Earth Observation (EO) plays a pivotal role in understanding and safeguarding historical and cultural sites, offering non-invasive methods for monitoring, risk assessment, landscape analysis, and the detection of illegal activities. While EO already provides time- and cost-efficient ways to study large areas, often including remote or hazardous locations, our research pushes these capabilities further. By harnessing vast EO datasets, we develop innovative Artificial Intelligence solutions that accelerate analysis, extend territorial coverage, and enhance computational capacity. This enables us to manage complex data with higher spatiotemporal resolution and longer temporal ranges, opening new possibilities for cultural heritage preservation.

• Looting Monitoring
  Monitoring illicit looting is essential in the ongoing effort to safeguard global cultural heritage. The clandestine excavation and trafficking of artefacts deprive communities not only of invaluable objects but also of precious insights into the past, while causing irreversible damage to landscapes of profound cultural significance. Remote sensing technologies are a powerful means of surveying and safeguarding archaeological sites worldwide. These technologies underpin initiatives such as ALCEO (Automatic Looting Classification from Earth Observation) and OPTIMAL (Optimal Transport for Identifying Marauder Activities on LiDAR), which have developed deep learning- and optimal transport-based change detection models, respectively.
• Image Fusion
  Our research focuses on developing image fusion and pansharpening techniques to enhance the spatial resolution of multispectral and hyperspectral satellite imagery. The aim is to achieve higher spatial resolution for detecting small-scale archaeological features, while simultaneously preserving spectral range and resolution, which are essential for highlighting specific archaeological proxy indicators. In our projects, we have applied these approaches to identify subsoil archaeological and palaeo-environmental features across various contexts in the Mediterranean and MENA regions, leveraging partnerships with the Italian Space Agency (ASI) and access to PRISMA satellite imagery.
• Archeaological Detection
  Our research develops advanced Artificial Intelligence solutions to detect subsurface archaeological features using EO data. By analysing large multimodal and multitemporal datasets, we support large-scale archaeological surveys and the study of cultural landscapes. Our methods identify patterns such as crop marks and soil marks, combining data from both passive sensors, like multispectral and hyperspectral imagery, and active sensors, such as LiDAR and SAR. We also collaborate with the European Space Agency (ESA), the Italian Space Agency (ASI), and the European Union Satellite Centre (SatCen) to enhance the scope and impact of our projects.

Our research is advancing novel learning algorithms to automatically identify and map the spectral features of materials within hyperspectral datacubes. These methods address the challenge of high dimensionality inherent to hyperspectral data, enabling more effective analysis and interpretation. Hyperspectral imaging, which collects hundreds of contiguous spectral bands and provides detailed spectral information for every pixel, offers great potential for characterising the composition and distribution of materials. By enhancing the analysis of this data through our algorithms, we are unlocking new possibilities for the study and preservation of cultural heritage.

Our research aims to advance Artificial Intelligence solutions for the automatic transcription of historical documents, particularly those from the Middle Ages and Early Modern periods written in the Latin alphabet. These invaluable documents present challenges due to the frequent use of abbreviations, where context plays a crucial role. Incorporating contextual information into the models is therefore essential, as it enhances object classification in images. We also investigate explainability, seeking optimal configurations for neural network encoders based on input data, with the goal of minimising their size while maintaining optimal performance.