![Hier klicken, um den Treffer aus der Auswahl zu entfernen](images/unchecked.gif) |
Titel |
Nationwide lithological interpretation of cone penetration tests using neural networks |
VerfasserIn |
Peter-Paul van Maanen, Jeroen Schokker, Ronald Harting, Renée de Bruijn |
Konferenz |
EGU General Assembly 2017
|
Medientyp |
Artikel
|
Sprache |
en
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 19 (2017) |
Datensatznummer |
250144624
|
Publikation (Nr.) |
EGU/EGU2017-8473.pdf |
|
|
|
Zusammenfassung |
The Geological Survey of the Netherlands (GSN) systematically produces 3D stochastic
geological models of the Dutch subsurface. These voxel models are regarded essential in
answering subsurface-related questions on, for example, aggregate resource potential,
groundwater flow, land subsidence hazard and the planning and realization of large-scale
infrastructural works. GeoTOP is the most recent and detailed generation of 3D voxel
models. This model describes 3D stratigraphical and lithological variability up to a
depth of 50 m using voxels of 100 × 100 × 0.5 m. Currently, visually described
borehole samples are the primary input of these large-scale 3D geological models, both
when modeling architecture and composition. Although tens of thousands of cone
penetration tests (CPTs) are performed each year, mainly in the reconnaissance
phase of construction activities, these data are hardly used as geological model
input.
There are many reasons why it is of interest to utilize CPT data for geological and
lithological modeling of the Dutch subsurface, such as: 1) CPTs are more abundant than
borehole descriptions, 2) CPTs are cheaper and easier to gather, and 3) CPT data are more
quantitative and uniform than visual sample descriptions.
This study uses CPTs and the lithological descriptions of associated nearby undisturbed
drilling cores collected by the GSN to establish a nationwide reference dataset for
physical and chemical properties of the shallow subsurface. The 167 CPT-core
pairs were collected at 160 locations situated in the North, West and South of the
Netherlands. These locations were chosen to cover the full extent of geological units and
lithological composition in the upper 30 to 40 m of the subsurface in these areas.
The distance between the CPT location and associated borehole is small, varying
between 0 and 30 m, with an average of 6 m. For each 2 cm CPT interval the data
was automatically annotated with the lithoclass from the associated core using a
lithological classification script that is also used in GeoTOP to classify the visual sample
descriptions.
Based on this data a three-layer feedforward neural network was trained containing
5 different inputs: cone resistance, friction ratio, coordinates x and y, and interval depth z.
Previous training attempts showed an increased performance when using additional inputs
such as pore water pressure, but since these variables are not measured in the majority of
CPTs, these were left out in the training procedure. The Newton conjugate-gradient algorithm
was applied to train the network. 20-Fold cross-validation yielded 20 different trained
nets and independent performance outcomes. Significant performance increase
was found as compared to performances of conventional lithological classification
charts.
A similar neural network was then applied to new CPT data from a pilot area in the city of
Rotterdam. This area has a limited number of visual sample descriptions and therefore,
additional lithological information of the subsurface is desirable. The results of an
evaluation of the neural network’s outcomes in this area by geological experts are
positive, which paves the way for future nationwide application of this method. |
|
|
|
|
|