From 3D city modelling to digital twin smart city : a study of machine learning of private residential estate valuation using 3D environmental features in Hong Kong

Abstract

Valuation is both an art and a science. Professional surveyors often use subjective judgement to assess a property’s hedonic features in order to make percentage adjustments based on comparisons with similar properties. However, traditional market comparison methods may lack accuracy due to asymmetric access to information and subjectivity when evaluating neighbourhood attributes, such as views, sunlight, and natural ventilation. Therefore, this study integrates quantitative 3D Environmental Features derived from digital twins to improve valuation accuracy and reduce reliance on subjective assessments in Hong Kong. Notably, this dissertation is the first-ever study, as far as is known, to systematically incorporate comprehensive 3D Environmental Features, specifically simulated window views under the City Information Model (CIM), sunlight, and ventilation, into both conventional and machine-learning-based valuation models. The approach and findings mark a significant advancement in General Practice Surveying Industry. Methodologically, this study expanded the traditional hedonic pricing model, a well-known method for valuing property characteristics, by integrating novel 3D urban features. The primary dataset encompassed transaction records from over 14,034 residential property transactions of selected housing estates, with the most frequent transactions across Hong Kong Island and Kowloon. Additionally, new 3D view indices were generated by positioning virtual cameras within photorealistic 3D city models at precise locations corresponding to actual building windows. These generated images were then analysed through semantic segmentation, an advanced image content analysis, to quantify scenery elements, including greenery, water bodies, sky openness, and built structures, yielding measurable Window View Indices (WVIs) and Window View Openness Index (WVOI). Sunlight exposure was accurately quantified using Rhino3D and the Ladybug simulation tool. Meanwhile, natural ventilation intake was estimated based on Windrose Diagrams while considering prevailing wind directions, building orientations, and surrounding urban densities. Regression analyses and various machine-learning algorithms, such as Random Forest, k-Nearest Neighbours, Decision Trees, and Artificial Neural Network, were employed to evaluate these new variables’ predictive capability. Empirical findings confirmed that all introduced 3D Environmental Features significantly influence property prices, with results aligning closely with existing valuation literature. Incorporating these features markedly improved valuation precision, reducing Mean Absolute Percentage Error (MAPE), a measure of prediction accuracy, from 5.96% to 4.85% (i.e., by 18.62% fewer errors), indicating these variables effectively capture variations in environmental conditions. Furthermore, machine-learning algorithms, particularly Random Forest, outperformed traditional hedonic models, achieving the strongest explanatory power (R-squared: 0.985) and lowest MAPEs (3.8%). The experiments’ increased accuracy demonstrates their capability to capture the complex, nonlinear relationships between environmental attributes and property prices. By integrating new 3D Environmental Features into the analysis, this study offers unprecedented insights into the locational variation of environmental determinants on property prices in Hong Kong. For instance, green views are significant in housing prices, and open views yield incremental price premiums for properties remote from coastal districts and areas with higher building density. Direct sunlight only reduces prices for properties at high altitudes, and property orientations are mainly related to sunlight exposure when determining prices. These nuanced insights underscore the profound professional implications of the research. Equally significant is the demonstration that automating environmental assessments through digital twin technologies enhances valuation efficiency by significantly reducing the need for subjective judgement, costly site visits, or questionnaire-based evaluations. This reduction improves consistency and cost-effectiveness. Ultimately, this study introduces as a pioneering effort the application of digital twins for quantitative real estate valuation, supporting smart and sustainable urban development both in Hong Kong and internationally. Future research could expand the analysis to additional housing estates territory-wide, contingent on the availability of an up-to-date, high-quality, territorial-level 3D city.