Optimal Land Cover Mapping and Change
Analysis in Northeastern Oregon Using
Landsat Imagery
Michael Campbell, Russell G. Congalton, Joel Hartter, and Mark Ducey
Abstract
The necessity for the development of repeatable, efficient,
and accurate monitoring of land cover change is paramount
to successful management of our planet’s natural resources.
This study evaluated a number of remote sensing methods
for classifying land cover and land cover change throughout
a two-county area in northeastern Oregon (1986 to 2011).
In the past three decades, this region has seen significant
changes in forest management that have affected land use
and land cover. This study employed an accuracy assess-
ment-based empirical approach to test the optimality of a
number of advanced digital image processing techniques
that have recently emerged in the field of remote sensing.
The accuracies are assessed using traditional error matrices,
calculated using reference data obtained in the field. We
found that, for single-time land cover classification, Bayes
pixel-based classification using samples created with scale
and shape segmentation parameters of 8 and 0.3, respective-
ly, resulted in the highest overall accuracy. For land cover
change detection, using Landsat-5 TM band 7 with a change
threshold of 1.75 standard deviations resulted in the highest
accuracy for forest harvesting and regeneration mapping.
Introduction
Remote sensing technologies are unparalleled in their ability
to monitor and analyze Earth’s natural resources rapidly,
cost-effectively, and with ever-increasing levels of precision
and accuracy (Jensen, 2005). Although a number of high
spatial resolution imagery platforms have emerged in recent
years (e.g., Ikonos, QuickBird), the Landsat program has
greatly benefited the remote sensing community by providing
consistently high quality, medium spatial resolution imagery
since 1972 (Green, 2006). Landsat-5 Thematic Mapper (
TM
)
has proven particularly valuable, having contributed almost
30 years worth of essentially uninterrupted data (well beyond
its expected life span of three years) at a bi-monthly temporal
resolution (Chander and Markham, 2003). With Landsat data
now freely available, the potential for remote sensing studies
of all kinds has exploded as indicated by a 60-fold increase in
data downloads since January, 2009 (NASA).
Central to the study of natural resource management is the
ability to monitor changes in the landscape over time. The
remote sensing community is constantly seeking newer and
better ways to accomplish this very goal. Programs like the
National Land Cover Database (
NLCD
) are extremely valu-
able in providing a baseline of data which can be utilized in
studies spanning an array of disciplines (Homer
et al.
, 2004).
Additionally, the
NLCD
provides a generalized framework by
which similar land cover assessments can be accomplished,
including a tried-and-true methodology for land cover change
analysis (Xian
et al.
, 2009). Similarly, the National Oceanic
and Atmospheric Administration’s (
NOAA
) Coastal Change
Analysis Program (C-CAP) has informed this study and others
by suggesting a number of standardized techniques by which
land cover change can be monitored (Dobson
et al.
, 1995).
Traditionally, land cover mapping and analysis was
performed on a pixel basis, i.e., a purely spectral approach
wherein reflectance values for each pixel (and derivative
information) of an image are the sole basis for classifying
the imagery into a map. Within the last decade, object-based
image analysis (
OBIA
, also called
GEOBIA
) has gained momen-
tum in the remote sensing community (Blaschke, 2010).
OBIA
is based on segmenting images (i.e., grouping of pixels) into
meaningful areas of spatial and spectral homogeneity called
“objects” (Jensen, 2005). There is a great degree of user flexi-
bility in generating these objects, guided by the manipulation
of three parameters: scale, shape, and compactness to produce
the optimal segmentation (e.g., Moller
et al.
, 2007). While the
results tend to be case-specific, there appears to be general
agreement that images can be over-segmented (objects are too
small) and under-segmented (objects are too large) (Kim
et
al.
, 2008; Holt
et al.
, 2009; Liu and Xia, 2010; MacLean and
Congalton, 2011).
While the majority of
OBIA
studies tend to focus on feature
extraction from high-resolution image data (e.g., Moran, 2010;
Alganci
et al.
, 2013), a few have explored its applications
on medium-resolution data sources such as Landsat (e.g.,
Geneletti and Gorte, 2003; Gamanya, 2009). An increasing
number of studies are inquiring into the feasibility of using
OBIA
techniques to analyze land cover change (e.g., Im
et al.
,
2008; Chen
et al.
, 2012), but we have found few studies that
link object-based land cover change and Landsat-5
TM
data;
Robertson and King (2011) is a notable exception.
While the remote sensing community has consistently
pushed the limits of technical and computational capacity,
seeking to develop new and improved methodologies, there is
a critical need for the implementation of broad-scale monitor-
ing operations that employ relatively simple, repeatable, and
comprehensible processes. The focus of this study is precisely
that: to establish an analytical and processing workflow for
Michael Campbell, Russell G. Congalton, and Mark Ducey are
with the Department of Natural Resources & the Environment,
56 College Road, 114 James Hall, University of New Hamp-
shire, Durham, NH 03824 (
.
Joel Hartter is with the Department of Geography, Universi-
ty of New Hampshire, Durham, NH 03824, and currently at
the Environmental Studies Program, University of Colorado
Boulder, Boulder, CO 80309.
Photogrammetric Engineering & Remote Sensing
Vol. 81, No. 1, January 2015, pp. 37–47.
0099-1112/15/811–37
© 2014 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.81.1.37
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
January 2015
37
