Estimating environmental phenomena with informed deep learning over Earth Observation data (SPHERE)

Period: 01. 03. 2026 – 28. 02. 2029

PROJECT LEADER: assist. prof. dr. Marko Bizjak

Abstract:

Environmental modeling with Earth Observation (EO) data is essential for understanding, predicting, and managing environmental phenomena on a large scale. EO data provides invaluable insights into diverse environmental systems, capturing information such as land use, vegetation dynamics, and urban development. By analyzing this data, researchers and decision-makers can identify critical trends, assess potential risks, and develop effective strategies to mitigate and adapt to environmental challenges. As global environmental challenges intensify, leveraging EO data becomes increasingly indispensable. The growing availability of high-resolution, multi-modal EO datasets offers significant opportunities, yet current methodologies fail to fully exploit these data. Traditional approaches often treat spatial, temporal, and relational data independently, limiting their ability to capture interconnected dynamics. While Deep Learning (DL) has shown promise, existing methods struggle with overfitting, lack of interpretability, and inefficiencies in processing large-scale datasets due to data dependency and training limitations. These limitations highlight the need for innovative frameworks that fully exploit the potential of EO data.

The proposed SPHERE (Estimating environmental phenomena with informed deep learning over Earth Observation data) project addresses these challenges by introducing an innovative framework for informed DL algorithms that bridges the gap between traditional physics-driven simulations and modern data-driven DL methods. By integrating domain-specific knowledge into DL workflows and combining high-resolution spatial data with graph-based relational modeling, SPHERE aims to establish a robust foundation for scalable, interpretable, and accurate environmental models. This paradigm shift allows the seamless incorporation of physical laws, boundary conditions, spatiotemporal dynamics, and relational insights into modeling workflows, advancing scientific understanding and setting a new standard in environmental modeling.

SPHERE’s potential will be demonstrated through three diverse pilot applications addressing critical environmental challenges. The first pilot targets urban heat islands (UHIs), modeling their spatial and temporal dynamics to inform strategies that mitigate public health risks and reduce energy demand. The second pilot examines forest growth, offering insights into vegetation’s role in climate change mitigation. The third pilot explores glacier and snowfield dynamics, providing crucial data for water resource management and monitoring the consequences of climate change.

To support these applications, SPHERE will introduce advanced geospatial data integration techniques to unify diverse EO datasets. Its innovative data structuring seamlessly integrates these datasets, capturing detailed spatial features and complex interactions across scales. Unlike traditional DL approaches, SPHERE aims to enhance interpretability by incorporating domain-specific knowledge while also improving the accuracy of environmental phenomena assessment, even in the presence of data gaps or heterogeneous input data. By leveraging high-performance computing (HPC) and General-Purpose GPU (GPGPU) paradigms, SPHERE develops tailored parallelization strategies for informed deep learning algorithms, accelerating the processing of large datasets at higher resolutions.

By addressing critical limitations of existing approaches, SPHERE will establish a solid foundation for estimating environmental phenomena on a large scale. It will equip researchers and decision-makers with essential tools to address the challenges posed by climate change and urbanization, enabling more effective strategies for mitigating environmental impacts and adapting to evolving urban conditions.