Modelling with data types

For Hoogheemraadschap De Stichtse Rijnlanden (HDSR), we created a 3D model of the subsurface below the northern Lekdijk, between Schoonhoven and Amerongen. Between 2017 and 2029, HDSR was and will be reinforcing the Lekdijk so that cities like Utrecht and Amsterdam remain optimally protected when the water level rises. HDSR is using the subsurface model to determine which reinforcement measures are appropriate in which locations, which locations require more data, and explain to local residents why measures are necessary.

The subsurface in this area is complex but easy to map, as there are many boreholes and core penetration tests (CPTs). And it offers opportunities to add more detail to the model. (Using all available data for this project, we were able to model at high resolution.)

Subsurface variation demands high resolution

To model the subsurface under the Lekdijk, we used 3D cells of 25 x 25 x 0.25 metres. This is remarkable, as the national GeoTOP model uses cells of 100 x 100 x 0.5 metres. (The resolution of this project is 32 times higher!) This higher resolution makes it possible to model the subsurface in greater detail. For an area featuring a high subsurface variation, like in this river region, this is of great importance.

Interpreting CPTs through machine learning

Modelling with this much detail does require more data than usual. Normally, we only use boreholes for modelling, but this time we also used CPTs. Like boreholes, CPTs provide information about the subsurface, but in a very different form. This meant we needed to handle the CPTs differently than boreholes. There are quite a few places in the area where boreholes and CPTs are right next to each other. At these locations, you can see how the measurements in a CPT translate into a described borehole. Using machine learning techniques, we trained an artificial neural network with these borehole-CPT pairs, to then be able to translate all the CPTs into information that would ordinarily be obtained from a borehole (i.e. what material the subsurface consists of). This time, we could use the CPTs together with the boreholes to, on the one hand, determine where the transitions between different geological units are located, and, on the other, to determine which materials (such as sand, clay, or peat) most likely make up the 3D cells of the model. The use of machine learning to analyse subsurface information generates good results and is a technique we want to develop further.

This is the model that predicts the materials in the subsurface. The colour green signifies clay; brown is peat; light yellow is fine sand; dark yellow is coarse sand; and orange is gravel.

More innovations

Besides these innovations, we investigated how to enrich an existing subsurface model with new data points to make the model’s predictions even better. And of course, after adding new data points, we wanted to know how well they fit the model, so we developed a statistical method for that. Finally, we are also looking closely at how we model riverbeds. For these elongated, narrow elements, we now include the direction of the river course (anisotropy) in the statistical calculations that predict the nature of the subsurface.

Cooperate with GDN?

Are you interested in using neural network analysis or machine learning in subsurface modelling? Or do you feel like a high-resolution subsurface model might answer your questions? Then please contact Willem Dabekaussen via the blue ‘mail directly’ button below.

Also see