Neural networks based linear (PCA) and nonlinear (ISOMAP) feature extraction for soil swelling pressure prediction (North East Algeria)

Abstract

The swelling pressure (SP) of expansive soils is crucial for both geotechnical studies as well as practitioners. Multiple attempts have been made to correlate the SP with the properties of soil due to the difficulty of determining it in the laboratory. However, the large number of environmental and physical governing parameters makes accurate SP predictions difficult. In this paper, Artificial Neural Networks (ANNs) are used to assess accurate prediction of SP of soil. Dimension reduction techniques are intensely required for ANNs inputs. Feature extraction (FE) based dimension reduction (DR) methods map original multidimensional space into a space of reduced dimensionality. This paper presents a comparative study of linear FE using Principal Component Analysis (PCA) and nonlinear FE using ISOmetric MAPping (ISOMAP) for feed forward neural models to predict SP. Results showed that FE technique improves ANNs models compared to multiple linear regression (MLR) and ANNs model without DR. Moreover, nonlinear ISOMAP based DR technique has proven its effectiveness regarding performance metrics for five dimensions inputs (Dims), Determination coefficient (R2 = 0.923), Mean absolute percentage error (MAPE = 0.072), and Root mean square error (RMSE = 54.937) and Root relative squared error (RRSE = 0.383). Therefore, ISOMAP-ANN models can be adopted to solve geotechnical problems specially those of expansive soils which have a very complex and nonlinear structure.