SMILE

Despite the efforts devoted to DRR at global, regional and national scale, the overall losses caused by natural hazards have been increasing in the last decades. This can be mostly attributed to the modification in number and distribution of exposed assets (e.g. rapid urban development on land prone to geo-hazards). The knowledge of the location and characteristics of exposed assets (namely, the exposure) supports the assessment of expected losses and the prioritization of actions during and after emergencies. Damages to buildings, in particular, cause a relevant fraction of both economic and life losses, while strategic buildings (e.g. schools, hospitals) play a crucial role during the emergency response and recovery phases. Buildings exposure is therefore paramount to reduce disaster risk and define mitigation and response measures. Despite the large efforts devoted to exposure development, current datasets have one substantial limitation: they rely on data collected sporadically at discrete time steps (e.g. decennial building census). Current DRR strategies suggest developing exposure dynamically to grasp rapid changes (e.g. modification of buildings, urban development). Recent scientific developments point out two potential data sources that can support dynamic exposure development: remote sensing and crowd-sourced data. In the last decade, remote sensing products increased both their temporal and spatial resolution. As for crowd-sourced approaches, the engagement of citizens is increasingly envisaged by DRR strategies (e.g. SENDAI Framework) and has proven effective in retrieving up-to-date data. However, both sources should be reliably integrated with current exposure data.

The SMILE (Statistical Machine Learning for Exposure development) project aims at exploring the potential of Machine Learning (ML) to dynamically develop up-to-date exposure layers combining remote sensing images, ancillary data (e.g. national census) and crowd-sourced data collected by trained citizens. The project will assess how ML techniques can enhance existing exposure development methods, focusing on selected test areas. Data visualization tools, methods and algorithms will be developed to produce multiple representations of the dynamic exposure, in which each point can be associated with nominal, ordinal or quantitative data. Visual outputs will be defined together with stakeholders and knowledge users who will provide feedback on their usefulness and comprehensibility. Results will be presented at different levels of comprehensiveness and complexity to identified knowledge users.

To do that, SMILE relies on a multidisciplinary team that comprises scientists with a background in natural hazards and exposure development (OGS), remote sensing (UNIFI), ML (UNIMIB), statistics and data visualization (CNR).