Informatica 41 (2017) 379–380 379
Classification of Vegetation in Aerial LiDAR Data
Denis Horvat
Faculty of Electrical Engineering and Computer Science, University of Maribor, Slovenia
E-mail: denis.horvat@um.si, Web: https://gemma.feri.um.si/
Thesis summary
Keywords: LiDAR, classification, vegetation, mathematical morphology, surface fitting
Received: April 12, 2017
This contribution summarises a doctoral dissertation which proposes an algorithm for the classification of
vegetation points in aerial LiDAR data. The algorithm characterizes vegetated areas based on statistically
large dispersion in elevations of points, and the context in which the points are located. The algorithm is
able to classify vegetation in both rural and urban areas with an average F1 score of 97.9% and 91.0%,
respectively. The point-clouds can contain different types of vegetation and various degrees of canopy
densities.
Povzetek: V predlaganem prispevku povzamemo doktorsko disertacijo, ki predlaga algoritem za klasi-
fikacijo točk vegetacije iz podatkov LiDAR. Algoritem ovrednoti območja vegetacije na podlagi statistično
visoke razpršenosti višin točk v kombinacji s kontekstom v katerem se točke nahajajo. Algoritem klasifi-
cira točke vegetacije v urbanih in neurbanih področjih s 97% in 91% povprečnim rezultatom F1. Oblaki
točk lahko vsebujejo različne tipe vegetetacije z različno gostoto olistanosti.
1 Introduction
The potential of the data obtained by the aerial LiDAR
(Light Detection and Ranging) systems has been utilised
increasingly by a variety of scientific and industrial appli-
cations. Data acquisition using LiDAR produces a point-
cloud, in which the individual points are calculated based
on the time delay between the emitted and detected laser
beam. While the obtained point-clouds from a plane-
mounted LiDAR can represent the underlying Earth’s sur-
face accurately, the entity to which a given point belongs
(e.g. ground, building, vegetation) is not known. This con-
textual knowledge is crucial for a variety of applications,
such as environmental simulations, urban planning, or the
generation of a canopy height model. Thus, preprocess-
ing, using a classification algorithm is usually applied, in
which each point is correlated with one of the predeter-
mined classes.
This paper is a summary of a PhD thesis [1] (and the
corresponding paper [4]), which proposed an algorithm for
classification of vegetation points within LiDAR data. Veg-
etation can be particularly challenging to identify, as the
classification model should be universal enough to cover
the various sizes, shapes and canopy densities of differ-
ent vegetation but, at the same time, still differentiate it
from other surface objects (e.g. houses, cars, fences). The
following section outlines the proposed classification algo-
rithm, while section 3 evaluates it. Section 4 concludes the
paper.
2 Classification algorithm
Clusters of points that represent vegetation are, in most
cases, defined by statistically large dispersions in eleva-
tion. This is caused by the vegetation’s non-linear shape
and porosity, as the laser beam can usually penetrate the
canopy and capture many points within, or even under, the
vegetation. The mentioned properties can be character-
ized efficiently by modifying the LoFS (Local Fitting of
Surfaces) method [2]. Namely, by locally fitting planes on
the LiDAR-derived surface and evaluating the fitting error
using all points in the fitted area, a distinction can be made
between vegetation and most of other man-made objects.
Larger fitting errors are expected in the former case, while
the latter ones usually produce errors that are identifiably
smaller.
The remaining non-vegetation that does not conform to
this definition (i.e. also produces a larger fitting error) is
handled using contextual analysis which defines: 1) At-
tached objects 2) Overgrowing vegetation and 3) Small ob-
jects. Attached objects represents a transition between ar-
eas (e.g. a wall, balcony or chimney). Such objects can
be identified (and subsequently removed) using spatially-
variant morphological dilation [3] where all non-ground
areas with small fitting errors are dilated. The extent of
the dilation is controlled locally, using a structuring ele-
ment with the radius equal to the distance from the near-
est ground. Spatially-variant dilation is used similarly on
areas with high fitting errors to identify overgrown vegeta-
tion. However, the radius, in this case, is dependent on the
height difference between a given area and the nearest non-
380 Informatica 41 (2017) 379–380 D. Horvat
ground area with a small fitting error. Lastly, small objects
are removed using connected operators. The final result
that includes the described fitting error evaluation and con-
textual analysis is then mapped onto the individual LiDAR
points to get the classification.
3 Results
The algorithm was tested on multiple rural and urban
datasets which contained different types of vegetation and
degrees of canopy densities. The results were evaluated by
counting the false positives, false negatives, true positives
and true negatives which served as an input for the calcula-
tion of the well-established F1 score. An average F1 score
of 97.9% was achieved for rural and 91.0% for urban en-
vironments which, because of the more complex geometry,
tend to be more difficult to classify. The use of contextual
analysis improves the results in urban areas significantly.
Namely, by removing the attached object, small object and
handling overgrown vegetation, the F1 score is improved,
on average, by 8.8%, 1.1% and 1.0%, respectively.
4 Conclusion
The PhD thesis presented an algorithm for the classification
of LiDAR data, which classifies vegetated areas with high
accuracy in characteristically different point-clouds (rural
environment, urban environment, different types of vege-
tation, various leaf-on conditions). Additionally, the al-
gorithm can be used in most classification scenarios that
contain aerial point-clouds, as it relies only on geometrical
features of the point-cloud and the subsequent contextual
analysis.
Acknowledgement
The authors acknowledge the financial support from the
Slovenian Research Agency (research core funding No.
P2-0041 and project J2-6764)
References
[1] D. Horvat (2017) Algorithm for classification of veg-
etation in LiDAR data, Doctoral dissertation, Fac-
ulty of Electrical Engineering and Computer Science,
University of Maribor (in Slovene).
[2] D. Mongus, N. Lukač and B. Žalik (2014) Ground
and building extraction from LiDAR data based on
differential morphological profiles and locally fitted
surfaces, ISPRS Journal of Photogrammetry and Re-
mote Sensing, Vol. 93, pp. 145–156.
[3] N. Bouaynaya and Mohammed Charif-Chefchaouni
(2008) Theoretical Foundations of Spatially-Variant
Mathematical Morphology Part I: Binary Images,
IEEE Transactions on Pattern Analysis and Machine
Intelligence, Vol. 30, No. 5, pp. 145–156.
[4] D. Horvat, B. Žalik and D. Mongus (2016) Context-
dependent detection of non-linearly distributed points
for vegetation classification in airborne LiDAR, IS-
PRS Journal of Photogrammetry and Remote Sens-
ing, Vol. 116, pp. 1–14.