Combining Hyperspectral and Lidar Data for
Vegetation Mapping in the Florida Everglades
Caiyun Zhang
Abstract
This study explored a combination of hyperspectral and lidar
systems for vegetation mapping in the Florida Everglades.
A framework was designed to integrate two remotely sensed
datasets and four data processing techniques. Lidar eleva-
tion and intensity features were extracted from the original
point cloud data to avoid the errors and uncertainties in
the raster-based lidar methods. Lidar significantly increased
the classification accuracy compared with the application
of hyperspectral data alone. Three lidar-derived features
(elevation, intensity, and topography) had the same con-
tributions in the classification. A synergy of hyperspectral
imagery with all lidar-derived features achieved the best
result with an overall accuracy of 86 percent and a Kappa
value of 0.82 based on an ensemble analysis of three ma-
chine learning classifiers. Ensemble analysis did not signifi-
cantly increase the classification accuracy, but it provided
a complementary uncertainty map for the final classified
map. The study shows the promise of the synergy of hyper-
spectral and lidar systems for mapping complex wetlands.
Introduction
The Importance of Vegetation Information in the Florida Everglades
The Florida Everglades is the largest subtropical wetland in
the United States. It has been designated as an Internation-
al Biosphere Reserve, a World Heritage Site, and a Wetland
of International Importance due to its unique combination
of hydrology and water-based ecology that supports many
threatened and endangered species (Davis
et al
., 1994). In
the past century, human activities have severely modified the
Everglades ecosystem, resulting in a variety of environmental
issues in South Florida (McPherson and Halley, 1996). To
protect this valuable resource the US Congress authorized the
Comprehensive Everglades Restoration Plan (
CERP
) in 2000
to restore the Everglades ecosystem (
CERP
, 2013).
CERP
is a
$10.5 billion
USD
mission that is expected to take 30 or more
years to complete. It contains a variety of pilot environmental
engineering projects, many of which require accurate and
informative vegetation maps, because the restoration will
cause dramatic modification of plant communities (Doren
et
al
., 1999). Monitoring changes of vegetation communities can
measure the progress and effects of restoration on environ-
mental health (Doren
et al
., 1999; Welch
et al
., 1999).
Vegetation Mapping Using Hyperspectral and Lidar Systems
Vegetation mapping efforts to support
CERP
have focused
on manual interpretation of large-scale aerial photographs
using analytical stereo plotters (Rutchey
et al
., 2008; Jones,
2011). This procedure is time-consuming and labor-intensive.
Automated classification of the digital aerial photograph
cannot produce the requisite accuracy due to its poor spectral
resolution (Zhang and Xie, 2013b). Two promising remote
sensing techniques, hyperspectral and the Light Detection
And Ranging (lidar) systems, offer significant advantages over
manual interpretation of aerial photographs.
Hyperspectral sensors collect data in hundreds of relative-
ly narrow spectral bands throughout the visible and infra-
red portions of the electromagnetic spectrum. Research has
demonstrated the merit of hyperspectral data in a range of
applications such as quantifying agricultural crops, classify-
ing vegetation types, and characterizing wetlands (Thenkabail
et al
., 2011). The application of hyperspectral systems has
become an important area of research for wetland mapping
in the past decade (Adam
et al
., 2010). Such research can
be grouped into two categories. The first is the employment
of hyperspectral data alone (e.g., Hunter and Power, 2002;
Hirano
et al
., 2003; Schmidt
et al
., 2004; Artigas and Yang,
2005; Harken and Sugumaran, 2005; Li
et al
., 2005; Rosso
et
al
., 2005; Pengra
et al
., 2007; Jollineau and Howarth, 2008;
Zhang and Xie, 2012; Zhang and Xie, 2013a). The second is
the application of a synergy of hyperspectral and other remote
sensing data for a better characterization of wetlands (e.g.,
Held
et al
., 2003; Yang and Artigas, 2010; Onojeghuo and
Blackburn, 2011; Zhang and Xie, 2013b).
Lidar systems were originally designed to facilitate the
collection of data for digital terrain modeling by using the
reflections from the ground. Studies have illustrated that lidar
can be used to characterize vegetation, especially forests, us-
ing non-ground reflections (Hyyppä
et al
., 2008; van Leeuwen
and Nieuwenhuis, 2010). Lidar can complement the spectral
information of optical imagery to improve vegetation classifi-
cation. Encouraging results have been achieved by integrating
lidar data and hyperspectral imagery (e.g., Hill and Thomson,
2005; Mundt
et al
., 2006; Geerling
et al
., 2007; Jones
et al
.,
2010; Onojeghuo and Blackburn, 2011; Zhang and Qiu, 2012;
Cho
et al
., 2012). Research efforts on vegetation characteriza-
tion using lidar is dominated by high-posting-density (i.e., >4
pts/m
2
) lidar data (Ke
et al
., 2010; Zhang
et al
., 2013). Appli-
cation of low-posting-density (i.e., <2 pts/m
2
) lidar focuses on
terrestrial topographic mapping, and its research in vegetation
mapping is limited (Ke
et al
., 2010). Little work has been
conducted to combine low-posting-density lidar data with
hyperspectral imagery for vegetation mapping in the complex
wetlands. In addition, most lidar vegetation studies have only
examined the contribution of elevation information. Few
studies have explored the combined contribution of all the
Department of Geosciences, Florida Atlantic University, 777
Glades Road, Florida 33431 (
).
Photogrammetric Engineering & Remote Sensing
Vol. 80, No. 8, August 2014, pp. 733–743.
0099-1112/14/8008–733
© 2014 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.80.8.733
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
August 2014
733
