Stochastic Windfarm Soil Property Estimation Via Deep Cnns

Estimating geotechnical properties of near-subsurface soils, such as cone-tip resistance, is crucial to the success of offshore windfarm development. While a common practice is to integrate 2D ultra-high-resolution (UHR) seismic lines and 1D cone-tip penetration testing (CPT) at sparse locations through seismic inversion and/or geostatistical analysis, the recent advance of deep learning, especially the convolutional neural network (CNN), has well established its values in various tasks of seismic image interpretation, including soil property estimation. However, most of the current CNN-based workflows fail to provide reliable uncertainty analysis, which is another essential component in ground model building from estimated soil properties. In this paper, we propose a new workflow that enables stochastic estimation of multiple soil properties and validate its performance on the public Borssele site within the Dutch Wind Farm Zone. Trained with 58 available CPT logs, the CNN successfully predicts the cone-tip resistance, sleeve friction, pore-water pressure and friction ratio as well as the associated posterior uncertainty for all 55 UHR seismic lines. As tested at the remining eight CPT locations, the machine prediction matches well with the actual measurements, confirming the capability of the proposed workflow in stochastic soil property estimation.