Flood waste management is important for reducing the damage and secondary environmental pollution caused by delays in disaster recovery. One key issue related to flood waste management concerns estimating the precise quantity of waste to plan recovery strategies and policy. Multilinear regression has been recognized as a viable technique to estimate the amount of waste generation from flood. There are two types of flood waste estimation methods: pre-event predictions using factors related to regional properties and rainfall hazards, and post-event predictions using damage variables due to floods, such as the number of damaged buildings. However, adapting the framework suggested in other countries did not work in Korean cases. In this study, an advanced flood waste estimation technique was devised using data grouping, Bayesian linear regression, incorporating interaction terms and nonlinear regression using deep-neural network. The aim of this study was 1) to develop the flood waste regression model using aforementioned four frameworks; 2) to evaluate the performance and limitation of each method; and 3) to suggest the most viable strategy by comparison of the empirical validation results of the models established upon the frameworks. Totally 90 flood cases (2008-2017) in South Korea collected from annual report on disaster published by National Emergency Management Agency in Korea were the scope of this study.

The data grouping was performed in post-event flood waste estimation following three grouping characteristics: administrative region (AR; equivalent to special city or province), urbanization rate (UR), and disaster type and offshore accessibility (DO). Such data grouping led to flood waste prediction improvement not only by the single stage grouping but also by successive groupings. Data grouping was effective both for identifying groups with similar contexts and for eliminating disparities in the dataset that impede accurate waste prediction. Among the grouping sequences tested, the grouping order resulted in the most improvement in flood waste prediction was UR, AR, and DO. This grouping order yielded enhanced waste prediction in 74 cases. However, grouping cannot explain every case because 16 cases were not fitted even after three-phase grouping. In addition, it is not clear whether the enhanced model fitness of some groups is associated with the screening effect of grouping or just a coincidence owing to the limitation of statistical analysis.

Conventional attempts to establish flood waste models used deterministic approaches; however, probabilistic methods have never been applied. Considering the large degrees of uncertainty in waste generation from floods, a probabilistic approach can provide a more accurate model compared to models developed by the conventional deterministic approach. One part of this study applied Bayesian inference to develop a flood waste regression model in South Korea. The aims of the study are as follows: (1) to analyze the characteristics of coefficients estimated by the Bayesian approach; (2) evaluate the performance of the prediction model by Bayesian inference; and (3) assess the effectiveness of Bayesian updating in a flood waste estimation. According to the results, the coefficients obtained via Bayesian inference showed a more significant p-value compared to those developed through the deterministic approach. Bayesian inference with a null prior distribution slightly reduced error compared to that by deterministic regression, specifically for post-event prediction. Bayesian updating did not effectively increase the accuracy of the model, while iterative updating required a complex calculation process. These results reveal the potential of the Bayesian approach in flood waste estimations, but also showed the error reduction by Bayesian approach is, in fact, limited.

Incorporating the interaction terms comprised of the products of two independent variables resulted in improved modeling with lower root mean square error and higher adjusted r². It seems interaction terms compensated over/underestimated contributions of independent variables and also explained combined effect of two variables in waste generation. The observation throughout the field surveying after typhoon Danas in 2019 revealed that damage in single aspect, such as flooded cropland did not always generate waste, and damage in one variables is sometimes linked to other variables, for example, one completely destructed building can physically affect the nearest building and lead to partial damage to the building. Incorporating interaction terms for flood waste modeling is simple approach not needing costly works and significantly enhances the model performance.

For the final task, the use of deep-neural network to estimate the flood waste generation was tried. Totally 22 kinds of parameters were utilized as input variables which were classified by flood damage variables, factors related to regional characteristics and meteorological parameters. Rectified Linear Unit (ReLU) was applied as an activation function in each node and the adjusted adaptive scaled gradient descend method was utilized as an optimizer. In order to optimize the performance, hyperparameter tunings were carried out in two phases: 1) grid search for discrete type hyperparameters and 2) Bayesian optimization of continuous type hyperparameters. As a results, the optimized hyper parameter set was (No. layer = 2, No. node = 83, Regularization number 1 = 10^-1.798, Regularization number 2 = 10^-4.113, Initial learning rate = 10^-4.696) and the testing set loss was 6,738,730 equivalent to 2,595.91 of RMSE. The RMSE in testing set by optimized deep-neural network was the most small among the tested four framework. Considering the empirical validation, the use of deep-neural network is the best viable option to measure flood waste generation. This may be related to the nature of flood waste generation; flood waste generation is actually not linear but has threshold in some extent which can be modeled by deep-neural network with ReLU activation function.

Keywods: disaster waste management, waste amount estimation, flood waste, regression analysis, statistical model