Effects of seismic filters on pretrained CNN models for fault detection: application to the Delft 3D seismic volume, the Netherlands

Accurate mapping of faults from 3D seismic volumes is critical for identifying structural traps, assessing reservoir compartmentalization, and guiding drilling decisions. This study evaluates how common seismic conditioning filters affect both the visual quality of a production Delft 3D poststack time-migrated volume and the outputs of two pretrained 3D CNN fault detectors: UNet3D and FaultNet. Seven filters were applied independently—AJAX, DeSmile, SimpleDenoise, Dip-Steered Median Filter (DSMF), Edge-Preserving Smoother (EPS), Fault Enhancement, and Ridge Enhancement—and results were inspected across inline, crossline, time-slice, and full-cube perspectives. The analysis is image-driven, supported by high-resolution comparison figures and view-specific interpretations.Filters that suppress incoherent noise while preserving dip-aligned continuity and edge gradients, particularly DSMF and EPS, produce the most interpretable inputs for UNet3D and yield thin, continuous fault traces with low internal clutter. SimpleDenoise provides conservative conditioning that improves signal-to-noise without creating artificial discontinuities, while DeSmile stabilizes outputs in areas affected by migration-smile curvature. In contrast, contrast-enhancing operators (AJAX and Fault/Ridge Enhancement) increase the detectability of weak lineaments but also broaden the apparent fault response and emphasize nonfault edges that require interpreter screening. Across all conditioning strategies, FaultNet produces a larger, higher-recall set of candidates that is useful for lead generation but requires curation, whereas UNet3D tends to return thinner, more connected masks when background noise is reduced, and edges are preserved. These findings provide practical guidance for selecting conditioning filters that improve the usability of pretrained CNN fault predictors on production seismic data.