A room with a view: Automatic assessment of window views for high-rise high-density areas using
City Information Models and deep transfer learning
Maosu Li 1, Fan Xue 2*, Yijie Wu 3, and Anthony G. O. Yeh 4
This is the peer-reviewed post-print version of the paper:
Li, M., Xue, F., Wu, Y., & Yeh, A.G.O. (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, 104505.
Doi: 10.1016/j.landurbplan.2022.104505
The final version of this paper is available at: https://doi.org/10.1016/j.landurbplan.2022.104505.
The use of this file must follow the Creative Commons Attribution Non-Commercial No
Derivatives License, as required by Elsevier’s policy.
Highlights
⋅
⋅
⋅
⋅
⋅

Four Window View Indices (WVIs) were defined for measuring outside greenery,
water-body, sky, and construction views.
WVIs complemented existing view indices from the ground, aircraft, and satellites for
urban computing.
City Information Model (CIM)-based view images were trustworthy data sources for
WVIs.
Automatic WVI assessment based on deep transfer learning with an ML regression
layer was performed.
Highly satisfactory (R2 > 0.95) and fast (3.08 s/view) assessment results from
experimental tests were obtained.

1

Abstract

2

Every windowed room has a view, which reflects the visibility of nature and landscape and

3

has a strong influence on the health, living satisfaction, and housing value of inhabitants.

4

Thus, automatic accurate window view assessment is vital in examining neighborhood

5

landscape and optimizing the social and physical settings for sustainable urban development.
1 Maosu LI, PhD Candidate.

Department of Urban Planning and Design, The University of Hong Kong, Pokfulam, Hong Kong SAR, China; Email:
maosulee@connect.hku.hk, : https://orcid.org/0000-0002-9970-4053

2 Fan XUE, Assistant Professor.
Department of Real Estate and Construction, The University of Hong Kong, Pokfulam, Hong Kong SAR, China; Email:
xuef@hku.hk, : http://orcid.org/0000-0003-2217-3693
*: Corresponding author, Tel: +852 3917 4174, Fax: +852 2559 9457; Email: xuef@hku.hk
3 Yijie WU, PhD Student.
Department of Real Estate and Construction, The University of Hong Kong, Pokfulam, Hong Kong SAR, China; Email:
yijiewu@connect.hku.hk, : https://orcid.org/0000-0003-1441-1583
4 Anthony G. O. YEH, Chair Professor.

Department of Urban Planning and Design, The University of Hong Kong, Pokfulam, Hong Kong SAR, China; Email:
hdxugoy@hku.hk, : http://orcid.org/0000-0002-0587-0588

1

6

However, existing methods are labor-intensive, inaccurate, and non-scalable to assess

7

window views in high-rise, high-density cities. This study aims to assess Window View

8

Indices (WVIs) quantitatively and automatically by using a photo-realistic City Information

9

Model (CIM). First, we define four WVIs to represent the outside (i) greenery, (ii) water-

10

body, (iii) sky, and (iv) construction views quantitatively. Then, we proposed a deep transfer

11

learning method to estimate the WVIs for the window views captured in the CIM.

12

Preliminary experimental tests in Wan Chai District, Hong Kong confirmed that our method

13

was highly satisfactory (R² > 0.95) and fast (3.08 s per view), and the WVIs were accurate

14

(RMSE < 0.042). The proposed approach can be used in computing city-scale window views

15

for landscape management, sustainable urban planning and design, and real estate valuation.

16

Keywords: Window view; View quality index; High-rise buildings; City information model;

17

Deep learning; Urban computing.

18

19

1 Introduction

20

High-quality views can promote the physical and mental health, satisfaction,

21

restoration, and productivity of inhabitants as shown by studies on psychology, physiology,

22

and urban health (Ulrich 1984; Lottrup et al. 2015; Waczynska et al. 2021). In general, high-

23

quality views often involve considerable proportions of natural features, such as greenery,

24

sky, and water body, which are preferred by people (Hellinga 2013). Although the world

25

population is migrating to cities (UNPD 2014), urban planners and citizens find it challenging

26

to optimize the visibility of nature and landscape for windows in cities, especially in high-

27

rise, high-density areas. The Covid-19 pandemic recently has limited people’s physical access

28

to nature in many places, further amplifying the benefits of high-quality window views.

29

Consequently, high-quality window views, as a scarce resource, have been found to have

30

considerable influence on real estate values and sustainable urban development in terms of

31

neighborhood satisfaction, psychological and physical well-being, and urban planning and

32

design (Benson et al. 1998; Bishop et al. 2004; Jim & Chen 2009; Baranzini & Schaerer

33

2011).

34
35

Researchers have developed a plethora of urban indices and computational methods to

36

assess various urban views from different dimensions and perspectives. For example, on the

37

global scale, satellite images can produce overhead view indices (Tucker 1979; McFeeters

38

1996), such as the Normalized Difference Vegetation Index (NDVI) for vegetation (Liu et al.
2

39

2016) and the Normalized Difference Water Index for blue space exposure (Helbich et al.

40

2019). At the city and neighborhood scale, photographs and videos taken by vehicle-borne

41

cameras can assess the street views (Li et al. 2015; Shen et al. 2017; Dong et al. 2018; Lu

42

2018). These quantified view indices considerably contribute to human-built-environment

43

studies, such as urban depression symptoms (Helbich et al. 2019). However, cities, especially

44

high-rise, high-density ones, are not flat, so overhead and street-view assessment methods

45

cannot correctly represent window views (Li et al. 2015). High-rise, high-density areas, like

46

Hong Kong urban areas, have characteristics of high-rise buildings, narrow compacted street

47

canyons, high-level plot ratios, and high building densities (Gong et al. 2018). Within this

48

kind of context, the view from the window of a 30/F of an apartment may be completely

49

different from that of a 3/F one.

50
51

Window view quality is receiving increased attention from researchers in the fields of

52

architecture, urban health, and property valuation. For example, an ideal architectural design

53

tends to assess the indoor design and the outdoor views holistically (Ko et al. 2021; Li &

54

Samuelson 2020). Urban health researchers often use survey, interview, or questionnaire

55

methods to classify the window view qualitatively (Lottrup et al. 2015; Masoudinejad &

56

Hartig 2020). Qualitative descriptions of housing quality such as “with deluxe sea view” and

57

“with hill view” have been popular in the housing and hostel market of urban areas such as

58

Hong Kong (Jim & Chen 2009), and Mediterranean coastal cities (Fleischer 2012). However,

59

the existing methods are challenging for the assessment of window views at the city level.

60

First, too many window views exist in a city to be represented and preprocessed (Li et al.

61

2015). Second, conventional methods are too laborious to assess millions of window views

62

for a city, and manual assessments are prone to various errors, such as preconceived notions

63

in questionnaires and subjective judgments in valuation (Helbich et al. 2019). Thus, an

64

automatic accurate assessment of window views can contribute to large-scale landscape and

65

urban studies, as well as related disciplines and industries, for billions of urban inhabitants.

66

The quantified views serving as a vertical view information hub can facilitate developers,

67

urban planners, and other decision-makers to make well-informed decisions in real estate

68

valuation, sustainable urban planning, e.g., green space planning for prioritized buildings, and

69

new town design for balanced natural view acquisition and high-quality landscape view

70

conservation, especially in high-rise, high-density cities.

71
72

This study aims to present a series of Window View Indices (WVIs) together with an
3

73

automatic assessment method based on City Information Models (CIMs) and deep transfer

74

learning. A CIM is a digital representation of the physical and functional characteristics of a

75

city, and it can serve as a shared-knowledge resource (Song et al. 2017; Xue et al. 2021).

76

With advanced remote sensing technologies, photo-realistic 3D CIMs become increasingly

77

accurate in geometry and affordable in price. Recently, researchers have applied virtual

78

cameras to CIMs to generate realistic images of 3D window views as needed (Li &

79

Samuelson 2020; Li et al. 2020). The proposed method in the presented study extends the

80

existing work with deep transfer learning to quantify massive quantities of view images.

81
82
83

The main contributions of this study are thus twofold:
i.

From a theoretical perspective, the WVIs and assessment methods in this study extend

84

the knowledge on computing window views in cities, especially in high-rise, high-

85

density areas. The WVIs complement the existing studies on overhead and street-level

86

urban views.

87

ii.

For urban planning and design, the assessment results of this study are automatic and

88

accurate for any window (or 3D viewport), thanks to the up-to-date CIM and deep

89

transfer learning model pre-trained on other urban datasets. The output WVIs can

90

facilitate planners, architects, and other decision-makers in optimizing the

91

neighborhood landscape, urban planning and design, and property valuation for

92

sustainable urban development.

93
94

The remainder of this study is organized as follows. The related work in literature is reviewed

95

in Section 2. The WVI definitions and the automatic assessment method are presented in

96

Section 3. Section 4 describes preliminary experiments and the results. The discussion and

97

conclusion are presented in Sections 5 and 6, respectively.

98

2 Literature review

99

2.1 Urban views

100

Numerous studies have been conducted to compute and analyze urban views, e.g., impacts on

101

human response (Roe et al. 2013) and economic development (Bishop et al. 2004; Jim &

102

Chen 2009). Examples are green, water, sky, and construction views. Such views are not only

103

of the interest in landscape and urban planning, but also attracting researchers in other

104

disciplines such as psychology, physiology, urban health, and real estate.

105

4

106

First, greenery is of great significance to urban dwellers’ psychological and physical health.

107

Classical theories such as the stress reduction theory (Ulrich 1983) and attentional restorative

108

theory (Kaplan S. 1995) have already shown this. For instance, green views can reportedly

109

heighten positive effects such as performance and vitality (Van den Berg et al. 2016), reduce

110

fears (Ulrich 1984), and block stressful thoughts (Roe et al. 2013). Other studies have shown

111

that people’s accessibility to greenery can increase their restorative potential (Pazhouhanfar

112

& Kamal 2014) and thus influence the recovery from surgery (Ulrich 1984), and promote

113

productivity and job satisfaction (Kaplan R. 2001; Lottrup et al. 2015). The green view

114

impacts have been more extensively related to topics such as mental fatigue, depression

115

(Helbich et al. 2019), and potential for violence and crime (Kuo & Sullivan 2001).

116
117

Water and sky views as blue elements enable housing to enhance human healthcare and

118

property value. High-quality water bodies benefit people by having better aesthetic

119

enjoyment and restorative potential (White et al. 2010), whereas viewing the sky offers

120

occupants the sight and feeling of openness and spaciousness (Kaya & Erkip 2001). They

121

found that exposures to water and sky, similar to green views, benefit health and well-being,

122

such as stress reduction (Ulrich 1981), increased physical activity (Gascon et al. 2017), high

123

restorative potential (Masoudinejad & Hartig 2020), and promotion of positive mood and

124

satisfaction (Kaplan R. 2001; Gascon et al. 2017). Meanwhile, as precious attributes of the

125

aesthetic landscape, water and sky views are of great value, especially in high-rise, high-

126

density areas. As a result, both are influential on the property price (Baranzini & Schaerer

127

2011; Fleischer 2012).

128
129

For construction views from buildings, streets, and roads, their aesthetics and scarcity affect

130

the preferences of humans. For instance, features such as constructed landmarks are desirable

131

in window views (Baranzini & Schaerer 2011; Damigos & Anyfantis 2011). By contrast,

132

studies also demonstrated that urban views with natural features are preferred by occupants

133

over plain and dull construction scenes (Ulrich 1981; Grinde & Patil 2009). In summary, the

134

four types of view features are worthy of assessment for windows in high-rise, high-density

135

areas.

136
137

2.2 Assessments of window views

138

Generally, window-view quality can be assessed by two methods, namely, subjective and

5

139

objective. First, numerous studies have utilized mostly a window view assessment according

140

to the participants’ subjective judgments on views (Lottrup et al. 2015; Li & Samuelson

141

2020; Masoudinejad & Hartig 2020). The window views are presented by physical forms,

142

such as photographs and virtual forms (e.g., virtual reality). Researchers and practitioners

143

first collect window views according to their research objects. Then, participants assess or

144

rank the window views by using interview forms and questionnaire tables. The assessment

145

results are not concrete owing to fuzzy scales and criteria. The assessment methods on the

146

participants’ subjective answers are also time-consuming (Helbich et al. 2019; Labib et al.

147

2021). Thus, subjective methods are limited to a small scale and cannot practically form

148

common standards to coalesce the window view information objectively and automatically.

149
150

Objective methods and indices have emerged in the last decade for quantifying vertical

151

views. An example is a simulation-based view index harnessing the power of techniques in

152

the Geographic Information System, Remote Sensing, and 3D modeling (Yu et al. 2016;

153

Labib et al. 2021). A traditional method, namely 3D visibility analysis, has been used to

154

examine neighborhood amenities at the site and ground levels (Turan et al. 2019; Labib et al.

155

2021). Particularly, Yu et al.’s (2016) method measures floor-level greenery view based on

156

the NDVI metric in a high-rise, high-density context, though the oversimplified 2.5D

157

greenery can lead to errors. Alternatively, view photography method can effectively compute

158

and analyze the real profile view of landscapes (Li et al. 2015; Shen et al. 2017; Dong et al.

159

2018). Recently, the method has also been used in Li et al.’s (2020) two-class window view

160

classification, i.e., “nature” and “construction,” based on Apriori rules and a transfer learning

161

model. However, Li et al.’s (2020) method relies on rigid classification rules and has only

162

two types of features. Thus, next-generation objective assessment methods should be able to

163

adapt to more urban scenes, with up-to-date machine learning (ML) technologies.

164
165

2.3 Deep learning and applications in urban studies

166

Deep learning is a group of multi-layer artificial neural networks involving multiple levels of

167

representation learning (LeCun et al. 2015). Deep learning models have shown strengths in

168

general pattern recognition tasks (LeCun et al. 2015). For instance, SegNet as one of the best

169

deep convolutional network models has been used in visible landscape segmentation and

170

quantification tasks (Liang et al. 2017; Shen et al. 2017). To study the relationship of natural

171

features including greenery and water with geriatric depression in Beijing, China, a fully

6

172

convolutional neural network (FCN-8s) was used to segment the street view images into

173

green parts and blue parts (Helbich et al. 2019).

174
175

Deep transfer learning adopts a pre-trained network, inductively or transductively, from a

176

source domain to the target domain on the basis of a mapping mechanism (Pan & Yang

177

2009). A small training dataset in the target domain can effectively map the variables’

178

relationships and transfer the pre-trained network. Deep transfer learning has become

179

prevalent for saving the time and resource costs in labeling training data with negligible

180

performance downgrades from the original model. Thus, it is widely used in the semantic

181

understanding of urban research, such as environmental management (Chen et al., "Looking

182

beneath the surface”: A visual-physical feature hybrid approach for unattended gauging of

183

construction waste composition 2021; 2022), urban morphology (Middel et al. 2019), and

184

perception (Yao et al. 2019; Li et al. 2020). For instance, fed by the street view images, FCN-

185

8s pre-trained on the ADE20K dataset was transferred in water and greenery extraction of

186

streetscape (Helbich et al. 2019). All previous studies have confirmed that deep transfer

187

learning can be a versatile and inexpensive instrument from one domain to a similar domain

188

application. Thus, for large-scale window view quality assessment and applications, deep

189

transfer learning can provide cost-effective support for the semantic segmentation of the

190

view.

191
192

In summary, large-scale window view assessment, especially the automatic method, has

193

previously been a conundrum owing to the poor availability of window data and immature

194

window view reconstruction and processing. Meanwhile, textured CIMs, deep transfer

195

learning, and other learning technologies may open a window of opportunity to improve the

196

automatic window view assessment for high-rise, high-density areas significantly.

197

198

3 Research methods

199

Figure 1 shows the proposed method as an Icam DEFinition for Function modeling (IDEF0)

200

diagram, which is a public-domain flowchart-like methodology for modeling processes and

201

functions (Colquhoun et al. 1993). The legend in Figure 1 explains the inputs, methods and

202

tools, control parameters, and final outputs of each sub-process. The proposed automatic

203

window view assessment method comprises three steps: (i) batch generation, (ii) semantic

204

segmentation of pixels, and (iii) estimation of view indices. Each step employs a specific
7

205

method and control parameters. Overall, the main inputs are a 3D photo-realistic CIM and

206

corresponding 2D building footprints in this study. The output is a set of quantified WVIs.

207

Finally, post-processing enriches the input CIM with the WVIs for smart decision-making for

208

landscape and urban planning and related disciplines. A practitioner can follow the same

209

methods and tools in Figure 1 and adjust the control parameters for specific application

210
211

scenarios.

212
213
214

Figure 1. IDEF0 (Colquhoun et al. 1993) diagram of the proposed method for assessing
window views.

215
216

3.1 Definitions of WVIs

217

3.1.1 Window view index

218

This study defines the WVI as the ratio of pixels for each view type. Given a view image v =

219

{ pij | 1 ≤ i ≤ M, 1 ≤ j ≤ N } of M × N pixels and a finite set L of views, as shown in Figure 2,

220

the WVI in an input window view image is the ratio:

221
222
223
224
225
226

WVIl =

|{p | p ∈ v, λ(p) = l}|

, l ∈ L,

M×N

(1)

where λ(p) = l is the semantic label of a pixel p, e.g., “green” or “waterbody”, and | · | is the

cardinality operator indicating the total number of pixels. Thus, all WVIs are scalars bounded
between 0 and 1:

WVIl ∈ [0, 1]

, l ∈ L.

(2)

227

We select the four major types of window views as summarized in Section 2.1. That is, L =

228

{‘green’, ‘waterbody (water)’, ‘sky’, ‘construction (const.)’}, as shown in Figure 2. Table 1

229

lists the common city objects’ mapping to L. For instance, the “green” view type covers all

230

kinds of greenery, including trees, bushes, and grasses. Four symbols, namely, WVIgreen,

231

WVIwater, WVIsky, and WVIconst., represent the scalar values, respectively. Despite the presence
8

232

of other possible city objects such as pedestrians, pets, vehicles, and aircraft, the four major

233

types are dominant in window views in our experiment, i.e., WVIgreen + WVIwater + WVIsky +

234

WVIconst. ≈ 1, as shown in Figure 2. Furthermore, the ratio-based definition is consistent and

235

robust for the comparison of views from different window sizes across districts and cities,

236

which is helpful for the proof-of-concept purpose in this study, e.g., a window with WVIgreen =

237

0.8 owns more proportions of greenery and can thus be regarded as a totally green-view

238

window, compared with another having WVIgreen = 0.4.

239
240

Figure 2. Examples of the four Window View Indices (WVIs).

241
242

Table 1. List of types of views and associated common city objects.
Type
Green
Waterbody
Sky
Construction

Symbol
WVIgreen
WVIwater
WVIsky
WVIconst.

Example objects
Trees, bushes, and grasses
Sea, lakes, ponds, and rivers
Sky, clouds, and fog
Building facades, roofs, walls, streets, houses, and roads

243
244

3.1.2 Window view ranking

245

Furthermore, the relative window view ranking (WVR) of a window’s WVI within a high-

246

rise, high-density area A can be defined as the percentile to the maximum WVI of the context:

247

Very high, WVRAl ∈ [0.8, 1.0]
⎧
A
⎪ High, WVRl ∈ [0.6, 0.8)
WVI
l
WVR Al =
= Average, WVRAl ∈ [0.4, 0.6)
max�WVIl𝐴𝐴 � ⎨
A
⎪ Low, WVRl ∈ [0.2, 0.4)
⎩ Very low, WVRAl ∈ [0.0, 0.2)
9

, l ∈ L.

(3)

248

Therefore, the WVR classifies all the windows in an area into five isometric groups. WVR

249

can resolve the issue of inconsistent upper bounds of different WVIs, which enables an inter-

250

view-type comparison. For instance, although max(WVIsky) is roughly 0.5 and max(WVIgreen)

251

is 1.0 theoretically, max(WVRsky) and max(WVRgreen) can still reach 1.0. Thus, one window

252

with absolute WVIsky = 0.5 and WVIgreen = 0.5 can be tagged as a “very high”-level sky view

253

but an “average”-level green view within the context. People’s decision-making is expected

254

to be associated more with WVRs than WVIs, e.g., in property valuation.

255

3.2 Proposed assessment method

256

3.2.1 Batch generation of window view images

257

The first step aims to generate the window view images in an urban area in batch. The image

258

extraction process, as shown in Figure 3, is automatic on 3D GIS platforms with camera

259

functions, such as Cesium (Cesium GS 2022). Figure 3a shows a window’s 3D geolocation

260

(lng, lat, height) and heading direction are computed on the facade of extruded footprints by

261

building height information, where the heading direction is assumed perpendicular to the

262

facade at (lng, lat). The field of view is set to 60° to represent the normal human field of

263

vision (FoV) (Tara et al. 2021), while the pose of the virtual camera is set on the window with

264

tilt = 0 and pitch = 0 to capture views. The image extraction extends Li et al. (2020) as the

265

camera’s view of the photo-realistic CIM’s textured appearance. The difference from Li et al.

266

(2020) is the full automation for massive windows using a JavaScript program as shown in

267

Figure 3b.

268
269
270

Figure 3. Batch generation of window view images. (a) Window location computation,
camera settings, and (b) image generation process.
10

271
272

However, neighboring windows on the same facade often share similar views. Thus,

273

sampling the facade with certain intervals, e.g., every 10 or 20 m, is a cost-effective method,

274

as shown in Figure 3b, which can considerably save computational effort without losing

275

notable WVI accuracy. Based on the efficient sampling and GIS-based view visualization, the

276

batch generation can extract view images for the windows of a high-rise, high-density area.

277

Learned from experiments and sensitivity analysis results in Section 4, we used 20 and 5 m to

278

obtain a location matrix of view sites within the large and small facades, respectively.

279
280

3.2.2 Deep transfer learning-based semantic segmentation

281

This step classifies every pixel in an input image to a semantic view label through deep

282

transfer learning. One of the most relevant deep learning datasets is the Cityscapes

283

benchmarking dataset (Cordts et al. 2016), which comprises 25,000 urban views annotated as

284

19 pixel-level labels from 50 cities in Germany. According to the study of Pan & Yang

285

(2009), the models trained in Germany can potentially be transferred to other areas like Hong

286

Kong. Table 2 lists the labels for Cityscapes in seven groups. Apparently, three types of

287

views, i.e., green, sky, and construction, can be directly mapped from Cityscapes’ definitions.

288
289

Table 2. List of labels for the Cityscapes dataset and for WVIs in this study.

290

Group
Labels for Cityscapes
Labels for WVIs
a
Nature
Vegetation, Terrain
Green
Sky
Sky
Sky
Construction Building, Wall, Fence
Const.
Paved
Road, Sidewalk
Const.
Object
Pole, Traffic sign, Traffic light
Const.
Human
Rider, Person
–b
Vehicle
Car, Truck, Bus, Motorcycle, Bicycle, On rails
–b
a: Including all kinds of horizontal vegetation in Cityscapes; b: Negligible in this study.

291
292

A Deeplab (Ver. 3+ with the Xception_65 backbone) model pre-trained on Cityscapes (Chen

293

et al. 2018; Xia et al. 2021) is transductively transferred to the segmentation of captured

294

window view images to the labels in Table 2. The off-the-shelf Deeplab model is one of the

295

top open-source deep learning models for urban views, where the training parameters can be

296

referred to (Chollet 2017) and (Chen et al. 2018). Xue et al. (2021) showed that transductive

297

transferring Deeplab leads to an efficient and low-cost semantic segmentation of view

298

images, even though the training and target datasets are from different contexts. As shown in
11

299

Figure 4, the incorporated version of Deeplab has a network architecture consisting of two

300

parts, i.e., an encoder and a decoder (Chen et al. 2018). The encoder mainly includes an

301

Atrous Spatial Pyramid Pooling Module (ASPP) for concatenated features from a low-level

302

Atrous convolution (Chollet 2017), while the decoder concatenates the ASPP outputs and

303

low-level features with convolution and upsampling.

304
305
306

Figure 4. Three types of semantic segmentation errors in direct transductive transferring of
Deeplab. (a) Undefined labels, (b) segmentation errors, and (c) from input noises.

307
308

However, as shown in Figure 4, the segmentation results of a direct transductive transferring

309

were erroneous and unsatisfactory for WVIs in the study area. The primary source of errors

310

was from the inconsistent labels, e.g., the water body, between the training dataset Cityscapes

311

and our view images. Besides, minor errors resulted from the segmentation and input noises.

312

Therefore, deep transfer learning can deliver pixel-level semantic segmentation with relevant

313

labels for window view images, but the results must be corrected for the errors to improve the

314

accuracy in computing WVIs and WVRs using the ML-based WVI regression layer described

315

in Section 3.2.3 below.

316

3.2.3 ML-based regression for WVIs

317

This step applies an ML-based WVI regression, as shown in Figure 5, to correct the errors

318

from deep transfer learning for computing WVIs. The input features to the regression are

319

Cityscapes labels in terms of proportions of pixels segmented by Deeplab in Figure 4. The

320

outputs are the four WVIs, i.e., WVIgreen, WVIwater, WVIsky, and WVIconst. We annotate a small

321

set of window view images with five labels, i.e., green, waterbody, sky, construction, and

322

others (e.g., terrain and vehicles), which provide ground truth WVIs for the training process.

323

The candidate ML models include Decision Trees, Linear Regression, Support Vector

324

Machines (SVMs), kNN, Artificial Neural Network (ANN), Random Forest, and Adaboost. A
12

325

standard train-compare-finetune pipeline is applied to select appropriate ML models to

326

estimate the WVIs through cross-validations. For each type of WVI, the most accurate ML

327

model (together with its parameters) is selected for the regression layer. As a result, the first

328
329

two types of errors shown in Figure 4 can be considerably reduced.

330
331

Figure 5. ML-based regression layer for estimating WVIs.

332
333

The results of ML training are compared with the actual values of the four view types from

334

view image annotation using root-mean-squared error (RMSE):

335

∑

RMSE = � l∈L

(Predl −WVIl )2
n

,

(4)

336

where WVIl indicates the actual value for the view type l, Predl is the estimated value, and n

337

denotes the number of window view images. The ML model trained with the minimum

338

RMSE is selected for WVI estimation. We utilize 10-fold cross-validation for unbiased

339

RMSEs.

340
341

3.3 Post-processing for semantic enrichment of CIM

342

The estimated WVIs are post-processed to enrich the semantics of 3D CIM, which can

343

conveniently support future applications in related domains as a common knowledge

344

platform. The detailed workflow is as follows. First, geocoded view sites with WVIs are

345

registered at the 3D globe. Then, regarding view sites within the same facade as a group of

346

vertices, we triangulate them to reconstruct the building facade through a classic Delaunay
13

347

method (Lee & Schachter 1980). Thereafter, a linear interpolation-based 3D rendering

348

(Akenine-Möller et al. 2019) of WVIs visualizes the whole building facades in a mesh

349

surface. The interpolation result also estimates the WVIs of all locations of the building

350

facades. Finally, the CIM is enriched with the WVI semantics for a spectrum of applications

351

in landscape management, sustainable urban planning and design, and real estate valuation.

352

4 Experimental tests

353

4.1 Experimental area and settings

354

The study area was Wan Chai in Hong Kong, as shown in Figure 6a. Wan Chai is one of the

355

highest residential density zones according to the Hong Kong Planning Standards and

356

Guidelines (HKPlanD 2018). The average of building heights is 35.5 m and the 75th

357

percentile is over 48 m. The study area owns a plot ratio at 8.0 and building density at 0.29

358

(the ratio of building site area to land area). Although the area enjoys considerable sky, sea,

359

and greenery view contents, the visibility of natural features is often blocked by other

360

buildings. The 2D footprint data with building height information were extracted from the

361

iB1000 digital topographic map of Hong Kong (HKLandsD 2014) as shown in Figure 6b, and

362

converted into the GeoJSON (Butler et al. 2016) format for batch attribute computations of

363

view sites’ locations and headings. The data source of 3D photorealistic CIM was produced

364

and freely shared by the Planning Department of Hong Kong (2019) as shown in Figure 6c.

365

We calibrated the CIM as 3D tiles to the correct geographical locations on the WGS-84 globe.

366

Then, 2D and 3D datasets were loaded and registered in an open-source 3D GIS platform

367

named “Cesium ion.” Ten buildings with typical different built environments from the seaside

368

to the mountain area were selected as case studies to examine the proposed approach, as

369

shown in Figure 6d.

14

370
371
372

Figure 6. Study area of Wan Chai, Hong Kong. (a) Location of Wan Chai, (b) building
footprints, (c) input CIM, and (d) location of 10 case study buildings.

373
374

The computational experiments were set up as follows. The workstation comprised an Intel

375

i7-10700 CPU (2.90GHz, 16 cores), 128 GB memory, one Nvidia GeForce RTX 2070

376

graphic card, and Ubuntu 20.04 (64-bit) operating system. Sample window views were

377

collected on the Cesium platform (ver. 1.75). Deep transfer learning was in the environment

378

of Tensorflow (ver. 2.4) and Python (ver. 3.6). We adopted the seven ML models

379

implemented on Orange (ver. 3.26), a Python ML platform. From the case study buildings,

380

110 training examples were selected for unbiased representation of diversified window views

381

and manually annotated with the WVIs for training the ML models. The one-off annotation

382

work consumed about 10 person-hours. The size of training examples satisfied the

383

requirements of deep transfer learning. We set each view image with 900 × 900 pixels to

384

represent the view features seen from the window.

385
386

4.2 Results

387

Results showed that the proposed method is automatic and efficient, as shown in Table 3. The

388

first step of batch generation returned 1,416 window view images from the 10 selected

389

buildings for the case study. The average time for generating one view image was 2.00 s. The

390

deep transfer learning processed the view images at an average time of 1.08 s in the second

391

step. The ML-based regression estimated the WVIs in <0.001 s on average for each image.
15

392
393

Table 3. Computational time of the proposed method for a window view image.
Step
1
2
3

394

Processing
CIM-based batch generation
Deep transfer learning
ML-based regression

Software library
Cesium (ver. 1.75)
Deeplab (ver. 3+)
Orange (ver. 3.26)

Average time (s)
2.00#
1.08
0.00*
Total
3.08
#: A pre-set value that can be fine-tuned by workstation performance; *: Less than 0.001 s.

395
396

The WVIs’ assessments of the proposed method were also satisfactory. Table 4 shows that for

397

the best model of the four view indices’ estimation, the R² values were 0.952, 0.965, 0.978,

398

and 0.977 respectively, which represented more than 95% of the variance in the dependent

399

variables. The RMSEs of the four training models were 0.021, 0.022, 0.025, and 0.042,

400

respectively. The optimal parameter of each best model was as follows. For WVIgreen, the

401

Linear Regression model was trained with Lasso (L1) regularization and strength at 0.0001.

402

For WVIwater, the SVM model performed the best, with kernel = RBF, C = 0.9, gamma = 0.05.

403

For WVIsky, a Linear Regression model with an elastic net regularization (L1:L2=0.50:0.50)

404

was utilized with the best accuracy of estimation, whereas for WVIconst., the best estimation

405

was observed from a Linear Regression model with a Ridge (L2) regularization (Alpha =

406

0.003).

407
408

Table 4. Training errors and time of the best model for four WVIs.
WVI
Green

Best model
Linear
Regression

Parameters
L1 = 0.0001

RMSE
0.021

R²
0.952

Training time (s)
0.077

Water

SVM

Kernel = RBF, C =
0.9, gamma = 0.05

0.022

0.965

0.154

Sky

Linear
Regression

L1:L2 = 0.50:0.50

0.025

0.978

0.070

Const.

Linear
Regression

L2 =0.003

0.042

0.977

0.091

409
410

WVRs were computed from the WVIs by the best model. Table 5 shows three typical window

411

views and their WVIs and WVRs. In Table 5, a WVR is represented in an array of stars,

412

showing the level from “very low” to “very high” in Eq. 3. The highest WVRs correctly

413

reflected the given dominant features for all the samples.

414

16

415

Table 5. Sample WVIs and WVRs for typical sample window views.

View images
Dominant feature
Feature
Max.
Green
0.5421
Water
0.4375
Sky
0.5505
Const.
1.0000

Sky
WVI
0.0165
0.3352
0.4682
0.1870

416

Green
WVI
0.4867
0.0024
0.3236
0.1704

WVR
⋆
⋆⋆⋆⋆
⋆⋆⋆⋆⋆
⋆

WVR
⋆⋆⋆⋆⋆
⋆
⋆⋆⋆
⋆

Construction
WVI
WVR
0.0130
⋆
0.0000
⋆
0.0928
⋆
0.9057
⋆⋆⋆⋆⋆

417

4.3 Post-processing for enriching CIMs

418

In the post-processing, the estimated WVIs and WVRs were registered for enriching the

419

semantics of input CIM. Figure 7 shows the 3D mesh model of the regional WVIs in the

420

study area. Generally, most rooms of the buildings owned a high WVIconst. in this area as

421

shown in Figure 7d. Figure 7b shows that only windows facing the seaside in the high-rise

422

buildings near the harbor can have high-level WVIwater values in Wan Chai. Great sky views

423

were scattered across the rooms with the high storeys as shown in Figure 7c. Figure 7a shows

424

the generally low and fluctuated WVIgreen, reflecting the varied amount of the surrounding

425

greenery at different locations. In summary, the disparity of possession of natural view

426

resources, i.e., greenery, water, and sky, is significant in the study area. The quantified

427

disparity can help the urban planners to make a more accurate and specific decision for future

428

landscape management and urban planning, e.g., prioritized greenery planning for buildings

429

without any nature views.

17

430
431

Figure 7. Regional patterns of WVIs. (a) WVIgreen, (b) WVIwater, (c) WVIsky, and (d) WVIconst..

432
433

Figure 8 shows a WVR-enriched comparison of two example north-facing facades, one

434

nearby and the other far away from the seafront, of which the locations are marked in Figure

435

8e. Holistically, water and sky views of the first facade were above the “average” levels in the

436

study area (≥ 40%), as shown in Figures 8b and 8c; in contrast, the levels of those views of

437

the second facade were consistently lower due to the inter-building obstruction. Figure 8a

438

shows the green views were both at a “very low” level (< 20%) due to the less visible

439

greenery. The construction view patterns of the two facades varied as shown in Figure 8d,

440

where construction views dominated the second facade. In comparison with WVI values, the

441

relativity in such WVR results is more convenient for certain applications such as real estate

442

valuation, since the levelization of the window view such as “very high” and “very low” can

443

intuitively inform developers and occupants of the room view quality within the local

444

context.

445

18

446
447

Figure 8. WVR patterns of two example building facades. (a) WVRgreen, (b) WVRwater, (c)
WVRsky, (d) WVRconst., and (e) their general locations.

448
449

4.4 Sensitivity analysis

450

4.4.1 View sampling interval in Step 1

451

A trade-off existed between processing time cost and accuracy when applying the view

452

sampling interval in Step 1. A sensitivity analysis was conducted to identify a cost-effective

453

sampling plan. In the experiments, the case was a facade area (120 m × 60 m) of the China

454

Resources Building, as shown in Figure 9a. The benchmark was set to the result of a 5 m

455

sampling interval. We tested a range of sampling intervals from 10 m to 60 m in an

456

approximately exponential increment. Figure 9b shows the example of WVIsky estimation

457

results resampled back to the 5 m scale through linear interpolation to compare the accuracies

458

in terms of RMSE. We found that with increased sampling interval, the time consumption of

459

the window view image processing from generation to estimation witnessed a sharp decline,

460

whereas the RMSEs of four WVIs increased accordingly, as shown in Figure 9c. From the

461

observation, the sample interval of 20 m can be a “sweet point,” in which an efficient and

462

accurate estimation of WVI (RMSE<0.015) was obtained without excessive processing time.

463

Thus, for the view image processing of case buildings, we used 20 m as the sampling interval

464

for large facades. For a building facade whose length or width was less than 20 m, 5 m was

465

used.

19

466

467
468
469

Figure 9. Sensitivity analysis of sampling intervals. (a) A case facade, (b) estimated WVIsky at
different sampling intervals, and (c) trade-off between time cost and four WVIs’ accuracy.

470
471

4.4.2 Input CIM in Step 1

472

Figure 10 compares view segmentation results using two different CIMs. The appearances of

473

the two 3D models were close but clearly distinguishable. First, the color contrast of Google

474

Earth’s CIM was softer than the model adopted in this study, and the low contrast resulted in

475

the misclassification of constructions and greenery highlighted in Figure 10a. Second, the

476

model fidelity also affected the stimulation effects. Figure 10b shows that some parts of the
20

477

vegetation view (as highlighted in the rectangles), which were wrongly segmented using our

478

CIM, can be corrected using Google Earth’s model. This finding was due to the higher

479

quality of Google Earth’s in expressing the vegetation features, especially in close range.

480

Lastly, the distortions in CIMs affected the segmentation accuracy. As shown in the red

481

rectangles in Figure 10c, the blurred facades in the left column resulted in inaccurate

482

segmentation, whereas the distortions in Google Earth’s model led to the wrong detection of

483

buildings to vegetation.

484
485
486

Figure 10. Comparison of window view image segmentation (Step 2) against different CIMs.
(a) View color, (b) view fidelity, and (c) view distortions.

487
488

4.4.3 ML models for regression in Step 3

489

Based on the R², the performance of trained models is examined, and results are shown in

490

Figure 11. For the estimation of the four WVIs, all ML models had R² values greater than 0.7.

491

For three types of WVIs, i.e., WVIgreen, WVIsky, and WVIconst., the best models were produced

492

by Linear Regression. For the WVIwater estimation, the best model was SVM, whereas the

493

Linear Regression returned R² > 0.93. The satisfactory results from Linear Regression might

494

echo the assumption that four window view types could be mapped directly from the urban

495

street view features in high-rise, high-density areas.

21

496
497

Figure 11. Comparison of R² performances of the seven ML models.

498

499

5 Discussion

500

5.1 Significance

501

Large-scale window view assessment has a great potential to support many smart city

502

applications. The window view quality is of great significance for residents in high-rise, high-

503

density areas. In the post-Covid-19 era, window view plays an important role in accessing

504

nature as people have to stay longer in their houses or offices. The quantitative window view

505

quality assessment at the city scale can provide an intuitive understanding of environmental

506

inequality. Planners can use the results to prioritize improvements of the poor living

507

environments, such as prioritized provision of more green space for neighborhoods with poor

508

window views. And government sectors and policymakers can make the regulations, e.g.,

509

minimum acquisition of nature views in the future sustainable urban development. The results

510

can also facilitate urban and architectural design by quantifying the window view quality at a

511

relatively low cost. Designers can integrate the quantified view results for more

512

comprehensive generative designs of building spaces (Laovisutthichai et al. 2021) and new

513

towns. In addition, the method can serve as a new indicator for the housing market and thus

514

has a great potential to the architecture, engineering, and construction development.

515
516

In the past, surveyors had to enter real rooms of buildings to capture the window views.

517

Owing to this time-consuming, labor-intensive task, the window view dataset is always

518

limited (Labib et al. 2021). Furthermore, accessing all window views manually at a large

519

scale becomes impossible in terms of cost, labor force, and privacy (Helbich et al. 2019).

520

Nowadays, with the advancement of remote sensing, photogrammetry, and digital twin

521

technology, mature 3D CIMs with high-quality textured appearances are becoming
22

522

increasingly available for detecting multiple groups of view features. CIM-based simulated

523

window views for the real world have been validated effectively (Li & Samuelson 2020; Li et

524

al. 2020). However, for an urban-scale window view quality evaluation, processing a large

525

number of views manually remains laborious and expensive for surveyors. The proposed

526

window view quality assessment method can free humans from repetitive and time-

527

consuming tasks, and provide a set of quantifiable indicators to support fundamental and

528

derivative applications in window view quality evaluation.

529
530

The proposed automatic assessment method can effectively generate four major view indices

531

for quantifying and analyzing the urban-scale window views. First, this study makes full use

532

of volumetric landscapes from 3D photo-realistic CIMs to further enrich the CIM with four

533

WVIs, thereby enabling many window-view-based digital twin city applications, such as 3D

534

city living environment assessment and housing scenic quality comparison. From a

535

practitioners’ point of view, the method is easy-to-use, low-cost, and accurate. For example,

536

the automation process can be implemented without considerable prior knowledge. The pre-

537

trained Deeplab model was shared freely. Based on the transfer learning theory, only a small

538

dataset is required for a satisfactory WVI assessment. Moreover, the experimental results

539

confirmed a high accuracy of assessing the window views (R² > 0.95). In summary, the

540

proposed method contributes to window view assessment using CIM and AI, and also

541

provides relatively low-cost and high-accuracy WVIs for applications in urban planning and

542

design, and property valuation.

543
544

5.2 Limitations and future work

545

Nevertheless, a few limitations exist in the work presented in this study. First, the assessed

546

window view quality in this study only involved limited contents, including greenery, sky,

547

water body, and construction. Movable city objects e.g., pedestrian, car, and rare urban

548

features e.g., bare soil surface were not involved. Other view elements exerting influence on

549

indoor living satisfaction and outdoor environment perception such as aesthetic and

550

environmental quality, view distance, and layer were not considered. Second, the horizontal

551

view was set to compute the WVIs, which might miss visible features from other directions,

552

e.g., the ground level. Next, another limitation was the high workload of 2D image

553

segmentation involving repeated computation. For instance, similar view images from

554

neighboring windows were independent without reusing the intermediate segmentations. The

555

computation cost could be slightly higher for irregular buildings due to more view samples
23

556

and processing. Last, the window sampling and interpolation also led to possible accuracy

557

losses.

558
559

Future directions to improve the presented study are as follows. The first is extending the 2D

560

image format of window views to incorporate high-dimensional factors (e.g., fine-scale

561

classified view features, view distance that influences residents’ feeling of spaciousness, and

562

aesthetics and environmental quality attributes that influence living satisfaction) for holistic

563

quality and optimization. More FoVs, such as 360-views, can extend the WVIs assessed in

564

the 60° horizontal views in this study. Well-labelled CIM for landscapes is proven effective

565

for large-scale view quantification (Yu et al. 2016). Thus, a 3D segmented CIM may

566

eliminate the repetitive and redundant 2D image segmentation and save considerable costs of

567

training and applying deep transfer learning, especially for irregular buildings. Another

568

direction is to identify the accurate 3D location and orientation for each physical window in

569

the CIM so that the assessed WVIs and WVRs can be associated with windows and rooms.

570

571

6 Conclusion

572

A high-quality window view with enough features such as greenery, sky, and water not only

573

has a good impact on residents’ health, well-being, and performance, but also can enrich the

574

value of the house, especially in high-rise, high-density areas. Traditional window view

575

assessment methods have common problems such as subjectivity, scalability, and efficiency.

576

To address these limitations, this study uses an automatic method for the large-scale window

577

view quality assessment through the use of CIM-based window view images of city

578

buildings.

579
580

This study defines an indicator named Window View Index (WVI) including four sub-indices

581

i.e. Green view index, Water view index, Sky view index, and Construction view index,

582

which are measured at one time efficiently. By implementing a fast-sampling method, outside

583

views are captured at each view site of the 3D CIM at the initial stage. Then, a pre-trained

584

deep transfer learning model is used to classify view images into multiple features efficiently.

585

To construct the regression between detected features and the WVI, seven traditional machine

586

learning models are tuned to achieve the best performance. Our method achieved highly

587

satisfactory results in estimating the WVIs for the high-rise, high-density area, in Wan Chai,

588

Hong Kong. The RMSEs of estimation did not exceed 0.042, whereas the average time of
24

589

processing each window was 3.08 s.

590
591

The proposed method provides intuitive indicators of the window view quality for high-rise,

592

high-density areas. The automatic, accurate method is scalable to the urban scale, thereby

593

enabling many window view-based applications in landscape management, sustainable urban

594

planning and design, and real estate valuation, which would benefit residents’ health, urban

595

optimization, and the housing industry. Future work includes extending the view indices, 3D

596

semantic segmentation of CIM, and mapping the WVIs to physical windows and rooms.

597

Acknowledgment

598

The work presented in this study was supported in part by the Hong Kong Research Grants

599

Council (RGC) under the Early Career Scheme (ECS) No. 27200520 and the Guangdong–

600

Hong Kong-Macau Joint Laboratory for Smart Cities of the 2020 Guangdong New

601

Innovative Strategic Research Fund, Guangdong Science and Technology Department

602

(Project No.: 2020B1212030009)

603

604

References

605

Akenine-Möller, T., Haines, E. & Hoffman, N. (2019). Real-Time Rendering (3rd ed.). New

606
607

York: A K Peters/CRC Press. doi:10.1201/9781315365459
Baranzini, A. & Schaerer, C. (2011). A sight for sore eyes: Assessing the value of view and

608

land use in the housing market. Journal of Housing Economics, 20(3), 191-199.

609

doi:10.1016/j.jhe.2011.06.001

610

Benson, E. D., Hansen, J. L., Schwartz, A. L. & Smersh, G. T. (1998). Pricing residential

611

amenities: the value of a view. The Journal of Real Estate Finance and Economics,

612

16(1), 55-73. doi:10.1023/A:1007785315925

613

Bishop, I. D., Lange, E. & Mahbubul, A. M. (2004). Estimation of the influence of view

614

components on high-rise apartment pricing using a public survey and GIS modeling.

615

Environment and Planning B: Planning and Design, 31(3), 439-452.

616

doi:10.1068/b3042

617
618
619
620

Butler, H., Daly, M., Doyle, A., Gillies, S., Hagen, S. & Schaub, T. (2016). The geojson
format. California, USA: Internet Engineering Task Force (IETF).
Cesium GS. (2022). The Cesium Platform. Philadelphia, USA: Cesium GS, Inc. Retrieved
from https://cesium.com/platform/

25

621

Chen, J., Lu, W. & Xue, F. (2021). "Looking beneath the surface”: A visual-physical feature

622

hybrid approach for unattended gauging of construction waste composition. Journal

623

of Environmental Management, 286, 112233. doi:10.1016/j.jenvman.2021.112233

624

Chen, J., Lu, W., Yuan, L., Wu, Y. & Xue, F. (2022). Estimating construction waste truck

625

payload volume using monocular vision. Resources, Conservation and Recycling,

626

177, 106013. doi:10.1016/j.resconrec.2021.106013

627

Chen, L.-C., Zhu, Y., Papandreou, G., Schroff, F. & Adam, H. (2018). Encoder-decoder with

628

atrous separable convolution for semantic image segmentation. Proceedings of the

629

European Conference on Computer Vision (ECCV) (pp. 801-818). Munich, Germany:

630

Springer Science. doi:10.1007/978-3-030-01234-2_49

631

Chollet, F. (2017). Xception: Deep learning with depthwise separable convolutions.

632

Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

633

(pp. 1251-1258). Honolulu, Hawaii: IEEE. doi:10.1109/CVPR.2017.195

634

Colquhoun, G. J., Baines, R. W. & Crossley, R. (1993). A state of the art review of IDEFO.

635

International Journal of Computer Integrated Manufacturing, 6(4), 252-264.

636

doi:10.1080/09511929308944576

637

Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U.,

638

Roth, S. & Schiele, B. (2016). The cityscapes dataset for semantic urban scene

639

understanding. Proceedings of the IEEE Conference on Computer Vision and Pattern

640

Recognition (pp. 3213-3223). Las Vegas, Nevada: IEEE. doi:10.1109/CVPR.2016.350

641

Damigos, D. & Anyfantis, F. (2011). The value of view through the eyes of real estate

642

experts: A Fuzzy Delphi Approach. Landscape and Urban Planning, 101(2), 171-178.

643

doi:10.1016/j.landurbplan.2011.02.009

644

Dong, R., Zhang, Y. & Zhao, J. (2018). How green are the streets within the sixth ring road of

645

Beijing? An analysis based on tencent street view pictures and the green view index.

646

International Journal of Environmental Research and Public Health, 15(7), 1367.

647

doi:10.3390/ijerph15071367

648

Fleischer, A. (2012). A room with a view—A valuation of the Mediterranean Sea view.

649

Tourism Management, 33(3), 598-602. doi:10.1016/j.tourman.2011.06.016

650

Gascon, M., Zijlema, W., Vert, C., White, M. P. & Nieuwenhuijsen, M. J. (2017). Outdoor

651

blue spaces, human health and well-being: A systematic review of quantitative

652

studies. International Journal of Hygiene and Environmental Health, 220(8), 1207-

653

1221. doi:10.1016/j.ijheh.2017.08.004

654

Gong, F.-Y., Zeng, Z.-C., Zhang, F., Li, X., Ng, E. & Norford, L. K. (2018). Mapping sky,
26

655

tree, and building view factors of street canyons in a high-density urban environment.

656

Building and Environment, 134, 155-167. doi:10.1016/j.buildenv.2018.02.042

657

Grinde, B. & Patil, G. G. (2009). Biophilia: does visual contact with nature impact on health

658

and well-being? International Journal of Environmental Research and Public Health,

659

6(9), 2332-2343. doi:10.3390/ijerph6092332

660

Helbich, M., Yao, Y., Liu, Y., Zhang, J., Liu, P. & Wang, R. (2019). Using deep learning to

661

examine street view green and blue spaces and their associations with geriatric

662

depression in Beijing, China. Environment International, 126, 107-11.

663

doi:10.1016/j.envint.2019.02.013

664

Hellinga, H. I. (2013). Daylight and View: The influence of windows on the visual quality of

665

indoor spaces. Architectural Engineering and Technology. Delft, Netherlands: Delft

666

University of Technology. doi:10.4233/uuid:2daeb534-9572-4c85-bf8f-308f3f6825fd

667
668
669

HKLandsD. (2014). iB1000 Digital Topographic Map. Hong Kong: Lands Department,
Goverment of Hong Kong SAR.
HKPlanD. (2018). Hong Kong Planning Standards and Guidelines. Hong Kong: Planning

670

department, Hong Kong SAR.

671

doi:https://www.pland.gov.hk/pland_en/tech_doc/hkpsg/full/pdf/ch2.pdf

672

HKPlanD. (2019). 3D Photo-realistic Model. Hong Kong: Planning Department, Goverment

673

of Hong Kong SAR. Retrieved from

674

https://www.pland.gov.hk/pland_en/info_serv/3D_models/download.htm

675

Jim, C. Y. & Chen, W. Y. (2009). Value of scenic views: Hedonic assessment of private

676

housing in Hong Kong. Landscape and Urban Planning, 91(4), 226-234.

677

doi:10.1016/j.landurbplan.2009.01.009

678
679

Kaplan, R. (2001). The nature of the view from home: Psychological benefits. Environment
and Behavior, 33(4), 507-542. doi:10.1177/00139160121973115

680

Kaplan, S. (1995). The restorative benefits of nature: Toward an integrative framework.

681

Journal of Environmental Psychology, 15(3), 169-182. doi:10.1016/0272-

682

4944(95)90001-2

683

Kaya, N. & Erkip, F. (2001). Satisfaction in a dormitory building: The effects of floor height

684

on the perception of room size and crowding. Environment and Behavior, 33(1), 35-

685

53. doi:10.1177/00139160121972855

686

Ko, W. H., Kent, M. G., Schiavon, S., Levitt, B. & Betti, G. (2021). A window view quality

687

assessment framework. Leukos, 1-26. doi:10.1080/15502724.2021.1965889

688

Kuo, F. E. & Sullivan, W. C. (2001). Environment and crime in the inner city: Does
27

689

vegetation reduce crime? Environment and Behavior, 33(3), 343-367.

690

doi:10.1177/0013916501333002

691

Labib, S. M., Huck, J. J. & Lindley, S. (2021). Modelling and mapping eye-level greenness

692

visibility exposure using multi-source data at high spatial resolutions. Science of the

693

Total Environment, 755, 143050. doi:10.1016/j.scitotenv.2020.143050

694

Laovisutthichai, V., Li, M., Xue, F., Lu, W., Tam, K. & Yeh, A. G. (2021). CIM-enabled

695

quantitative view assessment in architectural design and space planning. 38th

696

International Symposium on Automation and Robotics in Construction (ISARC 2021).

697

Dubai.

698
699
700

LeCun, Y., Bengio, Y. & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436-444.
doi:10.1038/nature14539
Lee, D.-T. & Schachter, B. J. (1980). Two algorithms for constructing a Delaunay

701

triangulation. International Journal of Computer & Information Sciences, 9(3), 219-

702

242. doi:10.1007/BF00977785

703

Li, M., Xue, F., Yeh, A. G. & Lu, W. (2020). Classification of photo-realistic 3D window

704

views in a high-density city: The case of Hong Kong. 25th International Symposium

705

on Advancement of Construction Management and Real Estate (pp. 1339-1350).

706

Singapore: Springer Nature.

707

Li, W. & Samuelson, H. (2020). A new method for visualizing and evaluating views in

708

architectural design. Developments in the Built Environment, 1, 100005.

709

doi:10.1016/j.dibe.2020.100005

710

Li, X., Zhang, C., Li, W., Ricard, R., Meng, Q. & Zhang, W. (2015). Assessing street-level

711

urban greenery using Google Street View and a modified green view index. Urban

712

Forestry & Urban Greening, 14(3), 675-685. doi:10.1016/j.ufug.2015.06.006

713

Liang, J., Gong, J., Sun, J., Zhou, J., Li, W., Li, Y., Liu, J. & Shen, S. (2017). Automatic sky

714

view factor estimation from street view photographs—A big data approach. Remote

715

Sensing, 9(5), 411. doi:10.3390/rs9050411

716

Liu, Y., Meng, Q., Zhang, J., Zhang, L., Jancso, T. & Vatseva, R. (2016). An effective

717

Building Neighborhood Green Index model for measuring urban green space.

718

International Journal of Digital Earth, 9(4), 387-409.

719

doi:10.1080/17538947.2015.1037870

720

Lottrup, L., Stigsdotter, U. K., Meilby, H. & Claudi, A. G. (2015). The workplace window

721

view: a determinant of office workers’ work ability and job satisfaction. Landscape

722

Research, 40(1), 57-75. doi:10.1080/01426397.2013.829806
28

723

Lu, Y. (2018). The association of urban greenness and walking behavior: Using google street

724

view and deep learning techniques to estimate residents’ exposure to urban greenness.

725

International Journal of Environmental Research and Public Health, 15(8), 1576.

726

doi:10.3390/ijerph15081576

727

Masoudinejad, S. & Hartig, T. (2020). Window view to the sky as a restorative resource for

728

residents in densely populated cities. Environment and Behavior, 52(4), 401-436.

729

doi:10.1177/0013916518807274

730

McFeeters, S. K. (1996). Difference Water Index (NDWI) in the delineation of open water

731

features. International Journal of Remote Sensing, 17(7), 1425-1432.

732

doi:10.1080/01431169608948714

733

Middel, A., Lukasczyk, J., Zakrzewski, S., Arnold, M. & Maciejewski, R. (2019). Urban form

734

and composition of street canyons: A human-centric big data and deep learning

735

approach. Landscape and Urban Planning, 183, 122-132.

736

doi:10.1016/j.landurbplan.2018.12.001

737
738
739

Pan, S. J. & Yang, Q. (2009). A survey on transfer learning. IEEE Transactions on Knowledge
and Data Engineering, 22(10), 1345-1359. doi:10.1109/TKDE.2009.191
Pazhouhanfar, M. & Kamal, M. (2014). Effect of predictors of visual preference as

740

characteristics of urban natural landscapes in increasing perceived restorative

741

potential. Urban Forestry & Urban Greening, 13(1), 145-151.

742

doi:10.1016/j.ufug.2013.08.005

743

Roe, J. J., Thompson, C. W., Aspinall, P. A., Brewer, M. J., Duff, E. I., Miller, D., Mitchell, R.

744

& Clow, A. (2013). Green space and stress: evidence from cortisol measures in

745

deprived urban communities. International Journal of Environmental Research and

746

Public Health, 10(9), 4086-4103. doi:10.3390/ijerph10094086

747

Shen, Q., Zeng, W., Ye, Y., Arisona, S. M., Schubiger, S., Burkhard, R. & Qu, H. (2017).

748

StreetVizor: Visual exploration of human-scale urban forms based on street views.

749

IEEE Transactions on Visualization and Computer Graphics, 24(1), 1004-1013.

750

doi:10.1109/TVCG.2017.2744159

751

Song, Y., Wang, X., Tan, Y., Wu, P., Sutrisna, M., Cheng, J. C. & Hampson, K. (2017). Trends

752

and opportunities of BIM-GIS integration in the architecture, engineering and

753

construction industry: a review from a spatio-temporal statistical perspective. ISPRS

754

International Journal of Geo-Information, 6(12), 397. doi:10.3390/ijgi6120397

755

Tara, A., Lawson, G. & Renata, A. (2021). Measuring magnitude of change by high-rise

756

buildings in visual amenity conflicts in Brisbane. Landscape and Urban Planning,
29

757
758

205, 103930. doi:10.1016/j.landurbplan.2020.103930
Tucker, C. J. (1979). Red and photographic infrared linear combinations for monitoring

759

vegetation. Remote sensing of Environment, 8(2), 127-150. doi:10.1016/0034-

760

4257(79)90013-0

761

Turan, I., Reinhart, C. & Kocher, M. (2019). Evaluating spatially-distributed views in open

762

plan work spaces. Proceedings of the IBPSA International Building Simulation

763

Conference, (pp. 1098-1105). Rome. doi:10.26868/25222708.2019.210755

764

Ulrich, R. S. (1981). Natural versus urban scenes: Some psychophysiological effects.

765

Environment and Behavior, 13(5), 523-556. doi:10.1177/0013916581135001

766

Ulrich, R. S. (1983). Aesthetic and affective response to natural environment. Behavior and

767

the Natural Environment (pp. 85-125). Boston: Springer. doi:10.1007/978-1-4613-

768

3539-9_4

769
770

Ulrich, R. S. (1984). View through a window may influence recovery from surgery. Science,
224(4647), 420-421. doi:10.1126/science.6143402

771

UNPD. (2014). World Urbanization Prospects: The 2014 Revision, Highlights. Population

772

Division, Department of Economic and Social Affairs, ST/ESA/SER.A/352. United

773

Nations. Retrieved from https://population.un.org/wup/Publications/Files/WUP2014-

774

Highlights.pdf

775

Van den Berg, M., Poppel, M. v., Kamp, I. v., Andrusaityte, S., Balseviciene, B., Cirach, M.

776

& Danileviciute, A. e. (2016). Visiting green space is associated with mental health

777

and vitality: A cross-sectional study in four european cities. Health & Place, 38, 8-15.

778

doi:10.1016/j.healthplace.2016.01.003

779

Waczynska, M., Sokol, N. & Martyniuk-Peczek, J. (2021). Computational and experimental

780

evaluation of view out according to European Standard EN17037. Building and

781

Environment, 188, 107414. doi:10.1016/j.buildenv.2020.107414

782

White, M., Smith, A., Humphryes, K., Pahl, S., Snelling, D. & Depledge, M. (2010). Blue

783

space: The importance of water for preference, affect, and restorativeness ratings of

784

natural and built scenes. Journal of Environmental Psychology, 30(4), 482-493.

785

doi:10.1016/j.jenvp.2010.04.004

786

Xia, Y., Yabuki, N. & Fukuda, T. (2021). Development of a system for assessing the quality

787

of urban street-level greenery using street view images and deep learning. Urban

788

Forestry & Urban Greening , 59, 126995. doi:10.1016/j.ufug.2021.126995

789

Xue, F., Li, X., Lu, W., Webster, C. J., Chen, Z. & Lin, L. (2021). Big data-driven pedestrian

790

analytics: unsupervised clustering and relational query based on Tencent Street View
30

791

photographs. ISPRS International Journal of Geo-Information, 10(8), 561.

792

doi:10.3390/ijgi10080561

793

Xue, F., Wu, L. & Lu, W. (2021). Semantic enrichment of building and city information

794

models: A ten-year review. Advanced Engineering Informatics, 47, 101245.

795

doi:10.1016/j.aei.2020.101245

796

Yao, Y., Liang, Z., Yuan, Z., Liu, P., Bie, Y., Zhang, J., Wang, R., Wang, J. & Guan, Q.

797

(2019). A human-machine adversarial scoring framework for urban perception

798

assessment using street-view images. International Journal of Geographical

799

Information Science, 33(12), 2363-2384. doi:10.1080/13658816.2019.1643024

800

Yu, S., Yu, B., Song, W., Wu, B., Zhou, J., Huang, Y., Wu, J., Zhao, F. & Mao, W. (2016).

801

View-based greenery: A three-dimensional assessment of city buildings’ green

802

visibility using Floor Green View Index. Landscape and Urban Planning , 152, 13-

803

26. doi:10.1016/j.landurbplan.2016.04.004

804

31