Multi-variate flood loss estimation model for residential sector
Date of Issue2016
School of Civil and Environmental Engineering
Depth-damage curves or loss functions which are used to assess the direct loss from flooding are highly inadequate for accurate and reliable analysis. These models have a simplified structure which only takes into account the inundation depth and building characteristics while defining flood losses. Although a plethora of factors which influence the residential flood loss have been identified in the literature, relatively few models and studies have been attempted in assessing their interaction in predicting flood loss. This dissertation demonstrates that notable enhancements could be made in the approaches used for developing flood loss/damage functions. To supplement the knowledge of the multi-variate flood loss estimation this thesis attempts to address the issues pertained to (i) identifying the key factors which influence flood loss and (ii) development of comprehensive loss estimation model in a data deficient environment. To develop a residential flood loss estimation model for Jakarta, the regions of Pluit and Kelapa Gading in North Jakarta and Cengkareng in West Jakarta were selected to conduct structured questionnaire survey. The household surveys determined the characteristics of flood and the loss influencing factors in the study area. Due to the lack of information on factors for segregating the residential sector in the study area for flood loss estimation, a standardized procedure for categorising the flood loss cases objectively into homogenous groups is presented. The first part of the approach, which is a combination of Principal component analysis (PCA) supplemented with correlation and Mutual information (MI), showed that monthly household income and people’s response measures against flood along with flood water depth essentially determine the extent of flood loss in Jakarta. Using monthly income as similarity measures in Hierarchical cluster analysis (HCA), the second part yielded 3 income groups with homogenous flood loss data. Depth-damage curves were developed for these groups and scaling factors were derived from the surveyed data to represent the effect of people’s response measures against floods. To identify the main factors which influence flood damage and loss, a novel hybrid heterogeneous data feature selection algorithm based on kernel methods is proposed herein. This algorithm captures the importance and relevance of the factors using Multiple kernel learning (MKL) and mutual information (MI) and capture the interaction among them using Multiple kernel support vector regression (MKr) in order to identify those factors whose single and joint interaction determine the flood loss . The outcome of the proposed approach when implemented in Jakarta showed that apart from water depth other variables belonging to the domain of flood response actions, socio-economic characteristics and flood warning play a significant role during flooding. Lastly this data driven multi-variate flood loss estimation model (MKr model) has been found to perform significantly better than the loss estimation models developed using existing methodologies. Finally a new data driven and statistical framework for modelling residential flood loss in a data deficient environment was developed and presented in the dissertation. Key component of the framework is the probabilistic generation of synthetic data for underrepresented variables using an enhanced Gibbs sampler. The Enhanced MKr model developed by employing the proposed framework in Jakarta displayed comparable performance with the MKr model. However the factor selection stability of the proposed algorithm increased substantially under the new framework. Moreover, the following precautionary and emergency measures such as temporary water barriers, waterproofing of external walls, raised floors and using low value & easily movable objects in flood prone floor were identified by the En-MKr model which has the potential in reducing the damage from floods in Jakarta. The whole framework is beneficial in developing a comprehensive multi-variate flood loss estimation model especially with small datasets.
DRNTU::Engineering::Civil engineering::Water resources