Leveraging Earth Observation and Machine Learning to Enhance Understanding of Impacts in the Mediterranean Basin

Climate change is intensifying the frequency, magnitude, and spatial extent of climate-related hazards. Hotspot regions such as the Mediterranean basin have experienced severe impacts in recent decades, with an alarming increase in the occurrence and intensity of compound and co-occurring hazards. Despite the recognized need for multi-risk approaches, the availability of comprehensive, harmonized, and representative data remains limited, constraining the understanding of contributing risk factors. In particular, challenges in impact assessment represent a key bottleneck for quantitative multi-risk modelling and for disentangling interactions among risk drivers.

Earth Observation (EO) offers a largely underexploited opportunity in the context of multi-risk assessment, capable of providing spatially explicit, temporally consistent, and relevant, globally comparable indicators. The integration of these capabilities into coherent multi-risk assessment frameworks is an active area of research; however, significant opportunities for improvement remain in exploiting the full spatio-temporal richness of EO data through innovative methods, including artificial intelligence.

Modern representation learning techniques, such as embeddings for large spatio-temporal datasets, project system states into high-dimensional latent spaces. This enables exploitation of the full information content available from EO data and supports analysis of the entire system in which hazards occur, rather than relying on targeted regressors that may fail to capture the complexity and completeness of the underlying processes.

Here we explore and propose an EO-driven framework for the assessment of multi-risk from climate-related hazards (such as compound hot-dry extremes, wildfires, and water scarcity), with an application to the Mediterranean basin. By leveraging large amounts of remotely sensed data that describe the dynamics of the case study in both the temporal and spatial domains, this research aims to incorporate into the analysis the EO-based long-term system trajectories of the area, rather than relying solely on closely related preconditions. To this end, the study will explore the possibilities of coupling representation-learning techniques with machine learning methods to model impacts from multiple hazards in selected Mediterranean basin case studies. 

The proposed approach is designed to be flexible and transferable across diverse riskscapes, including data-scarce regions, by complementing commonly used datasets and reducing reliance on incomplete impact records. By combining EO-based system representations with data-driven modelling frameworks, the research seeks to enhance the predictability of multi-risk consequences in the Mediterranean hotspot. 

By considering a wider focus instead of hazard-specific situations, the ongoing work aims to contribute to the development of methodologies for climate risk analysis that may help better represent complex risk dynamics that are currently difficult to capture within traditional risk assessment approaches, with potential implications for adaptation planning, early warning, and disaster risk reduction under current and future climate conditions.