2026.02.10 44ma Conferenza Nazionale GNGTS (Udine)
Nel 2026 il Congresso GNGTS si è svolto a Udine, in occasione del 50° anniversario del terremoto del Friuli, sottolineando l’importanza della ricerca e della collaborazione nella riduzione del rischio sismico. Il Congresso, organizzato da OGS dal 2018, è coordinato da un Comitato Tecnico Scientifico composto da sei enti principali, con il rappresentante dell’OGS nel ruolo di Presidente.
I seguenti abstract sono stati presentati alla conferenza:
1) Title: A building footprint segmentation dataset from the Friuli Venezia Giulia region
Authors: Claudio Rota, Flavio Piccoli, Rajesh Kumar, Gianluigi Ciocca, Chiara Scaini and the SMILE team
Accurate and up-to-date information on the spatial distribution and characteristics of the built environment is a fundamental component of exposure assessment, which in is turn is paramount for disaster risk reduction (DRR). Building footprints and associated geometric information are useful for land use planning and risk assessment and mitigation strategies. Building footprint segmentation (i.e. detect building footprint area) from aerial and satellite imagery is a thus a key component of numerous geospatial applications. Despite the rapid progress enabled by deep learning, the performance and generalization capability of modern segmentation models remain limited by the availability of large, diverse, and high-resolution annotated datasets. Existing resources typically lack one or more of these characteristics, constraining their utility for training robust models across heterogeneous geographical contexts. In this work, we introduce Segmentation Friuli Venezia Giulia (SegFVG), a large-scale dataset containing 15,403 aerial tiles at 0.1 m ground sampling distance, each with precise pixel-level building footprint annotations. Covering 616 km² of the Friuli Venezia Giulia region in northeastern Italy, SegFVG encompasses a wide variety of landscapes, including alpine rural zones, agricultural plains, suburban environments, and densely populated coastal areas, resulting in approximately 357,000 annotated buildings. This geographic and environmental diversity makes SegFVG a representative and challenging benchmark for evaluating building footprint segmentation models. Figure 1 shows the spatial distribution of the images within the region. We use SegFVG to build ML algorithms for automatic building footprint segmentation. We tested multiple deep learning architectures and achieve high performance, with 0.943 precision, 0.975 recall, 0.959 F1-score, and 0.921 IoU, demonstrating the dataset's suitability for training accurate building segmentation models. Beyond enabling benchmarking and model development, SegFVG can be useful for a wide range of practical applications, including (1) automated building detection, (2) comparison with official cadastral data to identify discrepancies, (3) detection of unauthorized constructions, and (4) multi-temporal analysis for monitoring changes in the built environment.
Acknowledgments: This research contributes to the PRIN 2022 PNRR project SMILE "Statistical Machine Learning for Exposure development" (code P202247PK9, CUP B53D23029430001) within the European Union-NextGenerationEU program. We are grateful to the students and teachers at several high schools in the Friuli Venezia Giulia region and to the Università della Terza Età of Trieste for their valuable contribution to data collection. Finally, we warmly acknowledge our colleagues from the SMILE team: Matteo Del Soldato, Silvia Bianchini, and Olga Nardini (Univ. Florence); Antonella Peresan, Piero Brondi, Hazem Badreldin, and Hany Mohammed Hassan Elsayed (OGS), Elisa Varini and Maria Teresa Artese (CNR-IMATI Milano).
2) Title: A web platform for crowdsourced building data collection and expert-based validation
Authors: Maria Teresa Artese, Elisa Varini, Chiara Scaini and the SMILE team
The SMILE (Statistical MachIne Learning for Exposure development) research project explores the use of machine learning to generate updated building exposure layers by integrating remote sensing imagery, census data, and validated crowdsourced information. Crowdsourced data are collected through targeted initiatives involving trained students and citizens and subsequently verified by experts to ensure reliability. To support these activities, we developed a web-based multimedia platform capable of guiding users through data collection, managing the data workflow, and storing georeferenced information and images in a structured database. The data collection webform encompasses multiple building characteristics that are deemed relevant for multi-hazard exposure assessment. The database currently holds 4,100 surveys; Fig. 1 shows the distribution of a subsample of surveyed buildings with respect of census areas building density, in the town of Monfalcone (UD), located in a coastal area prone to multiple hazards including floods and earthquakes. The platform features statistical and GIS-based visualization tools that enable different user groups - policymakers, planners, and citizens - to explore the collected exposure information, assess the quality of collected data. The collected data, once validated, are going to be compared with official datasets (e.g., the ISTAT building census). The collected building data and images have also been used to train machine learning models for the automatic detection of building attributes, such as roof type and number of floors. These models are integrated into the platform to facilitate and streamline future data collection. Another key objective of the project is to evaluate the reliability of the collected data and determine their potential to update and enrich available building exposure datasets. To this end, the platform offers specific functionalities for the validation of collected data. Experts from the SMILE team reviewed approximately 400 surveys on buildings in Udine, completed by high school students using the SMILE web platform. The expert-validated building sample allowed comparison with student-collected data, enabling identification of potential issues in the web-survey and statistical assessment of the reliability and quality of the collected data. Among the broader impacts of this pilot project is also the potential to engage local communities while enhancing existing exposure layers to support risk mitigation and preparedness strategies.
Acknowledgments: This research contributes to the PRIN 2022 PNRR project SMILE "Statistical Machine Learning for Exposure development" (code P202247PK9, CUP B53D23029430001) within the European Union-NextGenerationEU program. We are grateful to the students and teachers at several high schools in the Friuli Venezia Giulia region and to the Università della Terza Età of Trieste for their valuable contribution to data collection. Finally, we warmly acknowledge our colleagues from the SMILE team: Gianluigi Ciocca, Flavio Piccoli, Claudio Rota, and Rajesh Kumar (Univ. Milano-Bicocca); Matteo Del Soldato, Silvia Bianchini, and Olga Nardini (Univ. Florence); Antonella Peresan, Piero Brondi, Hazem Badreldin, and Hany Mohammed Hassan Elsayed (OGS).