A Study On Modeling Of Flood Risk Evaluation Based On Bayesian Networks

Flood as a common natural event has vital impacts on natural environment and human society. It is an import water source to survive and flourish riparian plants and animals, reduce the negative impacts of drought and solve the shortage issues of water resources. However, flood also has negative influences on human society, such as resulting in casualties and socio-economic losses. Therefore, the investigations of flood disaster have increasingly attracted the extensive attentions of governments and human society worldwide. The study on flood risk is an important part of forecast, prevention and response of the negative impacts of flood disaster. It consists of numerous research areas, such as the inundation extent of flood disaster, the probability or likelihood of flood occurrence and the forecast of flood disaster. The prediction of flood disaster is a likelihood inference of the flood occurrence in future. Based on this, different risk evaluation models on flood were presented and applied in flood risk. Additionally, remote sensing (RS) and geographic information system (GIS) have widely provided spatial data and spatial-processing techniques to the study on flood risk under the development of information technologies. But, there are some urgent problems to be solved and explored further in flood disaster studies.1) For a regional scale, the flood-related study not only explores whether the area will be inundated, but also analysis the probability or likelihood of flood occurrence in future.2) The traditional evaluations on flood mainly depend on observation data of high temporal resolutions, such as water flow speed, capacity and precipitation density during a flood event. However, the traditional methods will be paralyzed when the observation data is short which cannot meet investigation requirements. How to infer the likelihood of flood occurrence in future is a crucial issues to be solved in flood investigations.3) Current surveys in flood risk assessments still have absence of uncertainty investigation for flood factors. The uncertainties consist of the relationship between flood factors, and the causality uncertainties among the factors which have impacts on the occurrence and change of flood disaster.4) Current Bayesian Networks (BNs) based risk evaluation models only constructed the ordinary point-based BNs for whole study area. They did not integrate BNs into spatial technologies, such as RS and GIS. The spatial character of BNs is not obvious.On the basis of problems aforementioned, the study presents an integrated spatial framework by combination of hierarchy Bayesian networks (HBNs), RS and GIS. The framework consists of three modes to process inputs and outputs to and from the BNs-based flood risk evaluation models. In the data processing and indices derivation module, the input indices are standardized into grid datasets. In the model development and implementation module, maximum iterations were used to select the most accurate sampling table for generating conditional proability tables (CPTs) which are employed for probability inference cell by cell in a study area. The mapping and evaluation of likelihood of flood risk module maps likelihood of flood risk and evaluates flood risk distribution over a study area. The studies analyzed, constructed and applied the naive Bayes (NB), weighted naive Bayes (WNB) and HBNs into a Bowen Basin in central Queensland, Australia, separately. The study extracted the maximum inundation extent (MIE) from the historical Moderate Resolution Imaging Spectroradiometer (MODIS) imagery by using the Open Water Likelihood (OWL) algorithm. The MIE was used as a base map to generate the training table and to verify the resultant evaluations. Comparing three methods, the study shows the evaluation accuracy constantly increase with the complex of BNs, but the running efficiency continually decreases. The paper explored the sensitivity analysis of three models between model inputs and outputs as well.The main jobs and outputs of the study include the following parts.1) The model inputs:the study extracted the nine flood-related indices, including extreme rainfall, evapotranspiration, net-water index, soil water retention, elevation, slope, drainage proximity and density. They were generated from spatial environmental data representing climate, soil, vegetation, hydrology and topography. These indices normally affect the occurrence and change of the flood. The study standardized the indices into four levels from high risk to very low risk in terms of domain knowledge.2) The flood historical maximum inundation extent (MIE):The study extracted the MIE from MODIS imagery by using OWL algorithm. The MIE as a base map was used to generate the training table and to verify the resultant evaluations.3) The construction of HBNs:the study introduce and construct three bayes models, NB, WNB and HBNs. The models were employed into the two study areas, separately. The NB was used into a small study area with the five indices without the climate factors that are precipitation and evapotranspiration. The paper evaluated the accuracy and indices contributions to resultant evaluations. According to the limitations of NB, the study modified NB method to WNB method using entropy algrithom to calculate the weights for nine input indices. Comparing WNB with NB, the investigations found the WNB has a higher evaluation results than that of NB. On the basis of the NB and WNB methods, the study constructed an optimal HBNs according to domain knowledge. The HBNs is the complex model, but it can reflect the relationships among indices in flood risk assessments. In some trial areas, HBNs has a higher accuracy than that of WNB via comparsion between two the models.4) Model sensitivity analysis:the investigation explored the relationship between model inputs and outputs. The analysis reflects the maximum sampling times, the sampling ration of inundation and non-inundation and the size of a study area have impacts on the outputs of models.The main innovative ideas of this research in the field of flood risk assessment can be summarized as follows:1) This research combines Bayes’theorem with Remote Sensing (RS) and Geographical Information Science (GIS) technology, which improves the accuracy of naive Bayes (NB). By contrast, NB based on RS technology has higher evaluation accuracy than traditional NB.2) This paper adopts entropy method to evaluate contribute of different input indicators in flood risk assessment research. On the basis of that, Weighted Bayes theorem is raised to develop NB’s by correcting the imperfections in the independence hypothesis. In contrast with NB, Weighted Bayes theorem has a great improvement in accuracy. Entropy method is proved to be a scientific and logical way in upgrading NB theorem.3) This paper builds a multi-layer Bayes network model on the foundation of expert knowledge. The model clarifies mutual relationship between different indicators by constructing a topological network. On top of that, probability distribution of different indicators is calculated to indicate contributes in flood risk. Compared with Weighted Bayes theorem, Multi-layer Bayes Network shows high accuracy and a quick indicating ability in pointing out causal relationships between indicator, occurrence regularity, and development rule.4) This paper makes an attempt to assess flood risk using Bayesian Networks (BNs), and acquires a reasonable evaluation result. BNs can not only point out causal relationship between different indicators in flood risk research, but also express contributes of them in inducing and developing flood in the form of Conditional Probability Tables (CPTs). Network structure and CPTs better represent uncertainties in flood risk than traditional flood risk modeling approaches such as Analytic Hierarchy Process (AHP). BNs prove to be highly objective.5) Different from traditional Bayes method, this paper raises spatial BNs by integrating RS and GIS technology and Multi-layer BN model. Based on unit grid instead of the whole study area, possibility of flood occurrence in the future is predicted. Grid-based Bayes’Network model requires a much larger amount of calculation compared with traditional method. This paper makes rewarding attempts in linking spatial technology with BNs.