Multi-perspective assessment of greenery visibility from streets,
windows, and drones using deep learning and CIM

Siyuan Menga , Maosu Lib* , Jinfeng Xiec , Fan Xued
a Department of Real Estate and Construction, The University of Hong Kong, Pokfulam,

Hong Kong, China;
b Thrust of Urban Governance and Design, The Hong Kong University of Science and
Technology (Guangzhou), Nansha, Guangzhou, China;
c Department of Urban Planning and Design, The University of Hong Kong, Pokfulam, Hong
Kong, China;
d Department of Real Estate and Construction, The University of Hong Kong, Pokfulam,
Hong Kong, China.
National Center of Technology Innovation for Digital Construction Hong Kong Branch, The
University of Hong Kong, Pokfulam, Hong Kong, China

ARTICLE HISTORY
Compiled May 14, 2026

ABSTRACT
Urban greenery includes 3D volumetric greenery and provides visual benets to
billions of urban dwellers. Although high-rise buildings and drones are expanding
urban activities vertically, most greenery studies are limited to street-level observations, relying on 2D/2.5D view imagery. This paper presents a greenery volume
visibility (GVV) assessment method by integrating 3D deep learning and city information model (CIM). A 3D GVV index (GV V I ) is dened for three groups of
observing perspectives: street, window, and drone. Semantic segmentation using
deep learning detects greenery volumes in photorealistic 3D models, then GV V I s
are calculated across all perspective groups. Hot spot maps and disparity analysis
comparing the three sub-indices reveal the urban greenery volumes overlooked in traditional ground-level studies. Experimental results on a 4.92 km2 area of Kowloon,
Hong Kong, showcased the proposed method's eciency (23,490 GV V I s in 256.45
hours); Comparative results between street and drone/window perspectives revealed
37.6% and 31.7% greenery entirely overlooked in traditional ground-level spatial
analysis. This paper contributes in three aspects: a 3D GVV index denition, an
ecient deep-learning-based geospatial analysis method, and novel quantitative 3D
evidence for informed urban greenery planning and maintenance, optimal building
designs, and low-altitude drone route planning.

KEYWORDS

Urban green volumes; Visibility assessment; Deep learning; City information
model; Greenery planning and maintenance

Word Count: 6210
* Corresponding author. Email: maosuli@hkust-gz.edu.cn

This is the peer-reviewed author's version of the paper:
Meng, S., Li, M., Xie, J., & Xue, F. (2026). Multi-perspective assessment of greenery visibility from streets, windows,
and drones using deep learning and CIM. International Journal of Geographical Information Science, (in press).
DOI: https://doi.org/10.1080/13658816.2026.2667807

1

2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49

1. Introduction
Urban greenery encompasses vegetation in green spaces within urban agglomerations,
such as woodlands, parks, gardens, streets, and squares (Konijnendijk et al. 2006).
Urban greenery provides a range of amenity values, e.g., leisure opportunities and aesthetic enjoyment for residents (Kong et al. 2007). Viewing urban greenery improves
human health, well-being, life contentment, and productivity (Ulrich 1984, Jiang et al.
2021), whether from walking on the street or from homes (Wang et al. 2022, Elsadek
et al. 2020, Li et al. 2026). The value of dierent greenery types is associated with
economic factors, such as housing prices (Panduro and Veie 2013). Spatial planning
of green spaces integrated with greenery impact assessments (e.g., noise, air pollution, visual amenity value) on residents can enhance their life satisfaction (Neema and
Ohgai 2010). Therefore, measuring greenery visibility supports evidence-based decisionmaking in maintaining urban green spaces, improving city image, and revitalizing city
blocks.
Urban greenery visibility is usually assessed at the street level, whereas new perspectives have recently emerged for its assessment. The expanding urban activities
include drone-view-based activities and vertical living. Residents in high-rise, highdensity cities often experience less nature at street level but benet more from green
views observable from high-rise windows (Elsadek et al. 2020, Ko et al. 2020). Dronelevel views oer a valuable means to capture and assess views from above-ground
greenery visibility. They are increasingly used for collecting greenery information, exploring virtual environments, and experiencing urban sightseeing (Ecke et al. 2022,
Perperidou and Kirgianis 2022, Zhang et al. 2024). However, a greenery visibility assessment method, including street, window, and drone perspectives, is limited in the
current literature.
Urban greenery visibility is often imbalanced in terms of spatial distribution (Cimburova et al. 2023), especially in high-rise, high-density cities. Given the importance of
viewing greenery for comfort and general well-being, a quantitative assessment of this
uneven spatial distribution from dierent perspectives is necessary. This assessment
can further support applications such as greenery planning and maintenance, optimal
building design, and low-altitude drone route planning.
Traditional methods operate in 2D or 2.5D, neglecting the 3D nature of GVV. Manual assessment of GVV remains labor-intensive and cost-prohibitive. 2D and 2.5D data,
such as street view imagery (SVI) and Digital Surface Models (DSM), were applied
to estimate the perception of greenery from a specic observation point, i.e., green
view index, or to assess the GVV (Yang et al. 2020, Cimburova et al. 2023). However, current methods focus on a single perspective, usually at street level, neglecting
the multi-perspective GVV assessment. Besides, the 2D-based method failed to locate
specic greenery, and viewshed-based methods using 2.5D data may be insucient to
accurately represent the real physical environment and often require substantial computational costs due to the need for stacked computations for each GVV-observation
pair.
The emerging deep learning and CIM provide new opportunities for assessing 3D
GVV. CIM, which originates from digitalized systems of urban elements and environment (Xue et al. 2021), integrates with photorealistic urban mesh models, virtual
cameras from any perspective, and 3D rendering techniques for consistent color representation (Li et al. 2022, Liang et al. 2017). Deep learning has been widely employed
to address many geographical issues, from semantic segmentation of SVI to 3D photorealistic mesh models, such as Stratied Transformer (Biljecki and Ito 2021, Lai et al.
2

50
51
52
53
54
55
56
57
58
59
60

2022, Li et al. 2025). The integration of CIM and deep learning enables the ecient
computation of GVV from any perspective.
This paper, therefore, introduces a novel GVV assessment method using deep learning and CIM (Figure 1). The main focus includes assessing visibility from building
windows for urban living, low-altitude drones, and traditional pedestrian perspectives.
Firstly, the greenery volume visibility index (GV V I ) is dened. Secondly, a 3D semantic segmentation model detects dierent greenery volumes from photorealistic mesh
models, and the Greenery Volume Layer (GVL) with a grid-level color index database
is displayed in 3D CIM. Afterward, the GV V I of each volume is calculated using street,
window, and drone-view images. Finally, the `hot spots' map and disparities of GV V I
and three sub-indices are analyzed.
Stage 1

Stage 2

For individual greenery volume Stage 3
GVVID

𝜸2

Deep learning

DVI

Grid color index

WVI

… GVVIW

𝜸1

SVI

CIM platform

GVL generation

View capture

𝜸1 𝜸2 𝜸n

GVVIS

GVVI

Figure 1.: Conceptual illustration of GVV assessment

67

The contribution of this paper is threefold. First, a metric GV V I is dened to
measure GVV types, including street, park, and garden, from dierent perspectives
(street, window, and drone levels). Second, an automatic three-step assessment method
that is eective and ecient, utilizing deep learning and CIM. Third, GV V I hot spots
and disparities highlight new quantitative 3D evidence for informing urban greenery
planning and maintenance, optimizing building designs, and planning sightseeing 3D
routes for low-altitude drones.

68

2. Literature Review

69

2.1. Traditional methods to measure greenery visibility

61
62
63
64
65
66

70
71
72
73
74
75
76
77
78
79
80
81
82
83
84

Greenery valuation methods established the monetary value of trees, including individual tree conditions and amenity functions such as visual contribution (Helliwell
2008, Roberto Barbosa et al. 2010, Doick et al. 2018, Cimburova et al. 2023). The
amenity value of greenery refers to the characteristics that contribute to the aesthetic
appreciation of people, i.e., the visibility of greenery. Instead of the green view index,
which assesses the residents' greenery exposure degree at street/building level (Li et al.
2023b), greenery visibility highlights the visual contribution of greenery to surrounding
residents (Cimburova et al. 2023). Thus, assessing the visibility of greenery is key to
informed decision-making related to greenery maintenance, design, and renovation.
Greenery assessment method is transitioning from traditional overhead views to
relying more on human-centric visibility data sources. The Normalized Dierence Vegetation Index (NDVI) from satellite imagery eectively quanties large-scale greenery
(Huang et al. 2021), but its urban application is limited by the lack of vertical greenery data and low resolution (Yu et al. 2022). Researchers, therefore, have turned to
eye-level data, such as SVI, to assess the visual quality and amenity value of greenery
3

109

(Li et al. 2015, Yu et al. 2016, Cimburova et al. 2023).
Methods for assessing eye-level green view include eld surveys, image-based analysis, and geospatial modeling (Yu et al. 2016, Wang et al. 2023). Since eld surveys
are labor-intensive, requiring excessive time to collect data (Doick et al. 2018), imagebased methods, such as SVI and CIM-based window view image (WVI), oer eye-level
visual evidence (Li et al. 2015, Liu et al. 2023, Xie et al. 2025). However, the method
is proven to have both building and greenery coverage bias (Fan et al. 2025, Huang
et al. 2025). For example, SVI only provided information at photographic observation
points and failed to consider the greenery within large green spaces (e.g., gardens and
parks) (Huang et al. 2025). The issues of availability and data quality (e.g., camera
pose variation, weather conditions, blurriness) have been reported in previous literature (Rui and Cheng 2023, Tang et al. 2023). Furthermore, the current CIM-based
methods, which extract green view through 2D deep learning, are unable to locate
greenery in images (Li et al. 2024). Thus, the image-based methods suer from quality, reliability, and coverage limitations, apart from locating greenery, which together
make them unsuitable for GVV assessment.
Several studies have employed the viewshed-based method to simulate greenery visibility using 3D data sources, such as DSM and point clouds. For instance, greenery
assessments of street-level exposure and tree-level visibility were developed (Cimburova
and Blumentrath 2022, Cimburova et al. 2023, Tang et al. 2023, Pyka et al. 2022).
While viewshed simulations oer a human perspective in 3D, the accuracy of occlusion surfaces derived from 2.5D data sources is limited in reecting the real physical
environment and require large amount of computation cost (Cimburova and Blumentrath 2022). Besides, the acquisition of point clouds for a large-scale GVV assessment
is costly (Tang et al. 2023).

110

2.2. Recent advances in CIM and deep learning

85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108

111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131

A 3D CIM is a digital representation of the physical and functional attributes of a city,
serving as a collaborative knowledge repository (Song et al. 2017, Xue et al. 2021). CIMbased research has demonstrated its ability to simulate human perspectives from the
building level by leveraging its photorealistic texture and high-performance capability.
Therefore, photorealistic mesh models demonstrate certain advantages in representing
the 3D greenery volumes compared with traditional data sources.
The high-performance capability of CIM makes it feasible for quick positioning and
capturing virtual views from any location. Specically, WVI was generated to assess
dierent urban view types using deep learning, quantify eye-level green view from
upper-oor perspectives, and measure window view distance (Li et al. 2022, 2023b,
2025). Moreover, the level of detail 2 CityGML-based model was employed to generate
WVI and identify visible greenery (Bolte et al. 2024). Therefore, real-time greenery
observation from higher elevation is promising, thanks to increasingly accurate and
realistic urban environment datasets worldwide (Yu and Gong 2012, HKPlanD 2019,
HKLandsD 2022).
The advancing deep learning facilitates novel approaches for extracting data on urban greenery. Deep learning enables computational models with multiple processing
layers to learn multiple-level abstracted representations of data (LeCun et al. 2015).
Previous research achieved semantic segmentation in 2D (e.g., SVI) and 3D (e.g.,
point cloud) urban-level data (Biljecki and Ito 2021, Li et al. 2025). Multi-view-based,
volumetric-based, and point-based methods have been developed to segment 3D point
4

139

clouds (Guo et al. 2021), such as Stratied Transformer (Li et al. 2023a). Thus, the
Stratied Transformer can attach the photorealistic models with semantic information.
In summary, GVV is warranted in its own right, but also critical for evaluating its
impact on landscape and urban studies. Current literature lacks a quantitative assessment method of GVV from non-traditional perspectives, such as drones and windows.
3D CIM represents a promising data source providing realistic views from drone, window, and street level observations. Deep learning, meanwhile, can be incorporated for
the semantic segmentation of a CIM and in the analysis of greenery volumes.

140

3. GVV assessment method

132
133
134
135
136
137
138

141
142
143
144
145
146

Figure 2 shows an overview of the three-stage GVV assessment method in this paper.
In the rst stage, the method applies 3D deep learning to the photorealistic mesh
model to create a GVL of multi-class greenery. Then, street, window, and drone views
are captured from the GVL-enriched CIM using road networks, windows on building
footprints, and drone positions above DTM. Finally, the output GV V I values are
calculated and analyzed for urban greenery maintenance and other purposes.
Stratified
Transformer

CIM platform
Cesium

§ 3.3 Stage 1

Getis-Ord Gi* Spearman

Semantic segmentation

Grid color index
Photorealistic
mesh model
Road network
Bldg footprint
DTM

CIM with GVL

CIM with GVL

§ 3.4 Stage 2
SVIs generation

WVIs generation

DVIs generation

𝐺𝑉𝑉𝐼𝑊

𝐺𝑉𝑉𝐼𝐷

SVI
WVI
DVI

§ 3.5 Stage 3
𝐺𝑉𝑉𝐼𝑆

Linear weighted 𝐺𝑉𝑉𝐼
Land use map

Hot spots analysis

Input

Correlation analysis
Proposed method

Differentials analysis

GVVI DTM and
analysis
Output

Figure 2.: Flow chart of the proposed green volume visibility assessment method

147
148
149
150
151

3.1. GV V I denition
γ
γ
For a given greenery volume γ ∈ Γ, the GV V I γ = [GV V ISγ , GV V IW
, GV V ID
] is a
vector of three sub-indices. These sub-indices represent the aggregated visibilities of
γ -th volume from three distinct observing perspectives: streets (S), windows (W), and
drone positions (D), respectively. For each sub-index i ∈ {S, W, D}, Equation 1 shows

5

152

the calculation for the total number of visible pixels (Niγ ) in all views (Vi ) in group i:

Niγ = Σv∈Vi ||pixels(v, γ)||,
153
154
155
156

where pixels represents the function that returns all the pixel set of γ -th volume in
the view v , and || · || returns the cardinality of the set. Which means that, the pixels
of the γ -th volume will be counted v times in each sub-index i, to evaluate GVV from
all observation points. The GV V Iiγ is the normalized number of pixels in Equation 2.

GV V Iiγ =
157
158
159
160
161
162

(1)

Niγ − Nmin
Niγ − minj∈Γ (Nij )
=
.
Nmax − Nmin
maxj∈Γ (Nij ) − minj∈Γ (Nij )

(2)

Each i-th sub-index GV V Iiγ
∈
[0, 1], and the vector GV V I γ
=
γ
γ
γ
3
[GV V IS , GV V IW , GV V ID ] ∈ [0, 1] is well bounded value in the unit cube. Therefore,
GV V I values of all green volumes Γ are a list of vectors [GV V I 1 , GV V I 2 , . . . , GV V I J ],
where J = ||Γ|| is the cardinality of the greenery volumes.
A weighted index GV V I aggregates the three sub-indices using observingperspective-specic weights, as depicted in Equation 3.
γ

GV V I =

γ
γ
+ wD · GV V ID
wS · GV V ISγ + wW · GV V IW
,
wS + wW + wD

(3)

164

where wS , wW , and wD are non-negative weights denoting the relative weights of each
group of observing perspectives.

165

3.2. Stage 1: Greenery volume layer (GVL)

163

166
167
168

In this stage, all grid-level greenery volumes are detected and highlighted within the
input 3D CIM. This stage includes semantic segmentation, colorized indexing, and
integration with the 3D CIM platform.

Stratified
Transformer

Grid color index
database

Road
Terrain
Water
Greenery
Building
Car
Facility
Photorealistic mesh Sampled point cloud

Input:

Segmented point cloud

Semantic segmentation

CIM platform
Cesium

Tiles
Agg.
Colored greenery mesh

GVL

Indexing GVV

Output:

Figure 3.: Creation of greenery volume layer (GVL) using 3D deep learning
169
170
171
172
173

Firstly, an urban point cloud is sampled on the photorealistic mesh model at a
density of 10 points/m2 . Then, the Stratied Transformer is utilized for the 3D semantic
segmentation. The Stratied Transformer is a deep learning model capable of capturing
long-range contexts, showcasing robust generalization capabilities, and achieving high
performance (Lai et al. 2022). This model has proven to be eective in segmenting
6

193

large-scale urban point clouds (Li et al. 2023a). The HRHD-HK of high-rise, highdensity cities serves as the training dataset, encompassing seven classes (i.e., road,
terrain, water, greenery, building, car, and facility), where the greenery class refers to
vegetation such as trees, bushes, and grass. etc, as listed by Li et al. (2023a). The
training is performed for 500 epochs with a batch size of 1 to balance computational
cost and accuracy, while all other parameters follow the default settings of the Stratied
Transformer. Semantic segmentation is validated using mean intersection over union
(mIoU). One annotator manually assigned ground-truth labels for greenery to a random
sample comprising 10% of the point cloud tiles in the study area. The selected tiles
cover greenery within hillsides, parks, gardens, and streets that typically occur in highdensity cities.
We assign a grid index to the greenery area based on the segmented point cloud.
The size of the grid is set to 10m, corresponding to the average tree diameter in Hong
Kong (Kong et al. 2017). The photorealistic mesh model is segmented based on the
greenery point cloud and the grid index; simultaneously, a unique and opaque color,
serving as an indexing ID, is randomly attached to each greenery grid mesh. The
non-greenery mesh is colored white. The segmented mesh is converted into 3D tiles
via CesiumLab and then displayed on the CIM Cesium platform. The colored mesh
models are dierentiated by their tiles. The nal output of this stage is the city-scale
3D CIM with GVL, as shown in Figure 3.

194

3.3. Stage 2: Automatic views capture and indices generation

174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192

195
196

All three groups of views, i.e., SVIs, WVIs, and DVIs, are captured based on the
city-scale GVL enriched CIM and the sample points generated in this stage (Figure 4).
GVL
120°

60°

Grid color index database

180°

240°
0°
1.65m 300°

Road network

SVI

𝐺𝑉𝑉𝐼𝑆

(a)

𝐺𝑉𝑉𝐼𝑊

(b)

𝐺𝑉𝑉𝐼𝐷

(c)

Eqs 1 - 2
Bldg footprint

WVI
100m

120°
60°

180°

0°

240°
300°
50m

DTM

Detecting visible
greenery volume
pixels (1,2....γ)

300m
limit.

50m
buffer

Sample observer positions

DVI
Views capture Color decoding Calculation

GVVI

Figure 4.: The process of capturing views. (a) SVI; (b) WVI; and (c) DVI
197
198
199
200

The sample points of SVIs are generated based on the road network to simulate
both pedestrian and car eye-level perception simultaneously. We sample points at an
interval of 30m, which is the average street length in the research area (Gong et al.
2018). For each point, the z value is determined by adding 1.65m to the DTM (average
7

232

height of adults in Hong Kong) (NCD Risk Factor Collaboration (NCD-RisC) 2016).
The heading of the virtual camera is classied into six directions (0◦ , 60◦ , 120◦ , 180◦ ,
240◦ , 300◦ ), with a 60-degree eld of view (FOV) (Meng and Zheng 2023).
We generate the sample points of WVIs based on the building footprints. The midpoint of each simplied footprint edge is used as the sampling point. Three height
levels-low, medium, and high-are sampled to simulate the window view from dierent
building oors. Subsequently, a virtual camera with a 60-degree FOV is positioned at
specied locations on the CIMs to capture the exterior window views, according to Li
et al. (2022, 2023b).
The sample points of DVIs are generated based on the DTM and the 3D building
footprints. Firstly, the sample points' altitudes are set to mark a 50m buer away from
buildings and terrain, because of the noise and privacy concerns of the residents. The
upper limit of the altitude of DVI is set according to the local airspace classication.
DVI was captured at 50 m vertical intervals, recording visibility changes caused
by altitude while balancing computational cost and precision (Seifert et al. 2019). In
addition, sampling is performed at intervals of 100m in the horizontal direction because
of the high-ying speed of the drone, contrary to vehicles on the ground (Xiang et al.
2024). The virtual camera is oriented in six directions (0◦ , 60◦ , 120◦ , 180◦ , 240◦ , and
300◦ ) for capturing images. A 60-degree FOV is used to represent the central visual
eld of a human (Meng and Zheng 2023, Tara et al. 2021).
Mesh penetration artifacts occur in some images, particularly at sampling points
close to the mesh surface and in regions where mesh fusion occurs (e.g., greenery and
buildings) at the bottom of the photorealistic model. This leads to inaccuracies in the
display of greenery volumes and, consequently, miscalculations of GVV. Hence, the
images with model penetration are ltered using a Support Vector Machine based on
the image deep feature (totaling 2,048 features) extracted by Inception V3 in Orange
(version 3.36) (Szegedy et al. 2016). We used 400 views (200 true, 200 false), with 320
for training and 80 for testing, as the dataset.
Finally, we calculate the GV V IS , GV V IW , GV V ID , and GV V I using an automated
Python script, based on captured and ltered views for analysis. Each color captured in
the views represents the visible pixels of a unique GVV from the observing perspectives.
The results are calculated using Equations 1, 2, and 3 with equal weight combination.

233

3.4. Stage 3: GV V I analysis

201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231

234
235
236
237
238
239
240
241
242
243
244
245
246

The Getis-Ord Gi∗ analysis within ArcGIS Pro is utilized to reveal the spatial cluster
locations of high and low values, namely the greenery hot spot map. Each of the
three sub-indices is visualized and analyzed using the greenery hot spot maps. In the
statistical analysis, volumes within a 100m buer of the research area boundary are
excluded due to insucient observations. A sensitivity analysis of weight adjustment
is conducted to evaluate the impact of weight combinations on the distribution of
greenery hot spots. Spearman's coecient was used to analyze the correlations and
the independence of indices between the three sub-indices of GV V I and NDVI data
from 2018.
The dierentials between street and drones, street and windows, are calculated, as
shown in Equation 4, where i ∈ {W, D}. Following this, hot spot maps of dierentials
are computed. The identied hot spots are extracted to demonstrate the total number
of identied dierential hot spots.

8

GV V Ii−S = GV V Ii − GV V IS = [GV V Ii1 − GV V IS1 , GV V Ii2 − GV V IS2 , . . . , GV V IiJ − GV V ISJ ]

(4)

252

Greenery distribution across dierent land-use types has been shown to impact human well-being (Bahr 2024). Therefore, assessing the association between GV V I dierentials and land use type can further support such applications. The two dierentials
GV V ID−S and GV V IW−S indicate the extent to which the proposed GV V I values
could dierentiate themselves from the traditional studies based on land use type at
street and ground levels.

253

4. Experimental setting

254

4.1. Study area

247
248
249
250
251

263

The study area, located in Kowloon, Hong Kong, spans 4.92 km2 with an elevation
range of approximately 100m (Figure 5). The study area is a well-developed urban area
with small variations in buildings over the past 10 years, totaling 150 new buildings
HKBD (2023). It comprises 6,294 buildings, with an average height of 25.6m (the lowest
is 1.5m, and the highest is 229m). Natural landscapes (such as parks and gardens) are
distributed in dierent locations in the study area, such as hillsides, residential areas,
and along streets. Because of the line-of-sight blockage caused by terrain elevation
changes and buildings, the amenity value of greenery to individuals varies considerably,
making it an ideal location for this research.

264

4.2. Data

255
256
257
258
259
260
261
262

272

The photorealistic mesh model and land use map (Figures 5c and f) were provided by
the Planning Department of Hong Kong (HKPlanD 2019, 2023). The road network used
for SVI generation was produced by the Transport Department of Hong Kong (HKTD
2019). Building footprints in the study area were extracted from the iB1000 digital
map of Hong Kong (HKLandsD 2025), which has been continuously updated since
2014. A 300m upper limit of DVI sampling point is set according to the uncontrolled
airspace classication of China (CAAC 2023), given that Hong Kong has no ocial
documentation specifying vertical airspace divisions.

273

4.3. Validation metrics

265
266
267
268
269
270
271

274
275
276
277
278
279
280
281
282
283

The validation of the proposed method is conducted based on SVI and DVI. Realworld SVI captured over a span of three years (2014, 2017, 2023) is used to validate
the eye-level green view from the street perspective. Real-world SVI is segmented using
Mask2Former model with Swin Transformer backbone (Cheng et al. 2022) trained on
the Cityscapes dataset. The analysis involves calculating the average greenery coverage
at sample points along each street. Real-world DVI (totaling 48 DVIs covered eight
sampling points and 60◦ × 6 headings) is collected using a drone (DJI Mini3 Pro) in a
public garden within the study area. One annotator manually assigned ground-truth
labels for greenery in the study area. Spearman's coecient between the normalized
greenery ratio in DVI pairs is used to validate the drone-level accuracy.
9

(a)

(b)

(c)

(d)

(e)

(f)

Figure 5.: The study area is Kowloon, Hong Kong. (a) The location of the study area;
(b) the 4.92 km2 study area; (c) 2D road network (111.6km in total); (d) 6,294 building
footprints; (e) DTM indicating a mixture of hilly and at areas; (f) land uses map

292

A method-level validation with baseline methods was conducted to demonstrate
eectiveness and eciency. Spearman's coecient was calculated to assess the eectiveness of the GVV results, comparing our method with that of Cimburova et al.
(2023); and the street-level green view, comparing with Cimburova and Blumentrath
(2022). The validation ground truth included 18 volumes manually annotated by one
annotator, and the street-level green view, which was calculated based on real-world
SVI (2017). The 18 greenery volumes were selected based on combinations of three
distance levels to the road (close, middle, far) and three levels of surrounding building
density (high, middle, low), with two samples for each combination.

293

4.4. Computational environment

284
285
286
287
288
289
290
291

294
295
296
297
298
299
300
301

The deep learning training environment was set up as follows: a high-performance computing (HPC) server with Dual Intel Xeon 6226R (16-core), 384GB RAM, and NVIDIA
V100 (32GB) SZM2 GPU. The training process was implemented using Pytorch (version 1.8) and Python (version 3.7). The view capture phase for GVV was completed
by Cesium (version 1.99), the script for GV V I calculation was completed by Numpy
(version 1.24) and Pandas (version 2.0.3), the GV V I analysis was completed in ArcGIS Pro (version 3.3) based on a desktop computer with 13th Gen Intel(R) Core(TM)
i7-13700K 3.40 GHz, 128 GB RAM, and NVIDIA RTX A4000 GPU.

10

302

303
304
305
306
307

5. Results
A total of 109,081 views were captured and ltered from the GVL-enriched 3D CIM
(17,718 from streets, 73,939 from windows, and 17,424 from drones). The SVM used
to lter penetrated views achieved 95% precision, recall, and F1-scores, and removed
14% of views at the window and street levels. Figure 6 illustrates examples of captured
SVIs, WVIs, and DVIs, respectively.

VIs with:
GVL

VIs with:
Photorealistic mesh
(a)

(b)

(c)

Figure 6.: Comparison of SVI, WVI, and DVI captured based on GVL CIM and the
photorealistic model. (a) SVI; (b) WVI; (c) DVI
308
309
310
311
312
313

The segmentation accuracy of greenery achieved IoU = 86.27%, while other classes
achieved IoU = 95.94% (mIoU = 91.11%). The greenery types, such as street-side shrub
and roof-top greenery, were often misclassied as classes such as facility, building, and
road, after manual inspection. These may be caused by low precision at the bottom
of the photorealistic models. A total of 23,490 greenery volumes were nally obtained.
The greenery volumes within the buer zone were 20,655 (up to 2.06km2 ).
Table 1.: Computational time of the proposed automated assessment method
Stage

Processing

Software

Time (h)

1

Semantic segmentation
Grid color index
GVL integrating
View capture
Visibility statistics
GV V I computation and analysis
Total

Stratied Transformer
Python
CesiumLab
Cesium
Python, Orange 3
ArcGIS Pro

83.18
11.03
0.80
147.03
14.41
0.00
256.45

2
3

317

Table 1 lists the computational cost of each process in the proposed method. The
rst stage, generating GVL, took 12.06 hours to process 4.92 km2 of CIM mesh with
90 million triangles and 267 million sampled points. The most time-consuming stage
was the view capturing process, taking up to 57% of the total processing time.

318

5.1. GV V I and three sub-indices

314
315
316

319
320
321
322

Figure 7 demonstrates the spatial distribution and hot spot map of GV V I and subindices. The highest GV V I value of 0.44 implied that a high amenity value level from all
three perspectives (i.e., street, window, and drone) was uncommon to appear simultaneously. Greenery hot spots of GV V I value (7,056 in total) mainly comprised greenery
11

342

located close to the major roads, areas with few surrounding high-rise buildings, and
greenery with greater canopy height and elevation. The near-edge interiors of public
green spaces exhibited moderate levels of GV V I . In contrast, cold spots of GV V I were
concentrated in the grass, shrublands, and high-rise, high-density residential areas.
Figure 7a, c, and d demonstrate the spatial distribution of GVVI sub-indices. The
street-level greenery received higher visual attention along the major roads. Visibility
was also obtained for street greenery scattered in residential areas though at lower
values. This dierence can primarily be attributed to the challenge of distinguishing
central plants when they are surrounded by edge greenery at lower eye levels, as illustrated in Figure 6a. The most visible greenery at the window level was typically
located at the edges of natural green spaces or urban park areas, whereas the interior
greenery in those areas was less visible. Along-street greenery exhibited lower visibility
because of visual obstructions by surrounding buildings. At the drone level, the visible
greenery was concentrated in the interior of large green spaces at high altitudes and in
areas with fewer surrounding buildings that caused visual obstructions.
Figure 7f, g, and h illustrate three hot spot maps of sub-indices. The hot spots
(condence > 90%) of GV V IS and GV V IW were 5,427 (mean = 0.06) and 5,237
(mean = 0.07), respectively. GV V ID obtained the largest number of hot spots, with
ND = 7,922, mean = 0.17. This indicates that GV V ID had a higher continuity spatial
clustering result because the aerial perspective was less visually obstructed.

343

5.2. Validation

344

5.2.1. Validation of eye-level green view from street perspective

323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341

345
346
347
348
349
350
351
352
353

Table 2 presents a street-level comparison between CIM-based and real-world SVI captured in dierent years. All real-world SVIs exhibited strong and positive associations
with CIM-based SVI (p-value < 0.001), indicating that the proposed method eectively simulated eye-level green view. The 2017 real-world SVI exhibited the highest
coecient (0.93) and R2 (0.84), and the lowest root mean square error (0.06). The
dierences in coecients between 2017 and the other two data sources were less than
0.03, indicating that the greenery visibility simulated using CIM was relatively insensitive to temporal variation. Nonetheless, it is recommended to acquire up-to-date data
to reduce the greenery change because of urban renovation and expansion.
Table 2.: Comparison of eye-level green view between CIM-based SVI and real-world
SVI captured in dierent years.
Data source

2014
2017
2023

354
355
356
357
358
359
360

Spearm. ρ

0.90
0.93
0.92

p-value

<0.001
<0.001
<0.001

MAE

0.04
0.04
0.04

RMSE

0.07
0.06
0.07

R2

0.81
0.84
0.82

The dierence between real-world SVI and CIM-based SVI was caused by the following reasons: (1) the roughness of the photorealistic mesh at the ground level; (2)
the lack of thin objects in the photorealistic mesh, which are labeled in the Cityscapes
dataset (e.g., car, person, pole, and trac sign), and (3) the assumption of GVL that
all greenery is impenetrable, as shown in Figure 8, which is dierent from the realworld SVI. The impenetrable greenery will result in a higher average GVVI dierential
of 0.002 according to density parameters and height classication criteria of dierent
12

Distribution

Hot spot map

j

(i)

0 0.44

i

𝐺𝑉𝑉𝐼

(a)

(j)

(e)

l

0

k

(k)

1
𝐺𝑉𝑉𝐼 S

(b)

(f)

(l)

m
n

0

(m)

1
𝐺𝑉𝑉𝐼 W

(c)

(g)

(n)

(o)

o
p
0

1
𝐺𝑉𝑉𝐼 D

(d)
N

(h)
Hot Spots (99% Confid.)
Hot Spots (95% Confid.)
Hot Spots (90% Confid.)

(p)
Cold Spots (99% Confid.)
Cold Spots (95% Confid.)
Cold Spots (90% Confid.)

Not Significant
100m Buffer

Figure 7.: Scaled distribution and hot spot map results with visible and non-visible
areas. (a),(e) GV V I ; (b),(f) GV V IS ; (c),(g) GV V IW ; (d),(h) GV V ID ; (i), (k), (m),
(o) non-visible areas; (j), (l), (n), (p) visible areas.
13

363

greenery types provided by Li et al. (2025). Despite the above limitations, CIM can effectively simulate the perception of greenery as experienced by humans and can further
be utilized for GVV computation.

364

5.2.2. Validation of eye-level green view above ground

361
362

365
366
367
368
369
370
371

Figure 9 shows the comparison between the eye-level green view of CIM and real-world
DVI. The overall Spearman's coecient and R2 of totaling 48 DVI pairs reached 0.95
(p-value < 0.001) and 0.92, respectively, demonstrating a satisfactory simulation result
of ours. Outlier pairs are mainly caused by vegetation renewal (e.g., changes in height
and pattern). Furthermore, due to dierences in camera FOV, some distant vegetation
with fewer pixels may be obscured by buildings, leading to subtle variations between
pairs.

VIs with:
Photorealistic Mesh

VIs with:
GVL

1
𝛾5 𝛾
𝛾2
𝛾3
4 𝛾 10 12
𝛾7 𝛾6 𝛾 11𝛾 𝛾13
𝛾9 𝛾
𝛾8
𝛾14

𝛾19

𝛾16
𝛾18

𝛾15

𝛾20

𝛾22
𝛾21

𝛾23

𝛾24

𝛾25

𝛾42
𝛾53𝛾58
𝛾39
𝛾26
𝛾52 𝛾59
𝛾37
𝛾44 46
60
𝛾55𝛾
𝛾28 𝛾29𝛾34 36 38 𝛾41 𝛾43 𝛾45 𝛾𝛾48 𝛾50
51
𝛾
𝛾30 33 𝛾 𝛾
𝛾47 49 𝛾54
𝛾 35 𝛾40
𝛾
𝛾56
31
𝛾
𝛾
𝛾27 𝛾32
𝛾57

𝛾17

VIs with:
Segmented greenery

VIs with:
Real-world
Invisible objects

Figure 8.: Illustration of Photorealistic-mesh-based, GVL-based, segmented-greenerybased, and real-world SVIs.

372
373
374
375
376
377
378
379
380
381

5.2.3. Method-level validation
The proposed method exhibited high eciency, accuracy, and multi-perspective assessment (Table 3). Although the method of Cimburova and Blumentrath (2022) achieved
higher eciency, it failed to compute GVV, while that of Cimburova et al. (2023) required signicant computational cost for viewshed analysis. Both methods lack multiperspective assessment. Our proposed method required 24.6h to calculate GV V IS ,
leading to an approximately 40% saving in time. Our results showed a strong correlation with the real-world SVI for both GVV and eye-level green view, with Spearman's
ρ = 0.89 and 0.93 (p-values < 0.01), and approximately 0.2 higher than previous methods, indicating enhanced assessment accuracy.

14

(a)

(b)

Real-world

Photorealistic mesh

Ground truth greenery

GVL

Validation area

Figure 9.: Scatter plot and comparison between eye-level green view of CIM and realworld DVI. (a) Scatter plot; (b) illustrative examples.
Table 3.: Validation of the proposed method with viewshed-based GVV assessment
Method
Cimburova and Blumentrath (2022)
Cimburova et al. (2023)
Ours

Type
Viewshed
Viewshed
CIM

Time cost (h)
1.3
41.6
24.6

Spearm. ρ (GVV)
\
0.70
0.89

Spearm. ρ (eye-level)
0.78
\
0.93

(a) wS:wW :wD=1:2:3

(b) wS:wW :wD=1:3:2

(c) wS:wW :wD=2:1:3

(d) wS:wW :wD=2:3:1
Hot Spots (99% Confid.)
Cold Spots (99% Confid.)
Not Sigificant

(e) wS:wW :wD=3:1:2
Hot Spots (95% Confid.)
Cold Spots (95% Confid.)
100m Buffer

(f) wS:wW :wD=3:2:1
Hot Spots (90% Confid.)
Cold Spots (90% Confid.)

Figure 10.: Hot spot maps of greenery's GV V I with dierent weight combination
15

382

5.3. GVVI analysis

383

5.3.1. Sensitivity of weight adjustment

390

The linear weighted GV V I , with dierent weight groups, identied 6,840 hot spots
and 7,512 cold spots GVV on average, as shown in Figure 10a-f. Those groups that
emphasize SVI and WVI exhibited lower GV V I on average; for example, about 0.04 in
Figure 10d and f. This dierence indicates that GV V ID contributed more signicantly
to GV V I . The clusters from dierent weights share a similar spatial distribution pattern, but vary in certain areas, e.g., greenery in open space with relatively high elevation
or dense greenery along streets.

391

5.3.2. Correlations between the sub-indices

384
385
386
387
388
389

392
393
394
395
396
397
398
399

The Spearman's correlation analysis is shown in Figure 11, with all p-values < 0.0001.
The correlation between GV V IS and GV V IW (ρ = 0.34) was weak but signicant,
indicating that a green volume more visible from streets was likely to be more visible
from windows. The correlation between streets and drones was weaker (ρ = 0.19). This
is because pedestrians mainly observe the greenery's bottom part, whereas a drone's
perspective oers a clearer view of the tree canopy and less of the bottom part due
to high elevation. The coecient between GV V IW and GV V ID was 0.40, indicating a
moderate positive association between greenery visibility in WVI and DVI.

*99,6    
*99,:    
*99,'    

'







1'
9,

9,

:

*9

9,

*9

*9

9,

6

1'9,    



Figure 11.: Correlation metrics of the scaled GV V I sub-indices and NDVI, all p-values
< 0.0001

407

The correlation between NDVI and the three sub-indices shows weak associations,
especially at the street and window levels, as shown in Figure 11. The negligible correlation between NDVI and GV V IS (ρ = −0.09) and GV V IW (ρ = 0.08) indicates
that greenery visibility of pedestrians and residents was signicantly dierent from traditional top-view data sources. Although the correlation between NDVI and GV V ID
showed a moderate positive association (ρ = 0.51), greenery in areas such as open
spaces, streets, and high-density residential areas displayed a notable dierential compared with NDVI (ranging from -0.88 to 0.39) according to manual validation.

408

5.3.3. Dierentials and the correlations to land use

400
401
402
403
404
405
406

409
410
411

Figure 12a shows the number of clustered hot spots of GV V ID-S and GV V IW-S in
dierent land use types. The government land areas had the highest number of hot
spots of dierentials, mainly because of the larger area size and higher altitude (e.g.,
16

414
415
416
417



*99,' 6
*99,: 6


D





E



E

&R

3X

3UL

UH
VL

*99,GLIIHUHQWLDOV

1XPEHURI+RWVSRWV

418

L
PP
H
*R UFLDO
YH
UQP
2S HQW
HQ
V
7UD SDFH

QVS
RUW
D
5D WLRQ
LOZ
D\
V
8W
LOLW
LHV
9D
FDQ
WOD
QG
2W 
KH
UV
:R
RG
ODQ
6K G
UXE
ODQ
*U G
DVV
OD
$J QG
ULF
XOW
:D XUH
UHK
RX
VH

413

hilltops). GV V ID-S had 7,757 hot spots of 0.77km2 , equivalent to 37.6% of the total
greenery area within the buer zone. Except for government areas, open spaces, private
residential areas, utilities, and woodland contributed most signicantly to GV V ID-S
(N≳800). Meanwhile, GV V IW-S had fewer hot spots (N = 6,539, 31.7% of the total
greenery area), with government areas being the primary contributor, followed by private residential, open spaces, and transportation. Sensitivity analysis of DVI sampling
parameters' impact on GV V ID-S hot spots is shown in Appendix C1.

UHV

412

/DQGXVHW\SH

Figure 12.: Hot spots number and dierentials distribution of GV V ID-S and GV V IW-S .
(a) Number of hot spots; (b) dierential distribution.

427

Figure 12b demonstrates the GV V I dierentials distribution of GV V ID-S and
GV V IW-S against land-use types. More than 81% and 76% of greenery volumes exhibited positive values for GV V ID-S and GV V IW-S , respectively. Not surprisingly, the
transportation area was the primary land-use type with negative value dierentials.
In contrast, other land-use types predominantly showed positive dierentials for both
GV V ID-S and GV V IW-S , suggesting that greenery in these areas was more visible
from residential buildings and drones. Furthermore, the dierentials of GV V IW-S were
smaller than those for GV V ID-S , indicating that the contribution of GVV increased
with elevation of the observing points.

428

6. Discussion

429

6.1. GVV assessment method and results

419
420
421
422
423
424
425
426

430
431
432
433
434
435
436
437
438

The proposed GV V I metric includes multi-perspective greenery visibility, street, window, and drone, compared with the previous GVV assessment, which was limited to
street-level only. The proposed GVL based on deep learning and CIM can eciently
capture views from any perspectives and assess GVV. GVL enables localizing greenery volume from remote observer points (e.g., windows, drone views). Besides, the
proposed method can assess eye-level green view and GVVI simultaneously without
complex coding using a reusable and customizable GVL. The proposed method reduces the time cost by 40% and increases the correlation by 0.19 compared to the
viewshed-based method. Therefore, it enables ecient, eective, and multi-perspective
17

450

GVV assessment.
The analysis results conrmed the imbalanced spatial distribution of 3D GVV. A
total of 37.6% dierentials from drone, and 31.6% from window, compared to traditional street-level studies, were identied. Correlation metrics highlight the necessity
of assessing 3D greenery visibility, according to the correlation between GV V IS and
GV V IW (0.34), as well as GV V IS and GV V ID (0.19). A multi-level strategy that balances ground-level experience with overhead landscape aesthetics is therefore required
to ensure residents can access adequate and consistent visible greenery from dierent
perspectives. The hot spots, distribution, and weight adjustment map can serve as a
coarse-to-ne guide for greenery maintenance. The three sub-indices and NDVI can
further be combined as a comprehensive, complementary, and down-to-top assessment
strategy for assessing dierent levels of greenery visibility.

451

6.2. Potential applications

439
440
441
442
443
444
445
446
447
448
449

452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484

The GV V I and hot spot map identied the amenity value level of urban greenery
and provided new insights into priority areas for maintenance. The proposed GV V I
can serve as the foundational data for the government authorities to prioritize green
areas with high GV V I value and to make informed decisions regarding urban green
space management, including planting, removal, or maintenance activities. The multiperspectives GV V I provides a signicant reference framework across multiple elds,
including urban planning, landscape architecture, forestry management, and dronebased environmental assessment.
The application of GVL can be expanded. For instance, it can be used to determine
how the height and canopy size of greenery can best serve surrounding areas by simulating greenery volumes in CIM. Reorganization of the three-dimensional pattern of
the urban landscape may inuence the take-o, route trajectory, and landing of drones
(Perperidou and Kirgianis 2022). GV V ID could help recommend optimal drone takeo and landing sites for cities, balancing aesthetic and cultural factors with other drone
indicators, such as noise and wind speed (Wild 2024). The drone route planning can
be optimized based on the GV V ID to balance the observed greenery resolution and
cost (Ecke et al. 2022). The proposed method also inspires CIM-based simulations for
multi-perspective visions for agents, i.e., physical assets (e.g., windows, lookouts, and
surveillance), individuals (e.g., wheelchair users, children, and cyclists), and robotics
(e.g., drones, unmanned cars, and vessels). This capacity thus enables applications in
safety, urban accessibility, and potential conicts analysis in order to support a future
human-machine society (Meng et al. 2025).
The weights of GV V I can be determined through experts' experience, surveys, and
local policies. Greenery managers can assign a personalized weighting to alter the emphasis of GV V I according to the focus of management, for example, by prioritizing
greenery with higher GV V IW (Meng and Zheng 2023). Additionally, other indicators, such as surrounding air pollution, the noise impact of drones, greenery health
condition, greenery maintenance cost, urban heat island, and global warming eect
(Chauhan et al. 2025, Zheng et al. 2025), can be combined with GV V I to comprehensively evaluate the ecosystem value and maintenance priority of greenery in future
research. Optimization-based multiple-criteria decision analysis can also be used for
GV V I assessment to recognize the order of priority (Li et al. 2023b, Zhou and Xue
2023).

18

485

6.3. Limitations and future directions

506

This research has certain limitations. From the CIM perspective, the limitations include
low precision at the bottom of the photorealistic models, the absence of realistic occlusion caused by the window frame, opaque GVL assumption against the permeability
of greenery, and low eciency in view capture. From the view sampling perspective,
the sampling parameter based on experience may lead to assessment bias, and the
static views-capture approach failed to reect human-environment interactions. From
the drone views perspective, limitations include over-reliance on simulated data and
under-representation of actual trajectories of dierent types of drone views. Besides,
model segmentation accuracy may be aected when applied to CIM across regions outside Hong Kong due to dierences in feature distribution. Finally, this paper excludes
the impact factors (e.g., urban morphology, seasons, and temporal data availability)
on GV V I across dierent perspectives.
Therefore, future work can focus on several key aspects. These include the development of high-precision CIM models and the design of high-performance algorithms
for OpenGL-based rendering. Additionally, evaluating sampling parameter bias, simulating viewpoints driven by diverse user types and video data, and modeling realistic
drone trajectories are necessary to advance future research. Further, the applicability
validation of datasets (e.g., HRHD-HK) across regions using transfer learning (Weiss
et al. 2016) should be explored. Finally, systematic analyses of seasonality, transferability, and temporal data availability impact on GV V I assessment are required (Han
et al. 2023).

507

7. Conclusion

486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505

508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530

This paper presents an eective, ecient, and automatic method to evaluate GVV. In
theory, the proposed method denes a new GV V I metric to unify street, window, and
drone-level GVV. In methodology, the proposed method introduces GVL for enabling
the 3D CIM-based GVV assessment using deep learning and color index. In application,
the proposed method is implemented through a case study in Hong Kong. Linearweighted GV V I values and hot spot clustering, covering 2.06 km2 of greenery volumes,
provided an analysis of multi-perspective greenery visibility.
The results conrmed the uneven contribution of greenery visibility. The identication of an additional 37.6% of greenery at the drone level, and 31.7% at the window
leveloverlooked in traditional street-level studiesdemonstrates the value of the proposed method for greenery management. The Spearman's correlation between GV V IS
and GV V IW (0.34), and that with GV V ID (0.19) highlighted the necessity of 3D
greenery visibility. Greenery located closest to main roads, surrounded by a few highrise buildings, and with high canopy height, should be given priority for maintenance.
The quantitative dierentials of hot spot distributions across land-use types provided
a reliable reason for stakeholders to develop targeted maintenance strategies.
The results of the proposed method demonstrate the utility of assessing 3D GV V I
and can serve as foundational geospatial data for urban planners and managers to prioritize greenery design and maintenance. The GVL attached to the photorealistic mesh
can represent a close-to-reality urban built environment, enhance the semantic representation of greenery in 3D CIM, simulate greenery visibility from any perspective,
and provide ecient GVV assessment compared with previous methods, thereby further supporting applications such as drone route planning and greenery maintenance.
19

540

Limitations of this study include model penetration, the time-consuming process for
capturing views, a lack of simulation of human-environment interactions and drone
trajectories, as well as the exclusion of factors such as urban morphology, seasons, and
temporal data availability. Future research directions are needed, such as incorporating
a higher precision photorealistic mesh/CIM, developing high-performance algorithms
for OpenGL rendering, simulating real drone trajectories and human-environment interactions, evaluating sampling parameter bias, validating dataset applicability across
regions, analyzing seasonality, transferability, temporal data availability of CIM impact on GV V I assessment, integrating additional assessment indicators, and exploring
potential applications.

541

Acknowledgments

531
532
533
534
535
536
537
538
539

544

The work presented in this paper was supported by the Hong Kong Research Grants
Council (RGC) (No. T22-504/21-R), and in part by the Department of Science and
Technology of Guangdong Province (GDST) (2023A1515010757).

545

Declaration of interests

542
543

547

The authors declare that they have no known competing nancial interests or personal
relationships that could have appeared to inuence the work reported in this paper.

548

Data and codes availability statement

546

551

The code and data are shared privately at https://gshare.com/s/46caea113650a8bfc03c
for review purposes. The code and data will be made publicly available upon acceptance.

552

Declaration of generative AI in writing

549
550

556

A university self-hosted LLM GenAI was used to assist with proofreading the work to
correct grammatical and connection errors in the main text, and to improve readability
of Abstract, Introduction, and Conclusion sections. We declare no use of GenAI to
produce any new content in the work.

557

References

553
554
555

558
559
560
561
562
563
564
565

Bahr, S., 2024. The relationship between urban greenery, mixed land use and life satisfaction: An examination using remote sensing data and deep learning. Landscape and Urban
Planning, 251, 105174. doi:10.1016/j.landurbplan.2024.105174.
Biljecki, F. and Ito, K., 2021. Street view imagery in urban analytics and GIS: A review.
Landscape and Urban Planning, 215, 104217. doi:10.1016/j.landurbplan.2021.104217.
Bolte, A.M., et al., 2024. The green window view index: automated multi-source visibility
analysis for a multi-scale assessment of green window views. Landscape Ecology, 39 (3), 71.
doi:10.1007/s10980-024-01871-7.

20

566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619

CAAC, 2023. National airspace basic classication method. Civil Aviation Administration of
China, Government of PR China. Available from: https://www.caac.gov.cn/XXGK/XXGK/
TZTG/202312/P020231222621680839714.pdf.
Chauhan, S., et al., 2025. Urban heat stress, air quality and climate change adaptation strategies in UK cities. Frontiers of Engineering Management, 117. doi:10.1007/s42524-025-4029y.
Cheng, B., et al., 2022. Masked-attention mask transformer for universal image segmentation.
In : Proceedings of the IEEE/CVF conference on computer vision and pattern recognition.
12901299. doi:10.1109/CVPR52688.2022.00135.
Cimburova, Z. and Blumentrath, S., 2022. Viewshed-based modelling of visual exposure to
urban greeneryAn ecient GIS tool for practical planning applications. Landscape and
Urban Planning, 222, 104395. doi:10.1016/j.landurbplan.2022.104395.
Cimburova, Z., Blumentrath, S., and Barton, D.N., 2023. Making trees visible: A GIS method
and tool for modelling visibility in the valuation of urban trees. Urban Forestry & Urban
Greening, 81, 127839. doi:10.1016/j.ufug.2023.127839.
Doick, K.J., et al., 2018. CAVAT (Capital Asset Value for Amenity Trees):
valuing amenity trees as public assets. Arboricultural Journal, 40 (2), 6791.
doi:10.1080/03071375.2018.1454077.
Ecke, S., et al., 2022. UAV-based forest health monitoring: A systematic review. Remote Sensing, 14 (13), 3205. doi:10.3390/rs14133205.
Elsadek, M., Liu, B., and Xie, J., 2020. Window view and relaxation: Viewing green space from
a high-rise estate improves urban dwellers' wellbeing. Urban Forestry & Urban Greening,
55, 126846. doi:10.1016/j.ufug.2020.126846.
Fan, Z., Feng, C.C., and Biljecki, F., 2025. Coverage and bias of street view imagery in mapping the urban environment. Computers, Environment and Urban Systems, 117, 102253.
doi:10.1016/j.compenvurbsys.2025.102253.
Gong, F.Y., et al., 2018. Mapping sky, tree, and building view factors of street
canyons in a high-density urban environment. Building and Environment, 134, 155167.
doi:10.1016/j.buildenv.2018.02.042.
Guo, Y., et al., 2021. Deep learning for 3D point clouds: A survey. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 43 (12), 43384364.
doi:10.1109/TPAMI.2020.3005434.
Han, Y., et al., 2023. Mapping seasonal changes of street greenery using multi-temporal streetview images. Sustainable Cities and Society, 92, 104498. doi:10.1016/j.scs.2023.104498.
Helliwell, R., 2008. Amenity valuation of trees and woodlands. Arboricultural Journal, 31 (3),
161168. doi:10.1080/03071375.2008.9747532.
HKBD, 2023. Building information and age records. Building Department, Government of Hong Kong SAR. Available from: https://data.gov.hk/en-data/dataset/
hk-bd-opendata-building-information.
HKLandsD, 2022. 3D Photo-realistic Model. Lands Department, Government of Hong Kong
SAR. Available from: https://3d.map.gov.hk/.
HKLandsD, 2025. Digital topographic map iB1000. Lands Department, Government of
Hong Kong SAR. Available from: https://www.hkmapservice.gov.hk/OneStopSystem/
map-search?product=OSSCatB&series=iB1000.
HKPlanD, 2019. 3D Photo-realistic Model. Planning Department, Government of Hong Kong
SAR. Available from: https://www-pland-gov-hk.eproxy.lib.hku.hk/pland_en/info_
serv/3D_models/download.htm.
HKPlanD, 2023. Land utilization in Hong Kong. Planning Department, Government of
Hong Kong SAR. Available from: https://www.pland.gov.hk/pland_en/info_serv/
open_data/landu/.
HKTD, 2019. Road Network. Transport Department, Government of Hong Kong SAR. Available from: https://static.data.gov.hk/td/road-network-v2/RdNet_IRNP.gdb.zip.
Huang, S., et al., 2021. A commentary review on the use of normalized dierence vegetation
index (NDVI) in the era of popular remote sensing. Journal of Forestry Research, 32 (1),

21

620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673

16. doi:10.1007/s11676-020-01155-1.
Huang, Y., et al., 2025. No "true" greenery: Deciphering the bias of satellite and street
view imagery in urban greenery measurement. Building and Environment, 269, 112395.
doi:10.1016/j.buildenv.2024.112395.
Jiang, B., et al., 2021. Impacts of nature and built acoustic-visual environments on human's
multidimensional mood states: A cross-continent experiment. Journal of Environmental Psychology, 77, 101659. doi:10.1016/j.jenvp.2021.101659.
Ko, W.H., et al., 2020. The impact of a view from a window on thermal comfort, emotion, and cognitive performance. Building and Environment, 175, 106779.
doi:10.1016/j.buildenv.2020.106779.
Kong, F., Yin, H., and Nakagoshi, N., 2007. Using GIS and landscape metrics in the hedonic
price modeling of the amenity value of urban green space: A case study in Jinan City, China.
Landscape and Urban Planning, 79 (3), 240252. doi:10.1016/j.landurbplan.2006.02.013.
Kong, L., et al., 2017. Regulation of outdoor thermal comfort by trees in Hong Kong. Sustainable Cities and Society, 31, 1225. doi:10.1016/j.scs.2017.01.018.
Konijnendijk, C.C., et al., 2006. Dening urban forestry - A comparative perspective
of North America and Europe. Urban Forestry & Urban Greening, 4 (3), 93103.
doi:10.1016/j.ufug.2005.11.003.
Lai, X., et al., 2022. Stratied transformer for 3D point cloud segmentation. March.
doi:10.1109/CVPR52688.2022.00831.
LeCun, Y., Bengio, Y., and Hinton, G., 2015. Deep learning. Nature, 521 (7553), 436444.
doi:10.1038/nature14539.
Li, M., et al., 2026. Inuence of objective and perceived exposures to urban nature on people's
happiness. npj Urban Sustainability, 6 (1), 6. doi:10.1038/s42949-025-00306-9.
Li, M., et al., 2023a. HRHD-HK: A benchmark dataset of high-rise and high-density urban
scenes for 3D semantic segmentation of photogrammetric point clouds. In : 2023 IEEE International Conference on Image Processing Challenges and Workshops (ICIPCW), October.
37143718. doi:10.1109/ICIPC59416.2023.10328383.
Li, M., et al., 2022. A room with a view: Automatic assessment of window views for high-rise
high-density areas using City Information Models and deep transfer learning. Landscape and
Urban Planning, 226. doi:10.1016/j.landurbplan.2022.104505.
Li, M., Xue, F., and Yeh, A.G.O., 2023b. Bi-objective analytics of 3D visual-physical nature
exposures in high-rise high-density cities for landscape and urban planning. Landscape and
Urban Planning, 233, 104714. doi:10.1016/j.landurbplan.2023.104714.
Li, M., Xue, F., and Yeh, A.G., 2025. Ecient and accurate assessment of window view distance
using City Information Models and 3D Computer Vision. Landscape and Urban Planning,
260, 105389. doi:10.1016/j.landurbplan.2025.105389.
Li, M., Yeh, A.G., and Xue, F., 2024. CIM-WV: A 2D semantic segmentation dataset of rich
window view contents in high-rise, high-density Hong Kong based on photorealistic city
information models. Urban Informatics, 3 (1), 12. doi:10.1007/s44212-024-00039-7.
Li, X., et al., 2015. Assessing street-level urban greenery using Google Street View and
a modied green view index. Urban Forestry & Urban Greening, 14 (3), 675685.
doi:10.1016/j.ufug.2015.06.006.
Liang, J., et al., 2017. Embedding user-generated content into oblique airborne
photogrammetry-based 3D city model. International Journal of Geographical Information
Science, 31 (1), 116. doi:10.1080/13658816.2016.1180389.
Liu, D., et al., 2023. Establishing a citywide street tree inventory with street view images
and computer vision techniques. Computers, Environment and Urban Systems, 100, 101924.
doi:10.1016/j.compenvurbsys.2022.101924.
Meng, S., et al., 2025. From 3D pedestrian networks to wheelable networks: An automatic
wheelability assessment method for high-density urban areas using contrastive deep learning
of smartphone point clouds. Computers, Environment and Urban Systems, 117, 102255.
doi:10.1016/j.compenvurbsys.2025.102255.
Meng, S. and Zheng, H., 2023. A personalized bikeability-based cycling route recommendation

22

674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727

method with machine learning. International Journal of Applied Earth Observation and
Geoinformation, 121, 103373. doi:10.1016/j.jag.2023.103373.
NCD Risk Factor Collaboration (NCD-RisC), 2016. A century of trends in adult human height.
eLife, 5, e13410. doi:10.7554/eLife.13410.
Neema, M.N. and Ohgai, A., 2010. Multi-objective location modeling of urban parks and open
spaces: Continuous optimization. Computers, Environment and Urban Systems, 34 (5), 359
376. doi:10.1016/j.compenvurbsys.2010.03.001.
Panduro, T.E. and Veie, K.L., 2013. Classication and valuation of urban green spacesA hedonic house price valuation. Landscape and Urban Planning, 120, 119128.
doi:10.1016/j.landurbplan.2013.08.009.
Perperidou, D.G. and Kirgianis, D., 2022. Urban air mobility (UAM) integration to
urban planning. In : Conference on Sustainable Urban Mobility. Springer, 16761686.
doi:10.1007/978-3-031-23721-8_130.
Pyka, K., Piskorski, R., and Jasi«ska, A., 2022. LiDAR-based method for analysing landmark
visibility to pedestrians in cities: case study in Kraków, Poland. International Journal of
Geographical Information Science, 36 (3), 476495. doi:10.1080/13658816.2021.2015600.
Roberto Barbosa, M., et al., 2010. Forest re alert system: a Geo Web GIS prioritization
model considering land susceptibility and hotspotsa case study in the carajás national
forest, brazilian amazon. International Journal of Geographical Information Science, 24 (6),
873901. doi:10.1080/13658810903194264.
Rui, Q. and Cheng, H., 2023. Quantifying the spatial quality of urban streets with open street
view images: A case study of the main urban area of Fuzhou. Ecological Indicators, 156,
111204. doi:10.1016/j.ecolind.2023.111204.
Seifert, E., et al., 2019. Inuence of Drone Altitude, Image Overlap, and Optical Sensor
Resolution on Multi-View Reconstruction of Forest Images. Remote Sensing, 11 (10).
doi:10.3390/rs11101252.
Song, Y., et al., 2017. Trends and opportunities of BIM-GIS integration in the architecture, engineering and construction industry: A review from a spatio-temporal statistical perspective.
ISPRS International Journal of Geo-Information, 6 (12), 397. doi:10.3390/ijgi6120397.
Szegedy, C., et al., 2016. Rethinking the inception architecture for computer vision. In : Proceedings of the IEEE conference on computer vision and pattern recognition. 28182826.
doi:doi.org/10.48550/arXiv.1512.00567.
Tang, L., et al., 2023. Assessing the visibility of urban greenery using MLS LiDAR data.
Landscape and Urban Planning, 232, 104662. doi:10.1016/j.landurbplan.2022.104662.
Tara, A., Lawson, G., and Renata, A., 2021. Measuring magnitude of change by high-rise buildings in visual amenity conicts in Brisbane. Landscape and Urban Planning, 205, 103930.
doi:10.1016/j.landurbplan.2020.103930.
Ulrich, R.S., 1984. View through a window may inuence recovery from surgery. Science, 224
(4647), 420421. doi:10.1126/science.6143402.
Wang, L., et al., 2022. Measuring residents' perceptions of city streets to inform better street
planning through deep learning and space syntax. ISPRS Journal of Photogrammetry and
Remote Sensing, 190, 215230. doi:10.1016/j.isprsjprs.2022.06.011.
Wang, Z., et al., 2023. A view-tree method to compute viewsheds from digital elevation models. International Journal of Geographical Information Science, 37 (1), 6887.
doi:10.1080/13658816.2022.2094385.
Weiss, K., Khoshgoftaar, T.M., and Wang, D., 2016. A survey of transfer learning. Journal of
Big Data, 3 (1), 9. doi:10.1186/s40537-016-0043-6.
Wild, G., 2024. Urban aviation: The future aerospace transportation system for intercity and
intracity mobility. Urban Science, 8 (4), 218. doi:10.3390/urbansci8040218.
Xiang, S., et al., 2024. Autonomous eVTOL: A summary of researches and challenges. Green
Energy and Intelligent Transportation, 3 (1), 100140. doi:10.1016/j.geits.2023.100140.
Xie, J., et al., 2025. Semantic segmentation of building facade materials and colors for urban
conservation. npj Heritage Science, 13 (1), 378. doi:10.1038/s40494-025-01888-4.
Xue, F., Wu, L., and Lu, W., 2021. Semantic enrichment of building and city in-

23

728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751

formation models: A ten-year review. Advanced Engineering Informatics, 47, 101245.
doi:10.1016/j.aei.2020.101245.
Yang, Y., et al., 2020. Urban greenery, active school transport, and body weight among Hong
Kong children. Travel Behaviour and Society, 20, 104113. doi:10.1016/j.tbs.2020.03.001.
Yu, L. and Gong, P., 2012. Google Earth as a virtual globe tool for Earth science applications
at the global scale: progress and perspectives. International Journal of Remote Sensing, 33
(12), 39663986. doi:10.1080/01431161.2011.636081.
Yu, S., et al., 2016. View-based greenery: A three-dimensional assessment of city buildings'
green visibility using Floor Green View Index. Landscape and Urban Planning, 152, 1326.
doi:10.1016/j.landurbplan.2016.04.004.
Yu, X., et al., 2022. Spatio-temporal monitoring of urban street-side vegetation greenery using Baidu Street View images. Urban Forestry & Urban Greening, 73, 127617.
doi:10.1016/j.ufug.2022.127617.
Zhang, J., et al., 2024. Exploring geospatial digital twins: a novel panorama-based
method with enhanced representation of virtual geographic scenes in Virtual Reality
(VR). International Journal of Geographical Information Science, 38 (11), 23012324.
doi:10.1080/13658816.2024.2386064.
Zheng, L., et al., 2025. Harnessing geographic information system and street view
imagery for thermal gradient distribution auditing. Urban Climate, 59, 102248.
doi:10.1016/j.uclim.2024.102248.
Zhou, Q. and Xue, F., 2023. Pushing the boundaries of modular-integrated construction: A symmetric skeleton grammar-based multi-objective optimization of passive design for energy savings and daylight autonomy. Energy and Buildings, 296, 113417.
doi:10.1016/j.enbuild.2023.113417.

24

752

753
754
755
756
757
758
759

Appendix A. GVVI sub-indices distribution and correlation
Figure A1a illustrates the value distribution of GV V IS , GV V IW and GV V ID . The
majority of greenery had a GV V IS of 0 (mean = 0.02, stdev. = 0.07), showing that
the visual contribution of greenery volumes at the street level was highly imbalanced.
The value of GV V IW predominantly fell within the range of 0 to 0.05 (mean = 0.04,
stdev. = 0.06), suggesting that the majority of greenery had low visibility at the window
level. The distribution of GV V ID was concentrated between 0 and 0.14 (mean = 0.10,
stdev. = 0.12), which was reasonable because of the altitude change.

(a)

(b)

Figure A1.: Distribution of GV V I sub-indices and correlation matrix among subindices with dierent sampling heights and NDVI. (a) Distribution; (b) correlation
matrix, all p-values < 0.0001
760
761
762
763
764
765
766
767
768
769
770
771

The correlation of GV V I computed based on dierent height levels of WVI (low:
0-24m, middle: 25-52m, and high: 53-137m, based on Natural Break) and DVI (50 to
300m) is conducted to validate the spatial dependency of sub-indices. The correlation
between GV V IW with increasing height levels and GV V IS showed a diminishing trend,
which logically considers the proximity of the viewpoints at lower height for GV V IW
and GV V IS . All coecients of GV V IW at dierent height levels with other indicators
show positive but less moderate associations, indicating that GVV is less spatially
dependent at near-ground level. The diverse height levels of GV V ID exhibit strong
associations among themselves, meaning that the vertical disparity of greenery within
the research area was under 100m. It is therefore recommended to reduce the number
of DVI sampling intervals, especially in areas with small terrain height dierences, to
reduce the computation cost.

25

773
774
775

Figure B1 shows the drone views from FOVs equal to 60 and 120 degrees at dierent
sample points' height levels. A total of 17,424 DVIs (FOV = 60) and 8,712 DVIs (FOV
= 120) were captured. In general, the average visible pixel of FOV 120 was similar to
that of FOV 60, which was lower by 923 only (maximum = 111,929).

volume

776

Appendix B. Comparison of GVV with dierent DVI FOV

(FOV=120)
volume
(FOV=120)
Visible pixel per volume

772

(a)

Visible pixel per volume (FOV=60)

(b)
FOV=60

FOV=120

Extra
visible
volumes
(c)

Figure B1.: Visibility dierence between FOV equal to 60 and 120 at dierent heights.
(a) average visible pixel per volume; (b) scatter plot of visible pixel per volume's
dierence; and (c) the DVIs at the same sample point with dierent FOV, which are
displayed with GVL and photorealistic meshes, and DVI captured by real drone.
777
778
779
780
781
782
783
784
785
786
787

The result of the FOV of 60 degrees had a higher average visible pixel ND per
volume at the height of 50m, with an average of 7,878 pixels, compared to the FOV
of 120. However, the visibility of FOV 120 degrees exceeded that of FOV 60 with the
increasing height, as shown in Figure B1a. This was because FOV of 120 degrees could
capture greenery close to the bottom of the virtual camera as well as greenery further
away (Figure B1c). The Spearman's coecient was 0.92 for visible pixels per volume
at 50m height. The Spearman's coecient gradually stabilized at around 0.95 as the
height increased, as depicted in Figure B1b. Hence, an FOV of 120 is suggested as the
data source for calculating the drone-level visibility with high altitudes (e.g., altitude
> 100m), as it can capture visible greenery with reduced computational expenses (only
half the time required compared to an FOV of 60).

26

788

789
790
791
792
793
794
795
796
797
798
799

Appendix C. Sensitivity analysis of DVI sampling position parameters
Figure C1 shows the sensitivity analysis of DVI sampling position parameters impact
on GV V ID-S hot spot dierential. The horizontal and height intervals, as well as FOV
changing caused less impact on GV V ID-S hot spot dierential, with an average value of
36.3%. Dierent FOV mainly shared the similar trend on GV V ID-S , but FOV equal to
120 showed a dramatic change curve. For FOV 120, the hot spot dierential increased
with higher sampling heights and smaller horizontal intervals, while for FOV 60, it
decreased as the height increased; both FOVs showed similar trends with respect to
the horizontal interval, as shown in Figure C1c and d. This may be caused by the
building obstruction demonstrated in Figure B1c. Therefore, the proposed method can
robustly evaluate the dierential between dierent observer perspectives, and it is also
suggested to evaluate GV V ID using multiple sampling height levels.

(a)

(b)

(c)

(d)

Figure C1.: Sensitivity analysis of DVI sampling position parameters impact on
GV V ID-S hot spot dierential, where (a) and (b) shows the horizontal, height interval
comparison of FOV 60 and 120, (c) and (d) shows the horizontal, sampling height
comparison of FOV 60 and 120.

27