The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
1322
nm, ~1900 nm and -2200 nm were used for predicting
engineering parameters.
PLS1 analysis (predicting a single engineering parameter at a
time) method implemented in The Unscrambler software was
used for the multivariate calibration and validation. Outlier
detection was performed through analyzing residuals, leverages,
normal probability plots of residuals both in the X and Y spaces,
stability plots etc. Sufficient number of PLS factors were
determined based on significance tests and through examining
RMSE of residual variances of each PLS factor. Regression
coefficients that significantly contribute to the models were
selected based on uncertainty limit tests. Full cross validation
method was used for calibrating and validating models.
Measured LL
Correlation coefficient RMSEP SEP Bias offset
086 056 Ô57 0.003 L89
Relevant predictors (absorption feature parameters)
position, depth, width & area -1900 nm; depth , width & area -1400
nm; position , depth & width ~2200nm
Correlation coefficient RMSEP SEP Bias offset
0.85 0J5 015 0.000 011
Relevant predictors (absorption feature parameters)
position, depth, width & area -1900 nm; depth & area -1400 nm;
position & width ~2200nm
Figure 6. Results of PLSR modeling for liquid limit and cation
exchange capacity showing the regression overview (measured
versus predicted values) in the calibration and validation stages.
As indicated by the correlation coefficients (Figure 5) a large
portion of the variation in the engineering parameters could be
accounted for by the spectral parameters. Estimation errors,
both RMSEP and SEP are small indicates that expected error
will be within acceptable limits if prediction is done on new
samples using these models. Bias, which is average value of the
difference between predicted and measured values, is also small
for the given number of PLS factors indicating that effect of
bias in the modelling is negligible (Martens and Naes, 1989).
offse
Correlation coefficient
RMSEP
SEP
Bias
t
PL
0.68
0.60
0.60
0.000
1.65
PI
0.83
0.47
0.48
0.000
0.65
FS
0.64
0.22
0.22
0.001
0.97
Relevant predictors (absorption feature parameters)
PL Position, depth & area -1900 nm; depth & area -1400 nm;
position, depth & width - 2200 nm.
PI Position, depth & area -1900 nm; depth & area -1400 nm;
position, depth, width & area - 2200 nm.
FS Position & depth -1900 nm; depth -1400 nm; position & width -
2200 nm.
Table 1. Summary of model results for engineering parameters
(PL, PI and FS)
Position and depth -1900 nm; depth - 1400 nm; position and
width - 2200 nm are found to be significant predictors for all
the five engineering parameters: LL, PL, PI, CEC and FS
(Figure 6 and Table 1). Water bearing clay mineral groups,
smectites produce absorption features due to the OH stretching
at -1400 nm; and due to combination of H-O-H bending and
OH-stretching near -1900 nm. Combination of fundamental OH
stretching and bending with Al, Mg, or Fe ions are known to
produce absorption features at -2200 - 2300 nm , and these are
diagnostic of clay minerals (Clark, 1999). According to Clark,
(1999) absorption features - 1400 nm is due to OH and H 2 0
which can be caused by kaolinite/ smectite/ illite clay minerals;
absorption features - 1900 nm is associated with H 2 0 which
can be caused by smectite/ illite clay mineral species; and
absorption features - 2200 nm is related with Al-OH, Mg-OH,
or Fe-OH bending which can be caused by kaolinite/ smectite/
illite clay minerals.
4. CONCLUSIONS
In this paper, relationships between expansive soil engineering
parameters and absorption feature parameters calculated from
their respective reflectance spectra were examined. We were
able to quantify engineering parameters from soil reflectance
spectra employing a multivariate regression analysis, PLSR.
The results confirm spectroscopy’s potential in assessing
engineering characteristics of expansive soils both in
identification of expansive soils and subsequent quantification
of their engineering parameters (Figure 3 and 5). Apart from
supplying a great deal of information within a short period of
time and of being cheaper, accuracy of spectroscopic estimates
seem also reliable.
Laboratory spectroscopy can be utilized for measuring soil
samples from any depth. It require small volume of samples for