Multiattribute probabilistic neural network for near-surface field engineering application

Abstract
Unconfined compressive strength (UCS) is an important rock parameter required in the engineering design of structures built on top or within the interior of rock formations. In a site investigation project, UCS is typically obtained discretely (through point-to-point measurement) and interpolated. This method is less than optimal to resolve meter-scale UCS variations of heterogenous rock such as carbonate formations in which property changes occur within data spacing. We investigate the geotechnical application of multiattribute analysis based on near-surface reflection seismic data to probe rock formations for their strength attributes at meter-scale variability. Two Late Jurassic outcrops located in central Saudi Arabia serve as testing sites: the Hanifa Formation in Wadi Birk and the Jubaila Formation in Wadi Laban. The study uses core and 2D seismic profiles acquired in both sites, from which we constrain UCS, acoustic velocity, density, and gamma-ray values. A positive linear correlation between UCS and acoustic impedance along the core indicates that seismic attributes can be utilized as a method to laterally extrapolate the UCS away from the core location. Seismic colored inversion serves as input for neural network multiattribute analysis and is validated with a blind test. Results from data at both outcrop sites indicate a high degree of consistency with an absolute UCS error of approximately 5%. We also demonstrate the applicability of predicted UCS profiles to interpret mechanical stratigraphy and map lateral UCS heterogeneities. These findings provide a less expensive alternative to constrain UCS from limited core data on a field-scale site engineering project.