Efficient assessment of window views in high-rise, high-density urban areas using 3D color City Information Models Maosu Li, Fan Xue, Anthony G.O. Yeh Faculty of Architecture, HKU 18th International Conference on Computational Urban Planning and 1 Urban Management (CUPUM 2023), 20-22 June, Montreal, Canada Maosu Li Scan me. 👈👈 Final-year PhD Candidate, Depart of Urban Planning and Design, The University of Hong Kong. Supervisors: Prof. Anthony Gar On Yeh, Department of Urban Planning and Design; Dr. Fan Xue, Department of Real Estate and Construction. Research interest: Create automatic decision support methods and tools  quantify urban semantics through 3D City Information Modeling, Machine Learning, and Data Analytics  smarter urban planning and urban management. Human-perceived ① Openness Objective Feature Distance … Structure … Window view Step 1：Discovery and initial representation Step 2：Fine-scale modeling Step 3：Accuracy and efficiency 2 CONTENT 1 Introduction 2 Research methods 3 Initial results 4 Summary 3 1 Introduction 2 Research methods 3 Initial results 4 Summary 4 1.1 Background  A high-quality window view, • with more greenery, sky, waterbody, and fewer construction elements • treasured by urban dwellers especially in high-rise, high-density (HRHD) urban areas * * * Stress relief Living satisfaction Recognized benefits Productivity improvement * Increasingly long-term indoor occupation Note: * (Source: flaticon.com) 5 1.1 Background  Assessment of window views, quantified evidence for multiple urban applications  However, window views are numerous especially in HRHD areas,  change in large numbers with the vertical development of neighborhoods * * * So many views to be assessed Both manual and automatic methods Thus, both efficient and accurate assessment of window views • Can aid housing property agencies, architectural designers, and urban planners Need plenty of resources! • Significant advancing the window view assessment for urban-scale applications Housing selection inBuilding space Built environment and valuation planning improvement Should I still do it? Note: * (Source: flaticon.com) Urban-scale window view assessment (Source: wallpaperflare.com) 6 1.2 Literature review  Window view assessment: Manual measurement and Simulation Method type Example Field Status Problem Manual measurement  Onsite photo collection  Psychology  Built environment  Architectural design High cost and laborious Unscalable to the urban scale Simulation  Visibility analysis Still shows “preference” for oversimplified models (Fig. 1) Inaccurate  Urban planning and design  Geographic Thus, next generation of assessment methods Information  Model view System photography Embrace 3D photorealistic City Models (CIMs) (Fig. 2)  improve the processing efficiency for an accurate Information quantification, More accurate but inefficient Supporting urban-scale assessment models and update of window view Fig. indices. 2 Model view photography Fig. 1 Traditional visibility analysis: Oversimplified Simulated (Cimburova and Blumentrath. 2022) (S.M. Labib et al. 2021) Real a) High burden of texture image loading and rendering (Li, M. et al. 2021; 2022; 2023) b) Repetitive 2D segmentation 7 1 Introduction 2 Research methods 3 Initial results 4 Summary 8 2.1 Workflow of the proposed method  Input: Four datasets  Methods: 3D semantic segmentation + model view photography + color pixel counting  Output: Four Window View Indices (WVIs) • greenery, waterbody, sky, and construction. 9 2.2 Definition of WVIs  WVIs: Defined as a ratio ranging from 0 to 1 on an 8-bit RGB color view image • Extension of definition defined on a 3D photorealistic scene (Li et al. 2022). 3D photorealistic scene Waterbody (Water) Construction (Const.) 8-bit RGB color window view image c Greenery (Green) WVIl (Li et al.’s (2022) method) 3D color scene Eq. 1 WVIgreen WVIwater WVIsky WVIconst WVIl = |{p| p∈c, m(pcolor) = l}| ∕ n, l ∈L, L= {‘greenery’, ‘waterbody’, ‘sky’, ‘const.’}, (1) 3D scene colored by L 10 2.3 3D semantic segmentation of CIM for a 3D color scene  3D semantic segmentation: KPConv (Thomas et al. 2019) and A priori-based rules  3D scene color setting 3D semantic segmentation of photorealistic mesh models 3D semantic segmentation of DSM using NDVI Eq. 2 KPConv Greenery Const. Eq. 3 Waterbody Eq. 3 Greenery Waterbody Const. Eq. 2 RGB (0,255,0), l = greenery, RGB (0,0,255), l = waterbody, color(vl ) = � RGB (255,255,255), l = sky, RGB (255,0,0), l = construction, greenery, NDVIpg ≥ 0.1, lpg = �construction, 0 ≤ NDVIpg ≤ 0.1, waterbody, NDVIpg = no data, Greenery Waterbody Sky Const. 11 2.4 Batch computation of WVIs using color view images  Two-step computation process • Window view generation in the 3D color scene • Color pixel counting for WVIs 3D color scene from Section 2.3 Color pixel counting for WVIs Batch generation of color window view images 1 3 4 ③ ② ④ Li et al. (2022) 2 ① Greenery19,359 pixels Waterbody22,280 pixels sky 381,661 pixels const.  386,700 pixels greenery, pcolor = RGB (0,255,0), waterbody, pcolor = RGB (0,0,255), m(pcolor) = � sky, pcolor = RGB (255,255,255), construction, pcolor = RGB (255,0,0). Eq.1 WVIgreen 0.0239 WVIwater 0.0275 WVIsky 0.4712 WVIconst 0.4774 (3) 12 1 Introduction 2 Research methods 3 Initial results 4 Summary 13 3.1 Experimental settings  Using 100 random window views from 207 buildings in To Kwa Wan, Kowloon Peninsula of Hong Kong  To test the feasibility of the proposed method. Dataset 3D photorealistic mesh models and annotations (Lands Department, 2017; Li et al. 2023) Building information models (Urban Renewal Authority, 2022) Digital surface model (Lands Department, 2020) NDVI map (B. Morgan and B. Guénard, 2019) Software 3D segmentation of CIM KPConv (Thomas et al. 2019) for photorealistic mesh ArcGIS Pro (2.9.0) for DSM Batch computation of WVIs Cesium (1.99), Python (ver. 3.7.11) Baseline method (Li et al. 2022) Deeplab V3+ (Chen et al. 2018), Orange 3 (ver. 3.26) Deep learning environment Docker - Ubuntu (ver. 18.04.5), Python (ver. 3.7.11), Pytorch (ver. 1.10.0) Workstation 100 photorealistic views and color views Intel i9-11900K CPU (3.50 GHz, 16 cores) 64 GB memory 24G Nvidia GeForce RTX 3090 graphic card Windows 10 operating system 14 3.2 Accuracy and efficiency  Comparison with Li et al.’s (2022) 2D segmentation method, the proposed 3D segmentation method:   Accuracy  RMSE < 0.01; Improvement: 76.26%; Efficiency  Total < 0.6 s; Improvement: 73%. Li et al.’s (2022) 2D method Models without preparation and rendering of textures ① Models with textures ② More accurate 2D Segmentation segmentation results Our 3D segmentation method WVIs Pixel counting WVIs Only pixel counting 15 1 Introduction 2 Research methods 3 Initial results 4 Summary 16 4 Summary  This study proposes a both efficient and accurate window view assessment method • Using 3D semantic segmentation and 3D color CIM  Significance • Improvement of the accuracy and efficiency • For urban-scale quantification and update of four WVIs • RMSE < 0.01 and 3.68 times faster • Advancing urban-scale planning, design, and real estate applications to use quantified WVIs • Urban planners and architectural designers in urban planning and design • Housing purchasers, renters, property agencies in real estate market  Limitation • A full 3D color scene needed for assessing four WVIs • Small-scale quantification may not afford the large-scale but one-off preprocessing cost • Batch quantification of window views regardless of the similarity • Window view pattern mining for a more efficient assessment 17 References  Chen, L. C., Zhu, Y., Papandreou, G., Schroff, F., & Adam, H. (2018). Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV) (pp. 801-818).  Cimburova, Z., & Blumentrath, S. (2022). Viewshed-based modelling of visual exposure to urban greenery–An efficient GIS tool for practical planning applications. Landscape and Urban Planning, 222, 104395.  HKLandsD. (2017). 3D Visualization Map. Hong Kong: Lands Department, Government of Hong Kong SAR.  HKLandsD. (2020). Digital Surface Model from 2020 LiDAR Survey. Hong Kong: Lands Department, Government of Hong Kong SAR.  HKURA. (2022). 3D Intelligent Map of an area in To Kwa Wan. Hong Kong: Urban Renewal Authority.  Labib, S. M., Huck, J. J., & Lindley, S. (2021). Modelling and mapping eye-level greenness visibility exposure using multi-source data at high spatial resolutions. Science of the Total Environment, 755, 143050.  Li, M., Xue, F., Yeh, A. G., & Lu, W. (2021). Classification of photo-realistic 3D window views in a high-density city: The case of Hong Kong. In Proceedings of the 25th International Symposium on Advancement of Construction Management and Real Estate (pp. 1339-1350). Springer Singapore.  Li, M., Xue, F., Wu, Y. & Yeh, A. G. (2022). A room with a view: Automated assessment of window views for high-rise high-density areas using City Information Models and transfer deep learning. Landscape and Urban Planning, 226, 104505. doi:10.1016/j.landurbplan.2022.104505  Li, M., Xue, F., & Yeh, A. G. (2023). Bi-objective analytics of 3D visual-physical nature exposures in high-rise high-density cities for landscape and urban planning. Landscape and Urban Planning, 233, 104714.  Morgan, B., & Guénard, B. (2019). New 30 m resolution Hong Kong climate, vegetation, and topography rasters indicate greater spatial variation than global grids within an urban mosaic. Earth System Science Data, 11(3), 1083-1098.  Thomas, H., Qi, C. R., Deschaud, J. E., Marcotegui, B., Goulette, F., & Guibas, L. J. (2019). Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 6411-6420). 18 Thank you for your attention! Scan me. 👈👈 Li Maosu, PhD. Candidate Dept. of Urban Planning and Design, HKU maosulee@connect.hku.hk 19