Predictive modelling of grain-size distributions from marine electromagnetic profiling data using end-member analysis and a radial basis function network

Abstract

In this work, we present a new methodology to predict grain-size distributions from geophysical data. Specifically, electric conductivity and magnetic susceptibility of seafloor sediments recovered from electromagnetic profiling data are used to predict grain-size distributions along shelf-wide survey lines. Field data from the NW Iberian shelf are investigated and reveal a strong relation between the electromagnetic properties and grain-size distribution. The here presented workflow combines unsupervised and supervised machine-learning techniques. Non-negative matrix factorization is used to determine grain-size end-members from sediment surface samples. Four end-members were found, which well represent the variety of sediments in the study area. A radial basis function network modified for prediction of compositional data is then used to estimate the abundances of these end-members from the electromagnetic properties. The end-members together with their predicted abundances are finally back transformed to grain-size distributions. A minimum spatial variation constraint is implemented in the training of the network to avoid overfitting and to respect the spatial distribution of sediment patterns. The predicted models are tested via leave-one-out crossvalidation revealing high prediction accuracy with coefficients of determination (R 2) between 0.76 and 0.89. The predicted grain-size distributions represent the well-known sediment facies and patterns on the NW Iberian shelf and provide new insights into their distribution, transition and dynamics. This study suggests that electromagnetic benthic profiling in combination with machine learning techniques is a powerful tool to estimate grain-size distribution of marine sediments.

Quantitative interpretation of geoelectric inverted data with a robust probabilistic approach

“…Therefore, highly localized sampling of certain materials may lead to small classification zones in the immediate vicinity of a borehole. Although MLP presents large accuracy scores using the spatial training data, which indicates that the distribution of the boreholes reflected the real material distribution of the synthetic case, this is rarely the case in the field (Gahegan, 2000;Cracknell and Reading, 2014;Baasch et al, 2018), especially in heterogeneous landfills where abrupt vertical and lateral variations of materials may be present.…”

“…The soil is predicted at some small areas of high chargeability (which might be artifacts in the inverted model), and as it was the class found at the largest depth of a pit at x ∼ 0 m, then the soil was predicted in the area nearby at larger depths. This is a consequence of including spatial training data that are not distributed over the entire survey area (Baasch et al, 2018) and which may not reflect the real distribution of the materials deposited in the landfill. The waste also is predicted at larger depths in the model and might be influenced by the intermediate resistivity values.…”

Field and Synthetic Waveform Tests on Using Large‐Offset Seismic Streamer Data to Derive Shallow Seabed Shear‐Wave Velocity and Geotechnical Properties

“…Typical offshore site exploration has been done by borehole measurements, electromagnetics, seismic reflection, and multibeam surveys (Baasch et al., 2018; Brown et al., 2011; Han et al., 2020; Liao et al., 2008). These methods can be very localized, time‐consuming, and expensive.…”