and geometric pattern criteria can play an important role in
creating these segments. Segments are then classified using
one of a variety of classification approaches, based on sample
object, or according to class descriptions with a membership
function routine, or using a combination of both approaches
(Al-Khudhairy
et al
., 2005). Finally, in a map generalization
phase, the classified objects can be aggregated, by merging
and filtering (Blaschke, 2010; Jensen, 2006). Another poten-
tial advantage of
GEOBIA
techniques is their ability to handle
differential illumination and view angle effects for image
change analysis with multi-date airborne images (Stow
et al
.,
2007). Object-based approaches have been utilized in forestry
applications to detect and delineate stressed and dead trees
(Flanders
et al
., 2003; Guo
et al
., 2007; Wulder
et al
., 2006).
Artificial neural networks (
ANN
) have been successfully
used to classify remotely sensed data by simulating the
nonlinear thinking process of human beings. They do so by
assigning weights in a network of connecting neurons, based
on remotely sensed data or other geospatial data sets such as
terrain elevation or slope (Atkinson and Tatnall, 1997; Foody
and Arora, 1997). In a training phase, an analyst selects pixel
samples that represent classes of interest, and then during a
learning phase, typically a back-propagation algorithm is used
to adjust the weights in several iterations until the system
achieves convergence, and the classification can be conduct-
ed. Neural networks, utilizing machine learning algorithms,
are advantageous in that they are non-parametric and have the
ability to adaptively simulate complex and nonlinear patterns
(Jensen, 2006). They are also able to incorporate spatial and
contextual information into the classification, and have been
utilized for mapping tree crowns (Vanderzanden and Mor-
rison, 2002). Though pixel-based spatial contextual classifiers
are not truly
GEOBIA
approaches because they do not segment
an image before classifying it, they can ultimately produce de-
finitive image objects when neighborhood pattern information
is incorporated and therefore can be presented as an object-
based approach in the more general sense of an approach that
produces thematic objects for the final classification product.
The object-based approach software program that was
utilized for this study is eCognition, version 5, produced by
Definiens Imaging, which enables a user to segment objects
with a routine that incorporates contextual classifiers based
on fuzzy logic and segment-based statistical feature charac-
teristics. A region-growing multi-resolution segmentation is
conducted, which is based on several customizable inputs, in-
cluding scale factor, shape and color, and object compactness.
The eCognition routine can incorporate extensive artificial
intelligence aspects (Flanders
et al
., 2003). This object-based
multi-pixel approach iterates between processing and classify-
ing image objects. The scale and shape information for image
segments can be optimized and exploited for classification.
The artificial neural network software package that was
tested is Feature Analyst, by Visual Learning Systems, Inc., an
extension for both Esri ArcView
®
and ERDAS Imagine
®
software.
It implements a suite of machine learning algorithms to extract
object-specific geographic features from high resolution imagery
using both spectral and spatial characteristics that resemble tar-
get examples (i.e., training templates) (Kaiser
et al
., 2004). The
user can select pre-packaged input representations, or search
kernel patterns, to optimize object recognition, which in the
present study proved to be the “crosshair” pattern. The software
also employs a hierarchical learning process (called “clean up
clutter feature”) following initial classifications to improve
recognition of the objects that are extracted and thus reduce
commission error. A user provides training samples of represen-
tative objects, and objects of interest are extracted by matching
pixels with spectral and local spatial characteristics that are
similar to those of the training pixels (Kaiser
et al
., 2004).
GEOBIA
approaches to mapping tree mortality patterns were
tested to answer the following questions.
1. What is the utility of high spatial resolution image data
and an object-based routine and a spatial contextual
per-pixel classifier to detect, map and quantify tree
dieback?
2. What is the accuracy of image-derived tree mortality
maps as assessed through object-based accuracy assess-
ment approaches?
3. How do the accuracies compare for maps derived by
the two approaches, and is one of these classification
approaches more useful and reliable than the other, as
measured by processing times, ease of application, and
computed accuracy?
This study provides the first assessment of semi-automat-
ed, object-based classification routines for mapping dead co-
nifer trees from a time series of digital airborne orthoimagery
having high spatial resolution.
Study Area
Tree mortality was investigated within the montane mixed-
conifer forests of San Diego County, which are found on
Palomar Mountain, in the greater Julian area, including Har-
rison Park, Cuyamaca, Laguna Mountain, Descanso, and Pine
Valley, and in the Lost Valley area near Warner Springs. These
coniferous forests occur as “islands” above elevations of 1,100
m in areas that receive more than 50 cm of precipitation. The
forest areas are remnants from a much more widespread forest
that was prevalent during the wetter and cooler Pleistocene
period. They occur within the Peninsular Ranges that extend
north-south from the San Jacinto Wilderness area near Palm
Springs to San Pedro Martir in northern Baja California. The
Peninsular Ranges are comprised of a large westerly-tilted
fault block of Cretaceous granite (Pryde, 1992).
Three specific study sites (as illustrated in Figure 1) were
utilized. Site 1 extends from the Fry Creek Campground area
to Palomar State Park on Palomar Mountain and is comprised
of Sierran mixed coniferous forest (white fir, incense cedar,
Big cone Douglas fir, and Coulter pine), at an elevation of
1,500 m and with an annual average rainfall of 120 cm. Site
2 is located in the Volcan Mountains just north of the town
of Julian and consists of a mixture of white fir, incense cedar,
Big cone Douglas fir, and Coulter pine on the upper slopes of
the West side of the mountains (elevation 1,300 m; rainfall 60
cm). The trees in Sites 1 and 2 grow in tightly packed closed
canopy groves. Site 3 is located on Laguna mountain (1,800
m), and is almost exclusively an open canopy Jeffrey pine for-
est, with an average annual rainfall of 90 cm.
Data and Methods
Image data sets were acquired from government agency
archives for the period 1997 through 2007, and the imagery
from the dates 2002, 2005, and 2007 were selected to test the
remote sensing methods. The different band combinations,
i.e. 2005 imagery, true color (B-G-R); 2002 imagery, color
infrared (G-R-
NIR
); and 2007 imagery, multispectral (B-G-R-
NIR
) enabled the efficacy of the two software approaches to be
tested on data sets with differing spectral-radiometric quali-
ties. The imagery was subset for the three study areas, and
then after collecting calibration and validation data, classified
using both object-based and spatial-contextual software pro-
grams. The resultant tree mortality maps were then compared
using an object-based accuracy assessment approach. Pro-
cessing times and ease of software implementation were also
evaluated. Figure 2 shows the general image processing flow.
Remote sensing analyses were based on several extant
aerial orthorectified and mosaicked image data sets (see Table
572
July 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
