for tree and other vegetative covers were similar. In the late
May image,
NDVI
accuracy was higher than the rest of the year,
because foliage had returned to trees but many agricultural
fields remained unplanted or were newly planted and crops
were immature.
Additionally, soil is typically more visible in agriculture
than forest cover, especially before crops reach maturity, and
both wet and dry soil typically are higher in the red region
of the electromagnetic spectrum (Roberts
et al.
, 1993; Huete,
1986) than is green vegetation.
Therefore, a pixel that contains soil and green vegetation
has a higher red reflectance value, which was evident in the
spectral profiles of agricultural and other non-tree vegetative
pixels throughout the imagery (see Figure 2 for a comparison
of a tree spectral profile to a profile of other vegetative spec-
tra). This could have been another factor that led agriculture
to be spectrally separable from forest cover.
Limitations and Future Research Directions
Our study site was predominately conifer and deciduous trees
with closed canopies, and agricultural fields with annual
and perennial crops, which contributed to the high level of
accuracy of the FCI1 and FCI2. Additionally, the study site
had little topographic relief which minimized variations in
reflectance caused by the presence of both sunlit and shaded
slopes. Since the success of FC1 and FC2 in separating tree
cover from other vegetative covers is partially due to shadow-
ing, it is possible that topographic shadowing would com-
plicate this in a less flat environment. This research tested
WorldView-2 imagery that was between 11.5 and 31.5 degrees
off-nadir, and did not examine how increasing the off-nadir
view angle beyond 31.5 would impact the ability of the FCI
and FCI2 to detect forest and individual tree cover, as a larger
off-nadir view angle increases the pixel size and makes shad-
owing more extreme.
Future research will examine other environments with
variation in terrain, locations that contain a variety of veg-
etative land covers that could be misclassified as trees, and
areas with a less dense tree canopy and wider variety of tree
species. Additional research will explore optimizing thresh-
old derivation without user input to automatically identify
the ideal threshold, and a comparison of the FCI1 and FCI2
to other vegetation indices beyond
NDVI
. Histograms have
been used in prior research to threshold imagery (Kittler and
Illingworth, 1986).
Future research will also include testing of the FCI2 with
other sensors that contain bands in the red and near infrared
regions of the electromagnetic spectrum, such as Landsat.
Since Landsat imagery is lower in spatial resolution than
WorldView-2, it could be useful in identifying forest cover but
not for locating individual trees. Landsat also provides open
access to an archive with global coverage, which WorldView-2
does not.
Conclusions
A methodology was developed to identify trees to mask out
forest cover in commercial multispectral imagery. The method
is a straightforward binary approach using two newly de-
veloped indices, FCI1 and FCI2, where the user applies the
indices to WorldView-2 imagery and defines a threshold to
separate forest from other vegetative land covers. The results
indicate FCI2 and FCI2 are more effective than using a tradi-
tional vegetation index, like
NDVI
, and are accurate throughout
the year in differentiating between trees and other vegeta-
tive land covers. FCI1 and FCI2 are also effective in identi-
fying individual trees. Previous research in tree detection
has been less successful in doing so due to spectral overlap
between trees and other vegetative covers or spatial resolution
limitations. This methodology circumvents the need for
HSI
and lidar data to identify trees and will aid in efforts to
map agriculture and deforestation, identify trees in an urban
environment, and assist in vehicular route modeling. Future
research will focus on testing FCI1 and FCI2 on diverse land
covers and locations, employing lower resolution but freely
available Landsat imagery, and using histograms to optimize
threshold derivation between tree and other land covers with-
out user input.
Acknowledgments
Funds for were provided by the U.S. Army Corps of Engineers
(
USACE
) to support this research under the Geospatial Analy-
sis at the Tactical Edge, Geospatial Intelligence and Complex
Urban Environments, and Geo-enabled Augmented Intelli-
gence for Decisive Engagement work packages. The authors
would like to thank the following
USACE
staff: Jeremy Atkin-
son, S. Bruce Blundell, Michael Campbell, Richard Curran,
Jacqueline Hunke, Susan Lyon, Dwayne Russler, and Nicole
Wayant for help during the conception, execution, and review
of the project; and Sean Griffin and Matthew Voss for writ-
ing code to automate the FCI1 and FCI2 in
ENVI
and ArcGIS
®
.
Public release, distribution unlimited.
References
Adler-Golden, S., A. Berk, L.S. Bernstein, S. Richtsmeier, P.K.
Acharya, M.W. Matthew, G.P. Anderson, C.L. Allred, L.S.
Jeong, and J.H. Chetwynd, 1998. FLAASH, a MODTRAN4
atmospheric correction package for hyperspectral data retrievals
and simulations,
Proceedings of the Seventh JPL Airborne Earth
Science Workshop
, 12-16 January, Jet Propulsion Laboratory,
Pasadena, CA, pp. 9–14.
Akar, A., E. Gökalp, Ö. Akar, and V. Yılmaz, 2017. Improving
classification accuracy of spectrally similar land covers in the
rangeland and plateau areas with a combination of WorldView-2
and UAV images,
Geocarto International
, 32(9):990–1003.
Bandyopadhyay, M., J.A. van Aardt, and K. Cawse-Nicholson,
2013. Classification and extraction of trees and buildings from
urban scenes using discrete return LiDAR and aerial color
imagery,
Proceedings SPIE 8731, Laser Radar Technology and
Applications XVIII
, 873105.
Bauer, M.E., 1985. Spectral inputs to crop identification and
condition assessment,
Proceedings of the IEEE
, 73(6):1071–1085.
Biard, F., and F. Baret, 1997. Crop residue estimation using multiband
reflectance,
Remote Sensing of Environment
, 59(3):530–536.
Bunnik, N.J.J., 1978.
The Multispectral Reflectance of Shortwave
Radiation by Agricultural Crops in Relation with their
Morphological and Optical Properties
, Ph.D. dissertation,
Agricultural University, Wageningen, The Netherlands, 176 p.
Bunting, P., and R. Lucas, 2006. The delineation of tree crowns
in Australian mixed species forests using Compact Airborne
Spectrographic Imager (CASI) data,
Remote Sensing of
Environment
, 101(2):230–248.
Carter, G.A., 1994. Ratios of leaf reflectances in narrow wavebands
as indicators of plant stress,
International Journal of Remote
Sensing
, 15(3):697–703.
Chen, Q., D. Baldocchi, P. Gong, and M. Kelly, 2006. Isolating
individual trees in a savanna woodland using small footprint
lidar data,
Photogrammetric Engineering & Remote Sensing
,
72(8):923–932.
Congalton, R.G., and K. Green, 2008.
Assessing the Accuracy of
Remotely Sensed Data: Principles and Practices, Second Edition,
CRC press, Boca Raton, Florida, 183 p.
Congalton, R.G., R.G. Oderwald, and R.A. Mead, 1983. Assessing
Landsat classification accuracy using discrete multivariate
analysis statistical techniques,
Photogrammetric Engineering &
Remote Sensing
, 49(12):1671–1678.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
August 2018
511
