OBJECT BASED BUILDING EXTRACTION AND BUILDING PERIOD ESTIMATION FROM UNMANNED AERIAL VEHICLE DATA

The aim of this study is to examine whether it is possible to estimate the building periods with respect to the building heights in the urban scale seismic performance assessment studies by using the building height retrieved from the unmanned aerial vehicle (UAV) data. For this purpose, a small area, which includes eight residential reinforced concrete buildings, was selected in Eskisehir (Turkey) city center. In this paper, the possibilities of obtaining the building heights that are used in the estimation of building periods from UAV based data, have been investigated. The investigations were carried out in 3 stages; (i) Building boundary extraction with Object Based Image Analysis (OBIA), (ii) height calculation for buildings of interest from nDSM and accuracy assessment with the terrestrial survey. (iii) Estimation of building period using height information. The average difference between the periods estimated according to the heights obtained from field measurements and from the UAV data is 2.86% and the maximum difference is 13.2%. Results of this study have shown that the building heights retrieved from the UAV data can be used in the building period estimation in the urban scale vulnerability assessments.


INTRODUCTION
The determination of the fundamental vibration period of a building is essential for both earthquake design and seismic vulnerability assessment.The building period is dependent on the mass and the stiffness of the building (Salameh et al. 2016) and it can be computed by eigenvalue or pushover analyses, which are conducted to a numerical model of the building.However, when conducting seismic vulnerability assessment studies in urban scale, it is necessary to estimate the seismic performance of hundreds or thousands of buildings (Crowley and Pinho 2004).Obviously, carrying out computer analysis to determine the vibration period of each building, present in the considered urban area is not feasible.Therefore, simplified equations to estimate the building period with respect to building height can be used instead.In many seismic design codes, there are empirical and semiempirical equations, which can be used to estimate the building period according to building height (Goel and Chopra 1997), which is a key factor for the building period estimation.Measuring the height of each building in large building stocks with traditional field survey methods can be very time-consuming.Therefore, it is necessary to use innovative technologies like remote sensing applications for determining building height (Masi and Vona 2009).
Remote sensing offers many opportunities to obtain information on urban areas such as monitoring urban development and detection land use and land cover changes (Li et al. 2014) Building extraction or detection is one of the important topics for urban area studies.Many studies have been carried out to extract buildings using different data types that obtained from passive and active remote sensing sensors.Optical satellite imagery, aerial images, RADAR imagery, LIDAR data, Digital Surface Model (DSM) and Digital Terrain Model (DTM) are mostly used data types in this kind of studies (Haala and Brenner 1999;Li et al. 2014;Rottensteiner and Briese 2002;Sohn and Dowman 2007).Different methods are used for building extraction; Object-based Image Analysis (OBIA), which allows using spectral, contextual, textural and geometric properties of image objects, is one of the preferred methods, especially when working with high-resolution satellite images (Li et al. 2014;Liu et al. 2005;Shrivastava and Rai 2015).
Building height value is an important parameter for 3D urban modeling (Shiode 2000), post-disaster assessment (Brunner et al. 2010) and building period estimation (Salama 2015).Several different methods have been applied for estimation of building height information from remote sensed images or Digital Elevation Model (DEM).One of the used methods is the shadow-based building height estimation using satellite images (Shao et al. 2011;Shettigara and Sumerling 1998).Digital Height Model (also known as normalized DSM (nDSM)) which is the difference between DSM and DTM is one of the used methods for building height extraction (Haala and Brenner 1999;Yu et al. 2010).
Last decades, one of the new remote sensing platform -Unmanned Aerial Vehicle (UAV), gives valuable data source for mapping, 3D modeling, and land use/land cover studies (Nex and Remondino 2014).It is possible to obtain number of informations for urban areas using DSM, DTM and Orthomosaic data generated from UAV images.Building boundary, building height, roads, trees which belong to urban area can be extracted using these data.
In this paper, the possibilities of obtaining the building heights, which are used in estimation of building periods from UAV based data, have been investigated.The investigations were carried out in 3 stages; (i) Building boundary extraction with OBIA, (ii) height calculation for buildings of interest from nDSM and accuracy assessment with terrestrial survey.(iii) Estimation of building period using height information.

Study Area
As a study area in this paper, a small area that includes residential reinforced concrete buildings in the city center of Eskisehir (Turkey) was selected.The study area contains buildings, trees, roads, green area and cars.The buildings have different height values and different dwelling types such as adjoined or detached orders (Figure 1).

Specifications of UAV System
The aerial image collection for the study area was performed with fixed wing Sensefly eBee UAV systems.The UAV contains 16 megapixel RGB digital camera and it has been controlled with Emotion -2 flight management software that used for flight planning, simulation, flight control and preparation of images for post-processing.The UAV cruise speed changes between 40 km/h to 90 km/h and it has up to 45 km/h wind resistance.The UAV system is 0.7 kg with 96 cm wing span.Its flight time capability is nearly 50 minutes for single flight.

Aerial Image Collection
The aerial images for the study area were collected in single flight.The flight was planned with 80% lateral overlap and 75% longitudinal overlap with 5 cm ground sample distance (GSD).Images were taken nearly 160 meters altitude above the ground.Totally 146 images were taken.Before the flight 7 ground control points (GCP) were located in the study area.GCP coordinates were measured with RTK-GNSS systems.

Data Processing
The data processing in this paper, was performed using Pix4D Desktop software.As a result, a 5 cm resolution orthomosaic and DSM of the study area was produced.Also, 25 cm DTM was generated for the study area.Figure 1 shows the orthomosaic images of study area.

METHODOLOGY
The methodology applied in this study consists of three stages; (i) Object based building extraction, (ii) building height extraction, and (iii) building period estimation.Figure 2 indicates the main steps and sub-steps of applied methodology.

Object Based Building Extraction
Pixels in a high-resolution data have a heterogeneous structure, thus, when they are used in classification many meaningless pixels are generated into the class.Object Based Image Analysis (OBIA), which is consist of image segmentation and classification of image objects, gives more advantageous solutions for analysis of high-resolution images.In OBIA, image pixels are meaningfully grouped into homogeneous images objects and noise is consequently reduced (Lu et al. 2011).The UAV data used in this study has ultra-high spatial resolution.Therefore, OBIA was preferred method for buildings extraction.
Figure 2. Applied methodology in this study 3.1.1Image Segmentation: Image segmentation, which is used to generate homogenous image objects, is the first step of object based images analysis.There are several methods for remote sensing applications (Dey et al. 2010).In this study, Multiresolution segmentation (MRS) that is one of most used method in high-resolution image analysis was used for generating homogenous image objects.For obtaining optimum image object size in MRS, several parameters should be determined by users; scale, shape, compactness and layer weight.In this paper, parameters were determined with trial -error methods.

Image Object Classification:
Classification is the second step of OBIA.In this step, obtained image objects in segmentation phase attend to relate to classes (Lang 2008).This process can be performed using classification methods such as machine learning algorithms or rule based classification, which has been used for extraction of building polygons in image objects in this study.
Before the classification of the image objects, object metrics which were used at rulesets developing were calculated.For eliminating the ground objects such as cars, roads, gardens, etc., nDSM was used.
Where, B is Brightness value, which is calculated as the sum of objects means in visible band (c i(vis) ) divided by number of the corresponding bands (n (vis) ).EGI = 2* Green -Red -Blue (2) Where, Green, Red and Blue are spectral bands of the orthomosaic.
For extracted the buildings from the images objects, Red Blue Spectral Index (RBSI) was used (Eq.3).This spectral index is more useful for our study area, because nearly all buildings have brown coloured roofs, which give high spectral reflectance in Red band and low spectral reflectance in Blue band.

RBSI=
Red-Blue Red+ Blue (3) After segmentation of the layers, some image objects which have high slope value between ground and buildings were generated.Elimination of these kind of objects was done using standard deviation of slope which was obtained from DSM.

Building Height Extraction
There are two dwelling types in the selected urban area.One of them is adjoined dwelling type in which the buildings are attached to each other.The other one is detached type in which the buildings are separately located in the area.In the adjoined buildings, the buildings' vibrations influence each other and it is very hard to distinguish the fundamental vibration period of each building.Therefore, detached buildings were chosen for the building height extraction and the building period estimation in this study.
Building heights were obtained from nDSM data (Eq.4) (Haala and Brenner 1999;Yu et al. 2010).Extracted buildings polygons using OBIA include all area of building roof.It is possible to get an average building height from nDSM using these polygons.However, the roof height (h) is not considered for building period estimation.The height which should be considered for building period estimation is the building height (H), which is measured from the ground to the beginning of the roof (figure 3 (a)).Therefore, the height of the roof edges was obtained to find the approximate height values of the building body.For this process, firstly, building polygons features were converted to line and then line simplification was done.After that, points objects were created for vertex of the line.Height values for all points were extracted from nDSM data.Finally, building heights which were used to estimate building periods were calculated as the average of the heights of the points forming the building (Eq.5).
In Figure 3, H value stands for the building body height.h value represents the roof height.Black dots represent the intersection of the building boundary lines.
Where, H avg is the average of the building body height which is calculated as the sum of the points heights (H) forming the building divided by a number of corresponding points (n).

Accuracy Assessment:
As an accuracy assessment, the obtained height values with the applied method in this study were compared with terrestrial survey results.Eight buildings heights were extracted and these building heights were measured with laser distance meter at the field.

Building Period Estimation
A number of simplified equations that relate the height of a building to its fundamental period of vibration are available in many seismic design codes.Most of these equations have been developed for the force-based design of a new building.However, the aim of this study is to assess the seismic performance of existing buildings in the study area.The building period, which should be used in the seismic performance assessment, is the yield period.In this study, the equation (Eq.6) proposed by Crowley and Pinho (Crowley and Pinho 2004) has been used for the estimation of the yield period of a building.
T yield =0,1 H (6) Whereas, T yield , is the yield period of the building in seconds (s).H, is the building height in meters (m).Using only the RGB bands for the segmentation, caused some building roof misleading from the ground objects which have similar spectral features on rooftops.Therefore, DSM layer was used to prevent this problem.After several trial -error steps, optimum parameters for segmentation were determined as scale 50, shape 0.3, compactness 0.5.Also equal layer weight was given all input data.Since the UAV data is ultra -high resolution data, it is not possible to get buildings with one objects.Hence, one building was segmented more than one image objects, named as building candidates (Figure 4).After segmentation steps, building candidates were merged to get final shape of buildings.Before the calculation of nDSM, the DTM data was resampled at the resolution of DSM.Image objects with nDSM value ≤ 5 meter were classified as grounds objects, which includes roads, cars, gardens and small trees, and were eliminated from the building extraction.Secondly, big trees were separated from remaining images objects using mean value of Brightness and EGI parameters.Image objects with mean brightness value ≤ 130 and mean EGI value ≥ 0 were classified as big trees objects.Thirdly, images objects that belongs the buildings extracted using mean value of RBSI.image objects with mean RBSI value ≥ 0.36 were extracted as building candidates.These buildings objects include several image objects, which are between building and ground, and had higher slope values.For elimination of these objects standard deviation of slope value was used.Image objects with standard deviation of slope value ≥ 200 were separated from the building objects candidates.Finally, image objects that were classified as building candidates were merged into single image object in order to extract the final border of the buildings.The obtained building objects were then exported as polygons features.Final buildings that were extracted with OBIA are shown in Figure 5.In this study, building period estimation was made only for detached buildings.Those buildings are shown with red colour in figure 5.

Building Height Extraction
Building extraction process was done only for the detached buildings using ArcGIS Desktop 10.3 software.Firstly, building polygon features were converted to line features (Figure 6a).Secondly, line simplification was performed.
Because the UAV data has 5 cm meter resolution, building boundaries, which were extracted with OBIA, has many edges.Line simplification is required to get the actual building shape.In this study, building boundary line was simplified using edge distance ≤ 0.4 meter value (Figure 6b).Thirdly, points were generated for the line vertexes.Finally, points heights were extracted from nDSM and average building heights were calculated for every detached building.1. Obtained and measured building heights and their differences.

Building Period Estimation
In this part of the study, the height data retrieved from object based building height extraction were used to estimate the building periods (T yield -UAV).Additionally, the building periods were also estimated with respect to the heights from obtained from the field measurements (T yield ). Table 2 shows the period results and the differences between these two periods.
Building Number  As it is shown in Table 2, the maximum difference between the periods estimated according to the heights obtained from field measurements and from the UAV data is 13.2% and the average difference is 2.86%.These values can be considered as reasonable in building period estimation for seismic performance assessment of existing buildings.

CONCLUSION
The aim of this study was to estimate the buildings vibration periods using building height data obtained from UAV.The results show that the period values, which were estimated according to the heights, retrieved from the UAV data and from the field measurements are highly correlated.
Calculating the building heights from the UAV data is more practical than measuring the building heights in the field.
Results of this study have shown that the building heights obtained from the UAV data can be used for the building period estimation in the urban scale vulnerability assessment studies.For future study, the used methodology is planned to be applied to a bigger study area.

Figure 1 :
Figure 1: Orthomosaic images of the study area The segmentation and building extraction with OBIA were performed in eCognition Developer 9.0 software.During MRS steps Red (R), Green (G), Blue (B) band of orthomosaic and DSM data were used as input layers.

Figure 4 .
Figure 4. Created images objects with the MRS

Figure 6 .
Figure 6.Building height extraction steps from nDSM

Table 1
shows the building heights difference between applied methodology and field measurement results.According to results, the maximum height difference occurred in building 2. When we examine the building 2, three sides of the building are surrounded by trees.The forth side is more shadowed area, thus in cases like this difference in the results can be expected.Since aerial images that are taken from UAV are optical images, the presence of shadowed area and more vegetation cover areas, does not allow generation of highly accurate DTM.Due to the lack of low ground point generation during dense point cloud production.

Table 2 .
Estimated periods and their differences