PE&RS July 2016 Public - page 571

Object-based Image Mapping of Conifer Tree
Mortality in San Diego County based on
Multitemporal Aerial Ortho-imagery
Mary Pyott Freeman, Doulas A. Stow, and Dar A. Roberts
Abstract
Two GEOBIA approaches are compared for their effective-
ness in mapping dead trees within island montane forests
of Southern California: a spatial contextual approach using
an artificial neural network classifier, and a segmentation
and multi-pixel classification approach. Both approaches
are tested with multitemporal aerial orthoimagery having
varying spatial resolutions. Spectral transformation inputs
are also tested. An object-based accuracy assessment is con-
ducted. Accuracies range between 30 percent to 90 percent
for the dead tree class and are significantly higher for the
spatial-contextual approach. Inclusion of spectral transforms
increased accuracies by 5 percent for the true object-based
approach, up to 13 percent for the spatial contextual ap-
proach, and reduced commission error up to 10 percent for
both approaches. Masking techniques increased accuracies of
the spatial contextual approach by 20 percent. With manual
editing, the most accurate maps of individual live and dead
trees from the spatial contextual approach are suitable for
studying spatio-temporal trends in montane conifer mortality.
Introduction
The montane mixed-conifer forests of San Diego County,
California are currently experiencing extensive tree mortal-
ity. Remote sensing provides the potential for labor effective,
repeatable, and spatially comprehensive approach to map-
ping, assessing and monitoring tree mortality conditions.
High spatial resolution digital, aerial orthoimagery provides
an affordable source of image data for San Diego County and
much of Southern California that are available for multiple
dates over the past two decades, and at suitable spatial scales
for the study of population level vegetation change (Erikson,
2003; Graetz, 1990; Gustafson, 1998). There is a need for
developing more automated image-based approaches to map-
ping tree mortality, given the extensive areas of mortality, the
large data volume of high spatial resolution remotely sensed
data sets, and the time-consuming efforts required for manu-
ally generating tree mortality maps. To this end, advances in
geographic object-based image analysis (
GEOBIA
) techniques
may provide a means for rapidly obtaining spatially explicit
information about tree conditions for large areas of forests.
The overall research goal is to better understand the
spatial-temporal patterns of tree mortality for an eight-year
period from 1997 to 2005, and the climatic and topographic
influences on the mortality within the montane forests of
San Diego County. For this paper the emphasis is on assess-
ing the reliability of maps of dead tree objects generated with
advanced remote sensing techniques applied to high spatial
resolution airborne orthoimagery. The specific objective of
this paper is to test the effectiveness of two
GEOBIA
classifica-
tion approaches, to efficiently map tree mortality, utilizing ex-
isting high spatial resolution orthoimagery data sets covering
the study area and period. The ultimate goal for our follow-on
research will be to delineate individual trees as objects, clas-
sify them as alive or dead, and track the location of dead tree
objects through time.
Two commercial off-of-the-shelf (
COTS
) software programs
were evaluated, a true
GEOBIA
software, eCognition
®
, and
a per-pixel, artificial neural network (
ANN
) classifier that
exploits spatial contextual information, Feature Analyst.
Henceforth, the first software approach will be referred to as
“object-based” and the second as the “spatial contextual” ap-
proach. Given the usefulness of remote sensing applications,
the costs associated with these applications, the many types
of software programs available, as well as many types of high
spatial resolution data, it is an interesting and useful exercise
to examine how effective existing approaches and readily
available software are in achieving the task of mapping tree
mortality. Land managers and remote sensing service provid-
ers may be interested in which approaches might would be
the most cost effective while achieving reliable maps and
inventories of tree mortality.
GEOBIA
methods are becoming more prevalent, due to the
greater availability of commercial object-based imagery analy-
sis software, high power computing systems, and high spatial
resolution remote sensing data (Benz
et al
., 2004; Blaschke
and Strobl, 2001).
GEOBIA
methods have an advantage over
per-pixel approaches in their ability to incorporate spatial,
textural, shape, and contextual information for contiguous
groups of pixels, and have been shown to provide improved
classification accuracies over pixel-based classifications, due
in part to their ability to overcome the so called “salt and
pepper effect” (Blaschke, 2010; Guo
et al
., 2007). For mapping
tree canopies,
GEOBIA
approaches have been demonstrated to
yield overall higher accuracies than pixel based approaches
(Poznanovic
et al.
, 2014).
True
GEOBIA
approaches utilize a segmentation algorithm,
which incorporates both spectral and spatial information to
group homogeneous regions of pixels to create image ob-
jects called segments. Image spectral-radiometric similarity
Mary Pyott Freeman is with California State University
Fullerton, Fullerton College, Chaffey College, 1430 Carissa
St. Upland CA 91784, and formerly with the Department of
Geography, San Diego State University SDSU/UCSB Joint
Doctoral Program (
).
Doulas A. Stow is with the Department of Geography, San Diego
State University, 5500 Campanile Drive, San Diego, CA 92182.
Dar A. Roberts is with the Department of Geography, UC
Santa Barbara, Santa Barbara, CA 93106.
Photogrammetric Engineering & Remote Sensing
Vol. 82, No. 7, July 2016, pp. 571–580.
0099-1112/16/571–580
© 2016 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.82.7.571
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
July 2016
571
447...,561,562,563,564,565,566,567,568,569,570 572,573,574,575,576,577,578,579,580,581,...582
Powered by FlippingBook