dc.description.abstract | Knowledge of the subsurface is an important aspect for a countries
welfare. To gain an understanding of this, geologists use boreholes com-
bined with interpolating methods to build statistical 3D models. Due
to the stochastic properties of these models, domain experts perform a
quality control procedure to find errors, which can be a time consuming
endeavor.
In this thesis, we looked into methods for predicting areas of errors
in GeoTOP, a geological voxel model. Firstly, we show that a previously
used Attention Model performs well when we optimize parameters for
each participant, but with an AUC of 0.61, the algorithm lacked finding
optimal parameters for the combined participants. We show that variance
among experts in assessing errors is high, making generalizing predictions
hard. Secondly, we showed that entropy, the voxel models quantification
of uncertainty, is not a good indicator of where errors occur. With an
average AUC of 0.54, where some participants scored even under 0.5,
we show that there is no relation between entropy and the assessment of
experts. Finally, we introduced a Velocity-Threshold Identification (I-VT)
algorithm combined with tree-based classifiers and showed that with an
AUC of 0.8 over each participant, errors can be found regardless of the
differences among participants. We show why finding optimal parameters
for fixation algorithms is difficult due to a lack of ground truth, but despite
that our new algorithm performs better and faster, allowing for real-time
error predictions. These findings suggest that a geologist combined with
our introduced algorithm can decrease their time spent on quality control.
Furthermore, this thesis can provide as a framework for other fields with a
similar problem description, such as radiologists looking for malignancies. | |