The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
1320
Identifying expansive soils and quantifying their potential
expansiveness is crucial to ensure proper site selection,
environmentally compatible and economically feasible
designing and construction especially of lightly loaded
structures. However, common geotechnical practices of
characterizing expansive soils need dense sampling, thus are
costly, labour intensive, time consuming and difficult to get a
continuous representation of soil masses in space.
A great deal of effort has gone into investigating methods that
can be supporting or alternate tools of estimating soil properties.
Advances in remote sensing techniques have enabled
discrimination of clay minerals that cause swelling and
shrinkage in soils and mapping their abundances (Chabrillat et
ah, 2002; Goetz et ah, 2001; Kariuki et ah, 2003; Kariuki et ah,
2004; Van der Meer, 1999). Van der Meer (1999) reported
possibility of mapping clay soils from remotely sensed data
based on the dependence of spectral signatures on soil
constituent minerals. Goetz et ah, (2001) established
relationships between short wave infrared (SWIR) 1800 - 2400
nm spectral bands and soil swelling potential classes of Seed et
ah, (1962). Chabrillant et ah, (2002) identified and mapped
exposed clay minerals (the three most important clay minerals
with respect of soil expansion; smectite, illite, kaolinite) from
airborne remote sensing images based on diagnostic absorption
bands in the SWIR spectral region. Kariuki et ah (2004)
proposed models that made use of spectral parameters from
selected single wavelength regions. They established a one-to-
one link between engineering parameters and absorption feature
parameters (position, depth, width, asymmetry and area of
absorption band) at -1400 nm, -1900 nm and -2200 nm
wavelengths.
In this study we developed new empirical models for estimating
specific engineering parameters of expansive soils from their
respective reflectance spectra. A multivariate calibration
method, partial least squares regression (PLSR) analysis,
making use of all absorption feature parameters calculated from
three wavelength regions (~ 1400 nm, -1900 nm and -2200 nm)
was employed. Resulting models provide numerical estimates
of engineering parameters that can be directly used in practical
engineering applications.
2. MATERIALS AND METHODS
2.1 STUDY AREA
The study area is located in the eastern part of Addis Ababa city
(Figure 2). Climate is cool to temperate with a mean annual
temperature of 16 °C, and a mean annual rainfall of 1200 to
1600 millimetres (EMA, 1988). Elevation ranges from 2700
meters to 2300 meters above sea level.
2.2 SAMPLING AND LABORATORY ANALYSIS
Disturbed soil samples were collected through a stratified
random sampling technique. Stratification was done through
combining information on lithology and topography of the
study area.
Engineering parameters that are commonly used for
identification of and indirect estimation of soil expansiveness;
consistency limits (liquid limits (LL), plastic limits (PL) and
plasticity indices (PI)) following the standard test procedures of
AASHTO specifications (AASHTO, 2002) T89 and T90; cation
exchange capacity (CEC) using methylene blue adsorption test
‘spot’ method (Verhoef, 1992) and free swell (FS) in
accordance with the methods and procedures demonstrated by
Head (Head, 1994) were measured in a soil mechanics
laboratory.
Soil reflectance spectra were acquired using ASD fieldspec full
range spectrometer (http://www.asdi.com) that covers the 350
to 2500 nm wavelength region of the electromagnetic spectrum.
Figure 2. Location map of the study area with names of places
and distribution of sampling points.
2.3 MULTIVARIATE (PARTIAL LEAST SQUARES)
REGRESSION ANALYSIS
Partial least squares regression (PLSR) deals with prediction of
set of dependent (y) variables from set of independent (x)
variables. PLSR is particularly important when dealing with
large number of variables that express common information
(Brereton, 2000; Wold et al., 2001; Yeniay and Goktas, 2002).
Though multiple linear regression (MLR) analysis can be
employed to explore relationships between a number of
predictors and response variables, with an increase in number of
predictors it will not perform well due to multicollinearity
problems. MLR assumes x variables as linearly independent
and require smaller number of x variables than the number of
observations. Significant predictors should also be well known
in MLR (Brereton, 2000). Another multivariate approach,
principal component regression (PCR) analysis decomposes set
of predictors into eigen vectors and scores to overcome
collinearity. After achieving optimal projection of x variables in
few principal components, it regress them against the responses
in a separate step. Unlike PCR, PLSR decomposes both
predictors and responses simultaneously to capture their
common variation, which will be projected into a small number
of mutually independent factors. Decomposition and regression
is a single step, through fewer principal components than that
required by PCR. Hence PLSR reduces the impact of irrelevant
x variations in the calibration modeling by balancing the
information in the x and y spaces (Martens and Naes, 1989;
Wold et al., 2001). More information on the differences of the
three multivariate calibration methods and their algorithms can