Prediction of Flood Occurrences using Rough Sets Theory (Case Study: Halilrood River)

Flooding, which is an overflow of water, brings massive biological and economic problems all around the world and disrupts many people’s lives. Based on several effective factors in flood occurrence as well as unknown relationships between parameters, researchers have used various methods to forecast flood over the past decade. An investigation of the conducted studies on flood forecasting reveal that most studies have used measured data such as river flow, temperature or moisture for forecasting. These data are collected over several years and they are used in the above methods. Indeed, the point common to all studies is using data model for model training and flood forecasting. However, as issue that has received less attention is the presence of uncertainty in the data used for extracting flood occurrence model. Taking uncertainty into account, due to high data volume used for model extraction, improves the results.
 The current research aims to investigate the relationship between flood occurrence and effective parameters by selecting rough sets theory as well as taking into account the uncertainty present in the data during the forecasting process; moreover, following the extraction of these relationships, some rules are extracted which, in addition to their simplicity, present the simultaneous effect of effective parameters on flood occurrence. Then, using the existent relationships in this theory, the correlation between the parameters is investigated in multiple form and the most efficient rules for identifying the most probable conditions of flood occurrence are obtained.
For this purpose, the above parameters, spanning four years (2003-2007), related to Jiroft dam is used for analysis and extraction of rules. In this way, first, data processing is performed and intervals of flood occurrence are separated from each year. Then, data discrete manufacturing is performed followed by data approximation and data reduction; the most probable rules are then extracted for flood occurrence by identifying effective core features. The region under study for the current research is Halilrood river located in southeast of Iran, Kerman province.
 
Materials and Methods
Nowadays, flood occurrence is one of the major issues associated with natural disaster management. Accordingly, the current research proposes a method for prediction of flood occurrence on a daily basis by using rough sets theory to both manage its occurrence risk and analyze the uncertainty obtained from the used data.
Given that rules extraction using rough sets theory is done only through previous data analysis, data selection has high sensitivity. The parameters used in this study include precipitation amount, minimum daily temperature, evaporation and the recorded river flow. Application of rough sets theory and simultaneous model extraction of effective parameters in flood occurrence are among the main objectives of the current research. Among the data of four years (2003-7), monthly data, which are obtained from hydrometer devices installed at the entrance of Jiroft Dam for the time period of 2003-7, are selected for analysis. In this way, first, data processing is performed and intervals of flood occurrence are separated from each year. Then, data discrete manufacturing is performed followed by data approximation and data reduction; the most probable rules are then extracted for flood occurrence by identifying effective core features. Finally, the data related to 2009 is used to evaluate the power of the extracted rules;.
Rough sets theory, proposed in 1980 by Pawlak, is a powerful mathematical tool for dealing with uncertainty and ambiguity of data. It relies on analysis of data Tables. These Tables may be obtained from measurements or by experts. The approximation synthesis of concepts from the acquired data is the main objective of the rough set analysis. It also provides some methods for reducing imprecise or redundant data in data bases. The process of eliminating redundant data is performed based on training without losing basic data of data base. As a result of data reduction, a set of tabloid and meaningful rules is extracted, which makes the process of decision-making easier.
 
Discussion and Results
In the first stage which includes data selection, all data of the years 2003-7 were annually and completely selected for processing. In the next stage for data discrete manufacturing, since it is impossible to put all data in an information system Table due to different average evaporation and minimum temperature for different seasons of the year, therefore, the data were selected based on different seasons and then they were prepared as four tables of information system. After data selection, data discrete manufacturing was performed by taking their average into account. In the next stage, in order to evaluate the data, Lower and Upper Approximations of each information system were individually calculated. Given that the current research aims to extract flood occurrence rules, only the information systems can be selected for the next stage so that they can have maximum approximation accuracy for Class 1 of decision.
The information systems of spring and winter were first mixed in this research for Reduction of Attributes which created an information system. Then, Reduction of Attributes for this new information system was performed by using the discussed relations in rough sets theory, following which no attribute was eliminated, the reason of which may be attributed to a variety of situations as well as high volume of data in this system. Afterwards, in order to obtain simple decision rules, Reduction of Attributes was performed by dividing the new information system into seasons as well as forming individual information systems for each month of spring.
Finally, the hydrology data of 2007 was used to evaluate the obtained rules. This data evaluation shows that there are 29 river flow occurrences with more than 10 m³/s. The results obtained from data discrete manufacturing associated with flood occurrence of 2007 show 21 cases out of 29 floods based on the extracted rules for Class 1. The results show that 72% of the extracted rules are consistent with the cases of flood occurrence which shows the ability of these rules in identifying the probable cases of flood occurrence. However, it should be mentioned that the powerful rules extracted by rough sets theory are based on lower approximation of Class 1; therefore, the elements forming the border or uncertainty are not used in these rules; however, border elements and upper approximation of Class 1 are used only for approximation rules. Using the values obtained from confusion matrix, Kappa value was calculated to be 0.84. In addition, an overall accuracy rate of 95% was obtained for the research results regarding the fact that this parameter also involves forecasting of flood non-occurrence. Moreover, based on the frequency, the overall accuracy has increased for more cases of flood non-occurrence than occurrence. Therefore, the uncertainties present in the data are identified and then eliminated from the process of rules extraction. This issue was one objective of the current research which was obtained using rough sets theory.
 
Conclusions
Flood occurrence is one natural disaster which is a serious threat to social infrastructures and financial compensation of damage due to floods is impossible.  Given that this phenomenon depends on several factors, many researches have been performed to forecast flood over the recent years. In many cases, those researches employ flood occurrence model for forecasting by taking the previous data into account. One considerable point common to those researches is that they do not consider the uncertainty present in high volume of the data used for identifying flood occurrence model. Moreover, some of these methods individually investigate the correlation between effective parameters and flood occurrence; then, the effect coefficients were calculated for them; it is necessary to compare all parameters with each other as well as to investigate them for forecasting due to complexity of natural phenomena and interdependency and simultaneous impact of various factors.
However, the current research, taking into account data uncertainties and their elimination from the process, aims to extract the rules and investigate the correlation between parameters and flood occurrence simultaneously by using rough sets theory. The results of the current research show that 72% of flood occurrences recorded in 2007 are consistent with the obtained rules, which indicate the ability of rough sets theory in extracting forecasting rules of natural phenomena occurrence which have the highest complexity. The results of this research can be utilized in crisis management planning and natural disasters control.