Trans. Nonferrous Met. Soc. China 24(2014) 2636-2641

Rapid assessment of flood loss based on neural network ensemble

Xiao-sheng LIU¹, Xiao HU¹, Ting-li WANG²

1. School of Architecture and Survey Engineering, Jiangxi University of Science and Technology, Ganzhou 341000, China;

2. School of Applied Science, Jiangxi University of Science and Technology, Ganzhou 341000, China

Received 10 July 2013; accepted 12 December 2013

Abstract: Considering the defects of low accuracy and slow speed existing in traditional flood loss assessment, firstly, the technical route of flood loss assessment was presented based on the neural network ensemble. Secondly, through the study of certain country of Poyang Lake district, the flood loss assessment indicators of the test area were analyzed and extracted by utilizing analytic hierarchy process (AHP), and the weights of the impact factors were assigned. Subsequently, the approaches to generate individuals and conclusions of neural network ensemble model were also investigated. In the platform of C# language and neural network library under AForge.NET open source, a flood loss assessment program which could rapidly build neural network ensemble models was developed. Finally, the proposed method was tested and verified. The comparison results between the assessment results of the proposed method and the actual statistical flood loss proved the feasibility of this method, thus a new approach for flood loss assessment was provided.

Key words: neural network ensemble; flood loss; rapid assessment; AForge.NET

1 Introduction

Generally, the traditional assessment of flood economic losses conforms to the monetary loss values of the flood disaster. Therefore, it is always analyzed and investigated from the aspects of hazard inducing factors, hazard environment and vulnerability of disaster bearing body, which involves a large number of influence factors, such as rainfall, hydrological, labor and socio-economic situation. Taking a certain region such as Poyang Lake district for example, its submerged extent, flood forecasting period and flooding time are the most important quantitative factors. At the same time, there are uncertain relationships in these factors. Therefore, flood disaster loss assessment is a kind of concrete numerical problem which is difficult to express by accurate formulas [1]. In other words, it means that the assessment of flood loss can be considered an issue of seeking out complicated nonlinear function approximation between flood loss and their impact factors. As the artificial neural network is complex network extensively connected by a large number of neurons [2], it possesses the properties of self-organizing, self-learning and associative memory, and it can be utilized to simulate some functions and structures of the human brain neural network with distributed, parallel, robustness and redundancy fault-tolerant features [3]. Moreover, it is very suitable for solving the practical problems difficult to express by mathematical formulas with complex influence factors. In addition, as a kind of rapid quantitative analysis model, it has been adopted in the assessment of flood loss in recent years [4,5].

However, due to the fact that the learning effect of neural network is related to the initial value, in a certain structure, the training results will always fall into the local optimal solution and can’t converge. In addition, as the artificial neural network belongs to the black-box learning, its generalization ability may not be reliably guaranteed, so sometimes the application effect can not be fully realized. For the limitations of the typical artificial neural network, there is an urgent need to develop a new neural network model. Correspondingly, neural network ensemble algorithm is rapidly growing in recent years. Through simple training on multiple neural networks and appropriate combination of the output results of these neural networks, more comprehensive and reliable judgment can be gotten [6], which is because it not only has a strong ability to process non-linear problems, but also can obviously enhance the generalization performance of artificial neural network. Therefore, it is regarded as a very effective engineering neural calculation method. In view of the situation stated above, the neural network ensemble was applied to the assessment of flood loss, and a new method for fast flood loss assessment based on the neural network ensemble was proposed.

2 Technical route of flood loss assessment based on neural network ensemble

Theoretically, flood loss assessment can be considered the problem of identifying the flood loss assessment indicators. Therefore, the neural network ensemble model was employed to identify and process the flood loss assessment indicators. By choosing the flood impact factors such as disaster-causing factor, flooding extent and socio-economic situation as the input neuron nodes of the neural network individuals, and fixing the direct economic loss of flood disaster as the output neuron node, a neural network ensemble model can be constructed by several neural network individuals. Afterwards, the historical flood data were divided into two parts: training sample data and test sample data. After the training sample data were inputted into the neural network individuals for obtaining the optimal neural network ensemble individuals, the optimal combination method was adopted to calculate the optimal nonnegative weight coefficients of these ensemble individuals in order to optimize the integration and obtain the optimal neural network ensemble model. Next, the test sample data were imported so as to rapidly and accurately acquire the output results. Ultimately, with the procedure of anti-normalization, the direct economic loss of the study flood area can be achieved and comprehended. The technique route of flood loss assessment based on neural network ensemble is shown in Fig. 1.

Fig. 1 Technique route of flood loss assessment based on neural network ensemble

The information in Fig. 1 is illustrated in detail as follows. Firstly, the flood assessment indexes reflecting the characteristic of the research region were extracted from the numerous impact factors of the sophisticated flooding system, and the weight of each index of flood loss assessment was calculated by taking advantage of the analytical hierarchy process (AHP). Secondly, according to the investigation, collection and collation of the related historical flood data such as terrain, elevation, remote sensing image and rainfall data of the flooding year, hydrological data, submerged depth and extent, affected population, local total production of industry and agriculture, and the direct economic loss of flood disaster, were looked upon as the original sample data and divided into training data and test data through normalization. Thirdly, through constructing the neural network ensemble model, the training data were imported into the individual neural networks which have been established for training in advance. Finally, after the test data were inputted into the trained neural network ensemble model for calculation, the output conclusions of each individual network were combined, and with anti-normalization operation, the direct economic loss of flood disaster can be obtained.

3 Indicators of flood loss assessment and weight distribution

In order to make the flood loss assessment indicators more reasonable and comprehensive, exhaustive considerations were given to the various characteristics such as population, economics and society of the study area (Poyang Lake district) when the indicators were determined. As a result, disaster-causing factors (rainfall, flooding depth), flood control capacity (water storage capacity, affected area), socio-economic factors (affected population, total production of industry and agriculture) were selected as the indicators for reflecting the direct economic losses caused by floods.

Herein, the analytic hierarchy process [7] was adopted to describe the flood losses on human society. The definite calculation steps are elaborated as follows.

Firstly, a hierarchical model was constructed. Based on the data from the collected flood loss statistics table, the flood loss assessment indicator system with three-layer structure was established. The constructed flood loss assessment indicator hierarchy model is shown in Fig. 2.

Subsequently, it needs to construct comparative judgment matrixes in pairs and respectively calculate the weight vectors and the combination weight vectors. First of all, according to the principle of pairwise comparisons and the relationships of all elements in order of importance, comparative judgment matrixes can be constructed. Based on the hierarchical model, five judgment comparative judgment matrixes were established. Then, the single-level weight of each assessment indicator can be achieved on the basis of the solutions of these matrixes by the utilization of AHP. At the same time, consistency check was implemented. Eventually, the weights of the indicators of flood loss assessment in the research area were acquired, which are listed in Table 1.

Table 1 Weight of flood loss assessment indexes of study area

4 Neural network ensemble and rapid implementation

With superiorities of simple manipulation and prominent effect, it is easy to grasp and apply, even for the general engineering and technical personnel lacking of relevant experience in the use of neural network. The investigation about neural network ensemble cannot only promote the research of neural computation and statistical theory, but also stimulate the application of neural computation in engineering [8,9]. The structure of neural network ensemble is demonstrated in Fig. 3.

4.1 Neural network ensemble

At present, the research of neural network ensemble implementation mainly focuses on two aspects: the generation of neural network individual and the integration of output conclusion [10]. Generally, the typical neural network ensemble structure is composed of the following two parts.

The first part is how to generate the integrated individual neural networks which should be accurate as well as different in a certain degree. Currently, there are mainly three types of methods in generating individual neural networks.

Fig. 2 Flood loss assessment indicator hierarchy model

From the aspect of training set selection, the most important technology was Bagging and Boosting [11]. Bagging was initially proposed by BREIMAN. Its basic principle is to put and return the sample of n times in random from the n-size data set. Taking out a sample in random for n times constitutes a n-size training sets. Simultaneously, Boosting technology is the general name of a broad class of algorithms which was primarily put forward by SCHAPIRE and improved by FREUND afterwards. The fundamental idea is to make the second classifier pay more attention to the classification error of the first classifier during the process of training a series of component classifier, and combine voting results of each classifier to get the final classifier.

According to the types of individual neural networks, they can be mainly divided into BP, SOM, Hopfiled. The smaller the related degree of each individual network is, the lower the generalization error exists. As a result, the neural network ensemble can be constituted by different strategies of choosing individual network with high difference degree.

The third approach is to consider the structure and training algorithm of the individual neural network, which means building the neural network ensemble through changing the individual neural network structure (such as the layer of network, hidden layer neuron number, threshold, impulse value, learning rate) and transforming the individual network training algorithm.

The second part is how to synthesize the individual neural network output conclusions into the integrated ones. It can be determined according to the analysis of the specific practical circumstance.

When neural network ensemble aims at settling the classification issue, the output conclusions of the individual neural network are usually generated from relative or absolute majority vote. Relative majority vote refers to the method that if a certain kind gets the most votes after carrying on statistics about the classification results, it will be the integrated final output conclusion, while absolute majority vote means that only if more than half of the neural network individual output conclusions are the same, this kind could be the integrated final output conclusion.

When neural network ensemble is takes advantage of regression, the integrated output conclusion can be generated from the simple or weighed average of the individual neural network output conclusions. In simple average method, the weight of each individual neural network that involves in integration is equal. On the contrary, the weighed average method means that the weight of each individual neural network is assigned according to their performance.

4.2 Rapid construction of neural network ensemble

With the assistance of C# language of Microsoft .NET [12] and the open source framework AForge.NET [13], a program which could quickly realize neural network ensemble functions was developed. This program can simply and quickly realize the construction and training of the neural network ensemble model, as well as output an intuitive graphical result, thus it greatly reduced the difficulty for normal users to apply the neural network ensemble model.

According to the definition of neural network ensemble, in order to achieve neural network ensemble, first of all, the training sets should be employed to produce a number of individual neural networks. Usually, there are two algorithms Bagging and Boosting which are available for generating the training data sets. The generation steps of the individual neural network ensemble are illustrated as follows.

Fig. 3 Schematic diagram of neural network ensemble structure

The training data set was generated. Firstly, C# language was utilized to rewrite Bagging and Boosting in order to obtain the training sets.

The individual neural network ensemble model built and trained. Firstly, it needs to open VisualStudio2008 software and add AForge.Controls.dll, AForge.dll and AForge.Neuro.dll these three dynamic link libraries in its solution manager in order to use the classes and interfaces of AForge.Neuro to construct and train the individual neural network model. Secondly, in accordance with the actual requirements, the activation network or distance network class is adopted to instantiate a number of individual neural network models. Subsequently, it needs to choose a suitable learning algorithm for each model from the classes of BP Learning, Deltarule Learning, Perceptron Learning, SOM Learning or Elastic Network Learning, and set the parameters such as relevant learning rate and impulse value. Finally, by using the training data sets generated by Bagging or Boosting algorithm, each individual neural network model can be trained, and at the same time, the error curves in the training process can be drawn onto the Chart control provided by AForge.NET framework so as to view the data conveniently. As training the neural network may cost a longer period of time, therefore, the multi-threading technology was utilized, which means opening up a sub-thread to specially train neural network. When all of the errors of individual neural networks are less than the preset error or the number of training, the training process will be stopped.

According to the definition of neural network ensemble, it is necessary to integrate all the conclusions of the individual neural networks before the conclusion of neural network ensemble is generated. The concrete steps are demonstrated as follows.

Generating the conclusions of individual neural networks. When the training of each individual neural network model was finished, the corresponding test data sets should be imputed for computation. Then, their conclusions and error values should be outputted and recorded and the results should be drawn onto the Chart control provided by AForge.NET framework at the same time for visual effects.

The conclusion of neural network ensemble was generated. When the conclusions of each individual neural network were gotten, above all, the function of this neural network ensemble should be judged. If it is for regression estimation, the conclusion should be yielded by simple average. While for classification, the relative majority vote method may be superior. The integrated results were also presented onto the Chart control for intuitional view.

5 Application and conclusion of the assessment method

The proposed method in this work for rapid flood loss assessment was applied in one county of the Poyang Lake district. The statistic data of flood loss assessment indexes of the county in 2005-2010 are listed in Table 2.

The rapid neural network ensemble construction program compiled in this work can be run with these statistical data, and the specific steps and results are as follows.

The ensemble is formed by five BP neural networks and the Bagging algorithm is adopted for individual generation. These parameters are imputed into the program to complete the setup.

As five BP networks should be produced, the neurons number of input and output layer is 6 and 1, respectively, while the neural numbers of hidden layer were set as 6, 8, 10, 12 and 14. These parameters are inputted into the rapid neural network ensemble construction program to complete the set.

After dividing the flood disaster data of the county from 2005 to 2010 into the training data and testing data, they are stored in text files respectively. With the operation of opening the program to read the file, clicking the menu and choosing normalization function, the normalization and weights assignment is completed.

After clicking the training operation in the program menu, the normalized training data are started to be read. Then, the generated five BP networks are trained according to the preset parameters until the training complete tips are emerged.

Table 2 Statistic data of flood loss assessment indexes of research county in 2005-2010

With clicking the testing operation in the program menu, it is started to read the normalized data. Next, conclusions are outputted respectively from the constructed five BP networks. Then, by clicking the output ensemble results menu, the final ensemble output conclusion data can be obtained.

Finally, by the operation of the anti-normalization operation in menu, the direct flood economic loss value of the research county in 2010 is 63.9813×10⁶ RMB.

The above result shows that the direct economic loss of this research county from the flood loss rapid assessment model based on neural network ensemble in 2010 is almost the same with the real values (66.0009×10⁶ RMB) in 2010, which demonstrates that the flood loss rapid assessment model based on neural network ensemble has certain reliability as well as work efficiency. Therefore, it can be promoted to the practical application, thus providing a new approach for flood loss assessment.

References

[1] WANG Bao-hua, FU Qiang, FENG Yan, YANG Na. Mixed fuzzy nerve net model for fast evaluation on economic loss of flood [J]. Journal of Northeast Agricultural University, 2008(6): 47-51. (in Chinese)

[2] YANG Jian-gang. Practical course of the artificial neural network [M]. Hangzhou: Zhejiang University Press, 2001: 10-68. (in Chinese)

[3] LE Xiao-rong. Algorithm design and analysis on neural network ensemble [D]. Yangzhou: Yangzhou University, 2007: 6-53. (in Chinese)

[4] LIU Xiao-sheng, YU Hao-feng. Research on assessment model for flood losses based on GIS and BP neural network [J]. Engineering Investigation, 2009(4): 72-74. (in Chinese)

[5] HUANG Tao-zhen, WANG Xiao-dong. Fast evaluation of flood and water-logging losses by BP network [J]. Journal of Hohai University, 2003(4): 457-460. (in Chinese)

[6] LIU He-xiu. Research on neural network ensemble algorithm [D]. Ocean University of China, 2009: 23-34. (in Chinese)

[7] ZHANG Guo-lin. Quantitative evaluation of the expo 2010 on the basis of analytic hierarchy process [J]. Journal of Yichun College, 2011(33): 41-42. (in Chinese)

[8] LI Yan. Research of neural network ensemble and its application to classification and regression [D]. Beijing: North China Electric Power University, 2009: 33-43. (in Chinese)

[9] ZHOU Qi-ming, WANG Miao, REN Hong. The research of application of constructional engineering cost estimation by neural network ensemble [J]. Journal of Chongqing Jiaotong College, 2005(4): 129-132. (in Chinese)

[10] ZHAO Sheng-ying, GAO Guang-chun. Ant conlony optimization- based approach for selective neural network ensemble [J]. Journal of Zhejiang University, 2009(9): 1568-1573. (in Chinese)

[11] ZHOU Zhi-hua, CHEN Shi-fu. Neural network ensemble [J]. Chinese Journal of Computers, 2002, 25(1): 1-7. (in Chinese)

[12] MARK SCHMIDT, SIMON ROBINSON. Microsoft visual C#.NET 2003 development technology books [M]. Beijing: China Water Power Press, 2005: 33-53. (in Chinese)

[13] GONG Cheng-ying, HE Hui. Video-based vehicle motion detection method with AForge.NET [J]. Computer Knowledge and Technology, 2011, 7(1): 187-188. (in Chinese).

基于神经网络集成的洪灾损失快速评估

刘小生¹，胡 啸¹，王婷丽²

1. 江西理工大学 建筑与测绘工程学院，赣州 341000；

2. 江西理工大学 应用科学学院，赣州 341000

摘  要：针对传统洪灾损失评估存在精度低、速度慢等问题，首先提出基于神经网络集成的洪灾损失评估方法技术路线；其次，以鄱阳湖区某县为研究对象，运用层次分析法对影响该研究区域的洪灾损失评估指标进行分析和提取，并对影响因子的权重进行分配；然后，研究神经网络集成模型的个体生成和结论生成的实现方法，并利用C#编程语言和AForge.NET开源框架下的神经网络类库搭建一个能快速构建神经网络集成模型的程序；最后，对该方法进行应用，并将评估结果与实际统计的洪灾损失值进行对比分析，验证该评估方法的可行性，从而为洪灾损失评估提供一种新的方法。

关键词：神经网络集成；洪灾损失；快速评估；AForge.NET

(Edited by Chao WANG)

Foundation item: Project (41061041) supported by the National Natural Science Foundation of China; Project (2010gzs0084) supported by the Natural Science Foundation of Jiangxi Province, China

Corresponding author: Xiao-sheng LIU; Tel: +86-797-8312550; E-mail: lxs9103@163.com

DOI: 10.1016/S1003-6326(14)63393-8