In the port of Rotterdam, poor visibility has a direct impact on activities. In limited visibility, for example due to fog, tugboats are no longer allowed to sail. The sooner this is known, the better the Port Authority can manage it. With the DigiShape seedmoney project PortVision, HKV Lijn in Water, the Port of Rotterdam Authority and Pythia Energy Intelligence are working on an AI model that can predict visibility conditions up to 12 hours ahead. We spoke to project leader Jochem Caspers of HKV about the first results and the steps towards application in practice.
“The idea for PortVision arose as a result of the DigiShape seedmoney call in 2025,” says Jochem. “To be eligible for seed money, you have to work with at least one other DigiShape partner, so we talked to the Port of Rotterdam Authority to make an inventory of what we could add for them with data and AI. We first looked broadly at water levels and other hydrodynamic variables, but it soon became clear that there was a great need for predicting visibility. This is measured in real time, but it is still difficult to predict.”
Multiple industries have an interest in predictive visibility
During the pitch in October at the DigiShape day in Delft, it turned out that more sectors are struggling with this problem. “Fog is not only an issue in the port, but also at sea, on the road and in the energy sector. That gave us the confidence that this is a relevant topic to invest in and the project won €10,000 in seed money.”
With the seed money, the project partners HKV, the Port of Rotterdam Authority and Pythia Energy Intelligence were able to immediately start building on an initial model design. “That’s the beauty of seed money,” says Jochem. “You don’t have to set up a large project first, but quickly go from idea to implementation.” HKV contributed model knowledge and experience with data analysis, the Port Authority provided the practical question and local visibility measurements, and Pythia contributed with expertise in data processing and modelling. This data has been supplemented with data from a KNMI measuring station.
From measuring to predicting
“The first step was to map out available data.” The visibility measurements of the Port of Rotterdam Authority, over a period of 20 years, formed the basis, supplemented with meteorological data such as wind, temperature, humidity and air pressure. “Based on that, we tried to translate the measurements into reliable predictions with deep learning with our models.”
The first analyses showed that vision is strongly determined locally. “We have worked with measuring points throughout the port area up to tens of kilometres apart, and then you see that the picture can differ greatly. That makes predicting visibility much more complex than, for example, water levels or wind.”
From regression model to classification model
Jochem describes how the modellers approached the project: “We started by combining sight measurements from the Port Authority and meteorological data from the KNMI. First, we tested regression models to predict an exact vision value. This proved difficult and the level of detail did not fit in well with the port authority’s practice. That is why we have switched to classification models, in which we work with a number of classes, for example whether or not to have an operational limitation.
We then tried different model types, such as ‘TemporalFusionTransformer’ and ‘TabularLogisticRegression’. This allowed us to build a first working model fairly quickly. What you see is that the transition from good to poor vision is often picked up. For that distinction, we are around 90 percent accuracy with a prediction of 3 hours ahead.”
According to Jochem, the biggest challenge lies in time and space. “Predicting one hour ahead is good, but towards six to twelve o’clock the uncertainty increases. And we are still working with point locations, while you actually want an area-wide picture. The next step is therefore to move to a grid, so that you also take spatial differences into account.”
What steps are needed for an application in practice?
The project partners now have a proof of concept that shows that it can be done. The question now is: is this good enough for practice? “We are discussing this with the Port Authority. They have responded positively to our results and we have discussed a number of opportunities for expansion and improvement. More validation is also desirable.”
There are also talks with Rijkswaterstaat. “At Rijkswaterstaat, visibility information is sometimes still retrieved by simply calling a ship and asking what the visibility is like. That works, but also shows that there is still a lot to be gained. And if the KNMI has to install fewer vision sensors with these kinds of models, a clear business case is created.”
Better insight into fog can also be of value to the energy sector. “Fog has a direct impact on solar production. If you can predict that better, you can also estimate yields better.” For Pythia Intelligence, this was an important reason to get involved. “Here you can see that you can serve multiple applications with the same data and models.”
Technically, the model still needs to be operationalized: automatic retrieval of meteorological forecasts, making visibility forecasts and access via a dashboard or API.
DigiShape Seedmoney accelerates innovation
When asked whether this research would have been carried out without seed money, Jochem is down-to-earth. “Probably, but not in this way and not at this pace,” he says. “Within HKV we have room to explore these kinds of ideas, but seedmoney helps to make it concrete and to tackle it together with partners. We started talking to the Port of Rotterdam Authority about this, because we are both DigiShape partners. For us, that is a great added value of DigiShape.”
Learn more or join
Do you work on issues related to vision, sensors or predictive models? Or do you see opportunities to take these types of applications further within your organization? Please contact Jochem Caspers via J.Caspers@hkv.nl or Sophie de Roda Husman via S.deRodaHusman@hkv.nl.
