Flanders AI Research Programme

VITO is tackling major environmental challenges—air pollution and water scarcity—which threaten health and sustainability across Flanders and Europe. The central strategy is to enhance environmental models using advanced AI and machine learning.

These models process data to simulate complex physical and chemical processes, informing policy-making, planning, and evaluation. The ambition is to integrate AI at every stage of the modelling pipeline: data exploration, data engineering, model building and exploitation. This approach is cross-domain, aiming to transfer AI techniques across air quality, water, and remote sensing. 

Urban areas present particular complexity for environmental modelling because the configuration of buildings, roads, vegetation, and open spaces strongly affects local air quality and environmental dynamics. To improve predictive accuracy in these settings, VITO is working with high-resolution aerial imagery at 15–25 cm resolution in both RGB and multispectral bands, covering the entire Flanders region. The research team is developing innovative self-supervised learning techniques that produce dense, object-centric representations of this data. These methods move beyond traditional patch-based image analysis, capturing richer structural information about the urban landscape. Satellite data from the open Terrascope platform are integrated into this process to enhance the spatial detail and detect changes over time. This combination of remote sensing and machine learning is designed to generate more precise input data for environmental models, supporting applications from air quality forecasting to land use planning.

In the domain of water quality modelling, AI plays a critical role in overcoming gaps in monitoring data. Environmental sensor networks frequently suffer from failures or periods of missing data. VITO is developing AI-assisted preprocessing pipelines that can automatically detect and fill these gaps. Machine learning models are also being created to predict water quality time series in river systems, transforming sparse measurements into continuous, high-frequency estimates. Additionally, high-resolution mapping of soil and groundwater conditions is under way to better understand their spatiotemporal dynamics and how they interact with surface water quality. Similarly, air quality forecasting can start from a limited number of observations of the Belgian monitoring network to forecasts upcoming air quality in the next days. We aim to further improve the current ATMOSYS machine learning models in operational use.

Air quality forecasting is a particular focus area for the modelling teams. The current focus lies on the following: 

• ATMOSYS:

  • An operational air quality forecasting system that already uses machine learning (especially gated recurrent unit (GRU) neural networks) for multi-day predictions.

  • Benchmarking study: Comparing newer time series forecasting techniques with the existing GRU-based approach, using the OPAQ toolbox.

• Quick Urban Air Quality using Kernels (QUARK):

  • Kernel methods are explored to speed up urban air quality modelling, bypassing costly simulation of physical/chemical processes.

  • New techniques are being tested against the current fast-screening models.

  • This work supports agencies like IRCEL – VMM and contributes to rapid assessments of air quality.

• AI Surrogate Models: Developing surrogate machine learning models trained on the outputs of white-box models, such as ATMO-Flow simulations.

  • These surrogates enable accurate predictions while reducing computational cost.

  • They support faster scenario analysis and operational deployment.

• Integration of Remote Sensing and Spatially Aware Modelling: Methods are being developed to integrate remote sensing data and spatial modelling techniques to improve assessments in complex urban environments.

• Other advanced techniques under investigation: graph neural networks, which explicitly model spatial dependencies among observations, kernel methods applied to high-dimensional datasets to improve prediction accuracy, and decision-focused learning, an approach that combines predictive modelling with prescriptive optimisation, such as pollution control measures.

These efforts are carried out in close collaboration with other institutions, including KU Leuven (PSI and STADIUS) and UGent-KERMIT. Several PhD and postdoctoral researchers—Nikola Djukic, Alessandro Barbini, and Sonny Achten—are driving research forward. The work is closely linked to broader initiatives, such as Digital Flanders’ Earth Observation Data Analytics Services (EODAS), ESA’s AI4EO programme, and Destination Earth. The ultimate aim is to unlock the value of environmental data for digital twins of Flanders and to integrate these capabilities into platforms like MAPEO and other local infrastructures.