3.1 Damage observations
Damage samples were collected for 247 flood damaged residential buildings in Westport (Fig. 1). Water depths above ground level measured between 0.1 m to 1.4 m, and depths above floor level from − 0.1 m to 0.8 m (Fig. 2a). Nearly 90% of buildings were exposed to water depths above floor level < 0.5 m. Higher water depths above floor levels were observed for concrete slab (µ 0.27 m, σ 0.13 m) foundations compared to piles (µ 0.25 m; σ 0.16 m) and solid walls (µ 0.22 m; σ 0.15 m). Compared to other foundation construction types, lower floor heights for concrete slab foundations (Fig. 2b) caused higher water depths above floor level. Damage presence from other flood hazards was not (i.e., flow velocity (n = 0)) or rarely (i.e., debris (n = 2), contamination (n = 2)) observed.
Buildings represent a 120year construction period with most comprising a timber structural frame (96%) and one storey (93%). Over 60% were constructed on pile, solid wall or mixed foundations, with the remaining on concrete slab. Buildings constructed prior to 1960 were often constructed with native hardwood timber floor finishes and often incorporated modern building materials and services. Floor finishes in buildings constructed after 1960 with suspended flooring systems, regularly observed composite timber materials e.g., low or medium density fibreboard [25]. Concrete slab foundations were more frequently constructed after 1980 and the most common (75%) foundation after 2000. Building areas after 1980 were larger on average than earlier construction periods. Building area increased by 101 m2 between the 1900–1920 and 2000–2020 construction periods. Multistorey and nontimber structural frame buildings only accounted for 8% of observations.
Relative building (DRb) and component (CDRi) damage is presented in Fig. 3a and subcomponents (CDRij) in Fig. 3b to 3d. DRb ranged from 0.01 to 0.5 (µ 0.29, σ 0.1), with internal finishes contributing the highest CDRi (µ 0.22, σ 0.07). Internal finishes subcomponents observe a CDRi range between 0 and 0.35 (Fig. 3b). A high proportion of damage occurs for 0 m to 0.5 m water depth above floor level whereby several internal finishes subcomponents (i.e., internal doors, floor finishes, wall finishes, and fittings and fixtures) were damaged on water contact. Physical damage of composite timber materials used for floor finishes occurred, often resulting in indirect damage to other internal finishes (e.g., internal walls, ceiling finishes) and services (e.g., plumbing) during cleanup and repair activities. Native hardwood timber floor finishes, although more resistant to damage from direct water contact, usually required replacement upon drying due to warping. Internal finishes in buildings constructed with suspended floors sustained a slighly higher CDRi (µ 0.23, σ 0.07) compared to those with concrete slab foundations (µ 0.2, σ 0.07).
External finishes observed relativey low CDRi (µ 0.03; σ 0.02) in comparison to internal finishes (Fig. 3a). Subcomponents damage susceptibility varies with windows and doors more frequently damaged (90%) compared to external walls (28%). Materials such as, timber doors and glass panes were highly susceptible to physical damage on water contact. Despite external windows and doors being frequently damaged their relative contribution to DRb is low on average (µ 0.03, σ 0.01) compared to subcomponents comprising internal finishes (Fig. 3c).
Service finishes (i.e., electrical and heating) were frequently damaged below floor level due to water contact of nearto ground condenser units for heating and airconditioning systems. Electrical and heating services damaged below floor level have a minor contribution to CDRi (µ < 0.01) for services (Fig. 3d). Electrical fittings were replaced on water contact while electrical wiring in buildings constructed after 1960 was reusable. Services CDRi (µ 0.05, σ 0.01) between 0 m and 0.5 m water depth above floor level often resulted from indirect damage to electrical and heating and plumbing services caused during internal finishes cleanup and repair.
3.2 Vulnerability models
In preparation for empirical vulnerability model development hazard and exposure variable importance for DRb was evaluated. Water depth above ground and floor level demonstrated the highest importance (Fig. 4), which corroborated many international studies emphasising its primary influence for DRb [5, 26]. Water depth above floor level had higher importance though the variable is rarely considered in international damage models. Other hazard variables showed low importance indicating minimal influence on residential damage in Westport. Area and floor height showed relatively higher importance than dwelling type and structural frame which are variables often applied in univariable models to classify residential buildings into homogenous classes for DRb prediction [4].
Univariable models were configured to predict DRb in response to water depth above ground level. Water depth above floor level showed higher importance for DRb though unavailable floor height observations in many flood hazard areas requires the application of water depth above ground level in univariable models [5]. Precision and reliability metrics for univariable models are presented in Table 4. Precision (i.e., MSE, MAE) was relatively consistent for most models as demonstrated by the comparison of observed and predicted DRb in Fig. 5. The Random Forest model returned the lowest precision as demonstrated by higher MAE and overprediction from MBE. This observation diverges from several flood damage studies [9, 10] demonstrating that complex ensemblebased learning algorithms outperform more simple regressionbased curves. The square root function observes the highest precision and reliability overall for univariable models. The general high performance of simple regressionbased models may be influenced by residential building homogeneity with the damage sample. This suggests simple models may provide sufficient representation of Westport residential buildings in damage models for locationbased flood risk assessments.
Table 4
Performance metrics for univariable and multivariable regression models evaluated in this study. Equations for depthdamage curves represent water depth above ground level (Depth) and damage ratio (DRb). Note. the ‘depthdamage curve’ for the event model is presented as an array for linear interpolation.
Model

MSE

MAE

MBE

QR

HR

Equation

Event Model

0.009

0.073

0.009

0.09

0.86

Depth = [0.0, 0.25, 0.5, 1.0, 1.25, 1.5, 3.0]
DRb = [0.0, 0.18, 0.26, 0.30, 0.35, 0.36, 1.0]

Linear Function

0.009

0.074

0.0005

0.09

0.86

DRb = 0.21∙Depth

Power Function

0.009

0.071

0.0001

0.12

0.87

DRb = 0.46∙Depth0.22

2nd Order Polynomial Function

0.009

0.071

0.0004

0.12

0.86

DRb = 0.11∙Depth2 + 0.22∙Depth

Square Root Function

0.009

0.07

0.0001

0.12

0.88

DRb = 0.47∙\(\sqrt{\text{D}\text{e}\text{p}\text{t}\text{h}}\)

Random Forest:
Univariable

0.009

0.074

0.0011

0.17

0.84


Random Forest: All Variables

0.007

0.067

0.0006

0.13

0.89


Random Forest:
Important Variables

0.007

0.062

0.0008

0.14

0.89


Multivariable models analyse nonlinear interactions between hazard and exposure variables to evaluate damage from multiple interdependent relationships [12]. Here, the Random Forest algorithm was learned on: 1) all hazard and exposure variables; and 2), variables with high importance for DRb prediction (i.e., > 0.05 mean decrease accuracy). Multivariable models demonstrate higher predictive precision compared to univariable models (Table 4). This is illustrated in Fig. 6. where observed and predicted DRb demonstrate closer proximity to the identity line. The model learned on important variables showed a prediction precision (MAE) 10% higher than the other univariable or multivariable models evaluated. While these more complex models demonstrated higher prediction uncertainty (QR), prediction hit rates close to 0.9 indicates high reliability. Higher precision and reliability of models learned on important variables is consistent with [8] who observed a loss in multivariable performance occurs from inclusion of variables with lower importance for DRb.
Westport residential building damage models demonstrate the importance of empirical data representing local flood damage processes. The majority (92%) of the 247 buildings assessed in Westport comprised one storey and a timber structural frame. These attributes did not demonstrate importance for DRb. However, the presence of homogenous construction material and subcomponent susceptibility to flood damage resulted in relatively precise and reliable DRb predictions compared with simple regressionbased univariable models. Westport is a small urban area (population < 10,000) and univariable model transfer to larger urban areas with broader building heterogeneity could lead to lower model prediction performance. Multivariable models demonstrated improved prediction performance from a larger hazard and exposure explanatory variable range which could be considered for future damage assessment in Westport. The larger model explanatory variable range is more likely to represent local damage factors in other locations [23]. Therefore, models learned in Westport should be applied and tested in other urban areas to evaluate their capacity for spatial transfer.