Semi-Automated Disaggregation of a Conventional
Soil Map Using Knowledge Driven Data Mining
and Random Forests in the Sonoran Desert, USA
Travis W. Nauman, James A. Thompson, and Craig Rasmussen
Abstract
Conventional soil maps (
CSM
) have provided baseline soil
information for land use planning for over 100 years. Al-
though
CSM
have been widely used, they are not suitable to
meet growing demands for high resolution soil information
at field scales. We present a repeatable method to disaggre-
gate
CSM
data into ~30-meter resolution rasterized soil class
maps that include continuous representation of probabilis-
tic map uncertainty. Methods include training set creation
for original
CSM
component soil classes from soil-landscape
descriptions within the original survey database. Training sets
are used to build a random forest predictive model utilizing
environmental covariate maps derived from
ASTER
satellite
imagery and the
USGS
National Elevation Dataset. Results
showed agreement at 70 percent of independent field vali-
dation sites and equivalent accuracy between original
CSM
map units and the disaggregated map. Uncertainty of predic-
tions was mapped by relating prediction frequencies of the
random forest model and success rates at validation sites.
Introduction
The increased availability of both digital elevation data and
remote sensing data have prompted many studies that use
these data to improve soil property prediction and inventory
in a field that has been coined “digital soil mapping” (
DSM
)
(Grunwald
et al
., 2011; Grunwald, 2009; McBratney
et al
.,
2003; Scull
et al
., 2003). Many of these studies use elevation
data and remotely sensed imagery to represent one or more
soil forming factors that include climate, organisms, relief,
parent material, and time (Jenny, 1941). In this form, soil
classes or properties are predicted from topographic or spec-
tral indices derived from elevation and imagery.
Soil properties and functions influence many societal
challenges particularly the response of ecosystem services
such as carbon and nutrient cycling; water storage, purifica-
tion and cycling; pollutant transport; and vegetation growth
to climate change (Brady and Weil, 2008). However, our
knowledge of soils is imprecise as demonstrated by estimates
of global soil carbon stocks in the top meter of soil that range
from 1,400 to 3,250 petagrams (Grunwald
et al
., 2011). In
light of the projected challenges that climate change presents
to ecosystem services (IPCC, 2007), high quality soil informa-
tion is central to natural resource management and land use
planning. Although many soil inventories in the form of
CSM
have been carried out around the world, the scope and coarse
spatial resolution of many soil databases have been criticized
as limitations to effective incorporation of soil information
into models of ecosystem services and other earth surface
processes (Burrough, 1989; Burrough
et al
., 1997; Grunwald,
2009; Grunwald
et al
., 2011; McBratney
et al
., 2003). The
field of
DSM
has responded to this challenge with concerted
efforts to quantitatively improve
CSM
soil information using a
wide array of statistical, spatial, and information technology
(Behrens et al., 2005; Bui
et al
., 2009; Bui
et al
., 2006; Bui
et
al
., 1999; Bui and Moran, 2001; Cook
et al
., 1996a; Cook
et
al
., 1996b; de Bruin
et al
., 1999; Häring
et al
., 2012; Kempen
et al
., 2009; Kerry
et al
., 2012; McBratney, 1998; Minasny and
McBratney, 2010; Nauman and Thompson, 2014; Nauman
et
al
., 2012; Thompson
et al
., 2010; Yang
et al
., 2011; Zhu, 1997;
Zhu
et al
., 1997, 2001).
One of the main challenges to improving
CSM
data repre-
sentation is that the original intent of
CSM
was management
oriented, and properties attributed to soils were often esti-
mates based on sparse data at representative locations and
not quantifications based on rigorous statistical sampling and
interpolation (USDA-NRCS, 2013). A large part of the goals of
the original design of
CSM
was to provide somewhat quali-
tative interpretations intended to provide pragmatic initial
guidance to developers, farmers, and other land management
institutions (Soil Survey Staff, 1993). However, many current
users of soil information, particularly those not familiar with
CSM
history and evolution, have attempted to use
CSM
data
beyond their original purposes leading to the potential for
spurious relationships and possible incorrect data and pro-
cess interpretation.
Various models and analyses have been developed using
spatial soil information from
CSM
(e.g., Gatzke
et al
., 2011;
Lineback Gritzner
et al
., 2001; Thomas-Van Gundy
et al
.,
2012; Thomas-Van Gundy and Strager, 2012). In the US, both
the US General Soil Map (
STATSGO
2) and the Soil Survey
Travis W. Nauman is with the US Department of Agriculture,
Natural Resources Conservation Service, National Soil Survey
Center, Geospatial Research Unit, and the West Virginia Uni-
versity, Division of Plant and Soil Sciences, 1090 Agricultural
Sciences Building, PO Box 6108, Morgantown, WV 26506
(
).
James A. Thompson is with the West Virginia University, Di-
vision of Plant and Soil Sciences, 1090 Agricultural Sciences
Building, PO Box 6108, Morgantown, WV 26506.
Craig Rasmussen is with the University of Arizona, Depart-
ment of Soil, Water and Environmental Science, 429 Shantz
Building, University of Arizona, Tucson, AZ 85721.
Photogrammetric Engineering & Remote Sensing
Vol. 80, No. 4, April 2014, pp. 353–366.
0099-1112/14/8004–353
© 2014 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.80.4.353
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
April 2014
353
