A Comparative Study of Land Cover Classification
Techniques for “Farmscapes” Using Very
High Resolution Remotely Sensed Data
Niva Kiran Verma, David W. Lamb, Nick Reid, and Brian Wilson
Abstract
High spatial resolution images (~10 cm) are routinely avail-
able from airborne platforms. Few studies have examined the
applicability of using such data to characterize land cover in
“farmscapes” comprising open pasture and remnant vegeta-
tion communities of varying density. Very high spatial resolu-
tion remotely sensed imagery has been used to classify land
cover classes on a ~5000 ha extensive grazing farm in Austra-
lia. This “farmscape” consisted of open pasture fields, scat-
tered trees, and remnant vegetation (woodlands). The relative
performances of object-based and pixel-based approaches to
classification were tested for accuracy and applicability. Max-
imum likelihood classification (
MLC
) was used for pixel-based
classification while the k-nearest neighbor (k-NN) technique
was used for object-based classification. A range of image
sampling scales was tested for image segmentation. At an op-
timal sampling scale, the pixel-based classification resulted in
an overall accuracy of 77 percent, while the object-based clas-
sification achieved an overall accuracy of 86 percent. While
both the object- and pixel-based classification techniques
yielded higher quantitative accuracies, a “more realistic”
land cover classification, with few errors due to intermixing of
similar classes, was achieved using the object-based method.
Introduction
Remote sensing provides a useful source of data to extract
accurate land-use and land-cover (
LULC
) information for
planning and implementation of different land use practices
(Xiuwan, 2002; Falcucci
et al.
, 2007). The demand for ac-
curate and up-to-date
LULC
information, along with historical
change as well as future trajectories, has been acknowledged
by various researchers (e.g., Sobrino and Raissouni, 2000;
Hester
et al.
, 2010). Most studies in the past two decades were
based on medium resolution remote sensing data such as
Landsat
TM
(30 m),
SPOT
(20 m), etc. These were found useful
in regional and medium scale land cover mapping and change
detection analysis (e.g., Robertson and King, 2011). In recent
years, numerous studies have used high-resolution images of
meter to sub-meter spatial resolution from satellite systems
such as Ikonos and QuickBird, to identify small-scale features
in a time and cost-effective way (e.g., Puissant
et al.
, 2005; Jo-
hansen
et al.
, 2007). Today, very high spatial-resolution data,
in the order of tens of centimeters, is now routinely available
from airborne sources, offering the possibility of creating
land cover maps of greater detail for planning applications
(Dehvari and Heck, 2009) as well as for conducting above-
ground biomass or carbon stock assessments (Brown
et al.
,
2005; Hester
et al.
, 2010).
High-resolution imagery increases the information avail-
able on land cover at both local and national scales (Aplin
et
al.
, 1997), allowing improved delineation between features
(Thomas
et al.
, 2003). Landscapes, including those used
for extensive farming practices, so-called “farmscapes,” are
inherently spatially variable. Taken at the individual farm
level, farmscapes can include agricultural fields (crops and
pastures), remnant native vegetation (trees, open woodland),
water features, roads, buildings, orchards and other develop-
ments. In Australia, a single farm can include a diverse range
of
LULC
classes and can range in size from 10 ha to 10,000 ha.
Despite the considerable range in spatial extent, enterprise-
relevant farm management tools to assess biomass in its many
forms start at the individual field scale and land use classifi-
cation must necessarily occur at the sub-field scale, namely,
at the order of tens of meters (Bramley
et al.
, 2008; Cook and
Bramley, 2000; Lamb, 2000). At this spatial scale, medium-
resolution data may give erroneous results when classifying
the boundary pixels. Sampling resolution must always exceed
that of the proposed delineation (Woodcock and Strahler,
1987). This increases the probability of having purer pixels
(end-member pixels) available for whatever classification
procedure is in use (Mundt
et al.
, 2006) as well as reducing
co-registration errors (Weber
et al.
, 2008).
For so many years, pixel-based classifications (
PBC
) have
been used successfully in various applications. The conven-
tional pixel based supervised methods such as maximum
likelihood classifier (
MLC
), minimum distance from means
(
MDM
), and parallelepiped all examine only spectral informa-
tion of the image to produce a classification. Such parametric
classifiers work on two assumptions: (a) that the image data
is normally distributed, and (b) that the training samples’ sta-
tistical parameters (e.g., mean vector and covariance matrix)
truly represent the corresponding land cover class. However,
the image derived parameters are not always normally distrib-
uted, especially in complex landscapes, and uncertainty in
image classification can be exacerbated by a lack of sufficient
training data and multimodal training samples. Moreover,
with high spatial resolution multi-spectral images, classifica-
tion can also be confounded by spectral similarities between
Precision Agriculture Research Group and Cooperative Re-
search Centre for Spatial Information (CRCSI), University of
New England, Armidale, NSW 2351, Australia
(
).
Photogrammetric Engineering & Remote Sensing
Vol. 80, No. 5, May 2014, pp. 461–470.
0099-1112/14/8005–461
© 2014 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.80.5.461
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2014
461
