A Neural Network Approach to Soil Electrical
Conductivity Estimation on Earthen Levees Using
Spaceborne X-band SAR Imagery
Majid Mahrooghy, James Aanstoos, Rodrigo A. Nobrega, Khaled Hasan, and Nicolas H. Younan
Abstract
Monitoring of soil moisture over earthen levees can help in
revealing patterns of moisture variation which could indicate
potential slope failures or other vulnerabilities. In this work a
machine learning approach using a back propagation neural
network (
BPNN
) and also a wavelet basis neural network (
WBNN
)
is developed to estimate the soil electrical conductivity (
EC
) over
an earthen levee system. Three scenarios based on the extract-
ed features are investigated to estimate the conductivity. In sce-
nario one, only radar backscatter coefficients are considered.
In scenario 2, in addition to the backscatter features an average
of backscatter in a sliding 3 × 3 window is used. In scenario 3,
texture features are added, including statistical and wavelet
features. The results show that using all backscatter and texture
features (scenario 3) results in better correlation performance,
with an 11 to 27 percent improvement compared to scenario 1
and a 1 to 17 percent improvement over scenario 2.
Introduction
Field observation and laboratory experiments have shown
that slope failure on earthen embankments is directly linked
with increases in pore water pressure, typically caused by
higher than normal precipitation events (O’Neill, 1995;
Tohari, 2007). Increasing pore water pressure can be detected
through
in-situ
monitoring of soil moisture and can help in
indicating zones of high moisture that can gradually lead to
potential slope failures. Identifying slope instability in a river
levee system is of critical importance to safeguard the flood-
plain protected by the levee. More than 150,000 kilometers of
levees over the entire US play a significant role in protecting
large areas of populated and cultivated land from flooding.
In levees, slope instability is expressed in the form of slough
slides. i.e., most often on the riverward slope.
This study is an extension of our previous work (Mah-
rooghy, 2011). In this extension, we have used a wavelet neural
network, and also included a local incident angle feature,
and tested different scenarios for estimating soil conductiv-
ity for different areas of study. This study is part of a broader
investigation focused on detecting levee slope failures using
synthetic aperture radar (
SAR
) (Cui, 2007). The ultimate goal is
to be able to detect levee anomalies of fairly small extent (5 to
10 meters typical dimension). Zones of unusually high spatial
gradients of soil moisture on the levee system might indicate
potential damage such as seepage or high pore water pressure.
Early awareness of such problems can allow levee managers
to more cost-effectively maintain them and facilitate needed
repairs in a timely manner. Relative spatial variations in surface
soil moisture affect the radar backscatter to facilitate this detec-
tion. Although that application does not require actual soil
moisture estimates, it is useful to ascertain how good an esti-
mate could be obtained from the
SAR
data in this environment.
Remotely sensed observations (with airborne or satellite
platforms) using
SAR
and optical sensors offer an inexpensive
and efficient method which has been employed successfully in
many studies to directly quantify soil moisture from bare sur-
faces. Some of the pioneering studies began in the mid-1980s
using L, C, and X band radar and field measured moisture,
surface roughness and dielectric data. In most cases bare soil
has been investigated and satisfactory estimation results have
been attained using theoretical, empirical, and semi-empirical
approaches. Significant empirical models have been proposed
by Oh
et al
. (Oh, 1992; Oh, 2006), and Dubois
et al.
(Dobson,
1985). The study by Oh
et al.
(Oh, 1992; Oh, 2006) was one of
the first attempts to develop a theoretical solution to predict
soil moisture and roughness of bare soil from truck-mounted
polarimetric radar data at L, C, and X frequencies. They devel-
oped an empirical model using inversion techniques from radar
phase and backscatter measurement of
HH
,
VV
, and
HV
channels
over four test sites. The study by Dubois et. al. (Dobson 1985)
used two co-polarized radar channels (
HH
and
VV
) to calculate
soil moisture over bare and sparsely vegetated areas. In all
three studies, few details on the environmental conditions at
the test site are reported except that they are bare soil sites.
Modification and enhancement of these models led to the
development of others more suitable in different environ-
ments (Bindlish, 2001; De Roo, 2001; Oh, 2002; Prakash, 2012).
Bindlish and Barros (Bindlish, 2001) added a vegetation scat-
tering parameterization to improve soil moisture determination
from vegetated surfaces in Washita, Oklahoma from spaceborne
SIR-C
data. De Roo et al. (2001) conducted their study with a
truck-mounted radar system over soybean fields in Michigan
and found
VV
polarization for L-band and
VV
and
HV
polariza-
tion for C-band radar data to be the most effective in inverting
for soil moisture. Prakash
et al
. (2012) used Normalized Differ-
ence Vegetation Index (
NDVI
) values derived from optical data
Majid Mahrooghy is with Xoran Technologies, Ann Arbor,
MI 48108, and formerly with Geosystems Research Institute,
Mississippi State University, Mississippi State, MS 39762
(
).
James Aanstoos is with Geosystems Research Institute,
Mississippi State University, Mississippi State, MS 39762.
Rodrigo A. Nobrega is with the Institute of Geosciences,
Federal University of Minas Gerais, Belo Horizonte, Brazil.
Khaled Hasan is with the Department of Geology, Austin
Community College, Austin, Texas.
Nicolas H. Younan is with the Department of Electrical
and Computer Engineering, Mississippi State University,
Mississippi State, MS 39762.
Photogrammetric Engineering & Remote Sensing
Vol. 82, No. 7, July 2016, pp. 509–519.
0099-1112/16/509–519
© 2016 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.82.7.509
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
July 2016
509
