17:15 〜 18:30
[HTT14-P03] 頻度比、ロジスティック回帰、人工ニューラルネットワーク、GISを用いた中国深圳の斜面崩壊とシンクホールの発生要因分析
キーワード：斜面崩壊、シンクホール、発生要因分析、頻度比、ロジスティック回帰、人工ニューラルネットワーク
As major geologic hazards, landslides and sinkhole formation account for a great number of human casualties, an enormous amount of property loss, and significant damage to natural ecosystems and human-built infrastructures, especially in urban areas. Therefore, risk evaluation and reduction of such hazards are massive challenges for researchers and engineers to prevent disasters. Although many previous studies have attempted to conduct susceptibility maps to portray hazards’ spatial distribution, only a small number of them focused on both landslides and sinkholes in an urbanized area.
This study is to assess and compare the performances of three typical models for hazard susceptibility mapping, namely the frequency ratio model, the logistic regression model, and the artificial neural networks model for the study area in Shenzhen, south China. To focus on potential damage in populated areas, a typical area with high susceptibility values and high population density was chosen as the main target area for analyzing the susceptibility mapping results. For the artificial neural networks model, the radial basis function (RBF) algorithm was applied to construct the networks. The conditioning factors such as elevation, slope inclination, slope aspect, lithology, land cover, and NDVI were derived from various data sources to generate the geospatial database. Precipitation was chosen as the triggering factor of hazards. According to the available hazard inventory, 268 landslide events and 172 sinkhole events were identified and recorded. In addition, 400 points without geological hazards were randomly chosen to provide the absence data for susceptibility modeling.
Subsequently, for the evaluation and comparison of susceptibility models and resultant maps, the confusion matrix and the receiver operating characteristics (ROC) curve with the area under the curve (AUC) were used to evaluate the model performance. The results of landslide susceptibility analysis using land conditions indicate that the values of overall accuracy for the frequency ratio model, the logistic regression model, and the RBF neural networks model were 0.764, 0.803, and 0.810, respectively, and the values of AUC were 0.787, 0.815, and 0.850. In sinkhole susceptibility analysis, the values of overall accuracy for these three models were 0.648, 0.672, and 0.735, respectively, and the values of AUC were 0.707, 0.745, and 0.757. To discuss the probability of landslide or sinkhole occurrence, it is necessary to combine triggering factors like rainfall with land susceptibility. Logistic regression was employed to determine the relative weights of rainfall and land susceptibility, and the modified susceptibility maps considering the effect of precipitation were produced.
The results of this study demonstrate that the RBF neural networks model gave the best results in both landslide susceptibility analysis and sinkhole susceptibility analysis in terms of the values of AUC and the overall accuracy. Moreover, the RBF usually performed better than other commonly used artificial neural network algorithms, especially BP (back propagation). This study also has some advantages over previous studies conducted in the same city of Shenzhen.
This study is to assess and compare the performances of three typical models for hazard susceptibility mapping, namely the frequency ratio model, the logistic regression model, and the artificial neural networks model for the study area in Shenzhen, south China. To focus on potential damage in populated areas, a typical area with high susceptibility values and high population density was chosen as the main target area for analyzing the susceptibility mapping results. For the artificial neural networks model, the radial basis function (RBF) algorithm was applied to construct the networks. The conditioning factors such as elevation, slope inclination, slope aspect, lithology, land cover, and NDVI were derived from various data sources to generate the geospatial database. Precipitation was chosen as the triggering factor of hazards. According to the available hazard inventory, 268 landslide events and 172 sinkhole events were identified and recorded. In addition, 400 points without geological hazards were randomly chosen to provide the absence data for susceptibility modeling.
Subsequently, for the evaluation and comparison of susceptibility models and resultant maps, the confusion matrix and the receiver operating characteristics (ROC) curve with the area under the curve (AUC) were used to evaluate the model performance. The results of landslide susceptibility analysis using land conditions indicate that the values of overall accuracy for the frequency ratio model, the logistic regression model, and the RBF neural networks model were 0.764, 0.803, and 0.810, respectively, and the values of AUC were 0.787, 0.815, and 0.850. In sinkhole susceptibility analysis, the values of overall accuracy for these three models were 0.648, 0.672, and 0.735, respectively, and the values of AUC were 0.707, 0.745, and 0.757. To discuss the probability of landslide or sinkhole occurrence, it is necessary to combine triggering factors like rainfall with land susceptibility. Logistic regression was employed to determine the relative weights of rainfall and land susceptibility, and the modified susceptibility maps considering the effect of precipitation were produced.
The results of this study demonstrate that the RBF neural networks model gave the best results in both landslide susceptibility analysis and sinkhole susceptibility analysis in terms of the values of AUC and the overall accuracy. Moreover, the RBF usually performed better than other commonly used artificial neural network algorithms, especially BP (back propagation). This study also has some advantages over previous studies conducted in the same city of Shenzhen.