On Wednesday 12 June, Koen van Asselt from Deltares gave a DigiShape Online Technical Session on his thesis research on predicting dune erosion using convolutional neural networks (CNNs). There was a lot of interest in this topic and a lively discussion ensued. In this interview, Koen talks about this research and why he believes everyone in the water world can work with machine learning.
Koen, why is it important to predict dune erosion?
‘Dunes provide a natural barrier against water. They protect people, buildings and infrastructure from flooding and are important for our safety. During heavy storms, you get transport of sediment, which means parts of the dunes disappear into the water. This is called dune erosion, and for the safety of the hinterland, it is important that we can predict how much damage a storm will cause.’
Why use convolutional neural networks for this?
‘CNNs and other machine learning algorithms are much faster than the numerical models we use now. With an approaching storm, speed is crucial for predicting scenarios and making timely decisions. By integrating CNNs with traditional models such as XBeach for sediment transport, we combine speed with accuracy.’
How did you approach predicting dune erosion using CNNs?
‘It was my thesis research, so we only had a few months. Nevertheless, we were able to take quite a few steps. We started by exploring the technical possibilities and ended up using U-net as a CNN. We then used XBeach as a synthetic data source to provide input to the CNN. We started with one storm scenario and four coastal profiles of the Dutch coast. Once we got this data feed working, which was not without a struggle, we started experimenting with larger quantities of coastal profiles. In the end, we managed to simulate a realistic storm scenario for the Dutch coast, based on four hundred storm profiles from XBeach.’
What do these results mean?
‘That using CNNs to predict dune erosion is promising. We have so far trained the CNN with a pittance of what information and variation is possible in the real world. But the fact that we now know how to run this kind of data through a CNN, and get output that matches what we would expect based on physics, gives confidence that we can move forward with this in the future. The ultimate goal is that in the future we will really help water managers make decisions in operations.’
In this, what do you think are the biggest challenges?
‘For me, as a young water manager, there are numerous challenges. But the biggest challenge in the field of machine learning is in the availability of data. It is very difficult to obtain data during a storm, while that data is crucial for our knowledge development in the field of dune erosion and the use of machine learning. In addition, I do see reluctance around me to start working with machine learning. It is a different way of looking at problems, which we as engineers are perhaps not quite used to. Still, I think it can add a lot to our current way of working!’
How can DigiShape help with this?
‘Obviously by making agreements on data sharing, but also by sharing our examples with each other. At the DigiShape session on 12 June, a lively discussion ensued with people who also have some experience with this. This is valuable input for me to determine whether I am on the right track, or whether there are other possibilities. Machine learning is not as exciting as it seems. I knew nothing about it when I started and I set it up in a few months. I hope people who read this story or watch the session back think, ‘If he can do it, I can do it too!’’
More information
Read the recap and watch Koen’s online technical session back on this topic
Have a look at Koens LinkedIn profile
