GEOTECHNICAL HAZARD ASSESSMENT OF SOIL SLOPES IN HILLY REGION

Abstract

Stability assessment of slopes is a very important aspect of geotechnical engineering as slope instabilities in hilly regions in the form of landslides often pose serious threat to build environment. Conventionally, slope stability analyses are carried out using Limit Equilibrium Methods (LEM) (e.g. Ordinary method, Bishop’s Modified Method, Morgenstern and Price’s Method, etc). However, due to time consuming nature of these methods, the use of LEMs is not viable when there are very large number of slopes to analyze in a limited time, particularly for the rapid detection of post-earthquake landslide activity, and landslide early warning. In these cases, Artificial Neural Network (ANN) & Multiple Regression (MR) methods appear to be promising for the rapid assessment of stability of large number of slopes in hills. Nowadays, the application of ANN has been increasingly used to various problems in geotechnical engineering including slope stability assessment. However, the performance of ANN in the past studies on slope stability prediction is found to be poor, while the prediction of relative importance of various slope contributing factors is not reliable. This is primarily due to the use of limited number of real field data cases and/or synthetic data covering limited parametric variations, in the training process. While ANN is a soft computing tool, MR is a statistical technique, which provides a simplified equation that can be used to calculate critical factor of safety of slopes without adopting any iterative procedure, thereby reducing the time and complexity involved in the evaluation of slope stability. However, the MR models in the past studies have shown poor performance, probably due to the similar reasons as mentioned in the case of ANN. Hence in this study, an attempt has been made to study various issues associated with the use of these two methods (i.e. ANN and MR) for the prediction of slope stability by considering different types of soil slopes covering all possible slope configurations and soil characteristics under dry, wet and seismic conditions. Further, the Relative Importance (RI) of various factors contributing to stability of slopes have been obtained from both ANN and MR models. RI factors obtained here can have several applications like: assigning appropriate weights to various causative factors in landslide hazard zonation, sensitivity analysis for identifying the most influential parameters, etc. Performances of both the developed ANN and MR models have been evaluated and compared. iv A three layered feed forward ANN trained using back propagation algorithm has been considered in this study. ANN considered basically consists of input, hidden and output layers. The input layer consists of number of neurons representing various parameters. Parameters of slope geometry and soil properties such as unit weight (γ), cohesion (c), angle of internal friction (φ), slope angle (β), height (H) and pore water pressure ratio (ru) are inputs to the ANN model, while Factor of Safety (FoS) is its output. The output layer consists of a single neuron representing the FoS. Initially, to study the dependency of ANN’s accuracy on the quantity and quality of training data, two separate ANN models were trained using limited field data as well as extensive synthetic data . The ANN which was trained using extensive synthetic data has shown improved performance in the stability prediction. The extensive synthetic training data set consisting of about 46 thousand cases, covering a wide range of soil characteristics and geometry, was prepared using Michalowski stability chart method. Further to improve the accuracy of ANN and to obtain reliable estimates of importance factors, separate ANN models have been developed for cohesive, cohesionless and c- ϕ soils under dry, saturated & seismic conditions and their accuracy has been tested by applying untrained as well as real field slopes. The average mean square error of the FoS predicated using ANN trained with extensive dataset is significantly reduced compared to ANN trained with limited field data, which clearly shows the improvement in the stability prediction of slopes using present ANN over the past studies. Thus, it highlights the importance of ANN training with good quality and quantity of data, for the accurate prediction of stability of slopes using ANN. A method developed by Garson has been used here to obtain the relative importance factors. Results from the study on importance factors, showed that angle of slope (β) and angle of internal friction (φ) are the most significant parameters contributing to FoS in case of cohesionless dry soils, while height and cohesion are found to be the most influencing factors in case of cohesive soils. In case of seismic case, the seismic coefficient is found to be significantly influencing the stability condition. Its relative influence on slope stability is found to be 18%. In addition to ANN, multiple linear regression as well as multiple nonlinear regression models have been developed for the calculation of critical FoS using extensive data set, covering all possible slope configurations and soil characteristics. The same extensive data set that was used for the v development of ANN model, has been used here. Separate multiple regression models have been developed for cohesionless, cohesive and mixed soil slopes under dry, wet & seismic conditions. Then, the developed models have been validated by applying to unlearned slope data. Similar to ANN model, the accuracy of these developed models has been compared with respect to conventional limit equilibrium methods in terms of Mean Square Error (MSE) & regression coefficient (R2). Apart from slope stability prediction, Relative Importance (RI) of various parameters have also been obtained for all the soil types under both dry and saturated conditions. MSE of Multiple Linear Regression (MLR) models is considerably higher than MSE observed between limit equilibrium methods. R2 lower than 0.85 is observed in most of the cases analyzed. Probable reason for the observed higher MSE and lower R2 values in case of MLR models is due to the presence of some nonlinearity between FoS (dependent parameter) and independent parameters. This fact is clearly evident from the results of Multiple Nonlinear Regression (MNR) models, where R2 value of above 0.95 is observed in most of the cases. Both ANN and MR models appear to be promising to quickly assess the stability of large number of soil slopes of hilly regions in contrast to the time consuming conventional LEMs. However, due to certain assumptions and limitations involved in each of these methods, there is an ambiguity over the selection of a particular method for the assessment of slope stability. Hence, apart from comparing the accuracy of the developed ANN and MR models with LEMs, they have also been compared with respect to each other, for various soil types under dry, saturated and seismic loading conditions. The accuracy of slope stability prediction of these methods has been compared. It is clearly observed from these results that accurate prediction of stability of slopes is possible with a well-trained ANN, in comparison to both the types of multiple regression models. MNR models performed far better than MLR models. Poor performance of MLR can be ascribed to the partial nonlinearity observed between dependent and independent variables in certain cases. Relative importance factors estimated form ANN and MR models showed good correlations in terms of trends and values. Overall, it can be concluded that ANN models can be used as an alternative to conventional LEMs for the rapid stability assessment of large number of soil slopes of hills. Though MLR models provide simple relationship between the parameters, the accuracy is not found to be adequate. On other hand, MNR models are found to provide more accurate estimation than MLR models. vi However, the accuracy of MNR models is still lower than ANN models. Thus, multiple regression models can be used only for preliminary assessment of slopes as they can easily be used by a practicing geotechnical engineer, who doesn’t have much idea on computer programming to develop ANN models.