machine Learning
Legacy seismic data, especially when stored in 8-bit format, often suffers from severe clipping. This loss of dynamic range makes interpretation more challenging and limits the use of advanced workflows such as spectral decomposition, seismic inversion, and other quantitative techniques.
Using Machine Learning, we can restore clipped seismic data to its full dynamic range—unlocking its true potential.
In this video, we showcase a 2D VGG19 UNet model trained on 256 × 256 sample images extracted from an onshore 3D dataset in The Netherlands. The training inputs consist of seismic tiles clipped by 5% and 10% (inline and crossline directions), with targets being the corresponding unclipped data. We then apply the trained model to an onshore dataset from Canada, clipped by 10% and stored in 8-bit format. Since the model lacks direct knowledge of the original seismic scaling, we compare the restored and original seismic volumes after RMS scaling.
This trained model—UNet VGG19 2D Unclip—will soon be available in OpendTect’s Machine Learning library, making it accessible to all of OpendTect's ML users.
Stay tuned!
dGB Earth Sciences#MachineLearning #SeismicData #Geophysics #OpendTect #DataRestoration #seismic #geoscience
