REGRESSION AND ARTIFICIAL NEURAL NETWORK MODELS FOR RELATIONSHIP BETWEEN VARIOUS SOIL PROPERTIES AND ELECTRICAL RESISTIVITY

The primary objective of the current research work is to investigate the
relationship between electrical resistivity and various soil parameters of naturally
occurring soils around Universiti Teknologi, Petronas, Malaysia. The research work
consists of four major phases; field resistivity surveys, soil boring, laboratory
resistivity measurements, and soil characterization tests. Field survey includes ID
vertical electrical sounding (VES) and 2D resistivity imaging.
Total 79 soil samples (46 Silty-sand and 36 Sandy soils samples) were obtained
from ten C1CT) boreholes fBH-01 to BH-10) brought to geotechnical laboratory for
various soil characterization tests. Moisture content of soil samples ranges from
6.11% to 52.42%. Plasticity index ranges from 0% to 26.27%. Direct shear test results
indicates that cohesion ranges from 0.00 to 68.23 KPa. The friction angle values for
all soil samples ranges between 5.36° to 42.51°.
The correlations between electrical resistivity and various properties of soil
samples were evaluated using least-squares regression method. Relationship between
moisture content and resistivity values shows a good power correlation with
regression co-efficient R =0.56. Unit weight has poor relationship with resistivity
(R =0.10) for all soil samples. Results indicates a good correlation between plasticity
index and resistivity with regression coefficients R2=0.42, R2=0.19 and R2=0.24 for
all soil samples, silty-sand soils, and sandy soils. Cohesion indicated a weaker
relationship with resistivity for all types ofsoil. Friction angle and resistivity indicates
increasing logarithmic trend with R =0.29 for all soil samples.
Artificial neural network modeling was also performed using LM and SCG
learning rule upto 10 hidden neurons. Best network with particular learning algorithm
and optimum number of neuron in hidden layer presenting lowest root mean square
error RMSE was selected for prediction of various soil properties. ANN models
showed higher prediction accuracy for all soil properties.