Use cases

Living Environment

env

Use case description

Air quality and water scarcity are main environmental challenges Flanders and Europe are facing. Air pollution remains the biggest environmental health risk in Europe. Both water scarcity and water quality are increasing problems due to the ongoing climate change. Through this use case we want to enhance the environmental modelling to enable improved policy making and policy evaluation. 

Environmental models mostly address the topic of interest by processing available input data and running detailed physical/chemical models. Policy support and decision making are supported through these model outcomes and simulations of possible future scenarios. For each step in such an environmental modelling flow scheme, AI techniques can be integrated to further enhance the modelling capabilities, improving accuracy and resolution and unlocking new applications through reduced calculation times. The current state-of-the-art already applies some AI techniques. A thematically broad use case enables cross-domain learning and identification of successful and novel AI approaches between the different environmental domains. With this use case we have the ambition to integrate AI techniques in different aspects of how environmental models are built up and target the 4 data science quadrants of data exploration, data engineering, model building and exploitation. 

The structure and elements of urban areas, such as buildings, vegetation, open areas, and transportation systems, play a crucial role in determining the precision of environmnetal prediction models. Environmental monitoring devices can provide accurate measurements at specific urban locations over time, yet achieving detailed spatial analysis demands further insight into the urban landscape's attributes and their temporal variations. To further enable uptake of remote sensing data into these environmental models processing of large datasets for a variety of uses is needed.

In this use case we aim to create innovative self-supervised techniques for dense representation learning, with an emphasis on object-centric learning going beyond the current patch-based techniques. These techniques will be utilized on high-resolution (15-25 cm pixel resolution) RGB and/or multispectral aerial images across Flanders. Satellite data and earth observation services are offered on the open and free Terrascope platform. The developed representation learning model will be integrated into a remote sensing workflow. This work builts on-going work to use AI and maching learning to streamline object and change detection from aerial imagery.

Data engineering is focused in this use case on AI assisted pre-processing. This will be integrated for time series modelling of water quality to deal with failing sensors and significant data gaps. Machine learning is needed to make the step from sparse water sensors to frequent and continuous water quality predictions in river systems, and to create high resolution maps of soil and groundwater to better understand spatiotemporal dynamics. 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.

Model building focusses here on new AI algorithms for environmental models and optimization of the involved parameters. We aim to speed up model calculations, apply AI techniques to extrapolate results without the need of explicitly running the models and exploit the spatial structures of data to do better informed predictions. To achieve these goals, we’ll explore the use of kernel methods on high dimension datasets and graph neural networks as a way to include spatial dependency. 

The results envisaged in this use case are of direct use and value to the air quality agencies (IRCEL – VMM), drinking water companies and remote sensing community (Digitaal Vlaanderen). This use case can be linked to other initiatives along the same lines, such as Digital Flanders’ EODAS program or the ESA programs AI4EO and Destination Earth. The use case aims to further valorize data in digital twins of Flanders, using the MAPEO platform or other local initiatives.  

Results, Highlights & Research Outputs

  • Decision-focused learning for environmental applications: Decision Focused Learning involves both predictive and prescriptive analytics. In the predictive stage, the uncertain parameters are predicted from the feature variables using an ML model. In the prescriptive stage, a decision is prescribed by solving a CO problem using the predicted parameters (from Mandy, Kotary, et al.). An application of this technique to the water and air quality domains is being explored.

  • Self-supervised representation learning (SSL): Dense and object-centric SSL approaches are being tested to produce strong backbones for a wide variety of tasks in the context of remote sensing. Using these approaches we aim to derive valuable input data to environmental models.

  • Surrogate Modeling with Deep Learning (ATMO-NET): Physics-based models that simulate detailed physical and chemical processes for the dispersion and transformation of pollution provide high accuracy, but are often too computationally expensive for large-scale or time-sensitive analyses. We aim to develop surrogate models—particularly neural networks—as efficient alternatives to these simulations, enabling faster experimentation and higher-resolution predictions while remaining consistent with their physics-based counterparts. In parallel, we will build dedicated tooling to support surrogate modeling, ensuring reproducibility, transparent data lineage, and straightforward deployment. The current state-of-the-art for fast wind and air quality screening is being refined and will be challenged by novel, promising techniques applicable in air quality products. 

  • Benchmarking of AI techniques for air quality forecasting: Machine learning techniques are already for more than a decade operationally in use in Flanders to forecasts air quality several days ahead. A benchmarking study of recent timeseries forecasting techniques against the current operationally used GRU neural network models is initiated, using the OPAQ toolbox.

  • Exploiting knowledge in time series: Predict label from multivariate time series, exploiting symbolic knowledge applied to the water domain. Aim to predict overflow from sensor time series while exploiting symbolic knowledge and automatic segmentation of overflow patterns

  • Reasoning about learned models: Research on expressive power of Graph Neural Network GNN architectures, Further expand connection between GNNs and database query languages (QL), exploring applications to the water quality domain.

Outputs & Resources

Publications

  • Đukić, N; Lebailly, T; Tuytelaars, T (2025) Object-Centric Pretraining via Target Encoder Bootstrapping, 13th International Conference on Learning Representations (ICLR 2025), Singapore, 24-28 April 2025

  • Vitali, L; Cuvelier, K; Piersanti, A; Monteiro, A; Adani, M; Amorati, R; Bartocha, A; D’Ausilio, A; Durka, P; Gama, C; Giovannini, G; Janssen, S; Przybyła, T; Stortini, M; Vranckx, S; Thunis, P (2023) A standardized methodology for the validation of air quality forecast applications (F-MQO): lessons learnt from its application across Europe. Geoscientific Model Development, 16, 6029–6047.

  • Agarwal, S; Sharma, S; Suresh, R; Rahman, MDH; Vranckx, S; Maiheu, B; Blyth, L; Janssen, S; Gargava, P; Shukla, VK; Batra, S (2020) Air quality forecasting using artificial neural networks with real time dynamic error correction in highly polluted regions. Science of The Total Environment, 735, 139454.

Software & Tools

  • Terrascope platform.

  • ATMOSYS operational air-quality forecasting models.

  • OPAQ toolbox for benchmarking air-quality forecasting.

  • MAPEO platform for digital-twin valorisation.

Models / Methods

  • Decision-focused learning for environmental applications.

  • Self-supervised representation learning for remote sensing.

  • Surrogate modelling with deep learning for air-quality applications.

  • Benchmarking of AI techniques for air-quality forecasting.

  • Knowledge-based time-series modelling for water-domain applications.

  • Reasoning about learned models and graph neural network architectures.

Team & Contact

Contact

Contact Person: Stijn Vranckx (VITO - Environmental Intelligence)

PhD Researchers

Nikola Djukic (KULeuven-PSI, PhD), Alessandro Barbini (UGent-KERMIT, PhD), Klara Pakhomenko (UHasselt, PhD), Elina Syrri (KULeuven-DTAI, PhD)

Rseearch groups

VITO; KU Leuven - PSI; UGent - KERMIT; KU Leuven - STADIUS