The battle for the square meter

When one land use needs more space, it often comes at the expense of another, with social and economic consequences. One of our tasks at PBL is to map those consequences.

That work gets harder because fields of knowledge and transition challenges are more and more interconnected, and major events have effects that are hard to predict. As a result, analyses need more and more data, while the time pressure to publish keeps rising. We need new processing methods to open up data, get results faster, and run analyses that were not possible until now. We expect AI techniques can help with this.

Scenarios for an unknown future

To explore possible spatial developments, PBL uses scenario analysis, among other methods. In spatial scenarios, we assume that driving variables, such as social trends and features of the environment, help determine how an area is laid out. For example, peat soil can make certain types of farming more likely than new housing. We calculate the effects of these scenarios with a spatial model that describes these relationships with mathematical algorithms.

That research process is exactly where we are testing what AI can add. In 2024 we started a pilot to design prototypes for parts of a new spatial model that we plan to start using in 2027. The prototypes focus on automating manual work: searching for suitable data and converting it into spatial information. We are also developing more robust algorithms.

Data management matters in any research, but using AI makes it even more complex. That is why data management becomes an important topic in 2027. We are already preparing for it by making our analyses as objective and verifiable as possible.

From external data source to spatial information

One way we do this is by linking the spatial model to the Kadaster Knowledge Graph (Kadaster is the Dutch land registry). A knowledge graph is a structured database that also stores the relationships between topics. Because of this, those relationships stay intact and predictable at any data volume, and we can query the core registers of cadastral datasets automatically.

We also retrieve official notices through the API of the Omgevingsloket, the Dutch government's online portal for environmental and planning rules, permits and notifications. These notices describe, among other things, how an area may be used, so we can include them as restrictions in spatial scenarios. The Omgevingsloket holds hundreds of thousands of notices. Because of that volume, we could only use them selectively in the past.

With an AI language model, we can now select relevant notices and analyze whether rules from higher levels of government, such as provinces, add to or contradict rules from lower levels, such as municipalities. As a next step, we want to work with Utrecht University to convert the retrieved text into spatial information. Sometimes that conversion is fairly simple because the spatial geometry is already present. In other cases we want to approximate it with calculation rules. Suppose a local bylaw says that new construction is restricted within a 500 meter buffer around wind turbines. A geographic information system can turn that written rule into buffer zones that the spatial model blocks for new construction. This is explicitly an interpretation for a scenario study, not a legal elaboration of the rule. In the new system, both finding the rule and converting it into geographic information can happen automatically.

Automation does not mean we can take the results at face value. One key point is that we cannot guarantee that all relevant notices were retrieved, because they have no fixed format or structure. That is why we ask the language model several queries and use only the answers that are consistent. We also take spot checks to confirm the interpretation is correct. The language model is not allowed to invent missing information, or “hallucinate.”

Better algorithms for land-use change

Besides opening up sources, over the past year we worked on improving the mathematical algorithms in the spatial model. Before, we used a limited set of studies on the relationship between driving variables and spatial layout. The calculation rules derived from those studies proved too limited for the complexity of new research questions.

So we developed new algorithms with Random Forest, a type of machine learning. For this we used a large set of historical driving variables from 2000 to 2010. We then applied the trained prototype to a fully independent set of driving variables from 2010 onward. The quality of the predicted land-use changes was promising. The results appear in a scientific paper later this year.

Faster, more targeted research

We expect AI in our workflow to let us run more complex “what if” scenarios, and to run them faster. Analysis quality can also improve, because we can make more steps objective and get more ways to run checks. At the same time, this is still a work in progress. Before we can use these applications as standard practice, we need to analyze the prototype results carefully.