Precise Extraction of Citrus Fruit Trees from a
Digital Surface Model Using a Unified Strategy:
Detection, Delineation, and Clustering
Ali Ozgun Ok and Asli Ozdarici-Ok
Abstract
In this study, we present an original unified strategy for
the precise extraction of individual citrus fruit trees from
single digital surface model (
DSM
) input data. A probabilistic
method combining the circular shape information with the
knowledge of the local maxima in the
DSM
has been used
for the detection of the candidate trees. An active contour is
applied within each detected region to extract the borders
of the objects. Thereafter, all extracted objects are seam-
lessly divided into clusters considering a new feature data set
formed by (1) the properties of trees, (2) planting parameters,
and (3) neighborhood relations. This original clustering stage
has led to two new contributions: (1) particular objects or
clustered structures having distinctive characters and rela-
tionships other than the citrus objects can be identified and
eliminated, and (2) the information revealed by clustering
can be used to recover missing citrus objects within and/or
nearby each cluster. The main finding of this research is that
a successful clustering can provide valuable input for iden-
tifying incorrect and missing information in terms of citrus
tree extraction. The proposed strategy is validated in eight
test sites selected from the northern part of Mersin province
of Turkey. The results achieved are also compared with the
state-of-the-art methods developed for tree extraction, and the
success of the proposed unified strategy is clearly highlighted.
Introduction
Citrus orchards play an important role i
the world. As reported by the Food and
tion (2017), in excess of 124 million ton
were cultivated in 2016. Today, various types of citrus trees
(e.g., lemon, orange, grapefruit, tangerine, etc.) are cultivated
on millions of acres in more than 50 countries. These statis-
tics clearly indicate that the development of efficient strate-
gies supplying reliable and up-to-date yield information is
essential for citrus products worldwide.
Remote sensing combined with image processing knowl-
edge plays a critical part in extracting individual trees and
their coverages (Ozdarici-Ok 2015). The techniques devel-
oped so far mainly prefer to handle mainly three-dimensional
point clouds either from
LiDAR
or dense image matching,
particularly to generate a digital surface model (
DSM
) that is
then converted to a normalized form (
nDSM
) or a crown height
model (
CHM
) (Zhen
et al.
2016). It is now a standard way to
produce high-quality point clouds with very high levels of de-
tail from unmanned aerial vehicle (
UAV
) image data sets with
high overlaps. In addition, the use of
UAVs
for civil purposes
is rapidly increasing since it offers an initial investment cost
at affordable levels. Therefore, we favor a dense surface model
generated from a
UAV
platform as a single input source.
In this study, we present a new unified strategy for the
extraction of individual citrus fruit trees (Figure 1). A proba-
bilistic method (Ok and Ozdarici-Ok 2018a) combining the
circular shape of trees with the knowledge of local maxima
(
LM
) in a
DSM
has been used for the detection stage. An active
contour method with a negative contraction bias was applied
considering each object region to extract the boundaries of the
detected objects (Ok and Ozdarici-Ok 2018b). Thereafter, all
(tree) objects are seamlessly divided into clusters consider-
ing (1) the properties of trees, (2) planting parameters, and
(3) neighborhood relations. An agglomerative hierarchical
Ali Ozgun Ok is with the Department of Geomatics
Engineering, Hacettepe University, 06800, Beytepe, Ankara,
Turkey
;
).
Asli Ozdarici-Ok is with the College of Land Registry, Ankara
H.B.V. University, 06450, Ankara, Turkey.
Photogrammetric Engineering & Remote Sensing
Vol. 86, No. 9, September 2020, pp. 557–569.
0099-1112/20/557–569
© 2020 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.86.9.557
Figure 1. The proposed unified strategy.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
September 2020
557
