2018 Kerala Flood Damage: Survey, Identification, and Damage Prediction Models Using Machine Learning
Publication: ASCE OPEN: Multidisciplinary Journal of Civil Engineering
Volume 1, Issue 1
Abstract
The 2018 flood experienced by Kerala, the southernmost state of India, has affected thousands of lives, resulting in heavy financial loss (248,930 million INR) and structural damage (129,420 residential buildings have been reported to be damaged). Depth versus damage relationships have been used conventionally for flood emergency decision-making; however, these are approximate and do not account for all the significant factors involved in the flood damage assessment. An extensive questionnaire survey was conducted in this study to collect data comprising significant parameters of flood-damaged residential buildings due to the 2018 Kerala flood event. The damage to the buildings was expressed in terms of damage ratios, and the buildings were classified into various damage states according to their damage ratios. Machine learning (ML) techniques, such as random forest, naïve Bayes, decision tree, K-nearest neighbors, AdaBoost, XGBoost, support vector machine, LightGBM, and CatBoost, were used in this study to develop models to predict and classify the buildings into various flood damage states. Results indicated that almost all the tree-based ML techniques performed well, and the random forest model was obtained as the best-performing classification model for flood damage prediction with 84% accuracy. Further SHAP (SHapely Additive exPlanations) analysis was conducted to explore the relative importance of the different parameters involved in damage prediction.
Practical Applications
Flood damage assessment of residential buildings is crucial for emergency recovery operations and disaster management. Conventionally, postflood damage assessment around the world is carried out through field inspection by an expert team immediately after the event, which is physically tiresome, costly, and time-consuming. The machine learning (ML)-based methods proposed in the present study can be used effectively for rapid postflood damage assessment of residential buildings. The buildings with high-risk category can be identified easily with the proposed ML models. This enables the authorities for efficient decision-making for proper allocation of the emergency response and rescue operations. The same system can be used very effectively incorporating GIS tools for pre- and postflood damage assessment. The proposed method can contribute to the prediction of flood damage cost provided the data related to flood depth, duration, and the data related to building are available. While the flood depth and duration data can be made available from the flood forecasting models, the building related data can be obtained from the district administration.
Introduction
Flood is one of the worst natural catastrophes that cause significant damage to human life, infrastructure, and agriculture (Sun et al. 2021). Recent floods show that even developed nations are susceptible to serious devastation (Fig. 1). As the flood water depth around a structure increases, hydrostatic and dynamic loads exert lateral pressure on the walls (Kelman and Spence 2004). The weight of sediments deposited against the structure also build up further pressure on the walls. Prolonged exposure of the building to flood water may damage the building materials. Mitigation and planning for flood disasters are possible through reliable damage assessment. This can be achieved by employing more accurate flood risk models (Danso-Amoako et al. 2012) and implementing the latest available techniques.
Previous studies (Appelbaum 1985; Davis and Skaggs 1992; Smith 1994; Kelman 2000; Pistrika and Jonkman 2010; Pistrika et al. 2014; Yildirim 2017) employed different methods to evaluate the cost of damage due to flood in buildings. The flood depth versus damage function is the conventional tool used for estimating the cost of damage due to flood in structures. In this approach, the flood depth is assumed to be the only significant variable that affects the damage (Davis and Skaggs 1992; Kelman 2000; Komolafe et al. 2018; Pistrika et al. 2014; Pistrika and Jonkman 2010; Smith 1994; Yildirim 2017). Previous studies (Komolafe et al. 2019) have also identified other related variables (Marvi 2020) such as duration of flooding, material used for construction, and floor area of the building as the parameters that affect the damage (Vatteri et al. 2022). The flood damage models are improved by including the aforementioned parameters using multiple linear regression methods (Chinh et al. 2016; Komolafe et al. 2019). However, none of the earlier studies used machine learning (ML) techniques to develop flood damage prediction models.
In recent years, many researchers have implemented ML techniques for prediction of the failure mode of bridge columns (Mangalathu and Jeon 2019); beam–column joints (Mangalathu and Jeon 2018); columns and shear walls (Mangalathu et al. 2020b, c); classification of buildings based on earthquake damage (Mangalathu et al. 2020a); flood prediction (Mosavi et al. 2018); and damage analysis of infrastructures (Ganguly et al. 2019). These ML approaches were used in the present study to develop a flood damage classification model in terms of significant parameters. The 2018 Kerala flood damage data were collected and processed to train and test the selected machine learning models to classify the buildings to different flood damage states. Parameters such as flood depth, flood duration, floor area, age of the building, roof material, and wall material were considered in this study for flood damage prediction. SHapely Additive exPlanations (SHAP) (Lundberg and Lee 2017) was also employed in this study to interpret the performance of the best ML algorithm in predicting flood damage.
Research Significance
The flood depth–damage relationship has been developed in previous studies for flood damage prediction. Although it is simple and easy to use for estimating damage for future flood events, it considers only the flood depth as the significant parameter ignoring other parameters such as duration of flooding, age of the building, floor area of the building, and material used for construction that affect the damage. Studies dealing with the application of ML techniques for flood damage prediction are not available. The present study is the first attempt to propose a methodology based on ML classification algorithms to classify buildings into various flood damage states incorporating other related parameters mentioned previously. SHAP analysis, a tool to measure the importance of a variable in the outcome, has not yet been used in the flood damage assessment. SHAP analysis was used in the present study to rank various parameters involved according to their importance and also to justify the performance of ML in classifying buildings into various flood damage states. The methodology and results presented in this study will be helpful for the stakeholders to implement such tools in flood disaster management.
Proposed Methodology for Developing a Flood Damage Model Using Machine Learning Techniques
The proposed methodology for developing a flood damage assessment model is presented using a flowchart shown in Fig. 2. The method involves identifying significant flood and building damage parameters in the assessment. The next step is collecting those significant parameters pertaining to the affected residential buildings in the chosen 2018 Kerala flood event. Flood damage in the building was quantified using a damage ratio parameter as per an accepted method [explained in section “Damage Ratio (Dr)”]. Flood damage states were defined to classify the buildings into various damage levels. A database of input (significant flood damage parameters) and output (damage classification) was constructed and used for implementation in a chosen ML model. The performance of the chosen ML model using various parameters was assessed to obtain the best-performing ML algorithm. Finally, global and local interpretation of the best-performing machine learning algorithm was made using SHAP analysis.
Collection of Data
Kerala (a southern state of India) experienced 2,346.6 mm (92.4 in.) (42% higher than the average) of rain between June 1, 2018 and August 19, 2018, according to IMD (India Meteorological Department) statistics, compared to a predicted 1,649.5 mm (64.94 in.). Fig. 3 shows the photographs of flood levels and damage to some residential buildings in the 2018 flood. Four districts of Kerala, namely, Alappuzha, Kottayam, Malappuram, and Calicut (shown in Fig. 4), were selected for the flood data collection in this study. As shown in Fig. 5, a questionnaire was prepared to gather flood parameters (flood depth and duration of flood), exposure characteristics of the building (floor area of the building, wall material, roof material, and age of the building), and repair cost (cost of repair of the structure and the content).
Typical buildings in Kerala, being the extreme precipitation region, are provided with plinth heights varying between 0.20 m (7.9 in.) and 0.80 m (31.50 in.) using stone, laterite, and concrete to protect the building from frequent runoff. The total flood depth is taken as the height of water from the ground level due to the following two reasons: (1) unlike the buildings at other locations around the world, flood water up to the plinth level over a duration of 2–3 days may cause some damage to the foundation due to seepage, scouring, and debris accumulation, which is to be accounted; and (2) the other reason is due to the practical applicability of the proposed method with the flood forecasting models.
About 270 randomly selected owners were interviewed for data collection following the procedure proposed by Thakur et al. (2012). The maximum flood depth at the location of the surveyed buildings was found to be about 8 m. The maximum duration of the flood was observed to be about 15 days. Table 1 shows the number of buildings surveyed in each category according to the types of roof used for the construction, such as reinforced cement concrete (RCC), tiled roof, sheet roof, and so on. Of all the RCC buildings found during the survey, 67 buildings were constructed using brick walls, 17 using laterite walls, 10 using hollow brick walls, and 15 using solid block masonry walls.
| Wall material | Type of roof | ||
|---|---|---|---|
| RCC | Tile | Sheet | |
| Brick | 67 | 66 | 34 |
| Laterite | 17 | 15 | 3 |
| Hollow brick | 10 | 6 | 20 |
| Solid block | 15 | 7 | 10 |
The frequency distribution of the data collected with respect to various parameters related to flood and exposure conditions is depicted in Figs. 6(a–f). Fig. 6(a) indicates the total number of buildings in each range of flood depth. A total of 136 buildings were affected by a flood of less than 1.0 m (39.37 in.) depth (from the ground level). The number of buildings affected by a flood depth above 3 m (roughly equal to 118.11 in.) was found to be the least (seven).
Fig. 6(b) shows that 13 buildings experienced the flood for more than 8 days. Maximum numbers of buildings were found in the category that experienced flood for fewer than 4 days.
Fig. 6(c) shows that about 88% of the damaged buildings had a floor area of less than 140 m2. Fig. 6(d) indicates the number of buildings according to their age. It can be seen that about 64% of the surveyed buildings had an age of fewer than 21 years.
Fig. 6(e) shows the number of buildings as per the type of roof material used for the construction. The material commonly used in the roof is RCC, tiles, and GI (galvanized iron) sheets. The buildings with RCC, tiled, and sheet roofs are designated as “1,” “2,” and “3,” respectively. It can be observed that 109 (40%), 94 (35%), and 67 (25%) buildings were constructed with RCC, tiled, and sheet roofs, respectively.
Fig. 6(f) shows the number of buildings as per the type of wall material used for the construction. Different materials used to construct the wall are bricks, laterite, hollow blocks, and solid blocks. The buildings built with brick, laterite, hollow, and solid block walls are designated as “1,” “2,” “3,” and “4,” respectively. Fig. 6(f) shows that 167 (62%) buildings had brick-type walls. The statistical details of the various parameters of the surveyed buildings, such as flood depth (in m), duration of the flood (days), floor area (m2), age of the building (years), repair cost (in INR), and replacement cost in INR (T) are given in Table 2.
| Descriptive statistics | Flood depth (m) | Duration (days) | Floor area (m2) | Age of the building (years) | Repair cost (INR) | Replacement cost (INR) | Damage ratio (Dr) |
|---|---|---|---|---|---|---|---|
| Mean | 1.26 | 4.43 | 93.37 | 21.01 | 120,867 | 1,760,861 | 0.08 |
| Standard error | 0.06 | 0.16 | 37.89 | 0.98 | 17,100 | 65,880 | 0.01 |
| Median | 1 | 4 | 84 | 15 | 50,000 | 1,586,250 | 0.03 |
| Mode | 1.5 | 7 | 93 | 15 | 50,000 | 1,460,000 | 0 |
| Minimum | 0 | 0 | 20 | 0.5 | 0 | 68,000 | 0 |
| Maximum | 8 | 15 | 558 | 100 | 3,200,000 | 10,200,000 | 0.93 |
| Count | 270 | 270 | 270 | 270 | 270 | 270 | 270 |
Damage Ratio (Dr)
Flood damage in a building depends on the flood load and the resistance offered by different building components (Pistrika and Jonkman 2010; Kelman 2000). The damage percentage is assumed as the ratio of cost of damage in the structure and its content to the total cost.
In line with the studies of Pistrika et al. (2014) and Komolafe et al. (2019), the flood damage in a building was estimated in this study in terms of “damage ratio” (Dr), which is defined as the ratio of repair costs (denoted “R,” obtained from the survey) to the total replacement cost (denoted “T,” evaluated using the schedule of rates in India) as follows:
(1)
The damage ratios of all the buildings surveyed were calculated to prepare the data set. The statistical details of the damage ratios (Dr) obtained are given in Table 2. Three limit states (DL0, DL1, and DL2 given in Table 3) were identified to represent the damage states following the study reported by Nofal et al. (2020). Each building was classified into any of the three damage limit states (DL0, DL1, and DL2) based on their damage ratios. A database consisting of significant flood damage parameters as input and damage limit states (DL0, DL1, and DL2) as the output classification was constructed to be implemented in a machine learning framework. The correlation of various input parameters with the output (damage state) is summarized in Table 4.
| Damage level | Damage state | Dr |
|---|---|---|
| Mild | DL0 | 0.0–0.03 |
| Moderate | DL1 | 0.03–0.15 |
| Severe | DL2 | 0.15–1.0 |
| Parameters | Roof material | Wall material | Age of the building | Flood depth | Duration | Floor area | Damage state |
|---|---|---|---|---|---|---|---|
| Roof material | 1.000 | 0.098 | 0.165 | 0.152 | 0.154 | −0.213 | 0.257 |
| Wall material | 0.098 | 1.000 | −0.223 | −0.034 | −0.151 | 0.086 | −0.067 |
| Age of the building | 0.165 | −0.223 | 1.000 | 0.101 | 0.149 | 0.011 | 0.255 |
| Flood depth | 0.152 | −0.034 | 0.101 | 1.000 | 0.187 | 0.007 | 0.245 |
| Duration | 0.154 | −0.151 | 0.149 | 0.187 | 1.000 | −0.174 | 0.730 |
| Floor area | −0.213 | 0.086 | 0.011 | 0.007 | −0.174 | 1.000 | −0.293 |
| Damage state | 0.257 | −0.067 | 0.255 | 0.245 | 0.730 | −0.293 | 1.000 |
Machine Learning Algorithms
Nine ML algorithms were employed to check their performance and identify the best-performing ML model for classifying the buildings into the various damage limit states as per their flood- and building-related input parameters. The algorithms used were (1) random forest (RF), (2) naïve Bayes (NB), (3) decision tree (DT), (4) K-nearest neighbor (K-NN), (5) AdaBoost, (6) XGBoost, (7) support vector machine (SVM), (8) LightGBM, and (9) CatBoost, which are adopted from recently published work (Mangalathu et al. 2020c).
The NB algorithm is based on Bayes theorem, and it assumes that the variables are independent (Rish 2001). The K-NN method is used to classify data into different classes considering the distances from the neighbors (Awais et al. 2016; Goldberger et al. 2004). The DT algorithm is used for the classification and regression of data (Ghori et al. 2020). It is a treelike structure in which internal nodes denote the data sets, branches denote the decisions, and leaves denote the outcomes.
Random forest (RF) comprises several tree classifiers, each of which is constructed using a random vector tested separately from the input vector (Ghori et al. 2020; Mangalathu and Jeon 2018). RF employs two essential techniques: random feature subspace and out-of-bag estimations. The former creates the tree faster, while the latter allows assessing the relative importance (Pavlov 2019). AdaBoost, XGBoost, LightGBM, and CatBoost are the various methods (Chen and Guestrin 2016; Goldberger et al. 2004; Ke et al. 2017) used to increase the performance of a model (Mangalathu et al. 2020c; Mangalathu and Jeon 2018). In AdaBoost (adaptive boosting), the weights are reassigned at each instance. The primary use of boosting is to reduce bias and variance for supervised learning. CatBoost can handle categorical features in the input variables (Dorogush et al. 2018; Prokhorenkova et al. 2018).
In contrast to various DT-based methods, LightGBM is developed leafwise. XGBoost cannot convert categorical parameters into numerical values on its own. SVM develops a model from the available training data, which helps to allocate new data into a particular class. It builds a hyperplane that separates the feature space into two groups that can be extended to even multiclass problems (Mangalathu and Jeon 2018). A more detailed description of the aforementioned ML algorithms can be found in the literature (Dorogush et al. 2018; Freund and Schapire 1997; Ghori et al. 2020; Goldberger et al. 2004; Ke et al. 2017; Mangalathu et al. 2020c; Mangalathu and Jeon 2018; Pavlov 2019; Prokhorenkova et al. 2018).
Analysis of the Flood Data Using Machine Learning
A Python environment using Google Colab, a free Tensorflow-enabled Jupyter notebook, was used for machine learning. The implementation of the ML algorithms was done with the help of Scikit-Learn. Scikit-Learn provides different tools for selecting an appropriate model, fitting the model, data preprocessing, and data evaluation.
The development of a flood damage model in the present study involves data exploration, best model selection, and SHAP analysis to explain the significance of each parameter. The distribution of the data set into the different classes of damage states is illustrated in Fig. 7.
Seventy percent of the collected data was used for training the ML algorithm, and the remaining was used for testing. Data division into test and training sets was random, and the model’s performance on the test data indicated its performance on unknown data.
The entire data set consisted of 140 buildings in DL0, 85 in DL1, and 44 in DL2 damage states. This indicates a slight imbalance in the data collected. This kind of skewness is expected in a real-life scenario; hence, this cannot be fully omitted. Several methods for minimizing such data imbalance are reported in the literature (Kiani et al. 2019), which include oversampling, undersampling, synthetic data generation, class weight, and adjusting the threshold. The undersampling method causes loss of information from the omitted samples from the majority class, while oversampling might lead to overfitting due to the usage of the same samples in the data set. Synthetic data generation overcomes the imbalance problem by generating artificial data. Adjusting threshold is a manual technique to balance a data set. The present study employed weighted class method to balance the data. To calculate the weight of each class, the weight of the biggest class was assigned to one and the weight of the smallest class was assigned to the ratio of the number of samples of the biggest class to the number of samples of the smallest class.
All the input parameters were brought to the same scale using MinMaxScaler to obtain reliable results and accelerate learning. Also, the randomized grid search method was used to find the optimal parameters and also to hypertune the model. Optimizing hyperparameters using randomized grid search cross-validation prevents overfitting in the models. A 10-fold cross-validation method was adopted in this study. The prediction model was established using 10-fold cross validation of the training set (the training set was clustered into 10 folds). In each training set fold, a model was established with 90% of the training set, and its performance was evaluated with the remaining 10%. The procedure was repeated three times for the training set, and the average of the weights was used to establish the prediction model.
The ML models were trained using the training data sets, and the performance of each ML model was assessed in terms of the confusion matrix (CM). Each element of the confusion matrix, Cij, can be defined as the number of actual observations in a damage state “i” but predicted to damage state “j.” The confusion matrices for each ML algorithm for the training and test data sets are given in Figs. 8 and 9, respectively. The performance of the ML model (Friedman and Hastie 2001) was evaluated in terms of parameters, precision, recall, F1 score, and accuracy (in percentage), which are obtained from CM.
The precision parameter represents the number of correct predictions to the total number of predictions of a damage state expressed as a percentage. The fourth row of the CM shows the precision of each damage state. The first column of CM given in Fig. 8(b) shows the number of predictions of the DL0 damage state for the NB algorithm. Of 96 (82 + 11 + 3) predictions, 82 were correctly predicted to DL0. The precision of the DL0 damage state for the NB algorithm was obtained as 85% (100 × 82/96).
Recall is the ratio of the number of correct predictions of a damage state to the total number of actual observations in the same damage state expressed as a percentage, which is shown in the fourth column of the CM. The first row of CM given in Fig. 8(b) represents the observed class of the DL0 damage state. Of 100 (82 + 17 + 1) actual observations, 82 were correctly predicted to DL0. However, 17 and 1 were wrongly predicted as DL1 and DL2, respectively. The recall of the DL0 damage state for the NB algorithm was obtained as 82% (100 × 82/100).
The ratio of the number of correct predictions (sum of the diagonal elements) of all damage states to the total number of observations in all damage states expressed as a percentage is known as accuracy and is shown at index (4, 4) of the CM. Since accuracy is defined for all damage states, it is considered as a global performance measure. Fig. 8(b) shows that the NB algorithm had an accuracy of 75%, which is obtained by dividing 141 (82 + 45 + 14) by 189 (total observations of the training data set). Similarly, precision, recall, and accuracy of the selected ML algorithms can be read for training (Fig. 8) and test (Fig. 9) data sets.
The F1 score is defined as the weighted harmonic mean of the recall and precision as follows:
(2)
High values of the aforementioned performance measures indicate the efficiency of model in classifying the damage states. The performance assessment of nine different ML algorithms was carried out by comparing the four performance measures.
Figs. 8 and 9 show that all the ML algorithms perform well with a minimum accuracy of 75% (NB) and 65% (K-NN) for training and test data, respectively. It was observed that the accuracy of the model for training data was 100% for three algorithms, decision tree, random forest model, and CatBoost. In contrast, the performance on test data showed the highest accuracy for the random forest model (84%), followed by CatBoost (83%) and XGBoost (81%). The better performance of tree-based models was due to the complex nonlinear decision boundaries that separate the damage states. The RF model had recall values of 98%, 84%, and 50% for DL0, DL1, and DL2, respectively. Similarly, RF had 91%, 72%, and 89% precision for DL0, DL1, and DL2, respectively.
Table 5 shows the summary of the accuracy of all the ML models for training and test data. The accuracy RF model was about 100% for training and 84% for test data, which is the highest among all the ML models considered. F1 scores for training and test data are listed in Tables 6 and 7, respectively, for all the ML models. Evaluation of the performance of all the ML models for each damage state in terms of the F1 score showed that the CatBoost (79%) model had marginally higher F1 score than to the RF model (78%). Even though the CatBoost model had slightly more F1 score than RF, from a practical viewpoint, to use a single algorithm to classify all damage states, the present study employed RF (which has higher accuracy) for further analyses.
| Sl. no | Algorithm used | Training set | Test set |
|---|---|---|---|
| a | RF | 100 | 84 |
| b | Naïve Bayes | 75 | 78 |
| c | DT | 100 | 69 |
| d | K-NN | 80 | 65 |
| e | AdaBoost | 77 | 81 |
| f | XGBoost | 96 | 81 |
| g | SVM | 80 | 73 |
| h | LightGBM | 98 | 79 |
| i | CatBoost | 100 | 83 |
| Damage state | RF | NB | DT | KNN | AdaBoost | XGBoost | SVM | LightGBM | CatBoost |
|---|---|---|---|---|---|---|---|---|---|
| DL0 | 100 | 83.47 | 100 | 88.16 | 85.36 | 98.00 | 88.60 | 99.00 | 100 |
| DL1 | 100 | 67.28 | 100 | 75.79 | 64.19 | 93.99 | 72.50 | 97.50 | 100 |
| DL2 | 100 | 53.73 | 100 | 47.33 | 72.00 | 94.74 | 56.72 | 98.50 | 100 |
| Average | 100 | 68.16 | 100 | 70.43 | 73.85 | 95.58 | 72.61 | 98.33 | 100 |
| Damage state | RF | NB | DT | KNN | AdaBoost | XGBoost | SVM | LightGBM | CatBoost |
|---|---|---|---|---|---|---|---|---|---|
| DL0 | 94.37 | 86.95 | 78.99 | 79.12 | 88.99 | 90.43 | 86.01 | 89.36 | 91.50 |
| DL1 | 77.54 | 75.23 | 52.98 | 57.98 | 74.47 | 76.36 | 62.98 | 68.20 | 75.23 |
| DL2 | 64.03 | 57.93 | 69.00 | 28.86 | 73.66 | 64.03 | 48.50 | 69.0 | 71.63 |
| Average | 78.65 | 73.37 | 66.99 | 55.32 | 79.04 | 76.94 | 65.83 | 75.52 | 79.45 |
All the ML models can predict the DL0 damage state fairly well. However, only RF, XGBoost and CatBoost can predict the DL1 damage state accurately. All the ML models were found to have low recall values in identifying the DL2 damage state. This may be due to the low percentage of buildings in the DL2 damage state in the original data set.
Although the imbalance was minimized by employing the methods discussed earlier, the performance of the RF model for DL2 prediction was about 50%. Of 16 buildings in the DL2 damage state, the RF algorithm predicted 8 (50%) correctly. However, seven buildings (44%) were incorrectly predicted to the DL1 damage state and one building (6%) to the DL0 damage state. This means that in real-life use of the proposed algorithm, it may predict about 44% of buildings in the DL1 damage state that are actually in DL2. Although the prediction of the algorithm is incorrect for 50% of total buildings in DL2 damage state (i.e. 44% to DL1 and 6% to DL0), it will still predict 44% to DL1, which is very close to DL2 damage state. In terms of total flood damage, this prediction may underestimate the damage; however, the allocation process of the emergency response team can still work well. The response team capable of handling the DL1 damage state may be visiting the location of the DL2 damage state, which is even better than the current practice.
SHAP Analysis
The concept of the Shapley value (SHAP) was initially developed to assess the significance of an individual player in a team (Lundberg and Lee 2017). SHAP assumes the model output as a linear summation of various input variables and allocates a significance value to every feature for a particular prediction. SHAP analysis provides the feature importance (global interpretation) of the entire data set and the local interpretation of each sample in the data set. The magnitude and direction of influence (summary plot) of each input parameter in the prediction in terms of relative significance can be obtained from SHAP analysis. More details and applications of SHAP analysis are available elsewhere (Lundberg et al. 2018; Lundberg and Lee 2017; Mangalathu et al. 2020b, 2022).
The importance of each input parameter in predicting the output is shown in Fig. 10. Parameter duration of flood was found to have a significant impact (0.34) on the flood damage ratio (given a set of input parameters), followed by parameter floor area of the building (0.2). Wall material and roof material had negligible significance in predicting the output.
The global importance (in an average sense), which is calculated as the average of SHAP values of each input parameter on each damage state, is illustrated schematically in Fig. 11. The six input variables were arranged from top to bottom based on their importance (most important variable at the top and least at the bottom). If the mean SHAP value is high, the significance of that particular variable will also be high. It was observed that parameter duration has a high SHAP value for all damage states, which suggests that it plays a major role in predicting the damage state of the sample. The order of input parameters according to their significance (with the least significance for the parameter wall material) in the prediction of each damage states was found to be duration, floor area, flood depth, age of the building, roof material, and wall material. It was also observed that duration has more influence on the classification of the damage state into DL0, which is indicated by a large proportion of blue color in the bar representing parameter duration compared to other parameters.
Three samples with the same predicted and actual damage states were considered to study the factors contributing to the prediction in terms of force plots, as shown in Fig. 12. Fig. 12(a) shows the force plot of Sample no. 3 for the DL0 damage state. The base value is the percentage of samples that fall into a particular damage state. The base value of 0.5249 (=100/189, out of 189 samples, 100 belongs to DL0) indicates the proportion of the DL0 damage state in the data set. Since three damage states are considered, when the probability of prediction of a sample to a damage state is greater than 0.33 (=1/3), it is more likely to be classified into that damage state. In Fig. 12(a), the probability of DL0 is 0.83 (shown in bold), which shows that Sample no. 3 has more probability of being classified into DL0.
Similarly, Figs. 12(b and c) display the force plots of the 100th and 80th data samples. It can be seen that the probability of prediction of the 100th sample to be classified into the DL0 damage state is 0.76 and that of the 80th sample to be classified into the DL1 damage state is 0.78. Fig. 12(b) shows that for the 100th sample, parameters floor area and age of the building influence the prediction of the DL0 damage state negatively, while all other parameters positively influence the prediction of DL0. Similarly, Fig. 12(c) shows that only wall material negative influences the prediction of the DS1 damage state for the 80th sample.
Figs. 13(a–c) show the summary plots for the three damage states. Every point in the summary plot has three attributes, namely, the order of significance (Attribute 1) of the input parameter, the SHAP value (Attribute 2), and the value of the input parameter (Attribute 3). The selected input parameters are arranged based on their significance (Attribute 1). The most important parameter is shown at the top and the least at the bottom. Attribute 2 of each input parameter is shown in the abscissa. Positive SHAP values indicate that any increase in that parameter will have more probability of being classified into that damage state. Blue denotes the low value of Attribute 3 of the input variable, and red shows the high value. The conclusions obtained from Figs. 13(a–c) are explained as follows:
1.
Duration is the most significant parameter for all the damage states (DL0, DL1, and DL2).
2.
Fig. 13(a) shows that a decrease in duration will increase the probability of being classified into the DL0 damage state. The next important input parameter is the floor area. It can be seen that an increase in floor area will increase the probability of classifying it into DL0. Although it may contradict the common notion around the world, when the floor area increases, the quality of the residential building in terms of construction, materials used, and engineering involved becomes superior in the context of Kerala. Due to this, the damage cost as a percentage of the total cost of the building can be less for buildings with larger floor areas. The three important parameters are flood depth, building age, and roof material. For these three parameters, the higher-valued samples have negative SHAP values, while lower-valued samples have positive SHAP values. Therefore, any increase in the values of these three parameters may reduce the probability of being classified into DL0. Due to their high SHAP values, RCC roof material and solid block wall influence the DL0 damage state prediction.
3.
Similarly, duration is the most significant parameter, followed by floor area and flood depth for DL1, as shown in Fig. 13(b). An increase in the value of duration results in the classification of that data into DL1. Similarly, brick wall buildings have more probability of being classified into DL1.
4.
Since the SHAP values of samples having higher values of age (red color dots) are positive, they will have more probability of being classified into DL2, i.e., old houses are more likely to have severe damage [Fig 13(c)]. Parameter age increases the degradation in the strength of the building materials, which increases the damage. Also, an increase in flood depth results in the DL1 damage state.
Conclusions
Although the conventional flood depth-damage function is a popular tool for flood emergency decision-making, it is crude as it considers only flood depth as the independent variable ignoring all the contributing factors affecting the damage. The present study focuses on applying recently introduced ML classification models for the flood damage assessment of residential buildings, taking into account the essential parameters involved.
Flood data of buildings (flood depth, wall material, floor area, roof material, duration, and building age) are collected through an extensive survey from the four districts, Alappuzha, Kottayam, Malappuram, and Calicut of the Kerala state. The flood damages in the selected buildings are expressed in terms of damage ratios, and based on the damage ratios, the buildings are classified into three damage states (DL0, DL1, and DL2). Suitable measures available in literature are taken to address the imbalance in the data. Nine ML algorithms (random forest, naïve Bayes, decision tree, K-nearest neighbor, AdaBoost, XGBoost, support vector machine, LightGBM, and CatBoost) are used in this study to predict and classify the buildings to various damage states. The performance of the machine learning models is assessed through the parameters of recall, precision, accuracy, and F1 score to obtain the best ML model.
In general, tree-based machine learning models are found to perform well considering the parameters of recall, precision, accuracy, and F1 score. The accuracy of the RF model was the highest (84%), and the F1 score of the CatBoost model was the highest (79%). Considering the convenience in implementation for the classification of damage states, the present study considers RF as the best ML classification model for flood damage assessment. RF predicts DL0 and DL1 reasonably well, but due to the fewer number of samples in the DL2 damage state, 44% of the buildings actually in the DL2 damage state are incorrectly predicted in the DL1 damage state. In terms of the total flood damage, this prediction may underestimate the same slightly; however, the allocation process of emergency response and rescue team can still work well.
The input parameters are ranked based on their significance using an SHAP analysis. The results of the SHAP analysis of the RF model show that duration is the most critical parameter in predicting the damage. The summary plot obtained after the SHAP analysis is used to explain the impact of a particular variable in predicting the damage state. In general, it can be concluded that flood damage reduces with the decreasing values of duration, flood depth, and age of the building.
One of the limitations of this study is the use of relatively fewer sample points (270). However, with these limited data, the current study shows a potential way to use ML-based models in flood risk assessment. Currently, the damage identification is done using field survey to collect the data, which is very costly and time-consuming, and the proposed model shows a promising path for rapid damage assessment using ML-based models. Also, the model can be improved further by adding more data, and further research is needed in that direction. Although the application of the best-performing ML model is limited to the specific region selected in this study, the methodology proposed here can be applied to any other location. As an efficient, economical, and novel alternative, ML-based techniques can potentially simplify flood damage prediction models incorporating significant variables for rapid postflood damage assessment of residential buildings and indirectly contribute to disaster mitigation.
Data Availability Statement
Some or all data, models, or codes that support the findings of this study are available from the corresponding author upon reasonable request.
References
Appelbaum, S. J. 1985. “Determination of urban flood damages.” J. Water Resour. Plann. Manage. 111 (3): 269–283. https://doi.org/10.1061/(asce)0733-9496(1985)111:3(269).
Awais, M., L. Palmerini, and L. Chiari. 2016. “Physical activity classification using body-worn inertial sensors in a multi-sensor setup.” In Proc., 2nd Int. Forum on Research and Technologies for Society and Industry Leveraging a Better Tomorrow. Bologna, Italy: University of Bologna.
Chen, T., and C. Guestrin. 2016. “XGBoost: A scalable tree boosting system.” In Proc., 22nd ACM Sigkdd Int. Conf. on Knowledge Discovery and Data Mining, 785–794. New York: Association for Computing Machinery.
Chinh, D. T., A. K. Gain, N. V. Dung, D. Haase, and H. Kreibich. 2016. “Multi-variate analyses of flood loss in Can Tho city, Mekong delta.” Water (Switzerland) 8 (1): 1–21.
Danso-Amoako, E., M. Scholz, N. Kalimeris, Q. Yang, and J. Shao. 2012. “Predicting dam failure risk for sustainable flood retention basins: A generic case study for the wider Greater Manchester area.” Comput. Environ. Urban Syst. 36 (5): 423–433. https://doi.org/10.1016/j.compenvurbsys.2012.02.003.
Davis, S. A., and L. Skaggs. 1992. Catalog of residential depth-damage functions used by the army corps of engineers in flood damage estimation. Alexandria, VA: Army Engineer Institute for Water Resources.
Dorogush, A. V., V. Ershov, and A. Gulin. 2018. “CatBoost: Gradient boosting with categorical features support.” Preprint, submitted October 24, 2018. http://arxiv.org/abs/arXiv.1810.11363.
Freund, Y., and R. E. Schapire. 1997. “A decision-theoretic generalization of on-line learning and an application to boosting.” J. Comput. Syst. Sci. 55: 119–139. https://doi.org/10.1006/jcss.1997.1504.
Friedman, J., and T. T. R. Hastie. 2001. The elements of statistical learning. Dordrecht, Netherlands: Springer.
Ganguly, K. K., N. Nahar, and B. M. M. Hossain. 2019. “A machine learning-based prediction and analysis of flood affected households: A case study of floods in Bangladesh.” Int. J. Disaster Risk Reduct. 34: 283–294. https://doi.org/10.1016/j.ijdrr.2018.12.002.
Ghori, K. M., R. A. Abbasi, M. Awais, M. Imran, A. Ullah, and L. Szathmary. 2020. “Performance analysis of different types of machine learning classifiers for non-technical loss detection.” IEEE Access 8: 16033–16048. https://doi.org/10.1109/ACCESS.2019.2962510.
Goldberger, J., R. Sam, H. Geoff, and R. Salkhutdinov. 2004. “Neighbourhood components analysis.” In Proc., 17th Int. Conf. on Neural Information Processing Systems, 513–520. Toronto, ON: University of Toronto.
Ke, G., Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T. Y. Liu. 2017. “LightGBM: A highly efficient gradient boosting decision tree.” In Advances in neural information processing systems 30, edited by I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, 3147–3155. Red Hook, NY: Curran Associates.
Kelman, I. 2000. Physical flood vulnerability of residential properties in coastal eastern England. Cambridge, UK: Univ. of Cambridge.
Kelman, I., and R. Spence. 2004. “An overview of flood actions on buildings.” Eng. Geol. 73 (3–4): 297–309. https://doi.org/10.1016/j.enggeo.2004.01.010.
Kiani, J., C. Camp, and S. Pezeshk. 2019. “On the application of machine learning techniques to derive seismic fragility curves.” Comput. Struct. 218: 108–122. https://doi.org/10.1016/j.compstruc.2019.03.004.
Komolafe, A. A., S. Herath, and R. Avtar. 2018. “Sensitivity of flood damage estimation to spatial resolution.” J. Flood Risk Manage. 11: S370–S381. https://doi.org/10.1111/jfr3.12224.
Komolafe, A. A., S. Herath, and R. Avtar. 2019. “Establishment of detailed loss functions for the urban flood risk assessment in Chao Phraya River basin, Thailand.” Geomatics Nat. Hazards Risk 10 (1): 633–650. https://doi.org/10.1080/19475705.2018.1539038.
Lundberg, S. M., G. G. Erion, and S.-I. Lee. 2018. “Consistent individualized feature attribution for tree ensembles.” Preprint, submitted February 12, 2018. http://arxiv.org/abs/1706.06060.
Lundberg, S. M., and S. I. Lee. 2017. “A unified approach to interpreting model predictions.” In Advances in neural information processing systems 30, edited by I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, 4766–4775. Red Hook, NY: Curran Associates.
Mangalathu, S., M. Eeri, H. Sun, C. C. Nweke, Z. Yi, and H. V. Burton. 2020a. “Classifying earthquake damage to buildings using machine learning.” Earthquake Spectra 36 (1): 183–208. https://doi.org/10.1177/8755293019878137.
Mangalathu, S., S.-H. Hwang, and J.-S. Jeon. 2020b. “Failure mode and effects analysis of RC members based on machine-learning-based SHapley Additive exPlanations (SHAP) approach.” Eng. Struct. 219: 110927. https://doi.org/10.1016/j.engstruct.2020.110927.
Mangalathu, S., H. Jang, S.-H. Hwang, and J.-S. Jeon. 2020c. “Data-driven machine-learning-based seismic failure mode identification of reinforced concrete shear walls.” Eng. Struct. 208: 110331. https://doi.org/10.1016/j.engstruct.2020.110331.
Mangalathu, S., and J. Jeon. 2019. “Machine learning-based failure mode recognition of circular reinforced concrete bridge columns: Comparative study.” J. Struct. Eng. 145 (10). https://doi.org/10.1061/(ASCE)ST.1943-541X.00024.
Mangalathu, S., and J.-S. Jeon. 2018. “Classification of failure mode and prediction of shear strength for reinforced concrete beam–column joints using machine learning techniques.” Eng. Struct. 160: 85–94. https://doi.org/10.1016/j.engstruct.2018.01.008.
Mangalathu, S., K. Karthikeyan, D.-C. Feng, and J.-S. Jeon. 2022. “Machine-learning interpretability techniques for seismic performance assessment of infrastructure systems.” Eng. Struct. 250: 112883. https://doi.org/10.1016/j.engstruct.2021.112883.
Marvi, M. T. 2020. “A review of flood damage analysis for a building structure and contents.” Nat. Hazard. 102 (3): 967–995. https://doi.org/10.1007/s11069-020-03941-w.
Mosavi, A., P. Ozturk, and K. W. Chau. 2018. “Flood prediction using machine learning models: Literature review.” Water (Switzerland) 10 (11): 1–40.
Nofal, O. M., J. W. van de Lindt, and T. Q. Do. 2020. “Multi-variate and single-variable flood fragility and loss approaches for buildings.” In Reliability engineering and system safety. Toronto, ON: University of Toronto.
Pavlov, Y. L. 2019. Random forests, 1–122. Berlin, Boston: De Gruyter.
Pistrika, A., G. Tsakiris, and I. Nalbantis. 2014. “Flood depth-damage functions for built environment.” Environ. Processes 1: 553–572. https://doi.org/10.1007/s40710-014-0038-2.
Pistrika, A. K., and S. N. Jonkman. 2010. “Damage to residential buildings due to flooding of New Orleans after hurricane Katrina.” Nat. Hazard. 54 (2): 413–434. https://doi.org/10.1007/s11069-009-9476-y.
Prokhorenkova, L., G. Gusev, A. Vorobev, A. V. Dorogush, and A. Gulin. 2018. “CatBoost: Unbiased boosting with categorical features.” In Advances in neural information processing systems 31, edited by S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, 6638–6648. Red Hook, NY: Curran Associates.
Rish, I. 2001. “An empirical study of the naive Bayes classifier.” In IJCAI 2001 Workshop on Empirical Methods in Artificial Intelligence, 41–46. New York: T.J. Watson Research Center.
Smith, D. I. 1994. “Flood damage estimation—A review of urban stage-damage curves and loss functions.” Water SA 20 (3): 231–238.
Sun, X., R. Li, X. Shan, H. Xu, and J. Wang. 2021. “Assessment of climate change impacts and urban flood management schemes in central Shanghai.” Int. J. Disaster Risk Reduct. 65: 102563. https://doi.org/10.1016/j.ijdrr.2021.102563.
Thakur, P. K., S. Maiti, N. C. Kingma, V. Hari Prasad, S. P. Aggarwal, and A. Bhardwaj. 2012. “Estimation of structural vulnerability for flooding using geospatial tools in the rural area of Orissa, India.” Nat. Hazard. 61 (2): 501–520. https://doi.org/10.1007/s11069-011-9932-3.
Vatteri, A. P., D. D’Ayala, and P. Gehl. 2022. “Bayesian networks for assessment of disruption to school systems under combined hazards.” Int. J. Disaster Risk Reduct. 74: 102924. https://doi.org/10.1016/j.ijdrr.2022.102924.
Yildirim, E. 2017. “Hazus-MH flood loss estimation on a web-based system.” M.S. thesis, Dept. of Civil and Environmental Engineering,Univ. of Iowa.
Information & Authors
Information
Published In
ASCE OPEN: Multidisciplinary Journal of Civil Engineering
Volume 1 • Issue 1 • December 2023
Copyright
This work is made available under the terms of the Creative Commons Attribution 4.0 International license, https://creativecommons.org/licenses/by/4.0/.
History
Received: Feb 8, 2023
Accepted: Aug 14, 2023
Published online: Sep 12, 2023
Published in print: Dec 31, 2023
Discussion open until: Feb 12, 2024
Authors
Metrics & Citations
Metrics
Citations
Download citation
If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download.