The Study Of Regional Landslide Susceptibility Prediction Based On Rough Sets-Neural Network

The loss caused by landslide which is the most common in three major type geological disasters is huge. Recent years the number of geological disasters and the losses decreased, but there are still more than 10,000 every year, so regional landslide prediction has significance which can provide the basic forecasting information for the government and reduce the losses and casualties, to achieve the purpose of disaster prevention and reduction.Currently great progress had been made in the field of regional landslide susceptibility prediction, especially the prediction model. Many models had been built combining with mathematical methods. But there were some limitations such as could not descript each factor importantly. Therefore, the combination of various factors had been considered. In this way the predictors may be redundant and the work of forecasting would increase. It's not conducive to rapid and efficient forecasts. So the model which can descript the importance of each factor and avoid the redundant factors has been built in this paper.Because of the prediction mainly consider the combination of many factors, the research considered as a decision-making system. As the factors were usually inaccuracy, Rough Set (RS) theory can be used during data processing. The importance of each factor can be calculated by RS then reduced the redundant factor. The model which will reduce the involved factors and improve efficiency has been built through Neural Network to simulate the reduction data. Using RS to select predictors is more reasonable, and the result is more interpretable.Based on the project of "Research of the formative mechanism and evaluation of landslide in Enshi", geohazard had been investigated in Hubei Province. The relationship between various factors and landslide had been analyzed in this paper. Lithology, slope, slope structure, slope height, land using structure, road excavation, the affected areas and groundwater valley type had been selected as indicators of landslide prediction system in the study area. Sensitivity coefficient has been used for the state division of linear variable, the statistics method for the continuous and discrete variable. The high precision research area has been selected as sample area. In data processing, first delete the factor according to the importance from small to large, if the data system is coordination then delete the factor, otherwise keep it. The data system simulated by Neural Network, enter the forecast data into the model then the results will output. The study area has been divided into 25 500 by regular mesh grid (200m×200m) in this paper. Based on GIS's data detection capability, spatial analysis capabilities and graphics functions, the landslide degree subarea has been obtained. There are mainly results as follows:(1) The RS-Neural Network has been adopted based on the complexity of landslide susceptibility prediction. The model which exploit the ability of dealing with uncertain data by RS and the simulation and tolerance of training data by Neural Network optimizes the original predict model.(2) RS theory is applied to landslide hazard prediction process which provides a new method to select the effective prediction index system and avoid repetition and redundancy in the factors. According to the importance of each predictor, the relationship of landslide and predictor can be determined. The most favorable combination of predictors and the most compact predict index system will be used in practical work which will reduce the data processing and conducive to landslide hazard forecasting.(3) Landslide was 78.8 percent in the history geological hazard of Enshi city. For effective realization of disaster prevention and mitigation, predict the landslide is needed in the area. The model was applied to predict the landslide disaster in the area of 1020km2, the prediction result was closely agreement with the reality with higher accuracy. The study demonstrated that the model can be used for landslide disaster forecasting in some similar areas under the same scale.(4) The state of various factors classified according to the statistics and sensitive coefficient method which makes the division of state more reasonable. Land use type and slope height had been excluded according to the principle of coordination decision making system. The reduction data has been simulated by Neural Network, which can reduce the amount of training data and improve efficiency. The result is consistent with actual situation when using the model in the target area. The landslide is mainly affected by lithology, slope and slope structure which is indicated by the statistics of predicted results. This combination of factors can be conducted forecasting in field investigations.