A Bayesian network approach for multi-sectoral flood damage assessment and multi-scenario analysis

tesi vincitrice edizione 2022

Università di Venezia

Master’s Degree Programme in Environmental Sciences

Tesi di

Remi Harris

  • Biografia

    Remi Harris is a Doctoral student in Science and Management of Climate Change at Ca’ Foscari University of Venice, collaborating with the CMCC on the Horizon 2020 project MYRIAD-EU, where he is currently developing an approach for the assessment of environmental quality impacts under multiple hazards and global climate change. Originally from Jersey, he holds a BSc in Chemistry from the University of Bath (UK) where he specialized in sustainable chemistry during a study period abroad at Purdue University (USA). Later, he graduated from Ca’ Foscari University of Venice with an MSc in Environmental Sciences (Global Environmental Change). During and since his studies he worked as a junior researcher with the CMCC, and there he completed his thesis in the frame of the LODE project, designing a Bayesian Network approach for flooding damage prediction, with a particular focus on the case study of the Secchia river basin, Italy.

ABSTRACT

Extreme weather events, from river flooding to droughts and tropical cyclones, are likely to become both more severe and more common in the coming years, resulting from drivers such as climate change and mismanagement of natural resources. The damages caused by these extreme events will be felt across all sectors of society, particularly in the form of economic losses to productivity and physical assets. In the face of this threat, policy- and decision-makers are increasingly calling for new approaches and tools to support risk management and climate adaptation pathways that are able to capture the full extent of multi-sectoral damages. Beginning with a review of the state-of-the art literature concerning Machine Learning (ML) methods, this Thesis builds a GIS-based Bayesian Network (BN) approach that is capable of capturing and modelling multi-sectoral flooding damages against future ‘what-if’ scenarios (e.g. changes in land use/cover and flooding hazard patterns). In doing so the work addresses key knowledge gaps in the literature, while providing support and additional insight for decision making underpinning the definition of strategies and policies for Disaster Risk Reduction and Management. To enable this approach, a risk-based conceptual framework was developed in line with the current IPCC definitions, that highlights pathways of interaction between hazard, exposure, and vulnerability indicators related to flooding events, and the subsequent damages to the agricultural, industrial, and residential sectors. Building on this framework, the BN model was then trained and validated by exploiting data collected from the 2014 Secchia River flooding event, as well as other variables selected for the case study area. Moreover, in this stage, a novel approach to defining the structure of the BN was performed, reconfiguring the model according to expert judgment and data-based validation. During the validation process, the finalized BN model showed a good predictive capacity for flooding damages in the agricultural, industrial and residential sectors, predicting the severity of damages with a classification accuracy of about 60% for each of these assessment endpoints. Model accuracy would be improved with additional and more detailed observational data, as indicated in the validation stage showing as the range of uncertainty in the model prediction for each sector increases with decreasing quantities of data. However the model still performs successfully despite the constraints within the available dataset. This gives a greater picture of multi-sectoral flooding damages than that found in the current BN related literature, which focuses almost exclusively on the residential sector. In support of this, a sensitivity analysis was completed, examining the relative importance of the explanatory BN variables (e.g. area of reported damages, land use, and flooding hazard) towards the multi-sectoral damages, and highlighting those that should therefore be prioritized in a more detailed data collection, allowing for better capturing and modelling of multi-sectoral damages against future flooding events. Finally, based on the finalized BN model, ‘what-if’ scenario analysis was performed to understand the potential impacts of future changes in i) land use patterns, as well as ii) increasing flood depths resulting from more severe flood events. The output of the model showed a rising probability of experiencing large monetary damages under both scenarios, representing the increased damages likely to be faced in the future as extreme events become more common and as under the projected land development in the case study area. In summary, this Thesis presents a full narrative for the analysis of multi-sectoral flooding damages through a BN approach, not only offering a predictive damage model, but also describing its construction, as well the potential applications of the methodology for risk assessment purposes. In spite of constraints within the case study dataset, the results of the appraisal show good promise, and together with the designed BN model itself represent a valuable support for Disaster Risk Management and Reduction against extreme river flooding events, enabling more effective and reliable decision making.