Point cloud filtering on UAV based point cloud
Mustafa Zeybek
a,b,⇑, _Ismail S
ßanlıog˘lu
b,ca
Artvin Coruh University, Engineering Faculty, Geomatics Engineering, Artvin, Turkey
b
Selçuk University, Engineering Faculty, Department of Surveying Engineering, Turkey
c
Konya Technical University, Engineering and Earth Science Faculty, Geomatics Engineering, Konya, Turkey
a r t i c l e i n f o
Article history: Received 5 October 2017
Received in revised form 1 October 2018 Accepted 5 October 2018
Available online 7 October 2018 Keywords: UAV Point cloud Filtering Bare earth Extraction
a b s t r a c t
Nowadays, Unmanned Aerial Vehicles (UAVs) have been attracted wide attentions such as a new mea-surement equipment and mapping, which are capable of the high-resolution point cloud data collection. In addition, a massive point cloud data has brought about the data filtering and irregular data organiza-tion for the generaorganiza-tion of digital terrain models. Filtering of point clouds contains vegetaorganiza-tions and artifi-cial objects play a cruartifi-cial role for bare earth terrain modelling. Topographical maps rely on the data structures which are built on bare ground terrain points. The bare earth surface extraction is not the only crucial to the topographical maps but also decision-making processes such as natural hazards manage-ment, deformation analysis and interpretation.
In order to filter a UAV-based 3D raw point cloud data, in this paper, filtering performance of four dif-ferent algorithms using open source and commercial software’s have been investigated, (1) curvature based (Multiscale Curvature Classification-MCC), (2) surface-based filtering (FUSION), (3) progressive TIN based (LasTool-LasGround module-commercial) and (4) physical simulation processing (Cloth Simulation Filtering-CSF). The applied filtering results were validated with the reference data set classi-fied by operator. Although different filtering methodologies implemented on point clouds, these methods demonstrated similar results to extract ground on distinctive terrain feature such as dense vegetated, flat surface, rough and complex landscapes. The filtering algorithms’ results revealed that UAV-generated data suitable for extraction of bare earth surface feature on the different type of a terrain. Accuracy of the filtered point cloud reached the 93% true classification on flat surfaces from CSF filtering method.
Ó 2018 Elsevier Ltd. All rights reserved.
1. Introduction
Over the past decade, Unmanned Aerial Vehicle (UAV) plat-forms have been using for various purposes with significant devel-opments of communication systems, autopilot systems, geospatial global navigational satellite systems (GPS/GNSS). The UAV sys-tem’s great potential for various mapping applications which grows rapidly[1]. Using a UAV to obtain spatial data is highly ben-eficial for the users in terms of time, cost and accuracy[2].
The UAV systems are preferred as its low cost and a small num-ber of requirements for operation to take-off and landing sites. Fur-thermore, UAV platforms payload capacity have been increased due to technological improvements on sensors weights. Thus, dif-ferent sensors can integrate to the platform such as camera and light detection and ranging (LiDAR). The autonomous flying UAV
platforms which is able to determine flight direction and surveying area from web based map servers and continuously communicated with GNSS receivers on the platform that made possible to capture image data[3]. Briefly, UAV and remote sensors have giving great contribution to a survey grade mapping for users and research community. UAV’s are used wide range of applications such as nat-ural disasters, topographic mapping, volumetric calculations, building and highway engineering, precision agriculture and for-estry [4–11]. Some work demonstrated the potential of UAVs or the other call Unmanned Aerial System (UAS) platforms and pho-togrammetry as a usable technology for surveying fluvial sub-merged topography areas at the mesoscale[12].
UAV platforms are increasingly used to generate of high-resolution maps for geosciences studies which require unreachable areas instead of time-consuming ground-based traditional mea-surement techniques [13,14]. Also, various investigations were completed for generation point cloud such as effect of different ground height and camera angles on accuracy of orthomosaics [15]. Precise applications of UAVs such as in volume calculations have been investigated in previous studies[16–18]. In these
stud-https://doi.org/10.1016/j.measurement.2018.10.013 0263-2241/Ó 2018 Elsevier Ltd. All rights reserved.
⇑ Corresponding author at: Artvin Coruh University, Engineering Faculty, Geo-matics Engineering, Artvin, Turkey.
E-mail addresses:mzeybek@artvin.edu.tr(M. Zeybek),sanlioglu@selcuk.edu.tr (_I. Sßanlıog˘lu).
Contents lists available atScienceDirect
Measurement
ies, 3D point clouds have been produced through photogrammetri-cally work flow or also called as Structure from Motion (SfM) and Multiview-Stereo (MVS) techniques[19]. SfM includes basic tradi-tional photogrammetry techniques. However, it does not require any regular or parallel captured images to process. Because, these techniques automatically extract and matches image features. Fur-thermore, it implements the bundle block adjustment procedures [20]. In this study, SfM algorithm has been implemented to gener-ate millions of points clouds with images taken from the UAV platform.
Topographical maps and DTM productions are well-suited high resolution point cloud data to obtain various sensor and tech-niques [21]. LiDAR data filtering is challenging task for earth science research and being focused on many cases in the literature [22,23]. Mostly used point cloud filtering algorithms are applied into the LiDAR data processing software’s. Recent studies focused on UAV point cloud data filtering methodology [24–27]. LiDAR has been used by different disciplines efficiently. However, its high cost makes the technique hard to reach. For this reason, on small scale studies cannot be practical with LiDAR technology owing to their cost. Instead of airborne LiDAR system, UAV technology takes a place to generate high density of point cloud for many applica-tions. The real advantages of the UAV system that does not require a big aircraft, experienced workers and airport. Moreover, it can be ready to use at the study area in a short period of time when the air condition is suitable for flight.
Bare earth surface extraction is a challenging process for point clouds. Most of the surveying, mapping and topographical studies do not interest in the artificial objects and vegetation, apart from other disciplines and studies such as cadastral survey, base map-ping and Geographical Information Science (GIS) applications. Accordingly, filtering of objects such as trees, shrubs, buildings, vehicles, etc. is an important issue. Thus, the main purpose of this study is segmentation of point clouds as a ground and non-ground point cloud.
Currently many algorithms have been developed for filtering point cloud that acquired from LiDAR technique[26–30]. Ground filtering studies can be categorized into several class[31]. These algorithms do not work perfectly but they produce filtering desir-able level at over 90%. Axelsson[22]proposed progressive triangu-lation irregular network filtering algorithm. A lowest seed points based sparse TIN is created and densified in an iterative process. The TIN adapts each point in point cloud iteratively provided threshold below of curvature parameters.
Mathematical morphology approaches for filtering point clouds also developed with basic process that include opening, closing, dilatation and erosion based on grid operators[32,33]. The selec-tion of kernel sizes is the significant parameter for nonground point detection with different sizes[34].
According to Vosselman [35] slope based filtering mainly detects observations that a large height difference between two neighbour points is reason to high steep slope which is not possible in terrain conditions. Probably, low altitude point is a ground and the high point is non-ground point.
Point cloud filtering and determining the bare earth surface are crucial steps to generation of DTM. Active LiDAR or passive SfM algorithms can generate millions of data in a very short time. In active based methods, point cloud filtering procedure is relatively easier than passive systems. Because of the multi–echo informa-tion can be separated from different material reflecinforma-tion, it can also penetrate from vegetation in heavy forestry areas and able to get wide scanning angle to get measurements from ground surface. In general, the LiDAR data contain relatively a precise point cloud with small size outlier as compared to UAV point cloud. Here, pas-sive system generated point clouds were classified based on the geometric constraints.
Full waveform and multi echo specifications shows the advan-tage of active sensors and great contribution for filtering and seg-mentation process [28]. Among the filtering methods, some of them uses the integration of multiple specification not only geom-etry but also laser intensity. This neat solution could grow the abil-ity of to achieve an effective classification in the regions where a geometric specification of points does not provide sufficient seg-mentation. Image processing techniques have been also imple-mented onto 3D point clouds for filtering [36]. Surface-based algorithm with linear least square interpolation and surface gener-ation then stochastic characteristic determine of the class of points with this surface[37].
Filtering procedure include two basic errors. First is commission error which classifies non-ground points as ground points. The sec-ond is omission error that rejects ground points. The accuracy met-rics of the filtering are an issue to determine these error rates[38]. The purpose of this study is the investigation of point clouds fil-tering and extraction of bare earth surface from UAV derived point clouds by using open source software and a commercial software. In this study, different filtering algorithms for UAV based point clouds have been used for segmentation as ground and non-ground point cloud. The paper consists of sections as data acquisi-tion system, data processing, point clouds filtering and the results of the methods.
1.1. Study area
UAV flights were held in Konya Tasßkent province. This site was chosen since it is a landslide monitoring research area which involves different types of land cover, high-density vegetation, low vegetation, shrubs and different geomorphological features. The study area situated at the Middle of Taurus Mountains (Fig. 1). The region altitude has averagely 1400 a.s.l. The landslide size is 1800 m length and 150 m width which covers approxi-mately 50 ha area. The study terrain covers rock-slope, flat surface, forestry area and also includes urban area. Flight operations were performed for landslide mapping inventory and monitoring annu-ally in forestry region with the help of high resolution orthomo-saics and point cloud analyses. Thus, implemented algorithms focused on generally to forestry area for extraction of bare earth surface.
2. Material and methods 2.1. Data acquisition
In this study, two types of field data were obtained. Images have been captured by a UAV platform. UAV data acquisition following steps were given inFig. 2. The flight tracks were programmed in open source Mission Planner[39]and sent into the UAV Pixhawk autopilot.
Study area terrain has a high roughness with vary on height of range approximately 150 m. Thus, flights have been divided into three single flight mission. At the post-processing steps of divided flights were merged, aligned together and optimized for discrep-ancy part on height from different image scale errors.
Ground control points (GCPs) have been surveyed by differen-tial GNSS technique on benchmark. Field studies contain three dif-ferent steps. Those are mission planning pre-flight, flight operation and GCPs measurements. Here, GCPs benchmarks have to be painted on the ground before flight missions. These artificial tar-gets are large enough to be detectable in an aerial photograph. These benchmark points must be sharp and well defined, and lie in reachable locations on field. The navigation systems of UAV for take-off, landing and flight directions which have been
man-aged based on the UAV Inertial Navigation System (INS), GNSS, INS/ GPS combined system. The spatial coordinate system of the UAV is recorded on the position of designated GNSS antenna centres of platform in the UTM-WGS84 (World Geodetic System 1984) refer-ence frame. GNSS and an autopilot system registration calculate when the UAV stands on the ground for commands from ground control station (GCS) or aerial position of the platform on air at any time. For the tracking of platform on the air, GCS is mandatory.
GCS communicate with UAV for which part of the terrain remain to scan according to the flight plan. During the flight tracking, GCS beneficial monitoring continuously of platform without any human supervision by autopilot system correction. Flight planning requires to be specific for each flight operation. For the photogram-metric purposes and solutions, overlaps should be determined for-ward 80% and side overlap is 60%. These overlaps might come larger than the traditional airborne mapping system[40]. For hilly terrain conditions, where there are large fluctuations of height above and below the survey datum (average terrain height) a 60% of overlap for the datum could reduce the overlap to only 50% or 40% for the steep areas on top of mountains[40]. The main reason of larger values is the size and the weight of platforms which can easily affected from wind or fast manoeuvre of the plat-form. Small overlap percentages can cause the gaps between image blocks. The large overlap values prevent existence of gaps and increase the density of point cloud.
Different types of UAV platforms should select for different mis-sions. In this study, UAV platform have been selected TM-Geo2 quadcopter which is designed for mapping purposes. Sony Nex5 (24.3 megapixel) model consumer-grade digital RGB camera and U-blox low cost GNSS receiver integrated to the platform. The cam-era is installed below the body on central gravity of the platform. The captured images were 6000 pixels by 4000 pixels in size on 23.5–15.6 sensor size and image footprint on the ground was approximately 118 m x 78 m (averagely 100 m flying altitude). Due to long operations, UAV platform has a light designed platform of a 4-rotor UAV with a total weight of under 4 kg (Fig. 3). Detailed
Fig. 1. Location of study area.
Fig. 2. UAV data acquisition steps.
specification of TM-Geo2 series can be view onTable 1( https://ge-omaticsgroup.com/insansiz-hava-araci/). The quadcopter platform is electric powered with a li-po battery, multi-rotor system flight is approximately 55 min, which is enough to cover for a100 ha area with 10 m/sn speed.
GCPs were measured for higher accuracy by GNSS-RTK method pre-flight operation. ITRF-96 datum and 2005.00 reference epoch has been chosen as a reference frame. Projection system is based on the Transverse Mercator projection with 3° grid zone and cen-tral meridian 33°. GRS80 ellipsoidal heights were used for vertical component. Since the field area is landslide monitoring area thus GCPs distribution determined as pattern inside and outside of the active landslide region. The standard precise differential position-ing is used as one reference receiver located at a base station whose coordinates are known at the stable area and the rover receivers’ coordinates determined relative to this reference recei-ver. This is the main principle of RTK for short range of differential GNSS techniques[41].
In addition, rover points were distributed around the stable and active landslide area in high roughness, flat or forestry areas. RTK method provides simultaneous coordinate, thus further post-processing does not require. RTK accuracy metrics reach 5 cm + 5 ppm[42]. The significant limitation of RTK is about the distance between reference and the rover receiver due to distance depen-dent biases from satellite based and atmospheric effects which effects the signals during travel in the air. However, this issue occurs is the distance above 10 km. In this study we have longest distance 1.8 km from top to bottom of the area. The GCPs were located and surveyed in a randomly distribution in the study area which includes topographic variations.
2.2. Data processing
The collected aerial images were processed for the further anal-ysis to generate of point clouds. For this purpose, image processing techniques are performed with automatically algorithms. Com-puter vision techniques were carried out such as SIFT, SURF algo-rithms on which pictures were taken with sufficient overlap images to align matched parts of them. State of the art software procedures are possible for processing UAV images.
After, flight log data from autopilot system and image data from camera SD-disk saved to a computer. Log data and images imported to the software for processing. In this study, the digital camera calibration parameters were used from Pix4D library and optimized with the field based GCPs in the software. The georefer-encing step is crucial for the model data to be rotated, translated and scaled to geodetic coordinates. The NEU (North-East-Up) posi-tions of the GCPs were marked in Pix4D at least two images and
adjusted with least-squares sense in order to convert image coor-dinates to geodetic coorcoor-dinates with similarity transformation.
Basically, the workflow of the SfM processing steps are similar to the traditional photogrammetry. However, SfM uses more robust computer vision algorithms and user-friendly interfaces. General workflow of SfM based software includes 3 main steps as follows;
1. Image alignment (SIFT, SURF)
2. Dense point cloud and mesh model generation 3. DTM and Orthomosaicing
In this study, surface reconstruction pipelines were employed with Pix4D. Software uses a common SfM–MVS progression. SIFT key points detected automatically each characteristic point on the image and extracted them. Afterwards, using cross-correlation analysis of at least three overlapping images, the sparse key points of each image are aligned automatically to each other [13]. When two key points on stereo pair images are found correl-ative, they are matched for generation of sparse point cloud. Besides, image order is crucial to searching procedure of homolog points at least three images for matching. Thus, rough image cen-tre’s coordinates have considerably reduced the image alignment processing time cost.
Before image alignment procedure, flight log data is used for stamping of image shooting time using onboard GNSS time were imported for processing. The coordinate information added to the exchangeable image file format (EXIF) header which GNSS obtained coordinates of the camera for using image alignment step. These GNSS coordinates have relatively low accuracy (1– 10 m). However, they are sufficient for navigation of the UAV.
While sparse point cloud is generated, the image absolute ori-entations and positions were calculated. Image orientation estima-tion is based on the matched point set between image pairs. However, this orientation is not determined perfectly due to the image distortions.
Here, GCPs increase the absolute positioning accuracy and transform the image coordinates to reference frame. In this proce-dure, GCP coordinates are employed to optimize the calibration parameters in order to minimize radial and tangential distortions. Subsequently, tie points are automatically detected on the mul-tiple images to get a dataset of sparse 3D cloud points with RGB colour information by correlating neighbour images. The SfM pro-cessing performed dense and accurate point clouds that followed the theoretical principles of photogrammetry with the default parameter of Pix4D. Different dense point cloud generation meth-ods can be used for UAV applications. It is often signal based matching methods (least squares matching, cross correlation)
Table 1
Technical specifications of the UAV platform. UAV Body
Multicopter Model TM-Geo V2
Platform/Motor Quad
Time Endurance 55 min
Working Temperature 15 °C /+45 °C
Remote Control Distance 20 km (Max) / 15 km (Safe)
Telemetry and Real Time Video Transmitter 10 km
Single Flight Cover Area 100–400 ha (Depend on flying altitude and wind conditions)
Camera Specifications
Stabilizer Gimbal Camera Gimbal
Camera Sony Alpha Nex5
Resolution 6000x4000
Focal Length 16 mm
[43]. Pix4D default values achieved sufficient dense point clouds for the study site.
The dense matched points which have their 3-D coordinate can be used to create a DSM with interpolated form of triangulate irregular network (TIN) which is used to project every image pixel, thus allowing a georeferenced orthomosaic to be generated [44,45].
In this paper, GCPs were used for accuracy assessment; due to the single way to check accuracy of the produced models. Accuracy assessment of the GCPs were done by Root Mean Square Error (RMSE). RMSEX¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pn i¼1 ½XMXGPS n s RMSEY¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pn i¼1 ½YMYGPS n s RMSEZ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pn i¼1 ½ZMZGPS n s ð1Þ where:
X, Y, ZMis the coordinate measured on orthomosaics and DSM. X, Y, ZGPS: coordinate of the GCP measured with GPS.
n: total number of GCP used in comparison.
In this study UAV captured 688 images and 118 GCPs were pro-cessed with multi-view stereo image matching algorithm which was used to generate dense point cloud data by Pix4D commercial software. The generated data was approximately over one hundred twenty million points. According to the terrain feature, average Ground Sampling Distance (GSD) for DEM and orthomosaics were acquired 5, 6, 7 cm levels in three different flight. DSMs and ortho-photo mosaics were derived by interpolation from point cloud. For further process and processing time reduction, produced dense (170 points per m2, 5–7 cm point nearest distances) point cloud data down sampled by spaces minimum 0.10 m. Down sampling process reduced the total point cloud to ninety million, hence 25%-point data removed from raw point cloud.
UAV point clouds contain outliers due to the various error source. In order to improve the reliability of the point cloud before filtering or any further process, an outlier removal is an important step. In order to acquire the smooth earth surface, these outliers were automatically removed by statistical outlier removal (SOR). Firstly, it calculates mean distance of each point to its neighbours with considering of k nearest neighbours for each neighbour. Then it starts to remove points that are further than the mean distance plus a number of times the standard deviation[46]. Outliers were removed with statistical outlier removal tool according to neigh-bour points. Point neighneigh-bour parameter (k) determined as 50 points, alpha parameter set to 3. As a results of outlier removal pro-cess, 2%-point cloud were removed from raw point cloud.
In order to an improvement of the filtering performance and a reduction amount of time, point clouds have tiled by 500 m 500 m to 15 single part. Therefore, filtering process have been accomplished by batch processing.
3. Filtering of point clouds
In this section, we present investigation of different filtering methods for segmentation of ground and non-ground (object) points from UAV based point cloud.
Four filtering methods were conducted to validate the perfor-mance of these methods: in this study, implemented filtering algo-rithms for point clouds can be seen below,
1. Multiscale Curvature Classification (MCC), 2. Surface-based filtering
3. LasTool-LasGround module, Progressive TIN algorithm 4. Cloth Simulation Filtering (CSF)
MCC filtering algorithm procedure, firstly determines the verti-cal component (Z). Thin plate spline (TPS) interpolation algorithm generates surface model with scale parameter. The nearest neigh-bours are used to fitting variable windows. Raster vectors deter-mined as x(s), the curvature tolerance value added to x(s), conditional process starts and points determined either ground point or non-ground point. General steps are as follow,
TPS interpolation step, on Z component of points. User determi-nes the scale parameterk and the curvature tolerance t. 3 3 mean kernel pass over the interpolated surface and a
vec-tor x(s) are declared.
Curvature in scale domain l is calculated by,
c¼ xðsÞ þ t ð2Þ
where x(s) is mean elevation vector and t curvature is the tolerance parameter.
Point class determination depends on below condition,
ZðsÞ > c ð3Þ
then the decision can be given as the point is non-ground[47]. Each point in point cloud iteratively follow these steps and label-ling the class.
Surface-Based filtering method were used by FUSION software [48]. The basic principle of this algorithm is to determine an indi-vidual weight for whole points.
p¼ 1 1 1þaðvigÞb 0 8 > < > :
v
i6 g g<v
i6 g þ w gþ w <v
i ð4Þwhere, a and b show the degree of weight function which is the optimum (default) value between 1 and 4 produce reliable results. The shift value is g, cell size is
v
, offset parameter is denoted w. Pri-marily, all points are weighted as a ground point and iteratively residuals are evaluated for first and second condition, the points which satisfy the condition, they are segmented as ground, the last iteration removes the nonground points from point clouds[37]. The algorithm uses the linear estimation for segmentation of points by means of residuals from normal surface. Parameter values have to be defined by user experience on different field data set. This is the one major disadvantage of this method.LasTool software is developed for LiDAR data processing and LasGround module is developed for the classification of point as ground or non-ground point. This software uses improved version of developed algorithm by Axelsson[22]. All point data sets tiled to blocks and from each data block minimum elevation points are determined and these points are chosen as a seed point. These seed points are used for a triangulated irregular network (TIN) genera-tion. All points were evaluated to distance, height difference threshold value to these TIN and classification completed[30].
Cloth Simulation filtering (CSF) has a different method than aforementioned algorithms[29]. CSF algorithm has good solution for different characteristics mentioned by Zhang et al.[29]. CSF algorithm basically Z elevation axis determined as reference, all points mirrored according to reference plane and simulated the point cloud as artificial grid surface as cloth and dropped from the top on to the mirrored point clouds. Intersection points and the threshold value provided points which has limited to move to down classified as ground, the others classified as nonground points according to the position of cloth grid. Thus, the algorithm
Fig. 5. Reference data set, a) flat area, b) steep-slope, c) dense forest, d) urban area.
was relatively easier to be used by end users with less parameters as compared to other methods.
Fig. 4presents the filtering workflow for implemented methods. Each filtering algorithms are based on different parameters. MCC parameters were examined by experiment, scale parameter and curvature limit have been chosen 0.5, 0.10 m, respectively. CSF user-defined parameters were used as: grid resolution (GR) 0.5 m, which represents resolution of point cloud neighbour parti-cles; iteration parameter set to (dT) 1000, rigidness (RI) 0.1 m, which controls the limit of the cloth displacement; and steep slope factor (ST) parameter was set to steep slope (i.e. flat, relief, steep slope). Surface-based filtering algorithm were applied by FUSION software and its parameters were chosen as G, 0, w 0.1, tolerance value 0.5 and iteration 10. These parameters were revealed as opti-mum values for this study. It may change for different field condi-tions. LasGround module of LasTool software QGIS plugin were used with a trial license. Parameters were chosen land cover type; wilderness and the other settings were used as a default value. These parameters mostly were suit and optimum to this field data set. However, parameters can be changed for the user knowledge from the obtained results.
3.1. Filtering validation
The filtering accuracy metrics were conducted by reference point data set which were selected by visually inspection of oper-ator and profile sectioning on different terrain characteristics and cover (vegetation) type, to evaluate the quality and error of the classified UAV ground points (Fig. 5). Produced orthophotos were used to visual inspection to create reference data set and validation (Fig. 6). Orthomosaics were beneficial to segment building floor which is not possible to detect from manual cross-sectional filter-ing by operator. The filterfilter-ing algorithms cannot reach to 100% fil-tering accuracy thus error types have to find out for assessment of assurance the quality of the filtering. Comparison of filtering point clouds can also be carried out based on visual cross-sectional inspection. This evaluation method were used for model quality in previous works in literature[49]. The qualitative valida-tion of all filtering methods was made of a visual test and compar-ison of 2D cross-sections and 3D point clouds with the segmented and classified datasets. For complex forestry area visualization can be applied to the presentation of the segmented point clouds cross sectional inspection from 3D to 2D on interested area with line fea-ture as a ground and nonground points colorization virtually for better recognition of class. Filtering results may also determine
by visually however cannot guarantee the real accuracy metrics of filtering methods.
Table 2
Accuracy metrics of filtering.
Algorithm/Morphology and Land Cover Type I Error Type II Error F-score Overall Accuracy Kappa Index
CSF Steep Slope 0.119 0.268 0.70 0.844 0.595
CSF Flat Area 0.058 0.123 0.861 0.925 0.809
CSF Forest 0.250 0.069 0.936 0.897 0.671
CSF Urban Area 0.073 0.249 0.842 0.806 0.601
FUSION Steep Slope 0.062 0.560 0.541 0.814 0.431
FUSION Flat Area 0.049 0.372 0.712 0.865 0.626
FUSION Forest 0.337 0.261 0.813 0.725 0.311
FUSION Urban Area 0.160 0.513 0.624 0.598 0.259
LasGround Steep Slope 0.280 0.516 0.415 0.662 0.184
LasGround Flat Area 0.201 0.491 0.494 0.722 0.302
LasGround Forest 0.795 0.143 0.838 0.733 0.067
LasGround Urban Area 0.480 0.406 0.655 0.571 0.102
MCC Steep Slope 0.195 0.353 0.579 0.766 0.419
MCC Flat Area 0.087 0.263 0.746 0.866 0.655
MCC Forest 0.770 0.068 0.882 0.798 0.199
MCC Urban Area 0.344 0.312 0.745 0.678 0.314
Bold values show the high accuracy values of the filtering methods.
In this study, to evaluate the efficiency of the filtering methods, error type I(5), error type II(6)and overall accuracy(7)F-score(9) and Kappa(8) indexes were calculated from confusion matrix [24,30,38]. Type IERROR¼ a BE ð5Þ Type IIERROR¼ b OBJ ð6Þ O
v
erallACCURACY¼ aþ b BEþ OBJ ð7Þwhere, a, ground points classified as ground point, b, non-ground points classified as non-ground points, BE and OBJ non-ground and non-ground reference point set count. The performance of the filter-ing method is evaluated by comparfilter-ing the obtained classification against a reference data.Table 2shows the quantitative validations, there, five accuracy metrics the agreement of classification between predicted and reference point data sets were evaluated. R program-ming software were used to evaluate accuracy metrics[50–52]. The Kappa coefficient formulated as follow,
j
¼N Pn i¼1mii Pn i¼1ðPiRiÞ N2Pn i¼1ðPiRiÞ ð8Þ where:i is the classification number (Ground-Non-ground)
N is the total number of the compared data with reference data set
miiis the number of diagonals of the confusion matrices Piis the total number of predicted values according to class Riis the total number of true values from reference data set
The traditional F-score is the harmonic mean of precision and recall is calculated as below:
FScore¼ 2
Precision Recall
Precisionþ Recall ð9Þ
where precision and recall can be calculated from confusion matrix by true positive, false positive and false negative values.
To evaluate of classifier performance can be assessed by the ROC and Area under the Curve (AUC) applied to data set [53]. The ROC AUC statistic for model comparison were used for classi-fication performance. Therefore, the AUC equation is defined by:
AUC¼ ðTPR FPR þ 1Þ=2 ð10Þ
where, TPR denotes true positive rate, FPR denotes false positive rate.
4. Results and discussion
The accuracy of the generated point cloud and height model was investigated by comparing UAV data with traditional survey-ing methods. RMSE and standard deviation have been analysed. By this method, systematic errors influence was reduced. GCP’s acquisition quality were validated as RMSE in project and obtained 0.012 m. These GCPs points were used to exterior orientation step and initial error was a maximum 0.02 m in adjustment of aerial tri-angulation, scale and georeferencing. Agüera-Vega et al.[54] pro-posed that GCPs must be surveyed using conventional techniques such as GPS or tachymetry. In this study area, large number of GCPs due to large study area produce enough accuracy for survey grade accuracy requirements.
Visual inspection has been made by cross-sectional analysis. Fig. 7(a) shows the raw point cloud data with Red-Green-Blue (RGB) visualisation. For a detailed examination of this cross-section sample, rectangular area is determined as a region of inter-ested area (ROI). The type II errors of the method’s results are pre-sented inFig. 7(c). As a result of cross-sectional visual analysis in forestry area, CSF algorithm has given the most reliable results when compared to the other methods. Some points still have been remained as a ground point although these points appear a non-ground. These points were presented in red ellipses inFig. 8. Afore-mentioned, these points could be removed by changing of filtering parameter values. However, during changing these parameters, type I errors may occur. For instance, parameter determination is crucial task and if too much smoothness parameters were given, then surface may lead to removing specific topographical and mor-phological regions that ground points can be classified as non-ground. Thus, it will present different geomorphological feature from the original field.
Fig. 8shows the filtering results of the included all of filtering methods with different land cover and morphological data in cross
Fig. 9. a) RGB coloured point cloud, b) mesh of point cloud c) filtered point cloud and errors type I-II on steep slope and building roof.
sectional plot. As in the figure, the lowest omission and commis-sion errors were obtained from the CSF algorithm.Fig. 9shows that the LasGround filtering algorithm were succeeded on sparse tree vegetation classified as non-ground point on the slope area. How-ever, steep slopes mostly cause the source of error for filtering algorithms. Type I errors were occurred on steep-slope area. The building’s roof where low slope angle or flat roof has cause type II errors in urban areas. Again, these errors have to be removed manually by operator.
The visual inspection may come subjective and vary based on personal experience. For more reliable results, not only visual investigations should be analysed but also quantitative analyses should be made. Quantitative results are presented inTable 2. Cal-culated errors presented that the type I error ranged from 0.049 to 0.795. The FUSION algorithm produced the lowest type I error (0.049) on flat area. The largest type I error (0.795) was obtained from the LasGround on forest area. As seen in theTable 2, the type II error ranged from 0.068 to 0.560. The lowest type II error (0.068) was achieved by the MCC on forest area. The FUSION algorithm obtained the largest type II error (0.560) on steep-slope morpho-logical area. Type I errors in dataset was rarely high, because of there are some rough surfaces generated from UAV point cloud uncertainty on ground surface, thus some of points labelled incor-rectly as non-ground points. This situation will result to lower res-olution ground surface points.
Overall accuracy metrics present the overall effectiveness of the algorithm. It shows the probability of the ground truth of the clas-sification. CSF has the highest accuracy metrics on flat area as expected from its based to physical background methodology (92.5%).
Fig. 10shows the user and producer accuracy values for filtering methods. Producer accuracy shows the probability that a value in a reference class was segmented correctly. On the other hand, user accuracy is the probability of the predicted value that is in a certain segmented class. Ground and non-ground class were separated two types of accuracy percentage. These results present that the investigated filtering methods accurate and reliable for non-ground point determination. According to the figure, user and pro-ducer accuracy, FUSION, LasGround and MCC algorithms reached to high accuracy for non-ground detection on forest region. This result occurred due to the Type I error is higher than Type II error on these methods.
F-score and AUC values distributed different at various terrain type (Fig. 11). The result of method CSF clearly shows the best pre-dictive power among filtering methods. It shows that CSF can be useful for all type of point cloud data. MCC and LasGround module give the highest accuracy values for the forest area. Method selec-tion may vary by user. The effect in the filtering error, such as ter-rain morphology, land cover, was examined. In this sense, the accuracy metrics were showed that several error sources occurred
in the filtering methods. Errors due to the dense matching points occurred from low image and texture quality such as the shadow, bright soil or rock surface and the asphalt pavement surface, etc.
CSF algorithm has achieved the highest accuracy results on all type of datasets and the classifications when compared to the other algorithms. In mountain and forestry areas, when the filter-ing algorithms compared, CSF performed comparatively well, espe-cially in dense and large vegetation areas where ground measurements are usually less owing to the image-based point clouds were not able to generate point data under the large and merged one another trees.
These analyses showed that different algorithms produce vari-ous error rate for cover types and terrain slope. Such as LasGround algorithm was not achieved to expected accuracy on urban area. However, it has high accuracy to detect non-ground (vegetation) points due to its sensitive methodology, that means, it has high and sufficient to remove non-ground points from point clouds. Trial version may have added some noise effect from software developers, it should be noticed this point. As a summary, CSF algo-rithm fits to any kind of UAV point cloud data, FUSION can be used flat and steep-slope area, LasGround is sufficient forest and flat area, MCC is suitable for flat and forest area.
The flat areas were classified the highest accuracy by all the fil-tering algorithms that shows the terrain slope degree was the depended variable for filtering accuracy. All of used filtering algo-rithms were accurate over 60% apart from LasGround in urban areas. As a result, semi-automatic filtering procedures were obtained a desirable accuracy to detect non-ground points on UAV based point cloud compared to the operator accessibility of filtering on a complex areas and time reduction with automatiza-tion. Clearly, the grade of the outlier in the point clouds has a cru-cial effect on the filtering result.
5. Conclusion
This paper presented a performance of mostly used and acces-sible open source software for filtering of LiDAR point clouds to implement on UAV based point cloud data. Most notably, this is the first study to our knowledge to investigate the performance of these filtering algorithms implemented on the UAV point clouds. Our study provided the framework for classification of UAV based point clouds as ground and non-ground object points, using semi-automatic point filtering procedures. Our results are in gen-eral agreement with previous filtering and ground classification studies[29,30,34]. The results have shown that all of the investi-gated filtering algorithms are capable of high accuracy and reliable ground data extraction feature. However, a few algorithms pro-duced better results for the high noise level, complex and rough terrain conditions. While CSF, FUSION and MCC conducted reliable filtering result on flat areas, LasGround produced on forest area with highest filtering accuracy.
Today, filtering algorithms have some user determined param-eters to fit different field data. Thus, different paramparam-eters cannot guess the perfect filtering, it has to train with small scale and then apply the estimated parameters to the study field. Therefore, dif-ferent filtering method may produce difdif-ferent results. Especially, passive sensors cannot penetrate and obtain the ground truth sur-face on dense forestry areas. Thus, geometric specifications of point clouds are not well-suited for neighbouring point calculations, slope based or height and angle-based filtering applications. When the lowest value of accuracy investigated, it caused from either low vegetation which is outliers of UAV data or rough and flat building roof type of terrain. As experienced from this study and from liter-ature, there has not been produced any solution yet for this prob-lem in point cloud classification. These errors can be corrected by merging external data from other active sensors.
In this study, UAV platform acquired images processed by SfM algorithms and produced high resolution dense data filtered as
much as possible by the state-of-the-art methods. However, den-sity variations of data were not tested for filtering accuracies. Fur-ther researches will be focus on ground LiDAR combination to acquire ground truth data with aerial based data fusion on forestry areas to improve filtering and elimination of surface errors caused from UAV data roughness and outliers. We expect that filtering procedure to increase the accuracy for further analysis which is based on bare earth surface.
Acknowledgement
This paper is a part of Mustafa Zeybek’s PhD thesis. This work was partly supported by the Selcuk University Scientific Research Projects Coordination Unit (BAP Grant No. 15401017 and 2014-OYP-055).
References
[1] B. Ruzgiene, T. Berteska, S. Gecyte, E. Jakubauskiene, V.C. Aksamitauskas, The surface modelling based on UAV Photogrammetry and qualitative estimation, Measurement 73 (2015) 619–627, https://doi.org/10.1016/j. measurement.2015.04.018.
[2] A. Akar, Evaluation of accuracy of dems obtained from uav-point clouds for different topographical areas, Int. J. Eng. Geo. 2 (3) (2017) 110–117. https://doi.org/10.26833/ijeg.329717.
[3] M. Uysal, A.S. Toprak, N. Polat, DEM generation with UAV Photogrammetry and accuracy analysis in Sahitler hill, Measurement 73 (2015) 539–543,https://doi. org/10.1016/j.measurement.2015.06.010.
[4]P. Blistan, L. Kovanic, V. Zeliznakova, J. Palkova, Using UAV photogrammetry to document rock outcrops, Acta. Montan. Slovaca 21 (2) (2016) 154–161. [5] I. Aicardi, P. Dabove, A.M. Lingua, M. Piras, Integration between TLS and UAV
photogrammetry techniques for forestry applications, Iforest 10 (1) (2017) 41– 47,https://doi.org/10.3832/ifor1780-009.
[6] A.C. Birdal, U. Avdan, T. Turk, Estimating tree heights with images from an unmanned aerial vehicle, Geomat. Nat. Haz. Risk 8 (2) (2017) 1144–1156, https://doi.org/10.1080/19475705.2017.1300608.
[7] J. Fernandez-Hernandez, D. Gonzalez-Aguilera, P. Rodriguez-Gonzalvez, J. Mancera-Taboada, Image-based modelling from unmanned aerial vehicle (uav) photogrammetry: an effective low-cost tool for archaeological applications, Archaeometry 57 (1) (2015) 128–145,https://doi.org/10.1111/ arcm.12078.
[8] T. Fernandez, J.L. Perez, J. Cardenal, J.M. Gomez, C. Colomo, J. Delgado, Analysis of landslide evolution affecting olive groves using uav and photogrammetric techniques, Remote Sens-Basel 8 (10) (2016) 837,https://doi.org/10.3390/ rs8100837.
[9] A. Izumida, S. Uchiyama, T. Sugai, Application of UAV-SfM photogrammetry and aerial LiDAR to a disastrous flood: multitemporal topographic measurement of a newly formed crevasse splay of the Kinu River, central Japan, Nat. Hazards Earth Syst. Sci. Discuss. (2017) 1–22,https://doi.org/ 10.5194/nhess-2017-42.
[10] Y. Zhang, X.X. Yuan, Y. Fang, S.Y. Chen, UAV low altitude photogrammetry for power line inspection, Isprs Int. J. Geo-Inf. 6 (1) (2017) 14,https://doi.org/ 10.3390/ijgi6010014.
[11] M. Akgül, H. Yurtseven, M. Demir, A.E. Akay, S. Gülci, T. Öztürk, _Insansız hava araçları ile yüksek hassasiyette sayısal yükseklik modeli üretimi ve ormancılıkta kullanım olanakları, FORESTIST 66 (1) (2016), https://doi.org/ 10.17099/jffiu.23976.
[12] A.S. Woodget, P.E. Carbonneau, F. Visser, I.P. Maddock, Quantifying submerged fluvial topography using hyperspatial resolution UAS imagery and structure from motion photogrammetry, Earth. Surf. Proc. Land. 40 (1) (2015) 47–64, https://doi.org/10.1002/esp.3613.
[13] J. Suh, Y. Choi, Mapping hazardous mining-induced sinkhole subsidence using unmanned aerial vehicle (drone) photogrammetry, Environ. Earth Sci. 76 (4) (2017),https://doi.org/10.1007/s12665-017-6458-3.
[14] F. Agüera-Vega, F. Carvajal-Ramírez, P. Martínez-Carricondo, J. Sánchez-Hermosilla, F.J. López, A. Mesas-Carrascosa, F.J. García-Ferrer, Pérez-Porras, Reconstruction of extreme topography from UAV structure from motion photogrammetry, Measurement 121 (2018) 127–138,https://doi.org/10.1016/ j.measurement.2018.02.062.
[15] O. Öztürk, B.B. Bilgiliog˘lu, M.F. Çelik, S.S. Bilgiliog˘lu, R. Ulug˘, _Insanız Hava Aracı (_IHA) Görüntüleri _Ile Ortofoto Üretiminde Yükseklik Ve Kamera Açısının Dog ˘rulug˘a Etkisinin Arasßtırılması, Geomatik 2 (3) (2017) 135–142,https://doi.org/ 10.29128/geomatik.327049.
[16] M. Seki, _I. Tiryakiog˘lu, M. Uysal, Farklı Veri Toplama Yöntemleriyle Yapılan Hacim Hesaplarının Karsßılasßtırılması, Geomatik 2 (2) (2017) 106–111,https:// doi.org/10.29128/geomatik.322901.
[17] J.L. Stalin R.P.C. Gnanaprakasam Volume Calculation from UAV based DEM Int. J. Eng. Res. Tech. V6 06 2017 10.17577/IJERTV6IS060076
[18]A. Ulvi, Analysis of the utility of the unmanned aerial vehicle (Uav) in volume calculation by using photogrammetric techniques, Int. J. Eng. Geo. (2018).
[19] M.W. Smith, J.L. Carrivick, D.J. Quincey, Structure from motion photogrammetry in physical geography, 40 (2) (2015) 247–275, doi: 10.1177/0309133315615805.
[20] S.P. Bemis, S. Micklethwaite, D. Turner, M.R. James, S. Akciz, S.T. Thiele, H. A. Bangash, Ground-based and UAV-based photogrammetry: a multi-scale, high-resolution mapping tool for structural geology and paleoseismology, J. Struct. Geol. 69 (2014) 163–178, https://doi.org/10.1016/j.jsg.2014. 10.007.
[21]F.A. Sesli, M. Caniberk, Estimation of the coastline changes using LIDAR, Acta. Montan. Slovaca 20 (3) (2015) 225–233.
[22]P. Axelsson, DEM generation from laser scanner data using adaptive TIN models, ISPRS Archives 33 (B4/1; PART 4) (2000) 111–117.
[23] E. Bayram, P. Frossard, E. Vural, A. Alatan, Analysis of Airborne LiDAR Point Clouds With Spectral Graph Filtering, 15, (8), 2018, 1284-1288, doi: 10.1109/ lgrs.2018.2834626.
[24]C. Serifoglu Yilmaz, O. Gungor, Comparison of the performances of ground filtering algorithms and DTM generation from a UAV-based point cloud, Geocarto Int. 33 (5) (2016) 522–537, 10.1080/10106049.2016.1265599. [25] V. Yilmaz, B. Konakoglu, C. Serifoglu, O. Gungor, E. Gökalp, Image
classification-based ground filtering of point clouds extracted from UAV-based aerial photos, Geocarto Int. 33 (3) (2016) 310–320,https://doi.org/ 10.1080/10106049.2016.1250825.
[26] N. Polat, M. Uysal, Investigating performance of Airborne LiDAR data filtering algorithms for DTM generation, Measurement 63 (2015) 61–68,https://doi. org/10.1016/j.measurement.2014.12.017.
[27] N. Polat, M. Uysal, A.S. Toprak, An investigation of DEM generation process based on LiDAR data filtering, decimation, and interpolation methods for an urban area, Measurement 75 (2015) 50–56, https://doi.org/10.1016/j. measurement.2015.08.008.
[28] F. Pirotti, A. Guarnieri, A. Vettore, Ground filtering and vegetation mapping using multi-return terrestrial laser scanning, ISPRS J. Photogramm. 76 (2013) 56–63,https://doi.org/10.1016/j.isprsjprs.2012.08.003.
[29] W.M. Zhang, J.B. Qi, P. Wan, H.T. Wang, D.H. Xie, X.Y. Wang, G.J. Yan, An easy-to-use airborne LiDAR data filtering method based on cloth simulation, Remote Sens-Basel 8 (6) (2016) 501,https://doi.org/10.3390/rs8060501. [30]A.L. Montealegre, M.T. Lamelas, J. de la Riva, A comparison of open-source
LiDAR filtering algorithms in a mediterranean forest environment, IEEE J.-Stars 8 (8) (2015) 4072–4085, 10.1109/Jstars.2015.2436974.
[31] S. Ural, J. Shan, A Min-Cut Based Filter for Airborne Lidar Data, XLI-B3, 2016. 395–401, doi: 0.5194/isprsarchives-XLI-B3-395-2016.
[32] L. Xiaoye, Airborne LiDAR for DEM generation: some critical issues, 32 (1), 2008, 31–49, doi: 10.1177/0309133308089496.
[33] K.Q. Zhang, S.C. Chen, D. Whitman, M.L. Shyu, J.H. Yan, C.C. Zhang, A progressive morphological filter for removing nonground measurements from airborne LIDAR data, Ieee T. Geosci. Remote 41 (4) (2003) 872–882, https://doi.org/10.1109/Tgrs.2003.810682.
[34] P. Rashidi, H. Rastiveis, Ground, filtering lidar data based on multi-scale analysis of height difference threshold, ISPRS Arch. XLII-4/W4 (2017) 225–229, https://doi.org/10.5194/isprs-archives-XLII-4-W4-225-2017.
[35]G. Vosselman, Slope based filtering of laser altimetry data, Arch. Photogramm. Remote Sens. 3 (2000) 935–942.
[36] T.J. Pingel, K.C. Clarke, W.A. McBride, An improved simple morphological filter for the terrain classification of airborne LIDAR data, ISPRS J. Photogramm. 77 (2013) 21–30,https://doi.org/10.1016/j.isprsjprs.2012.12.002.
[37] K. Kraus, N. Pfeifer, Determination of terrain models in wooded areas with airborne laser scanner data, ISPRS J. Photogramm. 53 (4) (1998) 193–203, https://doi.org/10.1016/S0924-2716(98)00009-4.
[38] G. Sithole, G. Vosselman, Experimental comparison of filter algorithms for bare-Earth extraction from airborne laser scanning point clouds, ISPRS J. Photogramm. 59 (1–2) (2004) 85–101, https://doi.org/10.1016/j. isrsjprs.2004.05.004.
[39] M. Oborne, Mission Planner software, 2016, http://ardupilot.org/planner/ index.html, (accessed 24.09.2018).
[40]R. Graham, A. Koh, Digital Aerial Survey: Theory and Practice, CRC Press, 2002. [41] C. Rizos, Network RTK Research and Implementation - A Geodetic Perspective
1 (2), 2002, 144–150.
[42]B. Hofmann, H. Lichtenegger, E. Wasle, GNSS-global navigation satellite systems, Springer-Verlag, 2008.
[43] T. Rosnell, E. Honkavaara, Point cloud generation from aerial image data acquired by a quadrocopter type micro unmanned aerial vehicle and a digital still camera, Sensors (Basel) 12 (1) (2012) 453–480,https://doi.org/10.3390/ s120100453.
[44] D. Turner, A. Lucieer, L. Wallace, Direct Georeferencing of Ultrahigh-Resolution UAV Imagery, Ieee T. Geosci. Remote 52 (5) (2014) 2738–2745,https://doi.org/ 10.1109/Tgrs.2013.2265295.
[45]O. Küng, C. Strecha, A. Beyeler, J.-C. Zufferey, D. Floreano, P. Fua, F. Gervaix, The accuracy of automatic photogrammetric techniques on ultra-light UAV imagery, UAV-g 2011-Unmanned Aerial Vehicle in, Geomatics (2011). [46] R.B. Rusu, Z.C. Marton, N. Blodow, M. Dolha, M. Beetz, Towards 3D Point cloud
based object maps for household environments, Robot. Auton. Syst. 56 (11) (2008) 927–941,https://doi.org/10.1016/j.robot.2008.08.005.
[47] J.S. Evans, A.T. Hudak, A multiscale curvature algorithm for classifying discrete return LiDAR in forested environments, Ieee T. Geosci. Remote 45 (4) (2007) 1029–1038,https://doi.org/10.1109/Tgrs.2006.890412.
[48] R.J. McGaughey, FUSION/LDV: Software for LIDAR Data Analysis and Visualization, 2015, pp. 186.
[49] B. Krsak, P. Blistan, A. Paulikova, P. Puskarova, L. Kovanic, J. Palkova, V. Zeliznakova, Use of low-cost UAV photogrammetry to analyze the accuracy of a digital elevation model in a case study, Measurement 91 (2016) 276–287, https://doi.org/10.1016/j.measurement.2016.05.028.
[50] J.-R. Roussel, D. Auty, lidR: Airborne LiDAR Data Manipulation and Visualization for Forestry Applications, 2017.
[51] R.C. Team, R: A Language and Environment for Statistical Computing R Foundation for Statistical Computing, 2016, https://www.R-project.org/ (accessed 24.09.2018.).
[52] J. Evans, M. Murphy, rfUtilities. R package version 2.1-3, 2018, https://cran.r-project.org/package=rfUtilities (accessed 24.09.2018.).
[53]M. Sokolova, N. Japkowicz, S. Szpakowicz, Beyond accuracy, F-Score and ROC: A family of discriminant measures for performance evaluation, Lect. Notes Comput. Sci. 4304 (2006) 1015.
[54] F. Agüera-Vega, F. Carvajal-Ramírez, P. Martínez-Carricondo, Assessment of photogrammetric mapping accuracy based on variation ground control points number using unmanned aerial vehicle, Measurement 98 (2017) 221–227, https://doi.org/10.1016/j.measurement.2016.12.002.
