Access the full text.
Sign up today, get DeepDyve free for 14 days.
The formation of urban districts and the appeal of densely populated areas reflect a spatial equilibrium in which work - ers migrate to locations with greater urban vitality but diminished environmental qualities. However, the pandemic and associated health concerns have accelerated remote and hybrid work modes, altered people’s sense of place and appreciation of urban density, and transformed perceptions of desirable places to live and work. This study presents a systematic method for evaluating the trade-offs between perceived urban environmental qualities and urban ameni- ties by analysing post-pandemic urban residence preferences. By evaluating neighbourhood Street View Imagery (SVI) and urban amenity data, such as park sizes, the study collects subjective opinions from surveys on two working conditions (work-from-office or from-home). On this basis, several Machine Learning (ML) models were trained to predict the preference scores for both work modes. In light of the complexity of work-from-home preferences, the results demonstrate that the method predicts work-from-office scores with greater precision. In the post-pandemic era, the research aims to shed light on the development of a valuable instrument for driving and evaluating urban design strategies based on the potential self-organisation of work-life patterns and social profiles in designated neighbourhoods. Keywords Urban data, Sustainable communities, Post-pandemic, Deep learning, Convolutional neural network, Trade-off 1 Introduction The purpose of contemporary cities is to provide peo - ple with convenient and healthy neighbourhood settings while promoting efficiency, resilience, and sustainability. However, the high densities of urban settlements with *Correspondence: extensive commercial and industrial footprints have Dan Luo increased susceptibility to epidemic outbreaks and health firstname.lastname@example.org Arcadis IBI Group, Toronto, Canada concerns (Wang, 2021; X. Zhang et al., 2022). This has School of Geographic Information Science, University of Queensland, St. led to a rethinking of urban planning in the post-pan- Lucia, Australia 3 demic period (Bissell, 2021; Mehta, 2022a). In response Department of City and Regional Planning, Cornell University, Ithaca, USA to the pandemic, regulations encourage physical separa- Center for Spatial Information Science, The University of Tokyo, Tokyo, tion, and numerous industries now permit employees to Japan 5 work remotely permanently or in a hybrid capacity (M. School of Architecture, Tsinghua University, Beijing, China School of Architecture, The Chinese University of Hong Kong, Hong Hu et al., 2021). When work patterns shift to remote Kong SAR, China work, urban mobility, commuting patterns, neighbour- School of Architecture, University of Queensland, St.Lucia, Australia hood economies, housing prices, and street vitality are © The Author(s) 2023. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. Song et al. Architectural Intelligence (2023) 2:1 Page 2 of 16 significantly affected (Florida et al., 2021; M. Hu et al., 1. Urban design characteristics have effects at multi - 2021; Mehta, 2022a). ple scales, and both the built and perceived environ- People decide where to live based on a number of fac- ments are critical. When determining housing pref- tors, including accessibility to urban amenities (Rode erences, urban planning macro-level factors such as et al., 2017), access to public transportation (L. Yang the density of places of interest (POI) and micro-level et al., 2022), neighbourhood environment qualities (Y. factors such as perceptions of the neighbourhood Yang & Xiang, 2021), safety perceptions at the neigh- should be taken into account. bourhood scale (Qiu et al., 2022; Song, Li, Li, et al., 2022; 2. It is difficult to incorporate the subjective opinions X. Xu et al., 2022), affordability of housing (Kang et al., of citizens regarding their pre- and post-pandemic 2021a, 2021b), and education quality of nearby schools residence preferences into an effective evaluation (Wen et al., 2017). Consequently, in the post-pandemic model that is also explicable. It is currently unknown context, given the changes in work mode, a significant whether and how the pandemic and related work- mismatch has developed between the city’s existing spa- from-home scenario will affect housing preferences tial form and function and the residents’ life and work relative to the work-from-office condition. dynamics. When balancing desirable urban spatial quali- 3. To apply a model trained with existing variables ties, the value of amenities, and employment opportu- collected from the current context to future scenar- nities, it has subtly but inexorably altered the housing ios, additional experiments are required. preferences of individuals. 4. People’s actual actions and behaviours may differ Given this changing environment, the working popula- from their stated preferences, and people’s spontane- tion is likely to face two distinct scenarios. On the one ous activities may differ from their reported prefer - hand, people who primarily work from home may con- ences. tinue to reside in areas with shorter commute distances and travel times to offices, where this premium for con - Based on collected preference opinions from online venience will no longer be relevant. However, they must questionnaires using Hong Kong as a case study site, still pay exorbitant rent, which is not desirable. Alterna- this research aims to bridge the aforementioned knowl- tively, because they only need to be physically present edge gaps and supplement existing literature by integrat- in the office once or twice per week, these employees ing big data and the cutting-edge Computer Vision (CV) may choose to commute longer distances from farther and Machine Learning (ML) models to provide an oper- away (Florida et al., 2021). For these individuals, housing able workflow for the topic. First, learnable weights are prices or rent in the suburbs are less expensive than in applied to urban characteristics from planning data at city centres, albeit with fewer amenities. This preference the macro-level and Street View Imagery (SVI) reflect - shift may be accelerated by the reduced demand for office ing micro-level conditions in order to better compre- space, resulting in vacancies and scepticism regarding hend the trade-offs in housing preference decisions. the future desirability of office hubs, particularly in the The second step is to collect intuitive subjective rat - expensive downtown areas of cities (Mehta, 2022a). ings from users using visual pairwise selection surveys In the context of ensuring public health and well-being based on both macro- and micro-level qualities of the in response to urban disasters such as epidemics, the spa- neighbourhood to generate evaluation scores for each tial distribution of urban resources exhibits both scar- residence location in work-from-office and work-from- city and abundance patterns (Wang, 2021). Therefore, home conditions. Thirdly, we applied various ML and urban design and construction at the neighbourhood DL algorithms to investigate the effects of different built scale will need to reconsider changing needs and func- environment characteristics on residence preferences in tions, such as if a paradigm shift toward cultural and civic both situations and to evaluate the performance of vari- gathering activities necessitates more street spaces allo- ous models. Lastly, we validated the accuracy of stated cated for local neighbourhood uses (Florida et al., 2021; preferences from surveys using massive amounts of data Mehta, 2022a; Wang, 2021). To comprehend these poten- extracted from unplanned events such as real estate rent tial future policy changes and paradigm shifts, however, prices. This method can assess the accuracy with which there is a lack of understanding of people’s subjective subjective surveys reflect the preferences revealed and housing selection preferences in relation to the complex further validate the results. In conclusion, the research and interwoven factors of urban context across multiple aims to contribute to the general understanding of how geospatial scales. The construction of models to predict people’s perceptions of housing preferences may change neighbourhood development in the era of the "new nor- in response to different work modes and socioeconomic mal" faces a number of significant challenges due to the contexts when amenity and neighbourhood environment following gaps: trade-offs are considered. As preliminary research on this S ong et al. Architectural Intelligence (2023) 2:1 Page 3 of 16 topic, this study aims to provide meaningful inferences Regarding public parks, it is widely recognised that on evidence-based urban design strategies or guidelines they play an important role in the urban environment for creating a more equitable and resilient future through and provide numerous benefits for well-being, including urban transformation by gaining an understanding of recreational value and positive effects on physical and subjective housing preferences. It has great future poten- mental health (La Rosa, 2014). Recognized as an essen- tial and applicability to predict neighbourhoods that are tial criterion for assessing the livability of a location, ade- preferred in different work modes and to prioritise future quate access to green spaces can significantly influence redevelopment projects to meet the demand for poly- people’s residential decisions. Scholars discovered con- centric neighbourhoods. tradictory effects of green open spaces. On the one hand, in compact cities, proximity to urban parks is highly desirable (Netusil, 2013). However, in a North American city with much more sprawled urban fabric, no signifi - 2 Literature review cant correlation can be found using Salt Lake City as case 2.1 R esidence preferences and utilities study site (H. Li et al., 2016). Choosing a neighbourhood to live in is a complex process that requires a balance of various factors, such as prox- 2.2 Spatial environmental qualities imity to essential amenities (Ferreira & Wong, 2020). In According to numerous scholars, the success of the street a competitive real estate market, housing prices reflect environment is crucial to the development of a thriving the revealed preferences of people’s residential choices. urban centre (Jacobs, 1961; Y. Li et al., 2022; Montgom- Using housing price as an indicator of residential prefer- ery, 1998). As a key category of public open space, the ence, there is a large body of literature employing various street not only serves the utilitarian function of trans- models such as the hedonic pricing model (Rosen, 1974). portation, but has also become more adaptable in terms This model consists primarily of three attribute groups of its programs, providing places for exchange and social that are reported to influence the price, namely struc - interaction as well as venues for communities to reclaim tural, location, and neighbourhood attributes (Huang for informal and formal activities, especially during lock- et al., 2017). In addition to the structural attributes that downs (Mehta, 2020). Important factors that can signifi - capture the characteristics of the house’s living condition, cantly influence people’s physical activities (Lu, 2019), such as unit size and number of rooms, the location and pedestrian route choice preferences (Sevtsuk et al., 2021), neighbourhood attributes reflect the house’s spatial loca - and the willingness to pay for housing or rent prices are tions within the city and other functions that residents neighbourhood street environment qualities (Qiu et al., find appealing (H. Li et al., 2019). Among these factors, 2022; J. Yang et al., 2021). public transportation facilities, such as subway stations Previous research primarily relied on two metrics to and bus routes, facilitate access to employment and ser- assess the quality of built environment streets. Schol- vices. It is important to note that there is also a trade-off ars have adopted objective measures employing remote effect associated with housing preferences near public sensing data and used GIS programs to spatially reflect transportation, which is related to both convenience and the quality (Lee & Moudon, 2006). Typical objective noise and traffic intensity. For instance, research indi - characteristics of a dataset include building height, den- cates that highway construction is less desired than bus sity, ratio of floor area to total area, park area sizes, num - and rail transit facilities (Agostini & Palmucci, 2008). ber of specific facilities, access to public transportation, In addition, the access to service amenities has been etc. (L. Yang et al., 2022). However, these macro-level studied in the past (Rode et al., 2017). These amenities planning-oriented variables may be less effective at cap - include a variety of public service amenities, such as turing the daily human experience of site users in the schools and hospitals, as well as private service facili- micro-level street environment (Dubey et al., 2016). In ties, such as shopping malls and restaurants, as well as contrast to objective factors, subjective measurement cultural and recreational amenities, such as museums has been widely used in urban studies in the past, with and sports facilities (Glaeser et al., 2018). Younger gen- interviews or surveys reflecting people’s explicit opinions erations are attracted to the dense mating and friendship on the qualities of the spaces (Ewing et al., 2006; Lynch, markets that these amenities can provide, which can help 1960). However, this technique is limited to a small sam- to retain population in these central urban areas (Florida ple size and difficult to apply to a larger geographical area et al., 2021). In addition to proximity to schools, the qual- (H. Zhou et al., 2019). Consequently, with the availabil- ity of the school education is related to the residential ity of exploding urban big data such as SVI, recent urban location for families with children, which is an impor- studies have developed methods that use CV and ML tant factor that may influence people’s preferences (Wen algorithms to improve the efficacy of urban perceptions et al., 2017). Song et al. Architectural Intelligence (2023) 2:1 Page 4 of 16 mapping on a scale never before seen (Gong et al., 2018; Utilizing SVI, prior research has focused on the micro- Qiu et al., 2023). level characteristics and their effects on housing prefer - On the other hand, using CV semantic segmentation ences using housing prices as a proxy (Kang et al., 2021a, algorithms, researchers can extract the pixels of vari- 2021b; Kang et al., 2021a, 2021b; Qiu et al., 2022; J. Yang ous street features from SVI as indices for objectively et al., 2021). Scholars have determined, for instance, that constructing perception scores (Zhao et al., 2017). For street-level greenery contributes to higher prices; in example, X. Li et al. (2015) created the revised green other words, people prefer neighbourhoods with greater view index as a surrogate for the visual greenery, and the exposure to vegetation (Fu et al., 2019). And through building and sky view factors were measured in a simi- a place-based hedonic price model, Kang et al., (2021a, lar manner (Gong et al., 2018). Recently, researchers have 2021b) determined that people prefer to reside in a attempted to quantify urban design perceptions (e.g., neighbourhood with higher perceived safety and attrac- greenness, openness, enclosure, walkability, and image- tiveness. However, existing research focuses primarily ability) utilising complex formulas derived from design on the pre-COVID era (L. Yang et al., 2022). It does not theory and recombining the extracted view indices of key interpret how people’s subjective residential preferences urban street features (Ma et al., 2021). may have changed or which characteristics may influence Using globally collected perception surveys such as residents’ choices in the post-pandemic context (Florida Place Pulse Dataset 1.0 and 2.0, however, research- et al., 2021). ers could map psychological perceptual qualities such as safety, beauty, and boring (Dubey et al., 2016; Naik 3 Data and methodology et al., 2014; Salesses et al., 2013). F. Zhang et al. (2018) In order to examine the effect of shifts in work patterns predicted six perceptions in Beijing and Shanghai using on housing preferences, our research focuses on the extracted street characteristics as independent variables working population of design industries whose company and perception survey results as training labels. Through offices are located in the city centre. Hong Kong is used the high-level extracted features, the study also revealed as a case study example in response to contemporary which streetscapes may result in varying perceptual qual- issues. ities. The Place Pulse datasets were, however, collected We obtained the city’s SVIs and the opinions of its resi- globally, with limited results for China’s streetscapes. To dents through subjective rating of a randomly sampled avoid potential bias, studies focusing on Chinese cities SVI database. In this study, the innovative and crucial must rely on opinions gathered locally. Qiu et al. (2022), aspect of the workflow is that both macro-level planning- for instance, gathered expert opinions as training labels related objective spatial variables, such as points of inter- from a panel of experts. They subjectively measured five est, park size, commuting time to the city centre, and urban design-related perceptual qualities of Shanghai availability of public transportation, and micro-level per- neighbourhoods using CV and ML models. They discov - ceived street quality are integrated and presented to the ered that overall objective features outperform subjective participants in order to consider all related factors under perceptions in predicting housing prices, whereas sub- two working conditions comprehensively. The rating is jective perceptions have a greater impact on individual based on data that has been altered to include revised objective features. Moreover, Song, Li, Li, et al. (2022) neighbourhood conditions that simulate future develop- demonstrated that objective and subjective measurement ment scenarios that are not representative of currently of perceptions may exhibit contradictory spatial hetero- selected neighbourhoods. The variables from different geneity patterns even when mapping the same concept, scales reflect the complexity that real-world residents indicating the need for future research to investigate the must consider when selecting a residence. Using a CV discrepancy and consistency of these two methods. algorithm, the perceived street features are extracted People remain cautious about social distancing in from SVI data to reflect the physical neighbourhood the post-pandemic era as Covid-19 waves return every environment. Various ML algorithms are utilised to few months, and scholars have identified a new set of evaluate the accuracy of the trained model of the work- dimensions for social activities in the public sphere due from-home and work-in-office scores datasets. Last but to hygiene concerns (Mehta, 2020). Changes in spa- not least, real estate rental data, which reveals human tial distances between people, as well as shifts in work behaviour and housing preferences, is used to validate modes, have an impact on living preferences, travel, and its correlations with stated preferences, i.e., subjectively commuting patterns. Taking into account the trade- evaluated scores (Fig. 1). This result will indicate how off between accessibility to urban amenities and other the two types of working scenarios may affect the hous - assets, residents’ preferences and perceptions of their ing preferences of residents, which will have significant neighbourhood environment may be subject to change. implications for post-pandemic urban design strategies. S ong et al. Architectural Intelligence (2023) 2:1 Page 5 of 16 Fig. 1 Method and workflow It aims to provide a foundation for future research to pre- government has released the 2030 + strategic plan, which dict human behaviour in response to changes in the sense aims to improve the liveability in the compact, high-den- of place and preferences for future development. sity city through a series of policies such as the provision of additional living units to increase capacity for future sustainable growth (Hong Kong Development Bureau 3.1 Study area & Hong Kong Planning Bureau, 2021). Residents can The People’s Republic of China’s Hong Kong Special travel within one to two hours and live within walking Administrative Region is a semi-autonomous adminis- distance of accessible open space in Hong Kong, which trative territory located in south of Guangdong Province, continues to have one of the highest public transporta- connecting to Shenzhen. Hong Kong Island, Kowloon, tion usage rates among major developed cities. It enables and the New Territories comprise its three geographi- professionals who live far away from their workplaces to cal regions, and since 1999, the territory has been sub- commute via public transportation. However, as a result divided into 18 administrative districts. Approximately of the global pandemic, many companies have imple- 7.5 million people inhabit the 1100 square kilometre mented the WFH policy for the first time in an effort to land area, which is comprised of the historic urban core prevent the spread of infection, and the government has and a number of satellite towns. Due to its complex his- imposed social distancing mandates out of concern for torical, political, and topographical conditions, the built public health (Wut et al., 2022). This disruption of work - land comprises only about one-fifth of the total area, and ing-living spatial dynamics and its consequential effect residents’ living conditions are marked by high costs, on the reconsideration of the trade-off between utilities compact apartments, and limited public open space and neighbourhood environment qualities has received (Kwok et al., 2021; Liu et al., 2020). Additionally, previ- less attention. Understanding people’s subjective hous- ous research has demonstrated that there are serious ing preferences in this new context can inspire future social inequality and segregation issues (Monkkonen & urban redevelopment projects and policy initiatives Zhang, 2011). In response to these alarming social chal- lenges and in an effort to attract talent, the Hong Kong Song et al. Architectural Intelligence (2023) 2:1 Page 6 of 16 that contribute to a more sustainable and resilient met- that residents will experience daily in the neighbour- ropolitan region, as envisioned by the Hong Kong hood. And because our research site is Hong Kong, 2030 + vision. unlike previous studies conducted in northern climates such as Boston, where deciduous trees are the predomi- nant street trees, the seasonality of street greenery is not 3.2 Da ta collection and processing a concern for our study (X. Li et al., 2015). Consequently, We obtained information about Hong Kong’ s road net- removal of winter images is not necessary. We manually work from Open Street Map, which provides free vector replaced these invalid points with nearby points and then data. We sampled points along the road geometries at recalculated the SVI. 50-m intervals using QGIS to create shapefiles (Qiu et al., Significant criteria to measure street spatial quality that 2022; H. Zhou et al., 2019). This resulted in the crea - affect housing preferences are macro-level objective indi - tion of 1,000 points in each district, and we accumulated cators (Song, Liu, et al., 2022a, 2022b, 2022c). In addition 18,000 points throughout the city. To ensure that our to the SVIs, we selected the ’number of POIs’, ’number training images cover the majority of urban area types of bus stops’, ’commute time to the city’, and ’size of park and enhance the efficiency of the study, we randomly area’ as the key variables to represent the neighbourhood selected 578 points from the dataset (Fig.2). The sample qualities. Hong Kong Geodata Store (https:// geoda ta. size was determined by a compromise between the accu- gov. hk/ gs/) is used to retrieve the spatial distribution of racy of ML prediction, the reliability of the survey, and bus stations, POIs, and parks. And rent price informa- the workload of the participants. According to Beleites tion is crawled from the website of a real estate company, et al. (2013), to obtain accurate ML prediction results, namely Centaline (https:// hk. centa net. com/ info/index). at least 100 samples or ten times the required number of QGIS was used to process the macro-level datasets. independent variables must be collected. The traditional Around the centre of each selected point, a buffer zone ML models used in our study extract street features using with a 1.5 km radius was created. We obtained the num- semantic segmentation; the CV algorithm extracted ber of bus stops and points of interest, as well as park approximately 30 features from the scenes; therefore, area sizes, within each point’s buffer. The commute time our sample size must be at least 300 points; 578 points is was then determined based on the distance between nearly double the minimum requirement and is deemed the site and the city centre. These indicators are criti - adequate. Following previous urban perception mapping cal for enhancing the residents’ livability and have been research (Qiu et al., 2023), we initially randomly sampled reported to influence housing prices (Huang et al., 2017; a similar number of points within each administrative H. Li et al., 2019). The number of bus stops can be viewed district (approximately 32 points per district); however, as a proxy for the degree of mobility infrastructure, due to the unique geography and settlement layout, the which provides crucial access to a residence and place urban built land accounts for less than one-fifth of the of employment (Rode et al., 2017). The number of POIs total area, indicating that neighbourhoods can be con- reveals the urban spatial structure and provides data- centrated across several districts. Therefore, based on driven guidance for regional spatial regulation and opti- the spatial layout, we manually balanced the number of mization. POI data documents diverse types of economic sample points within the municipal districts to ensure and social sectors and can be used to identify the grid that they could geographically represent the street envi- units’ functional zones and inform the human-scale vital- ronment of Hong Kong’s urban and peri-urban neigh- ity (Y. Hu & Han, 2019; Yue et al., 2017). The park sizes bourhoods (Monkkonen & Zhang, 2011). We fed the represent the overall environmental quality and contrib- geographic coordinates of these points into the Google ute to subjective well-being and quality of life (M. Xu Street View Static API and retrieved the platform’s cor- et al., 2017). Lastly, commute time highlights the reflec - responding SVIs. Each scene is 800 by 400 pixels with a tion of travel behaviour of work-living experience as daily field of view (FOV) of 90 degrees and a pitch angle of 0. life; it is a crucial factor influencing residence experience Despite the fact that some previous studies have utilised on the long run and also becomes a safety concern in the panorama views (Lu, 2019), these are susceptible to dis- post-COVID era (Florida et al., 2021). tortion, making it difficult for individuals to express their In order for the parameters of each variable to be more preferences. In order to maintain spatial and temporal evenly distributed across the design parameter space, continuity for subsequent visual survey comparison and noise and variance were added to the existing data for a machine learning (ML) process and to minimise poten- number of POIs and park area size (Fig. 3). This enables tial biases, the 578 retrieved SVIs were cleaned to remove the data to include future possibilities that are currently blurry, grey (blank), and highway and indoor images (Ito absent. & Biljecki, 2021). This can help reflect the environment S ong et al. Architectural Intelligence (2023) 2:1 Page 7 of 16 Fig. 2 Five hundred seventy-eight sampled study points in Hong Kong 3.3 P erceived neighbourhood characteristics spatial qualities and utility access, but in different work at micro‑level scenarios. We created an online visual questionnaire platform that Previous research suggests that pairwise comparisons solicits respondents’ input on their preferences for liv- improve the accuracy of personal preferences (Salesses ing in two conditions (Fig. 4). On each page of the sur- et al., 2013). Using a dynamic comparison survey as vey, participants are asked to compare pairwise SVIs opposed to a static score survey reduces personal bias, and select their preferred scene by balancing the trade- such as the tendency of some individuals to assign higher off between perceived street environment quality based scores. The visual survey includes all 578 SVIs collected on a human vision of SVIs and macro-level urban spa- from Hong Kong. This sample size outperforms previous tial attributes such as commute time to the city centre. research, which also used a smaller training dataset; for This illustrates the complicated logic behind how people example, Qiu et al. (2022) used 300 SVIs to predict urban choose their homes. On the basis of their working mode design perceptions, and Ito and Biljecki (2021) collected (work-from-home or in-office), respondents will assume 400 SVIs. The respondents were architectural students that the office is located in the city centre and make a and young professionals between the ages of 20 and 35, decision regarding their preferred neighbourhood (by and 3,718 valid responses were collected. Using Place clicking the preferred scene). Each page of the visual Pulse Dataset 1.0, Salesses et al. (2013) demonstrated that survey is divided into two parts; the top half of each people’s preferences for street environment perceived page asks participants to evaluate the work-from-home qualities are not influenced by the demographic informa condition, while the bottom half asks about the work- tion of survey participants, such as age and gender, nor from-office condition. This means that each location by the geographical locations of the respondents. Con- will be evaluated under both the work-from-home and sequently, these responses can be generalised for future in-office criteria. It can help us compare how individu - research focusing on various geographical locations. als evaluate living preferences when evaluating the same Nevertheless, future research can consider incorporating Song et al. Architectural Intelligence (2023) 2:1 Page 8 of 16 Fig. 3 Study points scatter plot (original data VS modified data) Fig. 4 Questionnaire/Survey for Revealing Subjective Preference S ong et al. Architectural Intelligence (2023) 2:1 Page 9 of 16 responses from other professions, which may provide further establish a connection with ML models for pre- varying answers due to the nature of work in various diction (Qiu et al., 2022; Song, Li, Li, et al., 2022). We fields. The most frequently compared SVI in the working- used the pre-trained Pyramid Scene Parsing Network from-home scenario achieved 31 comparisons, while the (PSPNet) algorithm (Fig. 5) with the ADE20K dataset to highest SVI in the working-from-office scenario achieved efficiently segment the physical features from the SVI 27 comparisons. Each image was compared an average of scenes and compute the pixel ratio of each type of street 12.86 times, which exceeds the threshold of 12 compari- element (Zhao et al., 2017; B. Zhou et al., 2018). The CV sons suggested by Herbrich et al. (2006) for obtaining a algorithm achieved a pixel-level precision of 93.4% and reliable score estimate in the next step. It is close to the has been extensively utilised in previous urban studies 16 times average comparisons in Salesses et al.(2013)’s (Qiu et al., 2023; Song, Li, Qiu, et al., 2022). More than study utilising Place Pulse Dataset 1.0 and significantly 30 types of street features were extracted from the SVIs higher than Dubey et al. (2016)’s study utilising Place following the procedure. We chose eight major elements Pulse Dataset 2.0, which only achieved 3.4 times per (i.e., building, sky, road, signboard, wall, tree, sidewalk, image. Consequently, this dataset is deemed acceptable. and fence) for the next step of ML model training and The Microsoft True Skill algorithm, a Bayesian ranking prediction purposes, after removing ephemeral objects method, was utilised to faithfully convert survey prefer- such as automobiles and those with little significance in ences into scores based on a two-player game mechanism the scenes. (Minka et al., 2018). 3.5 ML, DL Model training and validation 3.4 Ex tracting micro‑level perceived neighbourhood Following Qiu et al. (2021)’s approach to using traditional characteristics ML algorithms for prediction, the dependent variables of SVI physical characteristics can serve as proxies for per- this study’s intended output are two types of preference ceived street characteristics (Ewing & Handy, 2009). scores (work-from-home and work-from-office). Inde - Prior research has quantified their numbers in order to pendent variables consist of the extracted pixel ratios of Fig. 5 SVI inputs and PSPNet semantic segmentation outputs Song et al. Architectural Intelligence (2023) 2:1 Page 10 of 16 selected eight features and objective spatial qualities. The issues. Convolutional layers culminate in layers with nine ML models utilised in this study are Linear Regres- complete connectivity. In order to use both planning spa- sion and Random Forest (RF), etc. tial quality variables and CNN output data readings from To select the optimal ML model, we evaluated the per- SVI as independent variables, the model concentrates formance of various ML algorithms using the Root Mean them and constructs densely interconnected layers with Square Error (RMSE) as the loss function, which has the final output of a score (Fig. 6). Both scores have sig- advantages in punishing significant errors. The lower the nificantly improved compared to the results of conven - RMSE value, the more accurate a prediction model. In tional ML models (Table 2). The RMSE for the Q1_Home addition, we calculated the R-Squared (R2) value, which score was 0.0195, and R2 was 0.34. Although the R2 of effectively explains the model fit by indicating how much the Work-from-Home scenario does not demonstrate of the variance of a dependent variable can be explained a significant correlation, CNN achieved much greater by independent variables in a regression model. The accuracy for the Q2_Office scenario, which had an RMSE model results are shown in Table 1. of 0.0045 and an R2 of 0.779. The R2 value of Q2_Office The best traditional machine learning method for outperforms previous studies that used SVI to predict predicting the Q1_home score is DT, which achieved subjective perception scores, whose R2 values range an RMSE of 0.18 and an R2 of 0.14. Four ML models from 0.47 to 0.61 (Qiu et al., 2022; Song, Li, Qiu, et al., had identical performance for the Q2 Office score, with 2022). Due to the ratio between the training and testing RMSE of 0.19 and R2 of 0.23; the R2 of these models datasets, the model may be susceptible to an overfitting explained more variances than the DT for Q1 scores. issue. The results indicate that for the Work-from-Office However, their results are regarded as having a low score, CNN has proven to be applicable for predicting degree of predictive accuracy because their criteria fall preference scores based on street scenes and neighbour- below typical threshold values (Qiu et al., 2022). hood spatial quality factors. Traditional machine learning methods rely heavily on training features at the highest level. Although applying 3.6 Correlation results between stated preferences and revealed preferences The stated preferences of work-from-home and work- from-office scores are compared to the revealed prefer - Table 1 Performance of traditional ML algorithms ences, which are the rent prices collected from a Hong Model Q1_Home Q2_Office Kong real estate website. The results (Fig. 7) demon- strated that the work-from-home score (Q1_home) R2 RMSE R2 RMSE and the work-from-office score (Q2_office) are highly a a Random Forest (RF) 0.11 0.18 0.23 0.19 positively correlated (coefficient = 0.51). Although the a a Decision Tree (DT ) 0.14 0.18 0.21 0.19 correlation between Q1, Q2 and rent price is less than a a Voting Selection (VS) 0.08 0.18 0.23 0.19 a a Gaussian (GS) 0.08 0.18 0.23 0.19 ADA Boost (ADAB) 0.04 0.18 0.22 0.19 Table 2 Performance of CNN Model K‑Nearest Neighbours (KNN) -0.11 0.2 0.06 0.21 Support Vector Machine (SVM) 0.06 0.18 0.14 0.2 Q1_Home Q2_Office a a Ordinary Least Square (OLS) 0.08 0.18 0.23 0.19 CNN Model R2 RMSE R2 RMSE Bagging Regression (BR) 0.07 0.18 0.18 0.19 Performance 0.34 0.0195 0.779 0.0045 a represents the best-performing model semantic segmentation to pre-process street scenes pro- 0.2, which indicates a weak correlation, the coefficient vides information on quantities of each type of feature between work from office score and rent is approximately and makes it interpretable (Song, Li, Qiu, et al., 2022), 64% higher than the coefficient between work from home it is less efficient to translate the overall ambience, style, score and rent price. and textures as well as the details of the streetscape; the information tends to be lost after extraction. Therefore, this study has advanced towards Deep Learning (DL) 4 Discussion by employing a custom Convolutional Neural Network 4.1 Interpretations of CNN Model (CNN) to train the dataset (F. Zhang et al., 2018). In general, the CNN model is more accurate than tradi- By repeatedly stacking a 3 × 3 convolution kernel and tional ML models at predicting both types of scores based a 2 × 2 max pooling layer three times, the dropout func- on the neighbourhood environment and urban amenities tion is employed for regularisation to prevent overfitting S ong et al. Architectural Intelligence (2023) 2:1 Page 11 of 16 Fig. 6 Customised CNN model architecture Fig. 7 Pearson correlation between scores and rent prices trade-off, and it captures the nuances of residents’ opin - children, work-from-home parents may take into account ions considerably better. It is reported, however, that the the educational quality of nearby schools, the availability model work-from-office score has a greater R2, which of day care facilities, the flexibility of their own jobs, and explains more score variance, and a lower RMSE, which other factors. Secondly, due to the disruption of mobil- indicates less error in predicting scores. This indicates ity patterns and the shift in work modes, depending on that our CNN model explains the work-from-office sce - job types, certain factors may have a greater influence on nario better than the work-from-home circumstance. choices. Access to a variety of amenities no longer con- Our validation using rent price as the revealed preference fers a premium on convenience for people who work demonstrates further that the work-from-home score remotely. Due to health concerns, previous research in correlates less with rent. This result demonstrates con - China revealed that the implicit price of metro access clusively that people’s housing preferences have changed around Chengdu decreased for a time. This indicates significantly in relation to their work mode. that the value of convenient access to public transporta- Multiple hypotheses can explain these results. Firstly, tion amenities is decreasing due to shift in work modes. housing preferences are jointly influenced by a large num - Nevertheless, based on the prediction model, the price is ber of factors; the spatial qualities captured by our model likely to rebound in the long term (L. Yang et al., 2022), could be expanded, particularly for the work-from-home and this requires further research. Thirdly, commuting scenario. Also, people may have different needs at vari - time may be the most important factor in deciding where ous stages of their lives; for instance, young adults may to live for people who need to commute to the office, as want to work in the city, enjoy more urban amenities, and those who use public transportation more frequently are socialise more frequently. While forming a family with more susceptible to epidemic diseases (X. Zhang et al., Song et al. Architectural Intelligence (2023) 2:1 Page 12 of 16 2022). In the current post-pandemic context, commuting people may desire to live in urban centres and enjoy the time may be the dominant factor that trumps all other convenience of accessing amenities without the need for indicators for reducing health-related risks. Depending public transportation. This may indicate a more polycen - on the office location, people may choose to continue liv - tric pattern of urbanisation that links diverse types of ing in urban areas if their homes do not provide sufficient neighbourhoods or communities. space for a functioning office (Wut et al., 2022). Fourthly, In Hong Kong, the WFH policy, despite being an indis- regarding access to amenities, a poll conducted during pensable component in addressing health concerns, is the first pandemic wave revealed that 40% of city dwellers not devoid of pitfalls. The lack of resources makes it considered moving to the suburbs to enjoy more private difficult for certain sociodemographic groups of Hong amenities, and they may never return (Hart, 2020). How Kong residents to adopt this policy on a long-term basis, people view this in the long run is unclear and requires in contrast to cities in North America, where residents further research. Fifthly, the cessation of mobility and typically live in much larger homes that allow for the transition to a hybrid work mode facilitates a shift in the allocation of space for work. Individuals who are able to sense of place and how people perceive their neighbour- achieve greater WFH effectiveness are more likely to pre - hood environment (Bissell, 2021). During the lockdown, fer this hybrid mode and more willing to continue after people must rely more on their neighbours, and after the the pandemic, according to studies (Wong et al., 2020). pandemic, these interactions have fostered a previously Finally, scholars argued that these types of neighbour- absent sense of social connectedness (Ottoni et al., 2022). hood choices are ideal in the real world because residents In addition, physical separation has spawned new prox- may have limited experience with residential choices and emics (Mehta, 2020), which foster new dimensions for rely on a limited social network for information. Such social space. In conclusion, urban parks are increasingly individual imperfection results in a certain bias in the recognised for their unique benefits to mental and physi - estimation of neighbourhood preferences; therefore, this cal health. Researchers discovered that greater access to imperfection factor must also be accounted for in the parks in neighbourhoods with higher population densi- housing preference prediction model (Ferreira & Wong, ties can encourage young people to remain physically 2020). active through exercise (Mitra et al., 2020). People may prefer communities with more open space, wider side- 4.3 Implications on policies walks, and ample bike lanes. Nonetheless, it is impor- The research findings on the shift in housing preferences tant to note that these dynamic changes are more likely have the potential to be applied in other regions of the to create new opportunities in neighbourhoods with a world, especially in comparable Asian metropolitan cit- medium-to-low population density, whereas this may ies with a high population density, such as Tokyo and not be the case in dense neighbourhoods. These obser - Singapore. This may also be appropriate for compact vations raise concerns about social inequality, which may and transit-oriented development. However, if the urban be exacerbated by the pandemic (Florida et al., 2021; Y. morphologies are located in a low-density context with Yang & Xiang, 2021). The Hong Kong 2030 + Vision Plan a starkly different urban–rural spatial pattern that is pri - has envisioned adding more green space; where and how marily car-oriented, it may have minimal effects due to to implement them through projects, taking into account the lifestyle differences. For instance, scholars have sug - the various work modes, are worthy of future research. gested that people in these neighbourhoods would rely on different means to access open space, such as pri - vate automobiles, and that a different strategy would be 4.2 F uture urban spatial patterns required to expand the public space around these neigh- FLORIDA et al. (2021) have predicted the effects of the bourhoods compared to urban centres (H. Li et al., 2016; global pandemic and the potential future of the world Mehta, 2022b). Furthermore, scholars hypothesised that post-pandemic. They predict that although cities will occupations with a higher proportion of knowledge- rebound in the long run and on a large scale, there may based work, such as some tech-related industry jobs, are be temporary or permanent morphological changes on likely to continue this work-from-home trend (Florida a smaller scale. Taking into account the urban amenity et al., 2021). Consequently, unlike Hong Kong, this shift trade-off and the neighbourhood environment, various in housing preferences may have already occurred in cer- groups of people may have contradictory housing prefer- tain cities, such as San Francisco, which are home to the ences. Some people may choose to live in suburban areas, major campuses of these technology companies. and we anticipate that their ’urban’ characteristics will These observations and findings could inform increase and their functional mix will evolve. Due to the changes in guidelines to reimagine urban morphology, reasons stated in the previous section, other groups of open spaces, and public gathering places, ultimately S ong et al. Architectural Intelligence (2023) 2:1 Page 13 of 16 contributing to the objective of creating a healthy city. the classification of case study sites can be based on According to research, elements of the built environ- macroscopic a priori data such as scale and accessibility ment, such as urban block morphologies, road networks, to amenities as evaluation indicators, and then random building configurations, and socioeconomic factors, sampling and subjective evaluation of similar spaces in influence COVID incidence and mortality rates (M. Hu different types of spaces. And to complement exist - et al., 2021; Kwok et al., 2021). Comparing the Compact ing quantitative frameworks, a mixed research method City principles and pedestrian-oriented development can provide additional value to reveal the dynamic cor- to other forms of sustainable urbanism with a moder- relation between spatial form and behavioural space. ate population density, the pandemic has raised many Third, the use of SVI images can be expanded by, for questions about which urban planning model ought to instance, providing participants with a series of differ - be promoted. There is no conclusive evidence linking ent neighbourhood images to compare and gain a bet- population density and the pandemic, as studies have ter intuitive understanding of perceived neighbourhood yielded conflicting results. According to studies, the qualities. Fourth, the survey can collect the opinions of decline of COVID is more closely related to mobility pat- more respondents, resulting in more accurate subjective terns than to building density, as people who can access scores. Fifth, ML models can be improved, as PSPNet’ s neighbourhood retail stores without having to travel far parsing of SVI images and use of ML results in inaccurate are safer (Y. Yang & Xiang, 2021). These findings further predictions. In our case, however, using CNN and real- substantiate the advantages of well-planned compact istic urban scenes reduces the error rate. In addition to developments in the direction of a self-sustaining model spatial objective quality variables, the prediction model (X. Zhang et al., 2022). On the other hand, they suggest may also incorporate high- and low-level features and the need for potential changes to land-use planning. The street images in order to predict preference scores (Song, pandemic demonstrates the fragility of retail main streets Li, Qiu, et al., 2022). Previous research has demonstrated and emphasises the importance of neighbourhood shops that subjective perception is superior to objective charac- for fulfilling daily shopping needs. The transformation of teristics for predicting housing prices. Thus, the function shopping centres to incorporate functions more suited of urban design features can be investigated further (Qiu to local needs and work-from-home professionals con- et al., 2022). Sixth, more objective factors can be added tributes to the development of a neighbourhood ecosys- to reflect the extent to which particular residents make tem and commons. There is an opportunity to create a well-informed decisions. Collecting pre- and post-pan- new image of a social and civic neighbourhood by add- demic rent or housing prices may serve as a validation ing more layers of urban vibrancy and activities (Florida dataset for comparing with subjective opinions (L. Yang et al., 2021; Mehta, 2022a). The pandemic necessitates et al., 2022). In addition to these limitations and oppor- adjustments to the city’s current planning and design tunities for future development, the methodology frame- methods. We should learn from these ongoing lessons, work presented in this paper has been demonstrated to construct future cities that are healthier and more resil- perform statistically and accurately. It can be applied to ient (Lak et al., 2021), and be more prepared for future other cities while conducting a systematic survey involv- uncertainties (Wang, 2021). ing a wide variety of groups and neighbourhood samples. 5 Limitations6 Conclusions This study is a proof-of-concept examination of a meth - This study fills in gaps in the understanding of urban resi - odology and workflow that has limitations. First, we col - dents’ residence preferences by examining the trade-offs lected data from design students and young professionals between urban spatial qualities and public facilities for to analyse the preferences of Hong Kong’s working popu- various work modes. It has utilised numerous ML and lation. In other words, our study focuses primarily on the DL models to predict reference scores for various scenar- working preferences of industry professions related to ios (work-from-home VS work-from-office). The research architectural design. To expand our knowledge, we intend is based on 578 sampled study areas from Hong Kong to solicit the opinions of workers or professionals from and uses SVI to quantify human neighbourhood environ- other fields for future projects. Due to the varied char - ment perceptions via pixel ratios in traditional ML mod- acteristics of the work, it may reveal diverse subjective els or DL using a custom CNN model. It has objectively preferences. To ensure the validity of the sampling data, measured urban spatial quality indicators and combined it is advantageous to classify and select various types of these with perceived scenes to elicit people’s real subjec- home and office populations based on the objectives of tive opinions through a survey gathering the preferences future studies. Second, we plan to use more scientific of the working population on different work modes. sampling methods in our future research. For example, Song et al. Architectural Intelligence (2023) 2:1 Page 14 of 16 Dubey, A., Naik, N., Parikh, D., Raskar, R., & Hidalgo, C. A. (2016). Deep Learning Our evaluation method innovatively integrates street the City: Quantifying Urban Perception at a Global Scale. In B. Leibe, J. scenes, objective neighbourhood data, and subjective Matas, N. Sebe, & M. Welling (Eds.), Computer Vision – ECCV 2016 ( Vol. opinions into a DL prediction model that can compre- 9905, pp. 196–212). Springer International Publishing. https:// doi. org/ 10. 1007/ 978-3- 319- 46448-0_ 12 hensively predict and inform neighbourhood streetscape Ewing, R., & Handy, S. (2009). Measuring the unmeasurable: urban design transformations to be better oriented towards a post- qualities related to walkability. Journal of Urban Design,14(1), 65–84. pandemic condition when selecting different work https:// doi. org/ 10. 1080/ 13574 80080 24511 55 Ewing, R., Handy, S., Brownson, R. C., Clemente, O., & Winston, E. (2006). modes. As long as SVI datasets and spatial quality data Identifying and measuring urban design qualities related to walkability. are accessible, the method can be applied to a variety of Journal of Physical Activity & Health,3(s1), S223–S240. https:// doi. org/ 10. global locations. It provides a workflow for examining 1123/ jpah.3. s1. s223 Ferreira, F. V., & Wong, M. (2020). Estimating Preferences for Neighborhood Ameni- housing preferences for other work population groups ties Under Imperfect Information ( Working Paper No. 28165). National and new research questions; for example, the objective Bureau of Economic Research. https:// doi. org/ 10. 3386/ w28165 attributes can include specific data, such as crime rates Fu, X., Jia, T., Zhang, X., Li, S., & Zhang, Y. (2019). Do street-level scene percep- tions affect housing prices in Chinese megacities? An analysis using open or rent prices, to expand the independent variables. access datasets and deep learning. PLoS ONE,14(5), e0217505. https:// doi. This research aims to one day enable the evaluation of org/ 10. 1371/ journ al. pone. 02175 05 neighbourhoods based on the preferences of a variety of Glaeser, E. L., Kominers, S. D., Luca, M., & Naik, N. (2018). Big data and big cities: the promises and limitations of improved measures of urban life. Eco- working population groups by predicting their competi- nomic Inquiry,56(1), 114–137. https:// doi. org/ 10. 1111/ ecin. 12364 tiveness in attracting target residents by toggling on and Gong, F.-Y., Zeng, Z.-C., Zhang, F., Li, X., Ng, E., & Norford, L. K. (2018). Mapping off selected work modes, thereby developing strategic sky, tree, and building view factors of street canyons in a high-density urban environment. Building and Environment,134, 155–167. https:// doi. economic sectors and preparing development plans. org/ 10. 1016/j. build env. 2018. 02. 042 Hart, K. (2020, April 30). Why coronavirus may prompt migration out of American cities. Axios. https:// www. axios. com/ 2020/ 04/ 30/ coron avirus- migra tion- Authors’ contributions ameri can- cities- survey Qiwei Song drafted the final and revised version of the publication, analysed TM Herbrich, R., Minka, T., & Graepel, T. (2006). TrueSkill : A Bayesian Skill Rating data, gathered additional data, and created diagrams. Zhiyi Dou drafted System. Advances in Neural Information Processing Systems, 19. https:// the initial version of the paper, gathered original data, and performed the proce edings. neuri ps. cc/ paper/ 2006/ hash/ f44ee 26395 2e65b 3610b 8ba51 preliminary analysis. Waishan Qiu supervised the project’s development and 229d1 f9- Abstr act. html contributed to the analysis framework. Wenjing Li contributed to data cleans- Hong Kong Development Bureau, & Hong Kong Planning Bureau. (2021). Hong ing and data enhancement. Jingsong Wang designed the interactive survey’s Kong 2030+: Final Recommentations—Booklet. https:// www. pland. gov. hk/ website. Jeroen van Ameijde provided site context and revised the draft. Dan pland_ en/p_ study/ comp_s/ hk203 0plus/ strat egy_a. htm Luo conceived the study, contributed to its conception and coordination, and Hu, M., Roberts, J. D., Azevedo, G. P., & Milner, D. (2021). The role of built and supervised the writing of the manuscript. The author(s) read and approved social environmental factors in Covid-19 transmission: a look at America’s the final manuscript. capital city. Sustainable Cities and Society,65, 102580. https:// doi. org/ 10. 1016/j. scs. 2020. 102580 Funding Hu, Y., & Han, Y. (2019). Identification of Urban Functional Areas Based on POI This work is supported by the University of Queensland Global Strategy and Data: A Case Study of the Guangzhou Economic and Technological Partnerships Seed Funding Scheme. Development Zone. Sustainability, 11(5), Article 5. https:// doi. org/ 10. 3390/ su110 51385 Availability of data and materials Huang, Z., Chen, R., Xu, D., & Zhou, W. (2017). Spatial and hedonic analysis of The datasets used and analysed during the current study are available from housing prices in Shanghai. Habitat International,67, 69–78. https:// doi. the corresponding author upon reasonable request. org/ 10. 1016/j. habit atint. 2017. 07. 002 Ito, K., & Biljecki, F. (2021). Assessing bikeability with street view imagery and computer vision. Transportation Research Part c: Emerging Technolo- Declarations gies,132, 103371. https:// doi. org/ 10. 1016/j. trc. 2021. 103371 Jacobs, J. (1961). The Death and Life of Great American Cities. Random House. Competing interests Kang, Y., Zhang, F., Gao, S., Peng, W., & Ratti, C. (2021a). Human settlement value Non-financial competing interests include (but are not limited to) political, assessment from a place perspective: Considering human dynamics and personal, religious, ideological, academic, and intellectual competing interests. perceptions in house price modeling. Cities,118, 103333. https:// doi. org/ 10. 1016/j. cities. 2021. 103333 Kang, Y., Zhang, F., Peng, W., Gao, S., Rao, J., Duarte, F., & Ratti, C. (2021b). Under- Received: 30 September 2022 Accepted: 30 December 2022 standing house price appreciation using multi-source big geo-data and machine learning. Land Use Policy,111, 104919. https:// doi. org/ 10. 1016/j. landu sepol. 2020. 104919 Kwok, C. Y. T., Wong, M. S., Chan, K. L., Kwan, M.-P., Nichol, J. E., Liu, C. H., Wong, J. Y. H., Wai, A. K. C., Chan, L. W. C., Xu, Y., Li, H., Huang, J., & Kan, Z. (2021). References Spatial analysis of the impact of urban geometry and socio-demographic Agostini, C. A., & Palmucci, G. A. (2008). The Anticipated Capitalisation Eec ff t characteristics on COVID-19, a study in Hong Kong. Science of the Total of a New Metro Line on Housing Prices*. Fiscal Studies,29(2), 233–256. Environment,764, 144455. https:// doi. org/ 10. 1016/j. scito tenv. 2020. 144455 https:// doi. org/ 10. 1111/j. 1475- 5890. 2008. 00074.x La Rosa, D. (2014). Accessibility to greenspaces: GIS based indicators for Beleites, C., Neugebauer, U., Bocklitz, T., Krafft, C., & Popp, J. (2013). Sample size sustainable planning in a dense urban context. Ecological Indicators,42, planning for classification models. Analytica Chimica Acta,760, 25–33. 122–134. https:// doi. org/ 10. 1016/j. ecoli nd. 2013. 11. 011 https:// doi. org/ 10. 1016/j. aca. 2012. 11. 007 Lak, A., Hakimian, P., & Sharifi, A. (2021). An evaluative model for assessing Bissell, D. (2021). A changing sense of place: Geography and COVID-19. pandemic resilience at the neighborhood level: The case of Tehran. Geographical Research,59(2), 150–159. https:// doi. org/ 10. 1111/ 1745- 5871. Sustainable Cities and Society,75, 103410. https:// doi. org/ 10. 1016/j. scs. 2021. 103410 S ong et al. Architectural Intelligence (2023) 2:1 Page 15 of 16 Lee, C., & Moudon, A. V. (2006). Correlates of walking for transportation or International Journal of Geo-Information, 10(8), Article 8. https:// doi. org/ recreation purposes. Journal of Physical Activity and Health,3(s1), S77–S98. 10. 3390/ ijgi1 00804 93 https:// doi. org/ 10. 1123/ jpah.3. s1. s77 R Florida A Rodríguez-Pose M Storper 2021 Cities in a post-COVID world Urban Li, X., Zhang, C., Li, W., Ricard, R., Meng, Q., & Zhang, W. (2015). Assessing street- Studies 00420980211018072 https:// doi. org/ 10. 1177/ 00420 98021 10180 level urban greenery using Google Street view and a modified green 72 view index. Urban Forestry & Urban Greening,14(3), 675–685. https:// doi. Rode, P., Floater, G., Thomopoulos, N., Docherty, J., Schwinger, P., Mahendra, A., org/ 10. 1016/j. ufug. 2015. 06. 006 & Fang, W. (2017). Accessibility in cities: Transport and urban form (G. Meyer Li, H., Wei, Y. D., Yu, Z., & Tian, G. (2016). Amenity, accessibility and housing & S. Shaheen, Eds.; pp. 239–273). Springer International Publishing. http:// values in metropolitan USA: a study of Salt Lake County, Utah. Cities,59, eprin ts. lse. ac. uk/ 100118/ 113–125. https:// doi. org/ 10. 1016/j. cities. 2016. 07. 001 Rosen, S. (1974). Hedonic prices and implicit markets: product differentiation Li, H., Wei, Y. D., Wu, Y., & Tian, G. (2019). Analyzing housing prices in Shanghai in pure competition. Journal of Political Economy. https:// doi. org/ 10. 1086/ with open data: amenity, accessibility and urban structure. Cities,91, 260169 165–179. https:// doi. org/ 10. 1016/j. cities. 2018. 11. 016 Salesses, P., Schechtner, K., & Hidalgo, C. A. (2013). The collaborative image Li, Y., Yabuki, N., & Fukuda, T. (2022). Exploring the association between street of the City: mapping the inequality of urban perception. PLoS ONE,8(7), built environment and street vitality using deep learning methods. e68400. https:// doi. org/ 10. 1371/ journ al. pone. 00684 00 Sustainable Cities and Society,79, 103656. https:// doi. org/ 10. 1016/j. scs. Sevtsuk, A., Basu, R., Li, X., & Kalvo, R. (2021). A big data approach to under- 2021. 103656 standing pedestrian route choice preferences: evidence from San Liu, J., Bi, H., & Wang, M. (2020). Using multi-source data to assess livability in Francisco. Travel Behaviour and Society,25, 41–51. https:// doi. org/ 10. 1016/j. Hong Kong at the community-based level: a combined subjective-objec-tbs. 2021. 05. 010 tive approach. Geography and Sustainability,1(4), 284–294. https:// doi. org/ Song, Q., Li, M., Qiu, W., Li, W., & Luo, D. (2022a). The Coherence and Divergence 10. 1016/j. geosus. 2020. 12. 001 Between the Objective and Subjective Measurement of Street Percep- Lu, Y. (2019). Using Google Street View to investigate the association between tions for Shanghai. In W. Chen, L. Yao, T. Cai, S. Pan, T. Shen, & X. Li (Eds.), street greenery and physical activity. Landscape and Urban Planning,191, Advanced Data Mining and Applications (pp. 244–256). Springer Nature 103435. https:// doi. org/ 10. 1016/j. landu rbplan. 2018. 08. 029 Switzerland. https:// doi. org/ 10. 1007/ 978-3- 031- 22064-7_ 19 Lynch, K. (1960). The image of the city. MIT Press. Song, Q., Li, W., Li, M., & Qiu, W. (2022b). Social Inequalities in Neighborhood-Level Ma, X., Ma, C., Wu, C., Xi, Y., Yang, R., Peng, N., Zhang, C., & Ren, F. (2021). Measur- Streetscape Perceptions in Shanghai: The Coherence and Divergence between ing human perceptions of streetscapes to better inform urban renewal: a the Objective and Subjective Measurements (SSRN Scholarly Paper No. perspective of scene semantic parsing. Cities,110, 103086. https:// doi. org/ 4179127). https:// doi. org/ 10. 2139/ ssrn. 41791 27 10. 1016/j. cities. 2020. 103086 Song, Q., Liu, Y., Qiu, W., Liu, R., & Li, M. (2022c). Investigating the Impact of Per- Mehta, V. (2020). The new proxemics: COVID-19, social distancing, and sociable ceived Micro-Level Neighborhood Characteristics on Housing Prices in space. Journal of Urban Design,25(6), 669–674. https:// doi. org/ 10. 1080/ Shanghai. Land, 11(11), Article 11. https:// doi. org/ 10. 3390/ land1 11120 02 13574 809. 2020. 17852 83 Wang, J. (2021). Vision of China’s future urban construction reform: In the Mehta, V. (2022a). Main street commons: from shopping street to neighbour- perspective of comprehensive prevention and control for multi disasters. hood commons. Journal of Urban Design,27(1), 59–62. https:// doi. org/ 10. Sustainable Cities and Society,64, 102511. https:// doi. org/ 10. 1016/j. scs. 1080/ 13574 809. 2022. 19947 662020. 102511 Minka, T., Cleven, R., & Zaykov, Y. (2018). TrueSkill 2: An improved Bayesian skill Wen, H., Xiao, Y., & Zhang, L. (2017). School district, education quality, and rating system. https:// www. micro soft. com/ en- us/ resea rch/ publi cation/ housing price: evidence from a natural experiment in Hangzhou, China. trues kill-2- impro ved- bayes ian- skill- rating- system/ Cities,66, 72–80. https:// doi. org/ 10. 1016/j. cities. 2017. 03. 008 Mitra, R., Moore, S. A., Gillespie, M., Faulkner, G., Vanderloo, L. M., Chulak-Bozzer, Wong, A. H. K., Cheung, J. O., & Chen, Z. (2020). Promoting effectiveness of T., Rhodes, R. E., Brussoni, M., & Tremblay, M. S. (2020). Healthy move- “working from home”: findings from Hong Kong working population ment behaviours in children and youth during the COVID-19 pandemic: under COVID-19. Asian Education and Development Studies,10(2), 210–228. exploring the role of the neighbourhood environment. Health & Place,65, https:// doi. org/ 10. 1108/ AEDS- 06- 2020- 0139 102418. https:// doi. org/ 10. 1016/j. healt hplace. 2020. 102418 Wut, T. M., Lee, S. W., Xu, J., & (Bill). (2022). Work from home challenges of the Monkkonen, P., & Zhang, X. (2011). Socioeconomic Segregation in Hong Kong: Pandemic Era in Hong Kong: a stimulus-organism-response perspective. Spatial and Ordinal Measures in a High-Density and Highly Unequal City. International Journal of Environmental Research and Public Health,19(6), https:// escho larsh ip. org/ uc/ item/ 5f097 4ww 3420. https:// doi. org/ 10. 3390/ ijerp h1906 3420 Montgomery, J. (1998). Making a city: Urbanity, vitality and urban design. Xu, M., Xin, J., Su, S., Weng, M., & Cai, Z. (2017). Social inequalities of park acces- Journal of Urban Design,3(1), 93–116. https:// doi. org/ 10. 1080/ 13574 80980 sibility in Shenzhen, China: the role of park quality, transport modes, and 87244 18 hierarchical socioeconomic characteristics. Journal of Transport Geogra- Naik, N., Philipoom, J., Raskar, R., & Hidalgo, C. (2014). Streetscore – predicting phy,62, 38–50. https:// doi. org/ 10. 1016/j. jtran geo. 2017. 05. 010 the perceived safety of one million streetscapes. IEEE Conference on Com- Xu, X., Qiu, W., Li, W. ., Liu, X., Zhang, Z., Li, X., & Luo, D. (2022). Associations puter Vision and Pattern Recognition Workshops,2014, 793–799. https:// doi. between street-view perceptions and housing prices: subjective vs. org/ 10. 1109/ CVPRW. 2014. 121 objective measures using computer vision and machine learning tech- Netusil, N. R. (2013). Urban environmental amenities and property values: Does niques. Remote Sensing, 14(4), 891. https:// doi. org/ 10. 3390/ rs140 40891 ownership matter? Land Use Policy,31, 371–377. https:// doi. org/ 10. 1016/j. Yang, Y., & Xiang, X. (2021). Examine the associations between perceived landu sepol. 2012. 07. 016 neighborhood conditions, physical activity, and mental health during the Ottoni, C. A., Winters, M., & Sims-Gould, J. (2022). “We see each other from COVID-19 pandemic. Health & Place,67, 102505. https:// doi. org/ 10. 1016/j. a distance”: Neighbourhood social relationships during the COVID-19 healt hplace. 2021. 102505 pandemic matter for older adults’ social connectedness. Health & Place,76, Yang, J., Rong, H., Kang, Y., Zhang, F., & Chegut, A. (2021). The financial impact 102844. https:// doi. org/ 10. 1016/j. healt hplace. 2022. 102844 of street-level greenery on New York commercial buildings. Landscape Qiu, W., Zhang, Z., Liu, X., Li, W., Li, X., Xu, X., & Huang, X. (2022). Subjective or and Urban Planning,214, 104162. https:// doi. org/ 10. 1016/j. landu rbplan. objective measures of street environment, which are more effective in 2021. 104162 explaining housing prices? Landscape and Urban Planning,221, 104358. Yang, L., Liang, Y., He, B., Lu, Y., & Gou, Z. (2022). COVID-19 effects on property https:// doi. org/ 10. 1016/j. landu rbplan. 2022. 104358 markets: the pandemic decreases the implicit price of metro accessibility. Qiu, W., Li, W., Liu, X., Zhang, Z., Li, X., & Huang, X. (2023). Subjective and objec- Tunnelling and Underground Space Technology,125, 104528. https:// doi. tive measures of streetscape perceptions: Relationships with property org/ 10. 1016/j. tust. 2022. 104528 value in Shanghai. Cities,132, 104037. https:// doi. org/ 10. 1016/j. cities. 2022. Yue, Y., Zhuang, Y., Yeh, A. G. O., Xie, J.-Y., Ma, C.-L., & Li, Q.-Q. (2017). Measure- 104037 ments of POI-based mixed use and their relationships with neigh- Qiu, W., Li, W., Liu, X., & Huang, X. (2021). Subjectively Measured Streetscape bourhood vibrancy. International Journal of Geographical Information Perceptions to Inform Urban Design Strategies for Shanghai. ISPRS Science,31(4), 658–675. https:// doi. org/ 10. 1080/ 13658 816. 2016. 12205 61 Song et al. Architectural Intelligence (2023) 2:1 Page 16 of 16 Mehta, V. (2022b) Unparalleled prospect: COVID-19 and the expansion of pub- lic space. Journal of Urbanism: International Research on Placemaking and Urban Sustainability. 1–4. https:// doi. org/ 10. 1080/ 17549 175. 2022. 21461 56 Zhang, F., Zhou, B., Liu, L., Liu, Y., Fung, H. H., Lin, H., & Ratti, C. (2018). Measuring human perceptions of a large-scale urban region using machine learning. Landscape and Urban Planning,180, 148–160. https:// doi. org/ 10. 1016/j. landu rbplan. 2018. 08. 020 Zhang, X., Sun, Z., Ashcroft, T., Dozier, M., Ostrishko, K., Krishan, P., McSwig- gan, E., Keller, M., & Douglas, M. (2022). Compact cities and the Covid-19 pandemic: systematic review of the associations between transmission of Covid-19 or other respiratory viruses and population density or other features of neighbourhood design. Health & Place,76, 102827. https:// doi. org/ 10. 1016/j. healt hplace. 2022. 102827 Zhao, H., Shi, J., Qi, X., Wang, X., & Jia, J. (2017). Pyramid scene parsing network. IEEE Conference on Computer Vision and Pattern Recognition (CVPR),2017, 6230–6239. https:// doi. org/ 10. 1109/ CVPR. 2017. 660 Zhou, H., He, S., Cai, Y., Wang, M., & Su, S. (2019). Social inequalities in neighbor- hood visual walkability: using street view imagery and deep learning technologies to facilitate healthy city planning. Sustainable Cities and Society,50, 101605. https:// doi. org/ 10. 1016/j. scs. 2019. 101605 Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., & Torralba, A. (2018). Semantic Understanding of Scenes through the ADE20K Dataset. Inter- national Journal of Computer Vision, 127(3), 302–321. https:// doi. org/ 10. 1007/ s11263- 018- 1140-0 Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in pub- lished maps and institutional affiliations.
Architectural Intelligence – Springer Journals
Published: Jan 27, 2023
Keywords: Urban data; Sustainable communities; Post-pandemic; Deep learning; Convolutional neural network; Trade-off
Access the full text.
Sign up today, get DeepDyve free for 14 days.