UAV-BASED 3 D URBAN ENVIRONMENT MONITORING

Unmanned Aerial Vehicle (UAV) based remote sensing can be used to make three-dimensions (3D) mapping with great flexibility, besides the ability to provide high resolution images. In this paper we propose a quick-change detection method on UAV images by combining altitude from Digital Surface Model (DSM) and texture analysis from images. Cases of UAV images with and without georeferencing are both considered. Research results show that the accuracy of change detection can be enhanced with georeferencing procedure, and the accuracy and precision of change detection on UAV images which are collected both vertically and obliquely but without georeferencing also have a good performance. * Corresponding author


INTRODUCTION
Unmanned Aerial Vehicle (UAV) based remotes sensing system is becoming very useful.It can be used to generate highresolution remote sensing data, Digital Surface Model (DSM), Digital Terrain Model (DTM), Digital Elevation Model (DEM) (Sazak and Savran, 2016) and so on, so it is now widely used in urban monitoring, urban planning in smart city (Sutanta et al., 2016), and it also shows great advantages in construction monitoring, road surface and runoff evaluation, road network planning and landscape changing (Díaz-Vilariño et al., 2016).In urban environment monitoring, 3D change detection based on remote sensing data are needed to detect land use changes or human activities (Chaabouni-Chouayakh et al., 2011;Chaabouni-Chouayakh and Reinartz, 2011).
For UAV-based images georeferencing, generally there are two methods, and the first is known as direct georeferencing, which requires high quality sensors, up to centimeter global navigation satellite system (GNSS) receivers and accurate inertial measurement units (IMU) (Gabrlik, 2015), while the second is known as registration method, which requires Ground Control Points (GCPs), and, the 3D GCPs should be of high positioning accuracy, usually collected by Static GNSS or Real-Time Kinematic (RTK) GNSS.In practice, Hugenholtz et al. (2016) demonstrated that direct georeferencing from onboard Global Positioning System (GPS) without GCPs has lower position and DSM accuracy compared to that from a survey-grade GPS receiver.Similarly, Tan et al. (2016) evaluated high cut slope area from DSM data which is produced from UAV image with and without GCPs, and results show that DSM data without GCPs can be used to detect high cut slope changes in emergency cases.Qin et al. (2016) gave a fairly complete review on the recent development and applications of 3D change detection, who classified 3D change detection into two rationales: 1) only by geometric comparison; 2) geometric and spectral analysis, which includes the computation of both geometric and spectral information for change analysis.Huang et al. (2017) applied multi-view ZY-3 satellite imagery to precise urban change analysis in a multi-level (pixel, grid, and city block) approach, and Tian et al. ( 2014 In this paper, we will work on images collected from a fixedwing and a Quad-Rotor for 3D urban environment monitoring.The contributions of this work are threefold.First, to evaluate the UAV mapping accuracy with and without georeferenced data from Fixed-wing and Quad-Rotor UAVs.Second, to apply DSM to change detection.Third, to demonstrate DSM and texture-based change detection.

Study area
Two study areas are selected, and one is in Chongqing and another is in Guangzhou (Figure 1): (a) First study area locates in the agriculture field near Zhujiang river, Guangzhou, which covers around 0.03 km 2 with varieties of land use such as bare land, tree, road, and agriculture; (b) Second study area locate in the garden of south gate of Yangtze Normal University, Chongqing, and it covers around 0.2 km 2 with varieties of land use such as trees, roads, buildings, grass, pool, and bare land.

Data preparation
The Fixed-wing and Quad-Rotor UAVs used are shown in figure 2. The fixed-wing UAV has the advantages of high ceiling flight, long time fly, multiple sensor installation, high positioning precision and wide area data acquisition.On the other hand, the Quad-Rotor UAV (DJI Phantom4) has the advantage of medium price with extremely device and autonomous flight.The data quality depends on onboard sensor and GPS, and it can capture data in medium and small areas.Thus, this research will evaluate the performance of these UAV platforms for generating DSMs.The difference of onboard capabilities and technology was shown in table 1.         0, ; ( , ) 1, where Hth (i,j) = the range of threshold 0,1 of pixel (i,j) H = height different value x = the range of changed area Second step is to conduct texture-based change detection.Uniform Local Binary Pattern (LBP U2 ) is used to calculate texture difference.The LBP texture operator has original 3 x 3 neighbourhood which is threshold at the center of kernel.There are 8 neighbouring pixels and perform 8 comparisons.Each comparison will calculate two pixels between center and neighbour pixel.If the center pixel value is greater or equal to neighbouring pixel value, it will set to 1. Else if the center pixel value is less than the neighbouring pixel value, it will set to 0.
Then each comparison will iterate in the counter-clockwise direction and multiplied by binomial weights.All neighbouring comparison will be calculated in equation ( 2).The Uniform mapping produces 59 output labels.In processing, kernel will move around and save these results into LBP U2 image as RLBP, GLBP, BLBP.
1 , ( , ) 0 ( )2 where LBPP,B (i,j) = Uniform LBP value of pixel (i,j) S( ) = the different of invariant changes of the images P = the total number of involved neighbours B = the radius of the neighbourhood gc = the value of the center pixel gp = the value of its neighbours Texture based change detection is calculated from the pairwise distance between two LBP U2 images following equation (3).This equation calculates arccos from two LBP U2 images and the result value will plot on histogram.The texture changed will defined by the range of thresholds (Lth) value including texture changed, which has L(i,j) more than Lth (set to 1) and no change, which has L(i,j) less than Lth (set to 0).
where L(i,j) = the range of cosine distance from pixel (i,j) A1 = the pixel value of LBP image 1 A2 = the pixel value of LBP image 2 DSM and texture-based change detection will be defined changed area as three conditions: 1) Changed area, it means that the plus value of same positioning between height change value (Hth (i,j)) and LBP changed value (L(i,j)) equal to 2; 2) changed in one condition, it means that other plus values equal to 1.It will be defined as changed area with one condition; 3) No changed area, it means that the plus values equal to 0, that will be defined as no changed area.The equation is shown in equation ( 4) Changed area

H L Changed in one condition
No changed area where Hth (i,j) = height different thresholds of pixel (i,j) L (i,j) = LBP changed thresholds of pixel (i,j)

CONCLUSION AND DISCUSSION
In this research fixed-wing UAV and quad-rotor UAV are both used to collect images, and processing results show that ortho image and DSM georeferenced with GCPs of high position accuracy can reach enough horizontal mapping accuracy even for work of very high requirement in accuracy.As to the mapping accuracy comparation of different camera orientations, result show that vertical image produces ortho image and DSM with highest accuracy, and DSM produced from vertical and oblique images has a good surface fitting and 3D shape of object.However, the most important criteria for producing ortho image and DSM with high position accuracy is GCPs.
In case of DSM mapping without georeferencing for change detection, result shows that DSM change detection from vertical mapping has limitation on map orientation and 3D shape.
Although vertical mapping has high position accuracy, vertical combined oblique mapping can produce less position error both horizontally and vertically.The proposed change detection method by combining DSM and LBP-based texture proved its good performance.So, in cases where the requirement in position accuracy is not very high, vertical combined with oblique UAV mapping can be used to make change detection.
) evaluated accuracy of DSM produced from IKONOS stereo imagery and propose a change detection method on height changes and Kullback-Leibler divergence, which confirmed that DSM can be used to improve the accuracy of change detection.Thus, in this research we will use the height difference from geometric comparison for change detection.Texture analysis is one of the main feature in computer vision, and it has been applied to remote sensing image classification, change detection, and segmentation.Local Binary Pattern (LBP) introduced by Ojala et al. (2002), is proposed as a twolevel version of the texture unit, and it describes the local textural patterns and comparative texture measures with classification based on feature distributions.Pietikainen et al. (2011) also introduced Local Binary Patterns for still images.

Figure 1 .
Figure 1.Study areas (a) Agriculture field in Panyu Qu, Guangzhou (b) Yangzte Normal University, Fuling, Chongqing Figure 2. (a) AI Bird KC-1600 Sky Hawk UAV model and (b) DJI Phantom 4 UAV properties/ UAV types -wing UAV mapping: This experiment includes two flights, and in each flight, all images will be divided into two datasets.Moreover, each processing will generate two outputs.The processing flow is shown in figure3, and the results of ortho image and DSM data are showed in figure4.

Figure 5 .
Figure 5.The experiment design of Quad-Rotor UAV data preparation: Change detection computes the difference of two mappings.In this function, map positioning (X, Y) and height value (Z) of the two mappings should have surface fitting.To deal with no referenced data, this research applied image to image registration to fit both mapping in horizontal positioning.At the first stage, the firsttime mapping without georeferenced is taken as the reference image and second-time image without georeferencing as the image to be adjusted.The following step is to adjust vertical positioning by computing height difference of 30 reference points.The evaluation of vertical error is shown in table7.Then RMSEZ are used to adjust height difference to fit the surface mappings in vertical position.2.5.3 DSM change detection:Geometric comparison is made to calculate height difference of two co-registered DSMs, which is done in ENVI.Results by Fixed-wing UAV is shown in figure7, and we can see that the changed area includes a car, shade of tree, and water level in pool.

Figure 7 .
Figure 7. DSM change detection from fixed-wing UAV

Table 1 .
The different of UAV platforms, onboard capabilities and technology According to the purpose of analysis, for each study area there will be two flights to collect data.Flight planning was conducted in terms of image content with high overlap.The mission planning was set in following table 2.

Table 2
. Flight and mission planning of this experiment GCPs were marked and measured with a Trimble R8 GNSS RTK GPS, based on local CORS station, which provided high positioning accuracy in horizontal and vertical.Table3shows the GCP properties and data collected.

Table 3
. GCP properties and data collected for this experiment Pix4D mapper is used in this research to generate ortho mosaic, 3D point cloud, DSM, DEM, and DTM.These results can be exported for further manipulation.

Table 4
. Map accuracy of fix-wing UAV 2.4.2Quad-Rotor mapping: The accuracy assessment of ortho mapping and DSM data is with 30 GCPs, and measurement also follow the digital ortho imagery accuracy in ASPRS positional accuracy standard.Results of first-and second-time mappings show that horizontal accuracy has RMSEX and RMSEY over 1-pixels, which is recommended in standard mapping and GIS work, and vertical accuracy for digital elevation data has absolute accuracy of RMSEZ in nonvegetated at 95% confidence level.The average GSD is 3.40 cm.In contrast, both vertical combined oblique images of firstand second-time mapping have horizontal accuracy in RMSEX and RMSEY less than 1-pixel which is recommended in highest accuracy work, and vertical accuracy has absolute accuracy of RMSEZ in non-vegetation level.The average GSD is 4.6 cm.The RMSE X, Y, Z are shown in table 5.

Table 5 .
Map accuracy of Quad-rotor UAV

2.4.3 Accuracy of ortho mapping and DSM data compared with testing
On the other hand, vertical combined oblique image, both first-and second-time mapping have horizontal and vertical accuracy less than 1-pixels standard.
GCP:The ortho mapping and DSM data will be evaluated with 30 referencing points from RTK-GNSS.The result of absolute accuracy of horizontal and vertical are shown in table 6.In vertical image, both first-and second-time mapping have over 1-pixels standard level of horizontal and vertical accuracy, but it is less than 95% confidence level.

4 To evaluate the accuracy of ortho mapping and DSM data without georeferencing:
To evaluate the positioning accuracy without GCPs georeferencing, RMS error from 30 testing points are calculated.Position errors of every testing point are measured, and the overall accuracy is calculated and shown in table 7. The result shows that horizontal error of vertical image first-and second-time mappings have distortion error about 7.8063 and 7.7430 meters, and vertical combined oblique image of first-and second-time mappings have distortion about 4.9651 and 1.9026 meters.
distance between the marked points on mapping and the referenced points from RTK-GNSS which should has the same location.The image distortion was presented by RMS Error (the more different, the larger distortion).All the RMS Error of distortions are less than 1 meter, as shown in table 8.