© 2019 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future
media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or
redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.

Semantic Enrichment for Rooftop Modeling using
Aerial LiDAR Reflectance
Tan Tan
Department of Real Estate and
Construction
The University of Hong Kong
Hong Kong SAR
tan.tan.17@ucl.ac.uk

Ke Chen
School of Civil Engineering and
Mechanics
Huazhong University of Science and
Technology
Wuhan, China
chenkecm@hust.edu.cn

Weisheng Lu
Department of Real Estate and
Construction
The University of Hong Kong
Hong Kong SAR
wilsonlu@hku.hk

Fan Xue*
Department of Real Estate and
Construction
The University of Hong Kong
Hong Kong SAR
*Corresponding author, xuef@hku.hk
Abstract—As demanded by smart city applications, the
recognition and enrichment of urban semantics from
unstructured spatial big data became an emerging trend for
the development of building information model (BIM) and city
information model (CIM). Rooftop constructs the essential
part of BIM and CIM and loads various new application
practices and scenarios. The recognition and enrichment of
rooftop elements represent the trending requirements. This
study develops a new approach for semantic enrichment of
aerial Light Detection and Ranging (LiDAR) point clouds. In
this paper, machine learning models such as decision tree are
applied to predict green roof elements based on the geometry
and laser reflectance, and was validated in a pilot zone in the
main campus of The University of Hong Kong. The recognized
rooftop elements could provide a solid foundation for further
research, such as rooftop landscape, rooftop energy, rooftop
farming.
Keywords—Rooftop, building information model,
information model, LiDAR reflectance, decision tree

city

I. INTRODUCTION
In the era of smart cities, rooftops have gone far beyond
their traditional functions to increasingly become the focal
point for more intriguing uses. This is particularly in highdensity urban areas. For example, in Hong Kong, rooftops
are increasingly part of discussions about solar energy
farming, urban agriculture [1], community connection, and
open space [2]. These applications of rooftops could benefit
from the availability of digital rooftop models with rich
semantics. Semantically-rich rooftop models enable
visualization and simulation, based on which more informed
decision could be made in order to make use of rooftop
spaces. Therefore, the recognition and enrichment of
semantics, e.g., category, materials, functions, of rooftop
elements, is a trending theme.
Nevertheless, semantic enrichment for rooftop modeling
is still a challenging problem. Manually complementing and
modeling the rooftop details are extremely expensive,
tedious, and hampered by accessibility problems. Automatic
or semi-automatic object recognition of photos and 3D point
clouds, the rooftop elements can form the uppermost part of

978-1-7281-1708-9/19/$31.00 ©2019

the created BIM or CIM. However, the current enrichment
practice of semantics, especially those non-geometric
semantics, of rooftops, are unsatisfactory. More solid
methods should be developed for developing semanticallyrich rooftop models.
This study aims to utilize the reflectance, a physical
property that denotes how much of laser beam is reflected by
an object’s surface materials. Reflectance is also a standard
data property of every point in the Log ASCII Standard
(LAS) format of LiDAR (Light Detection and Ranging), for
rooftop element recognition and semantic enrichment. It
extends previous research work that reconstructs the
geometries of rooftop elements by using point clouds [2]. In
this paper, the laser reflectance is introduced as a key feature
for rooftop element classification and semantic enrichment.
Decision tree, a typical machine learning model, is applied to
classify the rooftop elements using laser reflectance and
geometric features (e.g., area and height of the rooftop
element).
The contributions made by this study are two-fold.
Firstly, this study suggests that reliable urban object
recognition requires non-geometric information beyond the
spatial data. Second, this study confirms that laser
reflectance can be an important data source for recognizing
and enriching the urban semantics of rooftop elements.
The remainder of this paper is organized as follows.
Section II reviews the related methods. Section III describes
our semantic enrichment approach. A pilot case is introduced
in Section IV, and the experimental results appear in Section
V. Conclusion is given at the end of the paper.
II. LITERATURE REVIEW
Many methods, including data clustering, region
growing, energy minimization (optimization), and model
fitting, have been developed to automate the recognition of
the semantics of building rooftops from 3D point clouds [3],
[4]. Reference [5] proposed a two-step method to segment
and form rooftop boundary from raw boundary data
extracted from point clouds. Reference [6] proposed an
enhanced probability density clustering algorithm to cluster

Fig. 1. The overall framework of the proposed approach

(a) Input LiDAR point cloud of 55,000m² pilot area (298,126
(b) 158 reconstructed rooftop elements using [2], where color
points), where color indicates the laser reflectance (warmer = less)
indicates average reflectance (warmer = less)
Fig. 2. The input LiDAR data and reconstructed rooftop elements with reflectance

the rooftop primitives by taking into account the topological
consistency among primitives. Then, it employed a new
Voronoi subgraph based algorithm to trace the primitive
boundaries seamlessly. Reference [7] provided an
interpolation of rooftop area in which multi-surfaces intersect
generating non-manifold points. Reference [8] proposed a
triangle geometry method to help reconstruct virtual roofs
using point cloud data. Reference [2] employed architectural
regularity to fine-tune the recognized identified rooftop
elements with the building footprint.
Generally, these methods not only are capable of creating
complex building models but also, they can be applied for
the simulation of any roof styles including curved and freely
formed roofs. Nevertheless, the quality of raw data
influences the model reconstruction extremely, such as the
sensitive impact from the point density, noise and outliers.
Besides, the non-geometric semantic enrichment from point
clouds is still challenging [9]. As a result, the semantics, e.g.,
category, materials, functions, in the reconstructed models,
are unsatisfactory.
III. OVERVIEW OF PROPOSED METHOD
This paper aims to utilize the LiDAR reflectance, and
apply a machine learning-based semantic enrichment for
rooftop modeling. The value of laser reflectance can often
the albedo approximately, because many aerial LiDAR uses
ultraviolet (e.g., 340nm), vis, or near infrared laser beams
[10]. Laser reflectance, as well as the derived albedo, offer
vital semantics about the surface, type of objects, and
possible functions. Fig. 1 shows the overall framework of the
proposed approach. The reflectance is an innate property of
each laser-measured point, e.g., the property “Intensity” is
reserved for reflectance in the LAS (ver. 1.4) format. By
associating the reconstructed rooftop elements and a small
set of training example, the machine learning models in the

proposed approach can predict new semantics, e.g., type of
green roof elements, materials, and functions, to enrich the
reconstructed geometric rooftop models to a semantically
rich one.
Decision tree is a popular machine learning model, and
easy to interpret to human experts [11]. Various decision tree
models have successfully been trained and applied in many
research areas such as radar signal classification, character
recognition, remote sensing, medical diagnosis, and expert
systems. The most significant feature of a decision tree is the
ability to decompose a complicated decision process into a
collection of a more straightforward decision. As shown in
Fig. 1, this study applies the decision tree to predict the
semantics via classification based on their geometric and
reflectance features.
IV. A PILOT CASE IN HONG KONG
This study selects an area of The University of Hong
Kong. Totally, ten buildings at the center of the campus area
were selected for the case. There are three main reasons for
the selection. Firstly, the selected area is small but resides
tens of various buildings, ranging from Edwardian style to
post-modern style, from low-rise to high-rise, and with the
year of built from the 1910s to 2010s. Secondly, the results
can be compared to previous studies, such as [2], [5], and [8].
The last concern was about privacy and data ownership.
Fig. 2a shows the input LiDAR point cloud of 298,126
points as a part of the aerial survey done by [12]. The
reflectance in Fig. 2a was measured between 400-2000 nm
wavelength range, where the warmer color indicates less
reflectance; Fig. 2b shows 158 small rooftop elements
reconstructed using [2], where the blue color shows a high
reflectance and the yellow color shows a low reflectance.
Based on the observation and news reports, three areas of

green roof elements (of 7 rooftop elements), were assigned to
the training examples. In Fig. 2, there was only one lowreflective roof, of which the building is the Main Building, a
built heritage (built in 1911) in Hong Kong.

final accurate results. It can be found that the target attribute
in the root node is the “non-green”; the first level condition is
“height ≥ −0.18m”; the second level is “area < 71m²”; the
third level is “height ≥ 0.96m”; the fourth level is “height

The task of semantic enrichment in this paper is thus to
classify the green roof elements from the rest reconstructed
geometric primitives. In addition, two specific types of green
roof elements are distinguished in the pilot case: (i) turf on
which vegetation covers the whole surface of the roof
elements, and (ii) potted area on which plants live in pots or
other small containers. Table 1 lists an excerpt of the
summarized data table of the pilot case for roof element
classification and semantic enrichment. The data columns in
Table 1 include average reflectance in percentage, the top
area of each element in square meter, height (top surface
above the registered building roof level in government’s
digital map) in meter.
TABLE I.

EXCERPT OF THE DATA TABLE OF THE PILOT CASE FOR
CLASSIFICATION AND ENRICHMENT

Label
Non-green
Non-green
Non-green
Non-green
Non-green
Non-green
Non-green
Non-green
Non-green
potted
turf
turf
turf

Avg. reflectance
(%)
54.5
53.6
36.7
34.6
50.8
29.5
30.5
33.5
28.1
35.1
54.9
53.7
50.4

Top area
(m²)
123.6
66.2
400.5
58.6
12.5
5.0
9.5
29.1
5.3
74.0
61.9
529.3
74. 4

height
(m)
2.47
2.39
3.53
3.52
2.84
0.80
0.72
0.63
0.72
0.35
–0.35
–0.34
–0.39

The computational experiments were conducted on a
Windows 10 desktop PC. The decision tree is offered by the
rpart package (ver. 4.1), which is freely available on the R
platform (ver. 3.4). The parameters of the decision tree are:
“min split = 2,” “max depth = 5,” and “min bucket = 1.”
V. EXPERIMENTAL RESULTS
5.1 The trained decision tree model
Fig. 3 shows the rooftop elements recognition processed
by the decision tree, which used less than 0.01s to get the

(a) Prediction results (dark green = turf, light green = potted)

Fig. 3. A decision tree model trained in less than 0.01s

≥1.4m”; and the fifth level is average “reflectance ≥ 64%”.
Overall, the “potted” elements are distinguished by selecting
attribute between “0.96m ≤ height < 1.4m” and “area ≥
71m².” There are two rules embedded in Fig. 3 that lead to
“turf” elements. One rule is height < −0.18m; the other is
“height ≥ 1.4m” and “area > 71m²” and “average reflectance
≥ 64%”.
5.2 Semantic enrichment of green roof elements
Fig. 4 shows the comparison of predicted green roofs and
mesh models on Google Maps. The screenshot of the mesh
models on Google Maps illustrates the real distribution of
rooftop elements. The green areas, located mainly on top of
three buildings, are corresponded with the predicted green
roof elements. The semantic enrichment of green roof from
reflectance makes up the shortcoming of incomputable
Google’s mesh model even though they have details. The
comparison result shows the expecting accuracy for the
elements detection and division.

(b) Screenshot of the mesh models on Google Maps

Fig. 4. Comparison of predicted green roofs and the mesh models on Google Maps

5.3 Discussion
In comparison to previous results such as [3] and [5-8],
the proposed method has several advantages. Firstly, it
overcomes the drawback of the low quality of raw data, such
as the sensitive impact from the local point density, noise and
outliers. Second, it introduces laser reflectance as a key
factor for rooftop elements recognition. As a result, the
semantics, e.g., category, materials, functions, in the
automatically reconstructed BIM/CIM would be more
satisfied than traditional methods.
However, there also are some limitations in this paper.
This study used five-level attributes and limited data for the
small area detection, but this division technique and limited
scale of training data might not accurate enough when
applying it to city-scale rooftop elements recognition. The
sample data of the training may, therefore, require manual
debugging as the test area increases. In addition, the range of
detected green areas is mainly divided by rectangles.
However, many roof green areas may exhibit a distribution
of irregular patterns.
VI. CONCLUSION
A new criterion to rooftop elements recognition from
LiDAR data is proposed. The statistical fluctuations of the
laser reflectance, which roughly reflects the reflectance, is
proven successful for recognizing non-geometric
characteristics of rooftop elements. This study contributes to
the importance of non-geometric information regarding
reliable urban object recognition and confirms that laser
reflectance can be an important data source for recognizing
and enriching the urban semantics of rooftop elements. This
proposed method can overcome the shortcomings of
traditional recognition methods and increases detection
accuracy. In addition, the method overcomes the drawback
of the quality of raw data, such as the sensitive impact from
the local point density, noise and outliers. As a result, the
semantics, e.g., category, materials, functions, in the
automatically reconstructed BIM/CIM would be more
satisfied than traditional methods. The proposed utilization
of laser reflectance can also be helpful in BIM/CIM research
for indoor scenes and city areas. Further research could focus
on the widening of technique application scenarios, semantic
evidences such as architectural symmetry [13], and the
increasing of recognition accuracy in large-scale areas.

ACKNOWLEDGMENT
The work presented in this paper was supported by the
Hong Kong Research Grant Council (No. 17200218) and
The University of Hong Kong (Nos. 201711159016,
102009741).
REFERENCES
[1]

F. Orsini, D. Gasperi, L. Marchetti, C. Piovene, S. Draghetti, S.
Ramazzotti, G. Bazzocchi, and G. Gianquinto, “Exploring the
production capacity of rooftop gardens (RTGs) in urban agriculture:
the potential impact on food and nutrition security, biodiversity and
other ecosystem services in the city of Bologna,” Food Security, vol.
6, no. 6, pp. 781-792, 2014.
[2] K. Chen, W. Lu, F. Xue, P. Tang, and L. H. Li, “Automatic building
information model reconstruction in high-density urban areas:
Augmenting multi-source data with architectural knowledge,”
Automation in Construction, vol. 93, pp. 22-34, 2018.
[3] J. Yan, J. Shan, and W. Jiang, “A global optimization approach to
roof segmentation from airborne lidar point clouds,” ISPRS journal of
photogrammetry and remote sensing, vol. 94, pp. 183-193, 2014.
[4] F. Xue, W. Lu, and K. Chen, “Automatic Generation of Semantically
Rich As-Built Building Information Models Using 2D Images: A
Derivative-Free Optimization Approach,” Computer-Aided Civil and
Infrastructure Engineering, vol. 3, no. 11, pp. 926-942, 2018.
[5] Y.-H. Tseng, and H.-C. Hung, “Extraction of Building Boundary
Lines from Airborne LIDAR Point Clouds,” ISPRS-International
Archives of the Photogrammetry, Remote Sensing and Spatial
Information Sciences, vol. XLI-B3, pp. 957-962, 2016.
[6] D. Chen, R. Wang, and J. Peethambaran, “Topologically aware
building rooftop reconstruction from airborne laser scanning point
clouds,” IEEE Transactions on Geoscience and Remote Sensing, vol.
55, no. 12, pp. 7032-7052, 2017.
[7] S. A. N. Gilani, M. Awrangjeb, and G. Lu, “Segmentation of airborne
point cloud data for automatic building roof extraction,” GIScience &
Remote Sensing, vol. 55, no. 1, pp. 63-89, 2018.
[8] L. Cheng, L. Tong, W. Zhao, Y. Liu, and M. Li, “Dynamic triangle—
Based method for 3D building rooftop reconstruction from LiDAR
data,”, 19th IEEE International Conference on Geoinformatics. IEEE,
2011.
[9] Y. Sun, C. Wang, J. Li, Z. Zhang, D. Zai, P. Huang, and C. Wen,
“Automated segmentation of LiDAR point clouds for building rooftop
extraction,” 2016 IEEE International Geoscience and Remote Sensing
Symposium (IGARSS). IEEE, pp.1472-1475, 2016.
[10] R. Levinson, S. Chen, P. Berdahl, P. Rosado, and L. A. Medina,
“Reflectometer measurement of roofing aggregate albedo,” Solar
Energy, vol. 100, pp. 159-171, 2014.
[11] S. R. Safavian, and D. Landgrebe, “A survey of decision tree
classifier methodology,” IEEE transactions on systems, man, and
cybernetics, vol. 21, no. 3, pp. 660-674, 1991.
[12] CEDD. “The CEDD 2010 LiDAR Survey,” private communication,
Hong Kong: Civil Engineering and Development Department, 2015.
[13] F. Xue, W. Lu, C. J. Webster, and K. Chen, “A derivative-free
optimization-based approach for detecting architectural symmetries
from 3D point clouds,” ISPRS Journal of Photogrammetry and
Remote Sensing, vol. 148, pp. 32-40, 2019.