2018 5th NAFOSTED Conference on Information and Computer Science (NICS) Preliminary Result of 3D City Modelling For Hanoi, Vietnam Phan Anh Center of Multidisciplinary Integrated Technology for Field Monitoring University of Engineering and Technology, VNUH Hanoi, Vietnam anhp@fimo.edu.vn Nguyen Thi Nhat Thanh Center of Multidisciplinary Integrated Technology for Field Monitoring University of Engineering and Technology, VNUH Hanoi, Vietnam thanhntn@fimo.edu.vn Chu Thua Vu Center of Multidisciplinary Integrated Technology for Field Monitoring University of Engineering and Technology, VNUH Hanoi, Vietnam vuct@fimo.edu.vn Nguyen Viet Ha University of Engineering and Technology, VNUH Hanoi, Vietnam hanv@vnu.edu.vn Abstract² Hanoi is the one of the fastest-growing cities in Vietnam, which sets the target to turn into a smart city in 2030 Nowadays, 3D city models are being increasingly employed for many domains and tasks beyond visualization, then it will take an important role in smart city In order to develop 3D city models, 2D geographic data such as building footprint and building height attribute are required However, the lack of the height attribute for various types of building and low performance of rendering and visualizing 3D city models are two big remaining problems In this paper, available data from open sources is used to predict the building height The prediction has carried out with machine learning techniques using the combination of different attributes After that, the models will be created using 3D tiles specification to improve the visualization performance The preliminary results of the proposed method highlight the potential of generation of massive 3D city models from the available data in Vietnam Keywords— Hanoi, 3D city models, building heights, building footprint, 3D tiles I INTRODUCTION AND RELATED WORKS Using digital maps to describe the real world is no longer something new to us as we are moving into the 4th industrial revolution 2D maps now cannot meet the growing demand of people in urban management issues but a 3D map is a perfect replacement for that By representing objects in 3D environment, 3D maps always create compelling visibility into all relevant areas, especially in the Geographic Information System (GIS) GIS is a virtual geographic environment, including systems of hardware, software, geographic and human data, which are designed to capture, store, update, control, and analyze all types of information related to geographic information Through GIS, researchers can easily look at phenomena, geographic-related changes in the most general way, thus making the most accurate decisions and solutions to the reality situation 3D GIS technology can be considered as a virtual environment system which is full of information about all objects in real life This will bring benefits not only as a geographic information system but also as a basis for a smart city system in the future Currently, the use of 3D GIS technology to build virtual city has been realized in many cities in the world, such as New York [1], Berlin [2], etc These studies use 3D mapping technology to display 3D object models with information such 978-1-5386-7983-8/18/$31.00 ©2018 IEEE Bui Quang Hung Center of Multidisciplinary Integrated Technology for Field Monitoring University of Engineering and Technology, VNUH Hanoi, Vietnam hungbq@fimo.edu.vn as name, coordinates, elevation, etc of the object The 3D city maps of Berlin and New York City have hierarchical data display mechanisms based on user location and perspective The objects will not be displayed on the invisible part of the map This not only increases the accuracy of the user experiment but also reduces the amount of graphics processing and speeds up the system for a better experience In particular, more than 1.1 million buildings without texture in New York City and many textured buildings in scale 250 square kilometers in Berlin have been visualized in both systems, respectively In Vietnam, the problem of building virtual city is still quite new and no such a 3D virtual city is reported Available systems such as OpenStreetMap, Google Earth have also been providing 3D models, but the number of these models is limited and most of them are skyscrapers In 2017, Phan Anh et al developed The VNU Virtual Campus system which demonstrates the model of buildings and trees distribution in the campus of Hanoi National University [3] This system allows user movement as well as changing 360 degree viewing angles However, the system has many difficulties in loading and processing of large data In particular, manual data collection has been a major constraint to future largescale system development The lack of cadaster data and difficulties in automatic processing data such as footprint detection, building height detection, etc are becoming the most challenges needed to be addressed in Vietnam In addition, in 2011, Dinh Thi Bao Hoa et al [4] have also built 3D models for the West Lake area to promote tourism The advantage of this system is ability of displaying 3D models of houses, trees, traffic with high accuracy However, there still has some drawbacks such as: the system is only visible on specialized software, it has no online capability, and is not scalable due to the use of existing manual 3D models Through these two products, the main problem for 3D GIS system development for Vietnam is data deprivation and data processing efficiency for a large amount of 3D models The most important thing to build virtual city is to have enough input data with high accuracy Depending on the level of detail, data will contain the information of the building footprint, building height, story, roof type, roof height, etc In the previous works, the information is provided through a variety of sources with different quality for New York and 294 2018 5th NAFOSTED Conference on Information and Computer Science (NICS) Berlin However, acquiring this information in Vietnam is very difficult There is not such a complete set of free and public data, especially building height Moreover, Vietnamese houses have complex architectures and building footprints often have complex shapes Currently, LiDAR, highresolution satellite imagery and aerial images are used largely to gather information for 3D city modelling However, all of these methods require costly equipment and many related licenses In order to deal with the problem of building height deprivation, F Biljecki et al proposed a method for building height estimation based on related data [5] In the study, a prediction has been carried out with a machine learning technique (random forest) using 10 different attributes and their combinations which are footprints and attributes available in volunteered geoinformation and cadaster Results obtained in the study achieved a mean error of 0.8 m Recognizing the feasibility of this approach, we will apply the based method to deploy a machine learning model for building height prediction using information that can be collected with the capability and scale of Hanoi This will help to address the lack of building heights in Vietnam These data will then be used to build 3D models using open specification (3D tiles) provide by Cesium to enhance the performance of rendering and visualizing 3D models II STUDY AREA AND METHODOLOGY A Study area Hanoi is the capital of Vietnam and the country's second largest city by population The population in 2015 was estimated at 7.7 million people Hanoi covers an urban area of 319.56 km2 Hence this area is a good option for our case study (Fig 1) I They are inherited from the study of F Biljecki et al 2017, in which features including purpose of use, perimeter, the area of the building, the relationship between the perimeter and the area of the buildings are collected TABLE I ATTRIBUTES USED TO THE DEPLOY THE PREDICTOR MODEL No Attribute U Usage A Area P Perimeter NPI Normalized perimeter index NPI = Description ଶξగ C Prediction methodology The machine learning method that we use to build the model for predicting height is the Random Forest Regression This is a machine learning method built on the basis of a decision tree This method is used mainly for classifying, regression calculations It will make a lot of decision trees during training and then return the values as classes for the purpose of classifying or returning the predicted values of individual trees for the purpose of computation In fact, the use of the Random Forest in most cases is better than some of the other popular machine learning methods, such as decision trees, k-nearest neighbors, etc The Random Forest method can also work well with inadequate data and avoid the overlapping of the overflow model To apply Random Forest into predictor model, Scikit-learn library was implemented using the Python programming language After the model was deployed, the results will be evaluated in terms of Mean Absolute Error – MAE (1), Relative Error – RE (3), Root Mean Squared Error – RMSE (3) Mean absolute error or MAE is a measure of the difference between two consecutive variables The MAE unit is the unit of the variables of the observed object Formula for calculating MAE: σୀଵ ȁݕ െ ݔ ȁ (1) ݊ Where yi is the predicted height value, and xi is the actual height of ith measurement, n is the number of test values ܧܣܯൌ Fig RE is used to measure the degree of relative error of the estimated value with a real value in the same observation object This measure is expressed in terms of percentages and no units The formula for RE is as follows: Location of Hanoi in Vietnam B Data To build 3D models of buildings in Hanoi, two indispensable information are buildings heights and buildings footprints While building footprints can be obtained through open data sources on the Internet, building heights is failed to gather because of non-existing sources Therefore, building heights need to be collected in Hanoi by field observations/measurements to develop estimation models For the Hanoi area, through the consideration of the feasibility of gathering attributes, we collected 194 building’s information for training and testing the predictor Each building will be described by four attributes as stated in Table 295 ͳ ȁݕ െ ݔ ȁ ܴ ܧൌ ͲͲͳ כΨ ݔ (2) ୀଵ With yi as the predicted value of the object, xi is the real value in the same prediction, n is the predicted number, i = 0, 1, 2, , n RMSE is a frequently used measure of the differences between values predicted by a model or an estimator and the values observed The formula of RMSE is as follows: 2018 5th NAFOSTED Conference on Information and Computer Science (NICS) ܴ ܧܵܯൌ ඨ σୀଵሺݕ Ԣ െ ݕ ሻଶ ݊ perimeter index are collected and calculate The number of building for each type is listed in Table II (3) Where yi’ is the predicted height value, and yi is the actual height of ith measurement, n is the number of test values D CityGML CityGML is an open data model standard with XMLbased format for the storage and exchange of virtual 3D city models, issued by Open Geospatial Consortium (OGC) and the ISO TC211 The aim of the development of CityGML is to reach a common definition of the basic entities, attributes, and relations of a 3D city model This is especially important with respect to the cost-effective sustainable maintenance of 3D city models, allowing the reuse of the same data in different application fields CityGML defines ways to describe most of the common 3D features and objects found in cities as 13 thematic classes (such as buildings, roads, rivers, bridges, vegetation and city furniture) and the relationships between them It also defines different standard levels of detail (LoDs) for the 3D objects, which allows the representation of objects for different applications and purposes, such as simulations, urban data mining, facility management, and thematic inquiries Data standardizing give us the effective way to exchange and extent the scale of 3D Geographic Information System.[6] E Cesium platform Cesium –an open-source JavaScript library for the worldclass 3D globes and maps -raised by Cesium Consortium community Cesium provides powerful functions such as Virtual Globe, libraries for visualization image layers, 3D models Especially, Cesium provides high-fidelity timedynamic simulation and 4-D visualization It can be run on different browsers such as Google Chrome, Firefox We used Cesium open specification – 3D tiles and Cesium platform for high-performance of visualization Illustration of buildings footprints in Hanoi The data need to be pre-processed to remove the wrong buildings footprints These wrong polygons will cause the high error in the predicting process Fig Fig Illustration of building footprints in Hanoi (imagery maps) III EXPERIMENT AND RESULTS Experiments will be following up with these steps as follows: (i) collecting input data, (ii) building a model to predict building height, (iii) constructing a 3D model of buildings with LoD1 details, and (iv) developing a 3D GIS on Cesium platform A Collecting input data The input data to build the system should be divided into the following main types of data: %XLOGLQJ IRRWSULQW: Building footprint was downloaded via the API of the website: wikimapia.org (see Fig 3, Fig 4) The data cover an area of 8423 buildings with 14 types of buildings classified according to their intended use (Table II) Fig maps) %XLOGLQJ KHLJKW DQG IRXU IHDWXUHV were collected from field surveys 30 random points represent for 30 buildings are selected for each type of house in 14 classes (Table II) However, after removing points represented for wrong building polygons, the data was left with 186 points and divided into regions as shown in Fig.5 For each building, four feature including usage, area, perimeter, and Normalized 296 Illustration of building footprints in Hanoi (cadastre 2018 5th NAFOSTED Conference on Information and Computer Science (NICS) of the building with the building footprints collected from the open sources After the training and predicting, the results is shown in “Fig and Fig 10” Overall accuracy are considered using three measures MAE, and RMSE, and RE (see Table III) It is acceptable for 3D building with the height error less than or equal meters The error in each class is different and it is shown in Table IV TABLE III Fig Field trip locations We collected data from residential areas which are the crowed population TABLE II 14 CLASSES OF BUILDINGS IN HANOI EQUIVALENT TO ITS ATTRIBUTE VALUE THIS WILL BE USED TO DEPLOY THE BUILDING HEIGHT PREDICTOR Attribute Value Class Measure MAE Unit Metter (m) Value 7.12 RMSE Metter (m) 9.26 RE Percent (%) 37.55 TABLE IV Number of sampling points OVERALL ERROR OF THE PREDICTOR EVALUATE THE ERROR OF THE PREDICTOR WITH EACH CLASS Class MAE (m) RMSE (m) Apartment building Bar 3.403 3.714 RE (%) 22.173 6.314 8.324 36.332 Building 22.153 29.211 53.771 Café 5.606 6.761 45.963 Church 11.566 14.217 70.731 Hospital 8.025 9.898 45.363 Hotel 8.317 12.896 34.245 House 2.426 3.005 20.890 Museum 4.777 7.025 25.526 Office 9.101 11.452 42.426 Restaurant 3.861 5.731 35.594 Apartment building Bar Building 32 Cafe 11 Church Hospital 12 Hotel 12 House 23 Museum 10 Office 12 11 Restaurant 11 School 4.259 26.040 School 10 3.805 12 13 Store 21 Store 3.448 4.191 31.514 14 University 19 University 6.868 8.975 35.160 B Deploying the height predictor model In this section, we will present a machine learning model to predict the building height using the Random Forest method based on the data collected in the field trip as discussed in section A The data used to train the model will have four attributes (see Tabel I) Where the attribute "Usage" will be positive integers from to 14 corresponding to the classification of the building during fieldtrip, while remaining attributes which are "Area", "Perimeter" and "NPI" will be real numbers The results of the predictor model will be tested and evaluated through the MAE (1), RE (2), and RMSE (3) parameters The usage of building is a categorical feature so we first use onehot encoding to turn this into binary vectors to have a better job in prediction In order to make the model more accurate, we had to adjust the parameters for the model using cross-validation and gridsearch technique The predictor model will use the training data as the data obtained after the field trip to predict the height Table IV shows that building with large footprint such as class Building or Church classes often cause high errors Large buildings are often very tall and with some cases like Building and Church, which are large but not too high, can cause very high error C 3D modeling The predicted heights will then be merged with footprints for the next step 3D modeling 2D cascade building footprint will be extruded to obtain 3D building models according to 3D Tiles specification 3D Tiles enable adaptive spatial subdivision in 3D, including k-d trees, quadtrees, octrees, grids, and other spatial data structures (Fig 6) Thus, it will reduce the cost of rendering each model based on the distribution of models and result in a balanced data structure 3D Tiles is also more flexible when the user zoom in or zoom out, the visible map tiles will replace with new higherresolution map or lower-resolution map, respectively 297 2018 5th NAFOSTED Conference on Information and Computer Science (NICS) The building will be modeled LoD1 according to the CityGML specification [7] which means the building will be represented as a simple block This approach will not aim to create 3D models with high level of details due to the error in the predicting process To create 3D building in LoD1, the footprint will be extruded to its predicted height The whole 3D modeling is processed using FME which is a powerful software for geo-spatial data processing and transformation Fig Illustration of proportion of building usage in Hanoi From the chart in Fig 9, the “Building” is the highest and most distributed building type in the city of Hanoi with the average height is around 40m to 45m In addition, the classes of Restaurant, Food stores, Bar, House, and Cafe have the lowest height in the city with the average height is around 13m An adaptive quadtree-like subdivision based on the distribution of buildings Fig The results of generating 3D buildings models for Hanoi is shown in Fig With the rendering mechanism of 3D Tiles, 8423 buildings have been visualized with LoD1 according to CityGML standard Fig Illustration of 3D buildings in Hanoi Most of the buildings in the city are apartments and offices or shops according to the pie chart as in Fig This may indicate that modern buildings are being built in Hanoi tends to increase, recently Fig Building heights distributions according to 14 classes of building in Hanoi Fig 10 shows the average building height in each district of Hanoi High buildings are mainly located in the suburbs (Ha Dong Thanh Tri, Hoang Mai, Thanh Xuan, Tu Liem) as recently new apartment buildings and new urban areas have grown rapidly in these areas The buildings in the inner city have a lower average height, however, there are some tall skyscrapers, but the residential areas in the inner city are predominantly indigenous They mostly live in the same house through different generations so the height of these houses is not too high (around 10-13 meters high) while the houses in old quarter which are very famous for traveling are only around – meters high With the access to these kinds of information and the visualization of 3D building models, decision makers or real estate investors can be able to predict and the price estimation 298 2018 5th NAFOSTED Conference on Information and Computer Science (NICS) ACKNOWLEDGMENT This work has been supported by Vietnam National University, Hanoi (VNU), under Project No QG.18.36 REFERENCES [1] [2] [3] Fig 10 Average building height in each district IV CONCLUSION In this study, we have addressed two main issues remaining in limitation of 3D GIS in Vietnam which are the lack of 2D cadaster data especially building height data, and the performance of rendering 3D objects at a large scale With the cadastral data in 2D containing the footprint information including geometric and building information attributes, we have applied machine learning method to deploy a predictor model to gather the height which is an essential information for 3D city modeling at LoD1 After that, a virtual city for Hanoi had built using 3D Tiles to faster rendering process However, there are some limitation remaining It is noticed errors because of limited number of building samples and those feature leads to high prediction errors In the future, we plan to collected more samples and features of buildings using LiDAR, high-resolution satellite, and aerial images to improve the prediction quality 299 [4] [5] [6] [7] D J B Bloomberg, Michael R, “BIM Gidelines,” New York City, Dep Des Constr., no July, pp 1–57, 2012 M Kada, “The 3D Berlin Project,” Photogramm week, pp 331–340, 2009 A P, “Development of virtual campus using GIS data and 3D GIS technologyௗ: A case study for Development of Virtual Campus Using 3D GIS Technologyௗ: a case study for Vietnam National University , Hanoi,” 2017, no December H D, “A 3d gis to design tour for tourists in west lake and surrounding area , hanoi capital , vietnam,” pp 3–11, 2011 F Biljecki, H Ledoux, and J Stoter, “Computers , Environment and Urban Systems Generating 3D city models without elevation data,” Comput Environ Urban Syst., vol 64, pp 1–18, 2017 E Standard, T H Kolbe, C Nagel, and E Standard, “Open Geospatial Consortium OGC City Geography Markup Language ( CityGML ) En- coding Standard,” 2012 F Biljecki, “Level of detail in 3D city models,” 2017 ... standard with XMLbased format for the storage and exchange of virtual 3D city models, issued by Open Geospatial Consortium (OGC) and the ISO TC211 The aim of the development of CityGML is to reach... of buildings Fig The results of generating 3D buildings models for Hanoi is shown in Fig With the rendering mechanism of 3D Tiles, 8423 buildings have been visualized with LoD1 according to CityGML... footprints often have complex shapes Currently, LiDAR, highresolution satellite imagery and aerial images are used largely to gather information for 3D city modelling However, all of these methods